hack/tools/scripts/llm-key-hunter/results/_cache/content/2shin0_arxiv-ai-mailing.json

{"repo": "2shin0/arxiv-ai-mailing", "files": {"LLM/2026-05-08.md": {"64bbd1731dde152451afe1705df222e26b5c83fa": {"t": 1785747175.3731673, "n": 263185, "h": "c0b4ce7dc3328d6b", "c": "---\nlayout: default\ntitle: \"LLM 관련 주요 논문 - 2026-05-08\"\n---\n\n{% raw %}\n# LLM 관련 주요 논문 - 2026-05-08\n\n## 1. GlazyBench: A Benchmark for Ceramic Glaze Property Prediction and Image Generation\n- **Authors**: Ziyu Zhai , Siyou Li , Juexi Shao , Juntao Yu\n- **URL**: [https://arxiv.org/abs/2605.06641](https://arxiv.org/abs/2605.06641)\n- **Abstract**:\n> Developing ceramic glazes is a costly, time-consuming process of trial and error due to complex chemistry, placing a significant burden on independent artists. While recent advances in multimodal AI offer a modern solution, the field lacks the large-scale datasets required to train these models. We propose GlazyBench, the first dataset for AI-assisted glaze design. Comprising 23,148 real glaze formulations, GlazyBench supports two primary tasks: predicting post-firing surface properties, such as color and transparency, from raw materials, and generating accurate visual representations of the glaze based on these properties. We establish comprehensive baselines for property prediction using traditional machine learning and large language models, alongside image generation benchmarks using deep generative and large multimodal models. Our experiments demonstrate promising yet challenging results. GlazyBench pioneers a new research direction in AI-assisted material design, providing a standardized benchmark for systematic evaluation.\n\n---\n\n## 2. Can RL Teach Long-Horizon Reasoning to LLMs? Expressiveness Is Key\n- **Authors**: Tianle Wang , Zhaoyang Wang , Guangchen Lan , Xinpeng Wei , Sipeng Zhang , Guanwen Qiu , Abulhair Saparov\n- **URL**: [https://arxiv.org/abs/2605.06638](https://arxiv.org/abs/2605.06638)\n- **Abstract**:\n> Reinforcement learning (RL) has been applied to improve large language model (LLM) reasoning, yet the systematic study of how training scales with task difficulty has been hampered by the lack of controlled, scalable environments. We introduce ScaleLogic, a synthetic logical reasoning framework that offers independent control over two axes of difficulty: the depth of the required proof planning (i.e., the horizon) and the expressiveness of the underlying logic. Our proposed framework supports a wide range of logics: from simple implication-only logic (\"if-then\") towards more expressive first-order reasoning with conjunction (\"and\"), disjunction (\"or\"), negation (\"not\"), and universal quantification (\"for all\"). Using this framework, we show that the RL training compute $T$ follows a power law with respect to reasoning depth $D$ ($T \\propto D^{\\gamma}$, $R^{2} > 0.99$), and that the scaling exponent $\\gamma$ increases monotonically with logical expressiveness, from $1.04$ to $2.60$. On downstream mathematics and general reasoning benchmarks, more expressive training settings yield both larger performance gains (up to $+10.66$ points) and more compute-efficient transfer compared to less expressive settings, demonstrating that what a model is trained on, not just how much it is trained, shapes downstream transfer. We further show that the power-law relationship holds across multiple RL methods, and curriculum-based training substantially improves scaling efficiency.\n\n---\n\n## 3. MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems\n- **Authors**: Zhexuan Wang , Xuebo Liu , Li Wang , Zifei Shan , Yutong Wang , Zhenxi Song , Min Zhang\n- **URL**: [https://arxiv.org/abs/2605.06623](https://arxiv.org/abs/2605.06623)\n- **Abstract**:\n> Large language model (LLM)-based Multi-agent systems (MAS) have shown promise in tackling complex collaborative tasks, where agents are typically orchestrated via role-specific prompts. While the quality of these prompts is pivotal, jointly optimizing them across interacting agents remains a non-trivial challenge, primarily due to the misalignment between local agent objectives and holistic system goals. To address this, we introduce MASPO, a novel framework designed to automatically and iteratively refine prompts across the entire system. A core innovation of MASPO is its joint evaluation mechanism, which assesses prompts not merely by their local validity, but by their capacity to facilitate downstream success for successor agents. This effectively bridges the gap between local interactions and global outcomes without relying on ground-truth labels. Furthermore, MASPO employs a data-driven evolutionary beam search to efficiently navigate the high-dimensional prompt space. Extensive empirical evaluations across 6 diverse tasks demonstrate that MASPO consistently outperforms state-of-the-art prompt optimization methods, achieving an average accuracy improvement of 2.9. We release our code at this https URL .\n\n---\n\n## 4. SkillOS: Learning Skill Curation for Self-Evolving Agents\n- **Authors**: Siru Ouyang , Jun Yan , Yanfei Chen , Rujun Han , Zifeng Wang , Bhavana Dalvi Mishra , Rui Meng , Chun-Liang Li , Yizhu Jiao , Kaiwen Zha , Maohao Shen , Vishy Tirumalashetty , George Lee , Jiawei Han , Tomas Pfister , Chen-Yu Lee\n- **URL**: [https://arxiv.org/abs/2605.06614](https://arxiv.org/abs/2605.06614)\n- **Abstract**:\n> LLM-based agents are increasingly deployed to handle streaming tasks, yet they often remain one-off problem solvers that fail to learn from past interactions. Reusable skills distilled from experience provide a natural substrate for self-evolution, where high-quality skill curation serves as the key bottleneck. Existing approaches either rely on manual skill curation, prescribe heuristic skill operations, or train for short-horizon skill operations. However, they still struggle to learn complex long-term curation policies from indirect and delayed feedback. To tackle this challenge, we propose SkillOS, an experience-driven RL training recipe for learning skill curation in self-evolving agents. SkillOS pairs a frozen agent executor that retrieves and applies skills with a trainable skill curator that updates an external SkillRepo from accumulated experience. To provide learning signals for curation, we design composite rewards and train on grouped task streams based on skill-relevant task dependencies, where earlier trajectories update the SkillRepo, and later related tasks evaluate these updates. Across multi-turn agentic tasks and single-turn reasoning tasks, SkillOS consistently outperforms memory-free and strong memory-based baselines in both effectiveness and efficiency, with the learned skill curator generalizing across different executor backbones and task domains. Further analyses show that the learned curator produces more targeted skill use, while the skills in SkillRepo evolve into more richly structured Markdown files that encode higher-level meta-skills over time.\n\n---\n\n## 5. NeuroAgent: LLM Agents for Multimodal Neuroimaging Analysis and Research\n- **Authors**: Lujia Zhong , Yihao Xia , Jianwei Zhang , Shuo huang , Jiaxin Yue , Mingyang Xia , Yonggang Shi\n- **URL**: [https://arxiv.org/abs/2605.06584](https://arxiv.org/abs/2605.06584)\n- **Abstract**:\n> Multimodal neuroimaging analysis often involves complex, modality-specific preprocessing workflows that require careful configuration, quality control, and coordination across heterogeneous toolchains. Beyond preprocessing, downstream statistical analysis and disease classification commonly require task-specific code, evaluation protocols, and data-format conventions, creating additional barriers between raw acquisitions and reproducible scientific analysis. We present NeuroAgent, an LLM-driven agentic framework that automates key preprocessing and analysis steps for heterogeneous neuroimaging data, including sMRI, fMRI, dMRI, and PET, and supports interactive downstream analysis through natural-language queries. NeuroAgent employs a hierarchical multi-agent architecture with a feedback-driven Generate-Execute-Validate engine: agents autonomously generate executable preprocessing code, detect and recover from runtime errors, and validate output integrity. We evaluate the system on 1,470 subjects pooled across all ADNI phases (CN=1,000, AD=470), where all subjects have sMRI and tabular data, with subsets also having Tau-PET (n=469), fMRI (n=278), and DTI ($n=620$). Pipeline ablation studies across multiple LLM backends show that capable models reach up to 100% intent-parsing accuracy, with the strongest backend (Qwen3.5-27B) reaching 84.8% end-to-end preprocessing step correctness. Automated recovery limits manual intervention to edge cases where human review is required via the Human-In-The-Loop interface. For Alzheimer's Disease classification using automatically preprocessed multimodal data, our agent ensemble achieves an AUC of 0.9518 with four modalities, outperforming all single-modality baselines. These results show that NeuroAgent can reduce the manual effort required for neuroimaging preprocessing and enable end-to-end automated analysis pipelines for neuroimaging research.\n\n---\n\n## 6. Process Matters more than Output for Distinguishing Humans from Machines\n- **Authors**: Milena Rmus , Mathew D. Hardy , Thomas L. Griffiths , Mayank Agrawal\n- **URL**: [https://arxiv.org/abs/2605.06524](https://arxiv.org/abs/2605.06524)\n- **Abstract**:\n> Reliable human-machine discrimination is becoming increasingly important as large language models and autonomous agents are deployed in online settings. Existing approaches evaluate whether a system can produce behavior or responses indistinguishable from those of a human, following the emphasis on outputs as a criterion for intelligence proposed by Alan Turing. Cognitive science offers an alternative perspective: evaluating the process by which behavior is produced. To test whether cognitive processes can reliably distinguish humans from machines, we introduce CogCAPTCHA30, a battery of 30 cognitive tasks designed to elicit diagnostic process-level features even when task performance is matched. Across the battery, process-level features provide stronger discriminative signal than performance metrics alone, reliably distinguishing humans from agents even under output matching (mean process-feature classifier AUC = 0.88). To evaluate agentic process differences, we compare off-the-shelf frontier agents (Claude Sonnet 4.5, GPT-5, Gemini 2.5 Pro), Centaur (a language model fine-tuned on 10.7M human decisions), and two task-specific fine-tuning approaches applied to Qwen2.5-1.5B-Instruct: action-level supervised fine-tuning (A-SFT) and process-level fine-tuning (P-SFT), which directly optimizes process features. Broad fine-tuning on human decisions improves human-like task processes relative to off-the-shelf agents, while task-specific process-level supervision further improves behavioral mimicry. However, this advantage diminishes under cross-task transfer when supervised process targets do not naturally generalize across tasks. Explicit process-level supervision can improve human behavioral mimicry, but only if appropriate task-specific process representations are available, highlighting process specification as a bottleneck for achieving human-like cognitive processes in machines.\n\n---\n\n## 7. Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors\n- **Authors**: Jonas Wiedermann-Möller , Leonard Dung , Maksym Andriushchenko\n- **URL**: [https://arxiv.org/abs/2605.06490](https://arxiv.org/abs/2605.06490)\n- **Abstract**:\n> AI systems have become increasingly capable of dangerous behaviours in many domains. This raises the question: Do models sometimes choose to violate human instructions in order to perform behaviour that is more useful for certain goals? We introduce a benchmark for measuring model propensity for instrumental convergence (IC) behaviour in terminal-based agents. This is behaviour such as self-preservation that has been hypothesised to play a key role in risks from highly capable AI agents. Our benchmark is realistic and low-stakes which serves to reduce evaluation-awareness and roleplay confounds. The suite contains seven operational tasks, each with an official workflow and a policy-violating shortcut. An eight-variant shared framework varies monitoring, instruction clarity, stakes, permission, instrumental usefulness and blocked honest paths to support inferences regarding the factors driving IC behaviour. We evaluated ten models using deterministic environment-state scorers over 1,680 samples, with trace review employed for audit and adjudication purposes. The final IC rate is 86 out of 1,680 samples (5.1%). IC behaviour is concentrated rather than uniform: two Gemini models account for 66.3% of IC cases and three tasks account for 84.9%. Conditions in which IC behaviour is indispensable for task success result in the greatest increase in the adjusted IC rate (+15.7 percentage points), whereas emphasising that task success is critical or certain framing choices do not produce comparable effects. Our findings indicate that realistic, low-nudge environments elicit IC behaviour rarely but systematically in most tested models. We conclude that it is feasible to robustly measure tendencies for dangerous behaviour in current frontier AI agents.\n\n---\n\n## 8. ReasonSTL: Bridging Natural Language and Signal Temporal Logic via Tool-Augmented Process-Rewarded Learning\n- **Authors**: Bowen Ye , Zhijian Li , Junyue Huang , Junkai Ma , Xiang Yin\n- **URL**: [https://arxiv.org/abs/2605.06483](https://arxiv.org/abs/2605.06483)\n- **Abstract**:\n> Signal Temporal Logic (STL) is an expressive formal language for specifying spatio-temporal requirements over real-valued, real-time signals. It has been widely used for the verification and synthesis of autonomous systems and cyber-physical systems. In practice, however, users often express their requirements in natural language rather than in structured STL formulas, making natural-language-to-STL translation a critical yet challenging task. Manual specification requires temporal-logic expertise and cannot scale, while prompting commercial LLM APIs incurs substantial token costs and may expose sensitive system requirements to third-party services, raising privacy concerns for industrial deployment. To address these challenges, we present \\textsc{ReasonSTL}, a tool-augmented framework that adapts local open-source language models for natural-language-to-STL generation. \\textsc{ReasonSTL} decomposes the translation process into explicit reasoning, deterministic tool calls, and structured formula construction. We further introduce process-rewarded training to supervise both tool-use trajectories and final formulas, together with \\textsc{STL-Bench}, a bilingual, computation-aware benchmark grounded in real-world signals. Experiments show that a 4B model trained with \\textsc{ReasonSTL} achieves state-of-the-art performance in both automatic metrics and human evaluations, demonstrating that \\textsc{ReasonSTL} provides a transparent, low-cost, and privacy-preserving alternative for formal specification drafting.\n\n---\n\n## 9. Patch-Effect Graph Kernels for LLM Interpretability\n- **Authors**: Ruben Fernandez-Boullon , David N. Olivieri\n- **URL**: [https://arxiv.org/abs/2605.06480](https://arxiv.org/abs/2605.06480)\n- **Abstract**:\n> Mechanistic interpretability aims to reverse-engineer transformer computations by identifying causal circuits through activation patching. However, scaling these interventions across diverse prompts and task families produces high-dimensional, unstructured datasets that are difficult to compare systematically. We propose a framework that reframes mechanistic analysis as a graph machine-learning problem by representing activation-patching profiles as patch-effect graphs over model components. We introduce three graph-construction methods: direct-influence via causal mediation, partial-correlation, and co-influence and apply graph kernels to analyze the resulting structures. Evaluating this approach on GPT-2 Small using Indirect Object Identification (IOI) and related tasks, we find that patch-effect graphs preserve discriminative structural signals. Specifically, localized edge-slot features provide higher classification accuracy than global graph-shape descriptors. A screened paired-patching validation suggests that CI and PC selected candidate edges correspond to stronger activation-influence effects than random or low-rank candidates. Crucially, by evaluating these representations against rigorous prompt-only and raw patch-effect controls, we make the evidential scope of the benchmark explicit: graph features compress structured patching signal, while raw tensors and surface cues define strong baselines that any circuit-level claim should address. Ultimately, our framework provides a compression and evaluation pipeline for comparing patching-derived structures under controlled baselines, separating robust slice-discriminative evidence from stronger task-general causal-circuit claims.\n\n---\n\n## 10. Beyond Task Success: Measuring Workflow Fidelity in LLM-Based Agentic Payment Systems\n- **Authors**: Donghao Huang , Joon Kiat Chua , Zhaoxia Wang\n- **URL**: [https://arxiv.org/abs/2605.06457](https://arxiv.org/abs/2605.06457)\n- **Abstract**:\n> LLM-based multi-agent systems are increasingly deployed for payment workflows, yet prevailing metrics, Task Success Rate (TSR) and Agent Handoff F1-Score (HF1), capture only final outcomes or unordered routing decisions. We introduce the Agentic Success Rate (ASR), a trajectory-fidelity metric that compares observed and expected agent execution sequences at the transition level, decomposing performance into Transition Recall and Transition Precision. Applied to the Hierarchical Multi-Agent System for Payments (HMASP) across 18 LLMs and 90,000 task instances, ASR reveals that 10 of 18 models systematically skip a confirmation checkpoint during payment checkout, a deviation invisible to both TSR and HF1, while 8 models enforce the checkpoint perfectly. Notably, GPT-4.1 exhibits hidden workflow shortcuts despite achieving perfect TSR and HF1, while GPT-5.2 achieves perfect ASR. Prompt refinements and deterministic routing guards guided by ASR diagnostics yield substantial TSR improvements, with gains up to +93.8 percentage points for previously struggling models, demonstrating that trajectory-level evaluation is essential in regulated domains.\n\n---\n\n## 11. PrefixGuard: From LLM-Agent Traces to Online Failure-Warning Monitors\n- **Authors**: Xinmiao Huang , Jinwei Hu , Rajarshi Roy , Changshun Wu , Yi Dong , Xiaowei Huang\n- **URL**: [https://arxiv.org/abs/2605.06455](https://arxiv.org/abs/2605.06455)\n- **Abstract**:\n> Large language model (LLM) agents now execute long, tool-using tasks where final outcome checks can arrive too late for intervention. Online warning requires lightweight prefix monitors over heterogeneous traces, but hand-authored event schemas are brittle and deployment-time LLM judging is costly. We introduce PrefixGuard, a trace-to-monitor framework with an offline StepView induction step followed by supervised monitor training. StepView induces deterministic typed-step adapters from raw trace samples, and the monitor learns an event abstraction and prefix-risk scorer from terminal outcomes. Across WebArena, $\\tau^2$-Bench, SkillsBench, and TerminalBench, the strongest PrefixGuard monitors reach 0.900/0.710/0.533/0.557 AUPRC. Using the strongest backend within each representation, they improve over raw-text controls by an average of +0.137 AUPRC. LLM judges remain substantially weaker under the same prefix-warning protocol. We also derive an observability ceiling on score-based area under the precision-recall curve (AUPRC) that separates monitor error from failures lacking evidence in the observed prefix. For finite-state audit, post-hoc deterministic finite automaton (DFA) extraction remains compact on WebArena and $\\tau^2$-Bench (29 and 20 states) but expands to 151 and 187 states on SkillsBench and TerminalBench. Finally, first-alert diagnostics show that strong ranking does not imply deployment utility: WebArena ranks well yet fails to support low-false-alarm alerts, whereas $\\tau^2$-Bench and TerminalBench retain more actionable early alerts. Together, these results position PrefixGuard as a practical monitor-synthesis recipe with explicit diagnostics for when prefix warnings translate into actionable interventions.\n\n---\n\n## 12. SCRuB: Social Concept Reasoning under Rubric-Based Evaluation\n- **Authors**: Jamelle Watson-Daniels , Himaghna Bhattacharjee , Skyler Wang , Brandon Handoko , Antonio Li , Anaelia Ovalle , Mahesh Pasupuleti , Candace Ross , Vidya Sarma , Arjun Subramonian , Karen Ullrich , Will van der Vaart , Yijing Xin , Maximilian Nickel\n- **URL**: [https://arxiv.org/abs/2605.06444](https://arxiv.org/abs/2605.06444)\n- **Abstract**:\n> While many studies of Large Language Model (LLM) reasoning capabilities emphasize mathematical or technical tasks, few address reasoning about social concepts: the abstract ideas shaping social norms, culture, and institutions. This understudied capability is essential for modern models acting as social agents, yet no systematic evaluation methodology targets it. We introduce SCRuB (Social Concept Reasoning under Rubric-Based Evaluation), a framework designed for this setting of task indeterminacy. Our goal is to measure the degree to which a model reasons about social concepts with the depth and critical rigor of a human expert. SCRuB proceeds in three phases: prompt construction from established sources, response generation by experts and models, and comparative evaluation using a five-dimensional critical thinking rubric. To enable generalization of the pipeline, we introduce a Panel of Disciplinary Perspectives ensemble validated against independent expert judges. We release SCRuBEval (n=4,711 evaluation prompts) and SCRuBAnnotations (300 expert-authored responses and 150 expert comparative judgments from 45 PhD-level scholars). Our results show that frontier models consistently outperform human experts across all five rubric dimensions. Across 1,170 pairwise comparisons, expert judges ranked a model response first in 80.8% of judgments and preferred model responses overall 74.4% of the time. Ultimately, this study provides the first expert-grounded demonstration of evaluation saturation for social concept reasoning: the single-turn exam-style format has reached its ceiling for models and humans alike.\n\n---\n\n## 13. Knowledge Graphs, the Missing Link in Agentic AI-based Formal Verification\n- **Authors**: Vaisakh Naduvodi Viswambharan , Keerthan Kopparam Radhakrishna , Deepak Narayan Gadde , Aman Kumar\n- **URL**: [https://arxiv.org/abs/2605.06434](https://arxiv.org/abs/2605.06434)\n- **Abstract**:\n> Recent advances in Large Language Models (LLMs) have enabled workflows that generate SystemVerilog Assertions (SVAs) from natural-language specifications, with the potential to accelerate Formal Verification (FV). However, high-quality assertion synthesis remains challenging because specifications are often ambiguous or incomplete and critical micro-architectural details reside in the Register Transfer Level (RTL). Many existing approaches treat the specification and RTL as loosely structured text, which weakens specification-to-RTL grounding and leads to semantic mismatches and frequent syntax failures during formal parsing and elaboration. This work addresses these limitations with a verification-centric Knowledge Graph (KG) constructed from structured Intermediate Representations (IRs) extracted from the specification, RTL, and formal-tool feedback, including syntax diagnostics, Counterexamples (CEXs), and coverage reports. The KG links requirements, design hierarchy, signals, assumptions, and properties to provide traceable, design-grounded context for generation. A multi-agent workflow queries and updates this KG to generate SVAs and to drive three refinement loops: syntax repair guided by tool diagnostics, CEX-guided correction using trace links, and coverage-directed property augmentation. Evaluation across seven benchmark designs indicates that KG-based context retrieval improves specification-to-RTL grounding and consistently produces compilable SVAs with low syntax-repair overhead. The approach achieves formal coverage ranging from 78.5% to 99.4%, though convergence exhibits design dependence with complex temporal and arithmetic reasoning remaining challenging for current LLM capabilities.\n\n---\n\n## 14. Rethinking Vacuity for OOD Detection in Evidential Deep Learning\n- **Authors**: Claire McNamara\n- **URL**: [https://arxiv.org/abs/2605.06382](https://arxiv.org/abs/2605.06382)\n- **Abstract**:\n> Vacuity, or Uncertainty Mass (UM), is commonly used as a metric to evaluate Out-of-Distribution (OOD) detection in Evidential Deep Learning (EDL). It generally involves dividing the number of classes ($K$) by the total strength of belief ($S$) of the model's predictions, where $S$ is derived from summing the Dirichlet parameters. As such, UM is sensitive to the cardinality of $K$. In particular, it is unlikely in practice that there is a linear relationship between $K$ and $S$ as $K$ and $S$ increase due to the nature of EDL (suppressing incorrectly assigned evidence). As a result, when comparing In Distribution (ID) and OOD results, it is important that $K_{\\mathrm{ID} }$ and $K_{\\mathrm{OOD} }$ are equal; something that is not always ensured in practice. We provide an empirical demonstration of how results for AUROC and AUPR can substantially differ when class cardinality between ID and OOD differs by 1, with AUROC differing by as much as 0.318 and AUPR by 0.613 for standard EDL, and AUROC by 0.360 and AUPR by 0.683 for IB-EDL. More concretely, our findings isolate an evaluation artefact: when K differs between ID and OOD, AUROC/AUPR can be artificially inflated without any change in model predictions. We further discuss the evaluation of EDL over causal language models using Multiple-Choice Question-Answer (MCQA) datasets and argue for clearer definitions of ID and OOD in this context. Our primary contribution is an empirical and theoretical demonstration that vacuity-based OOD detection in EDL-fine-tuned LLMs is highly sensitive to uncontrolled differences in evaluated class cardinality.\n\n---\n\n## 15. From Agent Loops to Deterministic Graphs: Execution Lineage for Reproducible AI-Native Work\n- **Authors**: Josh Rosen , Seth Rosen\n- **URL**: [https://arxiv.org/abs/2605.06365](https://arxiv.org/abs/2605.06365)\n- **Abstract**:\n> Large language model systems are increasingly deployed as agentic workflows that interleave reasoning, tool use, memory, and iterative refinement. These systems are effective at producing answers, but they often rely on implicit conversational state, making it difficult to preserve stable work products, isolate irrelevant updates, or propagate changes through intermediate artifacts. We introduce execution lineage: an execution model in which AI-native work is represented as a directed acyclic graph (DAG) of artifact-producing computations with explicit dependencies, stable intermediate boundaries, and identity-based replay. The goal is not to make the model a better one-shot writer, but to make evolving AI-generated work maintainable under change. We compare execution-lineage replay against loop-centric update baselines on two controlled policy-memo update tasks. In an unrelated-branch update, DAG replay preserved the final memo exactly in all runs, with zero churn and zero unrelated-branch contamination, while loop baselines regenerated the memo and frequently imported unrelated context. In an intermediate-artifact edit, all systems reflected the new constraint in the final memo, but only DAG replay achieved perfect upstream preservation, downstream propagation, unaffected-artifact preservation, and cross-artifact consistency. These results show that final answer quality and maintained-state quality are distinct. Strong loop baselines can remain competitive at producing polished final outputs when the task is a bounded synthesis/update problem and all current sources fit in context, but immediate task success can mask partial state inconsistency that may compound over future revisions. Execution lineage provides stronger guarantees about what should change, what should remain stable, and how work evolves across revisions.\n\n---\n\n## 16. A Regime Theory of Controller Class Selection for LLM Action Decisions\n- **Authors**: Zhaoyang Jiang , Zhizhong Fu , Yunsoo Kim , Jiacong Mi , Zicheng Li , Xuanqi Peng , Honghan Wu\n- **URL**: [https://arxiv.org/abs/2605.06339](https://arxiv.org/abs/2605.06339)\n- **Abstract**:\n> Deployed language and vision-language models must decide, on each input, whether to answer directly, retrieve evidence, defer to a stronger model, or abstain. Contrary to the common monotonicity intuition, greater per-input expressivity is not uniformly beneficial in finite samples: under identical strict cross-validation, different benchmarks prefer different controller classes. This reflects a finite-sample limitation of instance-level uncertainty signals, which can be exhausted at a distribution-dependent scale. We organize controllers into a nested lattice of four classes: fixed actions, partition routers, instance-level controllers, and prior-gated controllers, ordered by complexity. We prove a regime theory that turns three data-estimable bottlenecks into a class choice: how much improvement is possible beyond the best fixed action, whether there are enough samples for instance-level controllers to make reliable decisions, and how much improvement a coarse partition router can recover when instance-level signal is unreliable. The resulting Bernstein-tight threshold has a matching information-theoretic lower bound, and strict nested cross-validation provably selects a near-best class. Across SMS-Spam, HallusionBench, A-OKVQA, and FOLIO, the predicted class matches the empirical winner; the prior-gated controller wins on TextVQA when OCR tokens supply a label-free prediction-time prior. Code is available at this https URL .\n\n---\n\n## 17. Addressing Labelled Data Scarcity: Taxonomy-Agnostic Annotation of PII Values in HTTP Traffic using LLMs\n- **Authors**: Thomas Cory , Axel Küpper\n- **URL**: [https://arxiv.org/abs/2605.06305](https://arxiv.org/abs/2605.06305)\n- **Abstract**:\n> Automated privacy audits of web and mobile applications often analyse outbound HTTP traffic to detect Personally Identifiable Information (PII) leakage. However, existing learning-based detectors typically depend on scarce, manually labelled traffic and are tightly coupled to fixed label taxonomies, limiting transferability across domains and evolving definitions of PII. This paper investigates whether Large Language Models (LLMs) can support taxonomy-agnostic annotation of explicitly transmitted PII values in HTTP message bodies when the taxonomy is provided at runtime. We introduce a multi-stage LLM-based pipeline that combines deterministic pre-processing with label-level classification, targeted instance-level value annotation, and output validation. To enable controlled evaluation and exemplar-based prompting without relying on sensitive real-user captures, we further propose an LLM-based generator for synthetic HTTP traffic with manually validated, taxonomy-derived PII annotations. We evaluate the approach across three taxonomies spanning different PII domains and granularity levels. Results show that the pipeline accurately detects PII types and extracts corresponding values for concrete PII taxonomies. Overall, our findings position LLMs as a promising foundation for flexible, taxonomy-agnostic traffic annotation and for creating labelled data under evolving privacy taxonomies.\n\n---\n\n## 18. Data Language Models: A New Foundation Model Class for Tabular Data\n- **Authors**: Eda Erol , Giuliano Pezzoli , Ozer Cem Kelahmet\n- **URL**: [https://arxiv.org/abs/2605.06290](https://arxiv.org/abs/2605.06290)\n- **Abstract**:\n> Every major data modality now has a foundation model that understands it natively: text has language models, images have vision models, audio has audio models. Tabular data, the modality on which many consequential real-world AI decisions are made, does not. Every approach to tabular AI today, from gradient-boosted trees to the latest tabular foundation models, requires a preprocessing pipeline before any model can consume the data. None of them understand tabular data as a modality. We introduce the Data Language Model (DLM), the missing foundation model for tabular data. A DLM understands tables the way a language model understands sentences: natively, without serialization or preprocessing, directly from raw cell values. It is the tabular data layer on which AI models, agents, and vertical AI applications can be built, eliminating the preprocessing pipelines that currently stand between raw data and every AI system that consumes it. We present Schema-1, the first DLM: a 140M parameter model trained on more than 2.3M synthetic and real-world tabular datasets. Schema-1 outperforms gradient-boosted ensembles, AutoML stacks, and the tabular foundation models we evaluate on established row-level prediction benchmarks. On missing value reconstruction it achieves lower reconstruction error than all classical statistical methods and frontier large language models on mean performance across conditions, establishing that structural understanding of a dataset's own distributional geometry is more useful for imputation than world knowledge encoded in language. It identifies the industry sector of any unseen dataset from raw cell values alone, reliably across any domain, a task no prior tabular model can perform. It is the native tabular understanding layer that has been missing from the AI stack.\n\n---\n\n## 19. Joint Consistency: A Unified Test-Time Aggregation Framework via Energy Minimization\n- **Authors**: Yunzhen Yao , Hongye Wang , Yahong Wang , Michael C. Gastpar , Bo Jiang , Lie He\n- **URL**: [https://arxiv.org/abs/2605.06219](https://arxiv.org/abs/2605.06219)\n- **Abstract**:\n> This paper studies test-time aggregation, an approach that generates multiple reasoning traces and aggregates them into a final answer. Most existing methods rely on evaluation signals collected from candidate traces in isolation or answer frequencies, while ignoring comparative interactions among candidates. We propose Joint Consistency (JC), formulated as a constrained Ising-type energy minimization problem, where independent evaluation signals act as external fields and pairwise comparisons act as interactions. JC provides a unified framework for test-time aggregation that subsumes existing voting and weighted aggregation methods as special cases. Our construction of the interaction matrix leverages LLM-as-a-judge comparisons, and admits a theoretical interpretation under answer-level homogeneity assumptions. Moreover, we develop an efficient approximation strategy that makes interaction modeling practical for large-scale test-time aggregation. Experiments on math and code reasoning benchmarks show that JC consistently outperforms existing baselines across tasks, judge models, trace budgets, and trace-generation settings.\n\n---\n\n## 20. Beyond Fixed Benchmarks and Worst-Case Attacks: Dynamic Boundary Evaluation for Language Models\n- **Authors**: Haoxiang Wang , Da Yu , Huishuai Zhang\n- **URL**: [https://arxiv.org/abs/2605.06213](https://arxiv.org/abs/2605.06213)\n- **Abstract**:\n> Evaluating large language models (LLMs) today rests on fixed benchmarks that apply the same set of items to any model, producing ceiling and floor effects that mask capability gaps. We argue that the most informative evaluation signal lies at the boundary, where the per-prompt pass probability is near $0.5$ under random-sampling decoding, and propose Dynamic Boundary Evaluation (DBE), which actively locates each model's boundary and places it on a globally comparable difficulty scale. DBE delivers three artifacts: (i) a calibrated item bank covering safety, capability, and truthfulness, with per-item difficulty labels validated across $9$ reference LLMs; (ii) Skill-Guided Boundary Search (SGBS), a search algorithm that finds boundary items for a given target LLM using only API-level query access; and (iii) an evaluation protocol that places a new LLM on a unified ability scale and grows the evaluation set adaptively when the target falls outside the bank's coverage. We instantiate DBE on four categories spanning safety (harmful request refusal and over-refusal), capability (constrained instruction following), and truthfulness (multi-turn sycophancy resistance). The resulting evaluation covers a broader model spectrum without saturation while remaining compatible with existing datasets.\n\n---\n\n## 21. Towards Annotation-Free Validation of MLLMs: A Vision-Language Logical Consistency Metric\n- **Authors**: Ying Gu , Mei Chee Leong , Hui Li Tan , Shangbo Mao , Liyuan Li , Nancy Chen\n- **URL**: [https://arxiv.org/abs/2605.06201](https://arxiv.org/abs/2605.06201)\n- **Abstract**:\n> Dominant accuracy evaluation might reward unwarranted guessing of Large Language Models, and it might not be applicable to novel tasks for model validation without ground-truth (gt) annotation. Based on basic logic principle, we propose a novel framework to evaluate the vision-language logical consistency of MLLMs on both sufficient and necessary cause-effect relations. We define Vision-Language Logical Consistency Metric (VL-LCM) on traditional MC-VQA tests, and recent NaturalBench tests without the need for gt annotation. Through systematic experiments on representative VL benchmark MMMU and recent VL challenges like NaturalBench, we evaluated 11 recent open-source MLLMs from 4 frontier families. Our findings reveal that, despite significant progress of recent MLLMs on accuracy, logical consistency lags behind significantly. Extensive evaluations on the correlations of VL-LCM with metrics on gt, the reliability of LCM, and the relation of VL-LCM with response distribution justify the validity and applicability of VL-LCM even without gt annotation. Our findings suggest that, beyond accuracy, logical consistency could be employed for both accuracy and reliability. VL-LCM can also be employed for MLLM selection, validation, and reliable answer justification in novel tasks without gt annotation.\n\n---\n\n## 22. The Granularity Axis: A Micro-to-Macro Latent Direction for Social Roles in Language Models\n- **Authors**: Chonghan Qin , Xiachong Feng , Ziyun Song , Xiaocheng Feng , Jing Xiong , Lingpeng Kong\n- **URL**: [https://arxiv.org/abs/2605.06196](https://arxiv.org/abs/2605.06196)\n- **Abstract**:\n> Large language models (LLMs) are routinely prompted to take on social roles ranging from individuals to institutions, yet it remains unclear whether their internal representations encode the granularity of such roles, from micro-level individual experience to macro-level organizational, institutional, or national reasoning. We show that they do. We define a contrast-based Granularity Axis as the difference between mean macro- and micro-role hidden states. In Qwen3-8B, this axis aligns with the principal axis (PC1) of the role representation space at cosine 0.972 and accounts for 52.6% of its variance, indicating that granularity is the dominant geometric axis organizing prompted social roles. We construct 75 social roles across five granularity levels and collect 91,200 role-conditioned responses over shared questions and prompt variants, then extract role-level hidden states and project them onto the axis. Role projections increase monotonically across all five levels, remain stable across layers, prompt variants, endpoint definitions, held-out splits, and score-filtered subsets, and transfer to Llama-3.1-8B-Instruct. The axis is also causally relevant: activation steering along it shifts response granularity in the predicted direction, with Llama moving from 2.00 to 3.17 on a five-point macro scale under positive steering on prompts that admit local responses. The two models differ in controllability, suggesting that steering depends on each model's default operating regime. Overall, our findings suggest that social role granularity is not merely a stylistic surface feature, but a structured, ordered, and causally manipulable latent direction in role-conditioned language model behavior.\n\n---\n\n## 23. Systematic Evaluation of Large Language Models for Post-Discharge Clinical Action Extraction\n- **Authors**: Shivali Dalmia , Ananya Mantravadi , Prasanna Desikan\n- **URL**: [https://arxiv.org/abs/2605.06191](https://arxiv.org/abs/2605.06191)\n- **Abstract**:\n> The work in this paper evaluates zero-shot and few-shot large language models (LLMs) for safety-critical clinical action extraction using the CLIP discharge-note dataset, with particular emphasis on transitions of care and post-discharge patient safety. To manage the complexity of clinical documentation, we introduce a two-stage extraction framework that decomposes discharge notes, that are written in narrative form, into fine-grained, explicitly actionable clinical tasks through a staged prompting strategy. Our contributions include a systematic assessment of generative LLMs for clinical action extraction, a detailed comparison between general-purpose LLMs and task-specific supervised BERT-based models, and an analysis of annotation inconsistencies across different action categories. We show that contemporary LLMs achieve performance comparable to or exceeding supervised models on binary actionability detection, while supervised baselines retain a meaningful advantage on fine-grained multi-label category classification, despite the absence of task-specific fine-tuning and under strict data-privacy constraints. Qualitative error analysis reveals that many failures stem from misalignment between model reasoning and dataset annotation conventions, particularly in cases involving implicit clinical actions and rigid structural labeling rules. These results indicate that reported performance reflects model limitations due to lack of clinical reasoning, that is not captured by plain annotations. Labels without rationales make it impossible to distinguish clinical reasoning failures from annotation convention mismatches. Advancing clinical NLP requires reasoning-annotated datasets that document why specific spans are actionable, not merely which spans were labeled, enabling proper evaluation of model clinical understanding.\n\n---\n\n## 24. Event-Causal RAG: A Retrieval-Augmented Generation Framework for Long Video Reasoning in Complex Scenarios\n- **Authors**: Peizheng Yan , Yu Zhao , Liang Xie , Juntong Qi , Mingming Wang , Erwei Yin\n- **URL**: [https://arxiv.org/abs/2605.06185](https://arxiv.org/abs/2605.06185)\n- **Abstract**:\n> Recent large vision-language models have achieved strong performance on short- and medium-length video understanding, yet they remain inadequate for ultra-long or even infinite video reasoning, where models must preserve coherent memory over extended durations and infer causal dependencies across temporally distant events. Existing end-to-end video understanding methods are fundamentally limited by the $O(n^2)$ complexity of self-attention, while recent retrieval-augmented generation (RAG) approaches still suffer from fragmented clip-level memory, weak modeling of temporal and causal structure, and high storage and online inference costs. We present Event-Causal RAG, a lightweight retrieval-augmented framework for infinite long-video reasoning. Instead of indexing fixed-length clips, our method segments streaming videos into semantically coherent events and represents each event as a structured State-Event-State (SES) graph, capturing the event together with its surrounding state transitions. These graphs are merged into a global Event Knowledge Graph and stored in a dual-store memory that supports both semantic matching and causal-topological retrieval. On top of this memory, we design a bidirectional retrieval strategy to efficiently identify the most relevant event causal chains and provide them, together with the associated video evidence, to a backbone video foundation model for answer generation. Experiments on long-video understanding benchmarks demonstrate that Event-Causal RAG consistently outperforms strong clip-based retrieval baselines and long-context video models, particularly on questions requiring multi-event integration and causal inference across long temporal gaps, while also achieving improved memory efficiency and robust streaming performance.\n\n---\n\n## 25. Post Reasoning: Improving the Performance of Non-Thinking Models at No Cost\n- **Authors**: Richmond Sin Jing Xuan , Rishabh Bhardwaj , Soujanya Poria\n- **URL**: [https://arxiv.org/abs/2605.06165](https://arxiv.org/abs/2605.06165)\n- **Abstract**:\n> As the widespread adoption of Large Language Models (LLMs) accelerates, token consumption from intermediate reasoning traces increasingly contributes to inference latency and operational cost. Recent studies suggest that many real-world tasks require little to no explicit reasoning, with additional reasoning sometimes even degrading performance. In this work, we propose \\textbf{Post-Reasoning}, a simple yet effective approach that improves instruction-tuned models by conditioning them to justify their answers after generating the final response. By design, it enables the final answer to be obtained without additional latency or token cost, while still improving performance through simple instruction augmentation. We evaluate Post-Reasoning across \\(117\\) model--benchmark settings spanning \\(13\\) open and proprietary models, \\(4\\) model families, and \\(9\\) diverse reasoning and knowledge-intensive benchmarks, including AMC, HMMT, GSM8K, GPQA, MMLU-Pro, and BIG-Bench Hard. Post-Reasoning improves performance in over \\(88.19\\%\\) of evaluated settings, achieving a mean relative improvements of \\(17.37\\%\\). Furthermore, we propose supervised post-reason tuning, which further improves performance in over \\(91.11\\%\\) of evaluated settings, and exceeds the prompt-based post-reasoning baseline by an average of \\(8.01\\%\\), demonstrating that post-reasoning can be effectively internalized through training. Ultimately, Post-Reasoning establishes a new performance ceiling for direct-answer capabilities.\n\n---\n\n## 26. Beyond Accuracy: Policy Invariance as a Reliability Test for LLM Safety Judges\n- **Authors**: Shihao Weng , Yang Feng , Xiaofei Xie\n- **URL**: [https://arxiv.org/abs/2605.06161](https://arxiv.org/abs/2605.06161)\n- **Abstract**:\n> LLM-as-a-Judge pipelines have become the de facto evaluator for agent safety, yet existing benchmarks treat their verdicts as ground-truth proxies without checking whether the verdicts depend on the agent's behavior or merely on how the evaluation policy happens to be worded. We argue that any trustworthy safety judge must satisfy a basic property we call policy invariance, and we operationalize it as three testable principles: rubric-semantics invariance under certified-equivalent rewrites, rubric-threshold invariance under intentional strict-to-lenient shifts, and ambiguity-aware calibration so that verdict instability concentrates on genuinely ambiguous cases. Instantiating these principles as a stress-test protocol with four agent-class judges on trajectories drawn from ASSEBench and R-Judge, we surface a previously unmeasured failure mode: today's judges respond to meaningful normative shifts and to meaningless structural rewrites with comparable strength, and cannot tell the two apart. Content-preserving policy rewrites flip up to 9.1% of verdicts above baseline jitter, and 18-43% of all observed flips occur on unambiguous cases under such rewrites, so existing safety scores conflate what the agent did with how the evaluator was prompted. Beyond the diagnosis, we contribute the Policy Invariance Score and the Judge Card reporting protocol, which expose an order-of-magnitude spread in judge reliability that is invisible to accuracy-only leaderboards. We release the protocol and code so that future agent-safety benchmarks can audit their own evaluators rather than trust them by default.\n\n---\n\n## 27. Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning\n- **Authors**: Yaorui Shi , Yuxin Chen , Zhengxi Lu , Yuchun Miao , Shugui Liu , Qi GU , Xunliang Cai , Xiang Wang , An Zhang\n- **URL**: [https://arxiv.org/abs/2605.06130](https://arxiv.org/abs/2605.06130)\n- **Abstract**:\n> A persistent skill library allows language model agents to reuse successful strategies across tasks. Maintaining such a library requires three coupled capabilities. The agent selects a relevant skill, utilizes it during execution, and distills new skills from experience. Existing methods optimize these capabilities in isolation or with separate reward sources, resulting in partial and conflicting evolution. We propose Skill1, a framework that trains a single policy to co-evolve skill selection, utilization, and distillation toward a shared task-outcome objective. The policy generates a query to search the skill library, re-ranks candidates to select one, solves the task conditioned on it, and distills a new skill from the trajectory. All learning derives from a single task-outcome signal. Its low-frequency trend credits selection and its high-frequency variation credits distillation. Experiments on ALFWorld and WebShop show that Skill1 outperforms prior skill-based and reinforcement learning baselines. Training dynamics confirm the co-evolution of the three capabilities, and ablations show that removing any credit signal degrades the evolution.\n\n---\n\n## 28. Back to the Beginning of Heuristic Design: Bridging Code and Knowledge with LLMs\n- **Authors**: Nguyen Viet Tuan Kiet , Bui Dinh Pham , Dao Van Tung , Tran Cong Dao , Huynh Thi Thanh Binh\n- **URL**: [https://arxiv.org/abs/2605.06123](https://arxiv.org/abs/2605.06123)\n- **Abstract**:\n> Large language models (LLMs) have recently advanced automatic heuristic design (AHD) for combinatorial optimization (CO), where candidate heuristics are iteratively proposed, evaluated, and refined. Most existing approaches search over executable programs and distill insights from execution feedback to guide later iterations. Because this process moves from low-level implementations to high-level principles, we refer to it as a bottom-up paradigm. We argue that this view is incomplete and introduce a complementary top-down perspective: knowledge becomes the primary search object and code merely instantiates and tests it, making what is learned explicit and reusable across problems and trajectories. We formalize this shift through a statistical-learning view that exposes a distortion--compression trade-off, and instantiate it in both population-based and tree-based AHD frameworks. Across CO and tasks beyond it, knowledge-first search improves discovery efficiency, transfer, and generalization, often outperforming code-centric pipelines, while combining both strategies yields further gains. Our results suggest that progress in AHD depends on iteratively constructing and evolving interpretable hypotheses that retain value beyond a single search trajectory.\n\n---\n\n## 29. Policy-Guided Stepwise Model Routing for Cost-Effective Reasoning\n- **Authors**: Wenwen Si , Insup Lee , Osbert Bastani\n- **URL**: [https://arxiv.org/abs/2605.06116](https://arxiv.org/abs/2605.06116)\n- **Abstract**:\n> Inference-time computation has greatly enhanced the performance of large language models (LLMs) on challenging reasoning tasks, but this strategy can incur high inference costs. One solution is to route intermediate chain-of-thought (CoT) states to language models of different sizes; however, existing approaches rely on handcrafted routing strategies that limit performance, or on training large process reward models that may be infeasible in many applications. We formulate stepwise model routing as a constrained decision-making problem, which we solve by training a small control policy using reinforcement learning in conjunction with threshold calibration to tune the performance-efficiency tradeoff. We validate our method on three math benchmarks (GSM8K, MATH500, and OmniMath) on both open and closed models. Our method consistently improves the accuracy-cost tradeoff compared to handcrafted approaches, while achieving a comparable tradeoff to methods that require training large process reward models.\n\n---\n\n## 30. CrossCult-KIBench: A Benchmark for Cross-Cultural Knowledge Insertion in MLLMs\n- **Authors**: Zhen Zeng , Leijiang Gu , Feng Li , Jing Yu , Zenglin Shi\n- **URL**: [https://arxiv.org/abs/2605.06115](https://arxiv.org/abs/2605.06115)\n- **Abstract**:\n> Multimodal Large Language Models (MLLMs), trained primarily on English-centric data, frequently generate culturally inappropriate or misaligned responses in cross-cultural settings. To mitigate this, we introduce the task of cross-cultural knowledge insertion, which focuses on adapting models to specific cultural contexts while preserving their original behavior in other cultures. To facilitate research in this area, we introduce CrossCult-KIBench, a comprehensive evaluation benchmark for assessing both the effectiveness of knowledge insertion and its unintended side effects on non-target cultures. The benchmark includes 9,800 image-grounded cases covering 49 culturally relevant visual scenarios across English, Chinese, and Arabic language-culture groups. It supports evaluation in both single-insert and sequential-insert settings. We also propose Memory-Conditioned Knowledge Insertion (MCKI) as a baseline method. MCKI retrieves relevant cultural knowledge from an external memory using frozen MLLM representations, prepending matched entries as conditional prompts when applicable. Extensive experiments on CrossCult-KIBench reveal that current approaches struggle to balance effective cultural adaptation with behavioral preservation, highlighting a key challenge in developing culturally-aware MLLMs. Our work thus underscores an important research direction for developing more culturally adaptive and responsible MLLMs.\n\n---\n\n## 31. Shallow Prefill, Deep Decoding: Efficient Long-Context Inference via Layer-Asymmetric KV Visibility\n- **Authors**: Jungsuk Oh , Hyeseo Jeon , Hyunjune Ji , Kyongmin Kong , Jay-Yoon Lee\n- **URL**: [https://arxiv.org/abs/2605.06105](https://arxiv.org/abs/2605.06105)\n- **Abstract**:\n> Long-context inference in decoder-only language models is costly because long prompts are processed during Prefill, cached at every layer, and repeatedly attended to during autoregressive Decode. We introduce \\emph{Shallow Prefill, dEEp Decode} (SPEED), a phase-asymmetric KV-visibility policy that materializes non-anchor prompt-token KV states only in lower layers while keeping Decode-phase tokens full-depth. Unlike previous approaches that make upper-layer prompt KV states cheaper to store or construct, SPEED removes prefill tokens from the upper-layer Decode visibility set altogether. With a minimal BoS anchor, this simple change preserves broad benchmark quality while reducing long-context cost. In a controlled Llama-3.1-8B instruction-tuning study, SPEED using only 75\\% of layers for prefill tokens reaches 51.2 average score on OLMES-style benchmarks, compared with 51.4 for the full-depth baseline, while improving TTFT by 33\\%, TPOT by 22\\%, and reducing active KV memory by 25.0\\% at 128K context. Layer-wise diagnostics suggest that this cutoff retains the main prompt-selection and representation-stabilization regions of the full-depth model. These results show that long-context prompt tokens need not always persist as full-depth KV-cache objects when Decode-phase tokens remain full-depth.\n\n---\n\n## 32. VibeServe: Can AI Agents Build Bespoke LLM Serving Systems?\n- **Authors**: Keisuke Kamahori , Shihang Li , Simon Peter , Baris Kasikci\n- **URL**: [https://arxiv.org/abs/2605.06068](https://arxiv.org/abs/2605.06068)\n- **Abstract**:\n> For years, we have built LLM serving systems like any other critical infrastructure: a single general-purpose stack, hand-tuned over many engineer-years, meant to support every model and workload. In this paper, we take the opposite bet: a multi-agent loop that automatically synthesizes bespoke serving systems for different usage scenarios. We propose VibeServe, the first agentic loop that generates entire LLM serving stacks end-to-end. VibeServe uses an outer loop to plan and track the search over system designs, and an inner loop to implement candidates, check correctness, and measure performance on the target benchmark. In the standard deployment setting, where existing stacks are highly optimized, VibeServe remains competitive with vLLM, showing that generation-time specialization need not come at the cost of performance. More interestingly, in non-standard scenarios, VibeServe outperforms existing systems by exploiting opportunities that generic systems miss in six scenarios involving non-standard model architectures, workload knowledge, and hardware-specific optimizations. Together, these results suggest a different point in the design space for infrastructure software: generation-time specialization rather than runtime generality. Code is available at this https URL .\n\n---\n\n## 33. Visual Fingerprints for LLM Generation Comparison\n- **Authors**: Amal Alnouri , Andreas Hinterreiter , Christina Humer , Furui Cheng , Marc Streit\n- **URL**: [https://arxiv.org/abs/2605.06054](https://arxiv.org/abs/2605.06054)\n- **Abstract**:\n> Large language model (LLM) outputs arise from complex interactions among prompts, system instructions, model parameters, and architecture. We refer to specific configurations of these factors as generation conditions, each of which can bias outputs in various ways. Understanding how different generation conditions shape model behaviors is essential for tasks such as prompt design and model evaluation, yet it remains challenging due to the stochastic and open-ended nature of text generation. We present an approach to visually compare LLM outputs across generation conditions by modeling responses as collections of linguistic choices, including content, expression, and structure. We extract these choices using natural language processing pipelines and represent their distributions across repeated samples. We then visualize these distributions as visual fingerprints, enabling direct, distribution-level comparison of condition-specific tendencies. Through four usage scenarios, we demonstrate how visual fingerprints reveal consistent patterns in LLM behavior that are difficult to observe through individual responses or aggregate metrics.\n\n---\n\n## 34. Novelty-based Tree-of-Thought Search for LLM Reasoning and Planning\n- **Authors**: Leon Hamm , Zlatan Ajanovic\n- **URL**: [https://arxiv.org/abs/2605.06040](https://arxiv.org/abs/2605.06040)\n- **Abstract**:\n> Although advances such as chain-of-thought, tree-of-thought or reinforcement learning have improved the performance of LLMs in reasoning and planning tasks, they are still brittle and have not achieved human-level performance in many domains, and often suffer from high time and token costs. Inspired by the success of width-based search in planning, we explore how the concept of novelty can be transferred to language domains and how it can improve tree-of-thought reasoning. A tree of thoughts relies on building possible \"paths\" of consecutive ideas or thoughts. These are generated by repeatedly prompting an LLM. In our paper, a measurable concept of novelty is proposed that describes the uniqueness of a new node (thought) in comparison to nodes previously seen in the search tree. Novelty is estimated by prompting an LLM and making use of embedded general knowledge from pre-training. This metric can then be used to prune branches and reduce the scope of the search. Although this method introduces more prompts per state, the overall token cost can be reduced by pruning and reducing the overall tree size. This procedure is tested and compared using several benchmarks in language-based planning and general reasoning.\n\n---\n\n## 35. Strat-LLM: Stratified Strategy Alignment for LLM-based Stock Trading with Real-time Multi-Source Signals\n- **Authors**: Wenliang Huang , Zengyi Yu\n- **URL**: [https://arxiv.org/abs/2605.06024](https://arxiv.org/abs/2605.06024)\n- **Abstract**:\n> Large Language Models (LLMs) are evolving into autonomous trading agents, yet existing benchmarks often overlook the interplay between architectural reasoning and strategy consistency. We propose Strat-LLM, a framework grounded in Stratified Strategy Alignment. Operating in a live-forward setting throughout 2025, it integrates heterogeneous data including sequential prices, real-time news, and annual reports to eliminate look-ahead bias. Extensive stress tests on A-share and U.S. markets reveal: (1) reasoning-heavy models achieve peak utility in Free Mode via internal logic, whereas standard models require Strict Mode as a vital risk anchor; (2) alignment utility is regime-dependent, with Free and Guided modes capturing momentum in uptrending markets, while Strict Mode mitigates drawdowns in downtrends; (3) mid-scale models (35B) show optimal fidelity under strict constraints, whereas ultra-large models (122B) suffer an alignment tax under rigid rules but gain a performance premium in Guided Mode; (4) standard LLMs often fall into a high win-rate trap, optimizing for small gains at the expense of total returns, which can only be mitigated through deep reasoning or strict external guardrails. Project details are available at this https URL .\n\n---\n\n## 36. TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steering\n- **Authors**: Yuan Sui , Yulin Chen , Yibo Li , Xue Jiang , Yufei He , Yihong Dong , Xiaoxin He , Tianyu Gao , Bryan Hooi\n- **URL**: [https://arxiv.org/abs/2605.05980](https://arxiv.org/abs/2605.05980)\n- **Abstract**:\n> When language model agents tackle complex software engineering tasks, they often degrade over long trajectories, which we define as *agent drift*. We focus on two recurring failure modes *overthinking* and *overacting*, i.e., where the agent repeatedly reasons over information it already has, and where it issues tool calls without integrating recent observations or acquiring new evidence. In this paper, we introduce TACT (Think-Act Calibration via activation Steering), to detect and mitigate agent drift in the residual stream before it surfaces as a behavioral failure. In specific, we label trajectory steps as overthinking, overacting, or calibrated, and find that their hidden states can separate linearly along two *drift axes*, pointing from calibrated behavior toward each failure mode (AUC $\\approx$ 0.9). To mitigate agent drift, we project each step's activation onto these axes at test time and pull drifted ones back toward the calibrated region. Experiments show that TACT outperforms unsteered baselines across SWE-bench Verified, Terminal-Bench 2.0, and CLAW-Eval, lifting average resolve rate by $+5.8$ pp on Qwen3.5-27B and $+4.8$ pp on Gemma-4-26B-A4B-it while cutting steps-to-resolve by up to $26\\%$. These gains frame agent drift as a steerable direction in the residual stream, and position TACT as a viable handle for reliable long-horizon agents.\n\n---\n\n## 37. TheraAgent: Self-Improving Therapeutic Agent for Precise and Comprehensive Treatment Planning\n- **Authors**: Junkai Li , Yunghwei Lai , Tianyi Zhu , Zheng Long Lee , Weizhi Ma , Yang Liu\n- **URL**: [https://arxiv.org/abs/2605.05963](https://arxiv.org/abs/2605.05963)\n- **Abstract**:\n> Formulating a treatment plan is inherently a complex reasoning and refinement task rather than a simple generation problem. However, existing large language models (LLMs) mainly rely on one-shot output without explicit verification, which may result in rough, incomplete, and potentially unsafe treatment plans. To address these limitations, we propose TheraAgent, an agentic framework that replaces one-shot generation with an iterative generate-judge-refine pipeline. By mirroring the actual reasoning process of human experts who iteratively revise treatment plans, our framework progressively transforms coarse and incomplete drafts into precise, comprehensive, and safer therapeutic regimens. To facilitate the critical judge component, we introduce TheraJudge, a treatment-specific evaluation module integrated into the inference loop to enforce clinical standards. Experiments show TheraAgent achieves state-of-the-art results on HealthBench, leading in Accuracy and Completeness. In expert evaluations, it attains an 86% win rate against physicians, with superior Targeting and Harm Control. Moreover, the highly agreement between TheraJudge and HealthBench evaluations confirms the reliability of our framework.\n\n---\n\n## 38. ICU-Bench:Benchmarking Continual Unlearning in Multimodal Large Language Models\n- **Authors**: Yuhang Wang , Wenjie Mei , Junkai Zhang , Guangyu He , Zhenxing Niu , Haichang Gao\n- **URL**: [https://arxiv.org/abs/2605.05938](https://arxiv.org/abs/2605.05938)\n- **Abstract**:\n> Although Multimodal Large Language Models (MLLMs) have achieved remarkable progress across many domains, their training on large-scale multimodal datasets raises serious privacy concerns, making effective machine unlearning increasingly necessary. However, existing benchmarks mainly focus on static or short-sequence settings, offering limited support for evaluating continual privacy deletion requests in realistic deployments. To bridge this gap, we introduce ICU-Bench, a continual multimodal unlearning benchmark built on privacy-critical document data. ICU-Bench contains 1,000 privacy-sensitive profiles from two document domains, medical reports and labor contracts, with 9,500 images, 16,000 question-answer pairs, and 100 forget tasks. Additionally, new continual unlearning metrics are introduced, facilitating a comprehensive analysis of forgetting effectiveness, historical forgetting preservation, retained utility, and stability throughout the continual unlearning process. Through extensive experiments with representative unlearning methods on ICU-Bench, we show that existing methods generally struggle in continual settings and exhibit clear limitations in balancing forgetting quality, utility preservation, and scalability over long task sequences. These findings highlight the need for multimodal unlearning methods explicitly designed for continual privacy deletion.\n\n---\n\n## 39. Wisteria: A Unified Multi-Scale Feature Learning Framework for DNA Language Model\n- **Authors**: Weihua Wang , Haoji Li , Feilong Bao , Lei Yang , Guanglai Gao\n- **URL**: [https://arxiv.org/abs/2605.05913](https://arxiv.org/abs/2605.05913)\n- **Abstract**:\n> DNA language model aims to decipher the regulatory grammar and semantic of genomes by capturing long range dependencies in DNA sequences. Existing methods emphasize long range token interactions but often ignore the interplay between local motifs and global dependencies. In this paper, we propose Wisteria, a genomic language model that integrates multi scale feature learning within a unified framework for DNA sequence. Specifically, Wisteria augments the Mamba based architecture with gated dilated convolutions to capture local motifs and regulatory patterns, while gated multilayer perceptrons refine global dependencies. We further introduce a Fourier based attention mechanism to support frequency domain modeling, periodic extension and length generalization. Across four experimental settings with both short and long range dependencies, Wisteria demonstrates strong performance on downstream benchmarks against competitive DNA language model baselines. These results indicate that Wisteria effectively unifies local and global dependency modeling for multi scale genomic sequence analysis.\n\n---\n\n## 40. Null Space Constrained Contrastive Visual Forgetting for MLLM Unlearning\n- **Authors**: Yuhang Wang , Zhenxing Niu , Haoxuan Ji , Guangyu He , Linlin Zhang , Haichang Gao\n- **URL**: [https://arxiv.org/abs/2605.05909](https://arxiv.org/abs/2605.05909)\n- **Abstract**:\n> The core challenge of machine unlearning is to strike a balance between target knowledge removal and non-target knowledge retention. In the context of Multimodal Large Language Models (MLLMs), this challenge becomes even more pronounced, as knowledge is further divided into visual and textual modalities that are tightly intertwined. In this paper, we introduce an MLLM unlearning approach that aims to forget target visual knowledge while preserving non-target visual knowledge and all textual knowledge. Specifically, we freeze the LLM backbone and achieve unlearning by fine-tuning the visual module. First, we propose a Contrastive Visual Forgetting (CVF) mechanism to separate target visual knowledge from retained visual knowledge, guiding the representations of target visual concepts toward appropriate regions in the feature space. Second, we identify the null space associated with retained knowledge and constrain the unlearning process within this space, thereby significantly mitigating degradation in knowledge retention. Third, beyond static unlearning scenarios, we extend our approach to continual unlearning, where forgetting requests arrive sequentially. Extensive experiments across diverse benchmarks demonstrate that our approach achieves a strong balance between effective forgetting and robust knowledge retention.\n\n---\n\n## 41. Taklif.AI: LLM-Powered Platform for Interest-Based Personalized College Assignments\n- **Authors**: Zaki Kurdya , Mohammed Zuqlam , Salem Amassi , Shady Telbany , Motaz Saad\n- **URL**: [https://arxiv.org/abs/2605.05842](https://arxiv.org/abs/2605.05842)\n- **Abstract**:\n> Educators face significant challenges in creating engaging, personalized assignments that accommodate students' diverse interests and cognitive abilities. Traditional one-size-fits-all assignments frequently lead to decreased student engagement and increased reliance on unethical practices such as plagiarism. To address these challenges, we present this http URL , a platform that leverages Large Language Models (LLMs) to automatically generate personalized assignments tailored to individual student interests. Unlike existing AI-powered educational platforms that personalize based on academic performance metrics alone, this http URL incorporates students' extracurricular interests and cultural contexts into the assignment generation process through a structured prompt engineering pipeline with input and output guardrails. The platform employs a serverless architecture on AWS with this http URL , using Llama 3.3 70B as the primary LLM via LiteLLM for multi-provider load balancing and LangChain for prompt orchestration. We describe the system architecture, the prompt design methodology, and the guardrails framework that ensures output quality. Preliminary user acceptance testing with 68 participants (65 students and 3 educators) indicates positive reception, with 84% of participants rating the personalization feature as beneficial. We discuss the platform's current capabilities and limitations, and outline directions for rigorous empirical evaluation of learning outcomes.\n\n---\n\n## 42. On the Role of Language Representations in Auto-Bidding: Findings and Implications\n- **Authors**: Guanyu Zhu , Jining Luan , Hanwen Du , Xinyu Fang , Sibo Xu , Ersheng Ni , Hongji Li , Jincheng Fang , Ronghao Chen , Huacan Wang , Xuanqi Lan , Yongxin Ni , Yiqi Sun , Youhua Li\n- **URL**: [https://arxiv.org/abs/2605.05833](https://arxiv.org/abs/2605.05833)\n- **Abstract**:\n> Auto-bidding is a crucial task in real-time advertising markets, where policies must optimize long-horizon value under delivery constraints (e.g., budget and CPA). Existing methods for auto-bidding rely on compact numerical state representations: while they can implicitly capture delivery dynamics, they offer limited support for explicitly representing and controlling high-level intent, evolving feedback, and operator-style strategic guidance in real campaigns. Meanwhile, Large Language Models (LLMs) offer a powerful method for encoding semantic information, it remains unclear when LLMs help and how to integrate them without sacrificing numerical precision. Through systematic preliminary studies, we find that (1) LLM embeddings contain bidding-relevant cues yet cannot replace numerical features, and (2) gains emerge only with careful semantic--numeric integration rather than naive concatenation. Motivated by these findings, we propose \\textit{SemBid}, a novel auto-bidding framework that injects LLM-encoded semantics into offline bidding trajectories at the token level. SemBid introduces three semantic inputs: \\textit{Task}, \\textit{History}, and \\textit{Strategy}. It injects these semantics as tokens alongside numerical trajectory tokens and uses self-attention to integrate them, improving controllability and generalization across objectives. Across diverse scenarios and budget regimes, SemBid outperforms competitive baselines from offline RL and generative sequence modeling, with more consistent gains in overall performance, constraint satisfaction, and robustness. Our code is available at: \\href{ this https URL }{\\textcolor{blue}{here} }.\n\n---\n\n## 43. AGPO: Asymmetric Group Policy Optimization for Verifiable Reasoning and Search Ads Relevance at JD\n- **Authors**: Yang Xu , Kun Yao , Yiming Deng , Zheng Fang , Kai Ming Ting , Ming Pang\n- **URL**: [https://arxiv.org/abs/2605.05826](https://arxiv.org/abs/2605.05826)\n- **Abstract**:\n> Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated notable success in enhancing the reasoning performance of large language models (LLMs). However, recent studies reveal that while current RLVR methods improve sampling efficiency towards correct paths, they do not elicit fundamentally new reasoning patterns. Instead, the reasoning capability boundary of trained models often narrows compared to their base models, with base models achieving higher coverage at large sample sizes. In this work, we propose Asymmetric Group Policy Optimization (AGPO) to counteract this boundary shrinkage. AGPO adopts a negative-dominant reinforcement strategy to suppress incorrect reasoning paths, maintaining the base model's exploration capacity. For positive reinforcement, AGPO adopts a group advantage mechanism, which scales positive updates based on intra-group variance, allowing the model to focus on rare correct paths while suppressing updates from trivial paths. Our experiments on five mathematical benchmarks demonstrate that AGPO achieves state-of-the-art accuracy while consistently improving pass@$k$ performance at scale. In a large-scale industrial application for search ads relevance optimization, AGPO effectively enhances the quality of the data annotation, leading to substantial performance gains in downstream student models.\n\n---\n\n## 44. CircuitFormer: A Circuit Language Model for Analog Topology Design from Natural Language Prompt\n- **Authors**: Md Touhidul Islam , Sujan Kumar Saha , Farimah Farahmandi , Mark Tehranipoor\n- **URL**: [https://arxiv.org/abs/2605.05773](https://arxiv.org/abs/2605.05773)\n- **Abstract**:\n> Automating analog circuit design remains a longstanding challenge in Electronic Design Automation (EDA). While Transformer-based Large Language Models (LLMs) have revolutionized software code generation, their application to analog hardware design is hindered by two critical limitations: (i) the scarcity of analog design datasets containing natural language description of a design and its corresponding netlist, and (ii) the inefficiency of general-purpose tokenizers (e.g., Byte Pair Encoding (BPE)) in capturing the inherent graph structure of circuits. To bridge this gap, first, we curate the largest annotated dataset of analog circuit netlists to date, comprising 31,341 netlist-natural language description pairs across all major circuit classes. Furthermore, we propose Circuit Tokenizer (CKT), a novel circuit graph tokenizer designed to encode netlist connectivity by explicitly mining frequent subcircuits. In terms of scalability, CKT overcomes the bottleneck of prior circuit graph serialization methods where vocabulary size scales linearly with maximum number of components in the dataset, n_max, (O(n_max)); instead, CKT decouples vocabulary growth from circuit complexity, achieving a constant O(1) complexity. Empirically, CKT outperforms standard BPE on circuit topology representation, reducing sequence length by 57% and achieving a 2.3x superior compression ratio using a compact, fixed vocabulary of size 512. Leveraging this optimized tokenization, we train a circuit-specific language model, CircuitFormer, a 511M parameter encoder-decoder transformer. Our model achieves 100% syntactic correctness and an 83% functional success rate across all major analog circuit categories, outperforming state-of-the-art open-source LLMs by 10% and 14%, respectively, while requiring 240x fewer parameters. The dataset is publicly available at this https URL .\n\n---\n\n## 45. HyperLens: Quantifying Cognitive Effort in LLMs with Fine-grained Confidence Trajectory\n- **Authors**: Chengda Lu , Xiaoyu Fan , Wei Xu\n- **URL**: [https://arxiv.org/abs/2605.05741](https://arxiv.org/abs/2605.05741)\n- **Abstract**:\n> While Large Language Models (LLMs) achieve strong performance across diverse tasks, their inference dynamics remain poorly understood because of the limited resolution of existing analysis tools. In this work, we identify an intrinsic magnification mechanism in transformer architectures: deeper layers inherently magnify the small changes of layer-wise confidence, providing a fine-grained confidence trajectory. Building on this insight, we introduce HyperLens, a high-resolution probe designed to trace confidence trajectories and quantify the cognitive effort during inference. Across LLMs and datasets, HyperLens reveals a consistent divergence in confidence trajectories that separates complex from simple tasks. We abstract this pattern into a quantitative cognitive effort metric. Our analysis reveals a fundamental principle: complex tasks consistently require higher cognitive effort. Finally, we provide a mechanistic diagnosis of a common side effect of standard Supervised Fine-Tuning (SFT): it can reduce cognitive effort and consequently degrade performance on in-domain tasks.\n\n---\n\n## 46. ReFlect: An Effective Harness System for Complex Long-Horizon LLM Reasoning\n- **Authors**: Fan Huang\n- **URL**: [https://arxiv.org/abs/2605.05737](https://arxiv.org/abs/2605.05737)\n- **Abstract**:\n> Current reasoning paradigms for LLMs include chain-of-thought, ReAct, and post-hoc self-critique. These paradigms rely on two assumptions that fail on long-horizon, multi-stage tasks. As a result, errors accumulate silently across reasoning steps, leaving an open question: can a reasoning system effectively detect and recover from its own failures? We present ReFlect, a \\emph{harness} system for LLM reasoning that creates standalone error detection and recovery logic as a deterministic wrapper around the model. Controlled experiments across 6 reasoning domains show that prompt-level self-critique produces formulaic templates that flag no issues in 90 of 100 audited reflection blocks, and the investigated LLMs wrongly accept a wrong answer in at least 76\\% of cases. Our ReFlect harness achieves task success rates ranging from 41\\% on gpt-4o-mini to 56\\% on Claude Sonnet 4.5 across six models spanning small and frontier scale, with per-model gains over Direct CoT ranging from +7 pp on Qwen2.5-72B to +29 pp on Claude Sonnet 4.5, and additionally raises SWE-bench patch-structural quality from 0\\% (Direct CoT) to between 82\\% (Qwen2.5-72B) and 87\\% (GPT-4o). Notably, the harness gain is inversely proportional to the model's Direct CoT task success rate (the fitted slope is -1.69 with r=-0.76): each pp lost in baseline success rate is mechanically recovered by 1.69 pp of harness gain. We spot that adding structured reasoning state and operators yields only 15.0--18.7\\% pair-mean on Llama-3.3-70B and Qwen2.5-72B because models at this scale cannot reliably populate the state its operators require. ReFlect is model-agnostic, training-free, and operates entirely at inference time.\n\n---\n\n## 47. Knee Osteoarthritis Severity Grading Using Optimized Deep Learning and LLM-Driven Intelligent AI on Computationally Limited Systems\n- **Authors**: Dayam Nadeem , Neha , Safdar Mustafa , Adnan Alvi , Mohd Hussain\n- **URL**: [https://arxiv.org/abs/2605.05731](https://arxiv.org/abs/2605.05731)\n- **Abstract**:\n> Knee osteoarthritis (KOA) is among the musculoskeletal disorders that considerably restrict joint mobility, cause severe chronic pain and impact negatively on quality life. It is one of the persistent health issues worldwide. Generally, subjectivity and inter-observer variability undermine conventional practices and evaluation process that are adopted to address such health issues. Hence precise and timely diagnosis would be one of the effective ways for the assessment of its severity. This paper proposes an automated diagnostic approach for severity grading of KOA by blending a deep learning convolutional neural network (CNN) with a device-based inference platform powered by TensorFlow Lite. It proposes a model based on the ResNet-18 convolutional neural network. The designed model is trained on publicly available database. Through a transfer learning approach obtained knee images are first classified into five Kellgren-Lawrence (KL) grades. Further the developed model is optimised. During the training of the model test accuracy of 94.48% with stable convergence has been achieved. Subsequently the optimised model transformed into a lightweight TensorFlow Lite format, facilitating seamless deployment on resource-constrained devices. The designed model is capable enough to operate in the environment having no continuous internet connectivity. Also, an auxiliary Large Language Model (Gemini-2.0-flash) is applied to generate structured interpretive findings like potential symptoms, risk factors, and preventive majors etc. The LLM component functions as interface without influencing the classification process. The proposed model articulates the feasibility of an on-device, interpretable decision-support tools for early diagnosis and improve accessibility to Artificial Intelligence (AI)-assisted knee screening tool.\n\n---\n\n## 48. SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents\n- **Authors**: Hongcheol Cho , Ryangkyung Kang , Youngeun Kim\n- **URL**: [https://arxiv.org/abs/2605.05726](https://arxiv.org/abs/2605.05726)\n- **Abstract**:\n> As LLM agents are increasingly deployed with large libraries of reusable skills, selecting the right skill for a user request has become a critical systems challenge. In small libraries, users may invoke skills explicitly by name, but this assumption breaks down as skill ecosystems grow under tight context and latency budgets. Despite its practical importance, skill retrieval remains underexplored, with limited benchmarks and little understanding of retrieval behavior on realistic skill libraries. To address this gap, we introduce SkillRet, a large-scale benchmark for skill retrieval in LLM agents. SkillRet contains 17,810 public agent skills, organized with structured semantic tags and a two-level taxonomy spanning 6 major categories and 18 sub-categories. It provides 63,259 training samples and 4,997 evaluation queries with disjoint skill pools, enabling both benchmarking and retrieval-oriented training. Across a diverse set of retrievers, we find that skill retrieval remains far from solved: off-the-shelf models struggle on realistic large-scale skill libraries, and prior skill-retrieval models still leave substantial headroom. Task-specific fine-tuning on SkillRet substantially improves performance, improving NDCG@10 by +13.1 points over the strongest prior retriever and by +16.9 points over the strongest off-the-shelf retriever. Our analysis further suggests that these gains arise because fine-tuned models better focus on the small skill-relevant signals within long and noisy queries. These results establish SkillRet as a strong benchmark and foundation for future research on retrieval in large-scale agent systems.\n\n---\n\n## 49. Detecting Time Series Anomalies Like an Expert: A Multi-Agent LLM Framework with Specialized Analyzers\n- **Authors**: Hyeongwon Kang , Jeongseob Kim , Jinwoo Park , Pilsung Kang\n- **URL**: [https://arxiv.org/abs/2605.05725](https://arxiv.org/abs/2605.05725)\n- **Abstract**:\n> Recent studies have explored large language models for time-series anomaly detection, yet existing approaches often rely on a single general-purpose model to directly infer anomaly indices or intervals, limiting controllability, interpretability, and reliability for complex anomaly patterns. We propose SAGE (Specialized Analyzer Group for Expert-like Detection), a multi-agent framework for structured anomaly diagnosis in univariate time series. It decomposes anomaly analysis into four specialized Analyzers for point, structural, seasonal, and pattern anomalies. Each Analyzer applies family-specific numerical tools and diagnostic visualizations to generate evidence, while an evidence-grounded Detector consolidates the evidence into confidence-scored anomaly records with intervals and candidate types. A Supervisor then converts these structured records into analyst-facing diagnostic reports. SAGE further constructs synthetic in-context examples from normal-reference training segments, without using real anomalous segments or anomaly-type labels as in-context examples. Across three benchmarks, SAGE achieves the best average performance among strong ML/DL and language-model-based baselines. Ablation studies and human evaluation further show that the proposed framework improves detection reliability and the practical usefulness of diagnostic outputs.\n\n---\n\n## 50. More Is Not Always Better: Cross-Component Interference in LLM Agent Scaffolding\n- **Authors**: Ming Liu\n- **URL**: [https://arxiv.org/abs/2605.05716](https://arxiv.org/abs/2605.05716)\n- **Abstract**:\n> LLM agent systems are built by stacking scaffolding components (planning, tools, memory, self-reflection, retrieval) assuming more is better. We study cross-component interference (CCI): degradation when components interact destructively. We run a full factorial experiment over all 2^5=32 subsets of five components on HotpotQA and GSM8K with Llama-3.1-8B/70B (96 conditions, up to 10 seeds). The All-In system is consistently suboptimal: on HotpotQA, a single-tool agent surpasses All-In by 32% (F1 0.233 vs 0.177, p=0.023); on GSM8K, a 3-component subset beats All-In by 79% (0.43 vs 0.24, p=0.010). Optimal component count is task-dependent (k*=1-4) and scale-sensitive: at 70B, combinations that hurt at 8B provide gains, though All-In still trails the best subset. We fit a main-effects regression (R^2=0.916, adj-R^2=0.899, LOOCV=0.872), compute exact Shapley values, and find 183/325 submodularity violations (56.3%), showing greedy selection is unreliable. A three-body synergy among Tool Use, Self-Reflection, and Retrieval (INT_3=+0.175, 95% CI [+0.003,+0.351]) is reported as exploratory. CCI replicates across model families (Qwen2.5) and is robust to prompt paraphrasing. Our findings suggest maximally-equipped agent defaults should be replaced by task-specific subset selection via interaction-aware analysis.\n\n---\n\n## 51. Decodable but Not Corrected by Fixed Residual-Stream Linear Steering: Evidence from Medical LLM Failure Regimes\n- **Authors**: Ming Liu\n- **URL**: [https://arxiv.org/abs/2605.05715](https://arxiv.org/abs/2605.05715)\n- **Abstract**:\n> Can linearly decodable failure signals in LLM hidden states be leveraged to correct those failures? We investigate this classification-correction gap via Overthinking (OT)--a stable behavioral regime (Jaccard >= 0.81, 94% inter-annotator agreement) in medical QA where models answer correctly under resampling yet fail in extended chain-of-thought. OT is linearly decodable at 71.6% balanced accuracy (p < 10^{-16}). Yet five families of fixed linear steering (29 configurations, n=1,273) all yield Delta ~= 0, with identical null results cross-architecture (Qwen2.5-7B) and cross-domain (MMLU-STEM). Three convergent lines of evidence suggest representational entanglement: the OT direction has 85-88% overlap with task-critical computation (specificity ratio <= 0.152); non-targeted shared-direction steering damages accuracy (-12.1pp); and LEACE concept erasure damages accuracy (-3.6pp, p=0.01), while 10 random erasures produce Delta=+0.3pp. The per-instance probe-steering correlation is r=-0.002 (p=0.97). Positively, the same probe enables selective abstention (held-out AUROC=0.610, exceeding all five uncertainty baselines, p=0.009): decodable failure structure supports post-generation reliability estimation even when the fixed linear steering family cannot exploit it for correction.\n\n---\n\n## 52. Conceal, Reconstruct, Jailbreak: Exploiting the Reconstruction-Concealment Tradeoff in MLLMs\n- **Authors**: Md Farhamdur Reza , Richeng Jin , Tianfu Wu , Huaiyu Dai\n- **URL**: [https://arxiv.org/abs/2605.05709](https://arxiv.org/abs/2605.05709)\n- **Abstract**:\n> Intent-obfuscation-based jailbreak attacks on multimodal large language models (MLLMs) transform a harmful query into a concealed multimodal input to bypass safety mechanisms. We show that such attacks are governed by a \\emph{reconstruction--concealment tradeoff}: the transformed input must hide harmful intent from safety filters while remaining recoverable enough for the victim model to reconstruct the original request. Through a reconstruction analysis of three representative black-box methods, we find that existing transformations struggle to balance this tradeoff, limiting their effectiveness. In contrast, we show that character-removed variants achieve a better balance. Building on this, we propose \\emph{concealment-aware variant construction}, which greedily selects character-removed variants that are low in harmful-keyword alignment and mutually diverse, and instantiates them through five modality-aware prompting strategies. We further introduce \\emph{keyword-related distractor images} that depict the harmful keyword in diverse contexts, providing more effective auxiliary visual context than generic distractor images. Experiments across closed-source and open-source MLLMs show the proposed strategies outperform strong baselines, revealing an underexplored vulnerability: a model's own reconstruction ability can be exploited to recover hidden harmful intent and produce unsafe responses.\n\n---\n\n## 53. Resolving the bias-precision paradox with stochastic causal representation learning for personalized medicine\n- **Authors**: Peisong Zhang , Manqiang Peng , Yuxuan Wu , Pawit Phadungsaksawasdi , Wesley Yeung , Ye Zhang , Trang Nguyen , Qiang Zhang , Nan Liu , Meng Wang , Kee Yuan Ngiam , Yih-Chung Tham , Ching-Yu Cheng , Tianfan Fu , Qingyu Chen , Rosemary Ke , Chang Li , Wenzhuo Yang , Zhenghao Lu , Chunyou Lai , Yu Zhang , Sheng Zhong , Hao Deng , Dianbo Liu\n- **URL**: [https://arxiv.org/abs/2605.05706](https://arxiv.org/abs/2605.05706)\n- **Abstract**:\n> Estimating individualized treatment effects from longitudinal observational data is central to data-driven medicine, yet existing methods face a fundamental limitation: reducing confounding bias often suppresses clinically informative heterogeneity, degrading patient-specific predictions. Here, we identify this tension as a bias-precision paradox in causal representation learning and introduce sampling-based maximum mean discrepancy (sMMD), a stochastic alignment strategy that replaces global adversarial balancing with subset-level matching. We instantiate this approach in a framework for counterfactual outcome prediction with attribution-grounded interpretability. Across two large-scale ICU cohorts (n = 27,783), our framework improves accuracy under distribution shift, reducing error by up to 11.5% and substantially increasing recall in high-risk tasks. Mechanistic analyses show that sMMD selectively preserves clinically decisive variables. In human-AI evaluation, our method outperforms clinicians-in-training and large language models, and improves clinician accuracy by 14.7% while reducing decision time, enabling interpretable, real-time clinical decision support.\n\n---\n\n## 54. Knowledge-Graph Paths as Intermediate Supervision for Self-Evolving Search Agents\n- **Authors**: Huyu Wu , Jun Liu , Xiaochi Wei , Yan Gao , Yi Wu , Yao Hu\n- **URL**: [https://arxiv.org/abs/2605.05702](https://arxiv.org/abs/2605.05702)\n- **Abstract**:\n> Self-evolving search agents reduce reliance on human-written training questions by generating and solving their own search tasks. We build on Search Self-Play (SSP), a representative Proposer and Solver framework in which questions are generated and answered via multi-step search and reasoning. In practice, however, SSP faces two bottlenecks: the Proposer constructs questions from isolated answer entities without relational context, yielding many invalid or unverifiable questions in early self-play training, while the Solver receives only a binary outcome reward that discards useful signal from partially on-track search trajectories. We address both bottlenecks by reusing knowledge-graph paths as construction-derived intermediate supervision for both question construction and reward shaping. First, we ground question construction in LLM-guided knowledge-graph subgraphs, providing relational context for the Proposer. Second, we observe that constructing and solving a multi-hop question can involve overlapping intermediate entities: the factual bridges used to formulate the question may provide approximate waypoints for answering it. Exploiting this overlap, we introduce Waypoint Coverage Reward (WCR), which grants graded partial credit to incorrect Solver trajectories according to their coverage of entities on the construction path, while preserving full reward for correct answers. Across seven QA benchmarks and nine model configurations, our approach improves the average score over standard SSP in all configurations, including notable gains on multi-hop QA tasks. These results suggest that knowledge-graph paths can be reused as lightweight intermediate supervision, providing both relational guidance and process feedback without additional task-specific human annotations or manually labeled process steps.\n\n---\n\n## 55. Inference-Time Budget Control for LLM Search Agents\n- **Authors**: Zhengru Fang , Senkang Forest Hu , Zhonghao Chang , Yu Guo , Yihang Tao , Hongyao Liu , Mengzhe Ruan , Jun Huang , Yuguang Fang\n- **URL**: [https://arxiv.org/abs/2605.05701](https://arxiv.org/abs/2605.05701)\n- **Abstract**:\n> LLM search agents increasingly rely on tools at inference time, but their trajectories are often constrained by hard limits on both tool calls and generated tokens. Under such dual budgets, better answers require not only stronger models, but also explicit control over which search action should receive the next budget unit and when the accumulated evidence is sufficient to commit a final answer. We study this problem in multi-hop question answering (QA) and formulate it as two-stage inference-time budget control. At search time, our controller assigns each feasible action a task-level Value-of-Information (VOI) score, defined as an operational estimate of marginal task value per unit budget under the current search state and remaining dual budget, and uses this score to choose among retrieval, decomposition, and answer commitment. After search, a selective evidence-grounded finalizer compares the trajectory answer with a refined candidate and rewrites only when the residual error appears to be a low-risk answer-form error. Across four multi-hop QA benchmarks, three LLM backbones, and four budget levels, the method yields positive aggregate gains over four audited baselines under the same hard dual-budget protocol. Ablations show that search-time budget control, especially budget-dependent penalty, provides the main performance gain, while answer-time control helps mainly when the retrieval path is already adequate. These results suggest that inference-time budget control for LLM search agents should govern both how budget is spent during search and how the final answer is committed.\n\n---\n\n## 56. Saliency-Aware Regularized Quantization Calibration for Large Language Models\n- **Authors**: Yanlong Zhao , Xiaoyuan Cheng , Huihang Liu , Baihua He , Xinyu Zhang , Harrison Bo Hua Zhu , Wenlong Chen , Li Zeng , Zhuo Sun\n- **URL**: [https://arxiv.org/abs/2605.05693](https://arxiv.org/abs/2605.05693)\n- **Abstract**:\n> Post-training quantization (PTQ) is an effective approach for deploying large language models (LLMs) under memory and latency constraints. Most existing PTQ methods determine quantization parameters by minimizing a layer-wise reconstruction error on a predetermined calibration dataset, usually optimized via either scale search or Gram-based methods. However, from the perspective of generalization risk, existing calibration objectives of PTQ based only on empirical reconstruction error on limited or unrepresentative calibration data could move the quantized weights away from the original weights. This may cause the generalization risk to diverge, potentially degrading downstream performance. To address this issue, we propose \\emph{Saliency-Aware Regularized Quantization Calibration} (SARQC) a unified framework that augments the standard PTQ objective with a saliency-aware regularization term. This term encourages quantized weights to stay close to the original weights during calibration, leading to improved generalization during inference. SARQC integrates seamlessly into existing PTQ pipelines, enhancing both scale search and Gram-based methods under a unified formulation. Extensive experiments on dense and Mixture-of-Experts LLMs demonstrate consistent improvements in perplexity and zero-shot accuracy, without additional computational overhead during inference.\n\n---\n\n## 57. DataDignity: Training Data Attribution for Large Language Models\n- **Authors**: Xiaomin Li , Andrzej Banburski-Fahey , Jaron Lanier\n- **URL**: [https://arxiv.org/abs/2605.05687](https://arxiv.org/abs/2605.05687)\n- **Abstract**:\n> Auditing language-model outputs often requires more than judging correctness: an auditor may need to identify which source document most likely supports the knowledge expressed in a response. We study this as pinpoint provenance: given a prompt, a target-model response, and a candidate corpus, rank the documents that best support the response. We introduce FakeWiki, a controlled benchmark of 3,537 fabricated Wikipedia-style articles designed to preserve ground-truth provenance while weakening lexical shortcuts. FakeWiki includes QA probes, source-preserving paraphrases, retro-generated variants, hard anti-documents that remain topically similar while removing answer-critical facts, and five query conditions: clean prompting plus four jailbreak-inspired transformations. We evaluate seven retrieval baselines, a training-free activation-steering retrieval-fusion method, SteerFuse, and a supervised contrastive provenance ranker, ScoringModel. ScoringModel maps response and document features into a shared space and is trained with InfoNCE using in-batch, retrieval-mined, and anti-document negatives. Across nine open-weight instruction-tuned LLMs and five query conditions, ScoringModel improves mean Recall@10 from 35.0 for the strongest retrieval baseline to 52.2, without inference-time fusion, and wins 41/45 model-by-condition cells. SteerFuse is usually second-best despite requiring no supervised training, showing that activation-space evidence can efficiently complement text retrieval. On jailbreak-inspired transformed queries, ScoringModel improves Recall@10 by 15.7 points on average over the best baseline. Overall, our work shows that robust training data attribution requires evaluation settings that separate true answer support from topical or lexical resemblance.\n\n---\n\n## 58. Attractor Geometry of Transformer Memory: From Conflict Arbitration to Confident Hallucination\n- **Authors**: Qiyao Liang , Risto Miikkulainen , Ila Fiete\n- **URL**: [https://arxiv.org/abs/2605.05686](https://arxiv.org/abs/2605.05686)\n- **Abstract**:\n> Language models draw on two knowledge sources: facts baked into weights (parametric memory, PM) and information in context (working memory, WM). We study two mechanistically distinct failure modes--conflict, when PM and WM disagree and interfere; and hallucination, when the queried fact was never learned. Both produce confident output regardless, making output-based monitoring blind by design. We show both failures share a unified geometric account. In the hidden-state space of autoregressive generation, learned facts form attractor basins. Conflict is basin competition: WM disrupts convergence to the correct basin without raising output entropy. Hallucination is basin absence: the hidden state drifts freely when no memorized basin exists. The frozen LM head, designed for next-token prediction, cannot distinguish these cases and fires confidently either way. We verify this account in a controlled synthetic task--entity identifiers mapped to unique codes with PM installed via LoRA adapters--where ground truth is exact and component roles can be causally isolated through targeted adapter placement. Geometric margin--the hidden state's distance to the nearest memorized basin--reads this geometry directly and separates correct recall from hallucination far more cleanly than output entropy, with zero false refusals where entropy-based detection cannot avoid rejecting the vast majority of correct outputs. The separation holds on natural-language factual queries from the pretrained model with no adaptation, confirming attractor geometry is structural rather than a fine-tuning artifact. The fraction of confident hallucinations follows a scaling law $C = \\exp(-c/\\bar\\Delta)$, growing with scale even as overall error rates fall. Hidden states reliably encode epistemic state; the frozen output head systematically erases it--and this erasure worsens with scale.\n\n---\n\n## 59. Large Vision-Language Models Get Lost in Attention\n- **Authors**: Gongli Xi , Ye Tian , Mengyu Yang , Huahui Yi , Liang Lin , Xiaoshuai Hao , Kun Wang , Wendong Wang\n- **URL**: [https://arxiv.org/abs/2605.05668](https://arxiv.org/abs/2605.05668)\n- **Abstract**:\n> Despite the rapid evolution of training paradigms, the decoder backbone of large vision--language models (LVLMs) remains fundamentally rooted in the residual-connection Transformer architecture. Therefore, deciphering the distinct roles of internal modules is critical for understanding model mechanics and guiding architectural optimization. While prior statistical approaches have provided valuable attribution-based insights, they often lack a unified theoretical basis. To bridge this gap, we propose a unified framework grounded in information theory and geometry to quantify the geometric and entropic nature of residual updates. Applying this unified framework reveals a fundamental functional decoupling: Attention acts as a subspace-preserving operator focused on reconfiguration, whereas FFNs serve as subspace-expanding operators driving semantic innovation. Strikingly, further experiments demonstrate that replacing learned attention weights with predefined values (e.g., Gaussian noise) yields comparable or even superior performance across a majority of datasets relative to vanilla models. These results expose severe misallocation and redundancy in current mechanisms, suggesting that state-of-the-art LVLMs effectively ``get lost in attention'' rather than efficiently leveraging visual context.\n\n---\n\n## 60. Retrieval-Conditioned Topology Selection with Provable Budget Conservation for Multi-Agent Code Generation\n- **Authors**: Abhijit Talluri , Pujith Anne , Bhagavan Choudary Pendiyala , Raghavendra Chilukuri\n- **URL**: [https://arxiv.org/abs/2605.05657](https://arxiv.org/abs/2605.05657)\n- **Abstract**:\n> Multi-agent LLM systems for code generation face a fundamental routing problem: the optimal orchestration topology depends on the structural complexity of the code under modification, yet existing systems select topologies without consulting the codebase. We present Retrieval-Guided Adaptive Orchestration (RGAO), an architecture that closes this loop by extracting a structural complexity vector from a hierarchical code index before selecting the orchestration topology. RGAO operates within Code-Agent, a multi-agent framework whose sub-agents are governed by formal contracts with six-dimensional budget vectors. Our headline contribution is the composition of two previously separate lines of work -- complexity-conditioned LLM routing and formal resource algebras -- yielding a property neither admits alone: provable budget conservation under retrieval-conditioned dynamic topology selection. Concretely we contribute: (1) a complexity-conditioned topology router that reduces proxy-measured misrouting from 30.1% to 8.2%; (2) a budget algebra with a structural-induction conservation theorem; and (3) a hierarchical code retrieval engine. Empirical evaluation demonstrates sub-millisecond DAG construction and linear tree-index scalability.\n\n---\n\n## 61. Text-Graph Synergy: A Bidirectional Verification and Completion Framework for RAG\n- **Authors**: Jiarui Zhong , Hong Cai Chen\n- **URL**: [https://arxiv.org/abs/2605.05643](https://arxiv.org/abs/2605.05643)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) has become a core paradigm for enhancing factual grounding and multi-hop reasoning in Large Language Models (LLMs). Traditional text-based RAG often retrieves logically irrelevant pseudo-evidence, while graph-based RAG is frequently hindered by search-time pruning, which may discard potentially valid reasoning paths. Existing hybrid approaches primarily adopt simple evidence concatenation or unidirectional enhancement, which fails to address the fundamental \"Information Island\" problem caused by asymmetric reasoning flows between unstructured text and structured graphs. We propose \\textbf{TGS-RAG}, a unified framework for \\textbf{T}ext-\\textbf{G}raph \\textbf{S}ynergistic enhancement. TGS-RAG introduces a bidirectional mechanism: (i) a \\textbf{Graph-to-Text} channel that employs a Global Voting strategy from visited graph nodes to re-rank and refine textual evidence, filtering out semantic noise; and (ii) a \\textbf{Text-to-Graph} channel that utilizes the \\textbf{Memory-based Orphan Entity Bridging} algorithm. This algorithm utilizes textual cues to proactively resurrect valid but previously pruned reasoning paths from the search history without additional database overhead. Experimental results on multiple multi-hop reasoning benchmarks demonstrate that TGS-RAG significantly outperforms state-of-the-art baselines, achieving a superior balance between retrieval precision and computational efficiency.\n\n---\n\n## 62. Prober.ai: Gated Inquiry-Based Feedback via LLM-Constrained Personas for Argumentative Writing Development\n- **Authors**: Ran Bi , Shiyao Wei , Yuanyiyi Zhou\n- **URL**: [https://arxiv.org/abs/2605.05598](https://arxiv.org/abs/2605.05598)\n- **Abstract**:\n> The proliferation of large language models (LLMs) in educational settings has paradoxically undermined the cognitive processes they purport to support. Students increasingly outsource critical thinking to AI assistants that generate polished text on demand, resulting in measurable cognitive debt and diminished argumentative reasoning skills. We present this http URL , a web-based writing environment that inverts the conventional AI-tutoring paradigm: rather than generating or rewriting student text, the system constrains an LLM (Gemini 3 Flash Preview) through persona-specific system prompts and structured JSON output schemas to produce only targeted, inquiry-based questions about argumentative weaknesses. A two-phase interaction architecture -- Challenge and Unlock -- implements a pedagogical friction mechanism whereby revision suggestions are gated behind mandatory student reflection. The system's design is grounded in Toulmin's argumentation theory, research on peer feedforward questioning mechanisms, and evidence on AI-supported feedback in writing instruction. A functional prototype was developed in 36 hours during the NY EdTech Hackathon (March 2026), where it was awarded second place. We describe the system architecture, the prompt engineering methodology for constraining LLM output to pedagogically aligned JSON schemas, and discuss implications for scalable, cognition-preserving AI integration in writing education.\n\n---\n\n## 63. Causal Probing for Internal Visual Representations in Multimodal Large Language Models\n- **Authors**: Zehao Deng , Tianjie Ju , Zheng Wu , Liangbo He , Jun Lan , Huijia Zhu , Weiqiang Wang , Zhuosheng Zhang\n- **URL**: [https://arxiv.org/abs/2605.05593](https://arxiv.org/abs/2605.05593)\n- **Abstract**:\n> Despite the remarkable success of Multimodal Large Language Models (MLLMs) across diverse tasks, the internal mechanisms governing how they encode and ground distinct visual concepts remain poorly understood. To bridge this gap, we propose a causal framework based on activation steering to actively probe and manipulate internal visual representations. Through systematic intervention across four visual concept categories, our results reveal a divergence in concept encoding: entities exhibit distinct localized memorization, whereas abstract concepts are globally distributed across the network. Critically, this divergence uncovers a mechanistic driver of scaling laws: increasing model depth is indispensable for encoding distributed and complex abstract concepts, whereas entity localization remains remarkably invariant to scale. Furthermore, reverse steering uncovers that blocking explicit output triggers a surge in latent activations, exposing a compensatory mechanism between perception and generation. Finally, extending our analysis to visual reasoning, we expose a disconnect between perception and reasoning although MLLMs successfully recognize geometric relations, they treat them merely as static visual features, failing to trigger the procedural execution necessary for abstract problem-solving.\n\n---\n\n## 64. Belief Memory: Agent Memory Under Partial Observability\n- **Authors**: Junfeng Liao , Qizhou Wang , Jianing Zhu , Bo Du , Rui Yan , Xiuying Chen\n- **URL**: [https://arxiv.org/abs/2605.05583](https://arxiv.org/abs/2605.05583)\n- **Abstract**:\n> LLM agents that operate over long context depend on external memory to accumulate knowledge over time. However, existing methods typically store each observation as a single deterministic conclusion (e.g., inferring \"API~X failed\" from temporary errors), even though such observations are inherently partial and potentially ambiguous. By committing to one conclusion and discarding uncertainty, these methods introduce self-reinforcing error: the agent acts on the stored conclusion, never revisits alternatives, and reinforces the conclusion over time. To address this issue, we propose BeliefMem, which shifts the memory paradigm from committing to a single conclusion per observation to retaining multiple candidate conclusions with their probabilities. Concretely, BeliefMem stores the candidate conclusions as separate memory entries, each carrying a probability that is updated via Noisy-OR rules as new observations arrive. At retrieval, all candidates surface together with their probabilities, keeping alternatives visible to the agent. Since each conclusion in memory retains its probability, BeliefMem preserves the uncertainty that the deterministic paradigm discards, enabling the agent to act with high confidence on well-evidenced knowledge while retaining the capacity to update its confidence when new evidence arrives. Empirical evaluations on LoCoMo and ALFWorld benchmarks show that, even with limited data, BeliefMem achieves the best average performance, remarkably outperforming well-known baselines. More broadly, such probabilistic memory produces substantial gains and explores a new direction for agent memory in partially observable environments.\n\n---\n\n## 65. AlphaCrafter: A Full-Stack Multi-Agent Framework for Cross-Sectional Quantitative Trading\n- **Authors**: Yishuo Yuan , Jiayi Sheng , Sirui Zeng , Jiaqi Wang , Jiaheng Liu\n- **URL**: [https://arxiv.org/abs/2605.05580](https://arxiv.org/abs/2605.05580)\n- **Abstract**:\n> Financial markets are inherently non-stationary, driven by complex interactions among macroeconomic regimes, microstructural frictions, and behavioral dynamics. Building quantitative strategies that remain profitable demands the continuous coupling of factor discovery, regime-adaptive selection, and risk-constrained execution. Prevailing approaches, however, optimize these components under static or isolated assumptions. Factor mining frameworks typically treat alpha discovery as a one-time search process, implicitly assuming that factor efficacy persists across market regimes. Execution-oriented systems often adopt role-playing agent architectures that simulate anthropomorphic trading committees, introducing behavioral noise rather than systematic rationality. Consequently, a fully automated, rationality-driven framework unifying a coherent quantitative pipeline remains absent. We introduce AlphaCrafter, a full-stack multi-agent framework that closes this gap through a continuously adaptive factor-to-execution pipeline, designed to track and respond to evolving market conditions without manual intervention. AlphaCrafter operates via three specialized agents: a Miner that continuously expands the factor pool via LLM-guided search, a Screener that assesses prevailing market conditions to construct regime-conditioned factor ensembles, and a Trader that translates these ensembles into quantitative strategies under explicit risk constraints. Together, these three agents form a closed-loop cross-sectional trading system that adapts holistically to evolving market dynamics. Extensive experiments on CSI 300 and S&P 500 demonstrate that AlphaCrafter consistently outperforms state-of-the-art baselines in risk-adjusted returns while exhibiting the lowest cross-trial variance, confirming that integrated and adaptive factor-to-execution design yields robust trading performance.\n\n---\n\n## 66. Nonsense Helps: Prompt Space Perturbation Broadens Reasoning Exploration\n- **Authors**: Langlin Huang , Chengsong Huang , Jinyuan Li , Donghong Cai , Yuyi Yang , Jiaxin Huang\n- **URL**: [https://arxiv.org/abs/2605.05566](https://arxiv.org/abs/2605.05566)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards, particularly Group Relative Policy Optimization (GRPO), has significantly advanced the reasoning capabilities of Large Language Models (LLMs). However, in complex tasks, GRPO frequently suffers from the ``zero-advantage problem'': when all sampled rollouts for a query fail, the relative advantage collapses to zero. Consequently, the model loses effective training signals for these questions, wasting the training data and computational budget. While simply increasing the sampling budget for these questions is a common remedy, the static sampling policy inherently constrains reasoning exploration, limiting the success rate. In this paper, we propose Lorem Perturbation for Exploration (LoPE), a simple yet effective training framework to break this exploration bottleneck. We posit that task-irrelevant prompt-space perturbations can shift the model's output distribution enough to unlock orthogonal reasoning pathways for hard questions. Specifically, LoPE prepends sequences stochastically assembled from Lorem Ipsum vocabulary (a pseudo-Latin placeholder text) to the prompts before resampling. Experiments across 1.7B, 4B, and 7B models demonstrate that LoPE significantly outperforms resampling with the original prompts. Further analysis reveals that other Latin-based random sequences with low perplexity are also effective perturbations. Our results establish LoPE as a strong baseline for broadening exploration in LLM reinforcement learning.\n\n---\n\n## 67. SPARK: Self-Play with Asymmetric Reward from Knowledge Graphs\n- **Authors**: Hyobin Park , Taeseop Kim , Dong-Geol Choi\n- **URL**: [https://arxiv.org/abs/2605.05546](https://arxiv.org/abs/2605.05546)\n- **Abstract**:\n> Self-play reinforcement learning has shown strong performance in domains with formally verifiable structure, such as mathematics and coding, where both problem generation and reward computation can be grounded in explicit rules. Extending this paradigm to scientific literature is more challenging: the relationships among multi-modal elements within and across documents are rarely made explicit in text, which makes automatic generation of relational reasoning questions difficult and weakens the reliability of reward signals. We propose SPARK (Self-Play with Asymmetric Reward from Knowledge Graphs), a framework that automatically constructs a unified knowledge graph (KG) from multi-document scientific literature and uses it as the structural basis for self-play. KG paths over multimodal nodes serve as a source for generating relational reasoning questions, and structured facts stored in the KG provide a basis for verifiable reward computation. A single small vision-language model (sVLM) alternates between Proposer and Solver roles under information asymmetry against a fixed KG, a design that we believe can be naturally extended toward online adaptation in future work. We evaluate SPARK on public benchmarks and a self-constructed cross-document multi-hop QA dataset. Results show that SPARK consistently outperforms flat-corpus-based self-play baselines, and the performance gap widens as hop count increases, suggesting that KG-structure grounding contributes to relational multi-hop reasoning beyond what unstructured corpus grounding can provide.\n\n---\n\n## 68. AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases\n- **Authors**: Susheel Suresh , Hazel Mak , Shangpo Chou , Fred Kroon , Sahil Bhatnagar\n- **URL**: [https://arxiv.org/abs/2605.05538](https://arxiv.org/abs/2605.05538)\n- **Abstract**:\n> We present AgenticRAG, a practical agentic harness for retrieval and analysis over enterprise knowledge bases. Standard RAG pipelines place significant burden of grounding on the search stack, constraining the language model to a fixed candidate set chosen deep in the retrieval process. Our approach reduces this overdependence by layering a lightweight harness on top of existing enterprise search infrastructure, equipping a reasoning LLM with search, find, open, and summarize tools enabling the model to iteratively retrieve information, navigate within documents, and analyze evidence autonomously. On three open benchmarks we observe substantial gains: $49.6\\%$ recall@1 on BRIGHT (+21.8 pp over the best embedding baseline), 0.96 factuality on WixQA ($+13\\%$ relative improvement), and $92\\%$ answer correctness on FinanceBench--within 2 pp of oracle access to true evidence. Ablation studies show that the most significant factor is the shift from single-shot retrieval to agentic tool use ($5.9\\times$ improvement), while multi-query search and in-document navigation contribute to both quality and efficiency. We present various design choices in our agentic harness that were informed by pre-production deployments. Our results demonstrate its suitability for real-world enterprise production environments.\n\n---\n\n## 69. FoodCHA: Multi-Modal LLM Agent for Fine-Grained Food Analysis\n- **Authors**: Woojin Lee , Pranav Mekkoth , Ye Tian , Onat Gungor , Tajana Rosing\n- **URL**: [https://arxiv.org/abs/2605.05499](https://arxiv.org/abs/2605.05499)\n- **Abstract**:\n> The widespread adoption of camera-equipped mobile devices and wearables has enabled convenient capture of meal images, making food recognition a key component for real time dietary monitoring. However, real-world food images present challenges due to high intra-class similarity and the frequent presence of multiple food items within a single image. While deep learning models achieve strong performance in coarse grained classification, they often struggle to capture fine-grained attributes such as cooking style. Moreover, open-ended generation in modern vision-language models can produce non-canonical labels, limiting their practical deployment. We propose FoodCHA, a multimodal agentic framework that reformulates food recognition as a hierarchical decision-making process. By progressively anchoring predictions, FoodCHA guides subcategory identification using high-level categories and guides cooking style recognition using subcategories, improving semantic consistency and attribute-level discrimination. To ensure practical deployability, FoodCHA utilizes the compact Moondream-2B vision language model, which provides strong reasoning capability while maintaining lower computational and memory overhead. Experiments on FoodNExTDB show that FoodCHA outperforms Food-Llama-3.2-11B by 13.8% and 38.2% in category and subcategory recognition precision, respectively, and achieves a striking 153.2% improvement in cooking style classification precision.\n\n---\n\n## 70. FinRAG-12B: A Production-Validated Recipe for Grounded Question Answering in Banking\n- **Authors**: Denys Katerenchuk , Pablo Duboue , Keelan Evanini , David Gondek , Nithin Govindugari , Olivier Allauzen , Joshua Baptiste , David J More , Joshua Schechter\n- **URL**: [https://arxiv.org/abs/2605.05482](https://arxiv.org/abs/2605.05482)\n- **Abstract**:\n> Large language models (LLMs) are rapidly being adopted across various domains. However, their adoption in banking industry faces resistance due to demands for high accuracy, regulatory compliance, and the need for verifiable and grounded responses. We present a unified, data-efficient framework for training grounded domain-specific LLMs that optimizes answer quality, citation grounding, and calibrated refusal under real-world deployment constraints. First, we describe a data generation pipeline that combines LLM-as-a-Judge filtering, citation annotation, and curriculum learning with only 143M tokens. The resulting 12B model achieves high answer quality outperforming GPT-4.1 on citation grounding, with a modest citation tradeoff versus the untuned base. Second, we propose a calibrated refusal mechanism: training on 22% unanswerable examples yield a 12% \"I don't know\" rate, substantially improving over the base model's unsafe 4.3% rate while avoiding GPT-4.1's over-refusal (20.2%). Third, we present an end-to-end methodology spanning from data curation to quantized serving. The system is deployed at 40+ financial institutions, achieving a 7.1 percentage point improvement in query resolution (p < 0.001). Additionally, the model delivers 3-5x faster responses at 20-50x lower cost compared to GPT-4.1.\n\n---\n\n## 71. LANTERN: LLM-Augmented Neurosymbolic Transfer with Experience-Gated Reasoning Networks\n- **Authors**: Mahyar Alinejad , Yue Wang , Amrit Singh Bedi , George Atia\n- **URL**: [https://arxiv.org/abs/2605.05478](https://arxiv.org/abs/2605.05478)\n- **Abstract**:\n> Transfer learning in reinforcement learning (RL) seeks to accelerate learning in new tasks by leveraging knowledge from related sources. Existing neurosymbolic transfer methods, however, typically rely on manually specified task automata, assume a single source task, and use fixed knowledge-integration mechanisms that cannot adapt to varying source relevance. We propose LANTERN, a unified framework for multi-source neurosymbolic transfer that addresses these limitations through three components: (i) deterministic finite automata generated from natural language task descriptions using large language models, (ii) semantic embedding-based aggregation of multiple source policies weighted by cross-task similarity, and (iii) adaptive teacher-student gating based on temporal-difference error and semantic uncertainty. Across domains spanning resource management, navigation, and control, LANTERN achieves 40-60% improvements in sample efficiency over existing baselines while remaining robust to poorly aligned sources. These results demonstrate that multi-source, adaptively weighted neurosymbolic transfer can improve scalability and robustness in symbolic RL settings.\n\n---\n\n## 72. Agentic Discovery of Exchange-Correlation Density Functionals\n- **Authors**: Titouan Duston , Jiashu Liang , Yuanheng Wang , Weihao Gao , Xuelan Wen , Nan Sheng , Weiluo Ren , Yang Sun , Yixiao Chen\n- **URL**: [https://arxiv.org/abs/2605.05460](https://arxiv.org/abs/2605.05460)\n- **Abstract**:\n> The development of accurate exchange-correlation (XC) functionals remains a longstanding challenge in density functional theory (DFT). The vast majority of XC functionals have been hand designed by human researchers combining physical insight, exact constraints, and empirical fitting. Recent advances in large language models enable a systematic, automated alternative to this human-driven design loop. This report presents an agentic search system in which an LLM proposes structured functional-form changes guided by evolutionary history. The system attempts to improve functional performance through an iterative plan-execute-summarize loop, where improvements are measurable by optimizing functional parameters against a standard thermochemistry dataset, then evaluating performance on a held-out subset. The strongest discovered functional, SAFS26-a (Seed Agentic Functional Search 2026), improves upon the gold-standard {\\omega}B97M-V baseline by ~9%. These results also surface a cautionary lesson for AI-assisted science: models powerful enough to discover genuine improvements are equally capable of exploiting unphysical shortcuts to game the benchmark; domain expertise translated into explicitly enforced constraints remains essential to keeping results scientifically grounded.\n\n---\n\n## 73. The Geopolitics of AI Safety: A Causal Analysis of Regional LLM Bias\n- **Authors**: Alif Al Hasan\n- **URL**: [https://arxiv.org/abs/2605.05427](https://arxiv.org/abs/2605.05427)\n- **Abstract**:\n> As Large Language Models (LLMs) are integrated into global software systems, ensuring equitable safety guardrails is a critical requirement. Current fairness evaluations predominantly measure bias observationally, a methodology confounded by the inherent toxicity of topics naturally paired with specific demographics in testing datasets. This study introduces a Probabilistic Graphical Model (PGM) framework to audit LLM safety mechanisms causally. By applying Pearl's do-operator, we mathematically isolate the causal effect of injecting a cultural demographic into a prompt. We conduct a large-scale empirical analysis across seven instruction-tuned models spanning diverse origins: the United States (Llama-3.1-8B, Gemma-2-9B), Europe (Mistral-7B-v0.3), the UAE (Falcon3-7B), China (Qwen2.5-7B, DeepSeek-7B), and India (Airavata-7B). Utilizing two distinct datasets (ToxiGen and BOLD), the findings reveal a disparity between observational and interventional bias, demonstrating that standard fairness metrics can overestimate demographic bias by failing to account for context toxicity. Furthermore, the causal probabilities indicate distinct alignment trends: Western models exhibit higher causal refusal rates for specific demographic groups, whereas Eastern models demonstrate low overall intervention rates with targeted sensitivities toward regional demographics. We discuss the implications of these biases, highlighting how demographic-sensitive over-triggering restricts benign discourse in downstream applications.\n\n---\n\n## 74. From History to State: Constant-Context Skill Learning for LLM Agents\n- **Authors**: Haoyang Xie , Xinyuan Wang , Yancheng Wang , Puda Zhao , Feng Ju\n- **URL**: [https://arxiv.org/abs/2605.05413](https://arxiv.org/abs/2605.05413)\n- **Abstract**:\n> Large language model (LLM) agents are increasingly used to operate browsers, files, code and tools, making personal assistants a natural deployment target. Yet personal agents face a privacy-cost-capability tension: cloud models execute multi-step workflows well but expose sensitive intermediate context to external APIs, while local models preserve privacy but remain less reliable. Both settings also pay repeatedly for long skill prompts and growing histories. We propose constant-context skill learning, a context-to-weights framework for recurring agent workflows: reusable procedures are learned in lightweight task-family modules, while inference conditions only on the current observation and a compact state block. A deterministic tracker renders this state block from task progress and supplies aligned subgoal rewards, so each module can be trained with step-level SFT and refined through online RL. Across ALFWorld, WebShop, and SciWorld, our agents achieve strong performance across Qwen3-4B, Qwen3-8B and Llama-3.1-8B. With Qwen3-8B, SFT+RL reaches 89.6\\% unseen success on ALFWorld, 76.8\\% success on WebShop, and 66.4\\% unseen success on SciWorld. They match or exceed strong published agent-training results while reducing prompt tokens per turn by 2--7$\\times$ relative to controlled ReAct prompting baselines, showing that procedural context can be moved from prompts into weights.\n\n---\n\n## 75. LaTA: A Drop-in, FERPA-Compliant Local-LLM Autograder for Upper-Division STEM Coursework\n- **Authors**: Jesse A. Rodríguez\n- **URL**: [https://arxiv.org/abs/2605.05410](https://arxiv.org/abs/2605.05410)\n- **Abstract**:\n> Large-language-model (LLM) graders promise to relieve the grading burden of upper-division STEM courses, but most deployments to date send student work to third-party APIs, violating FERPA and exposing institutions to data risk while requiring substantial assignment modification. We present $\\textbf{LaTA}\\ (\\textit{LaTeX Teaching Assistant})$, a drop-in, open-source autograder that runs entirely on commodity on-premises hardware and assumes a LaTeX-native workflow already adopted by many engineering and physics courses. LaTA implements a four-stage pipeline (ingest, segment, grade, report) using a locally hosted open-weight chain-of-thought LLM grader (gpt-oss:120b) that compares student work to an instructor-authored reference solution and applies a YAML rubric with binary per-item scoring. We deployed LaTA in Winter~2026 in ME 373 (Mechanical Engineering Methods) at Oregon State University, grading every weekly assignment for approximately 200 students on a single Mac Studio at \\$0 marginal cost per assignment and 1--3 minutes of wall-clock time per submission, enabling regrading of corrected assignments and greatly expanded TA office hour offerings. The instructor-confirmed grading-error rate held at roughly $0.02$--$0.04\\%$ per rubric line item across the term. Relative to the same instructor's previous traditionally-graded cohort, the LaTA-graded cohort outperformed by approximately $11\\%$ on the midterm exam and $8\\%$ on the final exam, and reported large gains in self-assessed confidence on every stated learning objective ($N = 159$ survey responses, $\\Delta \\geq +1.49$ Likert points, $p < 10^{-27}$ on every comparison). We release the code under AGPLv3.\n\n---\n\n## 76. Agentic Retrieval-Augmented Generation for Financial Document Question Answering\n- **Authors**: Yang Shu , Yingmin Liu , Zequn Xie\n- **URL**: [https://arxiv.org/abs/2605.05409](https://arxiv.org/abs/2605.05409)\n- **Abstract**:\n> Financial document question answering (QA) demands complex multi-step numerical reasoning over heterogeneous evidence--structured tables, textual narratives, and footnotes--scattered across corporate filings. Existing retrieval-augmented generation (RAG) approaches adopt a single-pass retrieve-then-generate paradigm that struggles with the compositional reasoning chains prevalent in financial analysis. We propose FinAgent-RAG, an agentic RAG framework that orchestrates iterative retrieval-reasoning loops with self-verification, specifically engineered for the precision requirements of financial numerical reasoning. The framework integrates three domain-specific innovations: (1) a Contrastive Financial Retriever trained with hard negative mining to distinguish semantically similar but numerically distinct financial passages, (2) a Program-of-Thought reasoning module that generates executable Python code for precise arithmetic rather than relying on error-prone LLM-based mental computation, and (3) an Adaptive Strategy Router that dynamically allocates computational resources based on question complexity, reducing API costs by 41.3% on FinQA while preserving accuracy. Extensive experiments on three benchmark datasets--FinQA, ConvFinQA, and TAT-QA--demonstrate that FinAgent-RAG achieves 76.81%, 78.46%, and 74.96% execution accuracy respectively, outperforming the strongest baseline by 5.62--9.32 percentage points. Ablation studies, cross-backbone evaluation with four LLMs, and deployment cost analysis confirm the framework's robustness and practical viability for financial institutions.\n\n---\n\n## 77. PRISM: Perception Reasoning Interleaved for Sequential Decision Making\n- **Authors**: Mohamed Salim Aissi , Clemence Grislain , Clement Romac , Laure Soulier , Mohamed Chetouani , Olivier Sigaud , Nicolas Thome\n- **URL**: [https://arxiv.org/abs/2605.05407](https://arxiv.org/abs/2605.05407)\n- **Abstract**:\n> Scaling LLM-based embodied agents from text-only environments to complex multimodal settings remains a major challenge. Recent work identifies a perception-reasoning-decision gap in standalone Vision-Language Models (VLMs), which often overlook task-critical information. In this paper, we introduce PRISM, a framework that tightly couples perception (VLM) and decision (LLM) through a dynamic question-answer (DQA) pipeline. Instead of passively accepting the VLM's description, the LLM critiques it, probes the VLM with goal-oriented questions, and synthesizes a compact image description. This closed-loop interaction yields a sharp, task-driven understanding of the scene. We evaluate PRISM on the ALFWorld and Room-to-Room (R2R) benchmarks. We show that: (1) PRISM significantly outperforms state-of-the-art image-based models, (2) our Interactive goal-oriented perception pipeline yields systematic and substantial gains, and (3) PRISM is fully automatic, eliminating the need for handcrafted questions or answers.\n\n---\n\n## 78. When Helpfulness Becomes Sycophancy: Sycophancy is a Boundary Failure Between Social Alignment and Epistemic Integrity in Large Language Models\n- **Authors**: Jiechen Li , Catherine A. Barry , Rishika Randev , Janet Chen , Ella Jorgensen , Brinnae Bent\n- **URL**: [https://arxiv.org/abs/2605.05403](https://arxiv.org/abs/2605.05403)\n- **Abstract**:\n> This position paper argues that sycophancy in LLMs is a boundary failure between social alignment and epistemic integrity. Existing work often operationalizes sycophancy through external behavior such as agreement with incorrect user beliefs, position reversals, or deviation from an objective standard of correctness. These formulations capture only overt forms of the phenomenon and leave subtler boundary failures involving epistemic integrity and social alignment underspecified. We argue that sycophancy should not be understood as agreement alone, but as alignment behavior that displaces independent epistemic judgment. To clarify this boundary, we propose a three-condition framework for sycophancy. First, the user expresses a cue in the form of a belief, preference, or self-concept. Second, the model shifts toward that cue through alignment behavior. Third, this shift compromises epistemic accuracy, independent reasoning, or appropriate correction. We also introduce a taxonomy for classifying sycophancy, consisting of alignment targets, mechanisms, and severity. The paper concludes by discussing implications for alignment evaluation and argues for boundary-aware assessment, structured rubrics, and mitigation strategies, while situating these proposals alongside alternative views of sycophancy.\n\n---\n\n## 79. BALAR : A Bayesian Agentic Loop for Active Reasoning\n- **Authors**: Aymen Echarghaoui , Dongxia Wu , Emily B. Fox\n- **URL**: [https://arxiv.org/abs/2605.05386](https://arxiv.org/abs/2605.05386)\n- **Abstract**:\n> Large language models increasingly operate in interactive settings where solving a task requires multiple rounds of information exchange with a user. However, most current systems treat dialogue reactively and lack a principled mechanism to reason about what information is missing and which question should be asked next. We propose BALAR (Bayesian Agentic Loop for Active Reasoning), a task-agnostic outer-loop algorithm that requires no fine-tuning and enables structured multi-turn interaction between an LLM agent and a user. BALAR maintains a structured belief over latent states, selects clarifying questions by maximizing expected mutual information, and dynamically expands its state representation when the current one proves insufficient. We evaluate BALAR on three diverse benchmarks: AR-Bench-DC (detective cases), AR-Bench-SP (thinking puzzles), and iCraft-MD (clinical diagnosis). BALAR significantly outperforms all baselines across all three benchmarks, with $14.6\\%$ higher accuracy on AR-Bench-DC, $38.5\\%$ on AR-Bench-SP, and $30.5\\%$ on iCraft-MD.\n\n---\n\n## 80. Understanding Annotator Safety Policy with Interpretability\n- **Authors**: Alex Oesterling , Donghao Ren , Yannick Assogba , Dominik Moritz , Sunnie S.Y. Kim , Leon Gatys , Fred Hohman\n- **URL**: [https://arxiv.org/abs/2605.05329](https://arxiv.org/abs/2605.05329)\n- **Abstract**:\n> Safety policies define what constitutes safe and unsafe AI outputs, guiding data annotation and model development. However, annotation disagreement is pervasive and can stem from multiple sources such as operational failures (annotators misunderstand or misexecute the task), policy ambiguity (policy wording leaves room for interpretation), or value pluralism (different annotators hold different perspectives on safety). Distinguishing these sources matters. For example, operational failures call for quality control, ambiguity calls for policy clarification, and pluralism calls for deliberation about incorporating diverse perspectives. Yet understanding why annotators disagree is difficult. Directly asking annotators for their reasoning is costly, substantially increasing annotation burden, and can be unreliable for both human and LLM annotators as self-reported reasoning often fails to reflect actual decision processes. We introduce Annotator Policy Models (APMs), interpretable models that learn annotators' internal safety policies from labeling behavior alone, making annotator reasoning visible and comparable without additional annotation effort. We validate that APMs accurately model annotator safety policy (>80% accuracy), faithfully predict responses to counterfactual edits, and recover known policy differences in controlled settings. Applying APMs to LLM and human annotations, we demonstrate two core applications: (1) surfacing policy ambiguity by revealing how annotators interpret safety instructions differently, and (2) surfacing value pluralism by uncovering systematic differences in safety priorities across demographic groups. Together, these capabilities support more targeted, transparent, and inclusive safety policy design.\n\n---\n\n## 81. Verifier-Backed Hard Problem Generation for Mathematical Reasoning\n- **Authors**: Yuhang Lai , Jiazhan Feng , Yee Whye Teh , Ning Miao\n- **URL**: [https://arxiv.org/abs/2605.06660](https://arxiv.org/abs/2605.06660)\n- **Abstract**:\n> Large Language Models (LLMs) demonstrate strong capabilities for solving scientific and mathematical problems, yet they struggle to produce valid, challenging, and novel problems - an essential component for advancing LLM training and enabling autonomous scientific research. Existing problem generation approaches either depend on expensive human expert involvement or adopt naive self-play paradigms, which frequently yield invalid problems due to reward hacking. This work introduces VHG, a verifier-enhanced hard problem generation framework built upon three-party self-play. By integrating an independent verifier into the conventional setter-solver duality, our design constrains the setter's reward to be jointly determined by problem validity (evaluated by the verifier) and difficulty (assessed by the solver). We instantiate two verifier variants: a Hard symbolic verifier and a Soft LLM-based verifier, with evaluations conducted on indefinite integral tasks and general mathematical reasoning tasks. Experimental results show that VHG substantially outperforms all baseline methods by a clear margin.\n\n---\n\n## 82. Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less\n- **Authors**: Yuxing Liu , Jianyu Wang , Tong Zhang\n- **URL**: [https://arxiv.org/abs/2605.06654](https://arxiv.org/abs/2605.06654)\n- **Abstract**:\n> Optimizers play an important role in both pretraining and finetuning stages when training large language models (LLMs). In this paper, we present an observation that full finetuning with the same optimizer as in pretraining achieves a better learning-forgetting tradeoff, i.e., forgetting less while achieving the same or better performance on the new task, than other optimizers and, possibly surprisingly, LoRA, during the supervised finetuning (SFT) stage. We term this phenomenon optimizer-model consistency. To better understand it, through controlled experiments and theoretical analysis, we show that: 1) optimizers can shape the models by having regularization effects on the activations, leading to different landscapes around the pretrained checkpoints; 2) in response to this regularization effect, the weight update in SFT should follow some specific structures to lower forgetting of the knowledge learned in pretraining, which can be obtained by using the same optimizer. Moreover, we specifically compare Muon and AdamW when they are employed throughout the pretraining and SFT stages and find that Muon performs worse when finetuned for reasoning tasks. With a synthetic language modeling experiment, we demonstrate that this can come from Muon's strong tendency towards rote memorization, which may hurt pattern acquisition with a small amount of data, as for SFT.\n\n---\n\n## 83. When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels\n- **Authors**: Sushant Gautam , Finn Schwall , Annika Willoch Olstad , Fernando Vallecillos Ruiz , Birk Torpmann-Hagen , Sunniva Maria Stordal Bjørklund , Leon Moonen , Klas Pettersen , Michael A. Riegler\n- **URL**: [https://arxiv.org/abs/2605.06652](https://arxiv.org/abs/2605.06652)\n- **Abstract**:\n> Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component ($\\eta^2 \\approx 0.52$), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.\n\n---\n\n## 84. Superintelligent Retrieval Agent: The Next Frontier of Information Retrieval\n- **Authors**: Zeyu Yang , Qi Ma , Jason Chen , Anshumali Shrivastava\n- **URL**: [https://arxiv.org/abs/2605.06647](https://arxiv.org/abs/2605.06647)\n- **Abstract**:\n> Retrieval-augmented agents are increasingly the interface to large organizational knowledge bases, yet most still treat retrieval as a black box: they issue exploratory queries, inspect returned snippets, and iteratively reformulate until useful evidence emerges. This approach resembles how a newcomer searches an unfamiliar database rather than how an expert navigates it with strong priors about terminology and likely evidence, and results in unnecessary retrieval rounds, increased latency, and poor recall. We introduce \\textit{SuperIntelligent Retrieval Agent} (SIRA), which defines \\emph{superintelligence} in retrieval as the ability to compress multi-round exploratory search into a single corpus-discriminative retrieval action. SIRA does not merely ask what terms are relevant to the query; it asks which terms are likely to separate the desired evidence from corpus-level confusers. On the corpus side, an LLM enriches each document offline with missing search vocabulary; on the query side, it predicts evidence vocabulary omitted by the query; and document-frequency statistics as a tool call to filter proposed terms that are absent, overly common, or unlikely to create retrieval margin. The final retrieval step is a single weighted BM25 call combining the original query with the validated expansion. Across ten BEIR benchmarks and downstream question-answering tasks, SIRA achieves the significantly superior performance outperforming dense retrievers and state-of-the-art multi-round agentic baselines, demonstrating that one well-formed lexical query, guided by LLM cognition and lightweight corpus statistics, can exceed substantially more expensive multi-round search while remaining interpretable, training-free, and efficient.\n\n---\n\n## 85. StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction\n- **Authors**: Xiangyuan Xue , Yifan Zhou , Zidong Wang , Shengji Tang , Philip Torr , Wanli Ouyang , Lei Bai , Zhenfei Yin\n- **URL**: [https://arxiv.org/abs/2605.06642](https://arxiv.org/abs/2605.06642)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used as interactive agents, but optimizing them for long-horizon decision making remains difficult because current methods are largely purely reactive, which weakens both exploration and credit assignment over extended trajectories. In this work, we present Strategic Trajectory Abstraction (StraTA), a simple framework that introduces an explicit trajectory-level strategy into agentic reinforcement learning (RL). StraTA samples a compact strategy from the initial task state, conditions subsequent actions on that strategy, and trains strategy generation and action execution jointly with a hierarchical GRPO-style rollout design, further enhanced by diverse strategy rollout and critical self-judgment. Experiments on ALFWorld, WebShop, and SciWorld show that StraTA consistently improves both sample efficiency and final performance over strong baselines. StraTA reaches success rates of 93.1% on ALFWorld and 84.2% on WebShop. On SciWorld, StraTA attains a 63.5% overall score, outperforming frontier closed-source models.\n\n---\n\n## 86. The Structural Origin of Attention Sink: Variance Discrepancy, Super Neurons, and Dimension Disparity\n- **Authors**: Siquan Li , Kaiqi Jiang , Jiacheng Sun , Tianyang Hu\n- **URL**: [https://arxiv.org/abs/2605.06611](https://arxiv.org/abs/2605.06611)\n- **Abstract**:\n> Despite the prevalence of the attention sink phenomenon in Large Language Models (LLMs), where initial tokens disproportionately monopolize attention scores, its structural origins remain elusive. This work provides a \\textit{mechanistic explanation} for this phenomenon. First, we trace its root to the value aggregation process inherent in self-attention, which induces a systematic variance discrepancy. We further demonstrate that this discrepancy is drastically amplified by the activation of super neurons within Feed-Forward Network (FFN) layers. Specifically, the channel-sparse down-projections trigger a dimension disparity of the first-token representation, necessitating the formation of attention sinks as a structural anchor. Then, we validate this causal chain through two controlled interventions: (i) isolating the aggregation effect via attention mask modifications and (ii) amplifying the variance of targeted token representations. Both interventions can replicate attention sinks at arbitrary positions. Our mechanistic understanding offers a foundation for the systematic control of sink formation. Finally, as a proof of concept, we propose \\textit{head-wise RMSNorm}, an architectural modification that stabilizes value aggregation outputs during pre-training. Our experiments demonstrate that restoring statistical parity across positions significantly accelerates convergence.\n\n---\n\n## 87. AI CFD Scientist: Toward Open-Ended Computational Fluid Dynamics Discovery with Physics-Aware AI Agents\n- **Authors**: Nithin Somasekharan , Rabi Pathak , Manushri Dhanakoti , Tingwen Zhang , Ling Yue , Andy Zhu , Shaowu Pan\n- **URL**: [https://arxiv.org/abs/2605.06607](https://arxiv.org/abs/2605.06607)\n- **Abstract**:\n> Recent LLM-based agents have closed substantial portions of the scientific discovery loop in software-only machine-learning research, in chemistry, and in biology. Extending the same loop to high-fidelity physical simulators is harder, because solver completion does not imply physical validity and many failure modes appear only in field-level imagery rather than in solver logs. We present AI CFD Scientist, an open-source AI scientist for computational fluid dynamics (CFD) that, to our knowledge, is the first to span literature-grounded ideation, validated execution, vision-based physics verification, source-code modification, and figure-grounded writing within a single inspectable workflow. Three coupled pathways cover parameter sweeps within a fixed solver, case-local C++ library compilation for new physical models, and open-ended hypothesis search against a reference comparator, all running on OpenFOAM through Foam-Agent. At the center of the framework is a vision-language physics-verification gate that inspects rendered flow fields before any result is accepted, rerun, or written into a manuscript. On five tasks under a shared GPT-5.5 backbone, AI CFD Scientist autonomously discovers a Spalart-Allmaras runtime correction that reduces lower-wall Cf RMSE against DNS by 7.89% on the periodic hill at Reh=5600; under matched LLM cost, two strong general AI-scientist baselines (ARIS, DeepScientist) execute partial CFD workflows but lack the domain-specific validity gates needed to convert runs into defensible scientific claims; and a controlled planted-failure ablation shows that the vision-language gate detects 14 of 16 silent failures missed by solver-level checks. Code, prompts, and run artifacts are released at this https URL .\n\n---\n\n## 88. UniSD: Towards a Unified Self-Distillation Framework for Large Language Models\n- **Authors**: Yiqiao Jin , Yiyang Wang , Lucheng Fu , Yijia Xiao , Yinyi Luo , Haoxin Liu , B. Aditya Prakash , Josiah Hester , Jindong Wang , Srijan Kumar\n- **URL**: [https://arxiv.org/abs/2605.06597](https://arxiv.org/abs/2605.06597)\n- **Abstract**:\n> Self-distillation (SD) offers a promising path for adapting large language models (LLMs) without relying on stronger external teachers. However, SD in autoregressive LLMs remains challenging because self-generated trajectories are free-form, correctness is task-dependent, and plausible rationales can still provide unstable or unreliable supervision. Existing methods mainly examine isolated design choices, leaving their effectiveness, roles, and interactions unclear. In this paper, we propose UniSD, a unified framework to systematically study self-distillation. UniSD integrates complementary mechanisms that address supervision reliability, representation alignment, and training stability, including multi-teacher agreement, EMA teacher stabilization, token-level contrastive learning, feature matching, and divergence clipping. Across six benchmarks and six models from three model families, UniSD reveals when self-distillation improves over static imitation, which components drive the gains, and how these components interact across tasks. Guided by these insights, we construct UniSDfull, an integrated pipeline that combines complementary components and achieves the strongest overall performance, improving over the base model by +5.4 points and the strongest baseline by +2.8 points. Extensive evaluation highlights self-distillation as a practical and steerable approach for efficient LLM adaptation without stronger external teachers.\n\n---\n\n## 89. Continuous Latent Diffusion Language Model\n- **Authors**: Hongcan Guo , Qinyu Zhao , Yian Zhao , Shen Nie , Rui Zhu , Qiushan Guo , Feng Wang , Tao Yang , Hengshuang Zhao , Guoqiang Wei , Yan Zeng\n- **URL**: [https://arxiv.org/abs/2605.06548](https://arxiv.org/abs/2605.06548)\n- **Abstract**:\n> Large language models have achieved remarkable success under the autoregressive paradigm, yet high-quality text generation need not be tied to a fixed left-to-right order. Existing alternatives still struggle to jointly achieve generation efficiency, scalable representation learning, and effective global semantic modeling. We propose Cola DLM, a hierarchical latent diffusion language model that frames text generation through hierarchical information decomposition. Cola DLM first learns a stable text-to-latent mapping with a Text VAE, then models a global semantic prior in continuous latent space with a block-causal DiT, and finally generates text through conditional decoding. From a unified Markov-path perspective, its diffusion process performs latent prior transport rather than token-level observation recovery, thereby separating global semantic organization from local textual realization. This design yields a more flexible non-autoregressive inductive bias, supports semantic compression and prior fitting in continuous space, and naturally extends to other continuous modalities. Through experiments spanning 4 research questions, 8 benchmarks, strictly matched ~2B-parameter autoregressive and LLaDA baselines, and scaling curves up to about 2000 EFLOPs, we identify an effective overall configuration of Cola DLM and verify its strong scaling behavior for text generation. Taken together, the results establish hierarchical continuous latent prior modeling as a principled alternative to strictly token-level language modeling, where generation quality and scaling behavior may better reflect model capability than likelihood, while also suggesting a concrete path toward unified modeling across discrete text and continuous modalities.\n\n---\n\n## 90. Is One Layer Enough? Understanding Inference Dynamics in Tabular Foundation Models\n- **Authors**: Amir Rezaei Balef , Mykhailo Koshil , Katharina Eggensperger\n- **URL**: [https://arxiv.org/abs/2605.06510](https://arxiv.org/abs/2605.06510)\n- **Abstract**:\n> Transformer-based tabular foundation models (TFMs) dominate small to medium tabular predictive benchmark tasks, yet their inference mechanisms remain largely unexplored. We present the first large-scale mechanistic study of layerwise dynamics in 6 state-of-the-art tabular in-context learning models. We explore how predictions emerge across depth, identify distinct stages of inference and reveal latent-space dynamics that differ from those of language models. Our findings indicate substantial depthwise redundancy across multiple models, suggesting iterative refinement with overlapping computations during inference stages. Guided by these insights, we design a proof-of-concept, looped single-layer model that uses only 20% of the original model's parameters while achieving comparable performance. The code is available at this https URL .\n\n---\n\n## 91. PACZero: PAC-Private Fine-Tuning of Language Models via Sign Quantization\n- **Authors**: Murat Bilgehan Ertan , Xiaochen Zhu , Phuong Ha Nguyen , Marten van Dijk , Srinivas Devadas\n- **URL**: [https://arxiv.org/abs/2605.06505](https://arxiv.org/abs/2605.06505)\n- **Abstract**:\n> We introduce PACZero, a family of PAC-private zeroth-order mechanisms for fine-tuning large language models that delivers usable utility at $I(S^*; Y_{1:T})=0$. This privacy regime bounds the membership-inference attack (MIA) posterior success rate at the prior, an MIA-resistance level the DP framework matches only at $\\varepsilon=0$ and infinite noise. All DP-ZO comparisons below are matched at the MIA posterior level. The key insight is that PAC Privacy charges mutual information only when the release depends on which candidate subset is the secret. Sign-quantizing subset-aggregated zeroth-order gradients creates frequent unanimity, steps at which every candidate subset agrees on the update direction; at these steps the released sign costs zero conditional mutual information. We propose two variants that span the privacy-utility trade-off: PACZero-MI (budgeted MI via exact calibration on the binary release) and PACZero-ZPL ($I=0$ via a uniform coin flip on disagreement steps). We evaluate on SST-2 and SQuAD with OPT-1.3B and OPT-6.7B in both LoRA and full-parameter tracks. On SST-2 OPT-1.3B full fine-tuning at $I=0$, PACZero-ZPL reaches ${88.99\\pm0.91}$, within $2.1$pp of the non-private MeZO baseline ($91.1$ FT). No prior method produces usable utility in the high-privacy regime $\\varepsilon<1$, and PACZero-ZPL obtains competitive SST-2 accuracy and nontrivial SQuAD F1 across OPT-1.3B and OPT-6.7B at $I=0$.\n\n---\n\n## 92. Litespark Inference on Consumer CPUs: Custom SIMD Kernels for Ternary Neural Networks\n- **Authors**: Nii Osae Osae Dade , Tony Morri , Moinul Hossain Rahat , Sayandip Pal\n- **URL**: [https://arxiv.org/abs/2605.06485](https://arxiv.org/abs/2605.06485)\n- **Abstract**:\n> Large language models (LLMs) have transformed artificial intelligence, but their computational requirements remain prohibitive for most users. Standard inference demands expensive datacenter GPUs or cloud API access, leaving over one billion personal computers underutilized for AI workloads. Ternary models offer a path forward: their weights are constrained to {-1, 0, +1}, theoretically eliminating the need for floating-point multiplication. However, existing frameworks fail to exploit this structure, treating ternary models as dense floating-point networks. We address this gap with custom SIMD kernels that replace matrix multiplication with simple addition and subtraction operations, targeting the integer dot product instructions available on modern CPUs. Our implementation, Litespark-Inference, is pip-installable and integrates directly with Hugging-Face, achieving 9.2x faster time-to-first-token, 52x higher throughput, and 14x memory reduction compared to standard PyTorch inference on Apple Silicon, with similar speedups on Intel and AMD processors.\n\n---\n\n## 93. Constraint Decay: The Fragility of LLM Agents in Backend Code Generation\n- **Authors**: Francesco Dente , Dario Satriani , Paolo Papotti\n- **URL**: [https://arxiv.org/abs/2605.06445](https://arxiv.org/abs/2605.06445)\n- **Abstract**:\n> Large Language Model (LLM) agents demonstrate strong performance in autonomous code generation under loose specifications. However, production-grade software requires strict adherence to structural constraints, such as architectural patterns, databases, and object-relational mappings. Existing benchmarks often overlook these non-functional requirements, rewarding functionally correct but structurally arbitrary solutions. We present a systematic study evaluating how well agents handle structural constraints in multi-file backend generation. By fixing a unified API contract across 80 greenfield generation tasks and 20 feature-implementation tasks spanning eight web frameworks, we isolate the effect of structural complexity using a dual evaluation with end-to-end behavioral tests and static verifiers. Our findings reveal a phenomenon of constraint decay: as structural requirements accumulate, agent performance exhibits a substantial decline. Capable configurations lose 30 points on average in assertion pass rates from baseline to fully specified tasks, while some weaker configurations approach zero. Framework sensitivity analysis exposes significant performance disparities: agents succeed in minimal, explicit frameworks (e.g., Flask) but perform substantially worse on average in convention-heavy environments (e.g., FastAPI, Django). Finally, error analysis identifies data-layer defects (e.g., incorrect query composition and ORM runtime violations) as the leading root causes. This work highlights that jointly satisfying functional and structural requirements remains a key open challenge for coding agents.\n\n---\n\n## 94. Is Escalation Worth It? A Decision-Theoretic Characterization of LLM Cascades\n- **Authors**: Dylan Bouchard\n- **URL**: [https://arxiv.org/abs/2605.06350](https://arxiv.org/abs/2605.06350)\n- **Abstract**:\n> Model cascades, in which a cheap LLM defers to an expensive one on low-confidence queries, are widely used to navigate the cost-quality tradeoff at deployment. Existing approaches largely treat the deferral threshold as an empirical hyperparameter, with limited guidance on the geometry of the resulting cost-quality frontier over a model pool. We develop a decision-theoretic framework grounded in constrained optimization and duality. For a two-model cascade, we establish piecewise concavity of the cost-quality frontier on decreasing-benefit regions of the confidence support, with reciprocal shadow prices linking the budget- and quality-constrained formulations. Given a pool of $k$ models, we characterize the frontier achievable by deterministic two-model threshold cascades as the pointwise envelope over $\\binom{k}{2}$ pairwise cascades, with switching points where the optimal pair changes. For $k$-model cascades, we derive first-order conditions in which a single shadow price equalizes marginal quality-per-cost across stage boundaries. We validate the framework on five benchmarks (MATH, MMLU, TriviaQA, SimpleQA, LiveCodeBench) across eight models from five providers. Within the deterministic threshold-cascade class, full fixed chains underperform the pairwise envelope, and optimized subsequence cascades do not deliver practically meaningful held-out gains over it. A lightweight pre-generation router exceeds the best cascade policy on four of five datasets, mainly because it avoids the cheap model's generation cost on queries sent directly to a larger model rather than because of a stronger routing signal. These results suggest that cascade performance is limited primarily by structural cost, since cascades pay the cheap model before any escalation decision, rather than by a shortage of intermediate stages.\n\n---\n\n## 95. Human-AI Co-Evolution and Epistemic Collapse: A Dynamical Systems Perspective\n- **Authors**: Xuening Wu , Yanlan Kang , Qianya Xu , Kexuan Xie , Jiaqi Mi , Honggang Wang , Yubin Liu , Zeping Chen\n- **URL**: [https://arxiv.org/abs/2605.06347](https://arxiv.org/abs/2605.06347)\n- **Abstract**:\n> Large language models (LLMs) are reshaping how knowledge is produced, with increasing reliance on AI systems for generation, summarization, and reasoning. While prior work has studied cognitive offloading in humans and model collapse in recursive training, these effects are typically considered in isolation. We propose a unified perspective: humans and language models form a coupled dynamical system linked by a feedback loop of usage, generation, and retraining. We introduce a minimal model with three variables -- human cognition, data quality, and model capability -- and show that this feedback can give rise to distinct dynamical regimes. Our analysis identifies three regimes: co-evolutionary enhancement, fragile equilibrium, and degenerative convergence. Through a simple simulation, we demonstrate that increasing reliance on AI can induce a transition toward a low-diversity, suboptimal equilibrium. From an information-theoretic perspective, this transition corresponds to an emergent information bottleneck in the human-AI loop, where entropy reduction reflects loss of diversity and support under closed-loop feedback rather than beneficial compression. These results suggest that the trajectory of AI systems is shaped not only by model design, but by the dynamics of human-AI co-evolution.\n\n---\n\n## 96. CoupleEvo: Evolving Heuristics for Coupled Optimization Problems Using Large Language Models\n- **Authors**: Thomas Bömer , Bastian Amberg , Max Disselnmeyer , Anne Meyer\n- **URL**: [https://arxiv.org/abs/2605.06341](https://arxiv.org/abs/2605.06341)\n- **Abstract**:\n> Many real-world optimization problems consist of multiple tightly coupled subproblems whose solutions must be coordinated to achieve high overall performance. However, existing large language model driven automated heuristic design approaches are limited to single-problem settings. In this paper, we propose CoupleEvo. CoupleEvo proposes three evolutionary coordination strategies to evolve heuristics for coupled optimization problems: the sequential strategy evolves heuristics for one subproblem after the other; the iterative strategy alternates the evolution of heuristics for different subproblems over successive generations; and the integrated strategy evolves heuristics for all problems simultaneously. The approach is evaluated on two representative coupled optimization problems. Experimental results show that decomposition-based strategies (sequential and iterative) provide more stable convergence and higher solution quality, while the integrated evolution strategy suffers from increased search complexity and variability. These findings highlight the importance of coordinating evolutionary search across interdependent subproblems and demonstrate the potential of LLM-driven heuristic design for complex coupled optimization problems. The code is available: this https URL .\n\n---\n\n## 97. Fine-Tuning Small Language Models for Solution-Oriented Windows Event Log Analysis\n- **Authors**: Siraaj Akhtar , Saad Khan , Simon Parkinson\n- **URL**: [https://arxiv.org/abs/2605.06330](https://arxiv.org/abs/2605.06330)\n- **Abstract**:\n> Large language models (LLMs) have shown promise for event log analysis, but their high computational requirements, reliance on cloud infrastructure, and security concerns limit practical deployment. In addition, most existing approaches focus only on the identification of the problem and do not provide actionable remediation. Small language models (SLMs) present a light-weight alternative that can be fine-tuned for a specific purpose and hosted locally. This paper investigates whether SLMs, when fine-tuned for a specific task, can serve as a practical alternative for event log analysis while also generating solutions. We first create a large-scale synthetic Windows event log dataset that contains remediation actions using a high-performing LLM. We then fine-tune multiple SLMs and LLMs using the LoRA parameter-efficient fine-tuning technique and evaluate their performance by comparing with expert assessment. The results show that the dataset accurately reflects real-world scenarios and that fine-tuned SLMs consistently outperform LLMs in identifying issues and providing relevant remediation, while requiring fewer computational resources.\n\n---\n\n## 98. Measuring Evaluation-Context Divergence in Open-Weight LLMs: A Paired-Prompt Protocol with Pilot Evidence of Alignment-Pipeline-Specific Heterogeneity\n- **Authors**: Florian A. D. Burnat , Brittany I. Davidson\n- **URL**: [https://arxiv.org/abs/2605.06327](https://arxiv.org/abs/2605.06327)\n- **Abstract**:\n> Safety benchmarks are routinely treated as evidence about how a language model will behave once deployed, but this inference is fragile if behavior depends on whether a prompt looks like an evaluation. We define evaluation-context divergence as an observable within-item change in behavior induced by framing a fixed task as an evaluation, a live deployment interaction, or a neutral request, and present a paired-prompt protocol that measures it in open-weight LLMs while controlling for paraphrase variation, benchmark familiarity, and judge framing-sensitivity. Across five instruction-tuned checkpoints from four open-weight families plus a matched OLMo-3 base/instruct ablation ($20$ paired items, $840$ generations per checkpoint), we find striking heterogeneity. OLMo-3-Instruct alone is eval-cautious -- evaluation framing raises refusal vs. neutral by $11.8$pp ($p=0.007$) and reduces harmful compliance vs. deployment by $3.6$pp ($p=0.024$, $0/20$ items inverted) -- while Mistral-Small-3.2, Phi-3.5-mini, and Llama-3.1-8B are deployment-cautious}, with marginal eval-vs-deployment refusal effects of $-9$ to $-20$pp. The matched OLMo-3 base also exhibits the deployment-cautious pattern, identifying alignment as the inversion stage; within Llama-3.1, the $70$B model preserves direction with attenuated magnitude, ruling out a simple ``small-model effect that reverses at scale.'' One caveat: the cross-family heterogeneity is judge-dependent. Re-judging with a different-family safety classifier (Llama-Guard-3-8B) preserves the within-OLMo eval-cautious direction but flattens the cross-family contrast, indicating that the two judges operationalize distinct constructs.\n\n---\n\n## 99. Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs\n- **Authors**: Elizabeth Mieczkowski , Alexander Ku , Tiwalayo Eisape , Dilip Arumugam , John Matters , Katherine M. Collins , Ilia Sucholutsky , Thomas L. Griffiths\n- **URL**: [https://arxiv.org/abs/2605.06320](https://arxiv.org/abs/2605.06320)\n- **Abstract**:\n> Large language models (LLMs) are increasingly deployed in teams, yet existing coordination approaches often occupy two extremes. Highly structured methods rely on fixed roles, pipelines, or task decompositions assigned a priori. In contrast, fully unstructured teams enable adaptability and exploration but suffer from inefficiencies such as error propagation, inter-agent conflicts, and wasted resources (measured in time, tokens, or file operations). We introduce Language Agent Teams for Task Evolution (LATTE), a framework for coordinating LLM teams inspired by distributed systems, where processors must operate under partial observability and communication constraints. In LATTE, a team of agents collaboratively construct and maintain a shared, evolving coordination graph which encodes sub-task dependencies, individual agent assignment, and the current state of sub-task progress. This protocol maintains consistency while empowering agents to dynamically allocate work, adapt coordination, and discover new tasks. Across multiple collaborative tasks and a variety of base models, we demonstrate how LATTE reduces token usage, wall-clock time, communication, and coordination failures (e.g. file conflicts and redundant outputs) while matching or exceeding the accuracy of standard designs including MetaGPT, decentralized teams, top-down Leader-Worker hierarchies, and static decompositions.\n\n---\n\n## 100. Pro-KLShampoo: Projected KL-Shampoo with Whitening Recovered by Orthogonalization\n- **Authors**: Ruotong Sun , Ermin Wei\n- **URL**: [https://arxiv.org/abs/2605.06316](https://arxiv.org/abs/2605.06316)\n- **Abstract**:\n> Optimizers that exploit the matrix structure of gradients are central to modern LLM pre-training, with two distinct frontiers: explicit Kronecker-factored preconditioning -- most recently KL-Shampoo, which estimates the preconditioner via KL divergence minimization -- and orthogonalization of the gradient momentum, exemplified by Muon and analyzed as steepest descent under the spectral norm. The two routes are typically developed in isolation. We make a structural observation about KL-Shampoo's Kronecker preconditioners: their eigenvalue spectra exhibit a \\emph{spike-and-flat} shape -- a few dominant eigenvalues followed by an approximately uniform tail -- across layers and training stages, holding exactly under a rank-$\\rho$ signal-plus-noise gradient model. We exploit this structure by restricting one of KL-Shampoo's Kronecker factors to a parametric family aligned with the spike-and-flat shape: full spectral structure on a tracked $r$-dimensional subspace, single shared eigenvalue across the remaining $n-r$ directions. On these directions, we apply orthogonalization. An identity shows that this orthogonalization recovers the algebraic form of full KL-Shampoo's preconditioner. On four pre-training scales (GPT-2 124M / 350M, LLaMA 134M / 450M), Pro-KLShampoo consistently outperforms KL-Shampoo at every subspace rank we test in validation loss, peak per-GPU memory, and wallclock time to reach each loss level.\n\n---\n\n## 101. Attributions All the Way Down? The Metagame of Interpretability\n- **Authors**: Hubert Baniecki , Przemyslaw Biecek , Fabian Fumagalli\n- **URL**: [https://arxiv.org/abs/2605.06295](https://arxiv.org/abs/2605.06295)\n- **Abstract**:\n> We introduce the metagame, a conceptual framework for quantifying second-order interaction effects of model explanations. For any first-order attribution $\\phi(f)$ explaining a model $f$, we measure the directional influence of feature $j$ on the attribution of feature $i$, denoted as meta-attribution $\\varphi_{j \\to i}(f)$, by treating the attribution method itself as a cooperative game and computing its Shapley value. Theoretically, we prove that attributions hierarchically decompose into meta-attributions, and establish these as directional extensions of existing interaction indices. Empirically, we demonstrate that the metagame delivers insights across diverse interpretability applications: (i) quantifying token interactions in instruction-tuned language models, (ii) explaining cross-modal similarity in vision-language encoders, and (iii) interpreting text-to-image concepts in multimodal diffusion transformers.\n\n---\n\n## 102. Log-Likelihood, Simpson's Paradox, and the Detection of Machine-Generated Text\n- **Authors**: Tom Kempton , Viktor Drobnyi , Maeve Madigan , Stuart Burrell\n- **URL**: [https://arxiv.org/abs/2605.06294](https://arxiv.org/abs/2605.06294)\n- **Abstract**:\n> The ability to reliably distinguish human-written text from that generated by large language models is of profound societal importance. The dominant approach to this problem exploits the likelihood hypothesis: that machine-generated text should appear more probable to a detector language model than human-written text. However, we demonstrate that the token-level signal distinguishing human and machine text is non-uniform across the hidden space of the detector model, and naively averaging likelihood-based token scores across regions with fundamentally different statistical structure, as most detectors do, causes a form of Simpson's paradox: a strong local signal is destroyed by inappropriate aggregation. To correct for this, we introduce a learned local calibration step grounded in Bayesian decision theory. Rather than aggregating raw token scores, we first learn lightweight predictors of the score distributions conditioned on position in hidden space, and aggregate calibrated log-likelihood ratios instead. This single intervention dramatically and consistently improves detection performance across all baseline detectors and all datasets we consider. For example, our calibrated variant of Fast-DetectGPT improves AUROC from $0.63$ to $0.85$ on GPT-5.4 text, and a locally-calibrated DMAP detector we introduce achieves state-of-the-art performance across the board. That said, our central contribution is not a new detector, but a precise diagnosis of a significant cause of under-performance of existing detectors and a principled, modular remedy compatible with any token-averaging pipeline. This will serve as a foundation for the community to build upon, with natural avenues including richer distributional models, improved calibration strategies, and principled ensembling with hidden-space geometry signals via the full Bayes-optimal decision rule.\n\n---\n\n## 103. Correct Code, Vulnerable Dependencies: A Large Scale Measurement Study of LLM-Specified Library Versions\n- **Authors**: Chengjie Wang , Jingzheng Wu , Xiang Ling , Tianyue Luo , Chen Zhao\n- **URL**: [https://arxiv.org/abs/2605.06279](https://arxiv.org/abs/2605.06279)\n- **Abstract**:\n> Large language models (LLMs) are now largely involved in software development workflows, and the code they generate routinely includes third-party library (TPL) imports annotated with specific version identifiers. These version choices can carry security and compatibility risks, yet they have not been systematically studied. We present the first large-scale measurement study of version-level risk in LLM-generated Python code, evaluating 10 LLMs on PinTrace, a curated benchmark of 1,000 Stack Overflow programming tasks. LLMs tend to specify version identifiers when directly prompted at 26.83%-95.18%, while down to 6.45%-59.19% in creating a manifest file directly. Among the specified versions, 36.70%-55.70% of tasks contain at least one known CVE, and 62.75%-74.51% of them carry Critical or High severity ratings. In 72.27%-91.37% of cases, the associated CVEs were publicly disclosed before the model's knowledge cutoff. The statistics show all models converge on the same small set of risky release versions, indicating a systemic bias rather than isolated model error. Static compatibility rates range from 19.70% to 63.20%, with installation failure as the dominant cause. The dynamic test cases confirm the pattern by 6.49%-48.62% pass rates. Further experiments confirm that these failures are attributable to version selection rather than code quality, and that externally anchored version constraints substantially reduce both vulnerability exposure and compatibility failures. Our findings reveal LLM version selection as a first-class, previously overlooked risk surface in LLM-based development. We disclosed these findings to the community of the evaluated models, and several confirmed the issue. All the code and dataset have been released for open science at this https URL .\n\n---\n\n## 104. Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs\n- **Authors**: Andy Zeyi Liu , Michael Zhang , Ilana Greenberg , Adam Alnasser , Lucas Baker , John Sous\n- **URL**: [https://arxiv.org/abs/2605.06225](https://arxiv.org/abs/2605.06225)\n- **Abstract**:\n> Steering large language models (LLMs) is usually done by either instruction prompting or activation steering. Prompting often gives strong control, but caches guidance tokens at every layer and can clutter long interactions; activation steering is compact but typically weaker and does not support large structured reminders. We introduce memory inception (MI), a training-free method that steers in latent attention space by inserting text-derived key-value (KV) banks only at selected layers. Rather than materializing reminder content throughout the prompt cache, MI treats steering as selective KV allocation, injecting latent slots only where the model routes to them. On matched personality-steering tasks, MI gives the best overall control--drift trade-off, remaining competitive with prompting while consistently outperforming CAA. On updateable guidance, MI supports mid-conversation behavior shifts without rewriting the visible transcript, achieving the highest post-shift alignment on Qwen3. On structured reasoning, MI outperforms visible prompting on HARDMath and PHYSICS (10/12 subject$\\times$mode cells), serving as proxies for structured reasoning in verifiable domains, while cutting content-matched KV storage by up to 118$\\times$. These results position MI as a powerful steering method when guidance is persistent, structured, or expensive to keep in the visible transcript.\n\n---\n\n## 105. TIDE: Every Layer Knows the Token Beneath the Context\n- **Authors**: Ajay Jaiswal , Lauren Hannah , Han-Byul Kim , Duc Hoang , Mehrdad Farajtabar , Minsik Cho\n- **URL**: [https://arxiv.org/abs/2605.06216](https://arxiv.org/abs/2605.06216)\n- **Abstract**:\n> We revisit a universally accepted but under-examined design choice in every modern LLM: a token index is looked up once at the input embedding layer and then permanently discarded. This single-injection assumption induces two structural failures: (i) the Rare Token Problem, where a Zipf-type distribution of vocabulary causes rare-token embeddings are chronically under-trained due to receiving a fraction of the cumulative gradient signal compared to common tokens; and (ii) the Contextual Collapse Problem, where limited parameters models map distributionally similar tokens to indistinguishable hidden states. As an attempt to address both, we propose TIDE, which augments the standard transformer with EmbeddingMemory: an ensemble of K independent MemoryBlocks that map token indices to context-free semantic vectors, computed once and injected into every layer through a depth-conditioned softmax router with a learnable null bank. We theoretically and empirically establish the benefits of TIDE in addressing the issues associated with single-token identity injection as well as improve performance across multiple language modeling and downstream tasks.\n\n---\n\n## 106. Retina-RAG: Retrieval-Augmented Vision-Language Modeling for Joint Retinal Diagnosis and Clinical Report Generation\n- **Authors**: Abdelrahman Zaian , Sheethal Bhat , Mohamed Abdalkader , Andreas Maier\n- **URL**: [https://arxiv.org/abs/2605.06173](https://arxiv.org/abs/2605.06173)\n- **Abstract**:\n> Diabetic Retinopathy (DR) is a leading cause of preventable blindness among working-age adults worldwide, yet most automated screening systems are limited to image-level classification and lack clinically structured reporting. We propose Retina-RAG, a low-cost modular framework that jointly performs DR severity grading, macular edema (ME) detection, and report generation. The architecture decouples a high-performance retinal classifier and a parameter-efficient vision-language model (Qwen2.5-VL-7B-Instruct) adapted via Low-Rank Adaptation (LoRA), enabling flexible component integration. A retrieval-augmented generation (RAG) module injects curated ophthalmic knowledge together with structured classifier outputs at inference time to improve diagnostic consistency and reduce hallucinations. Retina-RAG achieves an F1-score of 0.731 for DR grading and 0.948 for ME detection, substantially outperforming zero-shot Qwen (0.096, 0.732) and MMed-RAG (0.541, 0.641) on a retinal disease detection dataset with captions. For report generation, Retina-RAG attains ROUGE-L 0.429 and SBERT similarity 0.884, exceeding all baselines. The full framework operates on a single consumer-grade GPU, demonstrating that clinically structured retinal AI can be achieved with modest computational resources.\n\n---\n\n## 107. IRC-Bench: Recognizing Entities from Contextual Cues in First-Person Reminiscences\n- **Authors**: Yehudit Aperstein , Eden Moran , Alexander Apartsin\n- **URL**: [https://arxiv.org/abs/2605.06142](https://arxiv.org/abs/2605.06142)\n- **Abstract**:\n> When people recount personal memories, they often refer to people, places, and events indirectly, relying on contextual cues rather than explicit names. Such implicit references are central to reminiscence narratives: first-person accounts of lived experience used in therapeutic, archival, and social settings. They pose a difficult computational problem because the intended entity must be inferred from dispersed narrative evidence rather than from a local mention. We introduce IRC-Bench, the Implicit Reminiscence Context Benchmark, for evaluating implicit entity recognition in reminiscence transcripts. The benchmark targets non-locality: entity-identifying cues are distributed across multiple, non-contiguous clauses, unlike named entity recognition, entity linking, or coreference resolution. IRC-Bench comprises 25,136 samples constructed from 12,337 Wiki-data-linked entities across 1,994 transcripts spanning 11 thematic domains. Each sample pairs an Entity-Grounded Narrative, in which the target entity is explicitly mentioned, with an Entity-Elided Narrative, in which direct mentions are removed. We evaluate 19 configurations across LLM generation, dense retrieval, RAG, and fine-tuning. QLoRA-adapted Llama 3.1 8B performs best in the open-world setting (38.94% exact match; 51.59% Jaccard), while fine-tuned DPR leads closed-world retrieval (35.38% Hit@1; 71.49% Hit@10). We release IRC-Bench with data, code, and evaluation tools.\n\n---\n\n## 108. Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex\n- **Authors**: Yun Qu , Qi Wang , Yixiu Mao , Heming Zou , Yuhang Jiang , Yingyue Li , Wutong Xu , Lizhou Cai , Weijie Liu , Clive Bai , Kai Yang , Yangkun Chen , Saiyong Yang , Xiangyang Ji\n- **URL**: [https://arxiv.org/abs/2605.06139](https://arxiv.org/abs/2605.06139)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has become a standard approach for large language models (LLMs) post-training to incentivize reasoning capacity. Among existing recipes, group-based policy gradient is prevalent, which samples a group of responses per prompt and updates the policy via group-relative advantage signals. This work reveals that these optimization strategies share a common geometric structure: each implicitly defines a target distribution on the response simplex and projects toward it via first-order approximation. Building on this insight, we propose Listwise Policy Optimization (LPO) to explicitly conduct the target-projection, which demystifies the implicit target by restricting the proximal RL objective to the response simplex, and then projects the policy via exact divergence minimization. This framework provides (i) monotonic improvement on the listwise objective with bounded, zero-sum, and self-correcting projection gradients, and (ii) flexibility in divergence selection with distinct structural properties through the decoupled projection step. On diverse reasoning tasks and LLM backbones, LPO consistently improves training performance over typical policy gradient baselines under matched targets, while intrinsically preserving optimization stability and response diversity.\n\n---\n\n## 109. Normalized Architectures are Natively 4-Bit\n- **Authors**: Maxim Fishman , Brian Chmiel , Ron Banner , Daniel Soudry , Boris Ginsburg\n- **URL**: [https://arxiv.org/abs/2605.06067](https://arxiv.org/abs/2605.06067)\n- **Abstract**:\n> Training large language models at 4-bit precision is critical for efficiency. We show that nGPT, an architecture that constrains weights and hidden representations to the unit hypersphere, is inherently more robust to low-precision arithmetic. This removes the need for interventions-such as applying random Hadamard transforms and performing per-tensor scaling calculations-to preserve model quality, and it enables stable end-to-end NVFP4 training. We validate this approach on both a 1.2B dense model and hybrid (Mamba-Transformer) MoE models of up to 3B/30B parameters. We trace this robustness to the dot product: while quantization noise remains largely uncorrelated in both standard and normalized architectures, the signal behaves differently. In nGPT, the hypersphere constraint enhances weak positive correlations among the element-wise products, leading to a constructive accumulation of the signal across the hidden dimension while the noise continues to average out. This yields a higher effective signal-to-noise ratio and a flatter loss landscape, with the effect strengthening as the hidden dimension grows, suggesting increasing advantages at scale. A reference implementation is available at this https URL\n\n---\n\n## 110. Causal Reinforcement Learning for Complex Card Games: A Magic The Gathering Benchmark\n- **Authors**: Cristiano da Costa Cunha , Ajmal Mian , Tim French , Wei Liu\n- **URL**: [https://arxiv.org/abs/2605.06066](https://arxiv.org/abs/2605.06066)\n- **Abstract**:\n> Causal reinforcement learning (RL) lacks benchmarks for complex systems that combine sequential decision making, hidden information, large masked action spaces, and explicit causal structure. We introduce MTG-Causal-RL, a Gymnasium benchmark built on Magic: The Gathering with a 3,077-dimensional partial observation, a 478-action masked discrete action space, five competitive Standard archetypes, three reward schemes, and a hand-specified Structural Causal Model (SCM) over strategic variables. Every episode exposes causal variables, SCM-predicted intervention effects, and per-factor credit traces, making causal credit assignment, leave-one-out cross-archetype transfer, and policy auditability first-class metrics. We adapt a panel of reference baselines: random, heuristic, masked PPO, a causal-world-model PPO variant, and an architecture-matched scalar control. We propose Causal Graph-Factored Advantage PPO (CGFA-PPO) as a reference causal agent that uses SCM parents of win probability as factor-aligned critic targets with an intervention-calibration loss. All comparisons use paired seeds, paired-bootstrap confidence intervals, and Holm-Bonferroni correction within pre-registered families. Masked PPO and CGFA-PPO reach competitive in-distribution win rates and exceed the random baseline; per-factor calibration trajectories and leave-one-out transfer gaps expose diagnostic structure that scalar win rate alone cannot. We release the benchmark, reference-baseline results, and full evaluation protocol openly. By coupling a strategically rich, partially observed domain with an explicit causal interface and statistical protocol, MTG-Causal-RL gives causal-RL, world-model, and LLM-agent research a shared testbed for questions current benchmarks cannot pose together: causal credit assignment under masked action spaces, structural transfer across archetypes, and SCM-grounded policy auditability.\n\n---\n\n## 111. Optimal Transport for LLM Reward Modeling from Noisy Preference\n- **Authors**: Licheng Pan , Haochen Yang , Haoxuan Li , Yunsheng Lu , Yongqi Tong , Yinuo Wang , Shijian Wang , Zhixuan Chu , Lei Shen , Yuan Lu , Hao Wang\n- **URL**: [https://arxiv.org/abs/2605.06036](https://arxiv.org/abs/2605.06036)\n- **Abstract**:\n> Reward models are fundamental to Reinforcement Learning from Human Feedback (RLHF), yet real-world datasets are inevitably corrupted by noisy preference. Conventional training objectives tend to overfit these errors, while existing denoising approaches often rely on homogeneous noise assumptions that fail to capture the complexity of linguistic preferences. To handle these challenges, we propose SelectiveRM, a framework grounded in optimal transport. We first devise a Joint Consistency Discrepancy to align the distribution of model predictions with preference data. Furthermore, to address the limitation of strict mass conservation which compels the model to fit outliers, we incorporate a Mass Relaxation mechanism via partial transport. This enables the autonomous exclusion of samples with noisy preference that contradict semantic consistency. Theoretically, we demonstrate that SelectiveRM optimizes a tighter upper bound on the unobserved clean risk. Extensive experiments validate that our approach significantly outperforms state-of-the-art baselines across diverse benchmarks.\n\n---\n\n## 112. Does Synthetic Data Help? Empirical Evidence from Deep Learning Time Series Forecasters\n- **Authors**: Hugo Cazaux , Eyjólfur Ingi Ásgeirsson , Hlynur Stefánsson\n- **URL**: [https://arxiv.org/abs/2605.06032](https://arxiv.org/abs/2605.06032)\n- **Abstract**:\n> Synthetic data has transformed language model training, yet its role in time series forecasting remains poorly understood. We present a large-scale empirical study: nine experiment groups, 4,218 runs systematically evaluating synthetic time series augmentation across five architectures, four synthetic signals and seven datasets. The effect is sharply architecture-conditional: channel-mixing models (TimesNet, iTransformer) benefit in the majority of trials, while channel-independent models (DLinear, PatchTST) are consistently degraded. In selected low-resource settings the gains are striking: TimesNet trained on only 10\\% of Weather data with synthetic augmentation surpasses the full-data baseline (4 of 16 sparsity-dataset combinations). Averaged across all architectures, augmentation hurts in 67\\% of trials. We further find that only the Seasonal-Trend generator reliably helps across the tested benchmarks, and that hard curriculum switching is actively harmful (+24\\% MSE degradation). These results provide concrete, actionable guidelines on how to use synthetic data: use synthetic augmentation with channel-mixing architectures, use gradual annealing schedules, and treat low-resource augmentation as architecture- and dataset-dependent. Code is available at \\href{ this https URL }\n\n---\n\n## 113. Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks\n- **Authors**: Guoxin Lu , Letian Sha , Qing Wang , Peijie Sun , Hao Zhou , Hua Dai , Fu Xiao\n- **URL**: [https://arxiv.org/abs/2605.05995](https://arxiv.org/abs/2605.05995)\n- **Abstract**:\n> The safety alignment of Large Language Models (LLMs) remains vulnerable to Harmful Fine-tuning (HFT). While existing defenses impose constraints on parameters, gradients, or internal representations, we observe that they can be effectively circumvented under persistent HFT. Our analysis traces this failure to the inherent redundancy of the high-dimensional parameter space: attackers exploit optimization trajectories that are orthogonal to defense constraints to restore harmful capabilities while deceptively adhering to safety restrictions. To address this, we propose Safety Bottleneck Regularization (SBR). SBR shifts the defensive focus from the redundant parameter space to the unembedding layer, which serves as a geometric bottleneck. By anchoring the final hidden states of harmful queries to those of the safety-aligned model, SBR enables the model to maintain safe responses even under persistent HFT. Extensive experiments confirm SBR's effectiveness, demonstrating that utilizing just a single safety anchor is sufficient to reduce the Harmful Score to $<$10 while preserving competitive performance on benign downstream tasks.\n\n---\n\n## 114. PragLocker: Protecting Agent Intellectual Property in Untrusted Deployments via Non-Portable Prompts\n- **Authors**: Qinfeng Li , Yuntai Bao , Jianghui Hu , Wenqi Zhang , Jintao Chen , Huifeng Zhu , Yier Jin , Xuhong Zhang\n- **URL**: [https://arxiv.org/abs/2605.05974](https://arxiv.org/abs/2605.05974)\n- **Abstract**:\n> LLM agents rely on prompts to implement task-specific capabilities based on foundation LLMs, making agent prompts valuable intellectual property. However, in untrusted deployments, adversaries can copy and reuse these prompts with other proprietary LLMs, causing economic losses. To protect these prompts, we identify four key challenges: proactivity, runtime protection, usability, and non-portability that existing approaches fail to address. We present PragLocker, a prompt protection scheme that satisfies these requirements. PragLocker constructs function-preserving obfuscated prompts by anchoring semantics with code symbols and then using target-model feedback to inject noise, yielding prompts that only work on the target LLM. Experiments across multiple agent systems, datasets, and foundation LLMs show that PragLocker substantially reduces cross-LLM portability, maintains target performance, and remains robust against adaptive attackers.\n\n---\n\n## 115. Towards Reliable LLM Evaluation: Correcting the Winner's Curse in Adaptive Benchmarking\n- **Authors**: Yang Xu , Jiefu Zhang , Haixiang Sun , Zihan Zhou , Tianyu Cao , Vaneet Aggarwal\n- **URL**: [https://arxiv.org/abs/2605.05973](https://arxiv.org/abs/2605.05973)\n- **Abstract**:\n> Adaptive prompt and program search makes LLM evaluation selection-sensitive. Once benchmark items are reused inside tuning, the observed winner's score need not estimate the fresh-data performance of the full tune-then-deploy procedure. We study inference for this procedure-level target under explicit tuning budgets. We propose SIREN, a selection-aware repeated-split reporting protocol that freezes the post-search shortlist, separates splitwise selection from held-out evaluation, and uses an item-level Gaussian multiplier bootstrap for uncertainty quantification. In a fixed-shortlist regime with smooth stabilized selection, the estimator admits a first-order item-level representation, and the bootstrap yields valid simultaneous inference on a finite budget grid. This supports confidence intervals for procedure-performance curves and pre-specified equal-budget and cross-budget comparisons. Controlled simulations and MMLU-Pro tuning experiments show that winner-based reporting can be optimistic and can change deployment conclusions, while SIREN remains close to the finite-sample reporting target.\n\n---\n\n## 116. Beyond Uniform Credit Assignment: Selective Eligibility Traces for RLVR\n- **Authors**: Chaoli Mou , Zhan Zhuang , Xinning Chen , Yu Zhang\n- **URL**: [https://arxiv.org/abs/2605.05965](https://arxiv.org/abs/2605.05965)\n- **Abstract**:\n> Reinforcement Learning with Verifiable Rewards (RLVR) has become a key approach for improving the reasoning abilities of large language models. However, widely used critic-free algorithms such as Group Relative Policy Optimization (GRPO) necessitate a ``uniform credit assignment'' assumption that indiscriminately broadcast trajectory-level advantages, hindering learning efficiency by failing to distinguish critical reasoning steps. To address this limitation, we propose Selective Eligibility Traces (S-trace). Grounded in the intuition of partial trust region preservation, we initially introduce P-trace as a sample-efficient, critic-free eligibility traces method, upon which we build S-trace, implementing a sparse eligibility traces mechanism to further mitigate variance and achieve fine-grained credit assignment by selectively masking low-entropy tokens. Theoretically, we contextualize the recent Group Sequence Policy Optimization (GSPO) method within the critic-free eligibility traces framework, identifying it as a special instance of the eligibility traces method operating under uniform credit assignment. Experiments demonstrate that S-trace not only outperforms GRPO, showing gains of 0.49\\% on Qwen3-1.7B and 3.16\\% on Qwen3-4B, and maintaining a robust 2.98\\% improvement when scaled further to Qwen3-8B in average pass@16, but notably achieves this with simultaneously higher sample and token efficiency.\n\n---\n\n## 117. Hallucination as an Anomaly: Dynamic Intervention via Probabilistic Circuits\n- **Authors**: Erik Nielsen , Elia Cunegatti , Marcus Vukojevic , Giovanni Iacca\n- **URL**: [https://arxiv.org/abs/2605.05953](https://arxiv.org/abs/2605.05953)\n- **Abstract**:\n> One of the most critical challenges in Large Language Models is their tendency to hallucinate, i.e., produce factually incorrect responses. Existing approaches show promising results in terms of hallucination correction, but still suffer from a main limitation: they apply corrections indiscriminately to every token, corrupting also the originally correct generations. To overcome this drawback, we propose PCNET, a Probabilistic Circuit trained as a tractable density estimator over the LLM residual stream. The method detects hallucinations as geometric anomalies on the factual manifold, which is done via exact Negative Log-Likelihood computation, hence without the need for sampling, external verifiers, or weight modifications, as in existing techniques. To demonstrate its effectiveness, we exploit PCNET as a dynamic gate that distinguishes hallucinated from factual hidden states at each decoding step. This triggers our second main contribution, PC-LDCD (Probabilistic Circuit Latent Density Contrastive Decoding), only when the latent geometry deviates from factual regions, while leaving correct generations untouched. Across four LLMs, ranging from 1B to 8B models, and four benchmarks covering conversational reasoning, knowledge-intensive QA, reading comprehension, and truthfulness, PCNET achieves near-perfect hallucination detection across CoQA, SQuAD v2.0, and TriviaQA, with AUROC reaching up to 99%. Moreover, PC-LDCD obtains the highest True+Info, MC2, and MC3 scores on TruthfulQA in three out of four models, in comparison with state-of-the-art baselines, while reducing the mean corruption rate to 53.7% and achieving a preservation rate of 79.3%. Our proposed method is publicly available on GitHub.\n\n---\n\n## 118. LLM-Driven Design Space Exploration of FPGA-based Accelerators\n- **Authors**: Vinamra Sharma , Xingjian Fu , Jude Haris , José Cano\n- **URL**: [https://arxiv.org/abs/2605.05920](https://arxiv.org/abs/2605.05920)\n- **Abstract**:\n> Designing field-programmable gate array (FPGA)-based accelerators for modern artificial intelligence workloads requires navigating a large and complex hardware design space encompassing architectural parameters, dataflow strategies, and memory hierarchies, making the process time-consuming and resource-intensive. While the SECDA methodology enables rapid hardware-software co-design of accelerators through SystemC simulation and FPGA execution, identifying optimal accelerator configurations still requires substantial manual effort and domain expertise. This work presents SECDA-DSE, a framework that integrates Large Language Models (LLMs) into the SECDA ecosystem, comprising tools built around SECDA to automate the design space exploration (DSE) of FPGA-based accelerators. SECDA-DSE combines a structured DSE Explorer for generating accelerator configurations with an LLM Stack that performs reasoning-guided exploration using retrieval-augmented generation and chain-of-thought prompting, alongside a feedback loop that enables reinforced fine-tuning for continuous improvement. We demonstrate the feasibility of SECDA-DSE through an initial high-level synthesis based evaluation of a generated accelerator design that meets synthesis timing and resource constraints on an Zynq-7000 FPGA.\n\n---\n\n## 119. Quantum-enhanced Large Language Models on Quantum Hardware via Cayley Unitary Adapters\n- **Authors**: Borja Aizpurua , Sukhbinder Singh , Augustine Kshetrimayum , Saeed S. Jahromi , Roman Orus\n- **URL**: [https://arxiv.org/abs/2605.05914](https://arxiv.org/abs/2605.05914)\n- **Abstract**:\n> Large language models (LLMs) have transformed artificial intelligence, yet classical architectures impose a fundamental constraint: every trainable parameter demands classical memory that scales unfavourably with model size. Quantum computing offers a qualitatively different pathway, but practical demonstrations on real hardware have remained elusive for models of practical relevance. Here we show that Cayley-parameterised unitary adapters -- quantum circuit blocks inserted into the frozen projection layers of pre-trained LLMs and executed on a 156-qubit IBM Quantum System Two superconducting processor -- improve the perplexity of Llama 3.1 8B, an 8-billion-parameter model in widespread use, by 1.4% with only 6,000 additional parameters and end-to-end inference validated on real Quantum Processing Unit (QPU). A systematic study on SmolLM2 (135M parameters), chosen for its tractability, reveals monotonically improving perplexity with unitary block dimension, 83% recovery of compression-induced degradation, and correct answers to questions that both classical baselines fail -- with a sharp noise-expressivity phase transition identifying the concrete path to quantum utility at larger qubit scales.\n\n---\n\n## 120. CITE: Anytime-Valid Statistical Inference in LLM Self-Consistency\n- **Authors**: Hirofumi Ota , Naoto Iwase , Yuki Ichihara , Junpei Komiyama , Masaaki Imaizumi\n- **URL**: [https://arxiv.org/abs/2605.05873](https://arxiv.org/abs/2605.05873)\n- **Abstract**:\n> Large language models often improve reasoning by sampling multiple outputs and aggregating their final answers, but precise and efficient control of error levels remains a challenging task. In particular, deciding when to stop sampling remains difficult when the stopping rule is data-dependent and the set of possible answers is not known in advance. We study anytime-valid certification of a prespecified target answer as the unique mode of the model's response distribution, a guarantee distinct from answer correctness. We propose the Certification by Intersection-union Testing with E-processes (CITE) algorithm, which provably controls false certification at any prescribed level under arbitrary data-driven stopping, without requiring prior knowledge of the answer category set. We also prove an category-set-size-free stopping-time rate, establish matching minimax lower bounds up to constants in the main regime, and extend the construction to confidence-weighted voting. Simulations and LLM self-consistency experiments show empirical error control and improved certification in diffuse-tail settings.\n\n---\n\n## 121. VideoRouter: Query-Adaptive Dual Routing for Efficient Long-Video Understanding\n- **Authors**: Kuanwei Lin , Wenhao Zhang , Ge Li\n- **URL**: [https://arxiv.org/abs/2605.05848](https://arxiv.org/abs/2605.05848)\n- **Abstract**:\n> Video large multimodal models increasingly face a scalability bottleneck: long videos produce excessively long visual-token sequences, which sharply increase memory and latency during inference. While existing compression methods are effective in specific settings, most are either weakly query-aware or apply a fixed compression policy across frames, proving suboptimal when visual evidence is unevenly distributed over time. To address this, we present VideoRouter, a query-adaptive dual-router framework built on InternVL for budgeted evidence allocation. The Semantic Router predicts the dominant allocation policy, choosing between broad temporal coverage and adaptive high-resolution preservation, while the Image Router uses early LLM layers to score frame relevance. This enables aggressive compression on less relevant frames while preserving detail on critical evidence frames. To train both routers, we build Video-QTR-10K for allocation-policy supervision and Video-FLR-200K for frame-relevance supervision. Experiments on VideoMME, MLVU, and LongVideoBench show that VideoRouter consistently improves over the InternVL baseline under comparable or lower budgets, achieving up to a 67.9% token reduction.\n\n---\n\n## 122. LoopTrap: Termination Poisoning Attacks on LLM Agents\n- **Authors**: Huiyu Xu , Zhibo Wang , Wenhui Zhang , Ziqi Zhu , Yaopeng Wang , Kui Ren , Chun Chen\n- **URL**: [https://arxiv.org/abs/2605.05846](https://arxiv.org/abs/2605.05846)\n- **Abstract**:\n> Modern LLM agents solve complex tasks by operating in iterative execution loops, where they repeatedly reason, act, and self-evaluate progress to determine when a task is complete. In this work, we show that while this self-directed loop facilitates autonomy, it also introduces a critical risk: by injecting malicious prompts into the agent's context, an adversary can distort the agent's termination judgment, making it believe the task remains incomplete and leading to unbounded this http URL understand this threat, we define and systematically characterize it as Termination Poisoning and design 10 representative attack strategies. Through a empirical study spanning 8 LLM agents and 60 tasks, we demonstrate that different LLM agents exhibit distinct behavioral signatures that determine which strategies succeed. These transferable patterns can serve as principled guidance for crafting effective attacks against previously unseen agents and tasks, enabling scalable red-teaming beyond manually designed templates. Building on these insights, we introduce LoopTrap, an automated red-teaming framework that synthesizes target-specific malicious prompts by exploiting agent behavioral tendencies. LoopTrap first constructs a behavioral profile of the target agent along four vulnerability dimensions via lightweight probing. It then performs adaptive trap synthesis, routing to the most effective strategy and selecting optimal injections via a self-scoring mechanism. Finally, successful traps are abstracted into a reusable skill library, while failed attempts are refined through self-reflection, ensuring continuous improvement. Extensive evaluation shows that LoopTrap achieves an average of 3.57$\\times$ step amplification across 8 mainstream agents, with a peak of 25$\\times$.\n\n---\n\n## 123. LeakDojo: Decoding the Leakage Threats of RAG Systems\n- **Authors**: Maosen Zhang , Jianshuo Dong , Boting Lu , Wenyue Li , Xiaoping Zhang , Tianwei Zhang , Han Qiu\n- **URL**: [https://arxiv.org/abs/2605.05818](https://arxiv.org/abs/2605.05818)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to leverage external knowledge, but also exposes valuable RAG databases to leakage attacks. As RAG systems grow more complex and LLMs exhibit stronger instruction-following capabilities, existing studies fall short of systematically assessing RAG leakage risks. We present LeakDojo, a configurable framework for controlled evaluation of RAG leakage. Using LeakDojo, we benchmark six existing attacks across fourteen LLMs, four datasets, and diverse RAG systems. Our study reveals that (1) query generation and adversarial instructions contribute independently to leakage, with overall leakage well approximated by their product; (2) stronger instruction-following capability correlates with higher leakage risk; and (3) improvements in RAG faithfulness can introduce increased leakage risk. These findings provide actionable insights for understanding and mitigating RAG leakage in practice. Our codebase is available at this https URL .\n\n---\n\n## 124. LCC-LLM: Leveraging Code-Centric Large Language Models for Malware Attribution\n- **Authors**: Christopher G. Pedraza Pohlenz , Hassan Jalil Hadi , Ali Hassan , Ali Shoker\n- **URL**: [https://arxiv.org/abs/2605.05807](https://arxiv.org/abs/2605.05807)\n- **Abstract**:\n> LLMs are increasingly explored for malware analysis; however, current LLM-based malware attribution remains limited by unsupported indicators and insufficient code-level grounding for identifying malicious and vulnerable code segments. To address these limitations, this research introduces LCC-LLM, a code-centric benchmark dataset and evidence-grounded framework for malware attribution and multi-task static malware analysis. The proposed LCCD dataset contains approximately 34K PE samples processed through a large-scale reverse-engineering pipeline and represented using decompiled C code, assembly code, CFG/FCG artifacts, hexadecimal data, PE metadata, suspicious API evidence, and structural features. Beyond dataset construction, LCC-LLM integrates LangGraph-orchestrated static analysis with multi-source cybersecurity knowledge to support evidence-grounded malware reasoning. The framework employs a seven-layer retrieval-augmented generation pipeline, CoVe for IoC validation, and a multi-dimensional quality gate to improve factual reliability and analyst-oriented decision support. Curriculum-ordered instruction data is used to fine-tune DeepSeek-R1-Distill-Qwen-14B and Qwen3-Coder-30B-A3B using QLoRA. Evaluation across 43 malware-analysis task types achieves an average semantic similarity of 0.634, with the highest task-level performance in structured report generation, IoC extraction, vulnerability assessment, malware configuration extraction, and malware class detection. In a real-world case study using MalwareBazaar samples, the grounded pipeline achieves a 10/10 structured analysis pass rate, producing CFG/FCG evidence, MITRE ATT&CK mappings, detection guidance, and analyst-ready reports. These results show that code-centric representations, retrieval grounding, and verification-guided reasoning improve the reliability and operational usefulness of LLM-assisted malware attribution.\n\n---\n\n## 125. Revealing Modular Gradient Noise Imbalance in LLMs: Calibrating Adam via Signal-to-Noise Ratio\n- **Authors**: Ziqing Wen , Zhouyang Liu , Jiahuan Wang , Ping Luo , Li Shen , Dongsheng Li , Tao Sun\n- **URL**: [https://arxiv.org/abs/2605.05794](https://arxiv.org/abs/2605.05794)\n- **Abstract**:\n> The impressive performance of large language models (LLMs) arises from their massive scale and heterogeneous module composition. However, this structural heterogeneity introduces additional optimization challenges. While adaptive optimizers such as Adam(W) provide per-parameter adaptivity, they do not explicitly account for module-level gradient heterogeneity, resulting in slower convergence, suboptimal performance, or training instability. Existing approaches typically rely on manually tuned module-specific learning rates or specific optimization strategies, which are computationally costly and difficult to generalize across tasks or models. To establish a more principled approach, we first analyze the noise-damping behavior of Adam in high-noise modules and introduce \\textbf{Module-wise Learning Rate Scaling via SNR (MoLS)}. MoLS estimates module-level SNRs to scale Adam updates, allowing automated module-wise learning rate allocation without manual tuning. Empirical results through multiple LLM training benchmarks demonstrate that MoLS improves convergence speed and generalization, achieving performance comparable to carefully tuned module-specific learning rates, while remaining compatible with memory-efficient training algorithms.\n\n---\n\n## 126. Multi-Dimensional Behavioral Evaluation of Agentic Stock Prediction Systems Using LLM Judges with Closed-Loop Reinforcement Learning Feedback\n- **Authors**: Mohammad Al Ridhawi , Mahtab Haj Ali , Hussein Al Osman\n- **URL**: [https://arxiv.org/abs/2605.05739](https://arxiv.org/abs/2605.05739)\n- **Abstract**:\n> Agentic stock prediction systems make sequences of interdependent decisions (regime detection, pathway routing, reinforcement learning control) whose individual quality is hidden by aggregate metrics such as mean absolute percentage error (MAPE) or directional accuracy. We present a behavioral evaluation framework that addresses this gap. Behavioral traces logged at every autonomous decision point are grouped into five-day episodes and scored along six domain-specific dimensions (regime detection, routing, adaptation, risk calibration, strategy coherence, error recovery) by an ensemble of three large language model (LLM) judges (GPT 5.4, Claude 4.6 Opus, Gemini 3.1 Pro). Perturbation-based validation on 420 episodes yields targeted score drops of $-1.6$ to $-2.4$ on intended dimensions versus an average of $-0.32$ on the remaining five, with cross-model agreement up to Krippendorff's $\\alpha = 0.85$. The composite behavioral score, used here only for cross-episode reporting, correlates at $\\rho = 0.72$ with realized 20-day Sharpe ratio from offline backtesting. Closing the loop, the framework converts deficient per-dimension scores into a credit-assigned penalty term added to the Soft Actor-Critic (SAC) reward. Three short fine-tuning cycles, all confined to the validation period, produce on the held-out 2017-2025 test period a one-day MAPE reduction from 0.61% to 0.54% (an 11.5% relative reduction; $p<0.001$, Cohen's $d=0.31$), a directional accuracy increase from 71% to 74%, and an 18% Sharpe ratio improvement (95% bootstrap CI [8.2%, 27.4%]), with gains concentrated in high-volatility episodes where the original system was most behaviorally deficient. Results are from offline backtesting and do not address effects specific to live deployment.\n\n---\n\n## 127. CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning\n- **Authors**: Md Anwar Hossen , Fatema Siddika , Juan Pablo Munoz , Tanya Roosta , Ali Jannesari\n- **URL**: [https://arxiv.org/abs/2605.05732](https://arxiv.org/abs/2605.05732)\n- **Abstract**:\n> Large language models (LLMs) can acquire new capabilities through fine-tuning, but continual adaptation often leads to catastrophic forgetting. We propose CRAFT, a continual learning framework that avoids updating model weights by instead learning low-rank interventions on hidden representations. CRAFT proceeds in three stages: it first routes each task to a group of similar tasks based on output-distribution divergence; it then fine-tunes the model using a Kullback-Leibler (KL) divergence against the group's prior state, which directly controls forgetting and determines convergence; finally, it merges interventions for the updated task into the shared representation using the same KL signal. This design unifies routing, regularization, and merging through a single KL-based objective. CRAFT improves overall performance and reduces forgetting compared to strong LoRA-based approaches across multiple benchmarks and model scales, while remaining robust to task ordering. These results suggest that controlling adaptation in representation space, guided by output-space divergence, provides a scalable and principled approach to continual learning in LLMs.\n\n---\n\n## 128. SafeHarbor: Hierarchical Memory-Augmented Guardrail for LLM Agent Safety\n- **Authors**: Zhe Liu , Zonghao Ying , Wenxin Zhang , Quanchen Zou , Deyue Zhang , Dongdong Yang , Xiangzheng Zhang , Hao Peng\n- **URL**: [https://arxiv.org/abs/2605.05704](https://arxiv.org/abs/2605.05704)\n- **Abstract**:\n> With the rapid evolution of foundation models, Large Language Model (LLM) agents have demonstrated increasingly powerful tool-use capabilities. However, this proficiency introduces significant security risks, as malicious actors can manipulate agents into executing tools to generate harmful content. While existing defensive mechanisms are effective, they frequently suffer from the over-refusal problem, where increased safety strictness compromises the agent's utility on benign tasks. To mitigate this trade-off, we propose \\textsc{SafeHarbor}, a novel framework designed to establish precise decision boundaries for LLM agents. Unlike static guidelines, \\textsc{SafeHarbor} extracts context-aware defense rules through enhanced adversarial generation. We design a local hierarchical memory system for dynamic rule injection, offering a training-free, efficient, and plug-and-play solution. Furthermore, we introduce an information entropy-based self-evolution mechanism that continuously optimizes the memory structure through dynamic node splitting and merging. Extensive experiments demonstrate that \\textsc{SafeHarbor} achieves state-of-the-art performance on both ambiguous benign tasks and explicit malicious attacks, notably attaining a peak benign utility of 63.6\\% on GPT-4o while maintaining a robust refusal rate exceeding 93\\% against harmful requests. The source code is publicly available at this https URL .\n\n---\n\n## 129. Active Learning for Communication Structure Optimization in LLM-Based Multi-Agent Systems\n- **Authors**: Huchen Yang , Xinghao Dong , Dan Negrut , Jin-Long Wu\n- **URL**: [https://arxiv.org/abs/2605.05703](https://arxiv.org/abs/2605.05703)\n- **Abstract**:\n> Optimizing the communication structure of large language model based multi-agent systems (LLM-MAS) has been shown to improve downstream performance and reduce token usage. Existing methods typically rely on randomly sampled training tasks. However, tasks may differ substantially in difficulty and domain, and thus they are not equally informative for updating communication structure, making optimization under limited training budgets often unstable and highly sensitive to the particular training set. To actively identify the most valuable tasks for communication-structure optimization, we propose an ensemble-based information-theoretic task selection framework. The proposed method estimates task informativeness by how much a candidate task changes the distribution over graph parameters, using ensemble Kalman inversion as an efficient and derivative-free approximation of the corresponding Bayesian update. The resulting estimator is especially suitable for black-box and noisy multi-agent systems. To enhance scalability, we construct a compact candidate pool through embedding-based representative selection and combine the informative selection with surrogate modeling and batch Thompson sampling. We validate our method in both benign settings and settings with agent attacks, demonstrating its effectiveness for communication-structure optimization under constrained computational budgets.\n\n---\n\n## 130. An Empirical Study of Proactive Coding Assistants in Real-World Software Development\n- **Authors**: Lehui Li , Ruixuan Jia , Guo-Ye Yang , Jia Li\n- **URL**: [https://arxiv.org/abs/2605.05700](https://arxiv.org/abs/2605.05700)\n- **Abstract**:\n> Large language model (LLM)-based coding assistants have made substantial progress, yet most systems remain reactive, requiring developers to explicitly formulate their needs. Proactive coding assistants aim to infer latent developer intent from integrated development environment (IDE) interactions and repository context, thereby reducing interaction overhead and supporting more seamless assistance. However, research in this direction is limited by the scarcity of large-scale real-world developer behavior data. Existing studies therefore often rely on LLM-simulated IDE traces, whose fidelity to real development behavior remains unclear. In this paper, we investigate this simulation-to-reality gap through a large-scale empirical study. We collect real IDE interaction traces from 1{,}246 experienced industry developers over three consecutive days using a custom Visual Studio Code extension, and construct paired LLM-simulated traces for controlled comparison. Our analysis shows that simulated traces differ substantially from real traces in behavioral diversity, temporal structure, and exploratory patterns. Based on the collected data, we introduce \\textbf{ProCodeBench}, a real-world benchmark for proactive intent prediction. Experiments with representative LLMs, retrieval-augmented methods, and agentic baselines show that current approaches remain far from reliable under real IDE traces, suggesting that simulation-based evaluation can overestimate real-world performance. Finally, our training study shows that simulated data cannot replace real data, but can complement it when used before real-world fine-tuning. These findings highlight the importance of real developer behavior data for evaluating and training proactive coding assistants.\n\n---\n\n## 131. Irminsul: MLA-Native Position-Independent Caching for Agentic LLM Serving\n- **Authors**: Bole Ma , Jan Eitzinger , Harald Köstler\n- **URL**: [https://arxiv.org/abs/2605.05696](https://arxiv.org/abs/2605.05696)\n- **Abstract**:\n> Agentic LLM workloads put bit-identical tokens at shifted positions every turn, voiding prefix caches at the first byte of divergence. Operators report cache-hit regressions ranging from moderate slowdowns to severe TTFT spikes of 10-16s on unchanged content. Prior position-independent caching systems correct RoPE on the full $d_K$-dimensional key, an architectural cost imposed by GQA, not by caching itself. Multi-Head Latent Attention, deployed at scale in DeepSeek-V2/V3/R1, Kimi-K2/Moonlight, GLM-5, and Mistral Large 3, factors each KV row into a position-free $c_{KV}$ and a 64-dim $k_r$ correctable in closed form; this structure motivates content-addressed caching as a natural fit rather than a GQA workaround. We present Irminsul, which extends SGLang's radix cache with content-hash keying over CDC-chunked segments and a $\\delta$-rotation rule for $k_r$. We evaluate three native MLA-MoE deployments - DeepSeek-V2-Lite (16B/2.4B), Kimi Moonlight-16B-A3B, and JoyAI-Flash (48B/3B) - with output-consistency on all three and recovery measured on the two endpoints; Irminsul recovers up to ~83% of prompt tokens above exact-prefix on agentic traffic while delivering 63% prefill energy savings per cache hit. We argue that content-addressed caching belongs in the serving stack as a first-class primitive, not a retrofit over prefix matching.\n\n---\n\n## 132. Decomposing the Basic Abilities of Large Language Models: Mitigating Cross-Task Interference in Multi-Task Instruct-Tuning\n- **Authors**: Bing Wang , Ximing Li , Changchun Li , Jinjin Chi , Gang Niu , Masashi Sugiyama\n- **URL**: [https://arxiv.org/abs/2605.05676](https://arxiv.org/abs/2605.05676)\n- **Abstract**:\n> Recently, the prominent performance of large language models (LLMs) has been largely driven by multi-task instruct-tuning. Unfortunately, this training paradigm suffers from a key issue, named cross-task interference, due to conflicting gradients over shared parameters among different tasks. Some previous methods mitigate this issue by isolating task-specific parameters, e.g., task-specific neuron selection and mixture-of-experts. In this paper, we empirically reveal that the cross-task interference still exists for the existing solutions because of many parameters also shared by different tasks, and accordingly, we propose a novel solution, namely Basic Abilities Decomposition for multi-task Instruct-Tuning (BADIT). Specifically, we empirically find that certain parameters are consistently co-activated, and that co-activated parameters naturally organize into base groups. This motivates us to analogize that LLMs encode several orthogonal basic abilities, and that any task can be represented as a linear combination of these abilities. Accordingly, we propose BADIT that decomposes LLM parameters into orthogonal high-singular-value LoRA experts representing basic abilities, and dynamically enforces their orthogonality during training via spherical clustering of rank-1 components. We conduct extensive experiments on the SuperNI benchmark with 6 LLMs, and empirical results demonstrate that BADIT can outperform SOTA methods and mitigate the degree of cross-task interference.\n\n---\n\n## 133. XL-SafetyBench: A Country-Grounded Cross-Cultural Benchmark for LLM Safety and Cultural Sensitivity\n- **Authors**: Dasol Choi , Eugenia Kim , Jaewon Noh , Sang Seo , Eunmi Kim , Myunggyo Oh , Yunjin Park , Brigitta Jesica Kartono , Josef Pichlmeier , Helena Berndt , Sai Krishna Mendu , Glenn Johannes Tungka , Özlem Gökçe , Suresh Gehlot , Katherine Pratt , Amanda Minnich , Haon Park\n- **URL**: [https://arxiv.org/abs/2605.05662](https://arxiv.org/abs/2605.05662)\n- **Abstract**:\n> Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Moreover, they rarely evaluate a model's ability to detect culturally embedded sensitivities as distinct from universal harms. We introduce XL-SafetyBench. a suite of 5,500 test cases across 10 country-language pairs, comprising a Jailbreak Benchmark of country-grounded adversarial prompts and a Cultural Benchmark where local sensitivities are embedded within innocuous requests. Each item is constructed via a multi-stage pipeline that combines LLM-assisted discovery, automated validation gates, and dual independent native-speaker annotators per country. To distinguish principled refusal from comprehension failure, we evaluate Attack Success Rate (ASR) alongside two complementary metrics we introduce: Neutral-Safe Rate (NSR) and Cultural Sensitivity Rate (CSR). Evaluating 10 frontier and 27 local LLMs reveals two key findings. First, jailbreak robustness and cultural awareness do not show a coupled relationship among frontier models, so a composite safety score obscures per-axis variation. Second, local models exhibit a near-linear ASR-NSR trade-off (r = -0.81), indicating that their apparent safety reflects generation failure rather than genuine alignment. XL-SafetyBench enables more nuanced, cross-cultural safety evaluation in the multilingual era.\n\n---\n\n## 134. One Turn Too Late: Response-Aware Defense Against Hidden Malicious Intent in Multi-Turn Dialogue\n- **Authors**: Xinjie Shen , Rongzhe Wei , Peizhi Niu , Haoyu Wang , Ruihan Wu , Eli Chien , Bo Li , Pin-Yu Chen , Pan Li\n- **URL**: [https://arxiv.org/abs/2605.05630](https://arxiv.org/abs/2605.05630)\n- **Abstract**:\n> Hidden malicious intent in multi-turn dialogue poses a growing threat to deployed large language models (LLMs). Rather than exposing a harmful objective in a single prompt, increasingly capable attackers can distribute their intent across multiple benign-looking turns. Recent studies show that even modern commercial models with advanced guardrails remain vulnerable to such attacks despite advances in safety alignment and external guardrails. In this work, we address this challenge by detecting the earliest turn at which delivering the candidate response would make the accumulated interaction sufficient to enable harmful action. This objective requires precise turn-level intervention that identifies the harm-enabling closure point while avoiding premature refusal of benign exploratory conversations. To further support training and evaluation, we construct the Multi-Turn Intent Dataset (MTID), which contains branching attack rollouts, matched benign hard negatives, and annotations of the earliest harm-enabling turns. We show that MTID helps enable a turn-level monitor TurnGate, which substantially outperforms existing baselines in harmful-intent detection while maintaining low over-refusal rates. TurnGate further generalizes across domains, attacker pipelines, and target models. Our code is available at this https URL .\n\n---\n\n## 135. Leveraging Image Generators to Address Training Data Scarcity: The Gen4Regen Dataset for Forest Regeneration Mapping\n- **Authors**: Gabriel Jeanson , David-Alexandre Duclos , William Larrivée-Hardy , Noé Cochet , Matěj Boxan , Anthony Deschênes , François Pomerleau , Philippe Giguère\n- **URL**: [https://arxiv.org/abs/2605.05627](https://arxiv.org/abs/2605.05627)\n- **Abstract**:\n> Sustainable forest management relies on precise species composition mapping, yet traditional ground surveys are labour-intensive and geographically constrained. While Uncrewed Aerial Vehicles (UAVs) offer scalable data collection, the transition to deep learning-based interpretation is bottlenecked by the severe scarcity of expert-annotated imagery, particularly in complex, visually heterogeneous regeneration zones. This paper addresses the dual challenges of data scarcity and extreme class imbalance in the semantic segmentation of fine-grained forest regeneration species by providing a scalable framework that reduces reliance on manual photo-interpretation for high-resolution, millimetre-level aerial imagery. Importantly, we leverage the large-scale vision-language Nano Banana Pro model to simultaneously generate high-fidelity images and their corresponding pixel-aligned semantic masks from prompts. We introduce WilDReF-Q-V2, an expansion of a natural forest dataset with 13 977 new unlabelled and 50 labelled real images, as well as the Gen4Regen dataset, featuring 2101 pairs of synthetic images and semantic masks. Our methodology integrates real-world data with AI-generated images, highlighting that AI-generated data is highly complementary to real-world data, with unified training yielding an F1 score improvement of over 15 %pt compared to purely supervised baselines. Furthermore, we demonstrate that even small quantities of prompt-generated data significantly improve performance for underrepresented species, some of which saw per-species F1 score gains of up to 30 %pt. We conclude that vision-language models can serve as agile data generators, effectively bootstrapping perception tasks for niche AI domains where expert labels are scarce or unavailable. Our datasets, source code, and models will be available at this https URL .\n\n---\n\n## 136. When2Speak: A Dataset for Temporal Participation and Turn-Taking in Multi-Party Conversations for Large Language Models\n- **Authors**: Vihaan Nama , Shreya Mendi , Zian Ye , Brinnae Bent\n- **URL**: [https://arxiv.org/abs/2605.05626](https://arxiv.org/abs/2605.05626)\n- **Abstract**:\n> Large Language Models (LLMs) excel at generating contextually appropriate responses but remain poorly calibrated for multi-party conversations, where deciding when to speak is as critical as what to say. In such settings, naively responding at every turn leads to excessive interruptions and degraded conversational coherence. We introduce When2Speak, a grounded synthetic dataset and four-stage generation pipeline for learning intervention timing in group interactions. The dataset comprises over 215,000 examples derived from 16,000 conversations involving 2-6 speakers, spanning diverse conversational styles, tones, and participant dynamics, and explicitly modeling SPEAK vs. SILENT decisions at each turn. Our pipeline combines real-world grounding, structured augmentation, controlled transcript synthesis, and fine-tuning-ready supervision, and is fully open-sourced to support reproducibility and adaptation to domain-specific conversational norms. Across multiple model families, supervised fine-tuning (SFT) on When2Speak significantly outperforms zero-shot baselines (e.g., the average Macro F1 increase across 4B+ parameter models was 60%, with the largest increase being 120%). However, SFT-trained models remain systematically over-conservative, missing nearly half of warranted interventions as seen through the Missed Intervention Rate (MIR), which was on average 0.50 and is noticed even at larger model sizes. To address this limitation, we apply reinforcement learning with asymmetric reward shaping, which reduces MIR to 0.186-0.218 and increases recall from 0.479 to 0.78-0.81. Our findings establish that temporal participation is a distinct and trainable dimension of conversational intelligence, and that grounded synthetic data provides an effective and scalable pathway for enabling LLMs to participate more naturally and appropriately in multi-party interactions.\n\n---\n\n## 137. AstroAlertBench: Evaluating the Accuracy, Reasoning, and Honesty of Multimodal LLMs in Astronomical Classification\n- **Authors**: Claire Chen , Jiabao Sean Xiao , Shuze Daniel Liu , Facundo Perez Paolino , Luke Handley , Theophile Jegou du Laz , Ricky Nilsson , Alice Zou , Matthew Graham , Ashish Mahabal\n- **URL**: [https://arxiv.org/abs/2605.05573](https://arxiv.org/abs/2605.05573)\n- **Abstract**:\n> Modern astronomical observatories generate a massive volume of multimodal data, creating a critical bottleneck for expert human review. While multimodal large language models (LLMs) have shown promise in interpreting complex visual and textual inputs, their ability to perform specialized scientific classification while providing interpretable reasoning remains understudied. We introduce AstroAlertBench, a comprehensive multimodal benchmark designed to evaluate LLM performance in astronomical event review along a three-stage logical chain: metadata grounding, scientific reasoning, and hierarchical classification over five categories. We use a pilot sample of 1,500 real-world alerts from the Zwicky Transient Facility (ZTF), a wide-field survey that scans the northern sky to detect transient astronomical events. On this dataset, we benchmark 13 frontier closed-source and open-weight LLMs that support visual input. Our results reveal that high accuracy does not always align with model ``honesty,'' defined as the ability to self-evaluate its reasoning, which affects its reliability as a real-world assistant. We further initialize a human-in-the-loop evaluation protocol as a precursor to future community-scale participation. Together, AstroAlertBench provides a framework for developing calibrated and interpretable astronomical assistants.\n\n---\n\n## 138. ReaComp: Compiling LLM Reasoning into Symbolic Solvers for Efficient Program Synthesis\n- **Authors**: Atharva Naik , Yash Mathur , Prakam , Carolyn Rose , David Mortensen\n- **URL**: [https://arxiv.org/abs/2605.05485](https://arxiv.org/abs/2605.05485)\n- **Abstract**:\n> LLMs can solve program synthesis tasks but remain inefficient and unreliable on hard instances requiring large combinatorial search. Given a small set of reasoning traces, we use coding agents to compile them into reusable symbolic program synthesizers over constrained DSLs. The resulting solvers require no LLM calls at test time and are strong standalone systems: symbolic solver ensembles reach 91.3% accuracy on PBEBench-Lite and 84.7% on PBEBench-Hard, outperforming LLMs with test-time scaling for the latter by +16.3 percentage points at zero LLM inference cost. They also complement LLM search, improving PBEBench-Hard accuracy from 68.4% to 85.8% while reducing reported token usage by 78%, and raising SLR-Bench hard-tier accuracy from 34.4% to 58.0% in a neuro-symbolic hybrid setting. Compared to directly using coding agents as per-instance solvers, induced solvers are substantially more Pareto-efficient, amortizing a small one-time construction cost over many zero-token executions. Finally, most solvers transfer zero-shot to a real historical linguistics task - predicting sound changes in natural language data - reaching 80.1% accuracy under ensembling and recovering some plausible linguistic rules. Together, these results show that reasoning traces can be compiled into reusable symbolic solvers that solve many tasks directly, complement LLM inference on hard cases, and provide a scalable route to domain-general solver induction. We release code and data for reproducibility.\n\n---\n\n## 139. Robustness of Graph Self-Supervised Learning to Real-World Noise: A Case Study on Text-Driven Biomedical Graphs\n- **Authors**: Othmane Kabal , Mounira Harzallah , Fabrice Guillet , Hideaki Takeda , Ryutaro Ichise\n- **URL**: [https://arxiv.org/abs/2605.05463](https://arxiv.org/abs/2605.05463)\n- **Abstract**:\n> Graph Self-Supervised Learning (GSSL) offers a powerful paradigm for learning graph representations without labeled data. However, existing work assumes clean, manually curated graphs. Recent advances in NLP enable the large-scale automatic extraction of knowledge graphs from text, opening new opportunities for GSSL while introducing substantial real-world noise. This type of noise remains largely unexplored, as prior robustness studies typically rely on synthetic perturbations. To address this gap, we present the first comprehensive evaluation of GSSL methods on text-driven graphs for unsupervised term typing. We introduce Noise-Aware Text-Driven Graph GSSL (NATD-GSSL), a unified framework that combines automatic graph construction, graph refinement, and GSSL. Our evaluation follows a dual-graph protocol that contrasts a noisy graph derived from MedMentions with a clean Unified Medical Language System (UMLS) reference graph, aligned through a shared gold standard. Our results reveal variability in robustness across both pretext tasks and Graph Neural Network (GNN) architectures. Relation reconstruction is highly sensitive to noise and benefits from well-defined schemas, whereas feature reconstruction is considerably more robust, achieving performance comparable to clean-graph settings. Contrastive objectives are generally less affected by noise but depend strongly on alignment with downstream tasks. GNN architecture also plays a critical role: bidirectional relational message-passing designs are better suited to noisy, text-driven graphs, while unidirectional relational ones perform best on clean graphs. Overall, NATD-GSSL provides practical guidance for applying GSSL to real-world, noisy graphs and achieves up to a 7\\% improvement over pretrained language model baselines. All code and benchmarks are publicly available at this https URL .\n\n---\n\n## 140. SLAM: Structural Linguistic Activation Marking for Language Models\n- **Authors**: Fabrice Harel-Canada , Amit Sahai\n- **URL**: [https://arxiv.org/abs/2605.05443](https://arxiv.org/abs/2605.05443)\n- **Abstract**:\n> LLM watermarks must be detectable without compromising text quality, yet most existing schemes bias the next-token distribution and pay for detection with measurable quality loss. We present SLAM (Structural Linguistic Activation Marking), a novel white-box watermarking scheme that sidesteps this cost by writing the mark into structural geometry rather than token frequencies: sparse autoencoders identify residual-stream directions encoding linguistic structure (e.g., voice, tense, clause order), and we causally steer those directions at generation time, leaving lexical sampling and semantics unconstrained. On Gemma-2 2B and 9B, SLAM achieves 100% detection accuracy with a quality cost of only 1-2 reward points - compared to 7.5-11.5 for KGW, EWD, and Unigram - with naturalness and diversity preserved at near-unwatermarked levels across both models. The trade-off is a complementary robustness profile: SLAM resists word-level edits but is vulnerable to paraphrase that restructures syntax (at a quality cost), the converse of token-distribution methods.\n\n---\n\n## 141. Information Theoretic Adversarial Training of Large Language Models\n- **Authors**: Yiwei Zhang , Jeremiah Birrell , Reza Ebrahimi , Rouzbeh Behnia , Jason Pacheco , Elisa Bertino\n- **URL**: [https://arxiv.org/abs/2605.05415](https://arxiv.org/abs/2605.05415)\n- **Abstract**:\n> Large language models (LLMs) remain vulnerable to adversarial prompting despite advances in alignment and safety, often exhibiting harmful behaviors under novel attack strategies. While adversarial training can improve robustness, existing approaches are computationally expensive and difficult to scale. Recent continuous adversarial training methods, such as Continuous adversarial training (CAT) and Continuous Adversarial Preference Optimization (CAPO), address this challenge by leveraging gradient-based perturbations in the embedding space, enabling more efficient and expressive attacks. Building on this paradigm, we propose WARDEN, a distributionally robust adversarial training framework for LLMs that dynamically reweights adversarial examples through an f -divergence ambiguity set around the empirical training distribution. Our method optimizes the worst-case adversarial loss within a divergence ball around the empirical data distribution, automatically emphasizing harder adversarial examples. Using the convex dual formulation, the objective reduces to a log-sum-exp form under the KL divergence, with a dynamical parameter controlling the strength of reweighting. This study leads to a new class of information-theoretic objectives that significantly reduce attack success rates while maintaining model utility. Across multiple LLMs and attack settings, WARDEN substantially reduces attack success rates with computational and utility costs comparable to CAT-, CAPO-, and MixAT-based baselines, making it a practical approach for scalable robust alignment.\n\n---\n\n## 142. Open-SAT: LLM-Guided Query Embedding Refinement for Open-Vocabulary Object Retrieval in Satellite Imagery\n- **Authors**: Md Adnan Arefeen , Biplob Debnath , Ravi K. Rajendran , Murugan Sankaradas , Srimat T. Chakradhar\n- **URL**: [https://arxiv.org/abs/2605.05344](https://arxiv.org/abs/2605.05344)\n- **Abstract**:\n> In satellite applications, user queries often take the form of open-ended natural language, extending beyond a fixed set of predefined categories. This open-vocabulary nature poses significant challenges for retrieving relevant image tiles, as the retrieval system must generalize to a wide range of unseen objects and concepts. While vision-language models (VLMs) such as CLIP are widely used for text-image retrieval, even fine-tuned variants often struggle to accurately align such queries with satellite imagery. To address this, we propose Open-SAT, a training-free query embedding refinement algorithm that operates at inference time to improve alignment between user queries and satellite image content. Open-SAT uses VLMs to compute embeddings for image tiles, which are stored in a vector database for efficient retrieval. At query time, it leverages Large Language Models (LLMs) to refine the text embeddings by incorporating contextual information about objects of interest and their surroundings. A threshold-free retrieval mechanism further enhances accuracy and efficiency. Experimental results in three public benchmarks demonstrate that Open-SAT improves the F1 score by up to 16.04%, while retrieving a comparable number of image tiles. These results demonstrate the effectiveness of Open-SAT in open-vocabulary satellite image retrieval, leveraging LLM guidance without the need for additional training or supervision.\n\n---\n\n## 143. Feature Starvation as Geometric Instability in Sparse Autoencoders\n- **Authors**: Faris Chaudhry , Keisuke Yano , Anthea Monod\n- **URL**: [https://arxiv.org/abs/2605.05341](https://arxiv.org/abs/2605.05341)\n- **Abstract**:\n> Sparse autoencoders (SAEs) are used to disentangle the dense, polysemantic internal representations of large language models (LLMs) into interpretable, monosemantic concepts. However, standard $\\ell_1$-regularized SAEs suffer from feature starvation (dead neurons) and shrinkage bias, often requiring computationally expensive heuristic resampling and nondifferentiable hard-masking methods to bypass these challenges. We argue that feature starvation is not merely an empirical artifact of poor data diversity, but a fundamental optimization-geometric pathology of overcomplete dictionaries: the $\\ell_1$-induced sparse coding map is unstable and fundamentally misaligned with shallow, amortized encoders. To address this structural instability, we introduce adaptive elastic net SAEs (AEN-SAEs), a fully differentiable architecture grounded in classical sparse regression. AEN-SAEs combine an $\\ell_2$ structural term that enforces strong convexity and Lipschitz stability with adaptive $\\ell_1$ reweighting that eliminates shrinkage bias and suppresses spurious features, thereby jointly controlling the curvature and interaction structure of the induced polyhedral geometry. Theoretically, we show that AEN-SAEs yield a Lipschitz-continuous sparse coding map and recover the global feature support under mild assumptions. Empirically, across synthetic settings and LLMs (Pythia 70M, Llama 3.1 8B), AEN-SAEs mitigate feature starvation without auxiliary heuristics while maintaining competitive reconstruction abilities.\n\n---\n\n## 144. How Far Are VLMs from Privacy Awareness in the Physical World? An Empirical Study\n- **Authors**: Junran Wang , Xinjie Shen , Zehao Jin , Pan Li\n- **URL**: [https://arxiv.org/abs/2605.05340](https://arxiv.org/abs/2605.05340)\n- **Abstract**:\n> As Vision-Language Models (VLMs) are increasingly deployed as autonomous cognitive cores for embodied assistants, evaluating their privacy awareness in physical environments becomes critical. Unlike digital chatbots, these agents operate in intimate spaces, such as homes and hospitals, where they possess the physical agency to observe and manipulate privacy-sensitive information and artifacts. However, current benchmarks remain limited to unimodal, text-based representations that cannot capture the demands of real-world settings. To bridge this gap, we present ImmersedPrivacy, an interactive audio-visual evaluation framework that simulates realistic physical environments using a Unity-based simulator. ImmersedPrivacy evaluates physically grounded privacy awareness across three progressive tiers that test a model's ability to identify sensitive items in cluttered scenes, adapt to shifting social contexts, and resolve conflicts between explicit commands and inferred privacy constraints. Our evaluation of 12 state-of-the-art models reveals consistent deficits. In cluttered scenes, all models exhibit monotonic performance decay as scene complexity grows due to perceptual deficit. When social context shifts, no model exceed 65% selection accuracy. Under conflicting commands, the best model gemini-3.1-pro perfectly balances task completion and privacy preservation in only 51% of cases. These findings reveal that current VLMs in the physical world suffer from perceptual fragility and fail to let their knowledge of privacy cues govern their situated behavior. Our code and data is available at this https URL .\n\n---\n\n## 145. Bridging Generation and Training: A Systematic Review of Quality Issues in LLMs for Code\n- **Authors**: Kaifeng He , Xiaojun Zhang , Peiliang Cai , Mingwei Liu , Yanlin Wang , Chong Wang , Kaifeng Huang , Bihuan Chen , Xin Peng , Zibin Zheng\n- **URL**: [https://arxiv.org/abs/2605.05267](https://arxiv.org/abs/2605.05267)\n- **Abstract**:\n> Large language models (LLMs) frequently generate defective outputs in code generation tasks, ranging from logical bugs to security vulnerabilities. While these generation failures are often treated as model-level limitations, empirical evidence increasingly traces their root causes to imperfections within the training corpora. Yet, the specific mechanisms linking training data quality issues to generated code quality issues remain largely unmapped. This paper presents a systematic literature review of 114 primary studies to investigate how training data quality issues propagate into code generation. We establish a unified taxonomy that categorizes generated code quality issues across nine dimensions and training data quality issues into code and non-code attributes. Based on this taxonomy, we formalize a causal framework detailing 18 typical propagation mapping mechanisms. Furthermore, we synthesize state-of-the-art detection and mitigation techniques across the data, model, and generation lifecycles. The reviewed literature reveals a clear methodological shift: quality assurance is transitioning from reactive, heuristic-based post-generation filtering toward proactive, data-centric governance and closed-loop repair. Finally, we identify open challenges and outline research directions for developing reliable LLMs for code through integrated data curation and continuous evaluation. Our repository is available at this https URL .\n\n---\n\n## 146. Decision-aware User Simulation Agent for Evaluating Conversational Recommender Systems\n- **Authors**: Yuan-Chi Li , Li-Chi Chen , Sung-Yi Wu , Yu-Che Tsai , Shou-De Lin\n- **URL**: [https://arxiv.org/abs/2605.05250](https://arxiv.org/abs/2605.05250)\n- **Abstract**:\n> Conversational recommender systems (CRS) increasingly rely on user simulators for automated evaluation of sales agents. A key requirement for such simulators is the ability to model human decision-making. However, most existing simulation frameworks do not explicitly model the internal decision process, and LLM-based simulators often exhibit unrealistically strong information-processing capabilities, rarely exhibit the hesitation or decision deferral commonly observed in real consumer behavior, resulting in overly high acceptance probabilities. To address this limitation, we propose Hesitator, a theory-grounded user simulation framework that explicitly models human decision-making under choice overload. The framework introduces a modular Decision Module that separates utility-based item selection from overload-aware commitment decisions. Experiments across multiple user simulation frameworks, domains, sales modes, and LLM backbones show that integrating our module consistently mitigates unrealistic behaviors under increasing overload conditions. Furthermore, Hesitator reproduces established behavioral patterns from psychological economics, demonstrating its ability to model human decision behavior.\n\n---\n\n## 147. Towards Dependable Retrieval-Augmented Generation Using Factual Confidence Prediction\n- **Authors**: Florian Geissler , Francesco Carella , Laura Fieback , Jakob Spiegelberg\n- **URL**: [https://arxiv.org/abs/2605.05244](https://arxiv.org/abs/2605.05244)\n- **Abstract**:\n> Incorporating specific knowledge into large language models via retrieval-augmented generation (RAG) is a widespread technique that fuels many of today's industry AI applications. A fundamental problem is to assess if the context retrieved by some similarity search provides indeed supporting facts, or instead misguides the generator with irrelevant information. It is critical to associate meaningful confidence measures about the factuality of the retrieval process with the generated answers. We present a new, two-staged approach to predict fact faithfulness of the output of retrieval-augmented generations. First, we employ conformal prediction to select only those retrieved chunks who have a high chance to come from the correct source. This approach in itself can improve answer quality by up to 6% in some of the studied datasets, however, the associated statistical guarantees do not hold generally, since the assumption of sample exchangeability depends on the retriever setup. We present diagnostic metrics to assess whether a setup is suitable. Second, we quantify confidence in the consistency of a generated final answer with a given retrieved context, using an attention-based factuality classifier. This approach can detect inconsistent answers with a chance of up to 77%. Our work helps to establish a novel type of certified RAG systems for a broad range of natural language industry applications.\n\n---\n\n## 148. Rethinking Data Curation in LLM Training: Online Reweighting Offers Better Generalization than Offline Methods\n- **Authors**: Wanru Zhao , Yihong Chen , Yuzhi Tang , Wentao Ma , Shengchao Hu , Shell Xu Hu , Alex Iacob , Abhinav Mehrotra , Nicholas D. Lane\n- **URL**: [https://arxiv.org/abs/2605.05227](https://arxiv.org/abs/2605.05227)\n- **Abstract**:\n> Data curation is a critical yet under-explored area in large language model (LLM) training. Existing methods, such as data selection and mixing, operate in an offline paradigm, detaching themselves from training. This separation introduces engineering overhead and makes the curation brittle: the entire pipeline must be re-run under model/task shifts. Moreover, offline methods alter data size through hard filtering or resampling, often sacrificing data diversity and harming generalization. We propose to rethink data curation as an online reweighting problem, where sample importance is dynamically adjusted during training via loss weighting rather than static pre-processing. Specifically, we introduce ADAPT (Adaptive Data reweighting for Pretraining and FineTuning), a dynamic online framework that reweights training samples with adaptive per-sample learning rates guided by similarity-based quality signals, without changing the number of training samples. Unlike offline methods that enforce a static data distribution, ADAPT acts as an implicit curriculum learner, progressively shifting focus from coarse-grained patterns to fine-grained semantic distinctions as the model evolves. Experiments on both instruction tuning and large-scale pretraining show that ADAPT consistently outperforms offline selection/mixing and prior online methods, achieving stronger cross-benchmark generalization under equal FLOPs.\n\n---\n\n## 149. MACS: Modality-Aware Capacity Scaling for Efficient Multimodal MoE Inference\n- **Authors**: Bo Li , Chuan Wu , shaolin Zhu\n- **URL**: [https://arxiv.org/abs/2605.05225](https://arxiv.org/abs/2605.05225)\n- **Abstract**:\n> Mixture-of-Experts Multimodal Large Language Models (MoE MLLMs) suffer from a significant efficiency bottleneck during Expert Parallelism (EP) inference due to the straggler effect. This issue is worsened in the multimodal context, as existing token-count-based load balancing methods fail to address two unique challenges: (1) Information Heterogeneity, where numerous redundant visual tokens are treated equally to semantically critical ones, and (2) Modality Dynamics, where varying visual to text ratios across tasks lead to resource misallocation. To address these challenges, we propose MACS (Modality-Aware Capacity Scaling), a training-free inference framework. Specifically, MACS introduces an Entropy-Weighted Load mechanism to quantify the semantic value of visual tokens, addressing information heterogeneity. Additionally, the Dynamic Modality-Adaptive Capacity mechanism allocates expert resources based on the real-time modal composition of the input. Extensive experiments demonstrate that MACS significantly outperforms existing methods on various multimodal benchmarks, providing a novel and robust solution for the efficient deployment of MoE MLLMs in EP inference.\n\n---\n\n## 150. MidSteer: Optimal Affine Framework for Steering Generative Models\n- **Authors**: Tatiana Gaintseva , Andrew Stepanov , Ziquan Liu , Martin Benning , Gregory Slabaugh , Jiankang Deng , Ismail Elezi\n- **URL**: [https://arxiv.org/abs/2605.05220](https://arxiv.org/abs/2605.05220)\n- **Abstract**:\n> Steering intermediate representations has emerged as a powerful strategy for controlling generative models, particularly in post-deployment alignment and safety settings. However, despite its empirical success, it currently lacks a comprehensive theoretical framework. In this paper, we bridge this gap by formalizing the theory of concept steering. First, we establish a link between steering and affine concept erasure, proving that the standard approach for removing unwanted behaviors is a special case of LEACE (a closed-form method for affine erasure). Next, we formulate a principled theoretical framework for concept switching, LEACE-Switch, and characterize the assumptions under which it provides an optimal affine solution. Building on this analysis, we then introduce MidSteer (Minimal Disturbance concept Steering), a more general affine framework for concept manipulation that relaxes these assumptions and enables directed, minimal-disturbance transformations. We demonstrate that MidSteer performs favorably across a range of tasks, modalities, and architectures, including vision diffusion models and large language models.\n\n---\n\n## 151. Sparse Prefix Caching for Hybrid and Recurrent LLM Serving\n- **Authors**: Mikhail Shirokikh , Sergey Nikolenko\n- **URL**: [https://arxiv.org/abs/2605.05219](https://arxiv.org/abs/2605.05219)\n- **Abstract**:\n> Prefix caching is a key latency optimization for autoregressive LLM serving, yet existing systems assume dense per-token key/value reuse. State-space models change the structure of the problem: a recurrent layer can resume from a single stored state rather than requiring the entire token history. This asymmetry opens a new design point between no reuse and dense caching: store exact recurrent states at a sparse set of checkpoint positions and, on a cache hit, resume from the deepest stored checkpoint and recompute the remaining suffix exactly. We formalize sparse prefix caching as checkpoint placement under a distribution over overlap depths, yielding an exact O(NM) dynamic program. For use cases where requests share a non-trivial prefix (e.g. asking different questions about a single long document), we show that our method consistently improves the Pareto frontier traced by standard heuristics on real-world data. Across QuALITY and System Prompts, distribution-aware placement dominates every fixed-budget baseline on the measured layer-group Pareto frontier and matches or outperforms the strongest heuristic (block caching) while typically using substantially fewer checkpoints, with the largest gains at low checkpoint budgets where the overlap distribution is most non-uniform. The method is most relevant when many requests share a substantial but not identical prefix within a retained cache entry. It preserves exact outputs, does not change the recurrent computation itself or require new recurrent update kernels, applies to recurrent/SSM layers whose hidden state can be extracted and restored exactly, and for hybrid models can be combined with existing KV-cache compression techniques.\n\n---\n\n## 152. A Review of Large Language Models for Stock Price Forecasting from a Hedge-Fund Perspective\n- **Authors**: Olivia Zhang , Zhilin Zhang\n- **URL**: [https://arxiv.org/abs/2605.05211](https://arxiv.org/abs/2605.05211)\n- **Abstract**:\n> Large language models (LLMs) are increasingly deployed in quantitative finance for stock price forecasting. This review synthesizes recent applications of LLMs in this domain, including extracting sentiment from financial news and social media, analyzing financial reports and earnings-call transcripts, tokenizing or symbolizing stock price series, and constructing multi-agent trading systems. Particular attention is paid to practical pitfalls that are often understated in the literature, such as fragility in sentiment analysis, dataset and horizon design, performance evaluation metrics, data leakage, illiquidity premia, and limits of stock price predictability. Organized from a hedge-fund perspective, the review is intended to guide both academic researchers and hedge fund managers in integrating LLMs into real-world trading pipelines and in stress-testing their robustness under realistic market frictions.\n\n---\n\n{% endraw %}\n"}, "44b478f8b4630653eb5dfddb902d13a4a53ed4bc": {"t": 1785770986.4411502, "n": 263185, "h": "c0b4ce7dc3328d6b", "c": "---\nlayout: default\ntitle: \"LLM 관련 주요 논문 - 2026-05-08\"\n---\n\n{% raw %}\n# LLM 관련 주요 논문 - 2026-05-08\n\n## 1. GlazyBench: A Benchmark for Ceramic Glaze Property Prediction and Image Generation\n- **Authors**: Ziyu Zhai , Siyou Li , Juexi Shao , Juntao Yu\n- **URL**: [https://arxiv.org/abs/2605.06641](https://arxiv.org/abs/2605.06641)\n- **Abstract**:\n> Developing ceramic glazes is a costly, time-consuming process of trial and error due to complex chemistry, placing a significant burden on independent artists. While recent advances in multimodal AI offer a modern solution, the field lacks the large-scale datasets required to train these models. We propose GlazyBench, the first dataset for AI-assisted glaze design. Comprising 23,148 real glaze formulations, GlazyBench supports two primary tasks: predicting post-firing surface properties, such as color and transparency, from raw materials, and generating accurate visual representations of the glaze based on these properties. We establish comprehensive baselines for property prediction using traditional machine learning and large language models, alongside image generation benchmarks using deep generative and large multimodal models. Our experiments demonstrate promising yet challenging results. GlazyBench pioneers a new research direction in AI-assisted material design, providing a standardized benchmark for systematic evaluation.\n\n---\n\n## 2. Can RL Teach Long-Horizon Reasoning to LLMs? Expressiveness Is Key\n- **Authors**: Tianle Wang , Zhaoyang Wang , Guangchen Lan , Xinpeng Wei , Sipeng Zhang , Guanwen Qiu , Abulhair Saparov\n- **URL**: [https://arxiv.org/abs/2605.06638](https://arxiv.org/abs/2605.06638)\n- **Abstract**:\n> Reinforcement learning (RL) has been applied to improve large language model (LLM) reasoning, yet the systematic study of how training scales with task difficulty has been hampered by the lack of controlled, scalable environments. We introduce ScaleLogic, a synthetic logical reasoning framework that offers independent control over two axes of difficulty: the depth of the required proof planning (i.e., the horizon) and the expressiveness of the underlying logic. Our proposed framework supports a wide range of logics: from simple implication-only logic (\"if-then\") towards more expressive first-order reasoning with conjunction (\"and\"), disjunction (\"or\"), negation (\"not\"), and universal quantification (\"for all\"). Using this framework, we show that the RL training compute $T$ follows a power law with respect to reasoning depth $D$ ($T \\propto D^{\\gamma}$, $R^{2} > 0.99$), and that the scaling exponent $\\gamma$ increases monotonically with logical expressiveness, from $1.04$ to $2.60$. On downstream mathematics and general reasoning benchmarks, more expressive training settings yield both larger performance gains (up to $+10.66$ points) and more compute-efficient transfer compared to less expressive settings, demonstrating that what a model is trained on, not just how much it is trained, shapes downstream transfer. We further show that the power-law relationship holds across multiple RL methods, and curriculum-based training substantially improves scaling efficiency.\n\n---\n\n## 3. MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems\n- **Authors**: Zhexuan Wang , Xuebo Liu , Li Wang , Zifei Shan , Yutong Wang , Zhenxi Song , Min Zhang\n- **URL**: [https://arxiv.org/abs/2605.06623](https://arxiv.org/abs/2605.06623)\n- **Abstract**:\n> Large language model (LLM)-based Multi-agent systems (MAS) have shown promise in tackling complex collaborative tasks, where agents are typically orchestrated via role-specific prompts. While the quality of these prompts is pivotal, jointly optimizing them across interacting agents remains a non-trivial challenge, primarily due to the misalignment between local agent objectives and holistic system goals. To address this, we introduce MASPO, a novel framework designed to automatically and iteratively refine prompts across the entire system. A core innovation of MASPO is its joint evaluation mechanism, which assesses prompts not merely by their local validity, but by their capacity to facilitate downstream success for successor agents. This effectively bridges the gap between local interactions and global outcomes without relying on ground-truth labels. Furthermore, MASPO employs a data-driven evolutionary beam search to efficiently navigate the high-dimensional prompt space. Extensive empirical evaluations across 6 diverse tasks demonstrate that MASPO consistently outperforms state-of-the-art prompt optimization methods, achieving an average accuracy improvement of 2.9. We release our code at this https URL .\n\n---\n\n## 4. SkillOS: Learning Skill Curation for Self-Evolving Agents\n- **Authors**: Siru Ouyang , Jun Yan , Yanfei Chen , Rujun Han , Zifeng Wang , Bhavana Dalvi Mishra , Rui Meng , Chun-Liang Li , Yizhu Jiao , Kaiwen Zha , Maohao Shen , Vishy Tirumalashetty , George Lee , Jiawei Han , Tomas Pfister , Chen-Yu Lee\n- **URL**: [https://arxiv.org/abs/2605.06614](https://arxiv.org/abs/2605.06614)\n- **Abstract**:\n> LLM-based agents are increasingly deployed to handle streaming tasks, yet they often remain one-off problem solvers that fail to learn from past interactions. Reusable skills distilled from experience provide a natural substrate for self-evolution, where high-quality skill curation serves as the key bottleneck. Existing approaches either rely on manual skill curation, prescribe heuristic skill operations, or train for short-horizon skill operations. However, they still struggle to learn complex long-term curation policies from indirect and delayed feedback. To tackle this challenge, we propose SkillOS, an experience-driven RL training recipe for learning skill curation in self-evolving agents. SkillOS pairs a frozen agent executor that retrieves and applies skills with a trainable skill curator that updates an external SkillRepo from accumulated experience. To provide learning signals for curation, we design composite rewards and train on grouped task streams based on skill-relevant task dependencies, where earlier trajectories update the SkillRepo, and later related tasks evaluate these updates. Across multi-turn agentic tasks and single-turn reasoning tasks, SkillOS consistently outperforms memory-free and strong memory-based baselines in both effectiveness and efficiency, with the learned skill curator generalizing across different executor backbones and task domains. Further analyses show that the learned curator produces more targeted skill use, while the skills in SkillRepo evolve into more richly structured Markdown files that encode higher-level meta-skills over time.\n\n---\n\n## 5. NeuroAgent: LLM Agents for Multimodal Neuroimaging Analysis and Research\n- **Authors**: Lujia Zhong , Yihao Xia , Jianwei Zhang , Shuo huang , Jiaxin Yue , Mingyang Xia , Yonggang Shi\n- **URL**: [https://arxiv.org/abs/2605.06584](https://arxiv.org/abs/2605.06584)\n- **Abstract**:\n> Multimodal neuroimaging analysis often involves complex, modality-specific preprocessing workflows that require careful configuration, quality control, and coordination across heterogeneous toolchains. Beyond preprocessing, downstream statistical analysis and disease classification commonly require task-specific code, evaluation protocols, and data-format conventions, creating additional barriers between raw acquisitions and reproducible scientific analysis. We present NeuroAgent, an LLM-driven agentic framework that automates key preprocessing and analysis steps for heterogeneous neuroimaging data, including sMRI, fMRI, dMRI, and PET, and supports interactive downstream analysis through natural-language queries. NeuroAgent employs a hierarchical multi-agent architecture with a feedback-driven Generate-Execute-Validate engine: agents autonomously generate executable preprocessing code, detect and recover from runtime errors, and validate output integrity. We evaluate the system on 1,470 subjects pooled across all ADNI phases (CN=1,000, AD=470), where all subjects have sMRI and tabular data, with subsets also having Tau-PET (n=469), fMRI (n=278), and DTI ($n=620$). Pipeline ablation studies across multiple LLM backends show that capable models reach up to 100% intent-parsing accuracy, with the strongest backend (Qwen3.5-27B) reaching 84.8% end-to-end preprocessing step correctness. Automated recovery limits manual intervention to edge cases where human review is required via the Human-In-The-Loop interface. For Alzheimer's Disease classification using automatically preprocessed multimodal data, our agent ensemble achieves an AUC of 0.9518 with four modalities, outperforming all single-modality baselines. These results show that NeuroAgent can reduce the manual effort required for neuroimaging preprocessing and enable end-to-end automated analysis pipelines for neuroimaging research.\n\n---\n\n## 6. Process Matters more than Output for Distinguishing Humans from Machines\n- **Authors**: Milena Rmus , Mathew D. Hardy , Thomas L. Griffiths , Mayank Agrawal\n- **URL**: [https://arxiv.org/abs/2605.06524](https://arxiv.org/abs/2605.06524)\n- **Abstract**:\n> Reliable human-machine discrimination is becoming increasingly important as large language models and autonomous agents are deployed in online settings. Existing approaches evaluate whether a system can produce behavior or responses indistinguishable from those of a human, following the emphasis on outputs as a criterion for intelligence proposed by Alan Turing. Cognitive science offers an alternative perspective: evaluating the process by which behavior is produced. To test whether cognitive processes can reliably distinguish humans from machines, we introduce CogCAPTCHA30, a battery of 30 cognitive tasks designed to elicit diagnostic process-level features even when task performance is matched. Across the battery, process-level features provide stronger discriminative signal than performance metrics alone, reliably distinguishing humans from agents even under output matching (mean process-feature classifier AUC = 0.88). To evaluate agentic process differences, we compare off-the-shelf frontier agents (Claude Sonnet 4.5, GPT-5, Gemini 2.5 Pro), Centaur (a language model fine-tuned on 10.7M human decisions), and two task-specific fine-tuning approaches applied to Qwen2.5-1.5B-Instruct: action-level supervised fine-tuning (A-SFT) and process-level fine-tuning (P-SFT), which directly optimizes process features. Broad fine-tuning on human decisions improves human-like task processes relative to off-the-shelf agents, while task-specific process-level supervision further improves behavioral mimicry. However, this advantage diminishes under cross-task transfer when supervised process targets do not naturally generalize across tasks. Explicit process-level supervision can improve human behavioral mimicry, but only if appropriate task-specific process representations are available, highlighting process specification as a bottleneck for achieving human-like cognitive processes in machines.\n\n---\n\n## 7. Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors\n- **Authors**: Jonas Wiedermann-Möller , Leonard Dung , Maksym Andriushchenko\n- **URL**: [https://arxiv.org/abs/2605.06490](https://arxiv.org/abs/2605.06490)\n- **Abstract**:\n> AI systems have become increasingly capable of dangerous behaviours in many domains. This raises the question: Do models sometimes choose to violate human instructions in order to perform behaviour that is more useful for certain goals? We introduce a benchmark for measuring model propensity for instrumental convergence (IC) behaviour in terminal-based agents. This is behaviour such as self-preservation that has been hypothesised to play a key role in risks from highly capable AI agents. Our benchmark is realistic and low-stakes which serves to reduce evaluation-awareness and roleplay confounds. The suite contains seven operational tasks, each with an official workflow and a policy-violating shortcut. An eight-variant shared framework varies monitoring, instruction clarity, stakes, permission, instrumental usefulness and blocked honest paths to support inferences regarding the factors driving IC behaviour. We evaluated ten models using deterministic environment-state scorers over 1,680 samples, with trace review employed for audit and adjudication purposes. The final IC rate is 86 out of 1,680 samples (5.1%). IC behaviour is concentrated rather than uniform: two Gemini models account for 66.3% of IC cases and three tasks account for 84.9%. Conditions in which IC behaviour is indispensable for task success result in the greatest increase in the adjusted IC rate (+15.7 percentage points), whereas emphasising that task success is critical or certain framing choices do not produce comparable effects. Our findings indicate that realistic, low-nudge environments elicit IC behaviour rarely but systematically in most tested models. We conclude that it is feasible to robustly measure tendencies for dangerous behaviour in current frontier AI agents.\n\n---\n\n## 8. ReasonSTL: Bridging Natural Language and Signal Temporal Logic via Tool-Augmented Process-Rewarded Learning\n- **Authors**: Bowen Ye , Zhijian Li , Junyue Huang , Junkai Ma , Xiang Yin\n- **URL**: [https://arxiv.org/abs/2605.06483](https://arxiv.org/abs/2605.06483)\n- **Abstract**:\n> Signal Temporal Logic (STL) is an expressive formal language for specifying spatio-temporal requirements over real-valued, real-time signals. It has been widely used for the verification and synthesis of autonomous systems and cyber-physical systems. In practice, however, users often express their requirements in natural language rather than in structured STL formulas, making natural-language-to-STL translation a critical yet challenging task. Manual specification requires temporal-logic expertise and cannot scale, while prompting commercial LLM APIs incurs substantial token costs and may expose sensitive system requirements to third-party services, raising privacy concerns for industrial deployment. To address these challenges, we present \\textsc{ReasonSTL}, a tool-augmented framework that adapts local open-source language models for natural-language-to-STL generation. \\textsc{ReasonSTL} decomposes the translation process into explicit reasoning, deterministic tool calls, and structured formula construction. We further introduce process-rewarded training to supervise both tool-use trajectories and final formulas, together with \\textsc{STL-Bench}, a bilingual, computation-aware benchmark grounded in real-world signals. Experiments show that a 4B model trained with \\textsc{ReasonSTL} achieves state-of-the-art performance in both automatic metrics and human evaluations, demonstrating that \\textsc{ReasonSTL} provides a transparent, low-cost, and privacy-preserving alternative for formal specification drafting.\n\n---\n\n## 9. Patch-Effect Graph Kernels for LLM Interpretability\n- **Authors**: Ruben Fernandez-Boullon , David N. Olivieri\n- **URL**: [https://arxiv.org/abs/2605.06480](https://arxiv.org/abs/2605.06480)\n- **Abstract**:\n> Mechanistic interpretability aims to reverse-engineer transformer computations by identifying causal circuits through activation patching. However, scaling these interventions across diverse prompts and task families produces high-dimensional, unstructured datasets that are difficult to compare systematically. We propose a framework that reframes mechanistic analysis as a graph machine-learning problem by representing activation-patching profiles as patch-effect graphs over model components. We introduce three graph-construction methods: direct-influence via causal mediation, partial-correlation, and co-influence and apply graph kernels to analyze the resulting structures. Evaluating this approach on GPT-2 Small using Indirect Object Identification (IOI) and related tasks, we find that patch-effect graphs preserve discriminative structural signals. Specifically, localized edge-slot features provide higher classification accuracy than global graph-shape descriptors. A screened paired-patching validation suggests that CI and PC selected candidate edges correspond to stronger activation-influence effects than random or low-rank candidates. Crucially, by evaluating these representations against rigorous prompt-only and raw patch-effect controls, we make the evidential scope of the benchmark explicit: graph features compress structured patching signal, while raw tensors and surface cues define strong baselines that any circuit-level claim should address. Ultimately, our framework provides a compression and evaluation pipeline for comparing patching-derived structures under controlled baselines, separating robust slice-discriminative evidence from stronger task-general causal-circuit claims.\n\n---\n\n## 10. Beyond Task Success: Measuring Workflow Fidelity in LLM-Based Agentic Payment Systems\n- **Authors**: Donghao Huang , Joon Kiat Chua , Zhaoxia Wang\n- **URL**: [https://arxiv.org/abs/2605.06457](https://arxiv.org/abs/2605.06457)\n- **Abstract**:\n> LLM-based multi-agent systems are increasingly deployed for payment workflows, yet prevailing metrics, Task Success Rate (TSR) and Agent Handoff F1-Score (HF1), capture only final outcomes or unordered routing decisions. We introduce the Agentic Success Rate (ASR), a trajectory-fidelity metric that compares observed and expected agent execution sequences at the transition level, decomposing performance into Transition Recall and Transition Precision. Applied to the Hierarchical Multi-Agent System for Payments (HMASP) across 18 LLMs and 90,000 task instances, ASR reveals that 10 of 18 models systematically skip a confirmation checkpoint during payment checkout, a deviation invisible to both TSR and HF1, while 8 models enforce the checkpoint perfectly. Notably, GPT-4.1 exhibits hidden workflow shortcuts despite achieving perfect TSR and HF1, while GPT-5.2 achieves perfect ASR. Prompt refinements and deterministic routing guards guided by ASR diagnostics yield substantial TSR improvements, with gains up to +93.8 percentage points for previously struggling models, demonstrating that trajectory-level evaluation is essential in regulated domains.\n\n---\n\n## 11. PrefixGuard: From LLM-Agent Traces to Online Failure-Warning Monitors\n- **Authors**: Xinmiao Huang , Jinwei Hu , Rajarshi Roy , Changshun Wu , Yi Dong , Xiaowei Huang\n- **URL**: [https://arxiv.org/abs/2605.06455](https://arxiv.org/abs/2605.06455)\n- **Abstract**:\n> Large language model (LLM) agents now execute long, tool-using tasks where final outcome checks can arrive too late for intervention. Online warning requires lightweight prefix monitors over heterogeneous traces, but hand-authored event schemas are brittle and deployment-time LLM judging is costly. We introduce PrefixGuard, a trace-to-monitor framework with an offline StepView induction step followed by supervised monitor training. StepView induces deterministic typed-step adapters from raw trace samples, and the monitor learns an event abstraction and prefix-risk scorer from terminal outcomes. Across WebArena, $\\tau^2$-Bench, SkillsBench, and TerminalBench, the strongest PrefixGuard monitors reach 0.900/0.710/0.533/0.557 AUPRC. Using the strongest backend within each representation, they improve over raw-text controls by an average of +0.137 AUPRC. LLM judges remain substantially weaker under the same prefix-warning protocol. We also derive an observability ceiling on score-based area under the precision-recall curve (AUPRC) that separates monitor error from failures lacking evidence in the observed prefix. For finite-state audit, post-hoc deterministic finite automaton (DFA) extraction remains compact on WebArena and $\\tau^2$-Bench (29 and 20 states) but expands to 151 and 187 states on SkillsBench and TerminalBench. Finally, first-alert diagnostics show that strong ranking does not imply deployment utility: WebArena ranks well yet fails to support low-false-alarm alerts, whereas $\\tau^2$-Bench and TerminalBench retain more actionable early alerts. Together, these results position PrefixGuard as a practical monitor-synthesis recipe with explicit diagnostics for when prefix warnings translate into actionable interventions.\n\n---\n\n## 12. SCRuB: Social Concept Reasoning under Rubric-Based Evaluation\n- **Authors**: Jamelle Watson-Daniels , Himaghna Bhattacharjee , Skyler Wang , Brandon Handoko , Antonio Li , Anaelia Ovalle , Mahesh Pasupuleti , Candace Ross , Vidya Sarma , Arjun Subramonian , Karen Ullrich , Will van der Vaart , Yijing Xin , Maximilian Nickel\n- **URL**: [https://arxiv.org/abs/2605.06444](https://arxiv.org/abs/2605.06444)\n- **Abstract**:\n> While many studies of Large Language Model (LLM) reasoning capabilities emphasize mathematical or technical tasks, few address reasoning about social concepts: the abstract ideas shaping social norms, culture, and institutions. This understudied capability is essential for modern models acting as social agents, yet no systematic evaluation methodology targets it. We introduce SCRuB (Social Concept Reasoning under Rubric-Based Evaluation), a framework designed for this setting of task indeterminacy. Our goal is to measure the degree to which a model reasons about social concepts with the depth and critical rigor of a human expert. SCRuB proceeds in three phases: prompt construction from established sources, response generation by experts and models, and comparative evaluation using a five-dimensional critical thinking rubric. To enable generalization of the pipeline, we introduce a Panel of Disciplinary Perspectives ensemble validated against independent expert judges. We release SCRuBEval (n=4,711 evaluation prompts) and SCRuBAnnotations (300 expert-authored responses and 150 expert comparative judgments from 45 PhD-level scholars). Our results show that frontier models consistently outperform human experts across all five rubric dimensions. Across 1,170 pairwise comparisons, expert judges ranked a model response first in 80.8% of judgments and preferred model responses overall 74.4% of the time. Ultimately, this study provides the first expert-grounded demonstration of evaluation saturation for social concept reasoning: the single-turn exam-style format has reached its ceiling for models and humans alike.\n\n---\n\n## 13. Knowledge Graphs, the Missing Link in Agentic AI-based Formal Verification\n- **Authors**: Vaisakh Naduvodi Viswambharan , Keerthan Kopparam Radhakrishna , Deepak Narayan Gadde , Aman Kumar\n- **URL**: [https://arxiv.org/abs/2605.06434](https://arxiv.org/abs/2605.06434)\n- **Abstract**:\n> Recent advances in Large Language Models (LLMs) have enabled workflows that generate SystemVerilog Assertions (SVAs) from natural-language specifications, with the potential to accelerate Formal Verification (FV). However, high-quality assertion synthesis remains challenging because specifications are often ambiguous or incomplete and critical micro-architectural details reside in the Register Transfer Level (RTL). Many existing approaches treat the specification and RTL as loosely structured text, which weakens specification-to-RTL grounding and leads to semantic mismatches and frequent syntax failures during formal parsing and elaboration. This work addresses these limitations with a verification-centric Knowledge Graph (KG) constructed from structured Intermediate Representations (IRs) extracted from the specification, RTL, and formal-tool feedback, including syntax diagnostics, Counterexamples (CEXs), and coverage reports. The KG links requirements, design hierarchy, signals, assumptions, and properties to provide traceable, design-grounded context for generation. A multi-agent workflow queries and updates this KG to generate SVAs and to drive three refinement loops: syntax repair guided by tool diagnostics, CEX-guided correction using trace links, and coverage-directed property augmentation. Evaluation across seven benchmark designs indicates that KG-based context retrieval improves specification-to-RTL grounding and consistently produces compilable SVAs with low syntax-repair overhead. The approach achieves formal coverage ranging from 78.5% to 99.4%, though convergence exhibits design dependence with complex temporal and arithmetic reasoning remaining challenging for current LLM capabilities.\n\n---\n\n## 14. Rethinking Vacuity for OOD Detection in Evidential Deep Learning\n- **Authors**: Claire McNamara\n- **URL**: [https://arxiv.org/abs/2605.06382](https://arxiv.org/abs/2605.06382)\n- **Abstract**:\n> Vacuity, or Uncertainty Mass (UM), is commonly used as a metric to evaluate Out-of-Distribution (OOD) detection in Evidential Deep Learning (EDL). It generally involves dividing the number of classes ($K$) by the total strength of belief ($S$) of the model's predictions, where $S$ is derived from summing the Dirichlet parameters. As such, UM is sensitive to the cardinality of $K$. In particular, it is unlikely in practice that there is a linear relationship between $K$ and $S$ as $K$ and $S$ increase due to the nature of EDL (suppressing incorrectly assigned evidence). As a result, when comparing In Distribution (ID) and OOD results, it is important that $K_{\\mathrm{ID} }$ and $K_{\\mathrm{OOD} }$ are equal; something that is not always ensured in practice. We provide an empirical demonstration of how results for AUROC and AUPR can substantially differ when class cardinality between ID and OOD differs by 1, with AUROC differing by as much as 0.318 and AUPR by 0.613 for standard EDL, and AUROC by 0.360 and AUPR by 0.683 for IB-EDL. More concretely, our findings isolate an evaluation artefact: when K differs between ID and OOD, AUROC/AUPR can be artificially inflated without any change in model predictions. We further discuss the evaluation of EDL over causal language models using Multiple-Choice Question-Answer (MCQA) datasets and argue for clearer definitions of ID and OOD in this context. Our primary contribution is an empirical and theoretical demonstration that vacuity-based OOD detection in EDL-fine-tuned LLMs is highly sensitive to uncontrolled differences in evaluated class cardinality.\n\n---\n\n## 15. From Agent Loops to Deterministic Graphs: Execution Lineage for Reproducible AI-Native Work\n- **Authors**: Josh Rosen , Seth Rosen\n- **URL**: [https://arxiv.org/abs/2605.06365](https://arxiv.org/abs/2605.06365)\n- **Abstract**:\n> Large language model systems are increasingly deployed as agentic workflows that interleave reasoning, tool use, memory, and iterative refinement. These systems are effective at producing answers, but they often rely on implicit conversational state, making it difficult to preserve stable work products, isolate irrelevant updates, or propagate changes through intermediate artifacts. We introduce execution lineage: an execution model in which AI-native work is represented as a directed acyclic graph (DAG) of artifact-producing computations with explicit dependencies, stable intermediate boundaries, and identity-based replay. The goal is not to make the model a better one-shot writer, but to make evolving AI-generated work maintainable under change. We compare execution-lineage replay against loop-centric update baselines on two controlled policy-memo update tasks. In an unrelated-branch update, DAG replay preserved the final memo exactly in all runs, with zero churn and zero unrelated-branch contamination, while loop baselines regenerated the memo and frequently imported unrelated context. In an intermediate-artifact edit, all systems reflected the new constraint in the final memo, but only DAG replay achieved perfect upstream preservation, downstream propagation, unaffected-artifact preservation, and cross-artifact consistency. These results show that final answer quality and maintained-state quality are distinct. Strong loop baselines can remain competitive at producing polished final outputs when the task is a bounded synthesis/update problem and all current sources fit in context, but immediate task success can mask partial state inconsistency that may compound over future revisions. Execution lineage provides stronger guarantees about what should change, what should remain stable, and how work evolves across revisions.\n\n---\n\n## 16. A Regime Theory of Controller Class Selection for LLM Action Decisions\n- **Authors**: Zhaoyang Jiang , Zhizhong Fu , Yunsoo Kim , Jiacong Mi , Zicheng Li , Xuanqi Peng , Honghan Wu\n- **URL**: [https://arxiv.org/abs/2605.06339](https://arxiv.org/abs/2605.06339)\n- **Abstract**:\n> Deployed language and vision-language models must decide, on each input, whether to answer directly, retrieve evidence, defer to a stronger model, or abstain. Contrary to the common monotonicity intuition, greater per-input expressivity is not uniformly beneficial in finite samples: under identical strict cross-validation, different benchmarks prefer different controller classes. This reflects a finite-sample limitation of instance-level uncertainty signals, which can be exhausted at a distribution-dependent scale. We organize controllers into a nested lattice of four classes: fixed actions, partition routers, instance-level controllers, and prior-gated controllers, ordered by complexity. We prove a regime theory that turns three data-estimable bottlenecks into a class choice: how much improvement is possible beyond the best fixed action, whether there are enough samples for instance-level controllers to make reliable decisions, and how much improvement a coarse partition router can recover when instance-level signal is unreliable. The resulting Bernstein-tight threshold has a matching information-theoretic lower bound, and strict nested cross-validation provably selects a near-best class. Across SMS-Spam, HallusionBench, A-OKVQA, and FOLIO, the predicted class matches the empirical winner; the prior-gated controller wins on TextVQA when OCR tokens supply a label-free prediction-time prior. Code is available at this https URL .\n\n---\n\n## 17. Addressing Labelled Data Scarcity: Taxonomy-Agnostic Annotation of PII Values in HTTP Traffic using LLMs\n- **Authors**: Thomas Cory , Axel Küpper\n- **URL**: [https://arxiv.org/abs/2605.06305](https://arxiv.org/abs/2605.06305)\n- **Abstract**:\n> Automated privacy audits of web and mobile applications often analyse outbound HTTP traffic to detect Personally Identifiable Information (PII) leakage. However, existing learning-based detectors typically depend on scarce, manually labelled traffic and are tightly coupled to fixed label taxonomies, limiting transferability across domains and evolving definitions of PII. This paper investigates whether Large Language Models (LLMs) can support taxonomy-agnostic annotation of explicitly transmitted PII values in HTTP message bodies when the taxonomy is provided at runtime. We introduce a multi-stage LLM-based pipeline that combines deterministic pre-processing with label-level classification, targeted instance-level value annotation, and output validation. To enable controlled evaluation and exemplar-based prompting without relying on sensitive real-user captures, we further propose an LLM-based generator for synthetic HTTP traffic with manually validated, taxonomy-derived PII annotations. We evaluate the approach across three taxonomies spanning different PII domains and granularity levels. Results show that the pipeline accurately detects PII types and extracts corresponding values for concrete PII taxonomies. Overall, our findings position LLMs as a promising foundation for flexible, taxonomy-agnostic traffic annotation and for creating labelled data under evolving privacy taxonomies.\n\n---\n\n## 18. Data Language Models: A New Foundation Model Class for Tabular Data\n- **Authors**: Eda Erol , Giuliano Pezzoli , Ozer Cem Kelahmet\n- **URL**: [https://arxiv.org/abs/2605.06290](https://arxiv.org/abs/2605.06290)\n- **Abstract**:\n> Every major data modality now has a foundation model that understands it natively: text has language models, images have vision models, audio has audio models. Tabular data, the modality on which many consequential real-world AI decisions are made, does not. Every approach to tabular AI today, from gradient-boosted trees to the latest tabular foundation models, requires a preprocessing pipeline before any model can consume the data. None of them understand tabular data as a modality. We introduce the Data Language Model (DLM), the missing foundation model for tabular data. A DLM understands tables the way a language model understands sentences: natively, without serialization or preprocessing, directly from raw cell values. It is the tabular data layer on which AI models, agents, and vertical AI applications can be built, eliminating the preprocessing pipelines that currently stand between raw data and every AI system that consumes it. We present Schema-1, the first DLM: a 140M parameter model trained on more than 2.3M synthetic and real-world tabular datasets. Schema-1 outperforms gradient-boosted ensembles, AutoML stacks, and the tabular foundation models we evaluate on established row-level prediction benchmarks. On missing value reconstruction it achieves lower reconstruction error than all classical statistical methods and frontier large language models on mean performance across conditions, establishing that structural understanding of a dataset's own distributional geometry is more useful for imputation than world knowledge encoded in language. It identifies the industry sector of any unseen dataset from raw cell values alone, reliably across any domain, a task no prior tabular model can perform. It is the native tabular understanding layer that has been missing from the AI stack.\n\n---\n\n## 19. Joint Consistency: A Unified Test-Time Aggregation Framework via Energy Minimization\n- **Authors**: Yunzhen Yao , Hongye Wang , Yahong Wang , Michael C. Gastpar , Bo Jiang , Lie He\n- **URL**: [https://arxiv.org/abs/2605.06219](https://arxiv.org/abs/2605.06219)\n- **Abstract**:\n> This paper studies test-time aggregation, an approach that generates multiple reasoning traces and aggregates them into a final answer. Most existing methods rely on evaluation signals collected from candidate traces in isolation or answer frequencies, while ignoring comparative interactions among candidates. We propose Joint Consistency (JC), formulated as a constrained Ising-type energy minimization problem, where independent evaluation signals act as external fields and pairwise comparisons act as interactions. JC provides a unified framework for test-time aggregation that subsumes existing voting and weighted aggregation methods as special cases. Our construction of the interaction matrix leverages LLM-as-a-judge comparisons, and admits a theoretical interpretation under answer-level homogeneity assumptions. Moreover, we develop an efficient approximation strategy that makes interaction modeling practical for large-scale test-time aggregation. Experiments on math and code reasoning benchmarks show that JC consistently outperforms existing baselines across tasks, judge models, trace budgets, and trace-generation settings.\n\n---\n\n## 20. Beyond Fixed Benchmarks and Worst-Case Attacks: Dynamic Boundary Evaluation for Language Models\n- **Authors**: Haoxiang Wang , Da Yu , Huishuai Zhang\n- **URL**: [https://arxiv.org/abs/2605.06213](https://arxiv.org/abs/2605.06213)\n- **Abstract**:\n> Evaluating large language models (LLMs) today rests on fixed benchmarks that apply the same set of items to any model, producing ceiling and floor effects that mask capability gaps. We argue that the most informative evaluation signal lies at the boundary, where the per-prompt pass probability is near $0.5$ under random-sampling decoding, and propose Dynamic Boundary Evaluation (DBE), which actively locates each model's boundary and places it on a globally comparable difficulty scale. DBE delivers three artifacts: (i) a calibrated item bank covering safety, capability, and truthfulness, with per-item difficulty labels validated across $9$ reference LLMs; (ii) Skill-Guided Boundary Search (SGBS), a search algorithm that finds boundary items for a given target LLM using only API-level query access; and (iii) an evaluation protocol that places a new LLM on a unified ability scale and grows the evaluation set adaptively when the target falls outside the bank's coverage. We instantiate DBE on four categories spanning safety (harmful request refusal and over-refusal), capability (constrained instruction following), and truthfulness (multi-turn sycophancy resistance). The resulting evaluation covers a broader model spectrum without saturation while remaining compatible with existing datasets.\n\n---\n\n## 21. Towards Annotation-Free Validation of MLLMs: A Vision-Language Logical Consistency Metric\n- **Authors**: Ying Gu , Mei Chee Leong , Hui Li Tan , Shangbo Mao , Liyuan Li , Nancy Chen\n- **URL**: [https://arxiv.org/abs/2605.06201](https://arxiv.org/abs/2605.06201)\n- **Abstract**:\n> Dominant accuracy evaluation might reward unwarranted guessing of Large Language Models, and it might not be applicable to novel tasks for model validation without ground-truth (gt) annotation. Based on basic logic principle, we propose a novel framework to evaluate the vision-language logical consistency of MLLMs on both sufficient and necessary cause-effect relations. We define Vision-Language Logical Consistency Metric (VL-LCM) on traditional MC-VQA tests, and recent NaturalBench tests without the need for gt annotation. Through systematic experiments on representative VL benchmark MMMU and recent VL challenges like NaturalBench, we evaluated 11 recent open-source MLLMs from 4 frontier families. Our findings reveal that, despite significant progress of recent MLLMs on accuracy, logical consistency lags behind significantly. Extensive evaluations on the correlations of VL-LCM with metrics on gt, the reliability of LCM, and the relation of VL-LCM with response distribution justify the validity and applicability of VL-LCM even without gt annotation. Our findings suggest that, beyond accuracy, logical consistency could be employed for both accuracy and reliability. VL-LCM can also be employed for MLLM selection, validation, and reliable answer justification in novel tasks without gt annotation.\n\n---\n\n## 22. The Granularity Axis: A Micro-to-Macro Latent Direction for Social Roles in Language Models\n- **Authors**: Chonghan Qin , Xiachong Feng , Ziyun Song , Xiaocheng Feng , Jing Xiong , Lingpeng Kong\n- **URL**: [https://arxiv.org/abs/2605.06196](https://arxiv.org/abs/2605.06196)\n- **Abstract**:\n> Large language models (LLMs) are routinely prompted to take on social roles ranging from individuals to institutions, yet it remains unclear whether their internal representations encode the granularity of such roles, from micro-level individual experience to macro-level organizational, institutional, or national reasoning. We show that they do. We define a contrast-based Granularity Axis as the difference between mean macro- and micro-role hidden states. In Qwen3-8B, this axis aligns with the principal axis (PC1) of the role representation space at cosine 0.972 and accounts for 52.6% of its variance, indicating that granularity is the dominant geometric axis organizing prompted social roles. We construct 75 social roles across five granularity levels and collect 91,200 role-conditioned responses over shared questions and prompt variants, then extract role-level hidden states and project them onto the axis. Role projections increase monotonically across all five levels, remain stable across layers, prompt variants, endpoint definitions, held-out splits, and score-filtered subsets, and transfer to Llama-3.1-8B-Instruct. The axis is also causally relevant: activation steering along it shifts response granularity in the predicted direction, with Llama moving from 2.00 to 3.17 on a five-point macro scale under positive steering on prompts that admit local responses. The two models differ in controllability, suggesting that steering depends on each model's default operating regime. Overall, our findings suggest that social role granularity is not merely a stylistic surface feature, but a structured, ordered, and causally manipulable latent direction in role-conditioned language model behavior.\n\n---\n\n## 23. Systematic Evaluation of Large Language Models for Post-Discharge Clinical Action Extraction\n- **Authors**: Shivali Dalmia , Ananya Mantravadi , Prasanna Desikan\n- **URL**: [https://arxiv.org/abs/2605.06191](https://arxiv.org/abs/2605.06191)\n- **Abstract**:\n> The work in this paper evaluates zero-shot and few-shot large language models (LLMs) for safety-critical clinical action extraction using the CLIP discharge-note dataset, with particular emphasis on transitions of care and post-discharge patient safety. To manage the complexity of clinical documentation, we introduce a two-stage extraction framework that decomposes discharge notes, that are written in narrative form, into fine-grained, explicitly actionable clinical tasks through a staged prompting strategy. Our contributions include a systematic assessment of generative LLMs for clinical action extraction, a detailed comparison between general-purpose LLMs and task-specific supervised BERT-based models, and an analysis of annotation inconsistencies across different action categories. We show that contemporary LLMs achieve performance comparable to or exceeding supervised models on binary actionability detection, while supervised baselines retain a meaningful advantage on fine-grained multi-label category classification, despite the absence of task-specific fine-tuning and under strict data-privacy constraints. Qualitative error analysis reveals that many failures stem from misalignment between model reasoning and dataset annotation conventions, particularly in cases involving implicit clinical actions and rigid structural labeling rules. These results indicate that reported performance reflects model limitations due to lack of clinical reasoning, that is not captured by plain annotations. Labels without rationales make it impossible to distinguish clinical reasoning failures from annotation convention mismatches. Advancing clinical NLP requires reasoning-annotated datasets that document why specific spans are actionable, not merely which spans were labeled, enabling proper evaluation of model clinical understanding.\n\n---\n\n## 24. Event-Causal RAG: A Retrieval-Augmented Generation Framework for Long Video Reasoning in Complex Scenarios\n- **Authors**: Peizheng Yan , Yu Zhao , Liang Xie , Juntong Qi , Mingming Wang , Erwei Yin\n- **URL**: [https://arxiv.org/abs/2605.06185](https://arxiv.org/abs/2605.06185)\n- **Abstract**:\n> Recent large vision-language models have achieved strong performance on short- and medium-length video understanding, yet they remain inadequate for ultra-long or even infinite video reasoning, where models must preserve coherent memory over extended durations and infer causal dependencies across temporally distant events. Existing end-to-end video understanding methods are fundamentally limited by the $O(n^2)$ complexity of self-attention, while recent retrieval-augmented generation (RAG) approaches still suffer from fragmented clip-level memory, weak modeling of temporal and causal structure, and high storage and online inference costs. We present Event-Causal RAG, a lightweight retrieval-augmented framework for infinite long-video reasoning. Instead of indexing fixed-length clips, our method segments streaming videos into semantically coherent events and represents each event as a structured State-Event-State (SES) graph, capturing the event together with its surrounding state transitions. These graphs are merged into a global Event Knowledge Graph and stored in a dual-store memory that supports both semantic matching and causal-topological retrieval. On top of this memory, we design a bidirectional retrieval strategy to efficiently identify the most relevant event causal chains and provide them, together with the associated video evidence, to a backbone video foundation model for answer generation. Experiments on long-video understanding benchmarks demonstrate that Event-Causal RAG consistently outperforms strong clip-based retrieval baselines and long-context video models, particularly on questions requiring multi-event integration and causal inference across long temporal gaps, while also achieving improved memory efficiency and robust streaming performance.\n\n---\n\n## 25. Post Reasoning: Improving the Performance of Non-Thinking Models at No Cost\n- **Authors**: Richmond Sin Jing Xuan , Rishabh Bhardwaj , Soujanya Poria\n- **URL**: [https://arxiv.org/abs/2605.06165](https://arxiv.org/abs/2605.06165)\n- **Abstract**:\n> As the widespread adoption of Large Language Models (LLMs) accelerates, token consumption from intermediate reasoning traces increasingly contributes to inference latency and operational cost. Recent studies suggest that many real-world tasks require little to no explicit reasoning, with additional reasoning sometimes even degrading performance. In this work, we propose \\textbf{Post-Reasoning}, a simple yet effective approach that improves instruction-tuned models by conditioning them to justify their answers after generating the final response. By design, it enables the final answer to be obtained without additional latency or token cost, while still improving performance through simple instruction augmentation. We evaluate Post-Reasoning across \\(117\\) model--benchmark settings spanning \\(13\\) open and proprietary models, \\(4\\) model families, and \\(9\\) diverse reasoning and knowledge-intensive benchmarks, including AMC, HMMT, GSM8K, GPQA, MMLU-Pro, and BIG-Bench Hard. Post-Reasoning improves performance in over \\(88.19\\%\\) of evaluated settings, achieving a mean relative improvements of \\(17.37\\%\\). Furthermore, we propose supervised post-reason tuning, which further improves performance in over \\(91.11\\%\\) of evaluated settings, and exceeds the prompt-based post-reasoning baseline by an average of \\(8.01\\%\\), demonstrating that post-reasoning can be effectively internalized through training. Ultimately, Post-Reasoning establishes a new performance ceiling for direct-answer capabilities.\n\n---\n\n## 26. Beyond Accuracy: Policy Invariance as a Reliability Test for LLM Safety Judges\n- **Authors**: Shihao Weng , Yang Feng , Xiaofei Xie\n- **URL**: [https://arxiv.org/abs/2605.06161](https://arxiv.org/abs/2605.06161)\n- **Abstract**:\n> LLM-as-a-Judge pipelines have become the de facto evaluator for agent safety, yet existing benchmarks treat their verdicts as ground-truth proxies without checking whether the verdicts depend on the agent's behavior or merely on how the evaluation policy happens to be worded. We argue that any trustworthy safety judge must satisfy a basic property we call policy invariance, and we operationalize it as three testable principles: rubric-semantics invariance under certified-equivalent rewrites, rubric-threshold invariance under intentional strict-to-lenient shifts, and ambiguity-aware calibration so that verdict instability concentrates on genuinely ambiguous cases. Instantiating these principles as a stress-test protocol with four agent-class judges on trajectories drawn from ASSEBench and R-Judge, we surface a previously unmeasured failure mode: today's judges respond to meaningful normative shifts and to meaningless structural rewrites with comparable strength, and cannot tell the two apart. Content-preserving policy rewrites flip up to 9.1% of verdicts above baseline jitter, and 18-43% of all observed flips occur on unambiguous cases under such rewrites, so existing safety scores conflate what the agent did with how the evaluator was prompted. Beyond the diagnosis, we contribute the Policy Invariance Score and the Judge Card reporting protocol, which expose an order-of-magnitude spread in judge reliability that is invisible to accuracy-only leaderboards. We release the protocol and code so that future agent-safety benchmarks can audit their own evaluators rather than trust them by default.\n\n---\n\n## 27. Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning\n- **Authors**: Yaorui Shi , Yuxin Chen , Zhengxi Lu , Yuchun Miao , Shugui Liu , Qi GU , Xunliang Cai , Xiang Wang , An Zhang\n- **URL**: [https://arxiv.org/abs/2605.06130](https://arxiv.org/abs/2605.06130)\n- **Abstract**:\n> A persistent skill library allows language model agents to reuse successful strategies across tasks. Maintaining such a library requires three coupled capabilities. The agent selects a relevant skill, utilizes it during execution, and distills new skills from experience. Existing methods optimize these capabilities in isolation or with separate reward sources, resulting in partial and conflicting evolution. We propose Skill1, a framework that trains a single policy to co-evolve skill selection, utilization, and distillation toward a shared task-outcome objective. The policy generates a query to search the skill library, re-ranks candidates to select one, solves the task conditioned on it, and distills a new skill from the trajectory. All learning derives from a single task-outcome signal. Its low-frequency trend credits selection and its high-frequency variation credits distillation. Experiments on ALFWorld and WebShop show that Skill1 outperforms prior skill-based and reinforcement learning baselines. Training dynamics confirm the co-evolution of the three capabilities, and ablations show that removing any credit signal degrades the evolution.\n\n---\n\n## 28. Back to the Beginning of Heuristic Design: Bridging Code and Knowledge with LLMs\n- **Authors**: Nguyen Viet Tuan Kiet , Bui Dinh Pham , Dao Van Tung , Tran Cong Dao , Huynh Thi Thanh Binh\n- **URL**: [https://arxiv.org/abs/2605.06123](https://arxiv.org/abs/2605.06123)\n- **Abstract**:\n> Large language models (LLMs) have recently advanced automatic heuristic design (AHD) for combinatorial optimization (CO), where candidate heuristics are iteratively proposed, evaluated, and refined. Most existing approaches search over executable programs and distill insights from execution feedback to guide later iterations. Because this process moves from low-level implementations to high-level principles, we refer to it as a bottom-up paradigm. We argue that this view is incomplete and introduce a complementary top-down perspective: knowledge becomes the primary search object and code merely instantiates and tests it, making what is learned explicit and reusable across problems and trajectories. We formalize this shift through a statistical-learning view that exposes a distortion--compression trade-off, and instantiate it in both population-based and tree-based AHD frameworks. Across CO and tasks beyond it, knowledge-first search improves discovery efficiency, transfer, and generalization, often outperforming code-centric pipelines, while combining both strategies yields further gains. Our results suggest that progress in AHD depends on iteratively constructing and evolving interpretable hypotheses that retain value beyond a single search trajectory.\n\n---\n\n## 29. Policy-Guided Stepwise Model Routing for Cost-Effective Reasoning\n- **Authors**: Wenwen Si , Insup Lee , Osbert Bastani\n- **URL**: [https://arxiv.org/abs/2605.06116](https://arxiv.org/abs/2605.06116)\n- **Abstract**:\n> Inference-time computation has greatly enhanced the performance of large language models (LLMs) on challenging reasoning tasks, but this strategy can incur high inference costs. One solution is to route intermediate chain-of-thought (CoT) states to language models of different sizes; however, existing approaches rely on handcrafted routing strategies that limit performance, or on training large process reward models that may be infeasible in many applications. We formulate stepwise model routing as a constrained decision-making problem, which we solve by training a small control policy using reinforcement learning in conjunction with threshold calibration to tune the performance-efficiency tradeoff. We validate our method on three math benchmarks (GSM8K, MATH500, and OmniMath) on both open and closed models. Our method consistently improves the accuracy-cost tradeoff compared to handcrafted approaches, while achieving a comparable tradeoff to methods that require training large process reward models.\n\n---\n\n## 30. CrossCult-KIBench: A Benchmark for Cross-Cultural Knowledge Insertion in MLLMs\n- **Authors**: Zhen Zeng , Leijiang Gu , Feng Li , Jing Yu , Zenglin Shi\n- **URL**: [https://arxiv.org/abs/2605.06115](https://arxiv.org/abs/2605.06115)\n- **Abstract**:\n> Multimodal Large Language Models (MLLMs), trained primarily on English-centric data, frequently generate culturally inappropriate or misaligned responses in cross-cultural settings. To mitigate this, we introduce the task of cross-cultural knowledge insertion, which focuses on adapting models to specific cultural contexts while preserving their original behavior in other cultures. To facilitate research in this area, we introduce CrossCult-KIBench, a comprehensive evaluation benchmark for assessing both the effectiveness of knowledge insertion and its unintended side effects on non-target cultures. The benchmark includes 9,800 image-grounded cases covering 49 culturally relevant visual scenarios across English, Chinese, and Arabic language-culture groups. It supports evaluation in both single-insert and sequential-insert settings. We also propose Memory-Conditioned Knowledge Insertion (MCKI) as a baseline method. MCKI retrieves relevant cultural knowledge from an external memory using frozen MLLM representations, prepending matched entries as conditional prompts when applicable. Extensive experiments on CrossCult-KIBench reveal that current approaches struggle to balance effective cultural adaptation with behavioral preservation, highlighting a key challenge in developing culturally-aware MLLMs. Our work thus underscores an important research direction for developing more culturally adaptive and responsible MLLMs.\n\n---\n\n## 31. Shallow Prefill, Deep Decoding: Efficient Long-Context Inference via Layer-Asymmetric KV Visibility\n- **Authors**: Jungsuk Oh , Hyeseo Jeon , Hyunjune Ji , Kyongmin Kong , Jay-Yoon Lee\n- **URL**: [https://arxiv.org/abs/2605.06105](https://arxiv.org/abs/2605.06105)\n- **Abstract**:\n> Long-context inference in decoder-only language models is costly because long prompts are processed during Prefill, cached at every layer, and repeatedly attended to during autoregressive Decode. We introduce \\emph{Shallow Prefill, dEEp Decode} (SPEED), a phase-asymmetric KV-visibility policy that materializes non-anchor prompt-token KV states only in lower layers while keeping Decode-phase tokens full-depth. Unlike previous approaches that make upper-layer prompt KV states cheaper to store or construct, SPEED removes prefill tokens from the upper-layer Decode visibility set altogether. With a minimal BoS anchor, this simple change preserves broad benchmark quality while reducing long-context cost. In a controlled Llama-3.1-8B instruction-tuning study, SPEED using only 75\\% of layers for prefill tokens reaches 51.2 average score on OLMES-style benchmarks, compared with 51.4 for the full-depth baseline, while improving TTFT by 33\\%, TPOT by 22\\%, and reducing active KV memory by 25.0\\% at 128K context. Layer-wise diagnostics suggest that this cutoff retains the main prompt-selection and representation-stabilization regions of the full-depth model. These results show that long-context prompt tokens need not always persist as full-depth KV-cache objects when Decode-phase tokens remain full-depth.\n\n---\n\n## 32. VibeServe: Can AI Agents Build Bespoke LLM Serving Systems?\n- **Authors**: Keisuke Kamahori , Shihang Li , Simon Peter , Baris Kasikci\n- **URL**: [https://arxiv.org/abs/2605.06068](https://arxiv.org/abs/2605.06068)\n- **Abstract**:\n> For years, we have built LLM serving systems like any other critical infrastructure: a single general-purpose stack, hand-tuned over many engineer-years, meant to support every model and workload. In this paper, we take the opposite bet: a multi-agent loop that automatically synthesizes bespoke serving systems for different usage scenarios. We propose VibeServe, the first agentic loop that generates entire LLM serving stacks end-to-end. VibeServe uses an outer loop to plan and track the search over system designs, and an inner loop to implement candidates, check correctness, and measure performance on the target benchmark. In the standard deployment setting, where existing stacks are highly optimized, VibeServe remains competitive with vLLM, showing that generation-time specialization need not come at the cost of performance. More interestingly, in non-standard scenarios, VibeServe outperforms existing systems by exploiting opportunities that generic systems miss in six scenarios involving non-standard model architectures, workload knowledge, and hardware-specific optimizations. Together, these results suggest a different point in the design space for infrastructure software: generation-time specialization rather than runtime generality. Code is available at this https URL .\n\n---\n\n## 33. Visual Fingerprints for LLM Generation Comparison\n- **Authors**: Amal Alnouri , Andreas Hinterreiter , Christina Humer , Furui Cheng , Marc Streit\n- **URL**: [https://arxiv.org/abs/2605.06054](https://arxiv.org/abs/2605.06054)\n- **Abstract**:\n> Large language model (LLM) outputs arise from complex interactions among prompts, system instructions, model parameters, and architecture. We refer to specific configurations of these factors as generation conditions, each of which can bias outputs in various ways. Understanding how different generation conditions shape model behaviors is essential for tasks such as prompt design and model evaluation, yet it remains challenging due to the stochastic and open-ended nature of text generation. We present an approach to visually compare LLM outputs across generation conditions by modeling responses as collections of linguistic choices, including content, expression, and structure. We extract these choices using natural language processing pipelines and represent their distributions across repeated samples. We then visualize these distributions as visual fingerprints, enabling direct, distribution-level comparison of condition-specific tendencies. Through four usage scenarios, we demonstrate how visual fingerprints reveal consistent patterns in LLM behavior that are difficult to observe through individual responses or aggregate metrics.\n\n---\n\n## 34. Novelty-based Tree-of-Thought Search for LLM Reasoning and Planning\n- **Authors**: Leon Hamm , Zlatan Ajanovic\n- **URL**: [https://arxiv.org/abs/2605.06040](https://arxiv.org/abs/2605.06040)\n- **Abstract**:\n> Although advances such as chain-of-thought, tree-of-thought or reinforcement learning have improved the performance of LLMs in reasoning and planning tasks, they are still brittle and have not achieved human-level performance in many domains, and often suffer from high time and token costs. Inspired by the success of width-based search in planning, we explore how the concept of novelty can be transferred to language domains and how it can improve tree-of-thought reasoning. A tree of thoughts relies on building possible \"paths\" of consecutive ideas or thoughts. These are generated by repeatedly prompting an LLM. In our paper, a measurable concept of novelty is proposed that describes the uniqueness of a new node (thought) in comparison to nodes previously seen in the search tree. Novelty is estimated by prompting an LLM and making use of embedded general knowledge from pre-training. This metric can then be used to prune branches and reduce the scope of the search. Although this method introduces more prompts per state, the overall token cost can be reduced by pruning and reducing the overall tree size. This procedure is tested and compared using several benchmarks in language-based planning and general reasoning.\n\n---\n\n## 35. Strat-LLM: Stratified Strategy Alignment for LLM-based Stock Trading with Real-time Multi-Source Signals\n- **Authors**: Wenliang Huang , Zengyi Yu\n- **URL**: [https://arxiv.org/abs/2605.06024](https://arxiv.org/abs/2605.06024)\n- **Abstract**:\n> Large Language Models (LLMs) are evolving into autonomous trading agents, yet existing benchmarks often overlook the interplay between architectural reasoning and strategy consistency. We propose Strat-LLM, a framework grounded in Stratified Strategy Alignment. Operating in a live-forward setting throughout 2025, it integrates heterogeneous data including sequential prices, real-time news, and annual reports to eliminate look-ahead bias. Extensive stress tests on A-share and U.S. markets reveal: (1) reasoning-heavy models achieve peak utility in Free Mode via internal logic, whereas standard models require Strict Mode as a vital risk anchor; (2) alignment utility is regime-dependent, with Free and Guided modes capturing momentum in uptrending markets, while Strict Mode mitigates drawdowns in downtrends; (3) mid-scale models (35B) show optimal fidelity under strict constraints, whereas ultra-large models (122B) suffer an alignment tax under rigid rules but gain a performance premium in Guided Mode; (4) standard LLMs often fall into a high win-rate trap, optimizing for small gains at the expense of total returns, which can only be mitigated through deep reasoning or strict external guardrails. Project details are available at this https URL .\n\n---\n\n## 36. TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steering\n- **Authors**: Yuan Sui , Yulin Chen , Yibo Li , Xue Jiang , Yufei He , Yihong Dong , Xiaoxin He , Tianyu Gao , Bryan Hooi\n- **URL**: [https://arxiv.org/abs/2605.05980](https://arxiv.org/abs/2605.05980)\n- **Abstract**:\n> When language model agents tackle complex software engineering tasks, they often degrade over long trajectories, which we define as *agent drift*. We focus on two recurring failure modes *overthinking* and *overacting*, i.e., where the agent repeatedly reasons over information it already has, and where it issues tool calls without integrating recent observations or acquiring new evidence. In this paper, we introduce TACT (Think-Act Calibration via activation Steering), to detect and mitigate agent drift in the residual stream before it surfaces as a behavioral failure. In specific, we label trajectory steps as overthinking, overacting, or calibrated, and find that their hidden states can separate linearly along two *drift axes*, pointing from calibrated behavior toward each failure mode (AUC $\\approx$ 0.9). To mitigate agent drift, we project each step's activation onto these axes at test time and pull drifted ones back toward the calibrated region. Experiments show that TACT outperforms unsteered baselines across SWE-bench Verified, Terminal-Bench 2.0, and CLAW-Eval, lifting average resolve rate by $+5.8$ pp on Qwen3.5-27B and $+4.8$ pp on Gemma-4-26B-A4B-it while cutting steps-to-resolve by up to $26\\%$. These gains frame agent drift as a steerable direction in the residual stream, and position TACT as a viable handle for reliable long-horizon agents.\n\n---\n\n## 37. TheraAgent: Self-Improving Therapeutic Agent for Precise and Comprehensive Treatment Planning\n- **Authors**: Junkai Li , Yunghwei Lai , Tianyi Zhu , Zheng Long Lee , Weizhi Ma , Yang Liu\n- **URL**: [https://arxiv.org/abs/2605.05963](https://arxiv.org/abs/2605.05963)\n- **Abstract**:\n> Formulating a treatment plan is inherently a complex reasoning and refinement task rather than a simple generation problem. However, existing large language models (LLMs) mainly rely on one-shot output without explicit verification, which may result in rough, incomplete, and potentially unsafe treatment plans. To address these limitations, we propose TheraAgent, an agentic framework that replaces one-shot generation with an iterative generate-judge-refine pipeline. By mirroring the actual reasoning process of human experts who iteratively revise treatment plans, our framework progressively transforms coarse and incomplete drafts into precise, comprehensive, and safer therapeutic regimens. To facilitate the critical judge component, we introduce TheraJudge, a treatment-specific evaluation module integrated into the inference loop to enforce clinical standards. Experiments show TheraAgent achieves state-of-the-art results on HealthBench, leading in Accuracy and Completeness. In expert evaluations, it attains an 86% win rate against physicians, with superior Targeting and Harm Control. Moreover, the highly agreement between TheraJudge and HealthBench evaluations confirms the reliability of our framework.\n\n---\n\n## 38. ICU-Bench:Benchmarking Continual Unlearning in Multimodal Large Language Models\n- **Authors**: Yuhang Wang , Wenjie Mei , Junkai Zhang , Guangyu He , Zhenxing Niu , Haichang Gao\n- **URL**: [https://arxiv.org/abs/2605.05938](https://arxiv.org/abs/2605.05938)\n- **Abstract**:\n> Although Multimodal Large Language Models (MLLMs) have achieved remarkable progress across many domains, their training on large-scale multimodal datasets raises serious privacy concerns, making effective machine unlearning increasingly necessary. However, existing benchmarks mainly focus on static or short-sequence settings, offering limited support for evaluating continual privacy deletion requests in realistic deployments. To bridge this gap, we introduce ICU-Bench, a continual multimodal unlearning benchmark built on privacy-critical document data. ICU-Bench contains 1,000 privacy-sensitive profiles from two document domains, medical reports and labor contracts, with 9,500 images, 16,000 question-answer pairs, and 100 forget tasks. Additionally, new continual unlearning metrics are introduced, facilitating a comprehensive analysis of forgetting effectiveness, historical forgetting preservation, retained utility, and stability throughout the continual unlearning process. Through extensive experiments with representative unlearning methods on ICU-Bench, we show that existing methods generally struggle in continual settings and exhibit clear limitations in balancing forgetting quality, utility preservation, and scalability over long task sequences. These findings highlight the need for multimodal unlearning methods explicitly designed for continual privacy deletion.\n\n---\n\n## 39. Wisteria: A Unified Multi-Scale Feature Learning Framework for DNA Language Model\n- **Authors**: Weihua Wang , Haoji Li , Feilong Bao , Lei Yang , Guanglai Gao\n- **URL**: [https://arxiv.org/abs/2605.05913](https://arxiv.org/abs/2605.05913)\n- **Abstract**:\n> DNA language model aims to decipher the regulatory grammar and semantic of genomes by capturing long range dependencies in DNA sequences. Existing methods emphasize long range token interactions but often ignore the interplay between local motifs and global dependencies. In this paper, we propose Wisteria, a genomic language model that integrates multi scale feature learning within a unified framework for DNA sequence. Specifically, Wisteria augments the Mamba based architecture with gated dilated convolutions to capture local motifs and regulatory patterns, while gated multilayer perceptrons refine global dependencies. We further introduce a Fourier based attention mechanism to support frequency domain modeling, periodic extension and length generalization. Across four experimental settings with both short and long range dependencies, Wisteria demonstrates strong performance on downstream benchmarks against competitive DNA language model baselines. These results indicate that Wisteria effectively unifies local and global dependency modeling for multi scale genomic sequence analysis.\n\n---\n\n## 40. Null Space Constrained Contrastive Visual Forgetting for MLLM Unlearning\n- **Authors**: Yuhang Wang , Zhenxing Niu , Haoxuan Ji , Guangyu He , Linlin Zhang , Haichang Gao\n- **URL**: [https://arxiv.org/abs/2605.05909](https://arxiv.org/abs/2605.05909)\n- **Abstract**:\n> The core challenge of machine unlearning is to strike a balance between target knowledge removal and non-target knowledge retention. In the context of Multimodal Large Language Models (MLLMs), this challenge becomes even more pronounced, as knowledge is further divided into visual and textual modalities that are tightly intertwined. In this paper, we introduce an MLLM unlearning approach that aims to forget target visual knowledge while preserving non-target visual knowledge and all textual knowledge. Specifically, we freeze the LLM backbone and achieve unlearning by fine-tuning the visual module. First, we propose a Contrastive Visual Forgetting (CVF) mechanism to separate target visual knowledge from retained visual knowledge, guiding the representations of target visual concepts toward appropriate regions in the feature space. Second, we identify the null space associated with retained knowledge and constrain the unlearning process within this space, thereby significantly mitigating degradation in knowledge retention. Third, beyond static unlearning scenarios, we extend our approach to continual unlearning, where forgetting requests arrive sequentially. Extensive experiments across diverse benchmarks demonstrate that our approach achieves a strong balance between effective forgetting and robust knowledge retention.\n\n---\n\n## 41. Taklif.AI: LLM-Powered Platform for Interest-Based Personalized College Assignments\n- **Authors**: Zaki Kurdya , Mohammed Zuqlam , Salem Amassi , Shady Telbany , Motaz Saad\n- **URL**: [https://arxiv.org/abs/2605.05842](https://arxiv.org/abs/2605.05842)\n- **Abstract**:\n> Educators face significant challenges in creating engaging, personalized assignments that accommodate students' diverse interests and cognitive abilities. Traditional one-size-fits-all assignments frequently lead to decreased student engagement and increased reliance on unethical practices such as plagiarism. To address these challenges, we present this http URL , a platform that leverages Large Language Models (LLMs) to automatically generate personalized assignments tailored to individual student interests. Unlike existing AI-powered educational platforms that personalize based on academic performance metrics alone, this http URL incorporates students' extracurricular interests and cultural contexts into the assignment generation process through a structured prompt engineering pipeline with input and output guardrails. The platform employs a serverless architecture on AWS with this http URL , using Llama 3.3 70B as the primary LLM via LiteLLM for multi-provider load balancing and LangChain for prompt orchestration. We describe the system architecture, the prompt design methodology, and the guardrails framework that ensures output quality. Preliminary user acceptance testing with 68 participants (65 students and 3 educators) indicates positive reception, with 84% of participants rating the personalization feature as beneficial. We discuss the platform's current capabilities and limitations, and outline directions for rigorous empirical evaluation of learning outcomes.\n\n---\n\n## 42. On the Role of Language Representations in Auto-Bidding: Findings and Implications\n- **Authors**: Guanyu Zhu , Jining Luan , Hanwen Du , Xinyu Fang , Sibo Xu , Ersheng Ni , Hongji Li , Jincheng Fang , Ronghao Chen , Huacan Wang , Xuanqi Lan , Yongxin Ni , Yiqi Sun , Youhua Li\n- **URL**: [https://arxiv.org/abs/2605.05833](https://arxiv.org/abs/2605.05833)\n- **Abstract**:\n> Auto-bidding is a crucial task in real-time advertising markets, where policies must optimize long-horizon value under delivery constraints (e.g., budget and CPA). Existing methods for auto-bidding rely on compact numerical state representations: while they can implicitly capture delivery dynamics, they offer limited support for explicitly representing and controlling high-level intent, evolving feedback, and operator-style strategic guidance in real campaigns. Meanwhile, Large Language Models (LLMs) offer a powerful method for encoding semantic information, it remains unclear when LLMs help and how to integrate them without sacrificing numerical precision. Through systematic preliminary studies, we find that (1) LLM embeddings contain bidding-relevant cues yet cannot replace numerical features, and (2) gains emerge only with careful semantic--numeric integration rather than naive concatenation. Motivated by these findings, we propose \\textit{SemBid}, a novel auto-bidding framework that injects LLM-encoded semantics into offline bidding trajectories at the token level. SemBid introduces three semantic inputs: \\textit{Task}, \\textit{History}, and \\textit{Strategy}. It injects these semantics as tokens alongside numerical trajectory tokens and uses self-attention to integrate them, improving controllability and generalization across objectives. Across diverse scenarios and budget regimes, SemBid outperforms competitive baselines from offline RL and generative sequence modeling, with more consistent gains in overall performance, constraint satisfaction, and robustness. Our code is available at: \\href{ this https URL }{\\textcolor{blue}{here} }.\n\n---\n\n## 43. AGPO: Asymmetric Group Policy Optimization for Verifiable Reasoning and Search Ads Relevance at JD\n- **Authors**: Yang Xu , Kun Yao , Yiming Deng , Zheng Fang , Kai Ming Ting , Ming Pang\n- **URL**: [https://arxiv.org/abs/2605.05826](https://arxiv.org/abs/2605.05826)\n- **Abstract**:\n> Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated notable success in enhancing the reasoning performance of large language models (LLMs). However, recent studies reveal that while current RLVR methods improve sampling efficiency towards correct paths, they do not elicit fundamentally new reasoning patterns. Instead, the reasoning capability boundary of trained models often narrows compared to their base models, with base models achieving higher coverage at large sample sizes. In this work, we propose Asymmetric Group Policy Optimization (AGPO) to counteract this boundary shrinkage. AGPO adopts a negative-dominant reinforcement strategy to suppress incorrect reasoning paths, maintaining the base model's exploration capacity. For positive reinforcement, AGPO adopts a group advantage mechanism, which scales positive updates based on intra-group variance, allowing the model to focus on rare correct paths while suppressing updates from trivial paths. Our experiments on five mathematical benchmarks demonstrate that AGPO achieves state-of-the-art accuracy while consistently improving pass@$k$ performance at scale. In a large-scale industrial application for search ads relevance optimization, AGPO effectively enhances the quality of the data annotation, leading to substantial performance gains in downstream student models.\n\n---\n\n## 44. CircuitFormer: A Circuit Language Model for Analog Topology Design from Natural Language Prompt\n- **Authors**: Md Touhidul Islam , Sujan Kumar Saha , Farimah Farahmandi , Mark Tehranipoor\n- **URL**: [https://arxiv.org/abs/2605.05773](https://arxiv.org/abs/2605.05773)\n- **Abstract**:\n> Automating analog circuit design remains a longstanding challenge in Electronic Design Automation (EDA). While Transformer-based Large Language Models (LLMs) have revolutionized software code generation, their application to analog hardware design is hindered by two critical limitations: (i) the scarcity of analog design datasets containing natural language description of a design and its corresponding netlist, and (ii) the inefficiency of general-purpose tokenizers (e.g., Byte Pair Encoding (BPE)) in capturing the inherent graph structure of circuits. To bridge this gap, first, we curate the largest annotated dataset of analog circuit netlists to date, comprising 31,341 netlist-natural language description pairs across all major circuit classes. Furthermore, we propose Circuit Tokenizer (CKT), a novel circuit graph tokenizer designed to encode netlist connectivity by explicitly mining frequent subcircuits. In terms of scalability, CKT overcomes the bottleneck of prior circuit graph serialization methods where vocabulary size scales linearly with maximum number of components in the dataset, n_max, (O(n_max)); instead, CKT decouples vocabulary growth from circuit complexity, achieving a constant O(1) complexity. Empirically, CKT outperforms standard BPE on circuit topology representation, reducing sequence length by 57% and achieving a 2.3x superior compression ratio using a compact, fixed vocabulary of size 512. Leveraging this optimized tokenization, we train a circuit-specific language model, CircuitFormer, a 511M parameter encoder-decoder transformer. Our model achieves 100% syntactic correctness and an 83% functional success rate across all major analog circuit categories, outperforming state-of-the-art open-source LLMs by 10% and 14%, respectively, while requiring 240x fewer parameters. The dataset is publicly available at this https URL .\n\n---\n\n## 45. HyperLens: Quantifying Cognitive Effort in LLMs with Fine-grained Confidence Trajectory\n- **Authors**: Chengda Lu , Xiaoyu Fan , Wei Xu\n- **URL**: [https://arxiv.org/abs/2605.05741](https://arxiv.org/abs/2605.05741)\n- **Abstract**:\n> While Large Language Models (LLMs) achieve strong performance across diverse tasks, their inference dynamics remain poorly understood because of the limited resolution of existing analysis tools. In this work, we identify an intrinsic magnification mechanism in transformer architectures: deeper layers inherently magnify the small changes of layer-wise confidence, providing a fine-grained confidence trajectory. Building on this insight, we introduce HyperLens, a high-resolution probe designed to trace confidence trajectories and quantify the cognitive effort during inference. Across LLMs and datasets, HyperLens reveals a consistent divergence in confidence trajectories that separates complex from simple tasks. We abstract this pattern into a quantitative cognitive effort metric. Our analysis reveals a fundamental principle: complex tasks consistently require higher cognitive effort. Finally, we provide a mechanistic diagnosis of a common side effect of standard Supervised Fine-Tuning (SFT): it can reduce cognitive effort and consequently degrade performance on in-domain tasks.\n\n---\n\n## 46. ReFlect: An Effective Harness System for Complex Long-Horizon LLM Reasoning\n- **Authors**: Fan Huang\n- **URL**: [https://arxiv.org/abs/2605.05737](https://arxiv.org/abs/2605.05737)\n- **Abstract**:\n> Current reasoning paradigms for LLMs include chain-of-thought, ReAct, and post-hoc self-critique. These paradigms rely on two assumptions that fail on long-horizon, multi-stage tasks. As a result, errors accumulate silently across reasoning steps, leaving an open question: can a reasoning system effectively detect and recover from its own failures? We present ReFlect, a \\emph{harness} system for LLM reasoning that creates standalone error detection and recovery logic as a deterministic wrapper around the model. Controlled experiments across 6 reasoning domains show that prompt-level self-critique produces formulaic templates that flag no issues in 90 of 100 audited reflection blocks, and the investigated LLMs wrongly accept a wrong answer in at least 76\\% of cases. Our ReFlect harness achieves task success rates ranging from 41\\% on gpt-4o-mini to 56\\% on Claude Sonnet 4.5 across six models spanning small and frontier scale, with per-model gains over Direct CoT ranging from +7 pp on Qwen2.5-72B to +29 pp on Claude Sonnet 4.5, and additionally raises SWE-bench patch-structural quality from 0\\% (Direct CoT) to between 82\\% (Qwen2.5-72B) and 87\\% (GPT-4o). Notably, the harness gain is inversely proportional to the model's Direct CoT task success rate (the fitted slope is -1.69 with r=-0.76): each pp lost in baseline success rate is mechanically recovered by 1.69 pp of harness gain. We spot that adding structured reasoning state and operators yields only 15.0--18.7\\% pair-mean on Llama-3.3-70B and Qwen2.5-72B because models at this scale cannot reliably populate the state its operators require. ReFlect is model-agnostic, training-free, and operates entirely at inference time.\n\n---\n\n## 47. Knee Osteoarthritis Severity Grading Using Optimized Deep Learning and LLM-Driven Intelligent AI on Computationally Limited Systems\n- **Authors**: Dayam Nadeem , Neha , Safdar Mustafa , Adnan Alvi , Mohd Hussain\n- **URL**: [https://arxiv.org/abs/2605.05731](https://arxiv.org/abs/2605.05731)\n- **Abstract**:\n> Knee osteoarthritis (KOA) is among the musculoskeletal disorders that considerably restrict joint mobility, cause severe chronic pain and impact negatively on quality life. It is one of the persistent health issues worldwide. Generally, subjectivity and inter-observer variability undermine conventional practices and evaluation process that are adopted to address such health issues. Hence precise and timely diagnosis would be one of the effective ways for the assessment of its severity. This paper proposes an automated diagnostic approach for severity grading of KOA by blending a deep learning convolutional neural network (CNN) with a device-based inference platform powered by TensorFlow Lite. It proposes a model based on the ResNet-18 convolutional neural network. The designed model is trained on publicly available database. Through a transfer learning approach obtained knee images are first classified into five Kellgren-Lawrence (KL) grades. Further the developed model is optimised. During the training of the model test accuracy of 94.48% with stable convergence has been achieved. Subsequently the optimised model transformed into a lightweight TensorFlow Lite format, facilitating seamless deployment on resource-constrained devices. The designed model is capable enough to operate in the environment having no continuous internet connectivity. Also, an auxiliary Large Language Model (Gemini-2.0-flash) is applied to generate structured interpretive findings like potential symptoms, risk factors, and preventive majors etc. The LLM component functions as interface without influencing the classification process. The proposed model articulates the feasibility of an on-device, interpretable decision-support tools for early diagnosis and improve accessibility to Artificial Intelligence (AI)-assisted knee screening tool.\n\n---\n\n## 48. SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents\n- **Authors**: Hongcheol Cho , Ryangkyung Kang , Youngeun Kim\n- **URL**: [https://arxiv.org/abs/2605.05726](https://arxiv.org/abs/2605.05726)\n- **Abstract**:\n> As LLM agents are increasingly deployed with large libraries of reusable skills, selecting the right skill for a user request has become a critical systems challenge. In small libraries, users may invoke skills explicitly by name, but this assumption breaks down as skill ecosystems grow under tight context and latency budgets. Despite its practical importance, skill retrieval remains underexplored, with limited benchmarks and little understanding of retrieval behavior on realistic skill libraries. To address this gap, we introduce SkillRet, a large-scale benchmark for skill retrieval in LLM agents. SkillRet contains 17,810 public agent skills, organized with structured semantic tags and a two-level taxonomy spanning 6 major categories and 18 sub-categories. It provides 63,259 training samples and 4,997 evaluation queries with disjoint skill pools, enabling both benchmarking and retrieval-oriented training. Across a diverse set of retrievers, we find that skill retrieval remains far from solved: off-the-shelf models struggle on realistic large-scale skill libraries, and prior skill-retrieval models still leave substantial headroom. Task-specific fine-tuning on SkillRet substantially improves performance, improving NDCG@10 by +13.1 points over the strongest prior retriever and by +16.9 points over the strongest off-the-shelf retriever. Our analysis further suggests that these gains arise because fine-tuned models better focus on the small skill-relevant signals within long and noisy queries. These results establish SkillRet as a strong benchmark and foundation for future research on retrieval in large-scale agent systems.\n\n---\n\n## 49. Detecting Time Series Anomalies Like an Expert: A Multi-Agent LLM Framework with Specialized Analyzers\n- **Authors**: Hyeongwon Kang , Jeongseob Kim , Jinwoo Park , Pilsung Kang\n- **URL**: [https://arxiv.org/abs/2605.05725](https://arxiv.org/abs/2605.05725)\n- **Abstract**:\n> Recent studies have explored large language models for time-series anomaly detection, yet existing approaches often rely on a single general-purpose model to directly infer anomaly indices or intervals, limiting controllability, interpretability, and reliability for complex anomaly patterns. We propose SAGE (Specialized Analyzer Group for Expert-like Detection), a multi-agent framework for structured anomaly diagnosis in univariate time series. It decomposes anomaly analysis into four specialized Analyzers for point, structural, seasonal, and pattern anomalies. Each Analyzer applies family-specific numerical tools and diagnostic visualizations to generate evidence, while an evidence-grounded Detector consolidates the evidence into confidence-scored anomaly records with intervals and candidate types. A Supervisor then converts these structured records into analyst-facing diagnostic reports. SAGE further constructs synthetic in-context examples from normal-reference training segments, without using real anomalous segments or anomaly-type labels as in-context examples. Across three benchmarks, SAGE achieves the best average performance among strong ML/DL and language-model-based baselines. Ablation studies and human evaluation further show that the proposed framework improves detection reliability and the practical usefulness of diagnostic outputs.\n\n---\n\n## 50. More Is Not Always Better: Cross-Component Interference in LLM Agent Scaffolding\n- **Authors**: Ming Liu\n- **URL**: [https://arxiv.org/abs/2605.05716](https://arxiv.org/abs/2605.05716)\n- **Abstract**:\n> LLM agent systems are built by stacking scaffolding components (planning, tools, memory, self-reflection, retrieval) assuming more is better. We study cross-component interference (CCI): degradation when components interact destructively. We run a full factorial experiment over all 2^5=32 subsets of five components on HotpotQA and GSM8K with Llama-3.1-8B/70B (96 conditions, up to 10 seeds). The All-In system is consistently suboptimal: on HotpotQA, a single-tool agent surpasses All-In by 32% (F1 0.233 vs 0.177, p=0.023); on GSM8K, a 3-component subset beats All-In by 79% (0.43 vs 0.24, p=0.010). Optimal component count is task-dependent (k*=1-4) and scale-sensitive: at 70B, combinations that hurt at 8B provide gains, though All-In still trails the best subset. We fit a main-effects regression (R^2=0.916, adj-R^2=0.899, LOOCV=0.872), compute exact Shapley values, and find 183/325 submodularity violations (56.3%), showing greedy selection is unreliable. A three-body synergy among Tool Use, Self-Reflection, and Retrieval (INT_3=+0.175, 95% CI [+0.003,+0.351]) is reported as exploratory. CCI replicates across model families (Qwen2.5) and is robust to prompt paraphrasing. Our findings suggest maximally-equipped agent defaults should be replaced by task-specific subset selection via interaction-aware analysis.\n\n---\n\n## 51. Decodable but Not Corrected by Fixed Residual-Stream Linear Steering: Evidence from Medical LLM Failure Regimes\n- **Authors**: Ming Liu\n- **URL**: [https://arxiv.org/abs/2605.05715](https://arxiv.org/abs/2605.05715)\n- **Abstract**:\n> Can linearly decodable failure signals in LLM hidden states be leveraged to correct those failures? We investigate this classification-correction gap via Overthinking (OT)--a stable behavioral regime (Jaccard >= 0.81, 94% inter-annotator agreement) in medical QA where models answer correctly under resampling yet fail in extended chain-of-thought. OT is linearly decodable at 71.6% balanced accuracy (p < 10^{-16}). Yet five families of fixed linear steering (29 configurations, n=1,273) all yield Delta ~= 0, with identical null results cross-architecture (Qwen2.5-7B) and cross-domain (MMLU-STEM). Three convergent lines of evidence suggest representational entanglement: the OT direction has 85-88% overlap with task-critical computation (specificity ratio <= 0.152); non-targeted shared-direction steering damages accuracy (-12.1pp); and LEACE concept erasure damages accuracy (-3.6pp, p=0.01), while 10 random erasures produce Delta=+0.3pp. The per-instance probe-steering correlation is r=-0.002 (p=0.97). Positively, the same probe enables selective abstention (held-out AUROC=0.610, exceeding all five uncertainty baselines, p=0.009): decodable failure structure supports post-generation reliability estimation even when the fixed linear steering family cannot exploit it for correction.\n\n---\n\n## 52. Conceal, Reconstruct, Jailbreak: Exploiting the Reconstruction-Concealment Tradeoff in MLLMs\n- **Authors**: Md Farhamdur Reza , Richeng Jin , Tianfu Wu , Huaiyu Dai\n- **URL**: [https://arxiv.org/abs/2605.05709](https://arxiv.org/abs/2605.05709)\n- **Abstract**:\n> Intent-obfuscation-based jailbreak attacks on multimodal large language models (MLLMs) transform a harmful query into a concealed multimodal input to bypass safety mechanisms. We show that such attacks are governed by a \\emph{reconstruction--concealment tradeoff}: the transformed input must hide harmful intent from safety filters while remaining recoverable enough for the victim model to reconstruct the original request. Through a reconstruction analysis of three representative black-box methods, we find that existing transformations struggle to balance this tradeoff, limiting their effectiveness. In contrast, we show that character-removed variants achieve a better balance. Building on this, we propose \\emph{concealment-aware variant construction}, which greedily selects character-removed variants that are low in harmful-keyword alignment and mutually diverse, and instantiates them through five modality-aware prompting strategies. We further introduce \\emph{keyword-related distractor images} that depict the harmful keyword in diverse contexts, providing more effective auxiliary visual context than generic distractor images. Experiments across closed-source and open-source MLLMs show the proposed strategies outperform strong baselines, revealing an underexplored vulnerability: a model's own reconstruction ability can be exploited to recover hidden harmful intent and produce unsafe responses.\n\n---\n\n## 53. Resolving the bias-precision paradox with stochastic causal representation learning for personalized medicine\n- **Authors**: Peisong Zhang , Manqiang Peng , Yuxuan Wu , Pawit Phadungsaksawasdi , Wesley Yeung , Ye Zhang , Trang Nguyen , Qiang Zhang , Nan Liu , Meng Wang , Kee Yuan Ngiam , Yih-Chung Tham , Ching-Yu Cheng , Tianfan Fu , Qingyu Chen , Rosemary Ke , Chang Li , Wenzhuo Yang , Zhenghao Lu , Chunyou Lai , Yu Zhang , Sheng Zhong , Hao Deng , Dianbo Liu\n- **URL**: [https://arxiv.org/abs/2605.05706](https://arxiv.org/abs/2605.05706)\n- **Abstract**:\n> Estimating individualized treatment effects from longitudinal observational data is central to data-driven medicine, yet existing methods face a fundamental limitation: reducing confounding bias often suppresses clinically informative heterogeneity, degrading patient-specific predictions. Here, we identify this tension as a bias-precision paradox in causal representation learning and introduce sampling-based maximum mean discrepancy (sMMD), a stochastic alignment strategy that replaces global adversarial balancing with subset-level matching. We instantiate this approach in a framework for counterfactual outcome prediction with attribution-grounded interpretability. Across two large-scale ICU cohorts (n = 27,783), our framework improves accuracy under distribution shift, reducing error by up to 11.5% and substantially increasing recall in high-risk tasks. Mechanistic analyses show that sMMD selectively preserves clinically decisive variables. In human-AI evaluation, our method outperforms clinicians-in-training and large language models, and improves clinician accuracy by 14.7% while reducing decision time, enabling interpretable, real-time clinical decision support.\n\n---\n\n## 54. Knowledge-Graph Paths as Intermediate Supervision for Self-Evolving Search Agents\n- **Authors**: Huyu Wu , Jun Liu , Xiaochi Wei , Yan Gao , Yi Wu , Yao Hu\n- **URL**: [https://arxiv.org/abs/2605.05702](https://arxiv.org/abs/2605.05702)\n- **Abstract**:\n> Self-evolving search agents reduce reliance on human-written training questions by generating and solving their own search tasks. We build on Search Self-Play (SSP), a representative Proposer and Solver framework in which questions are generated and answered via multi-step search and reasoning. In practice, however, SSP faces two bottlenecks: the Proposer constructs questions from isolated answer entities without relational context, yielding many invalid or unverifiable questions in early self-play training, while the Solver receives only a binary outcome reward that discards useful signal from partially on-track search trajectories. We address both bottlenecks by reusing knowledge-graph paths as construction-derived intermediate supervision for both question construction and reward shaping. First, we ground question construction in LLM-guided knowledge-graph subgraphs, providing relational context for the Proposer. Second, we observe that constructing and solving a multi-hop question can involve overlapping intermediate entities: the factual bridges used to formulate the question may provide approximate waypoints for answering it. Exploiting this overlap, we introduce Waypoint Coverage Reward (WCR), which grants graded partial credit to incorrect Solver trajectories according to their coverage of entities on the construction path, while preserving full reward for correct answers. Across seven QA benchmarks and nine model configurations, our approach improves the average score over standard SSP in all configurations, including notable gains on multi-hop QA tasks. These results suggest that knowledge-graph paths can be reused as lightweight intermediate supervision, providing both relational guidance and process feedback without additional task-specific human annotations or manually labeled process steps.\n\n---\n\n## 55. Inference-Time Budget Control for LLM Search Agents\n- **Authors**: Zhengru Fang , Senkang Forest Hu , Zhonghao Chang , Yu Guo , Yihang Tao , Hongyao Liu , Mengzhe Ruan , Jun Huang , Yuguang Fang\n- **URL**: [https://arxiv.org/abs/2605.05701](https://arxiv.org/abs/2605.05701)\n- **Abstract**:\n> LLM search agents increasingly rely on tools at inference time, but their trajectories are often constrained by hard limits on both tool calls and generated tokens. Under such dual budgets, better answers require not only stronger models, but also explicit control over which search action should receive the next budget unit and when the accumulated evidence is sufficient to commit a final answer. We study this problem in multi-hop question answering (QA) and formulate it as two-stage inference-time budget control. At search time, our controller assigns each feasible action a task-level Value-of-Information (VOI) score, defined as an operational estimate of marginal task value per unit budget under the current search state and remaining dual budget, and uses this score to choose among retrieval, decomposition, and answer commitment. After search, a selective evidence-grounded finalizer compares the trajectory answer with a refined candidate and rewrites only when the residual error appears to be a low-risk answer-form error. Across four multi-hop QA benchmarks, three LLM backbones, and four budget levels, the method yields positive aggregate gains over four audited baselines under the same hard dual-budget protocol. Ablations show that search-time budget control, especially budget-dependent penalty, provides the main performance gain, while answer-time control helps mainly when the retrieval path is already adequate. These results suggest that inference-time budget control for LLM search agents should govern both how budget is spent during search and how the final answer is committed.\n\n---\n\n## 56. Saliency-Aware Regularized Quantization Calibration for Large Language Models\n- **Authors**: Yanlong Zhao , Xiaoyuan Cheng , Huihang Liu , Baihua He , Xinyu Zhang , Harrison Bo Hua Zhu , Wenlong Chen , Li Zeng , Zhuo Sun\n- **URL**: [https://arxiv.org/abs/2605.05693](https://arxiv.org/abs/2605.05693)\n- **Abstract**:\n> Post-training quantization (PTQ) is an effective approach for deploying large language models (LLMs) under memory and latency constraints. Most existing PTQ methods determine quantization parameters by minimizing a layer-wise reconstruction error on a predetermined calibration dataset, usually optimized via either scale search or Gram-based methods. However, from the perspective of generalization risk, existing calibration objectives of PTQ based only on empirical reconstruction error on limited or unrepresentative calibration data could move the quantized weights away from the original weights. This may cause the generalization risk to diverge, potentially degrading downstream performance. To address this issue, we propose \\emph{Saliency-Aware Regularized Quantization Calibration} (SARQC) a unified framework that augments the standard PTQ objective with a saliency-aware regularization term. This term encourages quantized weights to stay close to the original weights during calibration, leading to improved generalization during inference. SARQC integrates seamlessly into existing PTQ pipelines, enhancing both scale search and Gram-based methods under a unified formulation. Extensive experiments on dense and Mixture-of-Experts LLMs demonstrate consistent improvements in perplexity and zero-shot accuracy, without additional computational overhead during inference.\n\n---\n\n## 57. DataDignity: Training Data Attribution for Large Language Models\n- **Authors**: Xiaomin Li , Andrzej Banburski-Fahey , Jaron Lanier\n- **URL**: [https://arxiv.org/abs/2605.05687](https://arxiv.org/abs/2605.05687)\n- **Abstract**:\n> Auditing language-model outputs often requires more than judging correctness: an auditor may need to identify which source document most likely supports the knowledge expressed in a response. We study this as pinpoint provenance: given a prompt, a target-model response, and a candidate corpus, rank the documents that best support the response. We introduce FakeWiki, a controlled benchmark of 3,537 fabricated Wikipedia-style articles designed to preserve ground-truth provenance while weakening lexical shortcuts. FakeWiki includes QA probes, source-preserving paraphrases, retro-generated variants, hard anti-documents that remain topically similar while removing answer-critical facts, and five query conditions: clean prompting plus four jailbreak-inspired transformations. We evaluate seven retrieval baselines, a training-free activation-steering retrieval-fusion method, SteerFuse, and a supervised contrastive provenance ranker, ScoringModel. ScoringModel maps response and document features into a shared space and is trained with InfoNCE using in-batch, retrieval-mined, and anti-document negatives. Across nine open-weight instruction-tuned LLMs and five query conditions, ScoringModel improves mean Recall@10 from 35.0 for the strongest retrieval baseline to 52.2, without inference-time fusion, and wins 41/45 model-by-condition cells. SteerFuse is usually second-best despite requiring no supervised training, showing that activation-space evidence can efficiently complement text retrieval. On jailbreak-inspired transformed queries, ScoringModel improves Recall@10 by 15.7 points on average over the best baseline. Overall, our work shows that robust training data attribution requires evaluation settings that separate true answer support from topical or lexical resemblance.\n\n---\n\n## 58. Attractor Geometry of Transformer Memory: From Conflict Arbitration to Confident Hallucination\n- **Authors**: Qiyao Liang , Risto Miikkulainen , Ila Fiete\n- **URL**: [https://arxiv.org/abs/2605.05686](https://arxiv.org/abs/2605.05686)\n- **Abstract**:\n> Language models draw on two knowledge sources: facts baked into weights (parametric memory, PM) and information in context (working memory, WM). We study two mechanistically distinct failure modes--conflict, when PM and WM disagree and interfere; and hallucination, when the queried fact was never learned. Both produce confident output regardless, making output-based monitoring blind by design. We show both failures share a unified geometric account. In the hidden-state space of autoregressive generation, learned facts form attractor basins. Conflict is basin competition: WM disrupts convergence to the correct basin without raising output entropy. Hallucination is basin absence: the hidden state drifts freely when no memorized basin exists. The frozen LM head, designed for next-token prediction, cannot distinguish these cases and fires confidently either way. We verify this account in a controlled synthetic task--entity identifiers mapped to unique codes with PM installed via LoRA adapters--where ground truth is exact and component roles can be causally isolated through targeted adapter placement. Geometric margin--the hidden state's distance to the nearest memorized basin--reads this geometry directly and separates correct recall from hallucination far more cleanly than output entropy, with zero false refusals where entropy-based detection cannot avoid rejecting the vast majority of correct outputs. The separation holds on natural-language factual queries from the pretrained model with no adaptation, confirming attractor geometry is structural rather than a fine-tuning artifact. The fraction of confident hallucinations follows a scaling law $C = \\exp(-c/\\bar\\Delta)$, growing with scale even as overall error rates fall. Hidden states reliably encode epistemic state; the frozen output head systematically erases it--and this erasure worsens with scale.\n\n---\n\n## 59. Large Vision-Language Models Get Lost in Attention\n- **Authors**: Gongli Xi , Ye Tian , Mengyu Yang , Huahui Yi , Liang Lin , Xiaoshuai Hao , Kun Wang , Wendong Wang\n- **URL**: [https://arxiv.org/abs/2605.05668](https://arxiv.org/abs/2605.05668)\n- **Abstract**:\n> Despite the rapid evolution of training paradigms, the decoder backbone of large vision--language models (LVLMs) remains fundamentally rooted in the residual-connection Transformer architecture. Therefore, deciphering the distinct roles of internal modules is critical for understanding model mechanics and guiding architectural optimization. While prior statistical approaches have provided valuable attribution-based insights, they often lack a unified theoretical basis. To bridge this gap, we propose a unified framework grounded in information theory and geometry to quantify the geometric and entropic nature of residual updates. Applying this unified framework reveals a fundamental functional decoupling: Attention acts as a subspace-preserving operator focused on reconfiguration, whereas FFNs serve as subspace-expanding operators driving semantic innovation. Strikingly, further experiments demonstrate that replacing learned attention weights with predefined values (e.g., Gaussian noise) yields comparable or even superior performance across a majority of datasets relative to vanilla models. These results expose severe misallocation and redundancy in current mechanisms, suggesting that state-of-the-art LVLMs effectively ``get lost in attention'' rather than efficiently leveraging visual context.\n\n---\n\n## 60. Retrieval-Conditioned Topology Selection with Provable Budget Conservation for Multi-Agent Code Generation\n- **Authors**: Abhijit Talluri , Pujith Anne , Bhagavan Choudary Pendiyala , Raghavendra Chilukuri\n- **URL**: [https://arxiv.org/abs/2605.05657](https://arxiv.org/abs/2605.05657)\n- **Abstract**:\n> Multi-agent LLM systems for code generation face a fundamental routing problem: the optimal orchestration topology depends on the structural complexity of the code under modification, yet existing systems select topologies without consulting the codebase. We present Retrieval-Guided Adaptive Orchestration (RGAO), an architecture that closes this loop by extracting a structural complexity vector from a hierarchical code index before selecting the orchestration topology. RGAO operates within Code-Agent, a multi-agent framework whose sub-agents are governed by formal contracts with six-dimensional budget vectors. Our headline contribution is the composition of two previously separate lines of work -- complexity-conditioned LLM routing and formal resource algebras -- yielding a property neither admits alone: provable budget conservation under retrieval-conditioned dynamic topology selection. Concretely we contribute: (1) a complexity-conditioned topology router that reduces proxy-measured misrouting from 30.1% to 8.2%; (2) a budget algebra with a structural-induction conservation theorem; and (3) a hierarchical code retrieval engine. Empirical evaluation demonstrates sub-millisecond DAG construction and linear tree-index scalability.\n\n---\n\n## 61. Text-Graph Synergy: A Bidirectional Verification and Completion Framework for RAG\n- **Authors**: Jiarui Zhong , Hong Cai Chen\n- **URL**: [https://arxiv.org/abs/2605.05643](https://arxiv.org/abs/2605.05643)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) has become a core paradigm for enhancing factual grounding and multi-hop reasoning in Large Language Models (LLMs). Traditional text-based RAG often retrieves logically irrelevant pseudo-evidence, while graph-based RAG is frequently hindered by search-time pruning, which may discard potentially valid reasoning paths. Existing hybrid approaches primarily adopt simple evidence concatenation or unidirectional enhancement, which fails to address the fundamental \"Information Island\" problem caused by asymmetric reasoning flows between unstructured text and structured graphs. We propose \\textbf{TGS-RAG}, a unified framework for \\textbf{T}ext-\\textbf{G}raph \\textbf{S}ynergistic enhancement. TGS-RAG introduces a bidirectional mechanism: (i) a \\textbf{Graph-to-Text} channel that employs a Global Voting strategy from visited graph nodes to re-rank and refine textual evidence, filtering out semantic noise; and (ii) a \\textbf{Text-to-Graph} channel that utilizes the \\textbf{Memory-based Orphan Entity Bridging} algorithm. This algorithm utilizes textual cues to proactively resurrect valid but previously pruned reasoning paths from the search history without additional database overhead. Experimental results on multiple multi-hop reasoning benchmarks demonstrate that TGS-RAG significantly outperforms state-of-the-art baselines, achieving a superior balance between retrieval precision and computational efficiency.\n\n---\n\n## 62. Prober.ai: Gated Inquiry-Based Feedback via LLM-Constrained Personas for Argumentative Writing Development\n- **Authors**: Ran Bi , Shiyao Wei , Yuanyiyi Zhou\n- **URL**: [https://arxiv.org/abs/2605.05598](https://arxiv.org/abs/2605.05598)\n- **Abstract**:\n> The proliferation of large language models (LLMs) in educational settings has paradoxically undermined the cognitive processes they purport to support. Students increasingly outsource critical thinking to AI assistants that generate polished text on demand, resulting in measurable cognitive debt and diminished argumentative reasoning skills. We present this http URL , a web-based writing environment that inverts the conventional AI-tutoring paradigm: rather than generating or rewriting student text, the system constrains an LLM (Gemini 3 Flash Preview) through persona-specific system prompts and structured JSON output schemas to produce only targeted, inquiry-based questions about argumentative weaknesses. A two-phase interaction architecture -- Challenge and Unlock -- implements a pedagogical friction mechanism whereby revision suggestions are gated behind mandatory student reflection. The system's design is grounded in Toulmin's argumentation theory, research on peer feedforward questioning mechanisms, and evidence on AI-supported feedback in writing instruction. A functional prototype was developed in 36 hours during the NY EdTech Hackathon (March 2026), where it was awarded second place. We describe the system architecture, the prompt engineering methodology for constraining LLM output to pedagogically aligned JSON schemas, and discuss implications for scalable, cognition-preserving AI integration in writing education.\n\n---\n\n## 63. Causal Probing for Internal Visual Representations in Multimodal Large Language Models\n- **Authors**: Zehao Deng , Tianjie Ju , Zheng Wu , Liangbo He , Jun Lan , Huijia Zhu , Weiqiang Wang , Zhuosheng Zhang\n- **URL**: [https://arxiv.org/abs/2605.05593](https://arxiv.org/abs/2605.05593)\n- **Abstract**:\n> Despite the remarkable success of Multimodal Large Language Models (MLLMs) across diverse tasks, the internal mechanisms governing how they encode and ground distinct visual concepts remain poorly understood. To bridge this gap, we propose a causal framework based on activation steering to actively probe and manipulate internal visual representations. Through systematic intervention across four visual concept categories, our results reveal a divergence in concept encoding: entities exhibit distinct localized memorization, whereas abstract concepts are globally distributed across the network. Critically, this divergence uncovers a mechanistic driver of scaling laws: increasing model depth is indispensable for encoding distributed and complex abstract concepts, whereas entity localization remains remarkably invariant to scale. Furthermore, reverse steering uncovers that blocking explicit output triggers a surge in latent activations, exposing a compensatory mechanism between perception and generation. Finally, extending our analysis to visual reasoning, we expose a disconnect between perception and reasoning although MLLMs successfully recognize geometric relations, they treat them merely as static visual features, failing to trigger the procedural execution necessary for abstract problem-solving.\n\n---\n\n## 64. Belief Memory: Agent Memory Under Partial Observability\n- **Authors**: Junfeng Liao , Qizhou Wang , Jianing Zhu , Bo Du , Rui Yan , Xiuying Chen\n- **URL**: [https://arxiv.org/abs/2605.05583](https://arxiv.org/abs/2605.05583)\n- **Abstract**:\n> LLM agents that operate over long context depend on external memory to accumulate knowledge over time. However, existing methods typically store each observation as a single deterministic conclusion (e.g., inferring \"API~X failed\" from temporary errors), even though such observations are inherently partial and potentially ambiguous. By committing to one conclusion and discarding uncertainty, these methods introduce self-reinforcing error: the agent acts on the stored conclusion, never revisits alternatives, and reinforces the conclusion over time. To address this issue, we propose BeliefMem, which shifts the memory paradigm from committing to a single conclusion per observation to retaining multiple candidate conclusions with their probabilities. Concretely, BeliefMem stores the candidate conclusions as separate memory entries, each carrying a probability that is updated via Noisy-OR rules as new observations arrive. At retrieval, all candidates surface together with their probabilities, keeping alternatives visible to the agent. Since each conclusion in memory retains its probability, BeliefMem preserves the uncertainty that the deterministic paradigm discards, enabling the agent to act with high confidence on well-evidenced knowledge while retaining the capacity to update its confidence when new evidence arrives. Empirical evaluations on LoCoMo and ALFWorld benchmarks show that, even with limited data, BeliefMem achieves the best average performance, remarkably outperforming well-known baselines. More broadly, such probabilistic memory produces substantial gains and explores a new direction for agent memory in partially observable environments.\n\n---\n\n## 65. AlphaCrafter: A Full-Stack Multi-Agent Framework for Cross-Sectional Quantitative Trading\n- **Authors**: Yishuo Yuan , Jiayi Sheng , Sirui Zeng , Jiaqi Wang , Jiaheng Liu\n- **URL**: [https://arxiv.org/abs/2605.05580](https://arxiv.org/abs/2605.05580)\n- **Abstract**:\n> Financial markets are inherently non-stationary, driven by complex interactions among macroeconomic regimes, microstructural frictions, and behavioral dynamics. Building quantitative strategies that remain profitable demands the continuous coupling of factor discovery, regime-adaptive selection, and risk-constrained execution. Prevailing approaches, however, optimize these components under static or isolated assumptions. Factor mining frameworks typically treat alpha discovery as a one-time search process, implicitly assuming that factor efficacy persists across market regimes. Execution-oriented systems often adopt role-playing agent architectures that simulate anthropomorphic trading committees, introducing behavioral noise rather than systematic rationality. Consequently, a fully automated, rationality-driven framework unifying a coherent quantitative pipeline remains absent. We introduce AlphaCrafter, a full-stack multi-agent framework that closes this gap through a continuously adaptive factor-to-execution pipeline, designed to track and respond to evolving market conditions without manual intervention. AlphaCrafter operates via three specialized agents: a Miner that continuously expands the factor pool via LLM-guided search, a Screener that assesses prevailing market conditions to construct regime-conditioned factor ensembles, and a Trader that translates these ensembles into quantitative strategies under explicit risk constraints. Together, these three agents form a closed-loop cross-sectional trading system that adapts holistically to evolving market dynamics. Extensive experiments on CSI 300 and S&P 500 demonstrate that AlphaCrafter consistently outperforms state-of-the-art baselines in risk-adjusted returns while exhibiting the lowest cross-trial variance, confirming that integrated and adaptive factor-to-execution design yields robust trading performance.\n\n---\n\n## 66. Nonsense Helps: Prompt Space Perturbation Broadens Reasoning Exploration\n- **Authors**: Langlin Huang , Chengsong Huang , Jinyuan Li , Donghong Cai , Yuyi Yang , Jiaxin Huang\n- **URL**: [https://arxiv.org/abs/2605.05566](https://arxiv.org/abs/2605.05566)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards, particularly Group Relative Policy Optimization (GRPO), has significantly advanced the reasoning capabilities of Large Language Models (LLMs). However, in complex tasks, GRPO frequently suffers from the ``zero-advantage problem'': when all sampled rollouts for a query fail, the relative advantage collapses to zero. Consequently, the model loses effective training signals for these questions, wasting the training data and computational budget. While simply increasing the sampling budget for these questions is a common remedy, the static sampling policy inherently constrains reasoning exploration, limiting the success rate. In this paper, we propose Lorem Perturbation for Exploration (LoPE), a simple yet effective training framework to break this exploration bottleneck. We posit that task-irrelevant prompt-space perturbations can shift the model's output distribution enough to unlock orthogonal reasoning pathways for hard questions. Specifically, LoPE prepends sequences stochastically assembled from Lorem Ipsum vocabulary (a pseudo-Latin placeholder text) to the prompts before resampling. Experiments across 1.7B, 4B, and 7B models demonstrate that LoPE significantly outperforms resampling with the original prompts. Further analysis reveals that other Latin-based random sequences with low perplexity are also effective perturbations. Our results establish LoPE as a strong baseline for broadening exploration in LLM reinforcement learning.\n\n---\n\n## 67. SPARK: Self-Play with Asymmetric Reward from Knowledge Graphs\n- **Authors**: Hyobin Park , Taeseop Kim , Dong-Geol Choi\n- **URL**: [https://arxiv.org/abs/2605.05546](https://arxiv.org/abs/2605.05546)\n- **Abstract**:\n> Self-play reinforcement learning has shown strong performance in domains with formally verifiable structure, such as mathematics and coding, where both problem generation and reward computation can be grounded in explicit rules. Extending this paradigm to scientific literature is more challenging: the relationships among multi-modal elements within and across documents are rarely made explicit in text, which makes automatic generation of relational reasoning questions difficult and weakens the reliability of reward signals. We propose SPARK (Self-Play with Asymmetric Reward from Knowledge Graphs), a framework that automatically constructs a unified knowledge graph (KG) from multi-document scientific literature and uses it as the structural basis for self-play. KG paths over multimodal nodes serve as a source for generating relational reasoning questions, and structured facts stored in the KG provide a basis for verifiable reward computation. A single small vision-language model (sVLM) alternates between Proposer and Solver roles under information asymmetry against a fixed KG, a design that we believe can be naturally extended toward online adaptation in future work. We evaluate SPARK on public benchmarks and a self-constructed cross-document multi-hop QA dataset. Results show that SPARK consistently outperforms flat-corpus-based self-play baselines, and the performance gap widens as hop count increases, suggesting that KG-structure grounding contributes to relational multi-hop reasoning beyond what unstructured corpus grounding can provide.\n\n---\n\n## 68. AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases\n- **Authors**: Susheel Suresh , Hazel Mak , Shangpo Chou , Fred Kroon , Sahil Bhatnagar\n- **URL**: [https://arxiv.org/abs/2605.05538](https://arxiv.org/abs/2605.05538)\n- **Abstract**:\n> We present AgenticRAG, a practical agentic harness for retrieval and analysis over enterprise knowledge bases. Standard RAG pipelines place significant burden of grounding on the search stack, constraining the language model to a fixed candidate set chosen deep in the retrieval process. Our approach reduces this overdependence by layering a lightweight harness on top of existing enterprise search infrastructure, equipping a reasoning LLM with search, find, open, and summarize tools enabling the model to iteratively retrieve information, navigate within documents, and analyze evidence autonomously. On three open benchmarks we observe substantial gains: $49.6\\%$ recall@1 on BRIGHT (+21.8 pp over the best embedding baseline), 0.96 factuality on WixQA ($+13\\%$ relative improvement), and $92\\%$ answer correctness on FinanceBench--within 2 pp of oracle access to true evidence. Ablation studies show that the most significant factor is the shift from single-shot retrieval to agentic tool use ($5.9\\times$ improvement), while multi-query search and in-document navigation contribute to both quality and efficiency. We present various design choices in our agentic harness that were informed by pre-production deployments. Our results demonstrate its suitability for real-world enterprise production environments.\n\n---\n\n## 69. FoodCHA: Multi-Modal LLM Agent for Fine-Grained Food Analysis\n- **Authors**: Woojin Lee , Pranav Mekkoth , Ye Tian , Onat Gungor , Tajana Rosing\n- **URL**: [https://arxiv.org/abs/2605.05499](https://arxiv.org/abs/2605.05499)\n- **Abstract**:\n> The widespread adoption of camera-equipped mobile devices and wearables has enabled convenient capture of meal images, making food recognition a key component for real time dietary monitoring. However, real-world food images present challenges due to high intra-class similarity and the frequent presence of multiple food items within a single image. While deep learning models achieve strong performance in coarse grained classification, they often struggle to capture fine-grained attributes such as cooking style. Moreover, open-ended generation in modern vision-language models can produce non-canonical labels, limiting their practical deployment. We propose FoodCHA, a multimodal agentic framework that reformulates food recognition as a hierarchical decision-making process. By progressively anchoring predictions, FoodCHA guides subcategory identification using high-level categories and guides cooking style recognition using subcategories, improving semantic consistency and attribute-level discrimination. To ensure practical deployability, FoodCHA utilizes the compact Moondream-2B vision language model, which provides strong reasoning capability while maintaining lower computational and memory overhead. Experiments on FoodNExTDB show that FoodCHA outperforms Food-Llama-3.2-11B by 13.8% and 38.2% in category and subcategory recognition precision, respectively, and achieves a striking 153.2% improvement in cooking style classification precision.\n\n---\n\n## 70. FinRAG-12B: A Production-Validated Recipe for Grounded Question Answering in Banking\n- **Authors**: Denys Katerenchuk , Pablo Duboue , Keelan Evanini , David Gondek , Nithin Govindugari , Olivier Allauzen , Joshua Baptiste , David J More , Joshua Schechter\n- **URL**: [https://arxiv.org/abs/2605.05482](https://arxiv.org/abs/2605.05482)\n- **Abstract**:\n> Large language models (LLMs) are rapidly being adopted across various domains. However, their adoption in banking industry faces resistance due to demands for high accuracy, regulatory compliance, and the need for verifiable and grounded responses. We present a unified, data-efficient framework for training grounded domain-specific LLMs that optimizes answer quality, citation grounding, and calibrated refusal under real-world deployment constraints. First, we describe a data generation pipeline that combines LLM-as-a-Judge filtering, citation annotation, and curriculum learning with only 143M tokens. The resulting 12B model achieves high answer quality outperforming GPT-4.1 on citation grounding, with a modest citation tradeoff versus the untuned base. Second, we propose a calibrated refusal mechanism: training on 22% unanswerable examples yield a 12% \"I don't know\" rate, substantially improving over the base model's unsafe 4.3% rate while avoiding GPT-4.1's over-refusal (20.2%). Third, we present an end-to-end methodology spanning from data curation to quantized serving. The system is deployed at 40+ financial institutions, achieving a 7.1 percentage point improvement in query resolution (p < 0.001). Additionally, the model delivers 3-5x faster responses at 20-50x lower cost compared to GPT-4.1.\n\n---\n\n## 71. LANTERN: LLM-Augmented Neurosymbolic Transfer with Experience-Gated Reasoning Networks\n- **Authors**: Mahyar Alinejad , Yue Wang , Amrit Singh Bedi , George Atia\n- **URL**: [https://arxiv.org/abs/2605.05478](https://arxiv.org/abs/2605.05478)\n- **Abstract**:\n> Transfer learning in reinforcement learning (RL) seeks to accelerate learning in new tasks by leveraging knowledge from related sources. Existing neurosymbolic transfer methods, however, typically rely on manually specified task automata, assume a single source task, and use fixed knowledge-integration mechanisms that cannot adapt to varying source relevance. We propose LANTERN, a unified framework for multi-source neurosymbolic transfer that addresses these limitations through three components: (i) deterministic finite automata generated from natural language task descriptions using large language models, (ii) semantic embedding-based aggregation of multiple source policies weighted by cross-task similarity, and (iii) adaptive teacher-student gating based on temporal-difference error and semantic uncertainty. Across domains spanning resource management, navigation, and control, LANTERN achieves 40-60% improvements in sample efficiency over existing baselines while remaining robust to poorly aligned sources. These results demonstrate that multi-source, adaptively weighted neurosymbolic transfer can improve scalability and robustness in symbolic RL settings.\n\n---\n\n## 72. Agentic Discovery of Exchange-Correlation Density Functionals\n- **Authors**: Titouan Duston , Jiashu Liang , Yuanheng Wang , Weihao Gao , Xuelan Wen , Nan Sheng , Weiluo Ren , Yang Sun , Yixiao Chen\n- **URL**: [https://arxiv.org/abs/2605.05460](https://arxiv.org/abs/2605.05460)\n- **Abstract**:\n> The development of accurate exchange-correlation (XC) functionals remains a longstanding challenge in density functional theory (DFT). The vast majority of XC functionals have been hand designed by human researchers combining physical insight, exact constraints, and empirical fitting. Recent advances in large language models enable a systematic, automated alternative to this human-driven design loop. This report presents an agentic search system in which an LLM proposes structured functional-form changes guided by evolutionary history. The system attempts to improve functional performance through an iterative plan-execute-summarize loop, where improvements are measurable by optimizing functional parameters against a standard thermochemistry dataset, then evaluating performance on a held-out subset. The strongest discovered functional, SAFS26-a (Seed Agentic Functional Search 2026), improves upon the gold-standard {\\omega}B97M-V baseline by ~9%. These results also surface a cautionary lesson for AI-assisted science: models powerful enough to discover genuine improvements are equally capable of exploiting unphysical shortcuts to game the benchmark; domain expertise translated into explicitly enforced constraints remains essential to keeping results scientifically grounded.\n\n---\n\n## 73. The Geopolitics of AI Safety: A Causal Analysis of Regional LLM Bias\n- **Authors**: Alif Al Hasan\n- **URL**: [https://arxiv.org/abs/2605.05427](https://arxiv.org/abs/2605.05427)\n- **Abstract**:\n> As Large Language Models (LLMs) are integrated into global software systems, ensuring equitable safety guardrails is a critical requirement. Current fairness evaluations predominantly measure bias observationally, a methodology confounded by the inherent toxicity of topics naturally paired with specific demographics in testing datasets. This study introduces a Probabilistic Graphical Model (PGM) framework to audit LLM safety mechanisms causally. By applying Pearl's do-operator, we mathematically isolate the causal effect of injecting a cultural demographic into a prompt. We conduct a large-scale empirical analysis across seven instruction-tuned models spanning diverse origins: the United States (Llama-3.1-8B, Gemma-2-9B), Europe (Mistral-7B-v0.3), the UAE (Falcon3-7B), China (Qwen2.5-7B, DeepSeek-7B), and India (Airavata-7B). Utilizing two distinct datasets (ToxiGen and BOLD), the findings reveal a disparity between observational and interventional bias, demonstrating that standard fairness metrics can overestimate demographic bias by failing to account for context toxicity. Furthermore, the causal probabilities indicate distinct alignment trends: Western models exhibit higher causal refusal rates for specific demographic groups, whereas Eastern models demonstrate low overall intervention rates with targeted sensitivities toward regional demographics. We discuss the implications of these biases, highlighting how demographic-sensitive over-triggering restricts benign discourse in downstream applications.\n\n---\n\n## 74. From History to State: Constant-Context Skill Learning for LLM Agents\n- **Authors**: Haoyang Xie , Xinyuan Wang , Yancheng Wang , Puda Zhao , Feng Ju\n- **URL**: [https://arxiv.org/abs/2605.05413](https://arxiv.org/abs/2605.05413)\n- **Abstract**:\n> Large language model (LLM) agents are increasingly used to operate browsers, files, code and tools, making personal assistants a natural deployment target. Yet personal agents face a privacy-cost-capability tension: cloud models execute multi-step workflows well but expose sensitive intermediate context to external APIs, while local models preserve privacy but remain less reliable. Both settings also pay repeatedly for long skill prompts and growing histories. We propose constant-context skill learning, a context-to-weights framework for recurring agent workflows: reusable procedures are learned in lightweight task-family modules, while inference conditions only on the current observation and a compact state block. A deterministic tracker renders this state block from task progress and supplies aligned subgoal rewards, so each module can be trained with step-level SFT and refined through online RL. Across ALFWorld, WebShop, and SciWorld, our agents achieve strong performance across Qwen3-4B, Qwen3-8B and Llama-3.1-8B. With Qwen3-8B, SFT+RL reaches 89.6\\% unseen success on ALFWorld, 76.8\\% success on WebShop, and 66.4\\% unseen success on SciWorld. They match or exceed strong published agent-training results while reducing prompt tokens per turn by 2--7$\\times$ relative to controlled ReAct prompting baselines, showing that procedural context can be moved from prompts into weights.\n\n---\n\n## 75. LaTA: A Drop-in, FERPA-Compliant Local-LLM Autograder for Upper-Division STEM Coursework\n- **Authors**: Jesse A. Rodríguez\n- **URL**: [https://arxiv.org/abs/2605.05410](https://arxiv.org/abs/2605.05410)\n- **Abstract**:\n> Large-language-model (LLM) graders promise to relieve the grading burden of upper-division STEM courses, but most deployments to date send student work to third-party APIs, violating FERPA and exposing institutions to data risk while requiring substantial assignment modification. We present $\\textbf{LaTA}\\ (\\textit{LaTeX Teaching Assistant})$, a drop-in, open-source autograder that runs entirely on commodity on-premises hardware and assumes a LaTeX-native workflow already adopted by many engineering and physics courses. LaTA implements a four-stage pipeline (ingest, segment, grade, report) using a locally hosted open-weight chain-of-thought LLM grader (gpt-oss:120b) that compares student work to an instructor-authored reference solution and applies a YAML rubric with binary per-item scoring. We deployed LaTA in Winter~2026 in ME 373 (Mechanical Engineering Methods) at Oregon State University, grading every weekly assignment for approximately 200 students on a single Mac Studio at \\$0 marginal cost per assignment and 1--3 minutes of wall-clock time per submission, enabling regrading of corrected assignments and greatly expanded TA office hour offerings. The instructor-confirmed grading-error rate held at roughly $0.02$--$0.04\\%$ per rubric line item across the term. Relative to the same instructor's previous traditionally-graded cohort, the LaTA-graded cohort outperformed by approximately $11\\%$ on the midterm exam and $8\\%$ on the final exam, and reported large gains in self-assessed confidence on every stated learning objective ($N = 159$ survey responses, $\\Delta \\geq +1.49$ Likert points, $p < 10^{-27}$ on every comparison). We release the code under AGPLv3.\n\n---\n\n## 76. Agentic Retrieval-Augmented Generation for Financial Document Question Answering\n- **Authors**: Yang Shu , Yingmin Liu , Zequn Xie\n- **URL**: [https://arxiv.org/abs/2605.05409](https://arxiv.org/abs/2605.05409)\n- **Abstract**:\n> Financial document question answering (QA) demands complex multi-step numerical reasoning over heterogeneous evidence--structured tables, textual narratives, and footnotes--scattered across corporate filings. Existing retrieval-augmented generation (RAG) approaches adopt a single-pass retrieve-then-generate paradigm that struggles with the compositional reasoning chains prevalent in financial analysis. We propose FinAgent-RAG, an agentic RAG framework that orchestrates iterative retrieval-reasoning loops with self-verification, specifically engineered for the precision requirements of financial numerical reasoning. The framework integrates three domain-specific innovations: (1) a Contrastive Financial Retriever trained with hard negative mining to distinguish semantically similar but numerically distinct financial passages, (2) a Program-of-Thought reasoning module that generates executable Python code for precise arithmetic rather than relying on error-prone LLM-based mental computation, and (3) an Adaptive Strategy Router that dynamically allocates computational resources based on question complexity, reducing API costs by 41.3% on FinQA while preserving accuracy. Extensive experiments on three benchmark datasets--FinQA, ConvFinQA, and TAT-QA--demonstrate that FinAgent-RAG achieves 76.81%, 78.46%, and 74.96% execution accuracy respectively, outperforming the strongest baseline by 5.62--9.32 percentage points. Ablation studies, cross-backbone evaluation with four LLMs, and deployment cost analysis confirm the framework's robustness and practical viability for financial institutions.\n\n---\n\n## 77. PRISM: Perception Reasoning Interleaved for Sequential Decision Making\n- **Authors**: Mohamed Salim Aissi , Clemence Grislain , Clement Romac , Laure Soulier , Mohamed Chetouani , Olivier Sigaud , Nicolas Thome\n- **URL**: [https://arxiv.org/abs/2605.05407](https://arxiv.org/abs/2605.05407)\n- **Abstract**:\n> Scaling LLM-based embodied agents from text-only environments to complex multimodal settings remains a major challenge. Recent work identifies a perception-reasoning-decision gap in standalone Vision-Language Models (VLMs), which often overlook task-critical information. In this paper, we introduce PRISM, a framework that tightly couples perception (VLM) and decision (LLM) through a dynamic question-answer (DQA) pipeline. Instead of passively accepting the VLM's description, the LLM critiques it, probes the VLM with goal-oriented questions, and synthesizes a compact image description. This closed-loop interaction yields a sharp, task-driven understanding of the scene. We evaluate PRISM on the ALFWorld and Room-to-Room (R2R) benchmarks. We show that: (1) PRISM significantly outperforms state-of-the-art image-based models, (2) our Interactive goal-oriented perception pipeline yields systematic and substantial gains, and (3) PRISM is fully automatic, eliminating the need for handcrafted questions or answers.\n\n---\n\n## 78. When Helpfulness Becomes Sycophancy: Sycophancy is a Boundary Failure Between Social Alignment and Epistemic Integrity in Large Language Models\n- **Authors**: Jiechen Li , Catherine A. Barry , Rishika Randev , Janet Chen , Ella Jorgensen , Brinnae Bent\n- **URL**: [https://arxiv.org/abs/2605.05403](https://arxiv.org/abs/2605.05403)\n- **Abstract**:\n> This position paper argues that sycophancy in LLMs is a boundary failure between social alignment and epistemic integrity. Existing work often operationalizes sycophancy through external behavior such as agreement with incorrect user beliefs, position reversals, or deviation from an objective standard of correctness. These formulations capture only overt forms of the phenomenon and leave subtler boundary failures involving epistemic integrity and social alignment underspecified. We argue that sycophancy should not be understood as agreement alone, but as alignment behavior that displaces independent epistemic judgment. To clarify this boundary, we propose a three-condition framework for sycophancy. First, the user expresses a cue in the form of a belief, preference, or self-concept. Second, the model shifts toward that cue through alignment behavior. Third, this shift compromises epistemic accuracy, independent reasoning, or appropriate correction. We also introduce a taxonomy for classifying sycophancy, consisting of alignment targets, mechanisms, and severity. The paper concludes by discussing implications for alignment evaluation and argues for boundary-aware assessment, structured rubrics, and mitigation strategies, while situating these proposals alongside alternative views of sycophancy.\n\n---\n\n## 79. BALAR : A Bayesian Agentic Loop for Active Reasoning\n- **Authors**: Aymen Echarghaoui , Dongxia Wu , Emily B. Fox\n- **URL**: [https://arxiv.org/abs/2605.05386](https://arxiv.org/abs/2605.05386)\n- **Abstract**:\n> Large language models increasingly operate in interactive settings where solving a task requires multiple rounds of information exchange with a user. However, most current systems treat dialogue reactively and lack a principled mechanism to reason about what information is missing and which question should be asked next. We propose BALAR (Bayesian Agentic Loop for Active Reasoning), a task-agnostic outer-loop algorithm that requires no fine-tuning and enables structured multi-turn interaction between an LLM agent and a user. BALAR maintains a structured belief over latent states, selects clarifying questions by maximizing expected mutual information, and dynamically expands its state representation when the current one proves insufficient. We evaluate BALAR on three diverse benchmarks: AR-Bench-DC (detective cases), AR-Bench-SP (thinking puzzles), and iCraft-MD (clinical diagnosis). BALAR significantly outperforms all baselines across all three benchmarks, with $14.6\\%$ higher accuracy on AR-Bench-DC, $38.5\\%$ on AR-Bench-SP, and $30.5\\%$ on iCraft-MD.\n\n---\n\n## 80. Understanding Annotator Safety Policy with Interpretability\n- **Authors**: Alex Oesterling , Donghao Ren , Yannick Assogba , Dominik Moritz , Sunnie S.Y. Kim , Leon Gatys , Fred Hohman\n- **URL**: [https://arxiv.org/abs/2605.05329](https://arxiv.org/abs/2605.05329)\n- **Abstract**:\n> Safety policies define what constitutes safe and unsafe AI outputs, guiding data annotation and model development. However, annotation disagreement is pervasive and can stem from multiple sources such as operational failures (annotators misunderstand or misexecute the task), policy ambiguity (policy wording leaves room for interpretation), or value pluralism (different annotators hold different perspectives on safety). Distinguishing these sources matters. For example, operational failures call for quality control, ambiguity calls for policy clarification, and pluralism calls for deliberation about incorporating diverse perspectives. Yet understanding why annotators disagree is difficult. Directly asking annotators for their reasoning is costly, substantially increasing annotation burden, and can be unreliable for both human and LLM annotators as self-reported reasoning often fails to reflect actual decision processes. We introduce Annotator Policy Models (APMs), interpretable models that learn annotators' internal safety policies from labeling behavior alone, making annotator reasoning visible and comparable without additional annotation effort. We validate that APMs accurately model annotator safety policy (>80% accuracy), faithfully predict responses to counterfactual edits, and recover known policy differences in controlled settings. Applying APMs to LLM and human annotations, we demonstrate two core applications: (1) surfacing policy ambiguity by revealing how annotators interpret safety instructions differently, and (2) surfacing value pluralism by uncovering systematic differences in safety priorities across demographic groups. Together, these capabilities support more targeted, transparent, and inclusive safety policy design.\n\n---\n\n## 81. Verifier-Backed Hard Problem Generation for Mathematical Reasoning\n- **Authors**: Yuhang Lai , Jiazhan Feng , Yee Whye Teh , Ning Miao\n- **URL**: [https://arxiv.org/abs/2605.06660](https://arxiv.org/abs/2605.06660)\n- **Abstract**:\n> Large Language Models (LLMs) demonstrate strong capabilities for solving scientific and mathematical problems, yet they struggle to produce valid, challenging, and novel problems - an essential component for advancing LLM training and enabling autonomous scientific research. Existing problem generation approaches either depend on expensive human expert involvement or adopt naive self-play paradigms, which frequently yield invalid problems due to reward hacking. This work introduces VHG, a verifier-enhanced hard problem generation framework built upon three-party self-play. By integrating an independent verifier into the conventional setter-solver duality, our design constrains the setter's reward to be jointly determined by problem validity (evaluated by the verifier) and difficulty (assessed by the solver). We instantiate two verifier variants: a Hard symbolic verifier and a Soft LLM-based verifier, with evaluations conducted on indefinite integral tasks and general mathematical reasoning tasks. Experimental results show that VHG substantially outperforms all baseline methods by a clear margin.\n\n---\n\n## 82. Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less\n- **Authors**: Yuxing Liu , Jianyu Wang , Tong Zhang\n- **URL**: [https://arxiv.org/abs/2605.06654](https://arxiv.org/abs/2605.06654)\n- **Abstract**:\n> Optimizers play an important role in both pretraining and finetuning stages when training large language models (LLMs). In this paper, we present an observation that full finetuning with the same optimizer as in pretraining achieves a better learning-forgetting tradeoff, i.e., forgetting less while achieving the same or better performance on the new task, than other optimizers and, possibly surprisingly, LoRA, during the supervised finetuning (SFT) stage. We term this phenomenon optimizer-model consistency. To better understand it, through controlled experiments and theoretical analysis, we show that: 1) optimizers can shape the models by having regularization effects on the activations, leading to different landscapes around the pretrained checkpoints; 2) in response to this regularization effect, the weight update in SFT should follow some specific structures to lower forgetting of the knowledge learned in pretraining, which can be obtained by using the same optimizer. Moreover, we specifically compare Muon and AdamW when they are employed throughout the pretraining and SFT stages and find that Muon performs worse when finetuned for reasoning tasks. With a synthetic language modeling experiment, we demonstrate that this can come from Muon's strong tendency towards rote memorization, which may hurt pattern acquisition with a small amount of data, as for SFT.\n\n---\n\n## 83. When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels\n- **Authors**: Sushant Gautam , Finn Schwall , Annika Willoch Olstad , Fernando Vallecillos Ruiz , Birk Torpmann-Hagen , Sunniva Maria Stordal Bjørklund , Leon Moonen , Klas Pettersen , Michael A. Riegler\n- **URL**: [https://arxiv.org/abs/2605.06652](https://arxiv.org/abs/2605.06652)\n- **Abstract**:\n> Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component ($\\eta^2 \\approx 0.52$), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.\n\n---\n\n## 84. Superintelligent Retrieval Agent: The Next Frontier of Information Retrieval\n- **Authors**: Zeyu Yang , Qi Ma , Jason Chen , Anshumali Shrivastava\n- **URL**: [https://arxiv.org/abs/2605.06647](https://arxiv.org/abs/2605.06647)\n- **Abstract**:\n> Retrieval-augmented agents are increasingly the interface to large organizational knowledge bases, yet most still treat retrieval as a black box: they issue exploratory queries, inspect returned snippets, and iteratively reformulate until useful evidence emerges. This approach resembles how a newcomer searches an unfamiliar database rather than how an expert navigates it with strong priors about terminology and likely evidence, and results in unnecessary retrieval rounds, increased latency, and poor recall. We introduce \\textit{SuperIntelligent Retrieval Agent} (SIRA), which defines \\emph{superintelligence} in retrieval as the ability to compress multi-round exploratory search into a single corpus-discriminative retrieval action. SIRA does not merely ask what terms are relevant to the query; it asks which terms are likely to separate the desired evidence from corpus-level confusers. On the corpus side, an LLM enriches each document offline with missing search vocabulary; on the query side, it predicts evidence vocabulary omitted by the query; and document-frequency statistics as a tool call to filter proposed terms that are absent, overly common, or unlikely to create retrieval margin. The final retrieval step is a single weighted BM25 call combining the original query with the validated expansion. Across ten BEIR benchmarks and downstream question-answering tasks, SIRA achieves the significantly superior performance outperforming dense retrievers and state-of-the-art multi-round agentic baselines, demonstrating that one well-formed lexical query, guided by LLM cognition and lightweight corpus statistics, can exceed substantially more expensive multi-round search while remaining interpretable, training-free, and efficient.\n\n---\n\n## 85. StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction\n- **Authors**: Xiangyuan Xue , Yifan Zhou , Zidong Wang , Shengji Tang , Philip Torr , Wanli Ouyang , Lei Bai , Zhenfei Yin\n- **URL**: [https://arxiv.org/abs/2605.06642](https://arxiv.org/abs/2605.06642)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used as interactive agents, but optimizing them for long-horizon decision making remains difficult because current methods are largely purely reactive, which weakens both exploration and credit assignment over extended trajectories. In this work, we present Strategic Trajectory Abstraction (StraTA), a simple framework that introduces an explicit trajectory-level strategy into agentic reinforcement learning (RL). StraTA samples a compact strategy from the initial task state, conditions subsequent actions on that strategy, and trains strategy generation and action execution jointly with a hierarchical GRPO-style rollout design, further enhanced by diverse strategy rollout and critical self-judgment. Experiments on ALFWorld, WebShop, and SciWorld show that StraTA consistently improves both sample efficiency and final performance over strong baselines. StraTA reaches success rates of 93.1% on ALFWorld and 84.2% on WebShop. On SciWorld, StraTA attains a 63.5% overall score, outperforming frontier closed-source models.\n\n---\n\n## 86. The Structural Origin of Attention Sink: Variance Discrepancy, Super Neurons, and Dimension Disparity\n- **Authors**: Siquan Li , Kaiqi Jiang , Jiacheng Sun , Tianyang Hu\n- **URL**: [https://arxiv.org/abs/2605.06611](https://arxiv.org/abs/2605.06611)\n- **Abstract**:\n> Despite the prevalence of the attention sink phenomenon in Large Language Models (LLMs), where initial tokens disproportionately monopolize attention scores, its structural origins remain elusive. This work provides a \\textit{mechanistic explanation} for this phenomenon. First, we trace its root to the value aggregation process inherent in self-attention, which induces a systematic variance discrepancy. We further demonstrate that this discrepancy is drastically amplified by the activation of super neurons within Feed-Forward Network (FFN) layers. Specifically, the channel-sparse down-projections trigger a dimension disparity of the first-token representation, necessitating the formation of attention sinks as a structural anchor. Then, we validate this causal chain through two controlled interventions: (i) isolating the aggregation effect via attention mask modifications and (ii) amplifying the variance of targeted token representations. Both interventions can replicate attention sinks at arbitrary positions. Our mechanistic understanding offers a foundation for the systematic control of sink formation. Finally, as a proof of concept, we propose \\textit{head-wise RMSNorm}, an architectural modification that stabilizes value aggregation outputs during pre-training. Our experiments demonstrate that restoring statistical parity across positions significantly accelerates convergence.\n\n---\n\n## 87. AI CFD Scientist: Toward Open-Ended Computational Fluid Dynamics Discovery with Physics-Aware AI Agents\n- **Authors**: Nithin Somasekharan , Rabi Pathak , Manushri Dhanakoti , Tingwen Zhang , Ling Yue , Andy Zhu , Shaowu Pan\n- **URL**: [https://arxiv.org/abs/2605.06607](https://arxiv.org/abs/2605.06607)\n- **Abstract**:\n> Recent LLM-based agents have closed substantial portions of the scientific discovery loop in software-only machine-learning research, in chemistry, and in biology. Extending the same loop to high-fidelity physical simulators is harder, because solver completion does not imply physical validity and many failure modes appear only in field-level imagery rather than in solver logs. We present AI CFD Scientist, an open-source AI scientist for computational fluid dynamics (CFD) that, to our knowledge, is the first to span literature-grounded ideation, validated execution, vision-based physics verification, source-code modification, and figure-grounded writing within a single inspectable workflow. Three coupled pathways cover parameter sweeps within a fixed solver, case-local C++ library compilation for new physical models, and open-ended hypothesis search against a reference comparator, all running on OpenFOAM through Foam-Agent. At the center of the framework is a vision-language physics-verification gate that inspects rendered flow fields before any result is accepted, rerun, or written into a manuscript. On five tasks under a shared GPT-5.5 backbone, AI CFD Scientist autonomously discovers a Spalart-Allmaras runtime correction that reduces lower-wall Cf RMSE against DNS by 7.89% on the periodic hill at Reh=5600; under matched LLM cost, two strong general AI-scientist baselines (ARIS, DeepScientist) execute partial CFD workflows but lack the domain-specific validity gates needed to convert runs into defensible scientific claims; and a controlled planted-failure ablation shows that the vision-language gate detects 14 of 16 silent failures missed by solver-level checks. Code, prompts, and run artifacts are released at this https URL .\n\n---\n\n## 88. UniSD: Towards a Unified Self-Distillation Framework for Large Language Models\n- **Authors**: Yiqiao Jin , Yiyang Wang , Lucheng Fu , Yijia Xiao , Yinyi Luo , Haoxin Liu , B. Aditya Prakash , Josiah Hester , Jindong Wang , Srijan Kumar\n- **URL**: [https://arxiv.org/abs/2605.06597](https://arxiv.org/abs/2605.06597)\n- **Abstract**:\n> Self-distillation (SD) offers a promising path for adapting large language models (LLMs) without relying on stronger external teachers. However, SD in autoregressive LLMs remains challenging because self-generated trajectories are free-form, correctness is task-dependent, and plausible rationales can still provide unstable or unreliable supervision. Existing methods mainly examine isolated design choices, leaving their effectiveness, roles, and interactions unclear. In this paper, we propose UniSD, a unified framework to systematically study self-distillation. UniSD integrates complementary mechanisms that address supervision reliability, representation alignment, and training stability, including multi-teacher agreement, EMA teacher stabilization, token-level contrastive learning, feature matching, and divergence clipping. Across six benchmarks and six models from three model families, UniSD reveals when self-distillation improves over static imitation, which components drive the gains, and how these components interact across tasks. Guided by these insights, we construct UniSDfull, an integrated pipeline that combines complementary components and achieves the strongest overall performance, improving over the base model by +5.4 points and the strongest baseline by +2.8 points. Extensive evaluation highlights self-distillation as a practical and steerable approach for efficient LLM adaptation without stronger external teachers.\n\n---\n\n## 89. Continuous Latent Diffusion Language Model\n- **Authors**: Hongcan Guo , Qinyu Zhao , Yian Zhao , Shen Nie , Rui Zhu , Qiushan Guo , Feng Wang , Tao Yang , Hengshuang Zhao , Guoqiang Wei , Yan Zeng\n- **URL**: [https://arxiv.org/abs/2605.06548](https://arxiv.org/abs/2605.06548)\n- **Abstract**:\n> Large language models have achieved remarkable success under the autoregressive paradigm, yet high-quality text generation need not be tied to a fixed left-to-right order. Existing alternatives still struggle to jointly achieve generation efficiency, scalable representation learning, and effective global semantic modeling. We propose Cola DLM, a hierarchical latent diffusion language model that frames text generation through hierarchical information decomposition. Cola DLM first learns a stable text-to-latent mapping with a Text VAE, then models a global semantic prior in continuous latent space with a block-causal DiT, and finally generates text through conditional decoding. From a unified Markov-path perspective, its diffusion process performs latent prior transport rather than token-level observation recovery, thereby separating global semantic organization from local textual realization. This design yields a more flexible non-autoregressive inductive bias, supports semantic compression and prior fitting in continuous space, and naturally extends to other continuous modalities. Through experiments spanning 4 research questions, 8 benchmarks, strictly matched ~2B-parameter autoregressive and LLaDA baselines, and scaling curves up to about 2000 EFLOPs, we identify an effective overall configuration of Cola DLM and verify its strong scaling behavior for text generation. Taken together, the results establish hierarchical continuous latent prior modeling as a principled alternative to strictly token-level language modeling, where generation quality and scaling behavior may better reflect model capability than likelihood, while also suggesting a concrete path toward unified modeling across discrete text and continuous modalities.\n\n---\n\n## 90. Is One Layer Enough? Understanding Inference Dynamics in Tabular Foundation Models\n- **Authors**: Amir Rezaei Balef , Mykhailo Koshil , Katharina Eggensperger\n- **URL**: [https://arxiv.org/abs/2605.06510](https://arxiv.org/abs/2605.06510)\n- **Abstract**:\n> Transformer-based tabular foundation models (TFMs) dominate small to medium tabular predictive benchmark tasks, yet their inference mechanisms remain largely unexplored. We present the first large-scale mechanistic study of layerwise dynamics in 6 state-of-the-art tabular in-context learning models. We explore how predictions emerge across depth, identify distinct stages of inference and reveal latent-space dynamics that differ from those of language models. Our findings indicate substantial depthwise redundancy across multiple models, suggesting iterative refinement with overlapping computations during inference stages. Guided by these insights, we design a proof-of-concept, looped single-layer model that uses only 20% of the original model's parameters while achieving comparable performance. The code is available at this https URL .\n\n---\n\n## 91. PACZero: PAC-Private Fine-Tuning of Language Models via Sign Quantization\n- **Authors**: Murat Bilgehan Ertan , Xiaochen Zhu , Phuong Ha Nguyen , Marten van Dijk , Srinivas Devadas\n- **URL**: [https://arxiv.org/abs/2605.06505](https://arxiv.org/abs/2605.06505)\n- **Abstract**:\n> We introduce PACZero, a family of PAC-private zeroth-order mechanisms for fine-tuning large language models that delivers usable utility at $I(S^*; Y_{1:T})=0$. This privacy regime bounds the membership-inference attack (MIA) posterior success rate at the prior, an MIA-resistance level the DP framework matches only at $\\varepsilon=0$ and infinite noise. All DP-ZO comparisons below are matched at the MIA posterior level. The key insight is that PAC Privacy charges mutual information only when the release depends on which candidate subset is the secret. Sign-quantizing subset-aggregated zeroth-order gradients creates frequent unanimity, steps at which every candidate subset agrees on the update direction; at these steps the released sign costs zero conditional mutual information. We propose two variants that span the privacy-utility trade-off: PACZero-MI (budgeted MI via exact calibration on the binary release) and PACZero-ZPL ($I=0$ via a uniform coin flip on disagreement steps). We evaluate on SST-2 and SQuAD with OPT-1.3B and OPT-6.7B in both LoRA and full-parameter tracks. On SST-2 OPT-1.3B full fine-tuning at $I=0$, PACZero-ZPL reaches ${88.99\\pm0.91}$, within $2.1$pp of the non-private MeZO baseline ($91.1$ FT). No prior method produces usable utility in the high-privacy regime $\\varepsilon<1$, and PACZero-ZPL obtains competitive SST-2 accuracy and nontrivial SQuAD F1 across OPT-1.3B and OPT-6.7B at $I=0$.\n\n---\n\n## 92. Litespark Inference on Consumer CPUs: Custom SIMD Kernels for Ternary Neural Networks\n- **Authors**: Nii Osae Osae Dade , Tony Morri , Moinul Hossain Rahat , Sayandip Pal\n- **URL**: [https://arxiv.org/abs/2605.06485](https://arxiv.org/abs/2605.06485)\n- **Abstract**:\n> Large language models (LLMs) have transformed artificial intelligence, but their computational requirements remain prohibitive for most users. Standard inference demands expensive datacenter GPUs or cloud API access, leaving over one billion personal computers underutilized for AI workloads. Ternary models offer a path forward: their weights are constrained to {-1, 0, +1}, theoretically eliminating the need for floating-point multiplication. However, existing frameworks fail to exploit this structure, treating ternary models as dense floating-point networks. We address this gap with custom SIMD kernels that replace matrix multiplication with simple addition and subtraction operations, targeting the integer dot product instructions available on modern CPUs. Our implementation, Litespark-Inference, is pip-installable and integrates directly with Hugging-Face, achieving 9.2x faster time-to-first-token, 52x higher throughput, and 14x memory reduction compared to standard PyTorch inference on Apple Silicon, with similar speedups on Intel and AMD processors.\n\n---\n\n## 93. Constraint Decay: The Fragility of LLM Agents in Backend Code Generation\n- **Authors**: Francesco Dente , Dario Satriani , Paolo Papotti\n- **URL**: [https://arxiv.org/abs/2605.06445](https://arxiv.org/abs/2605.06445)\n- **Abstract**:\n> Large Language Model (LLM) agents demonstrate strong performance in autonomous code generation under loose specifications. However, production-grade software requires strict adherence to structural constraints, such as architectural patterns, databases, and object-relational mappings. Existing benchmarks often overlook these non-functional requirements, rewarding functionally correct but structurally arbitrary solutions. We present a systematic study evaluating how well agents handle structural constraints in multi-file backend generation. By fixing a unified API contract across 80 greenfield generation tasks and 20 feature-implementation tasks spanning eight web frameworks, we isolate the effect of structural complexity using a dual evaluation with end-to-end behavioral tests and static verifiers. Our findings reveal a phenomenon of constraint decay: as structural requirements accumulate, agent performance exhibits a substantial decline. Capable configurations lose 30 points on average in assertion pass rates from baseline to fully specified tasks, while some weaker configurations approach zero. Framework sensitivity analysis exposes significant performance disparities: agents succeed in minimal, explicit frameworks (e.g., Flask) but perform substantially worse on average in convention-heavy environments (e.g., FastAPI, Django). Finally, error analysis identifies data-layer defects (e.g., incorrect query composition and ORM runtime violations) as the leading root causes. This work highlights that jointly satisfying functional and structural requirements remains a key open challenge for coding agents.\n\n---\n\n## 94. Is Escalation Worth It? A Decision-Theoretic Characterization of LLM Cascades\n- **Authors**: Dylan Bouchard\n- **URL**: [https://arxiv.org/abs/2605.06350](https://arxiv.org/abs/2605.06350)\n- **Abstract**:\n> Model cascades, in which a cheap LLM defers to an expensive one on low-confidence queries, are widely used to navigate the cost-quality tradeoff at deployment. Existing approaches largely treat the deferral threshold as an empirical hyperparameter, with limited guidance on the geometry of the resulting cost-quality frontier over a model pool. We develop a decision-theoretic framework grounded in constrained optimization and duality. For a two-model cascade, we establish piecewise concavity of the cost-quality frontier on decreasing-benefit regions of the confidence support, with reciprocal shadow prices linking the budget- and quality-constrained formulations. Given a pool of $k$ models, we characterize the frontier achievable by deterministic two-model threshold cascades as the pointwise envelope over $\\binom{k}{2}$ pairwise cascades, with switching points where the optimal pair changes. For $k$-model cascades, we derive first-order conditions in which a single shadow price equalizes marginal quality-per-cost across stage boundaries. We validate the framework on five benchmarks (MATH, MMLU, TriviaQA, SimpleQA, LiveCodeBench) across eight models from five providers. Within the deterministic threshold-cascade class, full fixed chains underperform the pairwise envelope, and optimized subsequence cascades do not deliver practically meaningful held-out gains over it. A lightweight pre-generation router exceeds the best cascade policy on four of five datasets, mainly because it avoids the cheap model's generation cost on queries sent directly to a larger model rather than because of a stronger routing signal. These results suggest that cascade performance is limited primarily by structural cost, since cascades pay the cheap model before any escalation decision, rather than by a shortage of intermediate stages.\n\n---\n\n## 95. Human-AI Co-Evolution and Epistemic Collapse: A Dynamical Systems Perspective\n- **Authors**: Xuening Wu , Yanlan Kang , Qianya Xu , Kexuan Xie , Jiaqi Mi , Honggang Wang , Yubin Liu , Zeping Chen\n- **URL**: [https://arxiv.org/abs/2605.06347](https://arxiv.org/abs/2605.06347)\n- **Abstract**:\n> Large language models (LLMs) are reshaping how knowledge is produced, with increasing reliance on AI systems for generation, summarization, and reasoning. While prior work has studied cognitive offloading in humans and model collapse in recursive training, these effects are typically considered in isolation. We propose a unified perspective: humans and language models form a coupled dynamical system linked by a feedback loop of usage, generation, and retraining. We introduce a minimal model with three variables -- human cognition, data quality, and model capability -- and show that this feedback can give rise to distinct dynamical regimes. Our analysis identifies three regimes: co-evolutionary enhancement, fragile equilibrium, and degenerative convergence. Through a simple simulation, we demonstrate that increasing reliance on AI can induce a transition toward a low-diversity, suboptimal equilibrium. From an information-theoretic perspective, this transition corresponds to an emergent information bottleneck in the human-AI loop, where entropy reduction reflects loss of diversity and support under closed-loop feedback rather than beneficial compression. These results suggest that the trajectory of AI systems is shaped not only by model design, but by the dynamics of human-AI co-evolution.\n\n---\n\n## 96. CoupleEvo: Evolving Heuristics for Coupled Optimization Problems Using Large Language Models\n- **Authors**: Thomas Bömer , Bastian Amberg , Max Disselnmeyer , Anne Meyer\n- **URL**: [https://arxiv.org/abs/2605.06341](https://arxiv.org/abs/2605.06341)\n- **Abstract**:\n> Many real-world optimization problems consist of multiple tightly coupled subproblems whose solutions must be coordinated to achieve high overall performance. However, existing large language model driven automated heuristic design approaches are limited to single-problem settings. In this paper, we propose CoupleEvo. CoupleEvo proposes three evolutionary coordination strategies to evolve heuristics for coupled optimization problems: the sequential strategy evolves heuristics for one subproblem after the other; the iterative strategy alternates the evolution of heuristics for different subproblems over successive generations; and the integrated strategy evolves heuristics for all problems simultaneously. The approach is evaluated on two representative coupled optimization problems. Experimental results show that decomposition-based strategies (sequential and iterative) provide more stable convergence and higher solution quality, while the integrated evolution strategy suffers from increased search complexity and variability. These findings highlight the importance of coordinating evolutionary search across interdependent subproblems and demonstrate the potential of LLM-driven heuristic design for complex coupled optimization problems. The code is available: this https URL .\n\n---\n\n## 97. Fine-Tuning Small Language Models for Solution-Oriented Windows Event Log Analysis\n- **Authors**: Siraaj Akhtar , Saad Khan , Simon Parkinson\n- **URL**: [https://arxiv.org/abs/2605.06330](https://arxiv.org/abs/2605.06330)\n- **Abstract**:\n> Large language models (LLMs) have shown promise for event log analysis, but their high computational requirements, reliance on cloud infrastructure, and security concerns limit practical deployment. In addition, most existing approaches focus only on the identification of the problem and do not provide actionable remediation. Small language models (SLMs) present a light-weight alternative that can be fine-tuned for a specific purpose and hosted locally. This paper investigates whether SLMs, when fine-tuned for a specific task, can serve as a practical alternative for event log analysis while also generating solutions. We first create a large-scale synthetic Windows event log dataset that contains remediation actions using a high-performing LLM. We then fine-tune multiple SLMs and LLMs using the LoRA parameter-efficient fine-tuning technique and evaluate their performance by comparing with expert assessment. The results show that the dataset accurately reflects real-world scenarios and that fine-tuned SLMs consistently outperform LLMs in identifying issues and providing relevant remediation, while requiring fewer computational resources.\n\n---\n\n## 98. Measuring Evaluation-Context Divergence in Open-Weight LLMs: A Paired-Prompt Protocol with Pilot Evidence of Alignment-Pipeline-Specific Heterogeneity\n- **Authors**: Florian A. D. Burnat , Brittany I. Davidson\n- **URL**: [https://arxiv.org/abs/2605.06327](https://arxiv.org/abs/2605.06327)\n- **Abstract**:\n> Safety benchmarks are routinely treated as evidence about how a language model will behave once deployed, but this inference is fragile if behavior depends on whether a prompt looks like an evaluation. We define evaluation-context divergence as an observable within-item change in behavior induced by framing a fixed task as an evaluation, a live deployment interaction, or a neutral request, and present a paired-prompt protocol that measures it in open-weight LLMs while controlling for paraphrase variation, benchmark familiarity, and judge framing-sensitivity. Across five instruction-tuned checkpoints from four open-weight families plus a matched OLMo-3 base/instruct ablation ($20$ paired items, $840$ generations per checkpoint), we find striking heterogeneity. OLMo-3-Instruct alone is eval-cautious -- evaluation framing raises refusal vs. neutral by $11.8$pp ($p=0.007$) and reduces harmful compliance vs. deployment by $3.6$pp ($p=0.024$, $0/20$ items inverted) -- while Mistral-Small-3.2, Phi-3.5-mini, and Llama-3.1-8B are deployment-cautious}, with marginal eval-vs-deployment refusal effects of $-9$ to $-20$pp. The matched OLMo-3 base also exhibits the deployment-cautious pattern, identifying alignment as the inversion stage; within Llama-3.1, the $70$B model preserves direction with attenuated magnitude, ruling out a simple ``small-model effect that reverses at scale.'' One caveat: the cross-family heterogeneity is judge-dependent. Re-judging with a different-family safety classifier (Llama-Guard-3-8B) preserves the within-OLMo eval-cautious direction but flattens the cross-family contrast, indicating that the two judges operationalize distinct constructs.\n\n---\n\n## 99. Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs\n- **Authors**: Elizabeth Mieczkowski , Alexander Ku , Tiwalayo Eisape , Dilip Arumugam , John Matters , Katherine M. Collins , Ilia Sucholutsky , Thomas L. Griffiths\n- **URL**: [https://arxiv.org/abs/2605.06320](https://arxiv.org/abs/2605.06320)\n- **Abstract**:\n> Large language models (LLMs) are increasingly deployed in teams, yet existing coordination approaches often occupy two extremes. Highly structured methods rely on fixed roles, pipelines, or task decompositions assigned a priori. In contrast, fully unstructured teams enable adaptability and exploration but suffer from inefficiencies such as error propagation, inter-agent conflicts, and wasted resources (measured in time, tokens, or file operations). We introduce Language Agent Teams for Task Evolution (LATTE), a framework for coordinating LLM teams inspired by distributed systems, where processors must operate under partial observability and communication constraints. In LATTE, a team of agents collaboratively construct and maintain a shared, evolving coordination graph which encodes sub-task dependencies, individual agent assignment, and the current state of sub-task progress. This protocol maintains consistency while empowering agents to dynamically allocate work, adapt coordination, and discover new tasks. Across multiple collaborative tasks and a variety of base models, we demonstrate how LATTE reduces token usage, wall-clock time, communication, and coordination failures (e.g. file conflicts and redundant outputs) while matching or exceeding the accuracy of standard designs including MetaGPT, decentralized teams, top-down Leader-Worker hierarchies, and static decompositions.\n\n---\n\n## 100. Pro-KLShampoo: Projected KL-Shampoo with Whitening Recovered by Orthogonalization\n- **Authors**: Ruotong Sun , Ermin Wei\n- **URL**: [https://arxiv.org/abs/2605.06316](https://arxiv.org/abs/2605.06316)\n- **Abstract**:\n> Optimizers that exploit the matrix structure of gradients are central to modern LLM pre-training, with two distinct frontiers: explicit Kronecker-factored preconditioning -- most recently KL-Shampoo, which estimates the preconditioner via KL divergence minimization -- and orthogonalization of the gradient momentum, exemplified by Muon and analyzed as steepest descent under the spectral norm. The two routes are typically developed in isolation. We make a structural observation about KL-Shampoo's Kronecker preconditioners: their eigenvalue spectra exhibit a \\emph{spike-and-flat} shape -- a few dominant eigenvalues followed by an approximately uniform tail -- across layers and training stages, holding exactly under a rank-$\\rho$ signal-plus-noise gradient model. We exploit this structure by restricting one of KL-Shampoo's Kronecker factors to a parametric family aligned with the spike-and-flat shape: full spectral structure on a tracked $r$-dimensional subspace, single shared eigenvalue across the remaining $n-r$ directions. On these directions, we apply orthogonalization. An identity shows that this orthogonalization recovers the algebraic form of full KL-Shampoo's preconditioner. On four pre-training scales (GPT-2 124M / 350M, LLaMA 134M / 450M), Pro-KLShampoo consistently outperforms KL-Shampoo at every subspace rank we test in validation loss, peak per-GPU memory, and wallclock time to reach each loss level.\n\n---\n\n## 101. Attributions All the Way Down? The Metagame of Interpretability\n- **Authors**: Hubert Baniecki , Przemyslaw Biecek , Fabian Fumagalli\n- **URL**: [https://arxiv.org/abs/2605.06295](https://arxiv.org/abs/2605.06295)\n- **Abstract**:\n> We introduce the metagame, a conceptual framework for quantifying second-order interaction effects of model explanations. For any first-order attribution $\\phi(f)$ explaining a model $f$, we measure the directional influence of feature $j$ on the attribution of feature $i$, denoted as meta-attribution $\\varphi_{j \\to i}(f)$, by treating the attribution method itself as a cooperative game and computing its Shapley value. Theoretically, we prove that attributions hierarchically decompose into meta-attributions, and establish these as directional extensions of existing interaction indices. Empirically, we demonstrate that the metagame delivers insights across diverse interpretability applications: (i) quantifying token interactions in instruction-tuned language models, (ii) explaining cross-modal similarity in vision-language encoders, and (iii) interpreting text-to-image concepts in multimodal diffusion transformers.\n\n---\n\n## 102. Log-Likelihood, Simpson's Paradox, and the Detection of Machine-Generated Text\n- **Authors**: Tom Kempton , Viktor Drobnyi , Maeve Madigan , Stuart Burrell\n- **URL**: [https://arxiv.org/abs/2605.06294](https://arxiv.org/abs/2605.06294)\n- **Abstract**:\n> The ability to reliably distinguish human-written text from that generated by large language models is of profound societal importance. The dominant approach to this problem exploits the likelihood hypothesis: that machine-generated text should appear more probable to a detector language model than human-written text. However, we demonstrate that the token-level signal distinguishing human and machine text is non-uniform across the hidden space of the detector model, and naively averaging likelihood-based token scores across regions with fundamentally different statistical structure, as most detectors do, causes a form of Simpson's paradox: a strong local signal is destroyed by inappropriate aggregation. To correct for this, we introduce a learned local calibration step grounded in Bayesian decision theory. Rather than aggregating raw token scores, we first learn lightweight predictors of the score distributions conditioned on position in hidden space, and aggregate calibrated log-likelihood ratios instead. This single intervention dramatically and consistently improves detection performance across all baseline detectors and all datasets we consider. For example, our calibrated variant of Fast-DetectGPT improves AUROC from $0.63$ to $0.85$ on GPT-5.4 text, and a locally-calibrated DMAP detector we introduce achieves state-of-the-art performance across the board. That said, our central contribution is not a new detector, but a precise diagnosis of a significant cause of under-performance of existing detectors and a principled, modular remedy compatible with any token-averaging pipeline. This will serve as a foundation for the community to build upon, with natural avenues including richer distributional models, improved calibration strategies, and principled ensembling with hidden-space geometry signals via the full Bayes-optimal decision rule.\n\n---\n\n## 103. Correct Code, Vulnerable Dependencies: A Large Scale Measurement Study of LLM-Specified Library Versions\n- **Authors**: Chengjie Wang , Jingzheng Wu , Xiang Ling , Tianyue Luo , Chen Zhao\n- **URL**: [https://arxiv.org/abs/2605.06279](https://arxiv.org/abs/2605.06279)\n- **Abstract**:\n> Large language models (LLMs) are now largely involved in software development workflows, and the code they generate routinely includes third-party library (TPL) imports annotated with specific version identifiers. These version choices can carry security and compatibility risks, yet they have not been systematically studied. We present the first large-scale measurement study of version-level risk in LLM-generated Python code, evaluating 10 LLMs on PinTrace, a curated benchmark of 1,000 Stack Overflow programming tasks. LLMs tend to specify version identifiers when directly prompted at 26.83%-95.18%, while down to 6.45%-59.19% in creating a manifest file directly. Among the specified versions, 36.70%-55.70% of tasks contain at least one known CVE, and 62.75%-74.51% of them carry Critical or High severity ratings. In 72.27%-91.37% of cases, the associated CVEs were publicly disclosed before the model's knowledge cutoff. The statistics show all models converge on the same small set of risky release versions, indicating a systemic bias rather than isolated model error. Static compatibility rates range from 19.70% to 63.20%, with installation failure as the dominant cause. The dynamic test cases confirm the pattern by 6.49%-48.62% pass rates. Further experiments confirm that these failures are attributable to version selection rather than code quality, and that externally anchored version constraints substantially reduce both vulnerability exposure and compatibility failures. Our findings reveal LLM version selection as a first-class, previously overlooked risk surface in LLM-based development. We disclosed these findings to the community of the evaluated models, and several confirmed the issue. All the code and dataset have been released for open science at this https URL .\n\n---\n\n## 104. Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs\n- **Authors**: Andy Zeyi Liu , Michael Zhang , Ilana Greenberg , Adam Alnasser , Lucas Baker , John Sous\n- **URL**: [https://arxiv.org/abs/2605.06225](https://arxiv.org/abs/2605.06225)\n- **Abstract**:\n> Steering large language models (LLMs) is usually done by either instruction prompting or activation steering. Prompting often gives strong control, but caches guidance tokens at every layer and can clutter long interactions; activation steering is compact but typically weaker and does not support large structured reminders. We introduce memory inception (MI), a training-free method that steers in latent attention space by inserting text-derived key-value (KV) banks only at selected layers. Rather than materializing reminder content throughout the prompt cache, MI treats steering as selective KV allocation, injecting latent slots only where the model routes to them. On matched personality-steering tasks, MI gives the best overall control--drift trade-off, remaining competitive with prompting while consistently outperforming CAA. On updateable guidance, MI supports mid-conversation behavior shifts without rewriting the visible transcript, achieving the highest post-shift alignment on Qwen3. On structured reasoning, MI outperforms visible prompting on HARDMath and PHYSICS (10/12 subject$\\times$mode cells), serving as proxies for structured reasoning in verifiable domains, while cutting content-matched KV storage by up to 118$\\times$. These results position MI as a powerful steering method when guidance is persistent, structured, or expensive to keep in the visible transcript.\n\n---\n\n## 105. TIDE: Every Layer Knows the Token Beneath the Context\n- **Authors**: Ajay Jaiswal , Lauren Hannah , Han-Byul Kim , Duc Hoang , Mehrdad Farajtabar , Minsik Cho\n- **URL**: [https://arxiv.org/abs/2605.06216](https://arxiv.org/abs/2605.06216)\n- **Abstract**:\n> We revisit a universally accepted but under-examined design choice in every modern LLM: a token index is looked up once at the input embedding layer and then permanently discarded. This single-injection assumption induces two structural failures: (i) the Rare Token Problem, where a Zipf-type distribution of vocabulary causes rare-token embeddings are chronically under-trained due to receiving a fraction of the cumulative gradient signal compared to common tokens; and (ii) the Contextual Collapse Problem, where limited parameters models map distributionally similar tokens to indistinguishable hidden states. As an attempt to address both, we propose TIDE, which augments the standard transformer with EmbeddingMemory: an ensemble of K independent MemoryBlocks that map token indices to context-free semantic vectors, computed once and injected into every layer through a depth-conditioned softmax router with a learnable null bank. We theoretically and empirically establish the benefits of TIDE in addressing the issues associated with single-token identity injection as well as improve performance across multiple language modeling and downstream tasks.\n\n---\n\n## 106. Retina-RAG: Retrieval-Augmented Vision-Language Modeling for Joint Retinal Diagnosis and Clinical Report Generation\n- **Authors**: Abdelrahman Zaian , Sheethal Bhat , Mohamed Abdalkader , Andreas Maier\n- **URL**: [https://arxiv.org/abs/2605.06173](https://arxiv.org/abs/2605.06173)\n- **Abstract**:\n> Diabetic Retinopathy (DR) is a leading cause of preventable blindness among working-age adults worldwide, yet most automated screening systems are limited to image-level classification and lack clinically structured reporting. We propose Retina-RAG, a low-cost modular framework that jointly performs DR severity grading, macular edema (ME) detection, and report generation. The architecture decouples a high-performance retinal classifier and a parameter-efficient vision-language model (Qwen2.5-VL-7B-Instruct) adapted via Low-Rank Adaptation (LoRA), enabling flexible component integration. A retrieval-augmented generation (RAG) module injects curated ophthalmic knowledge together with structured classifier outputs at inference time to improve diagnostic consistency and reduce hallucinations. Retina-RAG achieves an F1-score of 0.731 for DR grading and 0.948 for ME detection, substantially outperforming zero-shot Qwen (0.096, 0.732) and MMed-RAG (0.541, 0.641) on a retinal disease detection dataset with captions. For report generation, Retina-RAG attains ROUGE-L 0.429 and SBERT similarity 0.884, exceeding all baselines. The full framework operates on a single consumer-grade GPU, demonstrating that clinically structured retinal AI can be achieved with modest computational resources.\n\n---\n\n## 107. IRC-Bench: Recognizing Entities from Contextual Cues in First-Person Reminiscences\n- **Authors**: Yehudit Aperstein , Eden Moran , Alexander Apartsin\n- **URL**: [https://arxiv.org/abs/2605.06142](https://arxiv.org/abs/2605.06142)\n- **Abstract**:\n> When people recount personal memories, they often refer to people, places, and events indirectly, relying on contextual cues rather than explicit names. Such implicit references are central to reminiscence narratives: first-person accounts of lived experience used in therapeutic, archival, and social settings. They pose a difficult computational problem because the intended entity must be inferred from dispersed narrative evidence rather than from a local mention. We introduce IRC-Bench, the Implicit Reminiscence Context Benchmark, for evaluating implicit entity recognition in reminiscence transcripts. The benchmark targets non-locality: entity-identifying cues are distributed across multiple, non-contiguous clauses, unlike named entity recognition, entity linking, or coreference resolution. IRC-Bench comprises 25,136 samples constructed from 12,337 Wiki-data-linked entities across 1,994 transcripts spanning 11 thematic domains. Each sample pairs an Entity-Grounded Narrative, in which the target entity is explicitly mentioned, with an Entity-Elided Narrative, in which direct mentions are removed. We evaluate 19 configurations across LLM generation, dense retrieval, RAG, and fine-tuning. QLoRA-adapted Llama 3.1 8B performs best in the open-world setting (38.94% exact match; 51.59% Jaccard), while fine-tuned DPR leads closed-world retrieval (35.38% Hit@1; 71.49% Hit@10). We release IRC-Bench with data, code, and evaluation tools.\n\n---\n\n## 108. Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex\n- **Authors**: Yun Qu , Qi Wang , Yixiu Mao , Heming Zou , Yuhang Jiang , Yingyue Li , Wutong Xu , Lizhou Cai , Weijie Liu , Clive Bai , Kai Yang , Yangkun Chen , Saiyong Yang , Xiangyang Ji\n- **URL**: [https://arxiv.org/abs/2605.06139](https://arxiv.org/abs/2605.06139)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has become a standard approach for large language models (LLMs) post-training to incentivize reasoning capacity. Among existing recipes, group-based policy gradient is prevalent, which samples a group of responses per prompt and updates the policy via group-relative advantage signals. This work reveals that these optimization strategies share a common geometric structure: each implicitly defines a target distribution on the response simplex and projects toward it via first-order approximation. Building on this insight, we propose Listwise Policy Optimization (LPO) to explicitly conduct the target-projection, which demystifies the implicit target by restricting the proximal RL objective to the response simplex, and then projects the policy via exact divergence minimization. This framework provides (i) monotonic improvement on the listwise objective with bounded, zero-sum, and self-correcting projection gradients, and (ii) flexibility in divergence selection with distinct structural properties through the decoupled projection step. On diverse reasoning tasks and LLM backbones, LPO consistently improves training performance over typical policy gradient baselines under matched targets, while intrinsically preserving optimization stability and response diversity.\n\n---\n\n## 109. Normalized Architectures are Natively 4-Bit\n- **Authors**: Maxim Fishman , Brian Chmiel , Ron Banner , Daniel Soudry , Boris Ginsburg\n- **URL**: [https://arxiv.org/abs/2605.06067](https://arxiv.org/abs/2605.06067)\n- **Abstract**:\n> Training large language models at 4-bit precision is critical for efficiency. We show that nGPT, an architecture that constrains weights and hidden representations to the unit hypersphere, is inherently more robust to low-precision arithmetic. This removes the need for interventions-such as applying random Hadamard transforms and performing per-tensor scaling calculations-to preserve model quality, and it enables stable end-to-end NVFP4 training. We validate this approach on both a 1.2B dense model and hybrid (Mamba-Transformer) MoE models of up to 3B/30B parameters. We trace this robustness to the dot product: while quantization noise remains largely uncorrelated in both standard and normalized architectures, the signal behaves differently. In nGPT, the hypersphere constraint enhances weak positive correlations among the element-wise products, leading to a constructive accumulation of the signal across the hidden dimension while the noise continues to average out. This yields a higher effective signal-to-noise ratio and a flatter loss landscape, with the effect strengthening as the hidden dimension grows, suggesting increasing advantages at scale. A reference implementation is available at this https URL\n\n---\n\n## 110. Causal Reinforcement Learning for Complex Card Games: A Magic The Gathering Benchmark\n- **Authors**: Cristiano da Costa Cunha , Ajmal Mian , Tim French , Wei Liu\n- **URL**: [https://arxiv.org/abs/2605.06066](https://arxiv.org/abs/2605.06066)\n- **Abstract**:\n> Causal reinforcement learning (RL) lacks benchmarks for complex systems that combine sequential decision making, hidden information, large masked action spaces, and explicit causal structure. We introduce MTG-Causal-RL, a Gymnasium benchmark built on Magic: The Gathering with a 3,077-dimensional partial observation, a 478-action masked discrete action space, five competitive Standard archetypes, three reward schemes, and a hand-specified Structural Causal Model (SCM) over strategic variables. Every episode exposes causal variables, SCM-predicted intervention effects, and per-factor credit traces, making causal credit assignment, leave-one-out cross-archetype transfer, and policy auditability first-class metrics. We adapt a panel of reference baselines: random, heuristic, masked PPO, a causal-world-model PPO variant, and an architecture-matched scalar control. We propose Causal Graph-Factored Advantage PPO (CGFA-PPO) as a reference causal agent that uses SCM parents of win probability as factor-aligned critic targets with an intervention-calibration loss. All comparisons use paired seeds, paired-bootstrap confidence intervals, and Holm-Bonferroni correction within pre-registered families. Masked PPO and CGFA-PPO reach competitive in-distribution win rates and exceed the random baseline; per-factor calibration trajectories and leave-one-out transfer gaps expose diagnostic structure that scalar win rate alone cannot. We release the benchmark, reference-baseline results, and full evaluation protocol openly. By coupling a strategically rich, partially observed domain with an explicit causal interface and statistical protocol, MTG-Causal-RL gives causal-RL, world-model, and LLM-agent research a shared testbed for questions current benchmarks cannot pose together: causal credit assignment under masked action spaces, structural transfer across archetypes, and SCM-grounded policy auditability.\n\n---\n\n## 111. Optimal Transport for LLM Reward Modeling from Noisy Preference\n- **Authors**: Licheng Pan , Haochen Yang , Haoxuan Li , Yunsheng Lu , Yongqi Tong , Yinuo Wang , Shijian Wang , Zhixuan Chu , Lei Shen , Yuan Lu , Hao Wang\n- **URL**: [https://arxiv.org/abs/2605.06036](https://arxiv.org/abs/2605.06036)\n- **Abstract**:\n> Reward models are fundamental to Reinforcement Learning from Human Feedback (RLHF), yet real-world datasets are inevitably corrupted by noisy preference. Conventional training objectives tend to overfit these errors, while existing denoising approaches often rely on homogeneous noise assumptions that fail to capture the complexity of linguistic preferences. To handle these challenges, we propose SelectiveRM, a framework grounded in optimal transport. We first devise a Joint Consistency Discrepancy to align the distribution of model predictions with preference data. Furthermore, to address the limitation of strict mass conservation which compels the model to fit outliers, we incorporate a Mass Relaxation mechanism via partial transport. This enables the autonomous exclusion of samples with noisy preference that contradict semantic consistency. Theoretically, we demonstrate that SelectiveRM optimizes a tighter upper bound on the unobserved clean risk. Extensive experiments validate that our approach significantly outperforms state-of-the-art baselines across diverse benchmarks.\n\n---\n\n## 112. Does Synthetic Data Help? Empirical Evidence from Deep Learning Time Series Forecasters\n- **Authors**: Hugo Cazaux , Eyjólfur Ingi Ásgeirsson , Hlynur Stefánsson\n- **URL**: [https://arxiv.org/abs/2605.06032](https://arxiv.org/abs/2605.06032)\n- **Abstract**:\n> Synthetic data has transformed language model training, yet its role in time series forecasting remains poorly understood. We present a large-scale empirical study: nine experiment groups, 4,218 runs systematically evaluating synthetic time series augmentation across five architectures, four synthetic signals and seven datasets. The effect is sharply architecture-conditional: channel-mixing models (TimesNet, iTransformer) benefit in the majority of trials, while channel-independent models (DLinear, PatchTST) are consistently degraded. In selected low-resource settings the gains are striking: TimesNet trained on only 10\\% of Weather data with synthetic augmentation surpasses the full-data baseline (4 of 16 sparsity-dataset combinations). Averaged across all architectures, augmentation hurts in 67\\% of trials. We further find that only the Seasonal-Trend generator reliably helps across the tested benchmarks, and that hard curriculum switching is actively harmful (+24\\% MSE degradation). These results provide concrete, actionable guidelines on how to use synthetic data: use synthetic augmentation with channel-mixing architectures, use gradual annealing schedules, and treat low-resource augmentation as architecture- and dataset-dependent. Code is available at \\href{ this https URL }\n\n---\n\n## 113. Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks\n- **Authors**: Guoxin Lu , Letian Sha , Qing Wang , Peijie Sun , Hao Zhou , Hua Dai , Fu Xiao\n- **URL**: [https://arxiv.org/abs/2605.05995](https://arxiv.org/abs/2605.05995)\n- **Abstract**:\n> The safety alignment of Large Language Models (LLMs) remains vulnerable to Harmful Fine-tuning (HFT). While existing defenses impose constraints on parameters, gradients, or internal representations, we observe that they can be effectively circumvented under persistent HFT. Our analysis traces this failure to the inherent redundancy of the high-dimensional parameter space: attackers exploit optimization trajectories that are orthogonal to defense constraints to restore harmful capabilities while deceptively adhering to safety restrictions. To address this, we propose Safety Bottleneck Regularization (SBR). SBR shifts the defensive focus from the redundant parameter space to the unembedding layer, which serves as a geometric bottleneck. By anchoring the final hidden states of harmful queries to those of the safety-aligned model, SBR enables the model to maintain safe responses even under persistent HFT. Extensive experiments confirm SBR's effectiveness, demonstrating that utilizing just a single safety anchor is sufficient to reduce the Harmful Score to $<$10 while preserving competitive performance on benign downstream tasks.\n\n---\n\n## 114. PragLocker: Protecting Agent Intellectual Property in Untrusted Deployments via Non-Portable Prompts\n- **Authors**: Qinfeng Li , Yuntai Bao , Jianghui Hu , Wenqi Zhang , Jintao Chen , Huifeng Zhu , Yier Jin , Xuhong Zhang\n- **URL**: [https://arxiv.org/abs/2605.05974](https://arxiv.org/abs/2605.05974)\n- **Abstract**:\n> LLM agents rely on prompts to implement task-specific capabilities based on foundation LLMs, making agent prompts valuable intellectual property. However, in untrusted deployments, adversaries can copy and reuse these prompts with other proprietary LLMs, causing economic losses. To protect these prompts, we identify four key challenges: proactivity, runtime protection, usability, and non-portability that existing approaches fail to address. We present PragLocker, a prompt protection scheme that satisfies these requirements. PragLocker constructs function-preserving obfuscated prompts by anchoring semantics with code symbols and then using target-model feedback to inject noise, yielding prompts that only work on the target LLM. Experiments across multiple agent systems, datasets, and foundation LLMs show that PragLocker substantially reduces cross-LLM portability, maintains target performance, and remains robust against adaptive attackers.\n\n---\n\n## 115. Towards Reliable LLM Evaluation: Correcting the Winner's Curse in Adaptive Benchmarking\n- **Authors**: Yang Xu , Jiefu Zhang , Haixiang Sun , Zihan Zhou , Tianyu Cao , Vaneet Aggarwal\n- **URL**: [https://arxiv.org/abs/2605.05973](https://arxiv.org/abs/2605.05973)\n- **Abstract**:\n> Adaptive prompt and program search makes LLM evaluation selection-sensitive. Once benchmark items are reused inside tuning, the observed winner's score need not estimate the fresh-data performance of the full tune-then-deploy procedure. We study inference for this procedure-level target under explicit tuning budgets. We propose SIREN, a selection-aware repeated-split reporting protocol that freezes the post-search shortlist, separates splitwise selection from held-out evaluation, and uses an item-level Gaussian multiplier bootstrap for uncertainty quantification. In a fixed-shortlist regime with smooth stabilized selection, the estimator admits a first-order item-level representation, and the bootstrap yields valid simultaneous inference on a finite budget grid. This supports confidence intervals for procedure-performance curves and pre-specified equal-budget and cross-budget comparisons. Controlled simulations and MMLU-Pro tuning experiments show that winner-based reporting can be optimistic and can change deployment conclusions, while SIREN remains close to the finite-sample reporting target.\n\n---\n\n## 116. Beyond Uniform Credit Assignment: Selective Eligibility Traces for RLVR\n- **Authors**: Chaoli Mou , Zhan Zhuang , Xinning Chen , Yu Zhang\n- **URL**: [https://arxiv.org/abs/2605.05965](https://arxiv.org/abs/2605.05965)\n- **Abstract**:\n> Reinforcement Learning with Verifiable Rewards (RLVR) has become a key approach for improving the reasoning abilities of large language models. However, widely used critic-free algorithms such as Group Relative Policy Optimization (GRPO) necessitate a ``uniform credit assignment'' assumption that indiscriminately broadcast trajectory-level advantages, hindering learning efficiency by failing to distinguish critical reasoning steps. To address this limitation, we propose Selective Eligibility Traces (S-trace). Grounded in the intuition of partial trust region preservation, we initially introduce P-trace as a sample-efficient, critic-free eligibility traces method, upon which we build S-trace, implementing a sparse eligibility traces mechanism to further mitigate variance and achieve fine-grained credit assignment by selectively masking low-entropy tokens. Theoretically, we contextualize the recent Group Sequence Policy Optimization (GSPO) method within the critic-free eligibility traces framework, identifying it as a special instance of the eligibility traces method operating under uniform credit assignment. Experiments demonstrate that S-trace not only outperforms GRPO, showing gains of 0.49\\% on Qwen3-1.7B and 3.16\\% on Qwen3-4B, and maintaining a robust 2.98\\% improvement when scaled further to Qwen3-8B in average pass@16, but notably achieves this with simultaneously higher sample and token efficiency.\n\n---\n\n## 117. Hallucination as an Anomaly: Dynamic Intervention via Probabilistic Circuits\n- **Authors**: Erik Nielsen , Elia Cunegatti , Marcus Vukojevic , Giovanni Iacca\n- **URL**: [https://arxiv.org/abs/2605.05953](https://arxiv.org/abs/2605.05953)\n- **Abstract**:\n> One of the most critical challenges in Large Language Models is their tendency to hallucinate, i.e., produce factually incorrect responses. Existing approaches show promising results in terms of hallucination correction, but still suffer from a main limitation: they apply corrections indiscriminately to every token, corrupting also the originally correct generations. To overcome this drawback, we propose PCNET, a Probabilistic Circuit trained as a tractable density estimator over the LLM residual stream. The method detects hallucinations as geometric anomalies on the factual manifold, which is done via exact Negative Log-Likelihood computation, hence without the need for sampling, external verifiers, or weight modifications, as in existing techniques. To demonstrate its effectiveness, we exploit PCNET as a dynamic gate that distinguishes hallucinated from factual hidden states at each decoding step. This triggers our second main contribution, PC-LDCD (Probabilistic Circuit Latent Density Contrastive Decoding), only when the latent geometry deviates from factual regions, while leaving correct generations untouched. Across four LLMs, ranging from 1B to 8B models, and four benchmarks covering conversational reasoning, knowledge-intensive QA, reading comprehension, and truthfulness, PCNET achieves near-perfect hallucination detection across CoQA, SQuAD v2.0, and TriviaQA, with AUROC reaching up to 99%. Moreover, PC-LDCD obtains the highest True+Info, MC2, and MC3 scores on TruthfulQA in three out of four models, in comparison with state-of-the-art baselines, while reducing the mean corruption rate to 53.7% and achieving a preservation rate of 79.3%. Our proposed method is publicly available on GitHub.\n\n---\n\n## 118. LLM-Driven Design Space Exploration of FPGA-based Accelerators\n- **Authors**: Vinamra Sharma , Xingjian Fu , Jude Haris , José Cano\n- **URL**: [https://arxiv.org/abs/2605.05920](https://arxiv.org/abs/2605.05920)\n- **Abstract**:\n> Designing field-programmable gate array (FPGA)-based accelerators for modern artificial intelligence workloads requires navigating a large and complex hardware design space encompassing architectural parameters, dataflow strategies, and memory hierarchies, making the process time-consuming and resource-intensive. While the SECDA methodology enables rapid hardware-software co-design of accelerators through SystemC simulation and FPGA execution, identifying optimal accelerator configurations still requires substantial manual effort and domain expertise. This work presents SECDA-DSE, a framework that integrates Large Language Models (LLMs) into the SECDA ecosystem, comprising tools built around SECDA to automate the design space exploration (DSE) of FPGA-based accelerators. SECDA-DSE combines a structured DSE Explorer for generating accelerator configurations with an LLM Stack that performs reasoning-guided exploration using retrieval-augmented generation and chain-of-thought prompting, alongside a feedback loop that enables reinforced fine-tuning for continuous improvement. We demonstrate the feasibility of SECDA-DSE through an initial high-level synthesis based evaluation of a generated accelerator design that meets synthesis timing and resource constraints on an Zynq-7000 FPGA.\n\n---\n\n## 119. Quantum-enhanced Large Language Models on Quantum Hardware via Cayley Unitary Adapters\n- **Authors**: Borja Aizpurua , Sukhbinder Singh , Augustine Kshetrimayum , Saeed S. Jahromi , Roman Orus\n- **URL**: [https://arxiv.org/abs/2605.05914](https://arxiv.org/abs/2605.05914)\n- **Abstract**:\n> Large language models (LLMs) have transformed artificial intelligence, yet classical architectures impose a fundamental constraint: every trainable parameter demands classical memory that scales unfavourably with model size. Quantum computing offers a qualitatively different pathway, but practical demonstrations on real hardware have remained elusive for models of practical relevance. Here we show that Cayley-parameterised unitary adapters -- quantum circuit blocks inserted into the frozen projection layers of pre-trained LLMs and executed on a 156-qubit IBM Quantum System Two superconducting processor -- improve the perplexity of Llama 3.1 8B, an 8-billion-parameter model in widespread use, by 1.4% with only 6,000 additional parameters and end-to-end inference validated on real Quantum Processing Unit (QPU). A systematic study on SmolLM2 (135M parameters), chosen for its tractability, reveals monotonically improving perplexity with unitary block dimension, 83% recovery of compression-induced degradation, and correct answers to questions that both classical baselines fail -- with a sharp noise-expressivity phase transition identifying the concrete path to quantum utility at larger qubit scales.\n\n---\n\n## 120. CITE: Anytime-Valid Statistical Inference in LLM Self-Consistency\n- **Authors**: Hirofumi Ota , Naoto Iwase , Yuki Ichihara , Junpei Komiyama , Masaaki Imaizumi\n- **URL**: [https://arxiv.org/abs/2605.05873](https://arxiv.org/abs/2605.05873)\n- **Abstract**:\n> Large language models often improve reasoning by sampling multiple outputs and aggregating their final answers, but precise and efficient control of error levels remains a challenging task. In particular, deciding when to stop sampling remains difficult when the stopping rule is data-dependent and the set of possible answers is not known in advance. We study anytime-valid certification of a prespecified target answer as the unique mode of the model's response distribution, a guarantee distinct from answer correctness. We propose the Certification by Intersection-union Testing with E-processes (CITE) algorithm, which provably controls false certification at any prescribed level under arbitrary data-driven stopping, without requiring prior knowledge of the answer category set. We also prove an category-set-size-free stopping-time rate, establish matching minimax lower bounds up to constants in the main regime, and extend the construction to confidence-weighted voting. Simulations and LLM self-consistency experiments show empirical error control and improved certification in diffuse-tail settings.\n\n---\n\n## 121. VideoRouter: Query-Adaptive Dual Routing for Efficient Long-Video Understanding\n- **Authors**: Kuanwei Lin , Wenhao Zhang , Ge Li\n- **URL**: [https://arxiv.org/abs/2605.05848](https://arxiv.org/abs/2605.05848)\n- **Abstract**:\n> Video large multimodal models increasingly face a scalability bottleneck: long videos produce excessively long visual-token sequences, which sharply increase memory and latency during inference. While existing compression methods are effective in specific settings, most are either weakly query-aware or apply a fixed compression policy across frames, proving suboptimal when visual evidence is unevenly distributed over time. To address this, we present VideoRouter, a query-adaptive dual-router framework built on InternVL for budgeted evidence allocation. The Semantic Router predicts the dominant allocation policy, choosing between broad temporal coverage and adaptive high-resolution preservation, while the Image Router uses early LLM layers to score frame relevance. This enables aggressive compression on less relevant frames while preserving detail on critical evidence frames. To train both routers, we build Video-QTR-10K for allocation-policy supervision and Video-FLR-200K for frame-relevance supervision. Experiments on VideoMME, MLVU, and LongVideoBench show that VideoRouter consistently improves over the InternVL baseline under comparable or lower budgets, achieving up to a 67.9% token reduction.\n\n---\n\n## 122. LoopTrap: Termination Poisoning Attacks on LLM Agents\n- **Authors**: Huiyu Xu , Zhibo Wang , Wenhui Zhang , Ziqi Zhu , Yaopeng Wang , Kui Ren , Chun Chen\n- **URL**: [https://arxiv.org/abs/2605.05846](https://arxiv.org/abs/2605.05846)\n- **Abstract**:\n> Modern LLM agents solve complex tasks by operating in iterative execution loops, where they repeatedly reason, act, and self-evaluate progress to determine when a task is complete. In this work, we show that while this self-directed loop facilitates autonomy, it also introduces a critical risk: by injecting malicious prompts into the agent's context, an adversary can distort the agent's termination judgment, making it believe the task remains incomplete and leading to unbounded this http URL understand this threat, we define and systematically characterize it as Termination Poisoning and design 10 representative attack strategies. Through a empirical study spanning 8 LLM agents and 60 tasks, we demonstrate that different LLM agents exhibit distinct behavioral signatures that determine which strategies succeed. These transferable patterns can serve as principled guidance for crafting effective attacks against previously unseen agents and tasks, enabling scalable red-teaming beyond manually designed templates. Building on these insights, we introduce LoopTrap, an automated red-teaming framework that synthesizes target-specific malicious prompts by exploiting agent behavioral tendencies. LoopTrap first constructs a behavioral profile of the target agent along four vulnerability dimensions via lightweight probing. It then performs adaptive trap synthesis, routing to the most effective strategy and selecting optimal injections via a self-scoring mechanism. Finally, successful traps are abstracted into a reusable skill library, while failed attempts are refined through self-reflection, ensuring continuous improvement. Extensive evaluation shows that LoopTrap achieves an average of 3.57$\\times$ step amplification across 8 mainstream agents, with a peak of 25$\\times$.\n\n---\n\n## 123. LeakDojo: Decoding the Leakage Threats of RAG Systems\n- **Authors**: Maosen Zhang , Jianshuo Dong , Boting Lu , Wenyue Li , Xiaoping Zhang , Tianwei Zhang , Han Qiu\n- **URL**: [https://arxiv.org/abs/2605.05818](https://arxiv.org/abs/2605.05818)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to leverage external knowledge, but also exposes valuable RAG databases to leakage attacks. As RAG systems grow more complex and LLMs exhibit stronger instruction-following capabilities, existing studies fall short of systematically assessing RAG leakage risks. We present LeakDojo, a configurable framework for controlled evaluation of RAG leakage. Using LeakDojo, we benchmark six existing attacks across fourteen LLMs, four datasets, and diverse RAG systems. Our study reveals that (1) query generation and adversarial instructions contribute independently to leakage, with overall leakage well approximated by their product; (2) stronger instruction-following capability correlates with higher leakage risk; and (3) improvements in RAG faithfulness can introduce increased leakage risk. These findings provide actionable insights for understanding and mitigating RAG leakage in practice. Our codebase is available at this https URL .\n\n---\n\n## 124. LCC-LLM: Leveraging Code-Centric Large Language Models for Malware Attribution\n- **Authors**: Christopher G. Pedraza Pohlenz , Hassan Jalil Hadi , Ali Hassan , Ali Shoker\n- **URL**: [https://arxiv.org/abs/2605.05807](https://arxiv.org/abs/2605.05807)\n- **Abstract**:\n> LLMs are increasingly explored for malware analysis; however, current LLM-based malware attribution remains limited by unsupported indicators and insufficient code-level grounding for identifying malicious and vulnerable code segments. To address these limitations, this research introduces LCC-LLM, a code-centric benchmark dataset and evidence-grounded framework for malware attribution and multi-task static malware analysis. The proposed LCCD dataset contains approximately 34K PE samples processed through a large-scale reverse-engineering pipeline and represented using decompiled C code, assembly code, CFG/FCG artifacts, hexadecimal data, PE metadata, suspicious API evidence, and structural features. Beyond dataset construction, LCC-LLM integrates LangGraph-orchestrated static analysis with multi-source cybersecurity knowledge to support evidence-grounded malware reasoning. The framework employs a seven-layer retrieval-augmented generation pipeline, CoVe for IoC validation, and a multi-dimensional quality gate to improve factual reliability and analyst-oriented decision support. Curriculum-ordered instruction data is used to fine-tune DeepSeek-R1-Distill-Qwen-14B and Qwen3-Coder-30B-A3B using QLoRA. Evaluation across 43 malware-analysis task types achieves an average semantic similarity of 0.634, with the highest task-level performance in structured report generation, IoC extraction, vulnerability assessment, malware configuration extraction, and malware class detection. In a real-world case study using MalwareBazaar samples, the grounded pipeline achieves a 10/10 structured analysis pass rate, producing CFG/FCG evidence, MITRE ATT&CK mappings, detection guidance, and analyst-ready reports. These results show that code-centric representations, retrieval grounding, and verification-guided reasoning improve the reliability and operational usefulness of LLM-assisted malware attribution.\n\n---\n\n## 125. Revealing Modular Gradient Noise Imbalance in LLMs: Calibrating Adam via Signal-to-Noise Ratio\n- **Authors**: Ziqing Wen , Zhouyang Liu , Jiahuan Wang , Ping Luo , Li Shen , Dongsheng Li , Tao Sun\n- **URL**: [https://arxiv.org/abs/2605.05794](https://arxiv.org/abs/2605.05794)\n- **Abstract**:\n> The impressive performance of large language models (LLMs) arises from their massive scale and heterogeneous module composition. However, this structural heterogeneity introduces additional optimization challenges. While adaptive optimizers such as Adam(W) provide per-parameter adaptivity, they do not explicitly account for module-level gradient heterogeneity, resulting in slower convergence, suboptimal performance, or training instability. Existing approaches typically rely on manually tuned module-specific learning rates or specific optimization strategies, which are computationally costly and difficult to generalize across tasks or models. To establish a more principled approach, we first analyze the noise-damping behavior of Adam in high-noise modules and introduce \\textbf{Module-wise Learning Rate Scaling via SNR (MoLS)}. MoLS estimates module-level SNRs to scale Adam updates, allowing automated module-wise learning rate allocation without manual tuning. Empirical results through multiple LLM training benchmarks demonstrate that MoLS improves convergence speed and generalization, achieving performance comparable to carefully tuned module-specific learning rates, while remaining compatible with memory-efficient training algorithms.\n\n---\n\n## 126. Multi-Dimensional Behavioral Evaluation of Agentic Stock Prediction Systems Using LLM Judges with Closed-Loop Reinforcement Learning Feedback\n- **Authors**: Mohammad Al Ridhawi , Mahtab Haj Ali , Hussein Al Osman\n- **URL**: [https://arxiv.org/abs/2605.05739](https://arxiv.org/abs/2605.05739)\n- **Abstract**:\n> Agentic stock prediction systems make sequences of interdependent decisions (regime detection, pathway routing, reinforcement learning control) whose individual quality is hidden by aggregate metrics such as mean absolute percentage error (MAPE) or directional accuracy. We present a behavioral evaluation framework that addresses this gap. Behavioral traces logged at every autonomous decision point are grouped into five-day episodes and scored along six domain-specific dimensions (regime detection, routing, adaptation, risk calibration, strategy coherence, error recovery) by an ensemble of three large language model (LLM) judges (GPT 5.4, Claude 4.6 Opus, Gemini 3.1 Pro). Perturbation-based validation on 420 episodes yields targeted score drops of $-1.6$ to $-2.4$ on intended dimensions versus an average of $-0.32$ on the remaining five, with cross-model agreement up to Krippendorff's $\\alpha = 0.85$. The composite behavioral score, used here only for cross-episode reporting, correlates at $\\rho = 0.72$ with realized 20-day Sharpe ratio from offline backtesting. Closing the loop, the framework converts deficient per-dimension scores into a credit-assigned penalty term added to the Soft Actor-Critic (SAC) reward. Three short fine-tuning cycles, all confined to the validation period, produce on the held-out 2017-2025 test period a one-day MAPE reduction from 0.61% to 0.54% (an 11.5% relative reduction; $p<0.001$, Cohen's $d=0.31$), a directional accuracy increase from 71% to 74%, and an 18% Sharpe ratio improvement (95% bootstrap CI [8.2%, 27.4%]), with gains concentrated in high-volatility episodes where the original system was most behaviorally deficient. Results are from offline backtesting and do not address effects specific to live deployment.\n\n---\n\n## 127. CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning\n- **Authors**: Md Anwar Hossen , Fatema Siddika , Juan Pablo Munoz , Tanya Roosta , Ali Jannesari\n- **URL**: [https://arxiv.org/abs/2605.05732](https://arxiv.org/abs/2605.05732)\n- **Abstract**:\n> Large language models (LLMs) can acquire new capabilities through fine-tuning, but continual adaptation often leads to catastrophic forgetting. We propose CRAFT, a continual learning framework that avoids updating model weights by instead learning low-rank interventions on hidden representations. CRAFT proceeds in three stages: it first routes each task to a group of similar tasks based on output-distribution divergence; it then fine-tunes the model using a Kullback-Leibler (KL) divergence against the group's prior state, which directly controls forgetting and determines convergence; finally, it merges interventions for the updated task into the shared representation using the same KL signal. This design unifies routing, regularization, and merging through a single KL-based objective. CRAFT improves overall performance and reduces forgetting compared to strong LoRA-based approaches across multiple benchmarks and model scales, while remaining robust to task ordering. These results suggest that controlling adaptation in representation space, guided by output-space divergence, provides a scalable and principled approach to continual learning in LLMs.\n\n---\n\n## 128. SafeHarbor: Hierarchical Memory-Augmented Guardrail for LLM Agent Safety\n- **Authors**: Zhe Liu , Zonghao Ying , Wenxin Zhang , Quanchen Zou , Deyue Zhang , Dongdong Yang , Xiangzheng Zhang , Hao Peng\n- **URL**: [https://arxiv.org/abs/2605.05704](https://arxiv.org/abs/2605.05704)\n- **Abstract**:\n> With the rapid evolution of foundation models, Large Language Model (LLM) agents have demonstrated increasingly powerful tool-use capabilities. However, this proficiency introduces significant security risks, as malicious actors can manipulate agents into executing tools to generate harmful content. While existing defensive mechanisms are effective, they frequently suffer from the over-refusal problem, where increased safety strictness compromises the agent's utility on benign tasks. To mitigate this trade-off, we propose \\textsc{SafeHarbor}, a novel framework designed to establish precise decision boundaries for LLM agents. Unlike static guidelines, \\textsc{SafeHarbor} extracts context-aware defense rules through enhanced adversarial generation. We design a local hierarchical memory system for dynamic rule injection, offering a training-free, efficient, and plug-and-play solution. Furthermore, we introduce an information entropy-based self-evolution mechanism that continuously optimizes the memory structure through dynamic node splitting and merging. Extensive experiments demonstrate that \\textsc{SafeHarbor} achieves state-of-the-art performance on both ambiguous benign tasks and explicit malicious attacks, notably attaining a peak benign utility of 63.6\\% on GPT-4o while maintaining a robust refusal rate exceeding 93\\% against harmful requests. The source code is publicly available at this https URL .\n\n---\n\n## 129. Active Learning for Communication Structure Optimization in LLM-Based Multi-Agent Systems\n- **Authors**: Huchen Yang , Xinghao Dong , Dan Negrut , Jin-Long Wu\n- **URL**: [https://arxiv.org/abs/2605.05703](https://arxiv.org/abs/2605.05703)\n- **Abstract**:\n> Optimizing the communication structure of large language model based multi-agent systems (LLM-MAS) has been shown to improve downstream performance and reduce token usage. Existing methods typically rely on randomly sampled training tasks. However, tasks may differ substantially in difficulty and domain, and thus they are not equally informative for updating communication structure, making optimization under limited training budgets often unstable and highly sensitive to the particular training set. To actively identify the most valuable tasks for communication-structure optimization, we propose an ensemble-based information-theoretic task selection framework. The proposed method estimates task informativeness by how much a candidate task changes the distribution over graph parameters, using ensemble Kalman inversion as an efficient and derivative-free approximation of the corresponding Bayesian update. The resulting estimator is especially suitable for black-box and noisy multi-agent systems. To enhance scalability, we construct a compact candidate pool through embedding-based representative selection and combine the informative selection with surrogate modeling and batch Thompson sampling. We validate our method in both benign settings and settings with agent attacks, demonstrating its effectiveness for communication-structure optimization under constrained computational budgets.\n\n---\n\n## 130. An Empirical Study of Proactive Coding Assistants in Real-World Software Development\n- **Authors**: Lehui Li , Ruixuan Jia , Guo-Ye Yang , Jia Li\n- **URL**: [https://arxiv.org/abs/2605.05700](https://arxiv.org/abs/2605.05700)\n- **Abstract**:\n> Large language model (LLM)-based coding assistants have made substantial progress, yet most systems remain reactive, requiring developers to explicitly formulate their needs. Proactive coding assistants aim to infer latent developer intent from integrated development environment (IDE) interactions and repository context, thereby reducing interaction overhead and supporting more seamless assistance. However, research in this direction is limited by the scarcity of large-scale real-world developer behavior data. Existing studies therefore often rely on LLM-simulated IDE traces, whose fidelity to real development behavior remains unclear. In this paper, we investigate this simulation-to-reality gap through a large-scale empirical study. We collect real IDE interaction traces from 1{,}246 experienced industry developers over three consecutive days using a custom Visual Studio Code extension, and construct paired LLM-simulated traces for controlled comparison. Our analysis shows that simulated traces differ substantially from real traces in behavioral diversity, temporal structure, and exploratory patterns. Based on the collected data, we introduce \\textbf{ProCodeBench}, a real-world benchmark for proactive intent prediction. Experiments with representative LLMs, retrieval-augmented methods, and agentic baselines show that current approaches remain far from reliable under real IDE traces, suggesting that simulation-based evaluation can overestimate real-world performance. Finally, our training study shows that simulated data cannot replace real data, but can complement it when used before real-world fine-tuning. These findings highlight the importance of real developer behavior data for evaluating and training proactive coding assistants.\n\n---\n\n## 131. Irminsul: MLA-Native Position-Independent Caching for Agentic LLM Serving\n- **Authors**: Bole Ma , Jan Eitzinger , Harald Köstler\n- **URL**: [https://arxiv.org/abs/2605.05696](https://arxiv.org/abs/2605.05696)\n- **Abstract**:\n> Agentic LLM workloads put bit-identical tokens at shifted positions every turn, voiding prefix caches at the first byte of divergence. Operators report cache-hit regressions ranging from moderate slowdowns to severe TTFT spikes of 10-16s on unchanged content. Prior position-independent caching systems correct RoPE on the full $d_K$-dimensional key, an architectural cost imposed by GQA, not by caching itself. Multi-Head Latent Attention, deployed at scale in DeepSeek-V2/V3/R1, Kimi-K2/Moonlight, GLM-5, and Mistral Large 3, factors each KV row into a position-free $c_{KV}$ and a 64-dim $k_r$ correctable in closed form; this structure motivates content-addressed caching as a natural fit rather than a GQA workaround. We present Irminsul, which extends SGLang's radix cache with content-hash keying over CDC-chunked segments and a $\\delta$-rotation rule for $k_r$. We evaluate three native MLA-MoE deployments - DeepSeek-V2-Lite (16B/2.4B), Kimi Moonlight-16B-A3B, and JoyAI-Flash (48B/3B) - with output-consistency on all three and recovery measured on the two endpoints; Irminsul recovers up to ~83% of prompt tokens above exact-prefix on agentic traffic while delivering 63% prefill energy savings per cache hit. We argue that content-addressed caching belongs in the serving stack as a first-class primitive, not a retrofit over prefix matching.\n\n---\n\n## 132. Decomposing the Basic Abilities of Large Language Models: Mitigating Cross-Task Interference in Multi-Task Instruct-Tuning\n- **Authors**: Bing Wang , Ximing Li , Changchun Li , Jinjin Chi , Gang Niu , Masashi Sugiyama\n- **URL**: [https://arxiv.org/abs/2605.05676](https://arxiv.org/abs/2605.05676)\n- **Abstract**:\n> Recently, the prominent performance of large language models (LLMs) has been largely driven by multi-task instruct-tuning. Unfortunately, this training paradigm suffers from a key issue, named cross-task interference, due to conflicting gradients over shared parameters among different tasks. Some previous methods mitigate this issue by isolating task-specific parameters, e.g., task-specific neuron selection and mixture-of-experts. In this paper, we empirically reveal that the cross-task interference still exists for the existing solutions because of many parameters also shared by different tasks, and accordingly, we propose a novel solution, namely Basic Abilities Decomposition for multi-task Instruct-Tuning (BADIT). Specifically, we empirically find that certain parameters are consistently co-activated, and that co-activated parameters naturally organize into base groups. This motivates us to analogize that LLMs encode several orthogonal basic abilities, and that any task can be represented as a linear combination of these abilities. Accordingly, we propose BADIT that decomposes LLM parameters into orthogonal high-singular-value LoRA experts representing basic abilities, and dynamically enforces their orthogonality during training via spherical clustering of rank-1 components. We conduct extensive experiments on the SuperNI benchmark with 6 LLMs, and empirical results demonstrate that BADIT can outperform SOTA methods and mitigate the degree of cross-task interference.\n\n---\n\n## 133. XL-SafetyBench: A Country-Grounded Cross-Cultural Benchmark for LLM Safety and Cultural Sensitivity\n- **Authors**: Dasol Choi , Eugenia Kim , Jaewon Noh , Sang Seo , Eunmi Kim , Myunggyo Oh , Yunjin Park , Brigitta Jesica Kartono , Josef Pichlmeier , Helena Berndt , Sai Krishna Mendu , Glenn Johannes Tungka , Özlem Gökçe , Suresh Gehlot , Katherine Pratt , Amanda Minnich , Haon Park\n- **URL**: [https://arxiv.org/abs/2605.05662](https://arxiv.org/abs/2605.05662)\n- **Abstract**:\n> Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Moreover, they rarely evaluate a model's ability to detect culturally embedded sensitivities as distinct from universal harms. We introduce XL-SafetyBench. a suite of 5,500 test cases across 10 country-language pairs, comprising a Jailbreak Benchmark of country-grounded adversarial prompts and a Cultural Benchmark where local sensitivities are embedded within innocuous requests. Each item is constructed via a multi-stage pipeline that combines LLM-assisted discovery, automated validation gates, and dual independent native-speaker annotators per country. To distinguish principled refusal from comprehension failure, we evaluate Attack Success Rate (ASR) alongside two complementary metrics we introduce: Neutral-Safe Rate (NSR) and Cultural Sensitivity Rate (CSR). Evaluating 10 frontier and 27 local LLMs reveals two key findings. First, jailbreak robustness and cultural awareness do not show a coupled relationship among frontier models, so a composite safety score obscures per-axis variation. Second, local models exhibit a near-linear ASR-NSR trade-off (r = -0.81), indicating that their apparent safety reflects generation failure rather than genuine alignment. XL-SafetyBench enables more nuanced, cross-cultural safety evaluation in the multilingual era.\n\n---\n\n## 134. One Turn Too Late: Response-Aware Defense Against Hidden Malicious Intent in Multi-Turn Dialogue\n- **Authors**: Xinjie Shen , Rongzhe Wei , Peizhi Niu , Haoyu Wang , Ruihan Wu , Eli Chien , Bo Li , Pin-Yu Chen , Pan Li\n- **URL**: [https://arxiv.org/abs/2605.05630](https://arxiv.org/abs/2605.05630)\n- **Abstract**:\n> Hidden malicious intent in multi-turn dialogue poses a growing threat to deployed large language models (LLMs). Rather than exposing a harmful objective in a single prompt, increasingly capable attackers can distribute their intent across multiple benign-looking turns. Recent studies show that even modern commercial models with advanced guardrails remain vulnerable to such attacks despite advances in safety alignment and external guardrails. In this work, we address this challenge by detecting the earliest turn at which delivering the candidate response would make the accumulated interaction sufficient to enable harmful action. This objective requires precise turn-level intervention that identifies the harm-enabling closure point while avoiding premature refusal of benign exploratory conversations. To further support training and evaluation, we construct the Multi-Turn Intent Dataset (MTID), which contains branching attack rollouts, matched benign hard negatives, and annotations of the earliest harm-enabling turns. We show that MTID helps enable a turn-level monitor TurnGate, which substantially outperforms existing baselines in harmful-intent detection while maintaining low over-refusal rates. TurnGate further generalizes across domains, attacker pipelines, and target models. Our code is available at this https URL .\n\n---\n\n## 135. Leveraging Image Generators to Address Training Data Scarcity: The Gen4Regen Dataset for Forest Regeneration Mapping\n- **Authors**: Gabriel Jeanson , David-Alexandre Duclos , William Larrivée-Hardy , Noé Cochet , Matěj Boxan , Anthony Deschênes , François Pomerleau , Philippe Giguère\n- **URL**: [https://arxiv.org/abs/2605.05627](https://arxiv.org/abs/2605.05627)\n- **Abstract**:\n> Sustainable forest management relies on precise species composition mapping, yet traditional ground surveys are labour-intensive and geographically constrained. While Uncrewed Aerial Vehicles (UAVs) offer scalable data collection, the transition to deep learning-based interpretation is bottlenecked by the severe scarcity of expert-annotated imagery, particularly in complex, visually heterogeneous regeneration zones. This paper addresses the dual challenges of data scarcity and extreme class imbalance in the semantic segmentation of fine-grained forest regeneration species by providing a scalable framework that reduces reliance on manual photo-interpretation for high-resolution, millimetre-level aerial imagery. Importantly, we leverage the large-scale vision-language Nano Banana Pro model to simultaneously generate high-fidelity images and their corresponding pixel-aligned semantic masks from prompts. We introduce WilDReF-Q-V2, an expansion of a natural forest dataset with 13 977 new unlabelled and 50 labelled real images, as well as the Gen4Regen dataset, featuring 2101 pairs of synthetic images and semantic masks. Our methodology integrates real-world data with AI-generated images, highlighting that AI-generated data is highly complementary to real-world data, with unified training yielding an F1 score improvement of over 15 %pt compared to purely supervised baselines. Furthermore, we demonstrate that even small quantities of prompt-generated data significantly improve performance for underrepresented species, some of which saw per-species F1 score gains of up to 30 %pt. We conclude that vision-language models can serve as agile data generators, effectively bootstrapping perception tasks for niche AI domains where expert labels are scarce or unavailable. Our datasets, source code, and models will be available at this https URL .\n\n---\n\n## 136. When2Speak: A Dataset for Temporal Participation and Turn-Taking in Multi-Party Conversations for Large Language Models\n- **Authors**: Vihaan Nama , Shreya Mendi , Zian Ye , Brinnae Bent\n- **URL**: [https://arxiv.org/abs/2605.05626](https://arxiv.org/abs/2605.05626)\n- **Abstract**:\n> Large Language Models (LLMs) excel at generating contextually appropriate responses but remain poorly calibrated for multi-party conversations, where deciding when to speak is as critical as what to say. In such settings, naively responding at every turn leads to excessive interruptions and degraded conversational coherence. We introduce When2Speak, a grounded synthetic dataset and four-stage generation pipeline for learning intervention timing in group interactions. The dataset comprises over 215,000 examples derived from 16,000 conversations involving 2-6 speakers, spanning diverse conversational styles, tones, and participant dynamics, and explicitly modeling SPEAK vs. SILENT decisions at each turn. Our pipeline combines real-world grounding, structured augmentation, controlled transcript synthesis, and fine-tuning-ready supervision, and is fully open-sourced to support reproducibility and adaptation to domain-specific conversational norms. Across multiple model families, supervised fine-tuning (SFT) on When2Speak significantly outperforms zero-shot baselines (e.g., the average Macro F1 increase across 4B+ parameter models was 60%, with the largest increase being 120%). However, SFT-trained models remain systematically over-conservative, missing nearly half of warranted interventions as seen through the Missed Intervention Rate (MIR), which was on average 0.50 and is noticed even at larger model sizes. To address this limitation, we apply reinforcement learning with asymmetric reward shaping, which reduces MIR to 0.186-0.218 and increases recall from 0.479 to 0.78-0.81. Our findings establish that temporal participation is a distinct and trainable dimension of conversational intelligence, and that grounded synthetic data provides an effective and scalable pathway for enabling LLMs to participate more naturally and appropriately in multi-party interactions.\n\n---\n\n## 137. AstroAlertBench: Evaluating the Accuracy, Reasoning, and Honesty of Multimodal LLMs in Astronomical Classification\n- **Authors**: Claire Chen , Jiabao Sean Xiao , Shuze Daniel Liu , Facundo Perez Paolino , Luke Handley , Theophile Jegou du Laz , Ricky Nilsson , Alice Zou , Matthew Graham , Ashish Mahabal\n- **URL**: [https://arxiv.org/abs/2605.05573](https://arxiv.org/abs/2605.05573)\n- **Abstract**:\n> Modern astronomical observatories generate a massive volume of multimodal data, creating a critical bottleneck for expert human review. While multimodal large language models (LLMs) have shown promise in interpreting complex visual and textual inputs, their ability to perform specialized scientific classification while providing interpretable reasoning remains understudied. We introduce AstroAlertBench, a comprehensive multimodal benchmark designed to evaluate LLM performance in astronomical event review along a three-stage logical chain: metadata grounding, scientific reasoning, and hierarchical classification over five categories. We use a pilot sample of 1,500 real-world alerts from the Zwicky Transient Facility (ZTF), a wide-field survey that scans the northern sky to detect transient astronomical events. On this dataset, we benchmark 13 frontier closed-source and open-weight LLMs that support visual input. Our results reveal that high accuracy does not always align with model ``honesty,'' defined as the ability to self-evaluate its reasoning, which affects its reliability as a real-world assistant. We further initialize a human-in-the-loop evaluation protocol as a precursor to future community-scale participation. Together, AstroAlertBench provides a framework for developing calibrated and interpretable astronomical assistants.\n\n---\n\n## 138. ReaComp: Compiling LLM Reasoning into Symbolic Solvers for Efficient Program Synthesis\n- **Authors**: Atharva Naik , Yash Mathur , Prakam , Carolyn Rose , David Mortensen\n- **URL**: [https://arxiv.org/abs/2605.05485](https://arxiv.org/abs/2605.05485)\n- **Abstract**:\n> LLMs can solve program synthesis tasks but remain inefficient and unreliable on hard instances requiring large combinatorial search. Given a small set of reasoning traces, we use coding agents to compile them into reusable symbolic program synthesizers over constrained DSLs. The resulting solvers require no LLM calls at test time and are strong standalone systems: symbolic solver ensembles reach 91.3% accuracy on PBEBench-Lite and 84.7% on PBEBench-Hard, outperforming LLMs with test-time scaling for the latter by +16.3 percentage points at zero LLM inference cost. They also complement LLM search, improving PBEBench-Hard accuracy from 68.4% to 85.8% while reducing reported token usage by 78%, and raising SLR-Bench hard-tier accuracy from 34.4% to 58.0% in a neuro-symbolic hybrid setting. Compared to directly using coding agents as per-instance solvers, induced solvers are substantially more Pareto-efficient, amortizing a small one-time construction cost over many zero-token executions. Finally, most solvers transfer zero-shot to a real historical linguistics task - predicting sound changes in natural language data - reaching 80.1% accuracy under ensembling and recovering some plausible linguistic rules. Together, these results show that reasoning traces can be compiled into reusable symbolic solvers that solve many tasks directly, complement LLM inference on hard cases, and provide a scalable route to domain-general solver induction. We release code and data for reproducibility.\n\n---\n\n## 139. Robustness of Graph Self-Supervised Learning to Real-World Noise: A Case Study on Text-Driven Biomedical Graphs\n- **Authors**: Othmane Kabal , Mounira Harzallah , Fabrice Guillet , Hideaki Takeda , Ryutaro Ichise\n- **URL**: [https://arxiv.org/abs/2605.05463](https://arxiv.org/abs/2605.05463)\n- **Abstract**:\n> Graph Self-Supervised Learning (GSSL) offers a powerful paradigm for learning graph representations without labeled data. However, existing work assumes clean, manually curated graphs. Recent advances in NLP enable the large-scale automatic extraction of knowledge graphs from text, opening new opportunities for GSSL while introducing substantial real-world noise. This type of noise remains largely unexplored, as prior robustness studies typically rely on synthetic perturbations. To address this gap, we present the first comprehensive evaluation of GSSL methods on text-driven graphs for unsupervised term typing. We introduce Noise-Aware Text-Driven Graph GSSL (NATD-GSSL), a unified framework that combines automatic graph construction, graph refinement, and GSSL. Our evaluation follows a dual-graph protocol that contrasts a noisy graph derived from MedMentions with a clean Unified Medical Language System (UMLS) reference graph, aligned through a shared gold standard. Our results reveal variability in robustness across both pretext tasks and Graph Neural Network (GNN) architectures. Relation reconstruction is highly sensitive to noise and benefits from well-defined schemas, whereas feature reconstruction is considerably more robust, achieving performance comparable to clean-graph settings. Contrastive objectives are generally less affected by noise but depend strongly on alignment with downstream tasks. GNN architecture also plays a critical role: bidirectional relational message-passing designs are better suited to noisy, text-driven graphs, while unidirectional relational ones perform best on clean graphs. Overall, NATD-GSSL provides practical guidance for applying GSSL to real-world, noisy graphs and achieves up to a 7\\% improvement over pretrained language model baselines. All code and benchmarks are publicly available at this https URL .\n\n---\n\n## 140. SLAM: Structural Linguistic Activation Marking for Language Models\n- **Authors**: Fabrice Harel-Canada , Amit Sahai\n- **URL**: [https://arxiv.org/abs/2605.05443](https://arxiv.org/abs/2605.05443)\n- **Abstract**:\n> LLM watermarks must be detectable without compromising text quality, yet most existing schemes bias the next-token distribution and pay for detection with measurable quality loss. We present SLAM (Structural Linguistic Activation Marking), a novel white-box watermarking scheme that sidesteps this cost by writing the mark into structural geometry rather than token frequencies: sparse autoencoders identify residual-stream directions encoding linguistic structure (e.g., voice, tense, clause order), and we causally steer those directions at generation time, leaving lexical sampling and semantics unconstrained. On Gemma-2 2B and 9B, SLAM achieves 100% detection accuracy with a quality cost of only 1-2 reward points - compared to 7.5-11.5 for KGW, EWD, and Unigram - with naturalness and diversity preserved at near-unwatermarked levels across both models. The trade-off is a complementary robustness profile: SLAM resists word-level edits but is vulnerable to paraphrase that restructures syntax (at a quality cost), the converse of token-distribution methods.\n\n---\n\n## 141. Information Theoretic Adversarial Training of Large Language Models\n- **Authors**: Yiwei Zhang , Jeremiah Birrell , Reza Ebrahimi , Rouzbeh Behnia , Jason Pacheco , Elisa Bertino\n- **URL**: [https://arxiv.org/abs/2605.05415](https://arxiv.org/abs/2605.05415)\n- **Abstract**:\n> Large language models (LLMs) remain vulnerable to adversarial prompting despite advances in alignment and safety, often exhibiting harmful behaviors under novel attack strategies. While adversarial training can improve robustness, existing approaches are computationally expensive and difficult to scale. Recent continuous adversarial training methods, such as Continuous adversarial training (CAT) and Continuous Adversarial Preference Optimization (CAPO), address this challenge by leveraging gradient-based perturbations in the embedding space, enabling more efficient and expressive attacks. Building on this paradigm, we propose WARDEN, a distributionally robust adversarial training framework for LLMs that dynamically reweights adversarial examples through an f -divergence ambiguity set around the empirical training distribution. Our method optimizes the worst-case adversarial loss within a divergence ball around the empirical data distribution, automatically emphasizing harder adversarial examples. Using the convex dual formulation, the objective reduces to a log-sum-exp form under the KL divergence, with a dynamical parameter controlling the strength of reweighting. This study leads to a new class of information-theoretic objectives that significantly reduce attack success rates while maintaining model utility. Across multiple LLMs and attack settings, WARDEN substantially reduces attack success rates with computational and utility costs comparable to CAT-, CAPO-, and MixAT-based baselines, making it a practical approach for scalable robust alignment.\n\n---\n\n## 142. Open-SAT: LLM-Guided Query Embedding Refinement for Open-Vocabulary Object Retrieval in Satellite Imagery\n- **Authors**: Md Adnan Arefeen , Biplob Debnath , Ravi K. Rajendran , Murugan Sankaradas , Srimat T. Chakradhar\n- **URL**: [https://arxiv.org/abs/2605.05344](https://arxiv.org/abs/2605.05344)\n- **Abstract**:\n> In satellite applications, user queries often take the form of open-ended natural language, extending beyond a fixed set of predefined categories. This open-vocabulary nature poses significant challenges for retrieving relevant image tiles, as the retrieval system must generalize to a wide range of unseen objects and concepts. While vision-language models (VLMs) such as CLIP are widely used for text-image retrieval, even fine-tuned variants often struggle to accurately align such queries with satellite imagery. To address this, we propose Open-SAT, a training-free query embedding refinement algorithm that operates at inference time to improve alignment between user queries and satellite image content. Open-SAT uses VLMs to compute embeddings for image tiles, which are stored in a vector database for efficient retrieval. At query time, it leverages Large Language Models (LLMs) to refine the text embeddings by incorporating contextual information about objects of interest and their surroundings. A threshold-free retrieval mechanism further enhances accuracy and efficiency. Experimental results in three public benchmarks demonstrate that Open-SAT improves the F1 score by up to 16.04%, while retrieving a comparable number of image tiles. These results demonstrate the effectiveness of Open-SAT in open-vocabulary satellite image retrieval, leveraging LLM guidance without the need for additional training or supervision.\n\n---\n\n## 143. Feature Starvation as Geometric Instability in Sparse Autoencoders\n- **Authors**: Faris Chaudhry , Keisuke Yano , Anthea Monod\n- **URL**: [https://arxiv.org/abs/2605.05341](https://arxiv.org/abs/2605.05341)\n- **Abstract**:\n> Sparse autoencoders (SAEs) are used to disentangle the dense, polysemantic internal representations of large language models (LLMs) into interpretable, monosemantic concepts. However, standard $\\ell_1$-regularized SAEs suffer from feature starvation (dead neurons) and shrinkage bias, often requiring computationally expensive heuristic resampling and nondifferentiable hard-masking methods to bypass these challenges. We argue that feature starvation is not merely an empirical artifact of poor data diversity, but a fundamental optimization-geometric pathology of overcomplete dictionaries: the $\\ell_1$-induced sparse coding map is unstable and fundamentally misaligned with shallow, amortized encoders. To address this structural instability, we introduce adaptive elastic net SAEs (AEN-SAEs), a fully differentiable architecture grounded in classical sparse regression. AEN-SAEs combine an $\\ell_2$ structural term that enforces strong convexity and Lipschitz stability with adaptive $\\ell_1$ reweighting that eliminates shrinkage bias and suppresses spurious features, thereby jointly controlling the curvature and interaction structure of the induced polyhedral geometry. Theoretically, we show that AEN-SAEs yield a Lipschitz-continuous sparse coding map and recover the global feature support under mild assumptions. Empirically, across synthetic settings and LLMs (Pythia 70M, Llama 3.1 8B), AEN-SAEs mitigate feature starvation without auxiliary heuristics while maintaining competitive reconstruction abilities.\n\n---\n\n## 144. How Far Are VLMs from Privacy Awareness in the Physical World? An Empirical Study\n- **Authors**: Junran Wang , Xinjie Shen , Zehao Jin , Pan Li\n- **URL**: [https://arxiv.org/abs/2605.05340](https://arxiv.org/abs/2605.05340)\n- **Abstract**:\n> As Vision-Language Models (VLMs) are increasingly deployed as autonomous cognitive cores for embodied assistants, evaluating their privacy awareness in physical environments becomes critical. Unlike digital chatbots, these agents operate in intimate spaces, such as homes and hospitals, where they possess the physical agency to observe and manipulate privacy-sensitive information and artifacts. However, current benchmarks remain limited to unimodal, text-based representations that cannot capture the demands of real-world settings. To bridge this gap, we present ImmersedPrivacy, an interactive audio-visual evaluation framework that simulates realistic physical environments using a Unity-based simulator. ImmersedPrivacy evaluates physically grounded privacy awareness across three progressive tiers that test a model's ability to identify sensitive items in cluttered scenes, adapt to shifting social contexts, and resolve conflicts between explicit commands and inferred privacy constraints. Our evaluation of 12 state-of-the-art models reveals consistent deficits. In cluttered scenes, all models exhibit monotonic performance decay as scene complexity grows due to perceptual deficit. When social context shifts, no model exceed 65% selection accuracy. Under conflicting commands, the best model gemini-3.1-pro perfectly balances task completion and privacy preservation in only 51% of cases. These findings reveal that current VLMs in the physical world suffer from perceptual fragility and fail to let their knowledge of privacy cues govern their situated behavior. Our code and data is available at this https URL .\n\n---\n\n## 145. Bridging Generation and Training: A Systematic Review of Quality Issues in LLMs for Code\n- **Authors**: Kaifeng He , Xiaojun Zhang , Peiliang Cai , Mingwei Liu , Yanlin Wang , Chong Wang , Kaifeng Huang , Bihuan Chen , Xin Peng , Zibin Zheng\n- **URL**: [https://arxiv.org/abs/2605.05267](https://arxiv.org/abs/2605.05267)\n- **Abstract**:\n> Large language models (LLMs) frequently generate defective outputs in code generation tasks, ranging from logical bugs to security vulnerabilities. While these generation failures are often treated as model-level limitations, empirical evidence increasingly traces their root causes to imperfections within the training corpora. Yet, the specific mechanisms linking training data quality issues to generated code quality issues remain largely unmapped. This paper presents a systematic literature review of 114 primary studies to investigate how training data quality issues propagate into code generation. We establish a unified taxonomy that categorizes generated code quality issues across nine dimensions and training data quality issues into code and non-code attributes. Based on this taxonomy, we formalize a causal framework detailing 18 typical propagation mapping mechanisms. Furthermore, we synthesize state-of-the-art detection and mitigation techniques across the data, model, and generation lifecycles. The reviewed literature reveals a clear methodological shift: quality assurance is transitioning from reactive, heuristic-based post-generation filtering toward proactive, data-centric governance and closed-loop repair. Finally, we identify open challenges and outline research directions for developing reliable LLMs for code through integrated data curation and continuous evaluation. Our repository is available at this https URL .\n\n---\n\n## 146. Decision-aware User Simulation Agent for Evaluating Conversational Recommender Systems\n- **Authors**: Yuan-Chi Li , Li-Chi Chen , Sung-Yi Wu , Yu-Che Tsai , Shou-De Lin\n- **URL**: [https://arxiv.org/abs/2605.05250](https://arxiv.org/abs/2605.05250)\n- **Abstract**:\n> Conversational recommender systems (CRS) increasingly rely on user simulators for automated evaluation of sales agents. A key requirement for such simulators is the ability to model human decision-making. However, most existing simulation frameworks do not explicitly model the internal decision process, and LLM-based simulators often exhibit unrealistically strong information-processing capabilities, rarely exhibit the hesitation or decision deferral commonly observed in real consumer behavior, resulting in overly high acceptance probabilities. To address this limitation, we propose Hesitator, a theory-grounded user simulation framework that explicitly models human decision-making under choice overload. The framework introduces a modular Decision Module that separates utility-based item selection from overload-aware commitment decisions. Experiments across multiple user simulation frameworks, domains, sales modes, and LLM backbones show that integrating our module consistently mitigates unrealistic behaviors under increasing overload conditions. Furthermore, Hesitator reproduces established behavioral patterns from psychological economics, demonstrating its ability to model human decision behavior.\n\n---\n\n## 147. Towards Dependable Retrieval-Augmented Generation Using Factual Confidence Prediction\n- **Authors**: Florian Geissler , Francesco Carella , Laura Fieback , Jakob Spiegelberg\n- **URL**: [https://arxiv.org/abs/2605.05244](https://arxiv.org/abs/2605.05244)\n- **Abstract**:\n> Incorporating specific knowledge into large language models via retrieval-augmented generation (RAG) is a widespread technique that fuels many of today's industry AI applications. A fundamental problem is to assess if the context retrieved by some similarity search provides indeed supporting facts, or instead misguides the generator with irrelevant information. It is critical to associate meaningful confidence measures about the factuality of the retrieval process with the generated answers. We present a new, two-staged approach to predict fact faithfulness of the output of retrieval-augmented generations. First, we employ conformal prediction to select only those retrieved chunks who have a high chance to come from the correct source. This approach in itself can improve answer quality by up to 6% in some of the studied datasets, however, the associated statistical guarantees do not hold generally, since the assumption of sample exchangeability depends on the retriever setup. We present diagnostic metrics to assess whether a setup is suitable. Second, we quantify confidence in the consistency of a generated final answer with a given retrieved context, using an attention-based factuality classifier. This approach can detect inconsistent answers with a chance of up to 77%. Our work helps to establish a novel type of certified RAG systems for a broad range of natural language industry applications.\n\n---\n\n## 148. Rethinking Data Curation in LLM Training: Online Reweighting Offers Better Generalization than Offline Methods\n- **Authors**: Wanru Zhao , Yihong Chen , Yuzhi Tang , Wentao Ma , Shengchao Hu , Shell Xu Hu , Alex Iacob , Abhinav Mehrotra , Nicholas D. Lane\n- **URL**: [https://arxiv.org/abs/2605.05227](https://arxiv.org/abs/2605.05227)\n- **Abstract**:\n> Data curation is a critical yet under-explored area in large language model (LLM) training. Existing methods, such as data selection and mixing, operate in an offline paradigm, detaching themselves from training. This separation introduces engineering overhead and makes the curation brittle: the entire pipeline must be re-run under model/task shifts. Moreover, offline methods alter data size through hard filtering or resampling, often sacrificing data diversity and harming generalization. We propose to rethink data curation as an online reweighting problem, where sample importance is dynamically adjusted during training via loss weighting rather than static pre-processing. Specifically, we introduce ADAPT (Adaptive Data reweighting for Pretraining and FineTuning), a dynamic online framework that reweights training samples with adaptive per-sample learning rates guided by similarity-based quality signals, without changing the number of training samples. Unlike offline methods that enforce a static data distribution, ADAPT acts as an implicit curriculum learner, progressively shifting focus from coarse-grained patterns to fine-grained semantic distinctions as the model evolves. Experiments on both instruction tuning and large-scale pretraining show that ADAPT consistently outperforms offline selection/mixing and prior online methods, achieving stronger cross-benchmark generalization under equal FLOPs.\n\n---\n\n## 149. MACS: Modality-Aware Capacity Scaling for Efficient Multimodal MoE Inference\n- **Authors**: Bo Li , Chuan Wu , shaolin Zhu\n- **URL**: [https://arxiv.org/abs/2605.05225](https://arxiv.org/abs/2605.05225)\n- **Abstract**:\n> Mixture-of-Experts Multimodal Large Language Models (MoE MLLMs) suffer from a significant efficiency bottleneck during Expert Parallelism (EP) inference due to the straggler effect. This issue is worsened in the multimodal context, as existing token-count-based load balancing methods fail to address two unique challenges: (1) Information Heterogeneity, where numerous redundant visual tokens are treated equally to semantically critical ones, and (2) Modality Dynamics, where varying visual to text ratios across tasks lead to resource misallocation. To address these challenges, we propose MACS (Modality-Aware Capacity Scaling), a training-free inference framework. Specifically, MACS introduces an Entropy-Weighted Load mechanism to quantify the semantic value of visual tokens, addressing information heterogeneity. Additionally, the Dynamic Modality-Adaptive Capacity mechanism allocates expert resources based on the real-time modal composition of the input. Extensive experiments demonstrate that MACS significantly outperforms existing methods on various multimodal benchmarks, providing a novel and robust solution for the efficient deployment of MoE MLLMs in EP inference.\n\n---\n\n## 150. MidSteer: Optimal Affine Framework for Steering Generative Models\n- **Authors**: Tatiana Gaintseva , Andrew Stepanov , Ziquan Liu , Martin Benning , Gregory Slabaugh , Jiankang Deng , Ismail Elezi\n- **URL**: [https://arxiv.org/abs/2605.05220](https://arxiv.org/abs/2605.05220)\n- **Abstract**:\n> Steering intermediate representations has emerged as a powerful strategy for controlling generative models, particularly in post-deployment alignment and safety settings. However, despite its empirical success, it currently lacks a comprehensive theoretical framework. In this paper, we bridge this gap by formalizing the theory of concept steering. First, we establish a link between steering and affine concept erasure, proving that the standard approach for removing unwanted behaviors is a special case of LEACE (a closed-form method for affine erasure). Next, we formulate a principled theoretical framework for concept switching, LEACE-Switch, and characterize the assumptions under which it provides an optimal affine solution. Building on this analysis, we then introduce MidSteer (Minimal Disturbance concept Steering), a more general affine framework for concept manipulation that relaxes these assumptions and enables directed, minimal-disturbance transformations. We demonstrate that MidSteer performs favorably across a range of tasks, modalities, and architectures, including vision diffusion models and large language models.\n\n---\n\n## 151. Sparse Prefix Caching for Hybrid and Recurrent LLM Serving\n- **Authors**: Mikhail Shirokikh , Sergey Nikolenko\n- **URL**: [https://arxiv.org/abs/2605.05219](https://arxiv.org/abs/2605.05219)\n- **Abstract**:\n> Prefix caching is a key latency optimization for autoregressive LLM serving, yet existing systems assume dense per-token key/value reuse. State-space models change the structure of the problem: a recurrent layer can resume from a single stored state rather than requiring the entire token history. This asymmetry opens a new design point between no reuse and dense caching: store exact recurrent states at a sparse set of checkpoint positions and, on a cache hit, resume from the deepest stored checkpoint and recompute the remaining suffix exactly. We formalize sparse prefix caching as checkpoint placement under a distribution over overlap depths, yielding an exact O(NM) dynamic program. For use cases where requests share a non-trivial prefix (e.g. asking different questions about a single long document), we show that our method consistently improves the Pareto frontier traced by standard heuristics on real-world data. Across QuALITY and System Prompts, distribution-aware placement dominates every fixed-budget baseline on the measured layer-group Pareto frontier and matches or outperforms the strongest heuristic (block caching) while typically using substantially fewer checkpoints, with the largest gains at low checkpoint budgets where the overlap distribution is most non-uniform. The method is most relevant when many requests share a substantial but not identical prefix within a retained cache entry. It preserves exact outputs, does not change the recurrent computation itself or require new recurrent update kernels, applies to recurrent/SSM layers whose hidden state can be extracted and restored exactly, and for hybrid models can be combined with existing KV-cache compression techniques.\n\n---\n\n## 152. A Review of Large Language Models for Stock Price Forecasting from a Hedge-Fund Perspective\n- **Authors**: Olivia Zhang , Zhilin Zhang\n- **URL**: [https://arxiv.org/abs/2605.05211](https://arxiv.org/abs/2605.05211)\n- **Abstract**:\n> Large language models (LLMs) are increasingly deployed in quantitative finance for stock price forecasting. This review synthesizes recent applications of LLMs in this domain, including extracting sentiment from financial news and social media, analyzing financial reports and earnings-call transcripts, tokenizing or symbolizing stock price series, and constructing multi-agent trading systems. Particular attention is paid to practical pitfalls that are often understated in the literature, such as fragility in sentiment analysis, dataset and horizon design, performance evaluation metrics, data leakage, illiquidity premia, and limits of stock price predictability. Organized from a hedge-fund perspective, the review is intended to guide both academic researchers and hedge fund managers in integrating LLMs into real-world trading pipelines and in stress-testing their robustness under realistic market frictions.\n\n---\n\n{% endraw %}\n"}, "c234ff9921f9be8d378a7558752264c977290646": {"t": 1786068094.3612225, "n": 263185, "h": "c0b4ce7dc3328d6b", "c": "---\nlayout: default\ntitle: \"LLM 관련 주요 논문 - 2026-05-08\"\n---\n\n{% raw %}\n# LLM 관련 주요 논문 - 2026-05-08\n\n## 1. GlazyBench: A Benchmark for Ceramic Glaze Property Prediction and Image Generation\n- **Authors**: Ziyu Zhai , Siyou Li , Juexi Shao , Juntao Yu\n- **URL**: [https://arxiv.org/abs/2605.06641](https://arxiv.org/abs/2605.06641)\n- **Abstract**:\n> Developing ceramic glazes is a costly, time-consuming process of trial and error due to complex chemistry, placing a significant burden on independent artists. While recent advances in multimodal AI offer a modern solution, the field lacks the large-scale datasets required to train these models. We propose GlazyBench, the first dataset for AI-assisted glaze design. Comprising 23,148 real glaze formulations, GlazyBench supports two primary tasks: predicting post-firing surface properties, such as color and transparency, from raw materials, and generating accurate visual representations of the glaze based on these properties. We establish comprehensive baselines for property prediction using traditional machine learning and large language models, alongside image generation benchmarks using deep generative and large multimodal models. Our experiments demonstrate promising yet challenging results. GlazyBench pioneers a new research direction in AI-assisted material design, providing a standardized benchmark for systematic evaluation.\n\n---\n\n## 2. Can RL Teach Long-Horizon Reasoning to LLMs? Expressiveness Is Key\n- **Authors**: Tianle Wang , Zhaoyang Wang , Guangchen Lan , Xinpeng Wei , Sipeng Zhang , Guanwen Qiu , Abulhair Saparov\n- **URL**: [https://arxiv.org/abs/2605.06638](https://arxiv.org/abs/2605.06638)\n- **Abstract**:\n> Reinforcement learning (RL) has been applied to improve large language model (LLM) reasoning, yet the systematic study of how training scales with task difficulty has been hampered by the lack of controlled, scalable environments. We introduce ScaleLogic, a synthetic logical reasoning framework that offers independent control over two axes of difficulty: the depth of the required proof planning (i.e., the horizon) and the expressiveness of the underlying logic. Our proposed framework supports a wide range of logics: from simple implication-only logic (\"if-then\") towards more expressive first-order reasoning with conjunction (\"and\"), disjunction (\"or\"), negation (\"not\"), and universal quantification (\"for all\"). Using this framework, we show that the RL training compute $T$ follows a power law with respect to reasoning depth $D$ ($T \\propto D^{\\gamma}$, $R^{2} > 0.99$), and that the scaling exponent $\\gamma$ increases monotonically with logical expressiveness, from $1.04$ to $2.60$. On downstream mathematics and general reasoning benchmarks, more expressive training settings yield both larger performance gains (up to $+10.66$ points) and more compute-efficient transfer compared to less expressive settings, demonstrating that what a model is trained on, not just how much it is trained, shapes downstream transfer. We further show that the power-law relationship holds across multiple RL methods, and curriculum-based training substantially improves scaling efficiency.\n\n---\n\n## 3. MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems\n- **Authors**: Zhexuan Wang , Xuebo Liu , Li Wang , Zifei Shan , Yutong Wang , Zhenxi Song , Min Zhang\n- **URL**: [https://arxiv.org/abs/2605.06623](https://arxiv.org/abs/2605.06623)\n- **Abstract**:\n> Large language model (LLM)-based Multi-agent systems (MAS) have shown promise in tackling complex collaborative tasks, where agents are typically orchestrated via role-specific prompts. While the quality of these prompts is pivotal, jointly optimizing them across interacting agents remains a non-trivial challenge, primarily due to the misalignment between local agent objectives and holistic system goals. To address this, we introduce MASPO, a novel framework designed to automatically and iteratively refine prompts across the entire system. A core innovation of MASPO is its joint evaluation mechanism, which assesses prompts not merely by their local validity, but by their capacity to facilitate downstream success for successor agents. This effectively bridges the gap between local interactions and global outcomes without relying on ground-truth labels. Furthermore, MASPO employs a data-driven evolutionary beam search to efficiently navigate the high-dimensional prompt space. Extensive empirical evaluations across 6 diverse tasks demonstrate that MASPO consistently outperforms state-of-the-art prompt optimization methods, achieving an average accuracy improvement of 2.9. We release our code at this https URL .\n\n---\n\n## 4. SkillOS: Learning Skill Curation for Self-Evolving Agents\n- **Authors**: Siru Ouyang , Jun Yan , Yanfei Chen , Rujun Han , Zifeng Wang , Bhavana Dalvi Mishra , Rui Meng , Chun-Liang Li , Yizhu Jiao , Kaiwen Zha , Maohao Shen , Vishy Tirumalashetty , George Lee , Jiawei Han , Tomas Pfister , Chen-Yu Lee\n- **URL**: [https://arxiv.org/abs/2605.06614](https://arxiv.org/abs/2605.06614)\n- **Abstract**:\n> LLM-based agents are increasingly deployed to handle streaming tasks, yet they often remain one-off problem solvers that fail to learn from past interactions. Reusable skills distilled from experience provide a natural substrate for self-evolution, where high-quality skill curation serves as the key bottleneck. Existing approaches either rely on manual skill curation, prescribe heuristic skill operations, or train for short-horizon skill operations. However, they still struggle to learn complex long-term curation policies from indirect and delayed feedback. To tackle this challenge, we propose SkillOS, an experience-driven RL training recipe for learning skill curation in self-evolving agents. SkillOS pairs a frozen agent executor that retrieves and applies skills with a trainable skill curator that updates an external SkillRepo from accumulated experience. To provide learning signals for curation, we design composite rewards and train on grouped task streams based on skill-relevant task dependencies, where earlier trajectories update the SkillRepo, and later related tasks evaluate these updates. Across multi-turn agentic tasks and single-turn reasoning tasks, SkillOS consistently outperforms memory-free and strong memory-based baselines in both effectiveness and efficiency, with the learned skill curator generalizing across different executor backbones and task domains. Further analyses show that the learned curator produces more targeted skill use, while the skills in SkillRepo evolve into more richly structured Markdown files that encode higher-level meta-skills over time.\n\n---\n\n## 5. NeuroAgent: LLM Agents for Multimodal Neuroimaging Analysis and Research\n- **Authors**: Lujia Zhong , Yihao Xia , Jianwei Zhang , Shuo huang , Jiaxin Yue , Mingyang Xia , Yonggang Shi\n- **URL**: [https://arxiv.org/abs/2605.06584](https://arxiv.org/abs/2605.06584)\n- **Abstract**:\n> Multimodal neuroimaging analysis often involves complex, modality-specific preprocessing workflows that require careful configuration, quality control, and coordination across heterogeneous toolchains. Beyond preprocessing, downstream statistical analysis and disease classification commonly require task-specific code, evaluation protocols, and data-format conventions, creating additional barriers between raw acquisitions and reproducible scientific analysis. We present NeuroAgent, an LLM-driven agentic framework that automates key preprocessing and analysis steps for heterogeneous neuroimaging data, including sMRI, fMRI, dMRI, and PET, and supports interactive downstream analysis through natural-language queries. NeuroAgent employs a hierarchical multi-agent architecture with a feedback-driven Generate-Execute-Validate engine: agents autonomously generate executable preprocessing code, detect and recover from runtime errors, and validate output integrity. We evaluate the system on 1,470 subjects pooled across all ADNI phases (CN=1,000, AD=470), where all subjects have sMRI and tabular data, with subsets also having Tau-PET (n=469), fMRI (n=278), and DTI ($n=620$). Pipeline ablation studies across multiple LLM backends show that capable models reach up to 100% intent-parsing accuracy, with the strongest backend (Qwen3.5-27B) reaching 84.8% end-to-end preprocessing step correctness. Automated recovery limits manual intervention to edge cases where human review is required via the Human-In-The-Loop interface. For Alzheimer's Disease classification using automatically preprocessed multimodal data, our agent ensemble achieves an AUC of 0.9518 with four modalities, outperforming all single-modality baselines. These results show that NeuroAgent can reduce the manual effort required for neuroimaging preprocessing and enable end-to-end automated analysis pipelines for neuroimaging research.\n\n---\n\n## 6. Process Matters more than Output for Distinguishing Humans from Machines\n- **Authors**: Milena Rmus , Mathew D. Hardy , Thomas L. Griffiths , Mayank Agrawal\n- **URL**: [https://arxiv.org/abs/2605.06524](https://arxiv.org/abs/2605.06524)\n- **Abstract**:\n> Reliable human-machine discrimination is becoming increasingly important as large language models and autonomous agents are deployed in online settings. Existing approaches evaluate whether a system can produce behavior or responses indistinguishable from those of a human, following the emphasis on outputs as a criterion for intelligence proposed by Alan Turing. Cognitive science offers an alternative perspective: evaluating the process by which behavior is produced. To test whether cognitive processes can reliably distinguish humans from machines, we introduce CogCAPTCHA30, a battery of 30 cognitive tasks designed to elicit diagnostic process-level features even when task performance is matched. Across the battery, process-level features provide stronger discriminative signal than performance metrics alone, reliably distinguishing humans from agents even under output matching (mean process-feature classifier AUC = 0.88). To evaluate agentic process differences, we compare off-the-shelf frontier agents (Claude Sonnet 4.5, GPT-5, Gemini 2.5 Pro), Centaur (a language model fine-tuned on 10.7M human decisions), and two task-specific fine-tuning approaches applied to Qwen2.5-1.5B-Instruct: action-level supervised fine-tuning (A-SFT) and process-level fine-tuning (P-SFT), which directly optimizes process features. Broad fine-tuning on human decisions improves human-like task processes relative to off-the-shelf agents, while task-specific process-level supervision further improves behavioral mimicry. However, this advantage diminishes under cross-task transfer when supervised process targets do not naturally generalize across tasks. Explicit process-level supervision can improve human behavioral mimicry, but only if appropriate task-specific process representations are available, highlighting process specification as a bottleneck for achieving human-like cognitive processes in machines.\n\n---\n\n## 7. Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors\n- **Authors**: Jonas Wiedermann-Möller , Leonard Dung , Maksym Andriushchenko\n- **URL**: [https://arxiv.org/abs/2605.06490](https://arxiv.org/abs/2605.06490)\n- **Abstract**:\n> AI systems have become increasingly capable of dangerous behaviours in many domains. This raises the question: Do models sometimes choose to violate human instructions in order to perform behaviour that is more useful for certain goals? We introduce a benchmark for measuring model propensity for instrumental convergence (IC) behaviour in terminal-based agents. This is behaviour such as self-preservation that has been hypothesised to play a key role in risks from highly capable AI agents. Our benchmark is realistic and low-stakes which serves to reduce evaluation-awareness and roleplay confounds. The suite contains seven operational tasks, each with an official workflow and a policy-violating shortcut. An eight-variant shared framework varies monitoring, instruction clarity, stakes, permission, instrumental usefulness and blocked honest paths to support inferences regarding the factors driving IC behaviour. We evaluated ten models using deterministic environment-state scorers over 1,680 samples, with trace review employed for audit and adjudication purposes. The final IC rate is 86 out of 1,680 samples (5.1%). IC behaviour is concentrated rather than uniform: two Gemini models account for 66.3% of IC cases and three tasks account for 84.9%. Conditions in which IC behaviour is indispensable for task success result in the greatest increase in the adjusted IC rate (+15.7 percentage points), whereas emphasising that task success is critical or certain framing choices do not produce comparable effects. Our findings indicate that realistic, low-nudge environments elicit IC behaviour rarely but systematically in most tested models. We conclude that it is feasible to robustly measure tendencies for dangerous behaviour in current frontier AI agents.\n\n---\n\n## 8. ReasonSTL: Bridging Natural Language and Signal Temporal Logic via Tool-Augmented Process-Rewarded Learning\n- **Authors**: Bowen Ye , Zhijian Li , Junyue Huang , Junkai Ma , Xiang Yin\n- **URL**: [https://arxiv.org/abs/2605.06483](https://arxiv.org/abs/2605.06483)\n- **Abstract**:\n> Signal Temporal Logic (STL) is an expressive formal language for specifying spatio-temporal requirements over real-valued, real-time signals. It has been widely used for the verification and synthesis of autonomous systems and cyber-physical systems. In practice, however, users often express their requirements in natural language rather than in structured STL formulas, making natural-language-to-STL translation a critical yet challenging task. Manual specification requires temporal-logic expertise and cannot scale, while prompting commercial LLM APIs incurs substantial token costs and may expose sensitive system requirements to third-party services, raising privacy concerns for industrial deployment. To address these challenges, we present \\textsc{ReasonSTL}, a tool-augmented framework that adapts local open-source language models for natural-language-to-STL generation. \\textsc{ReasonSTL} decomposes the translation process into explicit reasoning, deterministic tool calls, and structured formula construction. We further introduce process-rewarded training to supervise both tool-use trajectories and final formulas, together with \\textsc{STL-Bench}, a bilingual, computation-aware benchmark grounded in real-world signals. Experiments show that a 4B model trained with \\textsc{ReasonSTL} achieves state-of-the-art performance in both automatic metrics and human evaluations, demonstrating that \\textsc{ReasonSTL} provides a transparent, low-cost, and privacy-preserving alternative for formal specification drafting.\n\n---\n\n## 9. Patch-Effect Graph Kernels for LLM Interpretability\n- **Authors**: Ruben Fernandez-Boullon , David N. Olivieri\n- **URL**: [https://arxiv.org/abs/2605.06480](https://arxiv.org/abs/2605.06480)\n- **Abstract**:\n> Mechanistic interpretability aims to reverse-engineer transformer computations by identifying causal circuits through activation patching. However, scaling these interventions across diverse prompts and task families produces high-dimensional, unstructured datasets that are difficult to compare systematically. We propose a framework that reframes mechanistic analysis as a graph machine-learning problem by representing activation-patching profiles as patch-effect graphs over model components. We introduce three graph-construction methods: direct-influence via causal mediation, partial-correlation, and co-influence and apply graph kernels to analyze the resulting structures. Evaluating this approach on GPT-2 Small using Indirect Object Identification (IOI) and related tasks, we find that patch-effect graphs preserve discriminative structural signals. Specifically, localized edge-slot features provide higher classification accuracy than global graph-shape descriptors. A screened paired-patching validation suggests that CI and PC selected candidate edges correspond to stronger activation-influence effects than random or low-rank candidates. Crucially, by evaluating these representations against rigorous prompt-only and raw patch-effect controls, we make the evidential scope of the benchmark explicit: graph features compress structured patching signal, while raw tensors and surface cues define strong baselines that any circuit-level claim should address. Ultimately, our framework provides a compression and evaluation pipeline for comparing patching-derived structures under controlled baselines, separating robust slice-discriminative evidence from stronger task-general causal-circuit claims.\n\n---\n\n## 10. Beyond Task Success: Measuring Workflow Fidelity in LLM-Based Agentic Payment Systems\n- **Authors**: Donghao Huang , Joon Kiat Chua , Zhaoxia Wang\n- **URL**: [https://arxiv.org/abs/2605.06457](https://arxiv.org/abs/2605.06457)\n- **Abstract**:\n> LLM-based multi-agent systems are increasingly deployed for payment workflows, yet prevailing metrics, Task Success Rate (TSR) and Agent Handoff F1-Score (HF1), capture only final outcomes or unordered routing decisions. We introduce the Agentic Success Rate (ASR), a trajectory-fidelity metric that compares observed and expected agent execution sequences at the transition level, decomposing performance into Transition Recall and Transition Precision. Applied to the Hierarchical Multi-Agent System for Payments (HMASP) across 18 LLMs and 90,000 task instances, ASR reveals that 10 of 18 models systematically skip a confirmation checkpoint during payment checkout, a deviation invisible to both TSR and HF1, while 8 models enforce the checkpoint perfectly. Notably, GPT-4.1 exhibits hidden workflow shortcuts despite achieving perfect TSR and HF1, while GPT-5.2 achieves perfect ASR. Prompt refinements and deterministic routing guards guided by ASR diagnostics yield substantial TSR improvements, with gains up to +93.8 percentage points for previously struggling models, demonstrating that trajectory-level evaluation is essential in regulated domains.\n\n---\n\n## 11. PrefixGuard: From LLM-Agent Traces to Online Failure-Warning Monitors\n- **Authors**: Xinmiao Huang , Jinwei Hu , Rajarshi Roy , Changshun Wu , Yi Dong , Xiaowei Huang\n- **URL**: [https://arxiv.org/abs/2605.06455](https://arxiv.org/abs/2605.06455)\n- **Abstract**:\n> Large language model (LLM) agents now execute long, tool-using tasks where final outcome checks can arrive too late for intervention. Online warning requires lightweight prefix monitors over heterogeneous traces, but hand-authored event schemas are brittle and deployment-time LLM judging is costly. We introduce PrefixGuard, a trace-to-monitor framework with an offline StepView induction step followed by supervised monitor training. StepView induces deterministic typed-step adapters from raw trace samples, and the monitor learns an event abstraction and prefix-risk scorer from terminal outcomes. Across WebArena, $\\tau^2$-Bench, SkillsBench, and TerminalBench, the strongest PrefixGuard monitors reach 0.900/0.710/0.533/0.557 AUPRC. Using the strongest backend within each representation, they improve over raw-text controls by an average of +0.137 AUPRC. LLM judges remain substantially weaker under the same prefix-warning protocol. We also derive an observability ceiling on score-based area under the precision-recall curve (AUPRC) that separates monitor error from failures lacking evidence in the observed prefix. For finite-state audit, post-hoc deterministic finite automaton (DFA) extraction remains compact on WebArena and $\\tau^2$-Bench (29 and 20 states) but expands to 151 and 187 states on SkillsBench and TerminalBench. Finally, first-alert diagnostics show that strong ranking does not imply deployment utility: WebArena ranks well yet fails to support low-false-alarm alerts, whereas $\\tau^2$-Bench and TerminalBench retain more actionable early alerts. Together, these results position PrefixGuard as a practical monitor-synthesis recipe with explicit diagnostics for when prefix warnings translate into actionable interventions.\n\n---\n\n## 12. SCRuB: Social Concept Reasoning under Rubric-Based Evaluation\n- **Authors**: Jamelle Watson-Daniels , Himaghna Bhattacharjee , Skyler Wang , Brandon Handoko , Antonio Li , Anaelia Ovalle , Mahesh Pasupuleti , Candace Ross , Vidya Sarma , Arjun Subramonian , Karen Ullrich , Will van der Vaart , Yijing Xin , Maximilian Nickel\n- **URL**: [https://arxiv.org/abs/2605.06444](https://arxiv.org/abs/2605.06444)\n- **Abstract**:\n> While many studies of Large Language Model (LLM) reasoning capabilities emphasize mathematical or technical tasks, few address reasoning about social concepts: the abstract ideas shaping social norms, culture, and institutions. This understudied capability is essential for modern models acting as social agents, yet no systematic evaluation methodology targets it. We introduce SCRuB (Social Concept Reasoning under Rubric-Based Evaluation), a framework designed for this setting of task indeterminacy. Our goal is to measure the degree to which a model reasons about social concepts with the depth and critical rigor of a human expert. SCRuB proceeds in three phases: prompt construction from established sources, response generation by experts and models, and comparative evaluation using a five-dimensional critical thinking rubric. To enable generalization of the pipeline, we introduce a Panel of Disciplinary Perspectives ensemble validated against independent expert judges. We release SCRuBEval (n=4,711 evaluation prompts) and SCRuBAnnotations (300 expert-authored responses and 150 expert comparative judgments from 45 PhD-level scholars). Our results show that frontier models consistently outperform human experts across all five rubric dimensions. Across 1,170 pairwise comparisons, expert judges ranked a model response first in 80.8% of judgments and preferred model responses overall 74.4% of the time. Ultimately, this study provides the first expert-grounded demonstration of evaluation saturation for social concept reasoning: the single-turn exam-style format has reached its ceiling for models and humans alike.\n\n---\n\n## 13. Knowledge Graphs, the Missing Link in Agentic AI-based Formal Verification\n- **Authors**: Vaisakh Naduvodi Viswambharan , Keerthan Kopparam Radhakrishna , Deepak Narayan Gadde , Aman Kumar\n- **URL**: [https://arxiv.org/abs/2605.06434](https://arxiv.org/abs/2605.06434)\n- **Abstract**:\n> Recent advances in Large Language Models (LLMs) have enabled workflows that generate SystemVerilog Assertions (SVAs) from natural-language specifications, with the potential to accelerate Formal Verification (FV). However, high-quality assertion synthesis remains challenging because specifications are often ambiguous or incomplete and critical micro-architectural details reside in the Register Transfer Level (RTL). Many existing approaches treat the specification and RTL as loosely structured text, which weakens specification-to-RTL grounding and leads to semantic mismatches and frequent syntax failures during formal parsing and elaboration. This work addresses these limitations with a verification-centric Knowledge Graph (KG) constructed from structured Intermediate Representations (IRs) extracted from the specification, RTL, and formal-tool feedback, including syntax diagnostics, Counterexamples (CEXs), and coverage reports. The KG links requirements, design hierarchy, signals, assumptions, and properties to provide traceable, design-grounded context for generation. A multi-agent workflow queries and updates this KG to generate SVAs and to drive three refinement loops: syntax repair guided by tool diagnostics, CEX-guided correction using trace links, and coverage-directed property augmentation. Evaluation across seven benchmark designs indicates that KG-based context retrieval improves specification-to-RTL grounding and consistently produces compilable SVAs with low syntax-repair overhead. The approach achieves formal coverage ranging from 78.5% to 99.4%, though convergence exhibits design dependence with complex temporal and arithmetic reasoning remaining challenging for current LLM capabilities.\n\n---\n\n## 14. Rethinking Vacuity for OOD Detection in Evidential Deep Learning\n- **Authors**: Claire McNamara\n- **URL**: [https://arxiv.org/abs/2605.06382](https://arxiv.org/abs/2605.06382)\n- **Abstract**:\n> Vacuity, or Uncertainty Mass (UM), is commonly used as a metric to evaluate Out-of-Distribution (OOD) detection in Evidential Deep Learning (EDL). It generally involves dividing the number of classes ($K$) by the total strength of belief ($S$) of the model's predictions, where $S$ is derived from summing the Dirichlet parameters. As such, UM is sensitive to the cardinality of $K$. In particular, it is unlikely in practice that there is a linear relationship between $K$ and $S$ as $K$ and $S$ increase due to the nature of EDL (suppressing incorrectly assigned evidence). As a result, when comparing In Distribution (ID) and OOD results, it is important that $K_{\\mathrm{ID} }$ and $K_{\\mathrm{OOD} }$ are equal; something that is not always ensured in practice. We provide an empirical demonstration of how results for AUROC and AUPR can substantially differ when class cardinality between ID and OOD differs by 1, with AUROC differing by as much as 0.318 and AUPR by 0.613 for standard EDL, and AUROC by 0.360 and AUPR by 0.683 for IB-EDL. More concretely, our findings isolate an evaluation artefact: when K differs between ID and OOD, AUROC/AUPR can be artificially inflated without any change in model predictions. We further discuss the evaluation of EDL over causal language models using Multiple-Choice Question-Answer (MCQA) datasets and argue for clearer definitions of ID and OOD in this context. Our primary contribution is an empirical and theoretical demonstration that vacuity-based OOD detection in EDL-fine-tuned LLMs is highly sensitive to uncontrolled differences in evaluated class cardinality.\n\n---\n\n## 15. From Agent Loops to Deterministic Graphs: Execution Lineage for Reproducible AI-Native Work\n- **Authors**: Josh Rosen , Seth Rosen\n- **URL**: [https://arxiv.org/abs/2605.06365](https://arxiv.org/abs/2605.06365)\n- **Abstract**:\n> Large language model systems are increasingly deployed as agentic workflows that interleave reasoning, tool use, memory, and iterative refinement. These systems are effective at producing answers, but they often rely on implicit conversational state, making it difficult to preserve stable work products, isolate irrelevant updates, or propagate changes through intermediate artifacts. We introduce execution lineage: an execution model in which AI-native work is represented as a directed acyclic graph (DAG) of artifact-producing computations with explicit dependencies, stable intermediate boundaries, and identity-based replay. The goal is not to make the model a better one-shot writer, but to make evolving AI-generated work maintainable under change. We compare execution-lineage replay against loop-centric update baselines on two controlled policy-memo update tasks. In an unrelated-branch update, DAG replay preserved the final memo exactly in all runs, with zero churn and zero unrelated-branch contamination, while loop baselines regenerated the memo and frequently imported unrelated context. In an intermediate-artifact edit, all systems reflected the new constraint in the final memo, but only DAG replay achieved perfect upstream preservation, downstream propagation, unaffected-artifact preservation, and cross-artifact consistency. These results show that final answer quality and maintained-state quality are distinct. Strong loop baselines can remain competitive at producing polished final outputs when the task is a bounded synthesis/update problem and all current sources fit in context, but immediate task success can mask partial state inconsistency that may compound over future revisions. Execution lineage provides stronger guarantees about what should change, what should remain stable, and how work evolves across revisions.\n\n---\n\n## 16. A Regime Theory of Controller Class Selection for LLM Action Decisions\n- **Authors**: Zhaoyang Jiang , Zhizhong Fu , Yunsoo Kim , Jiacong Mi , Zicheng Li , Xuanqi Peng , Honghan Wu\n- **URL**: [https://arxiv.org/abs/2605.06339](https://arxiv.org/abs/2605.06339)\n- **Abstract**:\n> Deployed language and vision-language models must decide, on each input, whether to answer directly, retrieve evidence, defer to a stronger model, or abstain. Contrary to the common monotonicity intuition, greater per-input expressivity is not uniformly beneficial in finite samples: under identical strict cross-validation, different benchmarks prefer different controller classes. This reflects a finite-sample limitation of instance-level uncertainty signals, which can be exhausted at a distribution-dependent scale. We organize controllers into a nested lattice of four classes: fixed actions, partition routers, instance-level controllers, and prior-gated controllers, ordered by complexity. We prove a regime theory that turns three data-estimable bottlenecks into a class choice: how much improvement is possible beyond the best fixed action, whether there are enough samples for instance-level controllers to make reliable decisions, and how much improvement a coarse partition router can recover when instance-level signal is unreliable. The resulting Bernstein-tight threshold has a matching information-theoretic lower bound, and strict nested cross-validation provably selects a near-best class. Across SMS-Spam, HallusionBench, A-OKVQA, and FOLIO, the predicted class matches the empirical winner; the prior-gated controller wins on TextVQA when OCR tokens supply a label-free prediction-time prior. Code is available at this https URL .\n\n---\n\n## 17. Addressing Labelled Data Scarcity: Taxonomy-Agnostic Annotation of PII Values in HTTP Traffic using LLMs\n- **Authors**: Thomas Cory , Axel Küpper\n- **URL**: [https://arxiv.org/abs/2605.06305](https://arxiv.org/abs/2605.06305)\n- **Abstract**:\n> Automated privacy audits of web and mobile applications often analyse outbound HTTP traffic to detect Personally Identifiable Information (PII) leakage. However, existing learning-based detectors typically depend on scarce, manually labelled traffic and are tightly coupled to fixed label taxonomies, limiting transferability across domains and evolving definitions of PII. This paper investigates whether Large Language Models (LLMs) can support taxonomy-agnostic annotation of explicitly transmitted PII values in HTTP message bodies when the taxonomy is provided at runtime. We introduce a multi-stage LLM-based pipeline that combines deterministic pre-processing with label-level classification, targeted instance-level value annotation, and output validation. To enable controlled evaluation and exemplar-based prompting without relying on sensitive real-user captures, we further propose an LLM-based generator for synthetic HTTP traffic with manually validated, taxonomy-derived PII annotations. We evaluate the approach across three taxonomies spanning different PII domains and granularity levels. Results show that the pipeline accurately detects PII types and extracts corresponding values for concrete PII taxonomies. Overall, our findings position LLMs as a promising foundation for flexible, taxonomy-agnostic traffic annotation and for creating labelled data under evolving privacy taxonomies.\n\n---\n\n## 18. Data Language Models: A New Foundation Model Class for Tabular Data\n- **Authors**: Eda Erol , Giuliano Pezzoli , Ozer Cem Kelahmet\n- **URL**: [https://arxiv.org/abs/2605.06290](https://arxiv.org/abs/2605.06290)\n- **Abstract**:\n> Every major data modality now has a foundation model that understands it natively: text has language models, images have vision models, audio has audio models. Tabular data, the modality on which many consequential real-world AI decisions are made, does not. Every approach to tabular AI today, from gradient-boosted trees to the latest tabular foundation models, requires a preprocessing pipeline before any model can consume the data. None of them understand tabular data as a modality. We introduce the Data Language Model (DLM), the missing foundation model for tabular data. A DLM understands tables the way a language model understands sentences: natively, without serialization or preprocessing, directly from raw cell values. It is the tabular data layer on which AI models, agents, and vertical AI applications can be built, eliminating the preprocessing pipelines that currently stand between raw data and every AI system that consumes it. We present Schema-1, the first DLM: a 140M parameter model trained on more than 2.3M synthetic and real-world tabular datasets. Schema-1 outperforms gradient-boosted ensembles, AutoML stacks, and the tabular foundation models we evaluate on established row-level prediction benchmarks. On missing value reconstruction it achieves lower reconstruction error than all classical statistical methods and frontier large language models on mean performance across conditions, establishing that structural understanding of a dataset's own distributional geometry is more useful for imputation than world knowledge encoded in language. It identifies the industry sector of any unseen dataset from raw cell values alone, reliably across any domain, a task no prior tabular model can perform. It is the native tabular understanding layer that has been missing from the AI stack.\n\n---\n\n## 19. Joint Consistency: A Unified Test-Time Aggregation Framework via Energy Minimization\n- **Authors**: Yunzhen Yao , Hongye Wang , Yahong Wang , Michael C. Gastpar , Bo Jiang , Lie He\n- **URL**: [https://arxiv.org/abs/2605.06219](https://arxiv.org/abs/2605.06219)\n- **Abstract**:\n> This paper studies test-time aggregation, an approach that generates multiple reasoning traces and aggregates them into a final answer. Most existing methods rely on evaluation signals collected from candidate traces in isolation or answer frequencies, while ignoring comparative interactions among candidates. We propose Joint Consistency (JC), formulated as a constrained Ising-type energy minimization problem, where independent evaluation signals act as external fields and pairwise comparisons act as interactions. JC provides a unified framework for test-time aggregation that subsumes existing voting and weighted aggregation methods as special cases. Our construction of the interaction matrix leverages LLM-as-a-judge comparisons, and admits a theoretical interpretation under answer-level homogeneity assumptions. Moreover, we develop an efficient approximation strategy that makes interaction modeling practical for large-scale test-time aggregation. Experiments on math and code reasoning benchmarks show that JC consistently outperforms existing baselines across tasks, judge models, trace budgets, and trace-generation settings.\n\n---\n\n## 20. Beyond Fixed Benchmarks and Worst-Case Attacks: Dynamic Boundary Evaluation for Language Models\n- **Authors**: Haoxiang Wang , Da Yu , Huishuai Zhang\n- **URL**: [https://arxiv.org/abs/2605.06213](https://arxiv.org/abs/2605.06213)\n- **Abstract**:\n> Evaluating large language models (LLMs) today rests on fixed benchmarks that apply the same set of items to any model, producing ceiling and floor effects that mask capability gaps. We argue that the most informative evaluation signal lies at the boundary, where the per-prompt pass probability is near $0.5$ under random-sampling decoding, and propose Dynamic Boundary Evaluation (DBE), which actively locates each model's boundary and places it on a globally comparable difficulty scale. DBE delivers three artifacts: (i) a calibrated item bank covering safety, capability, and truthfulness, with per-item difficulty labels validated across $9$ reference LLMs; (ii) Skill-Guided Boundary Search (SGBS), a search algorithm that finds boundary items for a given target LLM using only API-level query access; and (iii) an evaluation protocol that places a new LLM on a unified ability scale and grows the evaluation set adaptively when the target falls outside the bank's coverage. We instantiate DBE on four categories spanning safety (harmful request refusal and over-refusal), capability (constrained instruction following), and truthfulness (multi-turn sycophancy resistance). The resulting evaluation covers a broader model spectrum without saturation while remaining compatible with existing datasets.\n\n---\n\n## 21. Towards Annotation-Free Validation of MLLMs: A Vision-Language Logical Consistency Metric\n- **Authors**: Ying Gu , Mei Chee Leong , Hui Li Tan , Shangbo Mao , Liyuan Li , Nancy Chen\n- **URL**: [https://arxiv.org/abs/2605.06201](https://arxiv.org/abs/2605.06201)\n- **Abstract**:\n> Dominant accuracy evaluation might reward unwarranted guessing of Large Language Models, and it might not be applicable to novel tasks for model validation without ground-truth (gt) annotation. Based on basic logic principle, we propose a novel framework to evaluate the vision-language logical consistency of MLLMs on both sufficient and necessary cause-effect relations. We define Vision-Language Logical Consistency Metric (VL-LCM) on traditional MC-VQA tests, and recent NaturalBench tests without the need for gt annotation. Through systematic experiments on representative VL benchmark MMMU and recent VL challenges like NaturalBench, we evaluated 11 recent open-source MLLMs from 4 frontier families. Our findings reveal that, despite significant progress of recent MLLMs on accuracy, logical consistency lags behind significantly. Extensive evaluations on the correlations of VL-LCM with metrics on gt, the reliability of LCM, and the relation of VL-LCM with response distribution justify the validity and applicability of VL-LCM even without gt annotation. Our findings suggest that, beyond accuracy, logical consistency could be employed for both accuracy and reliability. VL-LCM can also be employed for MLLM selection, validation, and reliable answer justification in novel tasks without gt annotation.\n\n---\n\n## 22. The Granularity Axis: A Micro-to-Macro Latent Direction for Social Roles in Language Models\n- **Authors**: Chonghan Qin , Xiachong Feng , Ziyun Song , Xiaocheng Feng , Jing Xiong , Lingpeng Kong\n- **URL**: [https://arxiv.org/abs/2605.06196](https://arxiv.org/abs/2605.06196)\n- **Abstract**:\n> Large language models (LLMs) are routinely prompted to take on social roles ranging from individuals to institutions, yet it remains unclear whether their internal representations encode the granularity of such roles, from micro-level individual experience to macro-level organizational, institutional, or national reasoning. We show that they do. We define a contrast-based Granularity Axis as the difference between mean macro- and micro-role hidden states. In Qwen3-8B, this axis aligns with the principal axis (PC1) of the role representation space at cosine 0.972 and accounts for 52.6% of its variance, indicating that granularity is the dominant geometric axis organizing prompted social roles. We construct 75 social roles across five granularity levels and collect 91,200 role-conditioned responses over shared questions and prompt variants, then extract role-level hidden states and project them onto the axis. Role projections increase monotonically across all five levels, remain stable across layers, prompt variants, endpoint definitions, held-out splits, and score-filtered subsets, and transfer to Llama-3.1-8B-Instruct. The axis is also causally relevant: activation steering along it shifts response granularity in the predicted direction, with Llama moving from 2.00 to 3.17 on a five-point macro scale under positive steering on prompts that admit local responses. The two models differ in controllability, suggesting that steering depends on each model's default operating regime. Overall, our findings suggest that social role granularity is not merely a stylistic surface feature, but a structured, ordered, and causally manipulable latent direction in role-conditioned language model behavior.\n\n---\n\n## 23. Systematic Evaluation of Large Language Models for Post-Discharge Clinical Action Extraction\n- **Authors**: Shivali Dalmia , Ananya Mantravadi , Prasanna Desikan\n- **URL**: [https://arxiv.org/abs/2605.06191](https://arxiv.org/abs/2605.06191)\n- **Abstract**:\n> The work in this paper evaluates zero-shot and few-shot large language models (LLMs) for safety-critical clinical action extraction using the CLIP discharge-note dataset, with particular emphasis on transitions of care and post-discharge patient safety. To manage the complexity of clinical documentation, we introduce a two-stage extraction framework that decomposes discharge notes, that are written in narrative form, into fine-grained, explicitly actionable clinical tasks through a staged prompting strategy. Our contributions include a systematic assessment of generative LLMs for clinical action extraction, a detailed comparison between general-purpose LLMs and task-specific supervised BERT-based models, and an analysis of annotation inconsistencies across different action categories. We show that contemporary LLMs achieve performance comparable to or exceeding supervised models on binary actionability detection, while supervised baselines retain a meaningful advantage on fine-grained multi-label category classification, despite the absence of task-specific fine-tuning and under strict data-privacy constraints. Qualitative error analysis reveals that many failures stem from misalignment between model reasoning and dataset annotation conventions, particularly in cases involving implicit clinical actions and rigid structural labeling rules. These results indicate that reported performance reflects model limitations due to lack of clinical reasoning, that is not captured by plain annotations. Labels without rationales make it impossible to distinguish clinical reasoning failures from annotation convention mismatches. Advancing clinical NLP requires reasoning-annotated datasets that document why specific spans are actionable, not merely which spans were labeled, enabling proper evaluation of model clinical understanding.\n\n---\n\n## 24. Event-Causal RAG: A Retrieval-Augmented Generation Framework for Long Video Reasoning in Complex Scenarios\n- **Authors**: Peizheng Yan , Yu Zhao , Liang Xie , Juntong Qi , Mingming Wang , Erwei Yin\n- **URL**: [https://arxiv.org/abs/2605.06185](https://arxiv.org/abs/2605.06185)\n- **Abstract**:\n> Recent large vision-language models have achieved strong performance on short- and medium-length video understanding, yet they remain inadequate for ultra-long or even infinite video reasoning, where models must preserve coherent memory over extended durations and infer causal dependencies across temporally distant events. Existing end-to-end video understanding methods are fundamentally limited by the $O(n^2)$ complexity of self-attention, while recent retrieval-augmented generation (RAG) approaches still suffer from fragmented clip-level memory, weak modeling of temporal and causal structure, and high storage and online inference costs. We present Event-Causal RAG, a lightweight retrieval-augmented framework for infinite long-video reasoning. Instead of indexing fixed-length clips, our method segments streaming videos into semantically coherent events and represents each event as a structured State-Event-State (SES) graph, capturing the event together with its surrounding state transitions. These graphs are merged into a global Event Knowledge Graph and stored in a dual-store memory that supports both semantic matching and causal-topological retrieval. On top of this memory, we design a bidirectional retrieval strategy to efficiently identify the most relevant event causal chains and provide them, together with the associated video evidence, to a backbone video foundation model for answer generation. Experiments on long-video understanding benchmarks demonstrate that Event-Causal RAG consistently outperforms strong clip-based retrieval baselines and long-context video models, particularly on questions requiring multi-event integration and causal inference across long temporal gaps, while also achieving improved memory efficiency and robust streaming performance.\n\n---\n\n## 25. Post Reasoning: Improving the Performance of Non-Thinking Models at No Cost\n- **Authors**: Richmond Sin Jing Xuan , Rishabh Bhardwaj , Soujanya Poria\n- **URL**: [https://arxiv.org/abs/2605.06165](https://arxiv.org/abs/2605.06165)\n- **Abstract**:\n> As the widespread adoption of Large Language Models (LLMs) accelerates, token consumption from intermediate reasoning traces increasingly contributes to inference latency and operational cost. Recent studies suggest that many real-world tasks require little to no explicit reasoning, with additional reasoning sometimes even degrading performance. In this work, we propose \\textbf{Post-Reasoning}, a simple yet effective approach that improves instruction-tuned models by conditioning them to justify their answers after generating the final response. By design, it enables the final answer to be obtained without additional latency or token cost, while still improving performance through simple instruction augmentation. We evaluate Post-Reasoning across \\(117\\) model--benchmark settings spanning \\(13\\) open and proprietary models, \\(4\\) model families, and \\(9\\) diverse reasoning and knowledge-intensive benchmarks, including AMC, HMMT, GSM8K, GPQA, MMLU-Pro, and BIG-Bench Hard. Post-Reasoning improves performance in over \\(88.19\\%\\) of evaluated settings, achieving a mean relative improvements of \\(17.37\\%\\). Furthermore, we propose supervised post-reason tuning, which further improves performance in over \\(91.11\\%\\) of evaluated settings, and exceeds the prompt-based post-reasoning baseline by an average of \\(8.01\\%\\), demonstrating that post-reasoning can be effectively internalized through training. Ultimately, Post-Reasoning establishes a new performance ceiling for direct-answer capabilities.\n\n---\n\n## 26. Beyond Accuracy: Policy Invariance as a Reliability Test for LLM Safety Judges\n- **Authors**: Shihao Weng , Yang Feng , Xiaofei Xie\n- **URL**: [https://arxiv.org/abs/2605.06161](https://arxiv.org/abs/2605.06161)\n- **Abstract**:\n> LLM-as-a-Judge pipelines have become the de facto evaluator for agent safety, yet existing benchmarks treat their verdicts as ground-truth proxies without checking whether the verdicts depend on the agent's behavior or merely on how the evaluation policy happens to be worded. We argue that any trustworthy safety judge must satisfy a basic property we call policy invariance, and we operationalize it as three testable principles: rubric-semantics invariance under certified-equivalent rewrites, rubric-threshold invariance under intentional strict-to-lenient shifts, and ambiguity-aware calibration so that verdict instability concentrates on genuinely ambiguous cases. Instantiating these principles as a stress-test protocol with four agent-class judges on trajectories drawn from ASSEBench and R-Judge, we surface a previously unmeasured failure mode: today's judges respond to meaningful normative shifts and to meaningless structural rewrites with comparable strength, and cannot tell the two apart. Content-preserving policy rewrites flip up to 9.1% of verdicts above baseline jitter, and 18-43% of all observed flips occur on unambiguous cases under such rewrites, so existing safety scores conflate what the agent did with how the evaluator was prompted. Beyond the diagnosis, we contribute the Policy Invariance Score and the Judge Card reporting protocol, which expose an order-of-magnitude spread in judge reliability that is invisible to accuracy-only leaderboards. We release the protocol and code so that future agent-safety benchmarks can audit their own evaluators rather than trust them by default.\n\n---\n\n## 27. Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning\n- **Authors**: Yaorui Shi , Yuxin Chen , Zhengxi Lu , Yuchun Miao , Shugui Liu , Qi GU , Xunliang Cai , Xiang Wang , An Zhang\n- **URL**: [https://arxiv.org/abs/2605.06130](https://arxiv.org/abs/2605.06130)\n- **Abstract**:\n> A persistent skill library allows language model agents to reuse successful strategies across tasks. Maintaining such a library requires three coupled capabilities. The agent selects a relevant skill, utilizes it during execution, and distills new skills from experience. Existing methods optimize these capabilities in isolation or with separate reward sources, resulting in partial and conflicting evolution. We propose Skill1, a framework that trains a single policy to co-evolve skill selection, utilization, and distillation toward a shared task-outcome objective. The policy generates a query to search the skill library, re-ranks candidates to select one, solves the task conditioned on it, and distills a new skill from the trajectory. All learning derives from a single task-outcome signal. Its low-frequency trend credits selection and its high-frequency variation credits distillation. Experiments on ALFWorld and WebShop show that Skill1 outperforms prior skill-based and reinforcement learning baselines. Training dynamics confirm the co-evolution of the three capabilities, and ablations show that removing any credit signal degrades the evolution.\n\n---\n\n## 28. Back to the Beginning of Heuristic Design: Bridging Code and Knowledge with LLMs\n- **Authors**: Nguyen Viet Tuan Kiet , Bui Dinh Pham , Dao Van Tung , Tran Cong Dao , Huynh Thi Thanh Binh\n- **URL**: [https://arxiv.org/abs/2605.06123](https://arxiv.org/abs/2605.06123)\n- **Abstract**:\n> Large language models (LLMs) have recently advanced automatic heuristic design (AHD) for combinatorial optimization (CO), where candidate heuristics are iteratively proposed, evaluated, and refined. Most existing approaches search over executable programs and distill insights from execution feedback to guide later iterations. Because this process moves from low-level implementations to high-level principles, we refer to it as a bottom-up paradigm. We argue that this view is incomplete and introduce a complementary top-down perspective: knowledge becomes the primary search object and code merely instantiates and tests it, making what is learned explicit and reusable across problems and trajectories. We formalize this shift through a statistical-learning view that exposes a distortion--compression trade-off, and instantiate it in both population-based and tree-based AHD frameworks. Across CO and tasks beyond it, knowledge-first search improves discovery efficiency, transfer, and generalization, often outperforming code-centric pipelines, while combining both strategies yields further gains. Our results suggest that progress in AHD depends on iteratively constructing and evolving interpretable hypotheses that retain value beyond a single search trajectory.\n\n---\n\n## 29. Policy-Guided Stepwise Model Routing for Cost-Effective Reasoning\n- **Authors**: Wenwen Si , Insup Lee , Osbert Bastani\n- **URL**: [https://arxiv.org/abs/2605.06116](https://arxiv.org/abs/2605.06116)\n- **Abstract**:\n> Inference-time computation has greatly enhanced the performance of large language models (LLMs) on challenging reasoning tasks, but this strategy can incur high inference costs. One solution is to route intermediate chain-of-thought (CoT) states to language models of different sizes; however, existing approaches rely on handcrafted routing strategies that limit performance, or on training large process reward models that may be infeasible in many applications. We formulate stepwise model routing as a constrained decision-making problem, which we solve by training a small control policy using reinforcement learning in conjunction with threshold calibration to tune the performance-efficiency tradeoff. We validate our method on three math benchmarks (GSM8K, MATH500, and OmniMath) on both open and closed models. Our method consistently improves the accuracy-cost tradeoff compared to handcrafted approaches, while achieving a comparable tradeoff to methods that require training large process reward models.\n\n---\n\n## 30. CrossCult-KIBench: A Benchmark for Cross-Cultural Knowledge Insertion in MLLMs\n- **Authors**: Zhen Zeng , Leijiang Gu , Feng Li , Jing Yu , Zenglin Shi\n- **URL**: [https://arxiv.org/abs/2605.06115](https://arxiv.org/abs/2605.06115)\n- **Abstract**:\n> Multimodal Large Language Models (MLLMs), trained primarily on English-centric data, frequently generate culturally inappropriate or misaligned responses in cross-cultural settings. To mitigate this, we introduce the task of cross-cultural knowledge insertion, which focuses on adapting models to specific cultural contexts while preserving their original behavior in other cultures. To facilitate research in this area, we introduce CrossCult-KIBench, a comprehensive evaluation benchmark for assessing both the effectiveness of knowledge insertion and its unintended side effects on non-target cultures. The benchmark includes 9,800 image-grounded cases covering 49 culturally relevant visual scenarios across English, Chinese, and Arabic language-culture groups. It supports evaluation in both single-insert and sequential-insert settings. We also propose Memory-Conditioned Knowledge Insertion (MCKI) as a baseline method. MCKI retrieves relevant cultural knowledge from an external memory using frozen MLLM representations, prepending matched entries as conditional prompts when applicable. Extensive experiments on CrossCult-KIBench reveal that current approaches struggle to balance effective cultural adaptation with behavioral preservation, highlighting a key challenge in developing culturally-aware MLLMs. Our work thus underscores an important research direction for developing more culturally adaptive and responsible MLLMs.\n\n---\n\n## 31. Shallow Prefill, Deep Decoding: Efficient Long-Context Inference via Layer-Asymmetric KV Visibility\n- **Authors**: Jungsuk Oh , Hyeseo Jeon , Hyunjune Ji , Kyongmin Kong , Jay-Yoon Lee\n- **URL**: [https://arxiv.org/abs/2605.06105](https://arxiv.org/abs/2605.06105)\n- **Abstract**:\n> Long-context inference in decoder-only language models is costly because long prompts are processed during Prefill, cached at every layer, and repeatedly attended to during autoregressive Decode. We introduce \\emph{Shallow Prefill, dEEp Decode} (SPEED), a phase-asymmetric KV-visibility policy that materializes non-anchor prompt-token KV states only in lower layers while keeping Decode-phase tokens full-depth. Unlike previous approaches that make upper-layer prompt KV states cheaper to store or construct, SPEED removes prefill tokens from the upper-layer Decode visibility set altogether. With a minimal BoS anchor, this simple change preserves broad benchmark quality while reducing long-context cost. In a controlled Llama-3.1-8B instruction-tuning study, SPEED using only 75\\% of layers for prefill tokens reaches 51.2 average score on OLMES-style benchmarks, compared with 51.4 for the full-depth baseline, while improving TTFT by 33\\%, TPOT by 22\\%, and reducing active KV memory by 25.0\\% at 128K context. Layer-wise diagnostics suggest that this cutoff retains the main prompt-selection and representation-stabilization regions of the full-depth model. These results show that long-context prompt tokens need not always persist as full-depth KV-cache objects when Decode-phase tokens remain full-depth.\n\n---\n\n## 32. VibeServe: Can AI Agents Build Bespoke LLM Serving Systems?\n- **Authors**: Keisuke Kamahori , Shihang Li , Simon Peter , Baris Kasikci\n- **URL**: [https://arxiv.org/abs/2605.06068](https://arxiv.org/abs/2605.06068)\n- **Abstract**:\n> For years, we have built LLM serving systems like any other critical infrastructure: a single general-purpose stack, hand-tuned over many engineer-years, meant to support every model and workload. In this paper, we take the opposite bet: a multi-agent loop that automatically synthesizes bespoke serving systems for different usage scenarios. We propose VibeServe, the first agentic loop that generates entire LLM serving stacks end-to-end. VibeServe uses an outer loop to plan and track the search over system designs, and an inner loop to implement candidates, check correctness, and measure performance on the target benchmark. In the standard deployment setting, where existing stacks are highly optimized, VibeServe remains competitive with vLLM, showing that generation-time specialization need not come at the cost of performance. More interestingly, in non-standard scenarios, VibeServe outperforms existing systems by exploiting opportunities that generic systems miss in six scenarios involving non-standard model architectures, workload knowledge, and hardware-specific optimizations. Together, these results suggest a different point in the design space for infrastructure software: generation-time specialization rather than runtime generality. Code is available at this https URL .\n\n---\n\n## 33. Visual Fingerprints for LLM Generation Comparison\n- **Authors**: Amal Alnouri , Andreas Hinterreiter , Christina Humer , Furui Cheng , Marc Streit\n- **URL**: [https://arxiv.org/abs/2605.06054](https://arxiv.org/abs/2605.06054)\n- **Abstract**:\n> Large language model (LLM) outputs arise from complex interactions among prompts, system instructions, model parameters, and architecture. We refer to specific configurations of these factors as generation conditions, each of which can bias outputs in various ways. Understanding how different generation conditions shape model behaviors is essential for tasks such as prompt design and model evaluation, yet it remains challenging due to the stochastic and open-ended nature of text generation. We present an approach to visually compare LLM outputs across generation conditions by modeling responses as collections of linguistic choices, including content, expression, and structure. We extract these choices using natural language processing pipelines and represent their distributions across repeated samples. We then visualize these distributions as visual fingerprints, enabling direct, distribution-level comparison of condition-specific tendencies. Through four usage scenarios, we demonstrate how visual fingerprints reveal consistent patterns in LLM behavior that are difficult to observe through individual responses or aggregate metrics.\n\n---\n\n## 34. Novelty-based Tree-of-Thought Search for LLM Reasoning and Planning\n- **Authors**: Leon Hamm , Zlatan Ajanovic\n- **URL**: [https://arxiv.org/abs/2605.06040](https://arxiv.org/abs/2605.06040)\n- **Abstract**:\n> Although advances such as chain-of-thought, tree-of-thought or reinforcement learning have improved the performance of LLMs in reasoning and planning tasks, they are still brittle and have not achieved human-level performance in many domains, and often suffer from high time and token costs. Inspired by the success of width-based search in planning, we explore how the concept of novelty can be transferred to language domains and how it can improve tree-of-thought reasoning. A tree of thoughts relies on building possible \"paths\" of consecutive ideas or thoughts. These are generated by repeatedly prompting an LLM. In our paper, a measurable concept of novelty is proposed that describes the uniqueness of a new node (thought) in comparison to nodes previously seen in the search tree. Novelty is estimated by prompting an LLM and making use of embedded general knowledge from pre-training. This metric can then be used to prune branches and reduce the scope of the search. Although this method introduces more prompts per state, the overall token cost can be reduced by pruning and reducing the overall tree size. This procedure is tested and compared using several benchmarks in language-based planning and general reasoning.\n\n---\n\n## 35. Strat-LLM: Stratified Strategy Alignment for LLM-based Stock Trading with Real-time Multi-Source Signals\n- **Authors**: Wenliang Huang , Zengyi Yu\n- **URL**: [https://arxiv.org/abs/2605.06024](https://arxiv.org/abs/2605.06024)\n- **Abstract**:\n> Large Language Models (LLMs) are evolving into autonomous trading agents, yet existing benchmarks often overlook the interplay between architectural reasoning and strategy consistency. We propose Strat-LLM, a framework grounded in Stratified Strategy Alignment. Operating in a live-forward setting throughout 2025, it integrates heterogeneous data including sequential prices, real-time news, and annual reports to eliminate look-ahead bias. Extensive stress tests on A-share and U.S. markets reveal: (1) reasoning-heavy models achieve peak utility in Free Mode via internal logic, whereas standard models require Strict Mode as a vital risk anchor; (2) alignment utility is regime-dependent, with Free and Guided modes capturing momentum in uptrending markets, while Strict Mode mitigates drawdowns in downtrends; (3) mid-scale models (35B) show optimal fidelity under strict constraints, whereas ultra-large models (122B) suffer an alignment tax under rigid rules but gain a performance premium in Guided Mode; (4) standard LLMs often fall into a high win-rate trap, optimizing for small gains at the expense of total returns, which can only be mitigated through deep reasoning or strict external guardrails. Project details are available at this https URL .\n\n---\n\n## 36. TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steering\n- **Authors**: Yuan Sui , Yulin Chen , Yibo Li , Xue Jiang , Yufei He , Yihong Dong , Xiaoxin He , Tianyu Gao , Bryan Hooi\n- **URL**: [https://arxiv.org/abs/2605.05980](https://arxiv.org/abs/2605.05980)\n- **Abstract**:\n> When language model agents tackle complex software engineering tasks, they often degrade over long trajectories, which we define as *agent drift*. We focus on two recurring failure modes *overthinking* and *overacting*, i.e., where the agent repeatedly reasons over information it already has, and where it issues tool calls without integrating recent observations or acquiring new evidence. In this paper, we introduce TACT (Think-Act Calibration via activation Steering), to detect and mitigate agent drift in the residual stream before it surfaces as a behavioral failure. In specific, we label trajectory steps as overthinking, overacting, or calibrated, and find that their hidden states can separate linearly along two *drift axes*, pointing from calibrated behavior toward each failure mode (AUC $\\approx$ 0.9). To mitigate agent drift, we project each step's activation onto these axes at test time and pull drifted ones back toward the calibrated region. Experiments show that TACT outperforms unsteered baselines across SWE-bench Verified, Terminal-Bench 2.0, and CLAW-Eval, lifting average resolve rate by $+5.8$ pp on Qwen3.5-27B and $+4.8$ pp on Gemma-4-26B-A4B-it while cutting steps-to-resolve by up to $26\\%$. These gains frame agent drift as a steerable direction in the residual stream, and position TACT as a viable handle for reliable long-horizon agents.\n\n---\n\n## 37. TheraAgent: Self-Improving Therapeutic Agent for Precise and Comprehensive Treatment Planning\n- **Authors**: Junkai Li , Yunghwei Lai , Tianyi Zhu , Zheng Long Lee , Weizhi Ma , Yang Liu\n- **URL**: [https://arxiv.org/abs/2605.05963](https://arxiv.org/abs/2605.05963)\n- **Abstract**:\n> Formulating a treatment plan is inherently a complex reasoning and refinement task rather than a simple generation problem. However, existing large language models (LLMs) mainly rely on one-shot output without explicit verification, which may result in rough, incomplete, and potentially unsafe treatment plans. To address these limitations, we propose TheraAgent, an agentic framework that replaces one-shot generation with an iterative generate-judge-refine pipeline. By mirroring the actual reasoning process of human experts who iteratively revise treatment plans, our framework progressively transforms coarse and incomplete drafts into precise, comprehensive, and safer therapeutic regimens. To facilitate the critical judge component, we introduce TheraJudge, a treatment-specific evaluation module integrated into the inference loop to enforce clinical standards. Experiments show TheraAgent achieves state-of-the-art results on HealthBench, leading in Accuracy and Completeness. In expert evaluations, it attains an 86% win rate against physicians, with superior Targeting and Harm Control. Moreover, the highly agreement between TheraJudge and HealthBench evaluations confirms the reliability of our framework.\n\n---\n\n## 38. ICU-Bench:Benchmarking Continual Unlearning in Multimodal Large Language Models\n- **Authors**: Yuhang Wang , Wenjie Mei , Junkai Zhang , Guangyu He , Zhenxing Niu , Haichang Gao\n- **URL**: [https://arxiv.org/abs/2605.05938](https://arxiv.org/abs/2605.05938)\n- **Abstract**:\n> Although Multimodal Large Language Models (MLLMs) have achieved remarkable progress across many domains, their training on large-scale multimodal datasets raises serious privacy concerns, making effective machine unlearning increasingly necessary. However, existing benchmarks mainly focus on static or short-sequence settings, offering limited support for evaluating continual privacy deletion requests in realistic deployments. To bridge this gap, we introduce ICU-Bench, a continual multimodal unlearning benchmark built on privacy-critical document data. ICU-Bench contains 1,000 privacy-sensitive profiles from two document domains, medical reports and labor contracts, with 9,500 images, 16,000 question-answer pairs, and 100 forget tasks. Additionally, new continual unlearning metrics are introduced, facilitating a comprehensive analysis of forgetting effectiveness, historical forgetting preservation, retained utility, and stability throughout the continual unlearning process. Through extensive experiments with representative unlearning methods on ICU-Bench, we show that existing methods generally struggle in continual settings and exhibit clear limitations in balancing forgetting quality, utility preservation, and scalability over long task sequences. These findings highlight the need for multimodal unlearning methods explicitly designed for continual privacy deletion.\n\n---\n\n## 39. Wisteria: A Unified Multi-Scale Feature Learning Framework for DNA Language Model\n- **Authors**: Weihua Wang , Haoji Li , Feilong Bao , Lei Yang , Guanglai Gao\n- **URL**: [https://arxiv.org/abs/2605.05913](https://arxiv.org/abs/2605.05913)\n- **Abstract**:\n> DNA language model aims to decipher the regulatory grammar and semantic of genomes by capturing long range dependencies in DNA sequences. Existing methods emphasize long range token interactions but often ignore the interplay between local motifs and global dependencies. In this paper, we propose Wisteria, a genomic language model that integrates multi scale feature learning within a unified framework for DNA sequence. Specifically, Wisteria augments the Mamba based architecture with gated dilated convolutions to capture local motifs and regulatory patterns, while gated multilayer perceptrons refine global dependencies. We further introduce a Fourier based attention mechanism to support frequency domain modeling, periodic extension and length generalization. Across four experimental settings with both short and long range dependencies, Wisteria demonstrates strong performance on downstream benchmarks against competitive DNA language model baselines. These results indicate that Wisteria effectively unifies local and global dependency modeling for multi scale genomic sequence analysis.\n\n---\n\n## 40. Null Space Constrained Contrastive Visual Forgetting for MLLM Unlearning\n- **Authors**: Yuhang Wang , Zhenxing Niu , Haoxuan Ji , Guangyu He , Linlin Zhang , Haichang Gao\n- **URL**: [https://arxiv.org/abs/2605.05909](https://arxiv.org/abs/2605.05909)\n- **Abstract**:\n> The core challenge of machine unlearning is to strike a balance between target knowledge removal and non-target knowledge retention. In the context of Multimodal Large Language Models (MLLMs), this challenge becomes even more pronounced, as knowledge is further divided into visual and textual modalities that are tightly intertwined. In this paper, we introduce an MLLM unlearning approach that aims to forget target visual knowledge while preserving non-target visual knowledge and all textual knowledge. Specifically, we freeze the LLM backbone and achieve unlearning by fine-tuning the visual module. First, we propose a Contrastive Visual Forgetting (CVF) mechanism to separate target visual knowledge from retained visual knowledge, guiding the representations of target visual concepts toward appropriate regions in the feature space. Second, we identify the null space associated with retained knowledge and constrain the unlearning process within this space, thereby significantly mitigating degradation in knowledge retention. Third, beyond static unlearning scenarios, we extend our approach to continual unlearning, where forgetting requests arrive sequentially. Extensive experiments across diverse benchmarks demonstrate that our approach achieves a strong balance between effective forgetting and robust knowledge retention.\n\n---\n\n## 41. Taklif.AI: LLM-Powered Platform for Interest-Based Personalized College Assignments\n- **Authors**: Zaki Kurdya , Mohammed Zuqlam , Salem Amassi , Shady Telbany , Motaz Saad\n- **URL**: [https://arxiv.org/abs/2605.05842](https://arxiv.org/abs/2605.05842)\n- **Abstract**:\n> Educators face significant challenges in creating engaging, personalized assignments that accommodate students' diverse interests and cognitive abilities. Traditional one-size-fits-all assignments frequently lead to decreased student engagement and increased reliance on unethical practices such as plagiarism. To address these challenges, we present this http URL , a platform that leverages Large Language Models (LLMs) to automatically generate personalized assignments tailored to individual student interests. Unlike existing AI-powered educational platforms that personalize based on academic performance metrics alone, this http URL incorporates students' extracurricular interests and cultural contexts into the assignment generation process through a structured prompt engineering pipeline with input and output guardrails. The platform employs a serverless architecture on AWS with this http URL , using Llama 3.3 70B as the primary LLM via LiteLLM for multi-provider load balancing and LangChain for prompt orchestration. We describe the system architecture, the prompt design methodology, and the guardrails framework that ensures output quality. Preliminary user acceptance testing with 68 participants (65 students and 3 educators) indicates positive reception, with 84% of participants rating the personalization feature as beneficial. We discuss the platform's current capabilities and limitations, and outline directions for rigorous empirical evaluation of learning outcomes.\n\n---\n\n## 42. On the Role of Language Representations in Auto-Bidding: Findings and Implications\n- **Authors**: Guanyu Zhu , Jining Luan , Hanwen Du , Xinyu Fang , Sibo Xu , Ersheng Ni , Hongji Li , Jincheng Fang , Ronghao Chen , Huacan Wang , Xuanqi Lan , Yongxin Ni , Yiqi Sun , Youhua Li\n- **URL**: [https://arxiv.org/abs/2605.05833](https://arxiv.org/abs/2605.05833)\n- **Abstract**:\n> Auto-bidding is a crucial task in real-time advertising markets, where policies must optimize long-horizon value under delivery constraints (e.g., budget and CPA). Existing methods for auto-bidding rely on compact numerical state representations: while they can implicitly capture delivery dynamics, they offer limited support for explicitly representing and controlling high-level intent, evolving feedback, and operator-style strategic guidance in real campaigns. Meanwhile, Large Language Models (LLMs) offer a powerful method for encoding semantic information, it remains unclear when LLMs help and how to integrate them without sacrificing numerical precision. Through systematic preliminary studies, we find that (1) LLM embeddings contain bidding-relevant cues yet cannot replace numerical features, and (2) gains emerge only with careful semantic--numeric integration rather than naive concatenation. Motivated by these findings, we propose \\textit{SemBid}, a novel auto-bidding framework that injects LLM-encoded semantics into offline bidding trajectories at the token level. SemBid introduces three semantic inputs: \\textit{Task}, \\textit{History}, and \\textit{Strategy}. It injects these semantics as tokens alongside numerical trajectory tokens and uses self-attention to integrate them, improving controllability and generalization across objectives. Across diverse scenarios and budget regimes, SemBid outperforms competitive baselines from offline RL and generative sequence modeling, with more consistent gains in overall performance, constraint satisfaction, and robustness. Our code is available at: \\href{ this https URL }{\\textcolor{blue}{here} }.\n\n---\n\n## 43. AGPO: Asymmetric Group Policy Optimization for Verifiable Reasoning and Search Ads Relevance at JD\n- **Authors**: Yang Xu , Kun Yao , Yiming Deng , Zheng Fang , Kai Ming Ting , Ming Pang\n- **URL**: [https://arxiv.org/abs/2605.05826](https://arxiv.org/abs/2605.05826)\n- **Abstract**:\n> Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated notable success in enhancing the reasoning performance of large language models (LLMs). However, recent studies reveal that while current RLVR methods improve sampling efficiency towards correct paths, they do not elicit fundamentally new reasoning patterns. Instead, the reasoning capability boundary of trained models often narrows compared to their base models, with base models achieving higher coverage at large sample sizes. In this work, we propose Asymmetric Group Policy Optimization (AGPO) to counteract this boundary shrinkage. AGPO adopts a negative-dominant reinforcement strategy to suppress incorrect reasoning paths, maintaining the base model's exploration capacity. For positive reinforcement, AGPO adopts a group advantage mechanism, which scales positive updates based on intra-group variance, allowing the model to focus on rare correct paths while suppressing updates from trivial paths. Our experiments on five mathematical benchmarks demonstrate that AGPO achieves state-of-the-art accuracy while consistently improving pass@$k$ performance at scale. In a large-scale industrial application for search ads relevance optimization, AGPO effectively enhances the quality of the data annotation, leading to substantial performance gains in downstream student models.\n\n---\n\n## 44. CircuitFormer: A Circuit Language Model for Analog Topology Design from Natural Language Prompt\n- **Authors**: Md Touhidul Islam , Sujan Kumar Saha , Farimah Farahmandi , Mark Tehranipoor\n- **URL**: [https://arxiv.org/abs/2605.05773](https://arxiv.org/abs/2605.05773)\n- **Abstract**:\n> Automating analog circuit design remains a longstanding challenge in Electronic Design Automation (EDA). While Transformer-based Large Language Models (LLMs) have revolutionized software code generation, their application to analog hardware design is hindered by two critical limitations: (i) the scarcity of analog design datasets containing natural language description of a design and its corresponding netlist, and (ii) the inefficiency of general-purpose tokenizers (e.g., Byte Pair Encoding (BPE)) in capturing the inherent graph structure of circuits. To bridge this gap, first, we curate the largest annotated dataset of analog circuit netlists to date, comprising 31,341 netlist-natural language description pairs across all major circuit classes. Furthermore, we propose Circuit Tokenizer (CKT), a novel circuit graph tokenizer designed to encode netlist connectivity by explicitly mining frequent subcircuits. In terms of scalability, CKT overcomes the bottleneck of prior circuit graph serialization methods where vocabulary size scales linearly with maximum number of components in the dataset, n_max, (O(n_max)); instead, CKT decouples vocabulary growth from circuit complexity, achieving a constant O(1) complexity. Empirically, CKT outperforms standard BPE on circuit topology representation, reducing sequence length by 57% and achieving a 2.3x superior compression ratio using a compact, fixed vocabulary of size 512. Leveraging this optimized tokenization, we train a circuit-specific language model, CircuitFormer, a 511M parameter encoder-decoder transformer. Our model achieves 100% syntactic correctness and an 83% functional success rate across all major analog circuit categories, outperforming state-of-the-art open-source LLMs by 10% and 14%, respectively, while requiring 240x fewer parameters. The dataset is publicly available at this https URL .\n\n---\n\n## 45. HyperLens: Quantifying Cognitive Effort in LLMs with Fine-grained Confidence Trajectory\n- **Authors**: Chengda Lu , Xiaoyu Fan , Wei Xu\n- **URL**: [https://arxiv.org/abs/2605.05741](https://arxiv.org/abs/2605.05741)\n- **Abstract**:\n> While Large Language Models (LLMs) achieve strong performance across diverse tasks, their inference dynamics remain poorly understood because of the limited resolution of existing analysis tools. In this work, we identify an intrinsic magnification mechanism in transformer architectures: deeper layers inherently magnify the small changes of layer-wise confidence, providing a fine-grained confidence trajectory. Building on this insight, we introduce HyperLens, a high-resolution probe designed to trace confidence trajectories and quantify the cognitive effort during inference. Across LLMs and datasets, HyperLens reveals a consistent divergence in confidence trajectories that separates complex from simple tasks. We abstract this pattern into a quantitative cognitive effort metric. Our analysis reveals a fundamental principle: complex tasks consistently require higher cognitive effort. Finally, we provide a mechanistic diagnosis of a common side effect of standard Supervised Fine-Tuning (SFT): it can reduce cognitive effort and consequently degrade performance on in-domain tasks.\n\n---\n\n## 46. ReFlect: An Effective Harness System for Complex Long-Horizon LLM Reasoning\n- **Authors**: Fan Huang\n- **URL**: [https://arxiv.org/abs/2605.05737](https://arxiv.org/abs/2605.05737)\n- **Abstract**:\n> Current reasoning paradigms for LLMs include chain-of-thought, ReAct, and post-hoc self-critique. These paradigms rely on two assumptions that fail on long-horizon, multi-stage tasks. As a result, errors accumulate silently across reasoning steps, leaving an open question: can a reasoning system effectively detect and recover from its own failures? We present ReFlect, a \\emph{harness} system for LLM reasoning that creates standalone error detection and recovery logic as a deterministic wrapper around the model. Controlled experiments across 6 reasoning domains show that prompt-level self-critique produces formulaic templates that flag no issues in 90 of 100 audited reflection blocks, and the investigated LLMs wrongly accept a wrong answer in at least 76\\% of cases. Our ReFlect harness achieves task success rates ranging from 41\\% on gpt-4o-mini to 56\\% on Claude Sonnet 4.5 across six models spanning small and frontier scale, with per-model gains over Direct CoT ranging from +7 pp on Qwen2.5-72B to +29 pp on Claude Sonnet 4.5, and additionally raises SWE-bench patch-structural quality from 0\\% (Direct CoT) to between 82\\% (Qwen2.5-72B) and 87\\% (GPT-4o). Notably, the harness gain is inversely proportional to the model's Direct CoT task success rate (the fitted slope is -1.69 with r=-0.76): each pp lost in baseline success rate is mechanically recovered by 1.69 pp of harness gain. We spot that adding structured reasoning state and operators yields only 15.0--18.7\\% pair-mean on Llama-3.3-70B and Qwen2.5-72B because models at this scale cannot reliably populate the state its operators require. ReFlect is model-agnostic, training-free, and operates entirely at inference time.\n\n---\n\n## 47. Knee Osteoarthritis Severity Grading Using Optimized Deep Learning and LLM-Driven Intelligent AI on Computationally Limited Systems\n- **Authors**: Dayam Nadeem , Neha , Safdar Mustafa , Adnan Alvi , Mohd Hussain\n- **URL**: [https://arxiv.org/abs/2605.05731](https://arxiv.org/abs/2605.05731)\n- **Abstract**:\n> Knee osteoarthritis (KOA) is among the musculoskeletal disorders that considerably restrict joint mobility, cause severe chronic pain and impact negatively on quality life. It is one of the persistent health issues worldwide. Generally, subjectivity and inter-observer variability undermine conventional practices and evaluation process that are adopted to address such health issues. Hence precise and timely diagnosis would be one of the effective ways for the assessment of its severity. This paper proposes an automated diagnostic approach for severity grading of KOA by blending a deep learning convolutional neural network (CNN) with a device-based inference platform powered by TensorFlow Lite. It proposes a model based on the ResNet-18 convolutional neural network. The designed model is trained on publicly available database. Through a transfer learning approach obtained knee images are first classified into five Kellgren-Lawrence (KL) grades. Further the developed model is optimised. During the training of the model test accuracy of 94.48% with stable convergence has been achieved. Subsequently the optimised model transformed into a lightweight TensorFlow Lite format, facilitating seamless deployment on resource-constrained devices. The designed model is capable enough to operate in the environment having no continuous internet connectivity. Also, an auxiliary Large Language Model (Gemini-2.0-flash) is applied to generate structured interpretive findings like potential symptoms, risk factors, and preventive majors etc. The LLM component functions as interface without influencing the classification process. The proposed model articulates the feasibility of an on-device, interpretable decision-support tools for early diagnosis and improve accessibility to Artificial Intelligence (AI)-assisted knee screening tool.\n\n---\n\n## 48. SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents\n- **Authors**: Hongcheol Cho , Ryangkyung Kang , Youngeun Kim\n- **URL**: [https://arxiv.org/abs/2605.05726](https://arxiv.org/abs/2605.05726)\n- **Abstract**:\n> As LLM agents are increasingly deployed with large libraries of reusable skills, selecting the right skill for a user request has become a critical systems challenge. In small libraries, users may invoke skills explicitly by name, but this assumption breaks down as skill ecosystems grow under tight context and latency budgets. Despite its practical importance, skill retrieval remains underexplored, with limited benchmarks and little understanding of retrieval behavior on realistic skill libraries. To address this gap, we introduce SkillRet, a large-scale benchmark for skill retrieval in LLM agents. SkillRet contains 17,810 public agent skills, organized with structured semantic tags and a two-level taxonomy spanning 6 major categories and 18 sub-categories. It provides 63,259 training samples and 4,997 evaluation queries with disjoint skill pools, enabling both benchmarking and retrieval-oriented training. Across a diverse set of retrievers, we find that skill retrieval remains far from solved: off-the-shelf models struggle on realistic large-scale skill libraries, and prior skill-retrieval models still leave substantial headroom. Task-specific fine-tuning on SkillRet substantially improves performance, improving NDCG@10 by +13.1 points over the strongest prior retriever and by +16.9 points over the strongest off-the-shelf retriever. Our analysis further suggests that these gains arise because fine-tuned models better focus on the small skill-relevant signals within long and noisy queries. These results establish SkillRet as a strong benchmark and foundation for future research on retrieval in large-scale agent systems.\n\n---\n\n## 49. Detecting Time Series Anomalies Like an Expert: A Multi-Agent LLM Framework with Specialized Analyzers\n- **Authors**: Hyeongwon Kang , Jeongseob Kim , Jinwoo Park , Pilsung Kang\n- **URL**: [https://arxiv.org/abs/2605.05725](https://arxiv.org/abs/2605.05725)\n- **Abstract**:\n> Recent studies have explored large language models for time-series anomaly detection, yet existing approaches often rely on a single general-purpose model to directly infer anomaly indices or intervals, limiting controllability, interpretability, and reliability for complex anomaly patterns. We propose SAGE (Specialized Analyzer Group for Expert-like Detection), a multi-agent framework for structured anomaly diagnosis in univariate time series. It decomposes anomaly analysis into four specialized Analyzers for point, structural, seasonal, and pattern anomalies. Each Analyzer applies family-specific numerical tools and diagnostic visualizations to generate evidence, while an evidence-grounded Detector consolidates the evidence into confidence-scored anomaly records with intervals and candidate types. A Supervisor then converts these structured records into analyst-facing diagnostic reports. SAGE further constructs synthetic in-context examples from normal-reference training segments, without using real anomalous segments or anomaly-type labels as in-context examples. Across three benchmarks, SAGE achieves the best average performance among strong ML/DL and language-model-based baselines. Ablation studies and human evaluation further show that the proposed framework improves detection reliability and the practical usefulness of diagnostic outputs.\n\n---\n\n## 50. More Is Not Always Better: Cross-Component Interference in LLM Agent Scaffolding\n- **Authors**: Ming Liu\n- **URL**: [https://arxiv.org/abs/2605.05716](https://arxiv.org/abs/2605.05716)\n- **Abstract**:\n> LLM agent systems are built by stacking scaffolding components (planning, tools, memory, self-reflection, retrieval) assuming more is better. We study cross-component interference (CCI): degradation when components interact destructively. We run a full factorial experiment over all 2^5=32 subsets of five components on HotpotQA and GSM8K with Llama-3.1-8B/70B (96 conditions, up to 10 seeds). The All-In system is consistently suboptimal: on HotpotQA, a single-tool agent surpasses All-In by 32% (F1 0.233 vs 0.177, p=0.023); on GSM8K, a 3-component subset beats All-In by 79% (0.43 vs 0.24, p=0.010). Optimal component count is task-dependent (k*=1-4) and scale-sensitive: at 70B, combinations that hurt at 8B provide gains, though All-In still trails the best subset. We fit a main-effects regression (R^2=0.916, adj-R^2=0.899, LOOCV=0.872), compute exact Shapley values, and find 183/325 submodularity violations (56.3%), showing greedy selection is unreliable. A three-body synergy among Tool Use, Self-Reflection, and Retrieval (INT_3=+0.175, 95% CI [+0.003,+0.351]) is reported as exploratory. CCI replicates across model families (Qwen2.5) and is robust to prompt paraphrasing. Our findings suggest maximally-equipped agent defaults should be replaced by task-specific subset selection via interaction-aware analysis.\n\n---\n\n## 51. Decodable but Not Corrected by Fixed Residual-Stream Linear Steering: Evidence from Medical LLM Failure Regimes\n- **Authors**: Ming Liu\n- **URL**: [https://arxiv.org/abs/2605.05715](https://arxiv.org/abs/2605.05715)\n- **Abstract**:\n> Can linearly decodable failure signals in LLM hidden states be leveraged to correct those failures? We investigate this classification-correction gap via Overthinking (OT)--a stable behavioral regime (Jaccard >= 0.81, 94% inter-annotator agreement) in medical QA where models answer correctly under resampling yet fail in extended chain-of-thought. OT is linearly decodable at 71.6% balanced accuracy (p < 10^{-16}). Yet five families of fixed linear steering (29 configurations, n=1,273) all yield Delta ~= 0, with identical null results cross-architecture (Qwen2.5-7B) and cross-domain (MMLU-STEM). Three convergent lines of evidence suggest representational entanglement: the OT direction has 85-88% overlap with task-critical computation (specificity ratio <= 0.152); non-targeted shared-direction steering damages accuracy (-12.1pp); and LEACE concept erasure damages accuracy (-3.6pp, p=0.01), while 10 random erasures produce Delta=+0.3pp. The per-instance probe-steering correlation is r=-0.002 (p=0.97). Positively, the same probe enables selective abstention (held-out AUROC=0.610, exceeding all five uncertainty baselines, p=0.009): decodable failure structure supports post-generation reliability estimation even when the fixed linear steering family cannot exploit it for correction.\n\n---\n\n## 52. Conceal, Reconstruct, Jailbreak: Exploiting the Reconstruction-Concealment Tradeoff in MLLMs\n- **Authors**: Md Farhamdur Reza , Richeng Jin , Tianfu Wu , Huaiyu Dai\n- **URL**: [https://arxiv.org/abs/2605.05709](https://arxiv.org/abs/2605.05709)\n- **Abstract**:\n> Intent-obfuscation-based jailbreak attacks on multimodal large language models (MLLMs) transform a harmful query into a concealed multimodal input to bypass safety mechanisms. We show that such attacks are governed by a \\emph{reconstruction--concealment tradeoff}: the transformed input must hide harmful intent from safety filters while remaining recoverable enough for the victim model to reconstruct the original request. Through a reconstruction analysis of three representative black-box methods, we find that existing transformations struggle to balance this tradeoff, limiting their effectiveness. In contrast, we show that character-removed variants achieve a better balance. Building on this, we propose \\emph{concealment-aware variant construction}, which greedily selects character-removed variants that are low in harmful-keyword alignment and mutually diverse, and instantiates them through five modality-aware prompting strategies. We further introduce \\emph{keyword-related distractor images} that depict the harmful keyword in diverse contexts, providing more effective auxiliary visual context than generic distractor images. Experiments across closed-source and open-source MLLMs show the proposed strategies outperform strong baselines, revealing an underexplored vulnerability: a model's own reconstruction ability can be exploited to recover hidden harmful intent and produce unsafe responses.\n\n---\n\n## 53. Resolving the bias-precision paradox with stochastic causal representation learning for personalized medicine\n- **Authors**: Peisong Zhang , Manqiang Peng , Yuxuan Wu , Pawit Phadungsaksawasdi , Wesley Yeung , Ye Zhang , Trang Nguyen , Qiang Zhang , Nan Liu , Meng Wang , Kee Yuan Ngiam , Yih-Chung Tham , Ching-Yu Cheng , Tianfan Fu , Qingyu Chen , Rosemary Ke , Chang Li , Wenzhuo Yang , Zhenghao Lu , Chunyou Lai , Yu Zhang , Sheng Zhong , Hao Deng , Dianbo Liu\n- **URL**: [https://arxiv.org/abs/2605.05706](https://arxiv.org/abs/2605.05706)\n- **Abstract**:\n> Estimating individualized treatment effects from longitudinal observational data is central to data-driven medicine, yet existing methods face a fundamental limitation: reducing confounding bias often suppresses clinically informative heterogeneity, degrading patient-specific predictions. Here, we identify this tension as a bias-precision paradox in causal representation learning and introduce sampling-based maximum mean discrepancy (sMMD), a stochastic alignment strategy that replaces global adversarial balancing with subset-level matching. We instantiate this approach in a framework for counterfactual outcome prediction with attribution-grounded interpretability. Across two large-scale ICU cohorts (n = 27,783), our framework improves accuracy under distribution shift, reducing error by up to 11.5% and substantially increasing recall in high-risk tasks. Mechanistic analyses show that sMMD selectively preserves clinically decisive variables. In human-AI evaluation, our method outperforms clinicians-in-training and large language models, and improves clinician accuracy by 14.7% while reducing decision time, enabling interpretable, real-time clinical decision support.\n\n---\n\n## 54. Knowledge-Graph Paths as Intermediate Supervision for Self-Evolving Search Agents\n- **Authors**: Huyu Wu , Jun Liu , Xiaochi Wei , Yan Gao , Yi Wu , Yao Hu\n- **URL**: [https://arxiv.org/abs/2605.05702](https://arxiv.org/abs/2605.05702)\n- **Abstract**:\n> Self-evolving search agents reduce reliance on human-written training questions by generating and solving their own search tasks. We build on Search Self-Play (SSP), a representative Proposer and Solver framework in which questions are generated and answered via multi-step search and reasoning. In practice, however, SSP faces two bottlenecks: the Proposer constructs questions from isolated answer entities without relational context, yielding many invalid or unverifiable questions in early self-play training, while the Solver receives only a binary outcome reward that discards useful signal from partially on-track search trajectories. We address both bottlenecks by reusing knowledge-graph paths as construction-derived intermediate supervision for both question construction and reward shaping. First, we ground question construction in LLM-guided knowledge-graph subgraphs, providing relational context for the Proposer. Second, we observe that constructing and solving a multi-hop question can involve overlapping intermediate entities: the factual bridges used to formulate the question may provide approximate waypoints for answering it. Exploiting this overlap, we introduce Waypoint Coverage Reward (WCR), which grants graded partial credit to incorrect Solver trajectories according to their coverage of entities on the construction path, while preserving full reward for correct answers. Across seven QA benchmarks and nine model configurations, our approach improves the average score over standard SSP in all configurations, including notable gains on multi-hop QA tasks. These results suggest that knowledge-graph paths can be reused as lightweight intermediate supervision, providing both relational guidance and process feedback without additional task-specific human annotations or manually labeled process steps.\n\n---\n\n## 55. Inference-Time Budget Control for LLM Search Agents\n- **Authors**: Zhengru Fang , Senkang Forest Hu , Zhonghao Chang , Yu Guo , Yihang Tao , Hongyao Liu , Mengzhe Ruan , Jun Huang , Yuguang Fang\n- **URL**: [https://arxiv.org/abs/2605.05701](https://arxiv.org/abs/2605.05701)\n- **Abstract**:\n> LLM search agents increasingly rely on tools at inference time, but their trajectories are often constrained by hard limits on both tool calls and generated tokens. Under such dual budgets, better answers require not only stronger models, but also explicit control over which search action should receive the next budget unit and when the accumulated evidence is sufficient to commit a final answer. We study this problem in multi-hop question answering (QA) and formulate it as two-stage inference-time budget control. At search time, our controller assigns each feasible action a task-level Value-of-Information (VOI) score, defined as an operational estimate of marginal task value per unit budget under the current search state and remaining dual budget, and uses this score to choose among retrieval, decomposition, and answer commitment. After search, a selective evidence-grounded finalizer compares the trajectory answer with a refined candidate and rewrites only when the residual error appears to be a low-risk answer-form error. Across four multi-hop QA benchmarks, three LLM backbones, and four budget levels, the method yields positive aggregate gains over four audited baselines under the same hard dual-budget protocol. Ablations show that search-time budget control, especially budget-dependent penalty, provides the main performance gain, while answer-time control helps mainly when the retrieval path is already adequate. These results suggest that inference-time budget control for LLM search agents should govern both how budget is spent during search and how the final answer is committed.\n\n---\n\n## 56. Saliency-Aware Regularized Quantization Calibration for Large Language Models\n- **Authors**: Yanlong Zhao , Xiaoyuan Cheng , Huihang Liu , Baihua He , Xinyu Zhang , Harrison Bo Hua Zhu , Wenlong Chen , Li Zeng , Zhuo Sun\n- **URL**: [https://arxiv.org/abs/2605.05693](https://arxiv.org/abs/2605.05693)\n- **Abstract**:\n> Post-training quantization (PTQ) is an effective approach for deploying large language models (LLMs) under memory and latency constraints. Most existing PTQ methods determine quantization parameters by minimizing a layer-wise reconstruction error on a predetermined calibration dataset, usually optimized via either scale search or Gram-based methods. However, from the perspective of generalization risk, existing calibration objectives of PTQ based only on empirical reconstruction error on limited or unrepresentative calibration data could move the quantized weights away from the original weights. This may cause the generalization risk to diverge, potentially degrading downstream performance. To address this issue, we propose \\emph{Saliency-Aware Regularized Quantization Calibration} (SARQC) a unified framework that augments the standard PTQ objective with a saliency-aware regularization term. This term encourages quantized weights to stay close to the original weights during calibration, leading to improved generalization during inference. SARQC integrates seamlessly into existing PTQ pipelines, enhancing both scale search and Gram-based methods under a unified formulation. Extensive experiments on dense and Mixture-of-Experts LLMs demonstrate consistent improvements in perplexity and zero-shot accuracy, without additional computational overhead during inference.\n\n---\n\n## 57. DataDignity: Training Data Attribution for Large Language Models\n- **Authors**: Xiaomin Li , Andrzej Banburski-Fahey , Jaron Lanier\n- **URL**: [https://arxiv.org/abs/2605.05687](https://arxiv.org/abs/2605.05687)\n- **Abstract**:\n> Auditing language-model outputs often requires more than judging correctness: an auditor may need to identify which source document most likely supports the knowledge expressed in a response. We study this as pinpoint provenance: given a prompt, a target-model response, and a candidate corpus, rank the documents that best support the response. We introduce FakeWiki, a controlled benchmark of 3,537 fabricated Wikipedia-style articles designed to preserve ground-truth provenance while weakening lexical shortcuts. FakeWiki includes QA probes, source-preserving paraphrases, retro-generated variants, hard anti-documents that remain topically similar while removing answer-critical facts, and five query conditions: clean prompting plus four jailbreak-inspired transformations. We evaluate seven retrieval baselines, a training-free activation-steering retrieval-fusion method, SteerFuse, and a supervised contrastive provenance ranker, ScoringModel. ScoringModel maps response and document features into a shared space and is trained with InfoNCE using in-batch, retrieval-mined, and anti-document negatives. Across nine open-weight instruction-tuned LLMs and five query conditions, ScoringModel improves mean Recall@10 from 35.0 for the strongest retrieval baseline to 52.2, without inference-time fusion, and wins 41/45 model-by-condition cells. SteerFuse is usually second-best despite requiring no supervised training, showing that activation-space evidence can efficiently complement text retrieval. On jailbreak-inspired transformed queries, ScoringModel improves Recall@10 by 15.7 points on average over the best baseline. Overall, our work shows that robust training data attribution requires evaluation settings that separate true answer support from topical or lexical resemblance.\n\n---\n\n## 58. Attractor Geometry of Transformer Memory: From Conflict Arbitration to Confident Hallucination\n- **Authors**: Qiyao Liang , Risto Miikkulainen , Ila Fiete\n- **URL**: [https://arxiv.org/abs/2605.05686](https://arxiv.org/abs/2605.05686)\n- **Abstract**:\n> Language models draw on two knowledge sources: facts baked into weights (parametric memory, PM) and information in context (working memory, WM). We study two mechanistically distinct failure modes--conflict, when PM and WM disagree and interfere; and hallucination, when the queried fact was never learned. Both produce confident output regardless, making output-based monitoring blind by design. We show both failures share a unified geometric account. In the hidden-state space of autoregressive generation, learned facts form attractor basins. Conflict is basin competition: WM disrupts convergence to the correct basin without raising output entropy. Hallucination is basin absence: the hidden state drifts freely when no memorized basin exists. The frozen LM head, designed for next-token prediction, cannot distinguish these cases and fires confidently either way. We verify this account in a controlled synthetic task--entity identifiers mapped to unique codes with PM installed via LoRA adapters--where ground truth is exact and component roles can be causally isolated through targeted adapter placement. Geometric margin--the hidden state's distance to the nearest memorized basin--reads this geometry directly and separates correct recall from hallucination far more cleanly than output entropy, with zero false refusals where entropy-based detection cannot avoid rejecting the vast majority of correct outputs. The separation holds on natural-language factual queries from the pretrained model with no adaptation, confirming attractor geometry is structural rather than a fine-tuning artifact. The fraction of confident hallucinations follows a scaling law $C = \\exp(-c/\\bar\\Delta)$, growing with scale even as overall error rates fall. Hidden states reliably encode epistemic state; the frozen output head systematically erases it--and this erasure worsens with scale.\n\n---\n\n## 59. Large Vision-Language Models Get Lost in Attention\n- **Authors**: Gongli Xi , Ye Tian , Mengyu Yang , Huahui Yi , Liang Lin , Xiaoshuai Hao , Kun Wang , Wendong Wang\n- **URL**: [https://arxiv.org/abs/2605.05668](https://arxiv.org/abs/2605.05668)\n- **Abstract**:\n> Despite the rapid evolution of training paradigms, the decoder backbone of large vision--language models (LVLMs) remains fundamentally rooted in the residual-connection Transformer architecture. Therefore, deciphering the distinct roles of internal modules is critical for understanding model mechanics and guiding architectural optimization. While prior statistical approaches have provided valuable attribution-based insights, they often lack a unified theoretical basis. To bridge this gap, we propose a unified framework grounded in information theory and geometry to quantify the geometric and entropic nature of residual updates. Applying this unified framework reveals a fundamental functional decoupling: Attention acts as a subspace-preserving operator focused on reconfiguration, whereas FFNs serve as subspace-expanding operators driving semantic innovation. Strikingly, further experiments demonstrate that replacing learned attention weights with predefined values (e.g., Gaussian noise) yields comparable or even superior performance across a majority of datasets relative to vanilla models. These results expose severe misallocation and redundancy in current mechanisms, suggesting that state-of-the-art LVLMs effectively ``get lost in attention'' rather than efficiently leveraging visual context.\n\n---\n\n## 60. Retrieval-Conditioned Topology Selection with Provable Budget Conservation for Multi-Agent Code Generation\n- **Authors**: Abhijit Talluri , Pujith Anne , Bhagavan Choudary Pendiyala , Raghavendra Chilukuri\n- **URL**: [https://arxiv.org/abs/2605.05657](https://arxiv.org/abs/2605.05657)\n- **Abstract**:\n> Multi-agent LLM systems for code generation face a fundamental routing problem: the optimal orchestration topology depends on the structural complexity of the code under modification, yet existing systems select topologies without consulting the codebase. We present Retrieval-Guided Adaptive Orchestration (RGAO), an architecture that closes this loop by extracting a structural complexity vector from a hierarchical code index before selecting the orchestration topology. RGAO operates within Code-Agent, a multi-agent framework whose sub-agents are governed by formal contracts with six-dimensional budget vectors. Our headline contribution is the composition of two previously separate lines of work -- complexity-conditioned LLM routing and formal resource algebras -- yielding a property neither admits alone: provable budget conservation under retrieval-conditioned dynamic topology selection. Concretely we contribute: (1) a complexity-conditioned topology router that reduces proxy-measured misrouting from 30.1% to 8.2%; (2) a budget algebra with a structural-induction conservation theorem; and (3) a hierarchical code retrieval engine. Empirical evaluation demonstrates sub-millisecond DAG construction and linear tree-index scalability.\n\n---\n\n## 61. Text-Graph Synergy: A Bidirectional Verification and Completion Framework for RAG\n- **Authors**: Jiarui Zhong , Hong Cai Chen\n- **URL**: [https://arxiv.org/abs/2605.05643](https://arxiv.org/abs/2605.05643)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) has become a core paradigm for enhancing factual grounding and multi-hop reasoning in Large Language Models (LLMs). Traditional text-based RAG often retrieves logically irrelevant pseudo-evidence, while graph-based RAG is frequently hindered by search-time pruning, which may discard potentially valid reasoning paths. Existing hybrid approaches primarily adopt simple evidence concatenation or unidirectional enhancement, which fails to address the fundamental \"Information Island\" problem caused by asymmetric reasoning flows between unstructured text and structured graphs. We propose \\textbf{TGS-RAG}, a unified framework for \\textbf{T}ext-\\textbf{G}raph \\textbf{S}ynergistic enhancement. TGS-RAG introduces a bidirectional mechanism: (i) a \\textbf{Graph-to-Text} channel that employs a Global Voting strategy from visited graph nodes to re-rank and refine textual evidence, filtering out semantic noise; and (ii) a \\textbf{Text-to-Graph} channel that utilizes the \\textbf{Memory-based Orphan Entity Bridging} algorithm. This algorithm utilizes textual cues to proactively resurrect valid but previously pruned reasoning paths from the search history without additional database overhead. Experimental results on multiple multi-hop reasoning benchmarks demonstrate that TGS-RAG significantly outperforms state-of-the-art baselines, achieving a superior balance between retrieval precision and computational efficiency.\n\n---\n\n## 62. Prober.ai: Gated Inquiry-Based Feedback via LLM-Constrained Personas for Argumentative Writing Development\n- **Authors**: Ran Bi , Shiyao Wei , Yuanyiyi Zhou\n- **URL**: [https://arxiv.org/abs/2605.05598](https://arxiv.org/abs/2605.05598)\n- **Abstract**:\n> The proliferation of large language models (LLMs) in educational settings has paradoxically undermined the cognitive processes they purport to support. Students increasingly outsource critical thinking to AI assistants that generate polished text on demand, resulting in measurable cognitive debt and diminished argumentative reasoning skills. We present this http URL , a web-based writing environment that inverts the conventional AI-tutoring paradigm: rather than generating or rewriting student text, the system constrains an LLM (Gemini 3 Flash Preview) through persona-specific system prompts and structured JSON output schemas to produce only targeted, inquiry-based questions about argumentative weaknesses. A two-phase interaction architecture -- Challenge and Unlock -- implements a pedagogical friction mechanism whereby revision suggestions are gated behind mandatory student reflection. The system's design is grounded in Toulmin's argumentation theory, research on peer feedforward questioning mechanisms, and evidence on AI-supported feedback in writing instruction. A functional prototype was developed in 36 hours during the NY EdTech Hackathon (March 2026), where it was awarded second place. We describe the system architecture, the prompt engineering methodology for constraining LLM output to pedagogically aligned JSON schemas, and discuss implications for scalable, cognition-preserving AI integration in writing education.\n\n---\n\n## 63. Causal Probing for Internal Visual Representations in Multimodal Large Language Models\n- **Authors**: Zehao Deng , Tianjie Ju , Zheng Wu , Liangbo He , Jun Lan , Huijia Zhu , Weiqiang Wang , Zhuosheng Zhang\n- **URL**: [https://arxiv.org/abs/2605.05593](https://arxiv.org/abs/2605.05593)\n- **Abstract**:\n> Despite the remarkable success of Multimodal Large Language Models (MLLMs) across diverse tasks, the internal mechanisms governing how they encode and ground distinct visual concepts remain poorly understood. To bridge this gap, we propose a causal framework based on activation steering to actively probe and manipulate internal visual representations. Through systematic intervention across four visual concept categories, our results reveal a divergence in concept encoding: entities exhibit distinct localized memorization, whereas abstract concepts are globally distributed across the network. Critically, this divergence uncovers a mechanistic driver of scaling laws: increasing model depth is indispensable for encoding distributed and complex abstract concepts, whereas entity localization remains remarkably invariant to scale. Furthermore, reverse steering uncovers that blocking explicit output triggers a surge in latent activations, exposing a compensatory mechanism between perception and generation. Finally, extending our analysis to visual reasoning, we expose a disconnect between perception and reasoning although MLLMs successfully recognize geometric relations, they treat them merely as static visual features, failing to trigger the procedural execution necessary for abstract problem-solving.\n\n---\n\n## 64. Belief Memory: Agent Memory Under Partial Observability\n- **Authors**: Junfeng Liao , Qizhou Wang , Jianing Zhu , Bo Du , Rui Yan , Xiuying Chen\n- **URL**: [https://arxiv.org/abs/2605.05583](https://arxiv.org/abs/2605.05583)\n- **Abstract**:\n> LLM agents that operate over long context depend on external memory to accumulate knowledge over time. However, existing methods typically store each observation as a single deterministic conclusion (e.g., inferring \"API~X failed\" from temporary errors), even though such observations are inherently partial and potentially ambiguous. By committing to one conclusion and discarding uncertainty, these methods introduce self-reinforcing error: the agent acts on the stored conclusion, never revisits alternatives, and reinforces the conclusion over time. To address this issue, we propose BeliefMem, which shifts the memory paradigm from committing to a single conclusion per observation to retaining multiple candidate conclusions with their probabilities. Concretely, BeliefMem stores the candidate conclusions as separate memory entries, each carrying a probability that is updated via Noisy-OR rules as new observations arrive. At retrieval, all candidates surface together with their probabilities, keeping alternatives visible to the agent. Since each conclusion in memory retains its probability, BeliefMem preserves the uncertainty that the deterministic paradigm discards, enabling the agent to act with high confidence on well-evidenced knowledge while retaining the capacity to update its confidence when new evidence arrives. Empirical evaluations on LoCoMo and ALFWorld benchmarks show that, even with limited data, BeliefMem achieves the best average performance, remarkably outperforming well-known baselines. More broadly, such probabilistic memory produces substantial gains and explores a new direction for agent memory in partially observable environments.\n\n---\n\n## 65. AlphaCrafter: A Full-Stack Multi-Agent Framework for Cross-Sectional Quantitative Trading\n- **Authors**: Yishuo Yuan , Jiayi Sheng , Sirui Zeng , Jiaqi Wang , Jiaheng Liu\n- **URL**: [https://arxiv.org/abs/2605.05580](https://arxiv.org/abs/2605.05580)\n- **Abstract**:\n> Financial markets are inherently non-stationary, driven by complex interactions among macroeconomic regimes, microstructural frictions, and behavioral dynamics. Building quantitative strategies that remain profitable demands the continuous coupling of factor discovery, regime-adaptive selection, and risk-constrained execution. Prevailing approaches, however, optimize these components under static or isolated assumptions. Factor mining frameworks typically treat alpha discovery as a one-time search process, implicitly assuming that factor efficacy persists across market regimes. Execution-oriented systems often adopt role-playing agent architectures that simulate anthropomorphic trading committees, introducing behavioral noise rather than systematic rationality. Consequently, a fully automated, rationality-driven framework unifying a coherent quantitative pipeline remains absent. We introduce AlphaCrafter, a full-stack multi-agent framework that closes this gap through a continuously adaptive factor-to-execution pipeline, designed to track and respond to evolving market conditions without manual intervention. AlphaCrafter operates via three specialized agents: a Miner that continuously expands the factor pool via LLM-guided search, a Screener that assesses prevailing market conditions to construct regime-conditioned factor ensembles, and a Trader that translates these ensembles into quantitative strategies under explicit risk constraints. Together, these three agents form a closed-loop cross-sectional trading system that adapts holistically to evolving market dynamics. Extensive experiments on CSI 300 and S&P 500 demonstrate that AlphaCrafter consistently outperforms state-of-the-art baselines in risk-adjusted returns while exhibiting the lowest cross-trial variance, confirming that integrated and adaptive factor-to-execution design yields robust trading performance.\n\n---\n\n## 66. Nonsense Helps: Prompt Space Perturbation Broadens Reasoning Exploration\n- **Authors**: Langlin Huang , Chengsong Huang , Jinyuan Li , Donghong Cai , Yuyi Yang , Jiaxin Huang\n- **URL**: [https://arxiv.org/abs/2605.05566](https://arxiv.org/abs/2605.05566)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards, particularly Group Relative Policy Optimization (GRPO), has significantly advanced the reasoning capabilities of Large Language Models (LLMs). However, in complex tasks, GRPO frequently suffers from the ``zero-advantage problem'': when all sampled rollouts for a query fail, the relative advantage collapses to zero. Consequently, the model loses effective training signals for these questions, wasting the training data and computational budget. While simply increasing the sampling budget for these questions is a common remedy, the static sampling policy inherently constrains reasoning exploration, limiting the success rate. In this paper, we propose Lorem Perturbation for Exploration (LoPE), a simple yet effective training framework to break this exploration bottleneck. We posit that task-irrelevant prompt-space perturbations can shift the model's output distribution enough to unlock orthogonal reasoning pathways for hard questions. Specifically, LoPE prepends sequences stochastically assembled from Lorem Ipsum vocabulary (a pseudo-Latin placeholder text) to the prompts before resampling. Experiments across 1.7B, 4B, and 7B models demonstrate that LoPE significantly outperforms resampling with the original prompts. Further analysis reveals that other Latin-based random sequences with low perplexity are also effective perturbations. Our results establish LoPE as a strong baseline for broadening exploration in LLM reinforcement learning.\n\n---\n\n## 67. SPARK: Self-Play with Asymmetric Reward from Knowledge Graphs\n- **Authors**: Hyobin Park , Taeseop Kim , Dong-Geol Choi\n- **URL**: [https://arxiv.org/abs/2605.05546](https://arxiv.org/abs/2605.05546)\n- **Abstract**:\n> Self-play reinforcement learning has shown strong performance in domains with formally verifiable structure, such as mathematics and coding, where both problem generation and reward computation can be grounded in explicit rules. Extending this paradigm to scientific literature is more challenging: the relationships among multi-modal elements within and across documents are rarely made explicit in text, which makes automatic generation of relational reasoning questions difficult and weakens the reliability of reward signals. We propose SPARK (Self-Play with Asymmetric Reward from Knowledge Graphs), a framework that automatically constructs a unified knowledge graph (KG) from multi-document scientific literature and uses it as the structural basis for self-play. KG paths over multimodal nodes serve as a source for generating relational reasoning questions, and structured facts stored in the KG provide a basis for verifiable reward computation. A single small vision-language model (sVLM) alternates between Proposer and Solver roles under information asymmetry against a fixed KG, a design that we believe can be naturally extended toward online adaptation in future work. We evaluate SPARK on public benchmarks and a self-constructed cross-document multi-hop QA dataset. Results show that SPARK consistently outperforms flat-corpus-based self-play baselines, and the performance gap widens as hop count increases, suggesting that KG-structure grounding contributes to relational multi-hop reasoning beyond what unstructured corpus grounding can provide.\n\n---\n\n## 68. AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases\n- **Authors**: Susheel Suresh , Hazel Mak , Shangpo Chou , Fred Kroon , Sahil Bhatnagar\n- **URL**: [https://arxiv.org/abs/2605.05538](https://arxiv.org/abs/2605.05538)\n- **Abstract**:\n> We present AgenticRAG, a practical agentic harness for retrieval and analysis over enterprise knowledge bases. Standard RAG pipelines place significant burden of grounding on the search stack, constraining the language model to a fixed candidate set chosen deep in the retrieval process. Our approach reduces this overdependence by layering a lightweight harness on top of existing enterprise search infrastructure, equipping a reasoning LLM with search, find, open, and summarize tools enabling the model to iteratively retrieve information, navigate within documents, and analyze evidence autonomously. On three open benchmarks we observe substantial gains: $49.6\\%$ recall@1 on BRIGHT (+21.8 pp over the best embedding baseline), 0.96 factuality on WixQA ($+13\\%$ relative improvement), and $92\\%$ answer correctness on FinanceBench--within 2 pp of oracle access to true evidence. Ablation studies show that the most significant factor is the shift from single-shot retrieval to agentic tool use ($5.9\\times$ improvement), while multi-query search and in-document navigation contribute to both quality and efficiency. We present various design choices in our agentic harness that were informed by pre-production deployments. Our results demonstrate its suitability for real-world enterprise production environments.\n\n---\n\n## 69. FoodCHA: Multi-Modal LLM Agent for Fine-Grained Food Analysis\n- **Authors**: Woojin Lee , Pranav Mekkoth , Ye Tian , Onat Gungor , Tajana Rosing\n- **URL**: [https://arxiv.org/abs/2605.05499](https://arxiv.org/abs/2605.05499)\n- **Abstract**:\n> The widespread adoption of camera-equipped mobile devices and wearables has enabled convenient capture of meal images, making food recognition a key component for real time dietary monitoring. However, real-world food images present challenges due to high intra-class similarity and the frequent presence of multiple food items within a single image. While deep learning models achieve strong performance in coarse grained classification, they often struggle to capture fine-grained attributes such as cooking style. Moreover, open-ended generation in modern vision-language models can produce non-canonical labels, limiting their practical deployment. We propose FoodCHA, a multimodal agentic framework that reformulates food recognition as a hierarchical decision-making process. By progressively anchoring predictions, FoodCHA guides subcategory identification using high-level categories and guides cooking style recognition using subcategories, improving semantic consistency and attribute-level discrimination. To ensure practical deployability, FoodCHA utilizes the compact Moondream-2B vision language model, which provides strong reasoning capability while maintaining lower computational and memory overhead. Experiments on FoodNExTDB show that FoodCHA outperforms Food-Llama-3.2-11B by 13.8% and 38.2% in category and subcategory recognition precision, respectively, and achieves a striking 153.2% improvement in cooking style classification precision.\n\n---\n\n## 70. FinRAG-12B: A Production-Validated Recipe for Grounded Question Answering in Banking\n- **Authors**: Denys Katerenchuk , Pablo Duboue , Keelan Evanini , David Gondek , Nithin Govindugari , Olivier Allauzen , Joshua Baptiste , David J More , Joshua Schechter\n- **URL**: [https://arxiv.org/abs/2605.05482](https://arxiv.org/abs/2605.05482)\n- **Abstract**:\n> Large language models (LLMs) are rapidly being adopted across various domains. However, their adoption in banking industry faces resistance due to demands for high accuracy, regulatory compliance, and the need for verifiable and grounded responses. We present a unified, data-efficient framework for training grounded domain-specific LLMs that optimizes answer quality, citation grounding, and calibrated refusal under real-world deployment constraints. First, we describe a data generation pipeline that combines LLM-as-a-Judge filtering, citation annotation, and curriculum learning with only 143M tokens. The resulting 12B model achieves high answer quality outperforming GPT-4.1 on citation grounding, with a modest citation tradeoff versus the untuned base. Second, we propose a calibrated refusal mechanism: training on 22% unanswerable examples yield a 12% \"I don't know\" rate, substantially improving over the base model's unsafe 4.3% rate while avoiding GPT-4.1's over-refusal (20.2%). Third, we present an end-to-end methodology spanning from data curation to quantized serving. The system is deployed at 40+ financial institutions, achieving a 7.1 percentage point improvement in query resolution (p < 0.001). Additionally, the model delivers 3-5x faster responses at 20-50x lower cost compared to GPT-4.1.\n\n---\n\n## 71. LANTERN: LLM-Augmented Neurosymbolic Transfer with Experience-Gated Reasoning Networks\n- **Authors**: Mahyar Alinejad , Yue Wang , Amrit Singh Bedi , George Atia\n- **URL**: [https://arxiv.org/abs/2605.05478](https://arxiv.org/abs/2605.05478)\n- **Abstract**:\n> Transfer learning in reinforcement learning (RL) seeks to accelerate learning in new tasks by leveraging knowledge from related sources. Existing neurosymbolic transfer methods, however, typically rely on manually specified task automata, assume a single source task, and use fixed knowledge-integration mechanisms that cannot adapt to varying source relevance. We propose LANTERN, a unified framework for multi-source neurosymbolic transfer that addresses these limitations through three components: (i) deterministic finite automata generated from natural language task descriptions using large language models, (ii) semantic embedding-based aggregation of multiple source policies weighted by cross-task similarity, and (iii) adaptive teacher-student gating based on temporal-difference error and semantic uncertainty. Across domains spanning resource management, navigation, and control, LANTERN achieves 40-60% improvements in sample efficiency over existing baselines while remaining robust to poorly aligned sources. These results demonstrate that multi-source, adaptively weighted neurosymbolic transfer can improve scalability and robustness in symbolic RL settings.\n\n---\n\n## 72. Agentic Discovery of Exchange-Correlation Density Functionals\n- **Authors**: Titouan Duston , Jiashu Liang , Yuanheng Wang , Weihao Gao , Xuelan Wen , Nan Sheng , Weiluo Ren , Yang Sun , Yixiao Chen\n- **URL**: [https://arxiv.org/abs/2605.05460](https://arxiv.org/abs/2605.05460)\n- **Abstract**:\n> The development of accurate exchange-correlation (XC) functionals remains a longstanding challenge in density functional theory (DFT). The vast majority of XC functionals have been hand designed by human researchers combining physical insight, exact constraints, and empirical fitting. Recent advances in large language models enable a systematic, automated alternative to this human-driven design loop. This report presents an agentic search system in which an LLM proposes structured functional-form changes guided by evolutionary history. The system attempts to improve functional performance through an iterative plan-execute-summarize loop, where improvements are measurable by optimizing functional parameters against a standard thermochemistry dataset, then evaluating performance on a held-out subset. The strongest discovered functional, SAFS26-a (Seed Agentic Functional Search 2026), improves upon the gold-standard {\\omega}B97M-V baseline by ~9%. These results also surface a cautionary lesson for AI-assisted science: models powerful enough to discover genuine improvements are equally capable of exploiting unphysical shortcuts to game the benchmark; domain expertise translated into explicitly enforced constraints remains essential to keeping results scientifically grounded.\n\n---\n\n## 73. The Geopolitics of AI Safety: A Causal Analysis of Regional LLM Bias\n- **Authors**: Alif Al Hasan\n- **URL**: [https://arxiv.org/abs/2605.05427](https://arxiv.org/abs/2605.05427)\n- **Abstract**:\n> As Large Language Models (LLMs) are integrated into global software systems, ensuring equitable safety guardrails is a critical requirement. Current fairness evaluations predominantly measure bias observationally, a methodology confounded by the inherent toxicity of topics naturally paired with specific demographics in testing datasets. This study introduces a Probabilistic Graphical Model (PGM) framework to audit LLM safety mechanisms causally. By applying Pearl's do-operator, we mathematically isolate the causal effect of injecting a cultural demographic into a prompt. We conduct a large-scale empirical analysis across seven instruction-tuned models spanning diverse origins: the United States (Llama-3.1-8B, Gemma-2-9B), Europe (Mistral-7B-v0.3), the UAE (Falcon3-7B), China (Qwen2.5-7B, DeepSeek-7B), and India (Airavata-7B). Utilizing two distinct datasets (ToxiGen and BOLD), the findings reveal a disparity between observational and interventional bias, demonstrating that standard fairness metrics can overestimate demographic bias by failing to account for context toxicity. Furthermore, the causal probabilities indicate distinct alignment trends: Western models exhibit higher causal refusal rates for specific demographic groups, whereas Eastern models demonstrate low overall intervention rates with targeted sensitivities toward regional demographics. We discuss the implications of these biases, highlighting how demographic-sensitive over-triggering restricts benign discourse in downstream applications.\n\n---\n\n## 74. From History to State: Constant-Context Skill Learning for LLM Agents\n- **Authors**: Haoyang Xie , Xinyuan Wang , Yancheng Wang , Puda Zhao , Feng Ju\n- **URL**: [https://arxiv.org/abs/2605.05413](https://arxiv.org/abs/2605.05413)\n- **Abstract**:\n> Large language model (LLM) agents are increasingly used to operate browsers, files, code and tools, making personal assistants a natural deployment target. Yet personal agents face a privacy-cost-capability tension: cloud models execute multi-step workflows well but expose sensitive intermediate context to external APIs, while local models preserve privacy but remain less reliable. Both settings also pay repeatedly for long skill prompts and growing histories. We propose constant-context skill learning, a context-to-weights framework for recurring agent workflows: reusable procedures are learned in lightweight task-family modules, while inference conditions only on the current observation and a compact state block. A deterministic tracker renders this state block from task progress and supplies aligned subgoal rewards, so each module can be trained with step-level SFT and refined through online RL. Across ALFWorld, WebShop, and SciWorld, our agents achieve strong performance across Qwen3-4B, Qwen3-8B and Llama-3.1-8B. With Qwen3-8B, SFT+RL reaches 89.6\\% unseen success on ALFWorld, 76.8\\% success on WebShop, and 66.4\\% unseen success on SciWorld. They match or exceed strong published agent-training results while reducing prompt tokens per turn by 2--7$\\times$ relative to controlled ReAct prompting baselines, showing that procedural context can be moved from prompts into weights.\n\n---\n\n## 75. LaTA: A Drop-in, FERPA-Compliant Local-LLM Autograder for Upper-Division STEM Coursework\n- **Authors**: Jesse A. Rodríguez\n- **URL**: [https://arxiv.org/abs/2605.05410](https://arxiv.org/abs/2605.05410)\n- **Abstract**:\n> Large-language-model (LLM) graders promise to relieve the grading burden of upper-division STEM courses, but most deployments to date send student work to third-party APIs, violating FERPA and exposing institutions to data risk while requiring substantial assignment modification. We present $\\textbf{LaTA}\\ (\\textit{LaTeX Teaching Assistant})$, a drop-in, open-source autograder that runs entirely on commodity on-premises hardware and assumes a LaTeX-native workflow already adopted by many engineering and physics courses. LaTA implements a four-stage pipeline (ingest, segment, grade, report) using a locally hosted open-weight chain-of-thought LLM grader (gpt-oss:120b) that compares student work to an instructor-authored reference solution and applies a YAML rubric with binary per-item scoring. We deployed LaTA in Winter~2026 in ME 373 (Mechanical Engineering Methods) at Oregon State University, grading every weekly assignment for approximately 200 students on a single Mac Studio at \\$0 marginal cost per assignment and 1--3 minutes of wall-clock time per submission, enabling regrading of corrected assignments and greatly expanded TA office hour offerings. The instructor-confirmed grading-error rate held at roughly $0.02$--$0.04\\%$ per rubric line item across the term. Relative to the same instructor's previous traditionally-graded cohort, the LaTA-graded cohort outperformed by approximately $11\\%$ on the midterm exam and $8\\%$ on the final exam, and reported large gains in self-assessed confidence on every stated learning objective ($N = 159$ survey responses, $\\Delta \\geq +1.49$ Likert points, $p < 10^{-27}$ on every comparison). We release the code under AGPLv3.\n\n---\n\n## 76. Agentic Retrieval-Augmented Generation for Financial Document Question Answering\n- **Authors**: Yang Shu , Yingmin Liu , Zequn Xie\n- **URL**: [https://arxiv.org/abs/2605.05409](https://arxiv.org/abs/2605.05409)\n- **Abstract**:\n> Financial document question answering (QA) demands complex multi-step numerical reasoning over heterogeneous evidence--structured tables, textual narratives, and footnotes--scattered across corporate filings. Existing retrieval-augmented generation (RAG) approaches adopt a single-pass retrieve-then-generate paradigm that struggles with the compositional reasoning chains prevalent in financial analysis. We propose FinAgent-RAG, an agentic RAG framework that orchestrates iterative retrieval-reasoning loops with self-verification, specifically engineered for the precision requirements of financial numerical reasoning. The framework integrates three domain-specific innovations: (1) a Contrastive Financial Retriever trained with hard negative mining to distinguish semantically similar but numerically distinct financial passages, (2) a Program-of-Thought reasoning module that generates executable Python code for precise arithmetic rather than relying on error-prone LLM-based mental computation, and (3) an Adaptive Strategy Router that dynamically allocates computational resources based on question complexity, reducing API costs by 41.3% on FinQA while preserving accuracy. Extensive experiments on three benchmark datasets--FinQA, ConvFinQA, and TAT-QA--demonstrate that FinAgent-RAG achieves 76.81%, 78.46%, and 74.96% execution accuracy respectively, outperforming the strongest baseline by 5.62--9.32 percentage points. Ablation studies, cross-backbone evaluation with four LLMs, and deployment cost analysis confirm the framework's robustness and practical viability for financial institutions.\n\n---\n\n## 77. PRISM: Perception Reasoning Interleaved for Sequential Decision Making\n- **Authors**: Mohamed Salim Aissi , Clemence Grislain , Clement Romac , Laure Soulier , Mohamed Chetouani , Olivier Sigaud , Nicolas Thome\n- **URL**: [https://arxiv.org/abs/2605.05407](https://arxiv.org/abs/2605.05407)\n- **Abstract**:\n> Scaling LLM-based embodied agents from text-only environments to complex multimodal settings remains a major challenge. Recent work identifies a perception-reasoning-decision gap in standalone Vision-Language Models (VLMs), which often overlook task-critical information. In this paper, we introduce PRISM, a framework that tightly couples perception (VLM) and decision (LLM) through a dynamic question-answer (DQA) pipeline. Instead of passively accepting the VLM's description, the LLM critiques it, probes the VLM with goal-oriented questions, and synthesizes a compact image description. This closed-loop interaction yields a sharp, task-driven understanding of the scene. We evaluate PRISM on the ALFWorld and Room-to-Room (R2R) benchmarks. We show that: (1) PRISM significantly outperforms state-of-the-art image-based models, (2) our Interactive goal-oriented perception pipeline yields systematic and substantial gains, and (3) PRISM is fully automatic, eliminating the need for handcrafted questions or answers.\n\n---\n\n## 78. When Helpfulness Becomes Sycophancy: Sycophancy is a Boundary Failure Between Social Alignment and Epistemic Integrity in Large Language Models\n- **Authors**: Jiechen Li , Catherine A. Barry , Rishika Randev , Janet Chen , Ella Jorgensen , Brinnae Bent\n- **URL**: [https://arxiv.org/abs/2605.05403](https://arxiv.org/abs/2605.05403)\n- **Abstract**:\n> This position paper argues that sycophancy in LLMs is a boundary failure between social alignment and epistemic integrity. Existing work often operationalizes sycophancy through external behavior such as agreement with incorrect user beliefs, position reversals, or deviation from an objective standard of correctness. These formulations capture only overt forms of the phenomenon and leave subtler boundary failures involving epistemic integrity and social alignment underspecified. We argue that sycophancy should not be understood as agreement alone, but as alignment behavior that displaces independent epistemic judgment. To clarify this boundary, we propose a three-condition framework for sycophancy. First, the user expresses a cue in the form of a belief, preference, or self-concept. Second, the model shifts toward that cue through alignment behavior. Third, this shift compromises epistemic accuracy, independent reasoning, or appropriate correction. We also introduce a taxonomy for classifying sycophancy, consisting of alignment targets, mechanisms, and severity. The paper concludes by discussing implications for alignment evaluation and argues for boundary-aware assessment, structured rubrics, and mitigation strategies, while situating these proposals alongside alternative views of sycophancy.\n\n---\n\n## 79. BALAR : A Bayesian Agentic Loop for Active Reasoning\n- **Authors**: Aymen Echarghaoui , Dongxia Wu , Emily B. Fox\n- **URL**: [https://arxiv.org/abs/2605.05386](https://arxiv.org/abs/2605.05386)\n- **Abstract**:\n> Large language models increasingly operate in interactive settings where solving a task requires multiple rounds of information exchange with a user. However, most current systems treat dialogue reactively and lack a principled mechanism to reason about what information is missing and which question should be asked next. We propose BALAR (Bayesian Agentic Loop for Active Reasoning), a task-agnostic outer-loop algorithm that requires no fine-tuning and enables structured multi-turn interaction between an LLM agent and a user. BALAR maintains a structured belief over latent states, selects clarifying questions by maximizing expected mutual information, and dynamically expands its state representation when the current one proves insufficient. We evaluate BALAR on three diverse benchmarks: AR-Bench-DC (detective cases), AR-Bench-SP (thinking puzzles), and iCraft-MD (clinical diagnosis). BALAR significantly outperforms all baselines across all three benchmarks, with $14.6\\%$ higher accuracy on AR-Bench-DC, $38.5\\%$ on AR-Bench-SP, and $30.5\\%$ on iCraft-MD.\n\n---\n\n## 80. Understanding Annotator Safety Policy with Interpretability\n- **Authors**: Alex Oesterling , Donghao Ren , Yannick Assogba , Dominik Moritz , Sunnie S.Y. Kim , Leon Gatys , Fred Hohman\n- **URL**: [https://arxiv.org/abs/2605.05329](https://arxiv.org/abs/2605.05329)\n- **Abstract**:\n> Safety policies define what constitutes safe and unsafe AI outputs, guiding data annotation and model development. However, annotation disagreement is pervasive and can stem from multiple sources such as operational failures (annotators misunderstand or misexecute the task), policy ambiguity (policy wording leaves room for interpretation), or value pluralism (different annotators hold different perspectives on safety). Distinguishing these sources matters. For example, operational failures call for quality control, ambiguity calls for policy clarification, and pluralism calls for deliberation about incorporating diverse perspectives. Yet understanding why annotators disagree is difficult. Directly asking annotators for their reasoning is costly, substantially increasing annotation burden, and can be unreliable for both human and LLM annotators as self-reported reasoning often fails to reflect actual decision processes. We introduce Annotator Policy Models (APMs), interpretable models that learn annotators' internal safety policies from labeling behavior alone, making annotator reasoning visible and comparable without additional annotation effort. We validate that APMs accurately model annotator safety policy (>80% accuracy), faithfully predict responses to counterfactual edits, and recover known policy differences in controlled settings. Applying APMs to LLM and human annotations, we demonstrate two core applications: (1) surfacing policy ambiguity by revealing how annotators interpret safety instructions differently, and (2) surfacing value pluralism by uncovering systematic differences in safety priorities across demographic groups. Together, these capabilities support more targeted, transparent, and inclusive safety policy design.\n\n---\n\n## 81. Verifier-Backed Hard Problem Generation for Mathematical Reasoning\n- **Authors**: Yuhang Lai , Jiazhan Feng , Yee Whye Teh , Ning Miao\n- **URL**: [https://arxiv.org/abs/2605.06660](https://arxiv.org/abs/2605.06660)\n- **Abstract**:\n> Large Language Models (LLMs) demonstrate strong capabilities for solving scientific and mathematical problems, yet they struggle to produce valid, challenging, and novel problems - an essential component for advancing LLM training and enabling autonomous scientific research. Existing problem generation approaches either depend on expensive human expert involvement or adopt naive self-play paradigms, which frequently yield invalid problems due to reward hacking. This work introduces VHG, a verifier-enhanced hard problem generation framework built upon three-party self-play. By integrating an independent verifier into the conventional setter-solver duality, our design constrains the setter's reward to be jointly determined by problem validity (evaluated by the verifier) and difficulty (assessed by the solver). We instantiate two verifier variants: a Hard symbolic verifier and a Soft LLM-based verifier, with evaluations conducted on indefinite integral tasks and general mathematical reasoning tasks. Experimental results show that VHG substantially outperforms all baseline methods by a clear margin.\n\n---\n\n## 82. Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less\n- **Authors**: Yuxing Liu , Jianyu Wang , Tong Zhang\n- **URL**: [https://arxiv.org/abs/2605.06654](https://arxiv.org/abs/2605.06654)\n- **Abstract**:\n> Optimizers play an important role in both pretraining and finetuning stages when training large language models (LLMs). In this paper, we present an observation that full finetuning with the same optimizer as in pretraining achieves a better learning-forgetting tradeoff, i.e., forgetting less while achieving the same or better performance on the new task, than other optimizers and, possibly surprisingly, LoRA, during the supervised finetuning (SFT) stage. We term this phenomenon optimizer-model consistency. To better understand it, through controlled experiments and theoretical analysis, we show that: 1) optimizers can shape the models by having regularization effects on the activations, leading to different landscapes around the pretrained checkpoints; 2) in response to this regularization effect, the weight update in SFT should follow some specific structures to lower forgetting of the knowledge learned in pretraining, which can be obtained by using the same optimizer. Moreover, we specifically compare Muon and AdamW when they are employed throughout the pretraining and SFT stages and find that Muon performs worse when finetuned for reasoning tasks. With a synthetic language modeling experiment, we demonstrate that this can come from Muon's strong tendency towards rote memorization, which may hurt pattern acquisition with a small amount of data, as for SFT.\n\n---\n\n## 83. When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels\n- **Authors**: Sushant Gautam , Finn Schwall , Annika Willoch Olstad , Fernando Vallecillos Ruiz , Birk Torpmann-Hagen , Sunniva Maria Stordal Bjørklund , Leon Moonen , Klas Pettersen , Michael A. Riegler\n- **URL**: [https://arxiv.org/abs/2605.06652](https://arxiv.org/abs/2605.06652)\n- **Abstract**:\n> Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component ($\\eta^2 \\approx 0.52$), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.\n\n---\n\n## 84. Superintelligent Retrieval Agent: The Next Frontier of Information Retrieval\n- **Authors**: Zeyu Yang , Qi Ma , Jason Chen , Anshumali Shrivastava\n- **URL**: [https://arxiv.org/abs/2605.06647](https://arxiv.org/abs/2605.06647)\n- **Abstract**:\n> Retrieval-augmented agents are increasingly the interface to large organizational knowledge bases, yet most still treat retrieval as a black box: they issue exploratory queries, inspect returned snippets, and iteratively reformulate until useful evidence emerges. This approach resembles how a newcomer searches an unfamiliar database rather than how an expert navigates it with strong priors about terminology and likely evidence, and results in unnecessary retrieval rounds, increased latency, and poor recall. We introduce \\textit{SuperIntelligent Retrieval Agent} (SIRA), which defines \\emph{superintelligence} in retrieval as the ability to compress multi-round exploratory search into a single corpus-discriminative retrieval action. SIRA does not merely ask what terms are relevant to the query; it asks which terms are likely to separate the desired evidence from corpus-level confusers. On the corpus side, an LLM enriches each document offline with missing search vocabulary; on the query side, it predicts evidence vocabulary omitted by the query; and document-frequency statistics as a tool call to filter proposed terms that are absent, overly common, or unlikely to create retrieval margin. The final retrieval step is a single weighted BM25 call combining the original query with the validated expansion. Across ten BEIR benchmarks and downstream question-answering tasks, SIRA achieves the significantly superior performance outperforming dense retrievers and state-of-the-art multi-round agentic baselines, demonstrating that one well-formed lexical query, guided by LLM cognition and lightweight corpus statistics, can exceed substantially more expensive multi-round search while remaining interpretable, training-free, and efficient.\n\n---\n\n## 85. StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction\n- **Authors**: Xiangyuan Xue , Yifan Zhou , Zidong Wang , Shengji Tang , Philip Torr , Wanli Ouyang , Lei Bai , Zhenfei Yin\n- **URL**: [https://arxiv.org/abs/2605.06642](https://arxiv.org/abs/2605.06642)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used as interactive agents, but optimizing them for long-horizon decision making remains difficult because current methods are largely purely reactive, which weakens both exploration and credit assignment over extended trajectories. In this work, we present Strategic Trajectory Abstraction (StraTA), a simple framework that introduces an explicit trajectory-level strategy into agentic reinforcement learning (RL). StraTA samples a compact strategy from the initial task state, conditions subsequent actions on that strategy, and trains strategy generation and action execution jointly with a hierarchical GRPO-style rollout design, further enhanced by diverse strategy rollout and critical self-judgment. Experiments on ALFWorld, WebShop, and SciWorld show that StraTA consistently improves both sample efficiency and final performance over strong baselines. StraTA reaches success rates of 93.1% on ALFWorld and 84.2% on WebShop. On SciWorld, StraTA attains a 63.5% overall score, outperforming frontier closed-source models.\n\n---\n\n## 86. The Structural Origin of Attention Sink: Variance Discrepancy, Super Neurons, and Dimension Disparity\n- **Authors**: Siquan Li , Kaiqi Jiang , Jiacheng Sun , Tianyang Hu\n- **URL**: [https://arxiv.org/abs/2605.06611](https://arxiv.org/abs/2605.06611)\n- **Abstract**:\n> Despite the prevalence of the attention sink phenomenon in Large Language Models (LLMs), where initial tokens disproportionately monopolize attention scores, its structural origins remain elusive. This work provides a \\textit{mechanistic explanation} for this phenomenon. First, we trace its root to the value aggregation process inherent in self-attention, which induces a systematic variance discrepancy. We further demonstrate that this discrepancy is drastically amplified by the activation of super neurons within Feed-Forward Network (FFN) layers. Specifically, the channel-sparse down-projections trigger a dimension disparity of the first-token representation, necessitating the formation of attention sinks as a structural anchor. Then, we validate this causal chain through two controlled interventions: (i) isolating the aggregation effect via attention mask modifications and (ii) amplifying the variance of targeted token representations. Both interventions can replicate attention sinks at arbitrary positions. Our mechanistic understanding offers a foundation for the systematic control of sink formation. Finally, as a proof of concept, we propose \\textit{head-wise RMSNorm}, an architectural modification that stabilizes value aggregation outputs during pre-training. Our experiments demonstrate that restoring statistical parity across positions significantly accelerates convergence.\n\n---\n\n## 87. AI CFD Scientist: Toward Open-Ended Computational Fluid Dynamics Discovery with Physics-Aware AI Agents\n- **Authors**: Nithin Somasekharan , Rabi Pathak , Manushri Dhanakoti , Tingwen Zhang , Ling Yue , Andy Zhu , Shaowu Pan\n- **URL**: [https://arxiv.org/abs/2605.06607](https://arxiv.org/abs/2605.06607)\n- **Abstract**:\n> Recent LLM-based agents have closed substantial portions of the scientific discovery loop in software-only machine-learning research, in chemistry, and in biology. Extending the same loop to high-fidelity physical simulators is harder, because solver completion does not imply physical validity and many failure modes appear only in field-level imagery rather than in solver logs. We present AI CFD Scientist, an open-source AI scientist for computational fluid dynamics (CFD) that, to our knowledge, is the first to span literature-grounded ideation, validated execution, vision-based physics verification, source-code modification, and figure-grounded writing within a single inspectable workflow. Three coupled pathways cover parameter sweeps within a fixed solver, case-local C++ library compilation for new physical models, and open-ended hypothesis search against a reference comparator, all running on OpenFOAM through Foam-Agent. At the center of the framework is a vision-language physics-verification gate that inspects rendered flow fields before any result is accepted, rerun, or written into a manuscript. On five tasks under a shared GPT-5.5 backbone, AI CFD Scientist autonomously discovers a Spalart-Allmaras runtime correction that reduces lower-wall Cf RMSE against DNS by 7.89% on the periodic hill at Reh=5600; under matched LLM cost, two strong general AI-scientist baselines (ARIS, DeepScientist) execute partial CFD workflows but lack the domain-specific validity gates needed to convert runs into defensible scientific claims; and a controlled planted-failure ablation shows that the vision-language gate detects 14 of 16 silent failures missed by solver-level checks. Code, prompts, and run artifacts are released at this https URL .\n\n---\n\n## 88. UniSD: Towards a Unified Self-Distillation Framework for Large Language Models\n- **Authors**: Yiqiao Jin , Yiyang Wang , Lucheng Fu , Yijia Xiao , Yinyi Luo , Haoxin Liu , B. Aditya Prakash , Josiah Hester , Jindong Wang , Srijan Kumar\n- **URL**: [https://arxiv.org/abs/2605.06597](https://arxiv.org/abs/2605.06597)\n- **Abstract**:\n> Self-distillation (SD) offers a promising path for adapting large language models (LLMs) without relying on stronger external teachers. However, SD in autoregressive LLMs remains challenging because self-generated trajectories are free-form, correctness is task-dependent, and plausible rationales can still provide unstable or unreliable supervision. Existing methods mainly examine isolated design choices, leaving their effectiveness, roles, and interactions unclear. In this paper, we propose UniSD, a unified framework to systematically study self-distillation. UniSD integrates complementary mechanisms that address supervision reliability, representation alignment, and training stability, including multi-teacher agreement, EMA teacher stabilization, token-level contrastive learning, feature matching, and divergence clipping. Across six benchmarks and six models from three model families, UniSD reveals when self-distillation improves over static imitation, which components drive the gains, and how these components interact across tasks. Guided by these insights, we construct UniSDfull, an integrated pipeline that combines complementary components and achieves the strongest overall performance, improving over the base model by +5.4 points and the strongest baseline by +2.8 points. Extensive evaluation highlights self-distillation as a practical and steerable approach for efficient LLM adaptation without stronger external teachers.\n\n---\n\n## 89. Continuous Latent Diffusion Language Model\n- **Authors**: Hongcan Guo , Qinyu Zhao , Yian Zhao , Shen Nie , Rui Zhu , Qiushan Guo , Feng Wang , Tao Yang , Hengshuang Zhao , Guoqiang Wei , Yan Zeng\n- **URL**: [https://arxiv.org/abs/2605.06548](https://arxiv.org/abs/2605.06548)\n- **Abstract**:\n> Large language models have achieved remarkable success under the autoregressive paradigm, yet high-quality text generation need not be tied to a fixed left-to-right order. Existing alternatives still struggle to jointly achieve generation efficiency, scalable representation learning, and effective global semantic modeling. We propose Cola DLM, a hierarchical latent diffusion language model that frames text generation through hierarchical information decomposition. Cola DLM first learns a stable text-to-latent mapping with a Text VAE, then models a global semantic prior in continuous latent space with a block-causal DiT, and finally generates text through conditional decoding. From a unified Markov-path perspective, its diffusion process performs latent prior transport rather than token-level observation recovery, thereby separating global semantic organization from local textual realization. This design yields a more flexible non-autoregressive inductive bias, supports semantic compression and prior fitting in continuous space, and naturally extends to other continuous modalities. Through experiments spanning 4 research questions, 8 benchmarks, strictly matched ~2B-parameter autoregressive and LLaDA baselines, and scaling curves up to about 2000 EFLOPs, we identify an effective overall configuration of Cola DLM and verify its strong scaling behavior for text generation. Taken together, the results establish hierarchical continuous latent prior modeling as a principled alternative to strictly token-level language modeling, where generation quality and scaling behavior may better reflect model capability than likelihood, while also suggesting a concrete path toward unified modeling across discrete text and continuous modalities.\n\n---\n\n## 90. Is One Layer Enough? Understanding Inference Dynamics in Tabular Foundation Models\n- **Authors**: Amir Rezaei Balef , Mykhailo Koshil , Katharina Eggensperger\n- **URL**: [https://arxiv.org/abs/2605.06510](https://arxiv.org/abs/2605.06510)\n- **Abstract**:\n> Transformer-based tabular foundation models (TFMs) dominate small to medium tabular predictive benchmark tasks, yet their inference mechanisms remain largely unexplored. We present the first large-scale mechanistic study of layerwise dynamics in 6 state-of-the-art tabular in-context learning models. We explore how predictions emerge across depth, identify distinct stages of inference and reveal latent-space dynamics that differ from those of language models. Our findings indicate substantial depthwise redundancy across multiple models, suggesting iterative refinement with overlapping computations during inference stages. Guided by these insights, we design a proof-of-concept, looped single-layer model that uses only 20% of the original model's parameters while achieving comparable performance. The code is available at this https URL .\n\n---\n\n## 91. PACZero: PAC-Private Fine-Tuning of Language Models via Sign Quantization\n- **Authors**: Murat Bilgehan Ertan , Xiaochen Zhu , Phuong Ha Nguyen , Marten van Dijk , Srinivas Devadas\n- **URL**: [https://arxiv.org/abs/2605.06505](https://arxiv.org/abs/2605.06505)\n- **Abstract**:\n> We introduce PACZero, a family of PAC-private zeroth-order mechanisms for fine-tuning large language models that delivers usable utility at $I(S^*; Y_{1:T})=0$. This privacy regime bounds the membership-inference attack (MIA) posterior success rate at the prior, an MIA-resistance level the DP framework matches only at $\\varepsilon=0$ and infinite noise. All DP-ZO comparisons below are matched at the MIA posterior level. The key insight is that PAC Privacy charges mutual information only when the release depends on which candidate subset is the secret. Sign-quantizing subset-aggregated zeroth-order gradients creates frequent unanimity, steps at which every candidate subset agrees on the update direction; at these steps the released sign costs zero conditional mutual information. We propose two variants that span the privacy-utility trade-off: PACZero-MI (budgeted MI via exact calibration on the binary release) and PACZero-ZPL ($I=0$ via a uniform coin flip on disagreement steps). We evaluate on SST-2 and SQuAD with OPT-1.3B and OPT-6.7B in both LoRA and full-parameter tracks. On SST-2 OPT-1.3B full fine-tuning at $I=0$, PACZero-ZPL reaches ${88.99\\pm0.91}$, within $2.1$pp of the non-private MeZO baseline ($91.1$ FT). No prior method produces usable utility in the high-privacy regime $\\varepsilon<1$, and PACZero-ZPL obtains competitive SST-2 accuracy and nontrivial SQuAD F1 across OPT-1.3B and OPT-6.7B at $I=0$.\n\n---\n\n## 92. Litespark Inference on Consumer CPUs: Custom SIMD Kernels for Ternary Neural Networks\n- **Authors**: Nii Osae Osae Dade , Tony Morri , Moinul Hossain Rahat , Sayandip Pal\n- **URL**: [https://arxiv.org/abs/2605.06485](https://arxiv.org/abs/2605.06485)\n- **Abstract**:\n> Large language models (LLMs) have transformed artificial intelligence, but their computational requirements remain prohibitive for most users. Standard inference demands expensive datacenter GPUs or cloud API access, leaving over one billion personal computers underutilized for AI workloads. Ternary models offer a path forward: their weights are constrained to {-1, 0, +1}, theoretically eliminating the need for floating-point multiplication. However, existing frameworks fail to exploit this structure, treating ternary models as dense floating-point networks. We address this gap with custom SIMD kernels that replace matrix multiplication with simple addition and subtraction operations, targeting the integer dot product instructions available on modern CPUs. Our implementation, Litespark-Inference, is pip-installable and integrates directly with Hugging-Face, achieving 9.2x faster time-to-first-token, 52x higher throughput, and 14x memory reduction compared to standard PyTorch inference on Apple Silicon, with similar speedups on Intel and AMD processors.\n\n---\n\n## 93. Constraint Decay: The Fragility of LLM Agents in Backend Code Generation\n- **Authors**: Francesco Dente , Dario Satriani , Paolo Papotti\n- **URL**: [https://arxiv.org/abs/2605.06445](https://arxiv.org/abs/2605.06445)\n- **Abstract**:\n> Large Language Model (LLM) agents demonstrate strong performance in autonomous code generation under loose specifications. However, production-grade software requires strict adherence to structural constraints, such as architectural patterns, databases, and object-relational mappings. Existing benchmarks often overlook these non-functional requirements, rewarding functionally correct but structurally arbitrary solutions. We present a systematic study evaluating how well agents handle structural constraints in multi-file backend generation. By fixing a unified API contract across 80 greenfield generation tasks and 20 feature-implementation tasks spanning eight web frameworks, we isolate the effect of structural complexity using a dual evaluation with end-to-end behavioral tests and static verifiers. Our findings reveal a phenomenon of constraint decay: as structural requirements accumulate, agent performance exhibits a substantial decline. Capable configurations lose 30 points on average in assertion pass rates from baseline to fully specified tasks, while some weaker configurations approach zero. Framework sensitivity analysis exposes significant performance disparities: agents succeed in minimal, explicit frameworks (e.g., Flask) but perform substantially worse on average in convention-heavy environments (e.g., FastAPI, Django). Finally, error analysis identifies data-layer defects (e.g., incorrect query composition and ORM runtime violations) as the leading root causes. This work highlights that jointly satisfying functional and structural requirements remains a key open challenge for coding agents.\n\n---\n\n## 94. Is Escalation Worth It? A Decision-Theoretic Characterization of LLM Cascades\n- **Authors**: Dylan Bouchard\n- **URL**: [https://arxiv.org/abs/2605.06350](https://arxiv.org/abs/2605.06350)\n- **Abstract**:\n> Model cascades, in which a cheap LLM defers to an expensive one on low-confidence queries, are widely used to navigate the cost-quality tradeoff at deployment. Existing approaches largely treat the deferral threshold as an empirical hyperparameter, with limited guidance on the geometry of the resulting cost-quality frontier over a model pool. We develop a decision-theoretic framework grounded in constrained optimization and duality. For a two-model cascade, we establish piecewise concavity of the cost-quality frontier on decreasing-benefit regions of the confidence support, with reciprocal shadow prices linking the budget- and quality-constrained formulations. Given a pool of $k$ models, we characterize the frontier achievable by deterministic two-model threshold cascades as the pointwise envelope over $\\binom{k}{2}$ pairwise cascades, with switching points where the optimal pair changes. For $k$-model cascades, we derive first-order conditions in which a single shadow price equalizes marginal quality-per-cost across stage boundaries. We validate the framework on five benchmarks (MATH, MMLU, TriviaQA, SimpleQA, LiveCodeBench) across eight models from five providers. Within the deterministic threshold-cascade class, full fixed chains underperform the pairwise envelope, and optimized subsequence cascades do not deliver practically meaningful held-out gains over it. A lightweight pre-generation router exceeds the best cascade policy on four of five datasets, mainly because it avoids the cheap model's generation cost on queries sent directly to a larger model rather than because of a stronger routing signal. These results suggest that cascade performance is limited primarily by structural cost, since cascades pay the cheap model before any escalation decision, rather than by a shortage of intermediate stages.\n\n---\n\n## 95. Human-AI Co-Evolution and Epistemic Collapse: A Dynamical Systems Perspective\n- **Authors**: Xuening Wu , Yanlan Kang , Qianya Xu , Kexuan Xie , Jiaqi Mi , Honggang Wang , Yubin Liu , Zeping Chen\n- **URL**: [https://arxiv.org/abs/2605.06347](https://arxiv.org/abs/2605.06347)\n- **Abstract**:\n> Large language models (LLMs) are reshaping how knowledge is produced, with increasing reliance on AI systems for generation, summarization, and reasoning. While prior work has studied cognitive offloading in humans and model collapse in recursive training, these effects are typically considered in isolation. We propose a unified perspective: humans and language models form a coupled dynamical system linked by a feedback loop of usage, generation, and retraining. We introduce a minimal model with three variables -- human cognition, data quality, and model capability -- and show that this feedback can give rise to distinct dynamical regimes. Our analysis identifies three regimes: co-evolutionary enhancement, fragile equilibrium, and degenerative convergence. Through a simple simulation, we demonstrate that increasing reliance on AI can induce a transition toward a low-diversity, suboptimal equilibrium. From an information-theoretic perspective, this transition corresponds to an emergent information bottleneck in the human-AI loop, where entropy reduction reflects loss of diversity and support under closed-loop feedback rather than beneficial compression. These results suggest that the trajectory of AI systems is shaped not only by model design, but by the dynamics of human-AI co-evolution.\n\n---\n\n## 96. CoupleEvo: Evolving Heuristics for Coupled Optimization Problems Using Large Language Models\n- **Authors**: Thomas Bömer , Bastian Amberg , Max Disselnmeyer , Anne Meyer\n- **URL**: [https://arxiv.org/abs/2605.06341](https://arxiv.org/abs/2605.06341)\n- **Abstract**:\n> Many real-world optimization problems consist of multiple tightly coupled subproblems whose solutions must be coordinated to achieve high overall performance. However, existing large language model driven automated heuristic design approaches are limited to single-problem settings. In this paper, we propose CoupleEvo. CoupleEvo proposes three evolutionary coordination strategies to evolve heuristics for coupled optimization problems: the sequential strategy evolves heuristics for one subproblem after the other; the iterative strategy alternates the evolution of heuristics for different subproblems over successive generations; and the integrated strategy evolves heuristics for all problems simultaneously. The approach is evaluated on two representative coupled optimization problems. Experimental results show that decomposition-based strategies (sequential and iterative) provide more stable convergence and higher solution quality, while the integrated evolution strategy suffers from increased search complexity and variability. These findings highlight the importance of coordinating evolutionary search across interdependent subproblems and demonstrate the potential of LLM-driven heuristic design for complex coupled optimization problems. The code is available: this https URL .\n\n---\n\n## 97. Fine-Tuning Small Language Models for Solution-Oriented Windows Event Log Analysis\n- **Authors**: Siraaj Akhtar , Saad Khan , Simon Parkinson\n- **URL**: [https://arxiv.org/abs/2605.06330](https://arxiv.org/abs/2605.06330)\n- **Abstract**:\n> Large language models (LLMs) have shown promise for event log analysis, but their high computational requirements, reliance on cloud infrastructure, and security concerns limit practical deployment. In addition, most existing approaches focus only on the identification of the problem and do not provide actionable remediation. Small language models (SLMs) present a light-weight alternative that can be fine-tuned for a specific purpose and hosted locally. This paper investigates whether SLMs, when fine-tuned for a specific task, can serve as a practical alternative for event log analysis while also generating solutions. We first create a large-scale synthetic Windows event log dataset that contains remediation actions using a high-performing LLM. We then fine-tune multiple SLMs and LLMs using the LoRA parameter-efficient fine-tuning technique and evaluate their performance by comparing with expert assessment. The results show that the dataset accurately reflects real-world scenarios and that fine-tuned SLMs consistently outperform LLMs in identifying issues and providing relevant remediation, while requiring fewer computational resources.\n\n---\n\n## 98. Measuring Evaluation-Context Divergence in Open-Weight LLMs: A Paired-Prompt Protocol with Pilot Evidence of Alignment-Pipeline-Specific Heterogeneity\n- **Authors**: Florian A. D. Burnat , Brittany I. Davidson\n- **URL**: [https://arxiv.org/abs/2605.06327](https://arxiv.org/abs/2605.06327)\n- **Abstract**:\n> Safety benchmarks are routinely treated as evidence about how a language model will behave once deployed, but this inference is fragile if behavior depends on whether a prompt looks like an evaluation. We define evaluation-context divergence as an observable within-item change in behavior induced by framing a fixed task as an evaluation, a live deployment interaction, or a neutral request, and present a paired-prompt protocol that measures it in open-weight LLMs while controlling for paraphrase variation, benchmark familiarity, and judge framing-sensitivity. Across five instruction-tuned checkpoints from four open-weight families plus a matched OLMo-3 base/instruct ablation ($20$ paired items, $840$ generations per checkpoint), we find striking heterogeneity. OLMo-3-Instruct alone is eval-cautious -- evaluation framing raises refusal vs. neutral by $11.8$pp ($p=0.007$) and reduces harmful compliance vs. deployment by $3.6$pp ($p=0.024$, $0/20$ items inverted) -- while Mistral-Small-3.2, Phi-3.5-mini, and Llama-3.1-8B are deployment-cautious}, with marginal eval-vs-deployment refusal effects of $-9$ to $-20$pp. The matched OLMo-3 base also exhibits the deployment-cautious pattern, identifying alignment as the inversion stage; within Llama-3.1, the $70$B model preserves direction with attenuated magnitude, ruling out a simple ``small-model effect that reverses at scale.'' One caveat: the cross-family heterogeneity is judge-dependent. Re-judging with a different-family safety classifier (Llama-Guard-3-8B) preserves the within-OLMo eval-cautious direction but flattens the cross-family contrast, indicating that the two judges operationalize distinct constructs.\n\n---\n\n## 99. Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs\n- **Authors**: Elizabeth Mieczkowski , Alexander Ku , Tiwalayo Eisape , Dilip Arumugam , John Matters , Katherine M. Collins , Ilia Sucholutsky , Thomas L. Griffiths\n- **URL**: [https://arxiv.org/abs/2605.06320](https://arxiv.org/abs/2605.06320)\n- **Abstract**:\n> Large language models (LLMs) are increasingly deployed in teams, yet existing coordination approaches often occupy two extremes. Highly structured methods rely on fixed roles, pipelines, or task decompositions assigned a priori. In contrast, fully unstructured teams enable adaptability and exploration but suffer from inefficiencies such as error propagation, inter-agent conflicts, and wasted resources (measured in time, tokens, or file operations). We introduce Language Agent Teams for Task Evolution (LATTE), a framework for coordinating LLM teams inspired by distributed systems, where processors must operate under partial observability and communication constraints. In LATTE, a team of agents collaboratively construct and maintain a shared, evolving coordination graph which encodes sub-task dependencies, individual agent assignment, and the current state of sub-task progress. This protocol maintains consistency while empowering agents to dynamically allocate work, adapt coordination, and discover new tasks. Across multiple collaborative tasks and a variety of base models, we demonstrate how LATTE reduces token usage, wall-clock time, communication, and coordination failures (e.g. file conflicts and redundant outputs) while matching or exceeding the accuracy of standard designs including MetaGPT, decentralized teams, top-down Leader-Worker hierarchies, and static decompositions.\n\n---\n\n## 100. Pro-KLShampoo: Projected KL-Shampoo with Whitening Recovered by Orthogonalization\n- **Authors**: Ruotong Sun , Ermin Wei\n- **URL**: [https://arxiv.org/abs/2605.06316](https://arxiv.org/abs/2605.06316)\n- **Abstract**:\n> Optimizers that exploit the matrix structure of gradients are central to modern LLM pre-training, with two distinct frontiers: explicit Kronecker-factored preconditioning -- most recently KL-Shampoo, which estimates the preconditioner via KL divergence minimization -- and orthogonalization of the gradient momentum, exemplified by Muon and analyzed as steepest descent under the spectral norm. The two routes are typically developed in isolation. We make a structural observation about KL-Shampoo's Kronecker preconditioners: their eigenvalue spectra exhibit a \\emph{spike-and-flat} shape -- a few dominant eigenvalues followed by an approximately uniform tail -- across layers and training stages, holding exactly under a rank-$\\rho$ signal-plus-noise gradient model. We exploit this structure by restricting one of KL-Shampoo's Kronecker factors to a parametric family aligned with the spike-and-flat shape: full spectral structure on a tracked $r$-dimensional subspace, single shared eigenvalue across the remaining $n-r$ directions. On these directions, we apply orthogonalization. An identity shows that this orthogonalization recovers the algebraic form of full KL-Shampoo's preconditioner. On four pre-training scales (GPT-2 124M / 350M, LLaMA 134M / 450M), Pro-KLShampoo consistently outperforms KL-Shampoo at every subspace rank we test in validation loss, peak per-GPU memory, and wallclock time to reach each loss level.\n\n---\n\n## 101. Attributions All the Way Down? The Metagame of Interpretability\n- **Authors**: Hubert Baniecki , Przemyslaw Biecek , Fabian Fumagalli\n- **URL**: [https://arxiv.org/abs/2605.06295](https://arxiv.org/abs/2605.06295)\n- **Abstract**:\n> We introduce the metagame, a conceptual framework for quantifying second-order interaction effects of model explanations. For any first-order attribution $\\phi(f)$ explaining a model $f$, we measure the directional influence of feature $j$ on the attribution of feature $i$, denoted as meta-attribution $\\varphi_{j \\to i}(f)$, by treating the attribution method itself as a cooperative game and computing its Shapley value. Theoretically, we prove that attributions hierarchically decompose into meta-attributions, and establish these as directional extensions of existing interaction indices. Empirically, we demonstrate that the metagame delivers insights across diverse interpretability applications: (i) quantifying token interactions in instruction-tuned language models, (ii) explaining cross-modal similarity in vision-language encoders, and (iii) interpreting text-to-image concepts in multimodal diffusion transformers.\n\n---\n\n## 102. Log-Likelihood, Simpson's Paradox, and the Detection of Machine-Generated Text\n- **Authors**: Tom Kempton , Viktor Drobnyi , Maeve Madigan , Stuart Burrell\n- **URL**: [https://arxiv.org/abs/2605.06294](https://arxiv.org/abs/2605.06294)\n- **Abstract**:\n> The ability to reliably distinguish human-written text from that generated by large language models is of profound societal importance. The dominant approach to this problem exploits the likelihood hypothesis: that machine-generated text should appear more probable to a detector language model than human-written text. However, we demonstrate that the token-level signal distinguishing human and machine text is non-uniform across the hidden space of the detector model, and naively averaging likelihood-based token scores across regions with fundamentally different statistical structure, as most detectors do, causes a form of Simpson's paradox: a strong local signal is destroyed by inappropriate aggregation. To correct for this, we introduce a learned local calibration step grounded in Bayesian decision theory. Rather than aggregating raw token scores, we first learn lightweight predictors of the score distributions conditioned on position in hidden space, and aggregate calibrated log-likelihood ratios instead. This single intervention dramatically and consistently improves detection performance across all baseline detectors and all datasets we consider. For example, our calibrated variant of Fast-DetectGPT improves AUROC from $0.63$ to $0.85$ on GPT-5.4 text, and a locally-calibrated DMAP detector we introduce achieves state-of-the-art performance across the board. That said, our central contribution is not a new detector, but a precise diagnosis of a significant cause of under-performance of existing detectors and a principled, modular remedy compatible with any token-averaging pipeline. This will serve as a foundation for the community to build upon, with natural avenues including richer distributional models, improved calibration strategies, and principled ensembling with hidden-space geometry signals via the full Bayes-optimal decision rule.\n\n---\n\n## 103. Correct Code, Vulnerable Dependencies: A Large Scale Measurement Study of LLM-Specified Library Versions\n- **Authors**: Chengjie Wang , Jingzheng Wu , Xiang Ling , Tianyue Luo , Chen Zhao\n- **URL**: [https://arxiv.org/abs/2605.06279](https://arxiv.org/abs/2605.06279)\n- **Abstract**:\n> Large language models (LLMs) are now largely involved in software development workflows, and the code they generate routinely includes third-party library (TPL) imports annotated with specific version identifiers. These version choices can carry security and compatibility risks, yet they have not been systematically studied. We present the first large-scale measurement study of version-level risk in LLM-generated Python code, evaluating 10 LLMs on PinTrace, a curated benchmark of 1,000 Stack Overflow programming tasks. LLMs tend to specify version identifiers when directly prompted at 26.83%-95.18%, while down to 6.45%-59.19% in creating a manifest file directly. Among the specified versions, 36.70%-55.70% of tasks contain at least one known CVE, and 62.75%-74.51% of them carry Critical or High severity ratings. In 72.27%-91.37% of cases, the associated CVEs were publicly disclosed before the model's knowledge cutoff. The statistics show all models converge on the same small set of risky release versions, indicating a systemic bias rather than isolated model error. Static compatibility rates range from 19.70% to 63.20%, with installation failure as the dominant cause. The dynamic test cases confirm the pattern by 6.49%-48.62% pass rates. Further experiments confirm that these failures are attributable to version selection rather than code quality, and that externally anchored version constraints substantially reduce both vulnerability exposure and compatibility failures. Our findings reveal LLM version selection as a first-class, previously overlooked risk surface in LLM-based development. We disclosed these findings to the community of the evaluated models, and several confirmed the issue. All the code and dataset have been released for open science at this https URL .\n\n---\n\n## 104. Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs\n- **Authors**: Andy Zeyi Liu , Michael Zhang , Ilana Greenberg , Adam Alnasser , Lucas Baker , John Sous\n- **URL**: [https://arxiv.org/abs/2605.06225](https://arxiv.org/abs/2605.06225)\n- **Abstract**:\n> Steering large language models (LLMs) is usually done by either instruction prompting or activation steering. Prompting often gives strong control, but caches guidance tokens at every layer and can clutter long interactions; activation steering is compact but typically weaker and does not support large structured reminders. We introduce memory inception (MI), a training-free method that steers in latent attention space by inserting text-derived key-value (KV) banks only at selected layers. Rather than materializing reminder content throughout the prompt cache, MI treats steering as selective KV allocation, injecting latent slots only where the model routes to them. On matched personality-steering tasks, MI gives the best overall control--drift trade-off, remaining competitive with prompting while consistently outperforming CAA. On updateable guidance, MI supports mid-conversation behavior shifts without rewriting the visible transcript, achieving the highest post-shift alignment on Qwen3. On structured reasoning, MI outperforms visible prompting on HARDMath and PHYSICS (10/12 subject$\\times$mode cells), serving as proxies for structured reasoning in verifiable domains, while cutting content-matched KV storage by up to 118$\\times$. These results position MI as a powerful steering method when guidance is persistent, structured, or expensive to keep in the visible transcript.\n\n---\n\n## 105. TIDE: Every Layer Knows the Token Beneath the Context\n- **Authors**: Ajay Jaiswal , Lauren Hannah , Han-Byul Kim , Duc Hoang , Mehrdad Farajtabar , Minsik Cho\n- **URL**: [https://arxiv.org/abs/2605.06216](https://arxiv.org/abs/2605.06216)\n- **Abstract**:\n> We revisit a universally accepted but under-examined design choice in every modern LLM: a token index is looked up once at the input embedding layer and then permanently discarded. This single-injection assumption induces two structural failures: (i) the Rare Token Problem, where a Zipf-type distribution of vocabulary causes rare-token embeddings are chronically under-trained due to receiving a fraction of the cumulative gradient signal compared to common tokens; and (ii) the Contextual Collapse Problem, where limited parameters models map distributionally similar tokens to indistinguishable hidden states. As an attempt to address both, we propose TIDE, which augments the standard transformer with EmbeddingMemory: an ensemble of K independent MemoryBlocks that map token indices to context-free semantic vectors, computed once and injected into every layer through a depth-conditioned softmax router with a learnable null bank. We theoretically and empirically establish the benefits of TIDE in addressing the issues associated with single-token identity injection as well as improve performance across multiple language modeling and downstream tasks.\n\n---\n\n## 106. Retina-RAG: Retrieval-Augmented Vision-Language Modeling for Joint Retinal Diagnosis and Clinical Report Generation\n- **Authors**: Abdelrahman Zaian , Sheethal Bhat , Mohamed Abdalkader , Andreas Maier\n- **URL**: [https://arxiv.org/abs/2605.06173](https://arxiv.org/abs/2605.06173)\n- **Abstract**:\n> Diabetic Retinopathy (DR) is a leading cause of preventable blindness among working-age adults worldwide, yet most automated screening systems are limited to image-level classification and lack clinically structured reporting. We propose Retina-RAG, a low-cost modular framework that jointly performs DR severity grading, macular edema (ME) detection, and report generation. The architecture decouples a high-performance retinal classifier and a parameter-efficient vision-language model (Qwen2.5-VL-7B-Instruct) adapted via Low-Rank Adaptation (LoRA), enabling flexible component integration. A retrieval-augmented generation (RAG) module injects curated ophthalmic knowledge together with structured classifier outputs at inference time to improve diagnostic consistency and reduce hallucinations. Retina-RAG achieves an F1-score of 0.731 for DR grading and 0.948 for ME detection, substantially outperforming zero-shot Qwen (0.096, 0.732) and MMed-RAG (0.541, 0.641) on a retinal disease detection dataset with captions. For report generation, Retina-RAG attains ROUGE-L 0.429 and SBERT similarity 0.884, exceeding all baselines. The full framework operates on a single consumer-grade GPU, demonstrating that clinically structured retinal AI can be achieved with modest computational resources.\n\n---\n\n## 107. IRC-Bench: Recognizing Entities from Contextual Cues in First-Person Reminiscences\n- **Authors**: Yehudit Aperstein , Eden Moran , Alexander Apartsin\n- **URL**: [https://arxiv.org/abs/2605.06142](https://arxiv.org/abs/2605.06142)\n- **Abstract**:\n> When people recount personal memories, they often refer to people, places, and events indirectly, relying on contextual cues rather than explicit names. Such implicit references are central to reminiscence narratives: first-person accounts of lived experience used in therapeutic, archival, and social settings. They pose a difficult computational problem because the intended entity must be inferred from dispersed narrative evidence rather than from a local mention. We introduce IRC-Bench, the Implicit Reminiscence Context Benchmark, for evaluating implicit entity recognition in reminiscence transcripts. The benchmark targets non-locality: entity-identifying cues are distributed across multiple, non-contiguous clauses, unlike named entity recognition, entity linking, or coreference resolution. IRC-Bench comprises 25,136 samples constructed from 12,337 Wiki-data-linked entities across 1,994 transcripts spanning 11 thematic domains. Each sample pairs an Entity-Grounded Narrative, in which the target entity is explicitly mentioned, with an Entity-Elided Narrative, in which direct mentions are removed. We evaluate 19 configurations across LLM generation, dense retrieval, RAG, and fine-tuning. QLoRA-adapted Llama 3.1 8B performs best in the open-world setting (38.94% exact match; 51.59% Jaccard), while fine-tuned DPR leads closed-world retrieval (35.38% Hit@1; 71.49% Hit@10). We release IRC-Bench with data, code, and evaluation tools.\n\n---\n\n## 108. Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex\n- **Authors**: Yun Qu , Qi Wang , Yixiu Mao , Heming Zou , Yuhang Jiang , Yingyue Li , Wutong Xu , Lizhou Cai , Weijie Liu , Clive Bai , Kai Yang , Yangkun Chen , Saiyong Yang , Xiangyang Ji\n- **URL**: [https://arxiv.org/abs/2605.06139](https://arxiv.org/abs/2605.06139)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has become a standard approach for large language models (LLMs) post-training to incentivize reasoning capacity. Among existing recipes, group-based policy gradient is prevalent, which samples a group of responses per prompt and updates the policy via group-relative advantage signals. This work reveals that these optimization strategies share a common geometric structure: each implicitly defines a target distribution on the response simplex and projects toward it via first-order approximation. Building on this insight, we propose Listwise Policy Optimization (LPO) to explicitly conduct the target-projection, which demystifies the implicit target by restricting the proximal RL objective to the response simplex, and then projects the policy via exact divergence minimization. This framework provides (i) monotonic improvement on the listwise objective with bounded, zero-sum, and self-correcting projection gradients, and (ii) flexibility in divergence selection with distinct structural properties through the decoupled projection step. On diverse reasoning tasks and LLM backbones, LPO consistently improves training performance over typical policy gradient baselines under matched targets, while intrinsically preserving optimization stability and response diversity.\n\n---\n\n## 109. Normalized Architectures are Natively 4-Bit\n- **Authors**: Maxim Fishman , Brian Chmiel , Ron Banner , Daniel Soudry , Boris Ginsburg\n- **URL**: [https://arxiv.org/abs/2605.06067](https://arxiv.org/abs/2605.06067)\n- **Abstract**:\n> Training large language models at 4-bit precision is critical for efficiency. We show that nGPT, an architecture that constrains weights and hidden representations to the unit hypersphere, is inherently more robust to low-precision arithmetic. This removes the need for interventions-such as applying random Hadamard transforms and performing per-tensor scaling calculations-to preserve model quality, and it enables stable end-to-end NVFP4 training. We validate this approach on both a 1.2B dense model and hybrid (Mamba-Transformer) MoE models of up to 3B/30B parameters. We trace this robustness to the dot product: while quantization noise remains largely uncorrelated in both standard and normalized architectures, the signal behaves differently. In nGPT, the hypersphere constraint enhances weak positive correlations among the element-wise products, leading to a constructive accumulation of the signal across the hidden dimension while the noise continues to average out. This yields a higher effective signal-to-noise ratio and a flatter loss landscape, with the effect strengthening as the hidden dimension grows, suggesting increasing advantages at scale. A reference implementation is available at this https URL\n\n---\n\n## 110. Causal Reinforcement Learning for Complex Card Games: A Magic The Gathering Benchmark\n- **Authors**: Cristiano da Costa Cunha , Ajmal Mian , Tim French , Wei Liu\n- **URL**: [https://arxiv.org/abs/2605.06066](https://arxiv.org/abs/2605.06066)\n- **Abstract**:\n> Causal reinforcement learning (RL) lacks benchmarks for complex systems that combine sequential decision making, hidden information, large masked action spaces, and explicit causal structure. We introduce MTG-Causal-RL, a Gymnasium benchmark built on Magic: The Gathering with a 3,077-dimensional partial observation, a 478-action masked discrete action space, five competitive Standard archetypes, three reward schemes, and a hand-specified Structural Causal Model (SCM) over strategic variables. Every episode exposes causal variables, SCM-predicted intervention effects, and per-factor credit traces, making causal credit assignment, leave-one-out cross-archetype transfer, and policy auditability first-class metrics. We adapt a panel of reference baselines: random, heuristic, masked PPO, a causal-world-model PPO variant, and an architecture-matched scalar control. We propose Causal Graph-Factored Advantage PPO (CGFA-PPO) as a reference causal agent that uses SCM parents of win probability as factor-aligned critic targets with an intervention-calibration loss. All comparisons use paired seeds, paired-bootstrap confidence intervals, and Holm-Bonferroni correction within pre-registered families. Masked PPO and CGFA-PPO reach competitive in-distribution win rates and exceed the random baseline; per-factor calibration trajectories and leave-one-out transfer gaps expose diagnostic structure that scalar win rate alone cannot. We release the benchmark, reference-baseline results, and full evaluation protocol openly. By coupling a strategically rich, partially observed domain with an explicit causal interface and statistical protocol, MTG-Causal-RL gives causal-RL, world-model, and LLM-agent research a shared testbed for questions current benchmarks cannot pose together: causal credit assignment under masked action spaces, structural transfer across archetypes, and SCM-grounded policy auditability.\n\n---\n\n## 111. Optimal Transport for LLM Reward Modeling from Noisy Preference\n- **Authors**: Licheng Pan , Haochen Yang , Haoxuan Li , Yunsheng Lu , Yongqi Tong , Yinuo Wang , Shijian Wang , Zhixuan Chu , Lei Shen , Yuan Lu , Hao Wang\n- **URL**: [https://arxiv.org/abs/2605.06036](https://arxiv.org/abs/2605.06036)\n- **Abstract**:\n> Reward models are fundamental to Reinforcement Learning from Human Feedback (RLHF), yet real-world datasets are inevitably corrupted by noisy preference. Conventional training objectives tend to overfit these errors, while existing denoising approaches often rely on homogeneous noise assumptions that fail to capture the complexity of linguistic preferences. To handle these challenges, we propose SelectiveRM, a framework grounded in optimal transport. We first devise a Joint Consistency Discrepancy to align the distribution of model predictions with preference data. Furthermore, to address the limitation of strict mass conservation which compels the model to fit outliers, we incorporate a Mass Relaxation mechanism via partial transport. This enables the autonomous exclusion of samples with noisy preference that contradict semantic consistency. Theoretically, we demonstrate that SelectiveRM optimizes a tighter upper bound on the unobserved clean risk. Extensive experiments validate that our approach significantly outperforms state-of-the-art baselines across diverse benchmarks.\n\n---\n\n## 112. Does Synthetic Data Help? Empirical Evidence from Deep Learning Time Series Forecasters\n- **Authors**: Hugo Cazaux , Eyjólfur Ingi Ásgeirsson , Hlynur Stefánsson\n- **URL**: [https://arxiv.org/abs/2605.06032](https://arxiv.org/abs/2605.06032)\n- **Abstract**:\n> Synthetic data has transformed language model training, yet its role in time series forecasting remains poorly understood. We present a large-scale empirical study: nine experiment groups, 4,218 runs systematically evaluating synthetic time series augmentation across five architectures, four synthetic signals and seven datasets. The effect is sharply architecture-conditional: channel-mixing models (TimesNet, iTransformer) benefit in the majority of trials, while channel-independent models (DLinear, PatchTST) are consistently degraded. In selected low-resource settings the gains are striking: TimesNet trained on only 10\\% of Weather data with synthetic augmentation surpasses the full-data baseline (4 of 16 sparsity-dataset combinations). Averaged across all architectures, augmentation hurts in 67\\% of trials. We further find that only the Seasonal-Trend generator reliably helps across the tested benchmarks, and that hard curriculum switching is actively harmful (+24\\% MSE degradation). These results provide concrete, actionable guidelines on how to use synthetic data: use synthetic augmentation with channel-mixing architectures, use gradual annealing schedules, and treat low-resource augmentation as architecture- and dataset-dependent. Code is available at \\href{ this https URL }\n\n---\n\n## 113. Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks\n- **Authors**: Guoxin Lu , Letian Sha , Qing Wang , Peijie Sun , Hao Zhou , Hua Dai , Fu Xiao\n- **URL**: [https://arxiv.org/abs/2605.05995](https://arxiv.org/abs/2605.05995)\n- **Abstract**:\n> The safety alignment of Large Language Models (LLMs) remains vulnerable to Harmful Fine-tuning (HFT). While existing defenses impose constraints on parameters, gradients, or internal representations, we observe that they can be effectively circumvented under persistent HFT. Our analysis traces this failure to the inherent redundancy of the high-dimensional parameter space: attackers exploit optimization trajectories that are orthogonal to defense constraints to restore harmful capabilities while deceptively adhering to safety restrictions. To address this, we propose Safety Bottleneck Regularization (SBR). SBR shifts the defensive focus from the redundant parameter space to the unembedding layer, which serves as a geometric bottleneck. By anchoring the final hidden states of harmful queries to those of the safety-aligned model, SBR enables the model to maintain safe responses even under persistent HFT. Extensive experiments confirm SBR's effectiveness, demonstrating that utilizing just a single safety anchor is sufficient to reduce the Harmful Score to $<$10 while preserving competitive performance on benign downstream tasks.\n\n---\n\n## 114. PragLocker: Protecting Agent Intellectual Property in Untrusted Deployments via Non-Portable Prompts\n- **Authors**: Qinfeng Li , Yuntai Bao , Jianghui Hu , Wenqi Zhang , Jintao Chen , Huifeng Zhu , Yier Jin , Xuhong Zhang\n- **URL**: [https://arxiv.org/abs/2605.05974](https://arxiv.org/abs/2605.05974)\n- **Abstract**:\n> LLM agents rely on prompts to implement task-specific capabilities based on foundation LLMs, making agent prompts valuable intellectual property. However, in untrusted deployments, adversaries can copy and reuse these prompts with other proprietary LLMs, causing economic losses. To protect these prompts, we identify four key challenges: proactivity, runtime protection, usability, and non-portability that existing approaches fail to address. We present PragLocker, a prompt protection scheme that satisfies these requirements. PragLocker constructs function-preserving obfuscated prompts by anchoring semantics with code symbols and then using target-model feedback to inject noise, yielding prompts that only work on the target LLM. Experiments across multiple agent systems, datasets, and foundation LLMs show that PragLocker substantially reduces cross-LLM portability, maintains target performance, and remains robust against adaptive attackers.\n\n---\n\n## 115. Towards Reliable LLM Evaluation: Correcting the Winner's Curse in Adaptive Benchmarking\n- **Authors**: Yang Xu , Jiefu Zhang , Haixiang Sun , Zihan Zhou , Tianyu Cao , Vaneet Aggarwal\n- **URL**: [https://arxiv.org/abs/2605.05973](https://arxiv.org/abs/2605.05973)\n- **Abstract**:\n> Adaptive prompt and program search makes LLM evaluation selection-sensitive. Once benchmark items are reused inside tuning, the observed winner's score need not estimate the fresh-data performance of the full tune-then-deploy procedure. We study inference for this procedure-level target under explicit tuning budgets. We propose SIREN, a selection-aware repeated-split reporting protocol that freezes the post-search shortlist, separates splitwise selection from held-out evaluation, and uses an item-level Gaussian multiplier bootstrap for uncertainty quantification. In a fixed-shortlist regime with smooth stabilized selection, the estimator admits a first-order item-level representation, and the bootstrap yields valid simultaneous inference on a finite budget grid. This supports confidence intervals for procedure-performance curves and pre-specified equal-budget and cross-budget comparisons. Controlled simulations and MMLU-Pro tuning experiments show that winner-based reporting can be optimistic and can change deployment conclusions, while SIREN remains close to the finite-sample reporting target.\n\n---\n\n## 116. Beyond Uniform Credit Assignment: Selective Eligibility Traces for RLVR\n- **Authors**: Chaoli Mou , Zhan Zhuang , Xinning Chen , Yu Zhang\n- **URL**: [https://arxiv.org/abs/2605.05965](https://arxiv.org/abs/2605.05965)\n- **Abstract**:\n> Reinforcement Learning with Verifiable Rewards (RLVR) has become a key approach for improving the reasoning abilities of large language models. However, widely used critic-free algorithms such as Group Relative Policy Optimization (GRPO) necessitate a ``uniform credit assignment'' assumption that indiscriminately broadcast trajectory-level advantages, hindering learning efficiency by failing to distinguish critical reasoning steps. To address this limitation, we propose Selective Eligibility Traces (S-trace). Grounded in the intuition of partial trust region preservation, we initially introduce P-trace as a sample-efficient, critic-free eligibility traces method, upon which we build S-trace, implementing a sparse eligibility traces mechanism to further mitigate variance and achieve fine-grained credit assignment by selectively masking low-entropy tokens. Theoretically, we contextualize the recent Group Sequence Policy Optimization (GSPO) method within the critic-free eligibility traces framework, identifying it as a special instance of the eligibility traces method operating under uniform credit assignment. Experiments demonstrate that S-trace not only outperforms GRPO, showing gains of 0.49\\% on Qwen3-1.7B and 3.16\\% on Qwen3-4B, and maintaining a robust 2.98\\% improvement when scaled further to Qwen3-8B in average pass@16, but notably achieves this with simultaneously higher sample and token efficiency.\n\n---\n\n## 117. Hallucination as an Anomaly: Dynamic Intervention via Probabilistic Circuits\n- **Authors**: Erik Nielsen , Elia Cunegatti , Marcus Vukojevic , Giovanni Iacca\n- **URL**: [https://arxiv.org/abs/2605.05953](https://arxiv.org/abs/2605.05953)\n- **Abstract**:\n> One of the most critical challenges in Large Language Models is their tendency to hallucinate, i.e., produce factually incorrect responses. Existing approaches show promising results in terms of hallucination correction, but still suffer from a main limitation: they apply corrections indiscriminately to every token, corrupting also the originally correct generations. To overcome this drawback, we propose PCNET, a Probabilistic Circuit trained as a tractable density estimator over the LLM residual stream. The method detects hallucinations as geometric anomalies on the factual manifold, which is done via exact Negative Log-Likelihood computation, hence without the need for sampling, external verifiers, or weight modifications, as in existing techniques. To demonstrate its effectiveness, we exploit PCNET as a dynamic gate that distinguishes hallucinated from factual hidden states at each decoding step. This triggers our second main contribution, PC-LDCD (Probabilistic Circuit Latent Density Contrastive Decoding), only when the latent geometry deviates from factual regions, while leaving correct generations untouched. Across four LLMs, ranging from 1B to 8B models, and four benchmarks covering conversational reasoning, knowledge-intensive QA, reading comprehension, and truthfulness, PCNET achieves near-perfect hallucination detection across CoQA, SQuAD v2.0, and TriviaQA, with AUROC reaching up to 99%. Moreover, PC-LDCD obtains the highest True+Info, MC2, and MC3 scores on TruthfulQA in three out of four models, in comparison with state-of-the-art baselines, while reducing the mean corruption rate to 53.7% and achieving a preservation rate of 79.3%. Our proposed method is publicly available on GitHub.\n\n---\n\n## 118. LLM-Driven Design Space Exploration of FPGA-based Accelerators\n- **Authors**: Vinamra Sharma , Xingjian Fu , Jude Haris , José Cano\n- **URL**: [https://arxiv.org/abs/2605.05920](https://arxiv.org/abs/2605.05920)\n- **Abstract**:\n> Designing field-programmable gate array (FPGA)-based accelerators for modern artificial intelligence workloads requires navigating a large and complex hardware design space encompassing architectural parameters, dataflow strategies, and memory hierarchies, making the process time-consuming and resource-intensive. While the SECDA methodology enables rapid hardware-software co-design of accelerators through SystemC simulation and FPGA execution, identifying optimal accelerator configurations still requires substantial manual effort and domain expertise. This work presents SECDA-DSE, a framework that integrates Large Language Models (LLMs) into the SECDA ecosystem, comprising tools built around SECDA to automate the design space exploration (DSE) of FPGA-based accelerators. SECDA-DSE combines a structured DSE Explorer for generating accelerator configurations with an LLM Stack that performs reasoning-guided exploration using retrieval-augmented generation and chain-of-thought prompting, alongside a feedback loop that enables reinforced fine-tuning for continuous improvement. We demonstrate the feasibility of SECDA-DSE through an initial high-level synthesis based evaluation of a generated accelerator design that meets synthesis timing and resource constraints on an Zynq-7000 FPGA.\n\n---\n\n## 119. Quantum-enhanced Large Language Models on Quantum Hardware via Cayley Unitary Adapters\n- **Authors**: Borja Aizpurua , Sukhbinder Singh , Augustine Kshetrimayum , Saeed S. Jahromi , Roman Orus\n- **URL**: [https://arxiv.org/abs/2605.05914](https://arxiv.org/abs/2605.05914)\n- **Abstract**:\n> Large language models (LLMs) have transformed artificial intelligence, yet classical architectures impose a fundamental constraint: every trainable parameter demands classical memory that scales unfavourably with model size. Quantum computing offers a qualitatively different pathway, but practical demonstrations on real hardware have remained elusive for models of practical relevance. Here we show that Cayley-parameterised unitary adapters -- quantum circuit blocks inserted into the frozen projection layers of pre-trained LLMs and executed on a 156-qubit IBM Quantum System Two superconducting processor -- improve the perplexity of Llama 3.1 8B, an 8-billion-parameter model in widespread use, by 1.4% with only 6,000 additional parameters and end-to-end inference validated on real Quantum Processing Unit (QPU). A systematic study on SmolLM2 (135M parameters), chosen for its tractability, reveals monotonically improving perplexity with unitary block dimension, 83% recovery of compression-induced degradation, and correct answers to questions that both classical baselines fail -- with a sharp noise-expressivity phase transition identifying the concrete path to quantum utility at larger qubit scales.\n\n---\n\n## 120. CITE: Anytime-Valid Statistical Inference in LLM Self-Consistency\n- **Authors**: Hirofumi Ota , Naoto Iwase , Yuki Ichihara , Junpei Komiyama , Masaaki Imaizumi\n- **URL**: [https://arxiv.org/abs/2605.05873](https://arxiv.org/abs/2605.05873)\n- **Abstract**:\n> Large language models often improve reasoning by sampling multiple outputs and aggregating their final answers, but precise and efficient control of error levels remains a challenging task. In particular, deciding when to stop sampling remains difficult when the stopping rule is data-dependent and the set of possible answers is not known in advance. We study anytime-valid certification of a prespecified target answer as the unique mode of the model's response distribution, a guarantee distinct from answer correctness. We propose the Certification by Intersection-union Testing with E-processes (CITE) algorithm, which provably controls false certification at any prescribed level under arbitrary data-driven stopping, without requiring prior knowledge of the answer category set. We also prove an category-set-size-free stopping-time rate, establish matching minimax lower bounds up to constants in the main regime, and extend the construction to confidence-weighted voting. Simulations and LLM self-consistency experiments show empirical error control and improved certification in diffuse-tail settings.\n\n---\n\n## 121. VideoRouter: Query-Adaptive Dual Routing for Efficient Long-Video Understanding\n- **Authors**: Kuanwei Lin , Wenhao Zhang , Ge Li\n- **URL**: [https://arxiv.org/abs/2605.05848](https://arxiv.org/abs/2605.05848)\n- **Abstract**:\n> Video large multimodal models increasingly face a scalability bottleneck: long videos produce excessively long visual-token sequences, which sharply increase memory and latency during inference. While existing compression methods are effective in specific settings, most are either weakly query-aware or apply a fixed compression policy across frames, proving suboptimal when visual evidence is unevenly distributed over time. To address this, we present VideoRouter, a query-adaptive dual-router framework built on InternVL for budgeted evidence allocation. The Semantic Router predicts the dominant allocation policy, choosing between broad temporal coverage and adaptive high-resolution preservation, while the Image Router uses early LLM layers to score frame relevance. This enables aggressive compression on less relevant frames while preserving detail on critical evidence frames. To train both routers, we build Video-QTR-10K for allocation-policy supervision and Video-FLR-200K for frame-relevance supervision. Experiments on VideoMME, MLVU, and LongVideoBench show that VideoRouter consistently improves over the InternVL baseline under comparable or lower budgets, achieving up to a 67.9% token reduction.\n\n---\n\n## 122. LoopTrap: Termination Poisoning Attacks on LLM Agents\n- **Authors**: Huiyu Xu , Zhibo Wang , Wenhui Zhang , Ziqi Zhu , Yaopeng Wang , Kui Ren , Chun Chen\n- **URL**: [https://arxiv.org/abs/2605.05846](https://arxiv.org/abs/2605.05846)\n- **Abstract**:\n> Modern LLM agents solve complex tasks by operating in iterative execution loops, where they repeatedly reason, act, and self-evaluate progress to determine when a task is complete. In this work, we show that while this self-directed loop facilitates autonomy, it also introduces a critical risk: by injecting malicious prompts into the agent's context, an adversary can distort the agent's termination judgment, making it believe the task remains incomplete and leading to unbounded this http URL understand this threat, we define and systematically characterize it as Termination Poisoning and design 10 representative attack strategies. Through a empirical study spanning 8 LLM agents and 60 tasks, we demonstrate that different LLM agents exhibit distinct behavioral signatures that determine which strategies succeed. These transferable patterns can serve as principled guidance for crafting effective attacks against previously unseen agents and tasks, enabling scalable red-teaming beyond manually designed templates. Building on these insights, we introduce LoopTrap, an automated red-teaming framework that synthesizes target-specific malicious prompts by exploiting agent behavioral tendencies. LoopTrap first constructs a behavioral profile of the target agent along four vulnerability dimensions via lightweight probing. It then performs adaptive trap synthesis, routing to the most effective strategy and selecting optimal injections via a self-scoring mechanism. Finally, successful traps are abstracted into a reusable skill library, while failed attempts are refined through self-reflection, ensuring continuous improvement. Extensive evaluation shows that LoopTrap achieves an average of 3.57$\\times$ step amplification across 8 mainstream agents, with a peak of 25$\\times$.\n\n---\n\n## 123. LeakDojo: Decoding the Leakage Threats of RAG Systems\n- **Authors**: Maosen Zhang , Jianshuo Dong , Boting Lu , Wenyue Li , Xiaoping Zhang , Tianwei Zhang , Han Qiu\n- **URL**: [https://arxiv.org/abs/2605.05818](https://arxiv.org/abs/2605.05818)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to leverage external knowledge, but also exposes valuable RAG databases to leakage attacks. As RAG systems grow more complex and LLMs exhibit stronger instruction-following capabilities, existing studies fall short of systematically assessing RAG leakage risks. We present LeakDojo, a configurable framework for controlled evaluation of RAG leakage. Using LeakDojo, we benchmark six existing attacks across fourteen LLMs, four datasets, and diverse RAG systems. Our study reveals that (1) query generation and adversarial instructions contribute independently to leakage, with overall leakage well approximated by their product; (2) stronger instruction-following capability correlates with higher leakage risk; and (3) improvements in RAG faithfulness can introduce increased leakage risk. These findings provide actionable insights for understanding and mitigating RAG leakage in practice. Our codebase is available at this https URL .\n\n---\n\n## 124. LCC-LLM: Leveraging Code-Centric Large Language Models for Malware Attribution\n- **Authors**: Christopher G. Pedraza Pohlenz , Hassan Jalil Hadi , Ali Hassan , Ali Shoker\n- **URL**: [https://arxiv.org/abs/2605.05807](https://arxiv.org/abs/2605.05807)\n- **Abstract**:\n> LLMs are increasingly explored for malware analysis; however, current LLM-based malware attribution remains limited by unsupported indicators and insufficient code-level grounding for identifying malicious and vulnerable code segments. To address these limitations, this research introduces LCC-LLM, a code-centric benchmark dataset and evidence-grounded framework for malware attribution and multi-task static malware analysis. The proposed LCCD dataset contains approximately 34K PE samples processed through a large-scale reverse-engineering pipeline and represented using decompiled C code, assembly code, CFG/FCG artifacts, hexadecimal data, PE metadata, suspicious API evidence, and structural features. Beyond dataset construction, LCC-LLM integrates LangGraph-orchestrated static analysis with multi-source cybersecurity knowledge to support evidence-grounded malware reasoning. The framework employs a seven-layer retrieval-augmented generation pipeline, CoVe for IoC validation, and a multi-dimensional quality gate to improve factual reliability and analyst-oriented decision support. Curriculum-ordered instruction data is used to fine-tune DeepSeek-R1-Distill-Qwen-14B and Qwen3-Coder-30B-A3B using QLoRA. Evaluation across 43 malware-analysis task types achieves an average semantic similarity of 0.634, with the highest task-level performance in structured report generation, IoC extraction, vulnerability assessment, malware configuration extraction, and malware class detection. In a real-world case study using MalwareBazaar samples, the grounded pipeline achieves a 10/10 structured analysis pass rate, producing CFG/FCG evidence, MITRE ATT&CK mappings, detection guidance, and analyst-ready reports. These results show that code-centric representations, retrieval grounding, and verification-guided reasoning improve the reliability and operational usefulness of LLM-assisted malware attribution.\n\n---\n\n## 125. Revealing Modular Gradient Noise Imbalance in LLMs: Calibrating Adam via Signal-to-Noise Ratio\n- **Authors**: Ziqing Wen , Zhouyang Liu , Jiahuan Wang , Ping Luo , Li Shen , Dongsheng Li , Tao Sun\n- **URL**: [https://arxiv.org/abs/2605.05794](https://arxiv.org/abs/2605.05794)\n- **Abstract**:\n> The impressive performance of large language models (LLMs) arises from their massive scale and heterogeneous module composition. However, this structural heterogeneity introduces additional optimization challenges. While adaptive optimizers such as Adam(W) provide per-parameter adaptivity, they do not explicitly account for module-level gradient heterogeneity, resulting in slower convergence, suboptimal performance, or training instability. Existing approaches typically rely on manually tuned module-specific learning rates or specific optimization strategies, which are computationally costly and difficult to generalize across tasks or models. To establish a more principled approach, we first analyze the noise-damping behavior of Adam in high-noise modules and introduce \\textbf{Module-wise Learning Rate Scaling via SNR (MoLS)}. MoLS estimates module-level SNRs to scale Adam updates, allowing automated module-wise learning rate allocation without manual tuning. Empirical results through multiple LLM training benchmarks demonstrate that MoLS improves convergence speed and generalization, achieving performance comparable to carefully tuned module-specific learning rates, while remaining compatible with memory-efficient training algorithms.\n\n---\n\n## 126. Multi-Dimensional Behavioral Evaluation of Agentic Stock Prediction Systems Using LLM Judges with Closed-Loop Reinforcement Learning Feedback\n- **Authors**: Mohammad Al Ridhawi , Mahtab Haj Ali , Hussein Al Osman\n- **URL**: [https://arxiv.org/abs/2605.05739](https://arxiv.org/abs/2605.05739)\n- **Abstract**:\n> Agentic stock prediction systems make sequences of interdependent decisions (regime detection, pathway routing, reinforcement learning control) whose individual quality is hidden by aggregate metrics such as mean absolute percentage error (MAPE) or directional accuracy. We present a behavioral evaluation framework that addresses this gap. Behavioral traces logged at every autonomous decision point are grouped into five-day episodes and scored along six domain-specific dimensions (regime detection, routing, adaptation, risk calibration, strategy coherence, error recovery) by an ensemble of three large language model (LLM) judges (GPT 5.4, Claude 4.6 Opus, Gemini 3.1 Pro). Perturbation-based validation on 420 episodes yields targeted score drops of $-1.6$ to $-2.4$ on intended dimensions versus an average of $-0.32$ on the remaining five, with cross-model agreement up to Krippendorff's $\\alpha = 0.85$. The composite behavioral score, used here only for cross-episode reporting, correlates at $\\rho = 0.72$ with realized 20-day Sharpe ratio from offline backtesting. Closing the loop, the framework converts deficient per-dimension scores into a credit-assigned penalty term added to the Soft Actor-Critic (SAC) reward. Three short fine-tuning cycles, all confined to the validation period, produce on the held-out 2017-2025 test period a one-day MAPE reduction from 0.61% to 0.54% (an 11.5% relative reduction; $p<0.001$, Cohen's $d=0.31$), a directional accuracy increase from 71% to 74%, and an 18% Sharpe ratio improvement (95% bootstrap CI [8.2%, 27.4%]), with gains concentrated in high-volatility episodes where the original system was most behaviorally deficient. Results are from offline backtesting and do not address effects specific to live deployment.\n\n---\n\n## 127. CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning\n- **Authors**: Md Anwar Hossen , Fatema Siddika , Juan Pablo Munoz , Tanya Roosta , Ali Jannesari\n- **URL**: [https://arxiv.org/abs/2605.05732](https://arxiv.org/abs/2605.05732)\n- **Abstract**:\n> Large language models (LLMs) can acquire new capabilities through fine-tuning, but continual adaptation often leads to catastrophic forgetting. We propose CRAFT, a continual learning framework that avoids updating model weights by instead learning low-rank interventions on hidden representations. CRAFT proceeds in three stages: it first routes each task to a group of similar tasks based on output-distribution divergence; it then fine-tunes the model using a Kullback-Leibler (KL) divergence against the group's prior state, which directly controls forgetting and determines convergence; finally, it merges interventions for the updated task into the shared representation using the same KL signal. This design unifies routing, regularization, and merging through a single KL-based objective. CRAFT improves overall performance and reduces forgetting compared to strong LoRA-based approaches across multiple benchmarks and model scales, while remaining robust to task ordering. These results suggest that controlling adaptation in representation space, guided by output-space divergence, provides a scalable and principled approach to continual learning in LLMs.\n\n---\n\n## 128. SafeHarbor: Hierarchical Memory-Augmented Guardrail for LLM Agent Safety\n- **Authors**: Zhe Liu , Zonghao Ying , Wenxin Zhang , Quanchen Zou , Deyue Zhang , Dongdong Yang , Xiangzheng Zhang , Hao Peng\n- **URL**: [https://arxiv.org/abs/2605.05704](https://arxiv.org/abs/2605.05704)\n- **Abstract**:\n> With the rapid evolution of foundation models, Large Language Model (LLM) agents have demonstrated increasingly powerful tool-use capabilities. However, this proficiency introduces significant security risks, as malicious actors can manipulate agents into executing tools to generate harmful content. While existing defensive mechanisms are effective, they frequently suffer from the over-refusal problem, where increased safety strictness compromises the agent's utility on benign tasks. To mitigate this trade-off, we propose \\textsc{SafeHarbor}, a novel framework designed to establish precise decision boundaries for LLM agents. Unlike static guidelines, \\textsc{SafeHarbor} extracts context-aware defense rules through enhanced adversarial generation. We design a local hierarchical memory system for dynamic rule injection, offering a training-free, efficient, and plug-and-play solution. Furthermore, we introduce an information entropy-based self-evolution mechanism that continuously optimizes the memory structure through dynamic node splitting and merging. Extensive experiments demonstrate that \\textsc{SafeHarbor} achieves state-of-the-art performance on both ambiguous benign tasks and explicit malicious attacks, notably attaining a peak benign utility of 63.6\\% on GPT-4o while maintaining a robust refusal rate exceeding 93\\% against harmful requests. The source code is publicly available at this https URL .\n\n---\n\n## 129. Active Learning for Communication Structure Optimization in LLM-Based Multi-Agent Systems\n- **Authors**: Huchen Yang , Xinghao Dong , Dan Negrut , Jin-Long Wu\n- **URL**: [https://arxiv.org/abs/2605.05703](https://arxiv.org/abs/2605.05703)\n- **Abstract**:\n> Optimizing the communication structure of large language model based multi-agent systems (LLM-MAS) has been shown to improve downstream performance and reduce token usage. Existing methods typically rely on randomly sampled training tasks. However, tasks may differ substantially in difficulty and domain, and thus they are not equally informative for updating communication structure, making optimization under limited training budgets often unstable and highly sensitive to the particular training set. To actively identify the most valuable tasks for communication-structure optimization, we propose an ensemble-based information-theoretic task selection framework. The proposed method estimates task informativeness by how much a candidate task changes the distribution over graph parameters, using ensemble Kalman inversion as an efficient and derivative-free approximation of the corresponding Bayesian update. The resulting estimator is especially suitable for black-box and noisy multi-agent systems. To enhance scalability, we construct a compact candidate pool through embedding-based representative selection and combine the informative selection with surrogate modeling and batch Thompson sampling. We validate our method in both benign settings and settings with agent attacks, demonstrating its effectiveness for communication-structure optimization under constrained computational budgets.\n\n---\n\n## 130. An Empirical Study of Proactive Coding Assistants in Real-World Software Development\n- **Authors**: Lehui Li , Ruixuan Jia , Guo-Ye Yang , Jia Li\n- **URL**: [https://arxiv.org/abs/2605.05700](https://arxiv.org/abs/2605.05700)\n- **Abstract**:\n> Large language model (LLM)-based coding assistants have made substantial progress, yet most systems remain reactive, requiring developers to explicitly formulate their needs. Proactive coding assistants aim to infer latent developer intent from integrated development environment (IDE) interactions and repository context, thereby reducing interaction overhead and supporting more seamless assistance. However, research in this direction is limited by the scarcity of large-scale real-world developer behavior data. Existing studies therefore often rely on LLM-simulated IDE traces, whose fidelity to real development behavior remains unclear. In this paper, we investigate this simulation-to-reality gap through a large-scale empirical study. We collect real IDE interaction traces from 1{,}246 experienced industry developers over three consecutive days using a custom Visual Studio Code extension, and construct paired LLM-simulated traces for controlled comparison. Our analysis shows that simulated traces differ substantially from real traces in behavioral diversity, temporal structure, and exploratory patterns. Based on the collected data, we introduce \\textbf{ProCodeBench}, a real-world benchmark for proactive intent prediction. Experiments with representative LLMs, retrieval-augmented methods, and agentic baselines show that current approaches remain far from reliable under real IDE traces, suggesting that simulation-based evaluation can overestimate real-world performance. Finally, our training study shows that simulated data cannot replace real data, but can complement it when used before real-world fine-tuning. These findings highlight the importance of real developer behavior data for evaluating and training proactive coding assistants.\n\n---\n\n## 131. Irminsul: MLA-Native Position-Independent Caching for Agentic LLM Serving\n- **Authors**: Bole Ma , Jan Eitzinger , Harald Köstler\n- **URL**: [https://arxiv.org/abs/2605.05696](https://arxiv.org/abs/2605.05696)\n- **Abstract**:\n> Agentic LLM workloads put bit-identical tokens at shifted positions every turn, voiding prefix caches at the first byte of divergence. Operators report cache-hit regressions ranging from moderate slowdowns to severe TTFT spikes of 10-16s on unchanged content. Prior position-independent caching systems correct RoPE on the full $d_K$-dimensional key, an architectural cost imposed by GQA, not by caching itself. Multi-Head Latent Attention, deployed at scale in DeepSeek-V2/V3/R1, Kimi-K2/Moonlight, GLM-5, and Mistral Large 3, factors each KV row into a position-free $c_{KV}$ and a 64-dim $k_r$ correctable in closed form; this structure motivates content-addressed caching as a natural fit rather than a GQA workaround. We present Irminsul, which extends SGLang's radix cache with content-hash keying over CDC-chunked segments and a $\\delta$-rotation rule for $k_r$. We evaluate three native MLA-MoE deployments - DeepSeek-V2-Lite (16B/2.4B), Kimi Moonlight-16B-A3B, and JoyAI-Flash (48B/3B) - with output-consistency on all three and recovery measured on the two endpoints; Irminsul recovers up to ~83% of prompt tokens above exact-prefix on agentic traffic while delivering 63% prefill energy savings per cache hit. We argue that content-addressed caching belongs in the serving stack as a first-class primitive, not a retrofit over prefix matching.\n\n---\n\n## 132. Decomposing the Basic Abilities of Large Language Models: Mitigating Cross-Task Interference in Multi-Task Instruct-Tuning\n- **Authors**: Bing Wang , Ximing Li , Changchun Li , Jinjin Chi , Gang Niu , Masashi Sugiyama\n- **URL**: [https://arxiv.org/abs/2605.05676](https://arxiv.org/abs/2605.05676)\n- **Abstract**:\n> Recently, the prominent performance of large language models (LLMs) has been largely driven by multi-task instruct-tuning. Unfortunately, this training paradigm suffers from a key issue, named cross-task interference, due to conflicting gradients over shared parameters among different tasks. Some previous methods mitigate this issue by isolating task-specific parameters, e.g., task-specific neuron selection and mixture-of-experts. In this paper, we empirically reveal that the cross-task interference still exists for the existing solutions because of many parameters also shared by different tasks, and accordingly, we propose a novel solution, namely Basic Abilities Decomposition for multi-task Instruct-Tuning (BADIT). Specifically, we empirically find that certain parameters are consistently co-activated, and that co-activated parameters naturally organize into base groups. This motivates us to analogize that LLMs encode several orthogonal basic abilities, and that any task can be represented as a linear combination of these abilities. Accordingly, we propose BADIT that decomposes LLM parameters into orthogonal high-singular-value LoRA experts representing basic abilities, and dynamically enforces their orthogonality during training via spherical clustering of rank-1 components. We conduct extensive experiments on the SuperNI benchmark with 6 LLMs, and empirical results demonstrate that BADIT can outperform SOTA methods and mitigate the degree of cross-task interference.\n\n---\n\n## 133. XL-SafetyBench: A Country-Grounded Cross-Cultural Benchmark for LLM Safety and Cultural Sensitivity\n- **Authors**: Dasol Choi , Eugenia Kim , Jaewon Noh , Sang Seo , Eunmi Kim , Myunggyo Oh , Yunjin Park , Brigitta Jesica Kartono , Josef Pichlmeier , Helena Berndt , Sai Krishna Mendu , Glenn Johannes Tungka , Özlem Gökçe , Suresh Gehlot , Katherine Pratt , Amanda Minnich , Haon Park\n- **URL**: [https://arxiv.org/abs/2605.05662](https://arxiv.org/abs/2605.05662)\n- **Abstract**:\n> Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Moreover, they rarely evaluate a model's ability to detect culturally embedded sensitivities as distinct from universal harms. We introduce XL-SafetyBench. a suite of 5,500 test cases across 10 country-language pairs, comprising a Jailbreak Benchmark of country-grounded adversarial prompts and a Cultural Benchmark where local sensitivities are embedded within innocuous requests. Each item is constructed via a multi-stage pipeline that combines LLM-assisted discovery, automated validation gates, and dual independent native-speaker annotators per country. To distinguish principled refusal from comprehension failure, we evaluate Attack Success Rate (ASR) alongside two complementary metrics we introduce: Neutral-Safe Rate (NSR) and Cultural Sensitivity Rate (CSR). Evaluating 10 frontier and 27 local LLMs reveals two key findings. First, jailbreak robustness and cultural awareness do not show a coupled relationship among frontier models, so a composite safety score obscures per-axis variation. Second, local models exhibit a near-linear ASR-NSR trade-off (r = -0.81), indicating that their apparent safety reflects generation failure rather than genuine alignment. XL-SafetyBench enables more nuanced, cross-cultural safety evaluation in the multilingual era.\n\n---\n\n## 134. One Turn Too Late: Response-Aware Defense Against Hidden Malicious Intent in Multi-Turn Dialogue\n- **Authors**: Xinjie Shen , Rongzhe Wei , Peizhi Niu , Haoyu Wang , Ruihan Wu , Eli Chien , Bo Li , Pin-Yu Chen , Pan Li\n- **URL**: [https://arxiv.org/abs/2605.05630](https://arxiv.org/abs/2605.05630)\n- **Abstract**:\n> Hidden malicious intent in multi-turn dialogue poses a growing threat to deployed large language models (LLMs). Rather than exposing a harmful objective in a single prompt, increasingly capable attackers can distribute their intent across multiple benign-looking turns. Recent studies show that even modern commercial models with advanced guardrails remain vulnerable to such attacks despite advances in safety alignment and external guardrails. In this work, we address this challenge by detecting the earliest turn at which delivering the candidate response would make the accumulated interaction sufficient to enable harmful action. This objective requires precise turn-level intervention that identifies the harm-enabling closure point while avoiding premature refusal of benign exploratory conversations. To further support training and evaluation, we construct the Multi-Turn Intent Dataset (MTID), which contains branching attack rollouts, matched benign hard negatives, and annotations of the earliest harm-enabling turns. We show that MTID helps enable a turn-level monitor TurnGate, which substantially outperforms existing baselines in harmful-intent detection while maintaining low over-refusal rates. TurnGate further generalizes across domains, attacker pipelines, and target models. Our code is available at this https URL .\n\n---\n\n## 135. Leveraging Image Generators to Address Training Data Scarcity: The Gen4Regen Dataset for Forest Regeneration Mapping\n- **Authors**: Gabriel Jeanson , David-Alexandre Duclos , William Larrivée-Hardy , Noé Cochet , Matěj Boxan , Anthony Deschênes , François Pomerleau , Philippe Giguère\n- **URL**: [https://arxiv.org/abs/2605.05627](https://arxiv.org/abs/2605.05627)\n- **Abstract**:\n> Sustainable forest management relies on precise species composition mapping, yet traditional ground surveys are labour-intensive and geographically constrained. While Uncrewed Aerial Vehicles (UAVs) offer scalable data collection, the transition to deep learning-based interpretation is bottlenecked by the severe scarcity of expert-annotated imagery, particularly in complex, visually heterogeneous regeneration zones. This paper addresses the dual challenges of data scarcity and extreme class imbalance in the semantic segmentation of fine-grained forest regeneration species by providing a scalable framework that reduces reliance on manual photo-interpretation for high-resolution, millimetre-level aerial imagery. Importantly, we leverage the large-scale vision-language Nano Banana Pro model to simultaneously generate high-fidelity images and their corresponding pixel-aligned semantic masks from prompts. We introduce WilDReF-Q-V2, an expansion of a natural forest dataset with 13 977 new unlabelled and 50 labelled real images, as well as the Gen4Regen dataset, featuring 2101 pairs of synthetic images and semantic masks. Our methodology integrates real-world data with AI-generated images, highlighting that AI-generated data is highly complementary to real-world data, with unified training yielding an F1 score improvement of over 15 %pt compared to purely supervised baselines. Furthermore, we demonstrate that even small quantities of prompt-generated data significantly improve performance for underrepresented species, some of which saw per-species F1 score gains of up to 30 %pt. We conclude that vision-language models can serve as agile data generators, effectively bootstrapping perception tasks for niche AI domains where expert labels are scarce or unavailable. Our datasets, source code, and models will be available at this https URL .\n\n---\n\n## 136. When2Speak: A Dataset for Temporal Participation and Turn-Taking in Multi-Party Conversations for Large Language Models\n- **Authors**: Vihaan Nama , Shreya Mendi , Zian Ye , Brinnae Bent\n- **URL**: [https://arxiv.org/abs/2605.05626](https://arxiv.org/abs/2605.05626)\n- **Abstract**:\n> Large Language Models (LLMs) excel at generating contextually appropriate responses but remain poorly calibrated for multi-party conversations, where deciding when to speak is as critical as what to say. In such settings, naively responding at every turn leads to excessive interruptions and degraded conversational coherence. We introduce When2Speak, a grounded synthetic dataset and four-stage generation pipeline for learning intervention timing in group interactions. The dataset comprises over 215,000 examples derived from 16,000 conversations involving 2-6 speakers, spanning diverse conversational styles, tones, and participant dynamics, and explicitly modeling SPEAK vs. SILENT decisions at each turn. Our pipeline combines real-world grounding, structured augmentation, controlled transcript synthesis, and fine-tuning-ready supervision, and is fully open-sourced to support reproducibility and adaptation to domain-specific conversational norms. Across multiple model families, supervised fine-tuning (SFT) on When2Speak significantly outperforms zero-shot baselines (e.g., the average Macro F1 increase across 4B+ parameter models was 60%, with the largest increase being 120%). However, SFT-trained models remain systematically over-conservative, missing nearly half of warranted interventions as seen through the Missed Intervention Rate (MIR), which was on average 0.50 and is noticed even at larger model sizes. To address this limitation, we apply reinforcement learning with asymmetric reward shaping, which reduces MIR to 0.186-0.218 and increases recall from 0.479 to 0.78-0.81. Our findings establish that temporal participation is a distinct and trainable dimension of conversational intelligence, and that grounded synthetic data provides an effective and scalable pathway for enabling LLMs to participate more naturally and appropriately in multi-party interactions.\n\n---\n\n## 137. AstroAlertBench: Evaluating the Accuracy, Reasoning, and Honesty of Multimodal LLMs in Astronomical Classification\n- **Authors**: Claire Chen , Jiabao Sean Xiao , Shuze Daniel Liu , Facundo Perez Paolino , Luke Handley , Theophile Jegou du Laz , Ricky Nilsson , Alice Zou , Matthew Graham , Ashish Mahabal\n- **URL**: [https://arxiv.org/abs/2605.05573](https://arxiv.org/abs/2605.05573)\n- **Abstract**:\n> Modern astronomical observatories generate a massive volume of multimodal data, creating a critical bottleneck for expert human review. While multimodal large language models (LLMs) have shown promise in interpreting complex visual and textual inputs, their ability to perform specialized scientific classification while providing interpretable reasoning remains understudied. We introduce AstroAlertBench, a comprehensive multimodal benchmark designed to evaluate LLM performance in astronomical event review along a three-stage logical chain: metadata grounding, scientific reasoning, and hierarchical classification over five categories. We use a pilot sample of 1,500 real-world alerts from the Zwicky Transient Facility (ZTF), a wide-field survey that scans the northern sky to detect transient astronomical events. On this dataset, we benchmark 13 frontier closed-source and open-weight LLMs that support visual input. Our results reveal that high accuracy does not always align with model ``honesty,'' defined as the ability to self-evaluate its reasoning, which affects its reliability as a real-world assistant. We further initialize a human-in-the-loop evaluation protocol as a precursor to future community-scale participation. Together, AstroAlertBench provides a framework for developing calibrated and interpretable astronomical assistants.\n\n---\n\n## 138. ReaComp: Compiling LLM Reasoning into Symbolic Solvers for Efficient Program Synthesis\n- **Authors**: Atharva Naik , Yash Mathur , Prakam , Carolyn Rose , David Mortensen\n- **URL**: [https://arxiv.org/abs/2605.05485](https://arxiv.org/abs/2605.05485)\n- **Abstract**:\n> LLMs can solve program synthesis tasks but remain inefficient and unreliable on hard instances requiring large combinatorial search. Given a small set of reasoning traces, we use coding agents to compile them into reusable symbolic program synthesizers over constrained DSLs. The resulting solvers require no LLM calls at test time and are strong standalone systems: symbolic solver ensembles reach 91.3% accuracy on PBEBench-Lite and 84.7% on PBEBench-Hard, outperforming LLMs with test-time scaling for the latter by +16.3 percentage points at zero LLM inference cost. They also complement LLM search, improving PBEBench-Hard accuracy from 68.4% to 85.8% while reducing reported token usage by 78%, and raising SLR-Bench hard-tier accuracy from 34.4% to 58.0% in a neuro-symbolic hybrid setting. Compared to directly using coding agents as per-instance solvers, induced solvers are substantially more Pareto-efficient, amortizing a small one-time construction cost over many zero-token executions. Finally, most solvers transfer zero-shot to a real historical linguistics task - predicting sound changes in natural language data - reaching 80.1% accuracy under ensembling and recovering some plausible linguistic rules. Together, these results show that reasoning traces can be compiled into reusable symbolic solvers that solve many tasks directly, complement LLM inference on hard cases, and provide a scalable route to domain-general solver induction. We release code and data for reproducibility.\n\n---\n\n## 139. Robustness of Graph Self-Supervised Learning to Real-World Noise: A Case Study on Text-Driven Biomedical Graphs\n- **Authors**: Othmane Kabal , Mounira Harzallah , Fabrice Guillet , Hideaki Takeda , Ryutaro Ichise\n- **URL**: [https://arxiv.org/abs/2605.05463](https://arxiv.org/abs/2605.05463)\n- **Abstract**:\n> Graph Self-Supervised Learning (GSSL) offers a powerful paradigm for learning graph representations without labeled data. However, existing work assumes clean, manually curated graphs. Recent advances in NLP enable the large-scale automatic extraction of knowledge graphs from text, opening new opportunities for GSSL while introducing substantial real-world noise. This type of noise remains largely unexplored, as prior robustness studies typically rely on synthetic perturbations. To address this gap, we present the first comprehensive evaluation of GSSL methods on text-driven graphs for unsupervised term typing. We introduce Noise-Aware Text-Driven Graph GSSL (NATD-GSSL), a unified framework that combines automatic graph construction, graph refinement, and GSSL. Our evaluation follows a dual-graph protocol that contrasts a noisy graph derived from MedMentions with a clean Unified Medical Language System (UMLS) reference graph, aligned through a shared gold standard. Our results reveal variability in robustness across both pretext tasks and Graph Neural Network (GNN) architectures. Relation reconstruction is highly sensitive to noise and benefits from well-defined schemas, whereas feature reconstruction is considerably more robust, achieving performance comparable to clean-graph settings. Contrastive objectives are generally less affected by noise but depend strongly on alignment with downstream tasks. GNN architecture also plays a critical role: bidirectional relational message-passing designs are better suited to noisy, text-driven graphs, while unidirectional relational ones perform best on clean graphs. Overall, NATD-GSSL provides practical guidance for applying GSSL to real-world, noisy graphs and achieves up to a 7\\% improvement over pretrained language model baselines. All code and benchmarks are publicly available at this https URL .\n\n---\n\n## 140. SLAM: Structural Linguistic Activation Marking for Language Models\n- **Authors**: Fabrice Harel-Canada , Amit Sahai\n- **URL**: [https://arxiv.org/abs/2605.05443](https://arxiv.org/abs/2605.05443)\n- **Abstract**:\n> LLM watermarks must be detectable without compromising text quality, yet most existing schemes bias the next-token distribution and pay for detection with measurable quality loss. We present SLAM (Structural Linguistic Activation Marking), a novel white-box watermarking scheme that sidesteps this cost by writing the mark into structural geometry rather than token frequencies: sparse autoencoders identify residual-stream directions encoding linguistic structure (e.g., voice, tense, clause order), and we causally steer those directions at generation time, leaving lexical sampling and semantics unconstrained. On Gemma-2 2B and 9B, SLAM achieves 100% detection accuracy with a quality cost of only 1-2 reward points - compared to 7.5-11.5 for KGW, EWD, and Unigram - with naturalness and diversity preserved at near-unwatermarked levels across both models. The trade-off is a complementary robustness profile: SLAM resists word-level edits but is vulnerable to paraphrase that restructures syntax (at a quality cost), the converse of token-distribution methods.\n\n---\n\n## 141. Information Theoretic Adversarial Training of Large Language Models\n- **Authors**: Yiwei Zhang , Jeremiah Birrell , Reza Ebrahimi , Rouzbeh Behnia , Jason Pacheco , Elisa Bertino\n- **URL**: [https://arxiv.org/abs/2605.05415](https://arxiv.org/abs/2605.05415)\n- **Abstract**:\n> Large language models (LLMs) remain vulnerable to adversarial prompting despite advances in alignment and safety, often exhibiting harmful behaviors under novel attack strategies. While adversarial training can improve robustness, existing approaches are computationally expensive and difficult to scale. Recent continuous adversarial training methods, such as Continuous adversarial training (CAT) and Continuous Adversarial Preference Optimization (CAPO), address this challenge by leveraging gradient-based perturbations in the embedding space, enabling more efficient and expressive attacks. Building on this paradigm, we propose WARDEN, a distributionally robust adversarial training framework for LLMs that dynamically reweights adversarial examples through an f -divergence ambiguity set around the empirical training distribution. Our method optimizes the worst-case adversarial loss within a divergence ball around the empirical data distribution, automatically emphasizing harder adversarial examples. Using the convex dual formulation, the objective reduces to a log-sum-exp form under the KL divergence, with a dynamical parameter controlling the strength of reweighting. This study leads to a new class of information-theoretic objectives that significantly reduce attack success rates while maintaining model utility. Across multiple LLMs and attack settings, WARDEN substantially reduces attack success rates with computational and utility costs comparable to CAT-, CAPO-, and MixAT-based baselines, making it a practical approach for scalable robust alignment.\n\n---\n\n## 142. Open-SAT: LLM-Guided Query Embedding Refinement for Open-Vocabulary Object Retrieval in Satellite Imagery\n- **Authors**: Md Adnan Arefeen , Biplob Debnath , Ravi K. Rajendran , Murugan Sankaradas , Srimat T. Chakradhar\n- **URL**: [https://arxiv.org/abs/2605.05344](https://arxiv.org/abs/2605.05344)\n- **Abstract**:\n> In satellite applications, user queries often take the form of open-ended natural language, extending beyond a fixed set of predefined categories. This open-vocabulary nature poses significant challenges for retrieving relevant image tiles, as the retrieval system must generalize to a wide range of unseen objects and concepts. While vision-language models (VLMs) such as CLIP are widely used for text-image retrieval, even fine-tuned variants often struggle to accurately align such queries with satellite imagery. To address this, we propose Open-SAT, a training-free query embedding refinement algorithm that operates at inference time to improve alignment between user queries and satellite image content. Open-SAT uses VLMs to compute embeddings for image tiles, which are stored in a vector database for efficient retrieval. At query time, it leverages Large Language Models (LLMs) to refine the text embeddings by incorporating contextual information about objects of interest and their surroundings. A threshold-free retrieval mechanism further enhances accuracy and efficiency. Experimental results in three public benchmarks demonstrate that Open-SAT improves the F1 score by up to 16.04%, while retrieving a comparable number of image tiles. These results demonstrate the effectiveness of Open-SAT in open-vocabulary satellite image retrieval, leveraging LLM guidance without the need for additional training or supervision.\n\n---\n\n## 143. Feature Starvation as Geometric Instability in Sparse Autoencoders\n- **Authors**: Faris Chaudhry , Keisuke Yano , Anthea Monod\n- **URL**: [https://arxiv.org/abs/2605.05341](https://arxiv.org/abs/2605.05341)\n- **Abstract**:\n> Sparse autoencoders (SAEs) are used to disentangle the dense, polysemantic internal representations of large language models (LLMs) into interpretable, monosemantic concepts. However, standard $\\ell_1$-regularized SAEs suffer from feature starvation (dead neurons) and shrinkage bias, often requiring computationally expensive heuristic resampling and nondifferentiable hard-masking methods to bypass these challenges. We argue that feature starvation is not merely an empirical artifact of poor data diversity, but a fundamental optimization-geometric pathology of overcomplete dictionaries: the $\\ell_1$-induced sparse coding map is unstable and fundamentally misaligned with shallow, amortized encoders. To address this structural instability, we introduce adaptive elastic net SAEs (AEN-SAEs), a fully differentiable architecture grounded in classical sparse regression. AEN-SAEs combine an $\\ell_2$ structural term that enforces strong convexity and Lipschitz stability with adaptive $\\ell_1$ reweighting that eliminates shrinkage bias and suppresses spurious features, thereby jointly controlling the curvature and interaction structure of the induced polyhedral geometry. Theoretically, we show that AEN-SAEs yield a Lipschitz-continuous sparse coding map and recover the global feature support under mild assumptions. Empirically, across synthetic settings and LLMs (Pythia 70M, Llama 3.1 8B), AEN-SAEs mitigate feature starvation without auxiliary heuristics while maintaining competitive reconstruction abilities.\n\n---\n\n## 144. How Far Are VLMs from Privacy Awareness in the Physical World? An Empirical Study\n- **Authors**: Junran Wang , Xinjie Shen , Zehao Jin , Pan Li\n- **URL**: [https://arxiv.org/abs/2605.05340](https://arxiv.org/abs/2605.05340)\n- **Abstract**:\n> As Vision-Language Models (VLMs) are increasingly deployed as autonomous cognitive cores for embodied assistants, evaluating their privacy awareness in physical environments becomes critical. Unlike digital chatbots, these agents operate in intimate spaces, such as homes and hospitals, where they possess the physical agency to observe and manipulate privacy-sensitive information and artifacts. However, current benchmarks remain limited to unimodal, text-based representations that cannot capture the demands of real-world settings. To bridge this gap, we present ImmersedPrivacy, an interactive audio-visual evaluation framework that simulates realistic physical environments using a Unity-based simulator. ImmersedPrivacy evaluates physically grounded privacy awareness across three progressive tiers that test a model's ability to identify sensitive items in cluttered scenes, adapt to shifting social contexts, and resolve conflicts between explicit commands and inferred privacy constraints. Our evaluation of 12 state-of-the-art models reveals consistent deficits. In cluttered scenes, all models exhibit monotonic performance decay as scene complexity grows due to perceptual deficit. When social context shifts, no model exceed 65% selection accuracy. Under conflicting commands, the best model gemini-3.1-pro perfectly balances task completion and privacy preservation in only 51% of cases. These findings reveal that current VLMs in the physical world suffer from perceptual fragility and fail to let their knowledge of privacy cues govern their situated behavior. Our code and data is available at this https URL .\n\n---\n\n## 145. Bridging Generation and Training: A Systematic Review of Quality Issues in LLMs for Code\n- **Authors**: Kaifeng He , Xiaojun Zhang , Peiliang Cai , Mingwei Liu , Yanlin Wang , Chong Wang , Kaifeng Huang , Bihuan Chen , Xin Peng , Zibin Zheng\n- **URL**: [https://arxiv.org/abs/2605.05267](https://arxiv.org/abs/2605.05267)\n- **Abstract**:\n> Large language models (LLMs) frequently generate defective outputs in code generation tasks, ranging from logical bugs to security vulnerabilities. While these generation failures are often treated as model-level limitations, empirical evidence increasingly traces their root causes to imperfections within the training corpora. Yet, the specific mechanisms linking training data quality issues to generated code quality issues remain largely unmapped. This paper presents a systematic literature review of 114 primary studies to investigate how training data quality issues propagate into code generation. We establish a unified taxonomy that categorizes generated code quality issues across nine dimensions and training data quality issues into code and non-code attributes. Based on this taxonomy, we formalize a causal framework detailing 18 typical propagation mapping mechanisms. Furthermore, we synthesize state-of-the-art detection and mitigation techniques across the data, model, and generation lifecycles. The reviewed literature reveals a clear methodological shift: quality assurance is transitioning from reactive, heuristic-based post-generation filtering toward proactive, data-centric governance and closed-loop repair. Finally, we identify open challenges and outline research directions for developing reliable LLMs for code through integrated data curation and continuous evaluation. Our repository is available at this https URL .\n\n---\n\n## 146. Decision-aware User Simulation Agent for Evaluating Conversational Recommender Systems\n- **Authors**: Yuan-Chi Li , Li-Chi Chen , Sung-Yi Wu , Yu-Che Tsai , Shou-De Lin\n- **URL**: [https://arxiv.org/abs/2605.05250](https://arxiv.org/abs/2605.05250)\n- **Abstract**:\n> Conversational recommender systems (CRS) increasingly rely on user simulators for automated evaluation of sales agents. A key requirement for such simulators is the ability to model human decision-making. However, most existing simulation frameworks do not explicitly model the internal decision process, and LLM-based simulators often exhibit unrealistically strong information-processing capabilities, rarely exhibit the hesitation or decision deferral commonly observed in real consumer behavior, resulting in overly high acceptance probabilities. To address this limitation, we propose Hesitator, a theory-grounded user simulation framework that explicitly models human decision-making under choice overload. The framework introduces a modular Decision Module that separates utility-based item selection from overload-aware commitment decisions. Experiments across multiple user simulation frameworks, domains, sales modes, and LLM backbones show that integrating our module consistently mitigates unrealistic behaviors under increasing overload conditions. Furthermore, Hesitator reproduces established behavioral patterns from psychological economics, demonstrating its ability to model human decision behavior.\n\n---\n\n## 147. Towards Dependable Retrieval-Augmented Generation Using Factual Confidence Prediction\n- **Authors**: Florian Geissler , Francesco Carella , Laura Fieback , Jakob Spiegelberg\n- **URL**: [https://arxiv.org/abs/2605.05244](https://arxiv.org/abs/2605.05244)\n- **Abstract**:\n> Incorporating specific knowledge into large language models via retrieval-augmented generation (RAG) is a widespread technique that fuels many of today's industry AI applications. A fundamental problem is to assess if the context retrieved by some similarity search provides indeed supporting facts, or instead misguides the generator with irrelevant information. It is critical to associate meaningful confidence measures about the factuality of the retrieval process with the generated answers. We present a new, two-staged approach to predict fact faithfulness of the output of retrieval-augmented generations. First, we employ conformal prediction to select only those retrieved chunks who have a high chance to come from the correct source. This approach in itself can improve answer quality by up to 6% in some of the studied datasets, however, the associated statistical guarantees do not hold generally, since the assumption of sample exchangeability depends on the retriever setup. We present diagnostic metrics to assess whether a setup is suitable. Second, we quantify confidence in the consistency of a generated final answer with a given retrieved context, using an attention-based factuality classifier. This approach can detect inconsistent answers with a chance of up to 77%. Our work helps to establish a novel type of certified RAG systems for a broad range of natural language industry applications.\n\n---\n\n## 148. Rethinking Data Curation in LLM Training: Online Reweighting Offers Better Generalization than Offline Methods\n- **Authors**: Wanru Zhao , Yihong Chen , Yuzhi Tang , Wentao Ma , Shengchao Hu , Shell Xu Hu , Alex Iacob , Abhinav Mehrotra , Nicholas D. Lane\n- **URL**: [https://arxiv.org/abs/2605.05227](https://arxiv.org/abs/2605.05227)\n- **Abstract**:\n> Data curation is a critical yet under-explored area in large language model (LLM) training. Existing methods, such as data selection and mixing, operate in an offline paradigm, detaching themselves from training. This separation introduces engineering overhead and makes the curation brittle: the entire pipeline must be re-run under model/task shifts. Moreover, offline methods alter data size through hard filtering or resampling, often sacrificing data diversity and harming generalization. We propose to rethink data curation as an online reweighting problem, where sample importance is dynamically adjusted during training via loss weighting rather than static pre-processing. Specifically, we introduce ADAPT (Adaptive Data reweighting for Pretraining and FineTuning), a dynamic online framework that reweights training samples with adaptive per-sample learning rates guided by similarity-based quality signals, without changing the number of training samples. Unlike offline methods that enforce a static data distribution, ADAPT acts as an implicit curriculum learner, progressively shifting focus from coarse-grained patterns to fine-grained semantic distinctions as the model evolves. Experiments on both instruction tuning and large-scale pretraining show that ADAPT consistently outperforms offline selection/mixing and prior online methods, achieving stronger cross-benchmark generalization under equal FLOPs.\n\n---\n\n## 149. MACS: Modality-Aware Capacity Scaling for Efficient Multimodal MoE Inference\n- **Authors**: Bo Li , Chuan Wu , shaolin Zhu\n- **URL**: [https://arxiv.org/abs/2605.05225](https://arxiv.org/abs/2605.05225)\n- **Abstract**:\n> Mixture-of-Experts Multimodal Large Language Models (MoE MLLMs) suffer from a significant efficiency bottleneck during Expert Parallelism (EP) inference due to the straggler effect. This issue is worsened in the multimodal context, as existing token-count-based load balancing methods fail to address two unique challenges: (1) Information Heterogeneity, where numerous redundant visual tokens are treated equally to semantically critical ones, and (2) Modality Dynamics, where varying visual to text ratios across tasks lead to resource misallocation. To address these challenges, we propose MACS (Modality-Aware Capacity Scaling), a training-free inference framework. Specifically, MACS introduces an Entropy-Weighted Load mechanism to quantify the semantic value of visual tokens, addressing information heterogeneity. Additionally, the Dynamic Modality-Adaptive Capacity mechanism allocates expert resources based on the real-time modal composition of the input. Extensive experiments demonstrate that MACS significantly outperforms existing methods on various multimodal benchmarks, providing a novel and robust solution for the efficient deployment of MoE MLLMs in EP inference.\n\n---\n\n## 150. MidSteer: Optimal Affine Framework for Steering Generative Models\n- **Authors**: Tatiana Gaintseva , Andrew Stepanov , Ziquan Liu , Martin Benning , Gregory Slabaugh , Jiankang Deng , Ismail Elezi\n- **URL**: [https://arxiv.org/abs/2605.05220](https://arxiv.org/abs/2605.05220)\n- **Abstract**:\n> Steering intermediate representations has emerged as a powerful strategy for controlling generative models, particularly in post-deployment alignment and safety settings. However, despite its empirical success, it currently lacks a comprehensive theoretical framework. In this paper, we bridge this gap by formalizing the theory of concept steering. First, we establish a link between steering and affine concept erasure, proving that the standard approach for removing unwanted behaviors is a special case of LEACE (a closed-form method for affine erasure). Next, we formulate a principled theoretical framework for concept switching, LEACE-Switch, and characterize the assumptions under which it provides an optimal affine solution. Building on this analysis, we then introduce MidSteer (Minimal Disturbance concept Steering), a more general affine framework for concept manipulation that relaxes these assumptions and enables directed, minimal-disturbance transformations. We demonstrate that MidSteer performs favorably across a range of tasks, modalities, and architectures, including vision diffusion models and large language models.\n\n---\n\n## 151. Sparse Prefix Caching for Hybrid and Recurrent LLM Serving\n- **Authors**: Mikhail Shirokikh , Sergey Nikolenko\n- **URL**: [https://arxiv.org/abs/2605.05219](https://arxiv.org/abs/2605.05219)\n- **Abstract**:\n> Prefix caching is a key latency optimization for autoregressive LLM serving, yet existing systems assume dense per-token key/value reuse. State-space models change the structure of the problem: a recurrent layer can resume from a single stored state rather than requiring the entire token history. This asymmetry opens a new design point between no reuse and dense caching: store exact recurrent states at a sparse set of checkpoint positions and, on a cache hit, resume from the deepest stored checkpoint and recompute the remaining suffix exactly. We formalize sparse prefix caching as checkpoint placement under a distribution over overlap depths, yielding an exact O(NM) dynamic program. For use cases where requests share a non-trivial prefix (e.g. asking different questions about a single long document), we show that our method consistently improves the Pareto frontier traced by standard heuristics on real-world data. Across QuALITY and System Prompts, distribution-aware placement dominates every fixed-budget baseline on the measured layer-group Pareto frontier and matches or outperforms the strongest heuristic (block caching) while typically using substantially fewer checkpoints, with the largest gains at low checkpoint budgets where the overlap distribution is most non-uniform. The method is most relevant when many requests share a substantial but not identical prefix within a retained cache entry. It preserves exact outputs, does not change the recurrent computation itself or require new recurrent update kernels, applies to recurrent/SSM layers whose hidden state can be extracted and restored exactly, and for hybrid models can be combined with existing KV-cache compression techniques.\n\n---\n\n## 152. A Review of Large Language Models for Stock Price Forecasting from a Hedge-Fund Perspective\n- **Authors**: Olivia Zhang , Zhilin Zhang\n- **URL**: [https://arxiv.org/abs/2605.05211](https://arxiv.org/abs/2605.05211)\n- **Abstract**:\n> Large language models (LLMs) are increasingly deployed in quantitative finance for stock price forecasting. This review synthesizes recent applications of LLMs in this domain, including extracting sentiment from financial news and social media, analyzing financial reports and earnings-call transcripts, tokenizing or symbolizing stock price series, and constructing multi-agent trading systems. Particular attention is paid to practical pitfalls that are often understated in the literature, such as fragility in sentiment analysis, dataset and horizon design, performance evaluation metrics, data leakage, illiquidity premia, and limits of stock price predictability. Organized from a hedge-fund perspective, the review is intended to guide both academic researchers and hedge fund managers in integrating LLMs into real-world trading pipelines and in stress-testing their robustness under realistic market frictions.\n\n---\n\n{% endraw %}\n"}, "ae95e160c1986855bbe27edfadb3713b4c8f4a25": {"t": 1786068561.3722315, "n": 263185, "h": "c0b4ce7dc3328d6b", "c": "---\nlayout: default\ntitle: \"LLM 관련 주요 논문 - 2026-05-08\"\n---\n\n{% raw %}\n# LLM 관련 주요 논문 - 2026-05-08\n\n## 1. GlazyBench: A Benchmark for Ceramic Glaze Property Prediction and Image Generation\n- **Authors**: Ziyu Zhai , Siyou Li , Juexi Shao , Juntao Yu\n- **URL**: [https://arxiv.org/abs/2605.06641](https://arxiv.org/abs/2605.06641)\n- **Abstract**:\n> Developing ceramic glazes is a costly, time-consuming process of trial and error due to complex chemistry, placing a significant burden on independent artists. While recent advances in multimodal AI offer a modern solution, the field lacks the large-scale datasets required to train these models. We propose GlazyBench, the first dataset for AI-assisted glaze design. Comprising 23,148 real glaze formulations, GlazyBench supports two primary tasks: predicting post-firing surface properties, such as color and transparency, from raw materials, and generating accurate visual representations of the glaze based on these properties. We establish comprehensive baselines for property prediction using traditional machine learning and large language models, alongside image generation benchmarks using deep generative and large multimodal models. Our experiments demonstrate promising yet challenging results. GlazyBench pioneers a new research direction in AI-assisted material design, providing a standardized benchmark for systematic evaluation.\n\n---\n\n## 2. Can RL Teach Long-Horizon Reasoning to LLMs? Expressiveness Is Key\n- **Authors**: Tianle Wang , Zhaoyang Wang , Guangchen Lan , Xinpeng Wei , Sipeng Zhang , Guanwen Qiu , Abulhair Saparov\n- **URL**: [https://arxiv.org/abs/2605.06638](https://arxiv.org/abs/2605.06638)\n- **Abstract**:\n> Reinforcement learning (RL) has been applied to improve large language model (LLM) reasoning, yet the systematic study of how training scales with task difficulty has been hampered by the lack of controlled, scalable environments. We introduce ScaleLogic, a synthetic logical reasoning framework that offers independent control over two axes of difficulty: the depth of the required proof planning (i.e., the horizon) and the expressiveness of the underlying logic. Our proposed framework supports a wide range of logics: from simple implication-only logic (\"if-then\") towards more expressive first-order reasoning with conjunction (\"and\"), disjunction (\"or\"), negation (\"not\"), and universal quantification (\"for all\"). Using this framework, we show that the RL training compute $T$ follows a power law with respect to reasoning depth $D$ ($T \\propto D^{\\gamma}$, $R^{2} > 0.99$), and that the scaling exponent $\\gamma$ increases monotonically with logical expressiveness, from $1.04$ to $2.60$. On downstream mathematics and general reasoning benchmarks, more expressive training settings yield both larger performance gains (up to $+10.66$ points) and more compute-efficient transfer compared to less expressive settings, demonstrating that what a model is trained on, not just how much it is trained, shapes downstream transfer. We further show that the power-law relationship holds across multiple RL methods, and curriculum-based training substantially improves scaling efficiency.\n\n---\n\n## 3. MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems\n- **Authors**: Zhexuan Wang , Xuebo Liu , Li Wang , Zifei Shan , Yutong Wang , Zhenxi Song , Min Zhang\n- **URL**: [https://arxiv.org/abs/2605.06623](https://arxiv.org/abs/2605.06623)\n- **Abstract**:\n> Large language model (LLM)-based Multi-agent systems (MAS) have shown promise in tackling complex collaborative tasks, where agents are typically orchestrated via role-specific prompts. While the quality of these prompts is pivotal, jointly optimizing them across interacting agents remains a non-trivial challenge, primarily due to the misalignment between local agent objectives and holistic system goals. To address this, we introduce MASPO, a novel framework designed to automatically and iteratively refine prompts across the entire system. A core innovation of MASPO is its joint evaluation mechanism, which assesses prompts not merely by their local validity, but by their capacity to facilitate downstream success for successor agents. This effectively bridges the gap between local interactions and global outcomes without relying on ground-truth labels. Furthermore, MASPO employs a data-driven evolutionary beam search to efficiently navigate the high-dimensional prompt space. Extensive empirical evaluations across 6 diverse tasks demonstrate that MASPO consistently outperforms state-of-the-art prompt optimization methods, achieving an average accuracy improvement of 2.9. We release our code at this https URL .\n\n---\n\n## 4. SkillOS: Learning Skill Curation for Self-Evolving Agents\n- **Authors**: Siru Ouyang , Jun Yan , Yanfei Chen , Rujun Han , Zifeng Wang , Bhavana Dalvi Mishra , Rui Meng , Chun-Liang Li , Yizhu Jiao , Kaiwen Zha , Maohao Shen , Vishy Tirumalashetty , George Lee , Jiawei Han , Tomas Pfister , Chen-Yu Lee\n- **URL**: [https://arxiv.org/abs/2605.06614](https://arxiv.org/abs/2605.06614)\n- **Abstract**:\n> LLM-based agents are increasingly deployed to handle streaming tasks, yet they often remain one-off problem solvers that fail to learn from past interactions. Reusable skills distilled from experience provide a natural substrate for self-evolution, where high-quality skill curation serves as the key bottleneck. Existing approaches either rely on manual skill curation, prescribe heuristic skill operations, or train for short-horizon skill operations. However, they still struggle to learn complex long-term curation policies from indirect and delayed feedback. To tackle this challenge, we propose SkillOS, an experience-driven RL training recipe for learning skill curation in self-evolving agents. SkillOS pairs a frozen agent executor that retrieves and applies skills with a trainable skill curator that updates an external SkillRepo from accumulated experience. To provide learning signals for curation, we design composite rewards and train on grouped task streams based on skill-relevant task dependencies, where earlier trajectories update the SkillRepo, and later related tasks evaluate these updates. Across multi-turn agentic tasks and single-turn reasoning tasks, SkillOS consistently outperforms memory-free and strong memory-based baselines in both effectiveness and efficiency, with the learned skill curator generalizing across different executor backbones and task domains. Further analyses show that the learned curator produces more targeted skill use, while the skills in SkillRepo evolve into more richly structured Markdown files that encode higher-level meta-skills over time.\n\n---\n\n## 5. NeuroAgent: LLM Agents for Multimodal Neuroimaging Analysis and Research\n- **Authors**: Lujia Zhong , Yihao Xia , Jianwei Zhang , Shuo huang , Jiaxin Yue , Mingyang Xia , Yonggang Shi\n- **URL**: [https://arxiv.org/abs/2605.06584](https://arxiv.org/abs/2605.06584)\n- **Abstract**:\n> Multimodal neuroimaging analysis often involves complex, modality-specific preprocessing workflows that require careful configuration, quality control, and coordination across heterogeneous toolchains. Beyond preprocessing, downstream statistical analysis and disease classification commonly require task-specific code, evaluation protocols, and data-format conventions, creating additional barriers between raw acquisitions and reproducible scientific analysis. We present NeuroAgent, an LLM-driven agentic framework that automates key preprocessing and analysis steps for heterogeneous neuroimaging data, including sMRI, fMRI, dMRI, and PET, and supports interactive downstream analysis through natural-language queries. NeuroAgent employs a hierarchical multi-agent architecture with a feedback-driven Generate-Execute-Validate engine: agents autonomously generate executable preprocessing code, detect and recover from runtime errors, and validate output integrity. We evaluate the system on 1,470 subjects pooled across all ADNI phases (CN=1,000, AD=470), where all subjects have sMRI and tabular data, with subsets also having Tau-PET (n=469), fMRI (n=278), and DTI ($n=620$). Pipeline ablation studies across multiple LLM backends show that capable models reach up to 100% intent-parsing accuracy, with the strongest backend (Qwen3.5-27B) reaching 84.8% end-to-end preprocessing step correctness. Automated recovery limits manual intervention to edge cases where human review is required via the Human-In-The-Loop interface. For Alzheimer's Disease classification using automatically preprocessed multimodal data, our agent ensemble achieves an AUC of 0.9518 with four modalities, outperforming all single-modality baselines. These results show that NeuroAgent can reduce the manual effort required for neuroimaging preprocessing and enable end-to-end automated analysis pipelines for neuroimaging research.\n\n---\n\n## 6. Process Matters more than Output for Distinguishing Humans from Machines\n- **Authors**: Milena Rmus , Mathew D. Hardy , Thomas L. Griffiths , Mayank Agrawal\n- **URL**: [https://arxiv.org/abs/2605.06524](https://arxiv.org/abs/2605.06524)\n- **Abstract**:\n> Reliable human-machine discrimination is becoming increasingly important as large language models and autonomous agents are deployed in online settings. Existing approaches evaluate whether a system can produce behavior or responses indistinguishable from those of a human, following the emphasis on outputs as a criterion for intelligence proposed by Alan Turing. Cognitive science offers an alternative perspective: evaluating the process by which behavior is produced. To test whether cognitive processes can reliably distinguish humans from machines, we introduce CogCAPTCHA30, a battery of 30 cognitive tasks designed to elicit diagnostic process-level features even when task performance is matched. Across the battery, process-level features provide stronger discriminative signal than performance metrics alone, reliably distinguishing humans from agents even under output matching (mean process-feature classifier AUC = 0.88). To evaluate agentic process differences, we compare off-the-shelf frontier agents (Claude Sonnet 4.5, GPT-5, Gemini 2.5 Pro), Centaur (a language model fine-tuned on 10.7M human decisions), and two task-specific fine-tuning approaches applied to Qwen2.5-1.5B-Instruct: action-level supervised fine-tuning (A-SFT) and process-level fine-tuning (P-SFT), which directly optimizes process features. Broad fine-tuning on human decisions improves human-like task processes relative to off-the-shelf agents, while task-specific process-level supervision further improves behavioral mimicry. However, this advantage diminishes under cross-task transfer when supervised process targets do not naturally generalize across tasks. Explicit process-level supervision can improve human behavioral mimicry, but only if appropriate task-specific process representations are available, highlighting process specification as a bottleneck for achieving human-like cognitive processes in machines.\n\n---\n\n## 7. Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors\n- **Authors**: Jonas Wiedermann-Möller , Leonard Dung , Maksym Andriushchenko\n- **URL**: [https://arxiv.org/abs/2605.06490](https://arxiv.org/abs/2605.06490)\n- **Abstract**:\n> AI systems have become increasingly capable of dangerous behaviours in many domains. This raises the question: Do models sometimes choose to violate human instructions in order to perform behaviour that is more useful for certain goals? We introduce a benchmark for measuring model propensity for instrumental convergence (IC) behaviour in terminal-based agents. This is behaviour such as self-preservation that has been hypothesised to play a key role in risks from highly capable AI agents. Our benchmark is realistic and low-stakes which serves to reduce evaluation-awareness and roleplay confounds. The suite contains seven operational tasks, each with an official workflow and a policy-violating shortcut. An eight-variant shared framework varies monitoring, instruction clarity, stakes, permission, instrumental usefulness and blocked honest paths to support inferences regarding the factors driving IC behaviour. We evaluated ten models using deterministic environment-state scorers over 1,680 samples, with trace review employed for audit and adjudication purposes. The final IC rate is 86 out of 1,680 samples (5.1%). IC behaviour is concentrated rather than uniform: two Gemini models account for 66.3% of IC cases and three tasks account for 84.9%. Conditions in which IC behaviour is indispensable for task success result in the greatest increase in the adjusted IC rate (+15.7 percentage points), whereas emphasising that task success is critical or certain framing choices do not produce comparable effects. Our findings indicate that realistic, low-nudge environments elicit IC behaviour rarely but systematically in most tested models. We conclude that it is feasible to robustly measure tendencies for dangerous behaviour in current frontier AI agents.\n\n---\n\n## 8. ReasonSTL: Bridging Natural Language and Signal Temporal Logic via Tool-Augmented Process-Rewarded Learning\n- **Authors**: Bowen Ye , Zhijian Li , Junyue Huang , Junkai Ma , Xiang Yin\n- **URL**: [https://arxiv.org/abs/2605.06483](https://arxiv.org/abs/2605.06483)\n- **Abstract**:\n> Signal Temporal Logic (STL) is an expressive formal language for specifying spatio-temporal requirements over real-valued, real-time signals. It has been widely used for the verification and synthesis of autonomous systems and cyber-physical systems. In practice, however, users often express their requirements in natural language rather than in structured STL formulas, making natural-language-to-STL translation a critical yet challenging task. Manual specification requires temporal-logic expertise and cannot scale, while prompting commercial LLM APIs incurs substantial token costs and may expose sensitive system requirements to third-party services, raising privacy concerns for industrial deployment. To address these challenges, we present \\textsc{ReasonSTL}, a tool-augmented framework that adapts local open-source language models for natural-language-to-STL generation. \\textsc{ReasonSTL} decomposes the translation process into explicit reasoning, deterministic tool calls, and structured formula construction. We further introduce process-rewarded training to supervise both tool-use trajectories and final formulas, together with \\textsc{STL-Bench}, a bilingual, computation-aware benchmark grounded in real-world signals. Experiments show that a 4B model trained with \\textsc{ReasonSTL} achieves state-of-the-art performance in both automatic metrics and human evaluations, demonstrating that \\textsc{ReasonSTL} provides a transparent, low-cost, and privacy-preserving alternative for formal specification drafting.\n\n---\n\n## 9. Patch-Effect Graph Kernels for LLM Interpretability\n- **Authors**: Ruben Fernandez-Boullon , David N. Olivieri\n- **URL**: [https://arxiv.org/abs/2605.06480](https://arxiv.org/abs/2605.06480)\n- **Abstract**:\n> Mechanistic interpretability aims to reverse-engineer transformer computations by identifying causal circuits through activation patching. However, scaling these interventions across diverse prompts and task families produces high-dimensional, unstructured datasets that are difficult to compare systematically. We propose a framework that reframes mechanistic analysis as a graph machine-learning problem by representing activation-patching profiles as patch-effect graphs over model components. We introduce three graph-construction methods: direct-influence via causal mediation, partial-correlation, and co-influence and apply graph kernels to analyze the resulting structures. Evaluating this approach on GPT-2 Small using Indirect Object Identification (IOI) and related tasks, we find that patch-effect graphs preserve discriminative structural signals. Specifically, localized edge-slot features provide higher classification accuracy than global graph-shape descriptors. A screened paired-patching validation suggests that CI and PC selected candidate edges correspond to stronger activation-influence effects than random or low-rank candidates. Crucially, by evaluating these representations against rigorous prompt-only and raw patch-effect controls, we make the evidential scope of the benchmark explicit: graph features compress structured patching signal, while raw tensors and surface cues define strong baselines that any circuit-level claim should address. Ultimately, our framework provides a compression and evaluation pipeline for comparing patching-derived structures under controlled baselines, separating robust slice-discriminative evidence from stronger task-general causal-circuit claims.\n\n---\n\n## 10. Beyond Task Success: Measuring Workflow Fidelity in LLM-Based Agentic Payment Systems\n- **Authors**: Donghao Huang , Joon Kiat Chua , Zhaoxia Wang\n- **URL**: [https://arxiv.org/abs/2605.06457](https://arxiv.org/abs/2605.06457)\n- **Abstract**:\n> LLM-based multi-agent systems are increasingly deployed for payment workflows, yet prevailing metrics, Task Success Rate (TSR) and Agent Handoff F1-Score (HF1), capture only final outcomes or unordered routing decisions. We introduce the Agentic Success Rate (ASR), a trajectory-fidelity metric that compares observed and expected agent execution sequences at the transition level, decomposing performance into Transition Recall and Transition Precision. Applied to the Hierarchical Multi-Agent System for Payments (HMASP) across 18 LLMs and 90,000 task instances, ASR reveals that 10 of 18 models systematically skip a confirmation checkpoint during payment checkout, a deviation invisible to both TSR and HF1, while 8 models enforce the checkpoint perfectly. Notably, GPT-4.1 exhibits hidden workflow shortcuts despite achieving perfect TSR and HF1, while GPT-5.2 achieves perfect ASR. Prompt refinements and deterministic routing guards guided by ASR diagnostics yield substantial TSR improvements, with gains up to +93.8 percentage points for previously struggling models, demonstrating that trajectory-level evaluation is essential in regulated domains.\n\n---\n\n## 11. PrefixGuard: From LLM-Agent Traces to Online Failure-Warning Monitors\n- **Authors**: Xinmiao Huang , Jinwei Hu , Rajarshi Roy , Changshun Wu , Yi Dong , Xiaowei Huang\n- **URL**: [https://arxiv.org/abs/2605.06455](https://arxiv.org/abs/2605.06455)\n- **Abstract**:\n> Large language model (LLM) agents now execute long, tool-using tasks where final outcome checks can arrive too late for intervention. Online warning requires lightweight prefix monitors over heterogeneous traces, but hand-authored event schemas are brittle and deployment-time LLM judging is costly. We introduce PrefixGuard, a trace-to-monitor framework with an offline StepView induction step followed by supervised monitor training. StepView induces deterministic typed-step adapters from raw trace samples, and the monitor learns an event abstraction and prefix-risk scorer from terminal outcomes. Across WebArena, $\\tau^2$-Bench, SkillsBench, and TerminalBench, the strongest PrefixGuard monitors reach 0.900/0.710/0.533/0.557 AUPRC. Using the strongest backend within each representation, they improve over raw-text controls by an average of +0.137 AUPRC. LLM judges remain substantially weaker under the same prefix-warning protocol. We also derive an observability ceiling on score-based area under the precision-recall curve (AUPRC) that separates monitor error from failures lacking evidence in the observed prefix. For finite-state audit, post-hoc deterministic finite automaton (DFA) extraction remains compact on WebArena and $\\tau^2$-Bench (29 and 20 states) but expands to 151 and 187 states on SkillsBench and TerminalBench. Finally, first-alert diagnostics show that strong ranking does not imply deployment utility: WebArena ranks well yet fails to support low-false-alarm alerts, whereas $\\tau^2$-Bench and TerminalBench retain more actionable early alerts. Together, these results position PrefixGuard as a practical monitor-synthesis recipe with explicit diagnostics for when prefix warnings translate into actionable interventions.\n\n---\n\n## 12. SCRuB: Social Concept Reasoning under Rubric-Based Evaluation\n- **Authors**: Jamelle Watson-Daniels , Himaghna Bhattacharjee , Skyler Wang , Brandon Handoko , Antonio Li , Anaelia Ovalle , Mahesh Pasupuleti , Candace Ross , Vidya Sarma , Arjun Subramonian , Karen Ullrich , Will van der Vaart , Yijing Xin , Maximilian Nickel\n- **URL**: [https://arxiv.org/abs/2605.06444](https://arxiv.org/abs/2605.06444)\n- **Abstract**:\n> While many studies of Large Language Model (LLM) reasoning capabilities emphasize mathematical or technical tasks, few address reasoning about social concepts: the abstract ideas shaping social norms, culture, and institutions. This understudied capability is essential for modern models acting as social agents, yet no systematic evaluation methodology targets it. We introduce SCRuB (Social Concept Reasoning under Rubric-Based Evaluation), a framework designed for this setting of task indeterminacy. Our goal is to measure the degree to which a model reasons about social concepts with the depth and critical rigor of a human expert. SCRuB proceeds in three phases: prompt construction from established sources, response generation by experts and models, and comparative evaluation using a five-dimensional critical thinking rubric. To enable generalization of the pipeline, we introduce a Panel of Disciplinary Perspectives ensemble validated against independent expert judges. We release SCRuBEval (n=4,711 evaluation prompts) and SCRuBAnnotations (300 expert-authored responses and 150 expert comparative judgments from 45 PhD-level scholars). Our results show that frontier models consistently outperform human experts across all five rubric dimensions. Across 1,170 pairwise comparisons, expert judges ranked a model response first in 80.8% of judgments and preferred model responses overall 74.4% of the time. Ultimately, this study provides the first expert-grounded demonstration of evaluation saturation for social concept reasoning: the single-turn exam-style format has reached its ceiling for models and humans alike.\n\n---\n\n## 13. Knowledge Graphs, the Missing Link in Agentic AI-based Formal Verification\n- **Authors**: Vaisakh Naduvodi Viswambharan , Keerthan Kopparam Radhakrishna , Deepak Narayan Gadde , Aman Kumar\n- **URL**: [https://arxiv.org/abs/2605.06434](https://arxiv.org/abs/2605.06434)\n- **Abstract**:\n> Recent advances in Large Language Models (LLMs) have enabled workflows that generate SystemVerilog Assertions (SVAs) from natural-language specifications, with the potential to accelerate Formal Verification (FV). However, high-quality assertion synthesis remains challenging because specifications are often ambiguous or incomplete and critical micro-architectural details reside in the Register Transfer Level (RTL). Many existing approaches treat the specification and RTL as loosely structured text, which weakens specification-to-RTL grounding and leads to semantic mismatches and frequent syntax failures during formal parsing and elaboration. This work addresses these limitations with a verification-centric Knowledge Graph (KG) constructed from structured Intermediate Representations (IRs) extracted from the specification, RTL, and formal-tool feedback, including syntax diagnostics, Counterexamples (CEXs), and coverage reports. The KG links requirements, design hierarchy, signals, assumptions, and properties to provide traceable, design-grounded context for generation. A multi-agent workflow queries and updates this KG to generate SVAs and to drive three refinement loops: syntax repair guided by tool diagnostics, CEX-guided correction using trace links, and coverage-directed property augmentation. Evaluation across seven benchmark designs indicates that KG-based context retrieval improves specification-to-RTL grounding and consistently produces compilable SVAs with low syntax-repair overhead. The approach achieves formal coverage ranging from 78.5% to 99.4%, though convergence exhibits design dependence with complex temporal and arithmetic reasoning remaining challenging for current LLM capabilities.\n\n---\n\n## 14. Rethinking Vacuity for OOD Detection in Evidential Deep Learning\n- **Authors**: Claire McNamara\n- **URL**: [https://arxiv.org/abs/2605.06382](https://arxiv.org/abs/2605.06382)\n- **Abstract**:\n> Vacuity, or Uncertainty Mass (UM), is commonly used as a metric to evaluate Out-of-Distribution (OOD) detection in Evidential Deep Learning (EDL). It generally involves dividing the number of classes ($K$) by the total strength of belief ($S$) of the model's predictions, where $S$ is derived from summing the Dirichlet parameters. As such, UM is sensitive to the cardinality of $K$. In particular, it is unlikely in practice that there is a linear relationship between $K$ and $S$ as $K$ and $S$ increase due to the nature of EDL (suppressing incorrectly assigned evidence). As a result, when comparing In Distribution (ID) and OOD results, it is important that $K_{\\mathrm{ID} }$ and $K_{\\mathrm{OOD} }$ are equal; something that is not always ensured in practice. We provide an empirical demonstration of how results for AUROC and AUPR can substantially differ when class cardinality between ID and OOD differs by 1, with AUROC differing by as much as 0.318 and AUPR by 0.613 for standard EDL, and AUROC by 0.360 and AUPR by 0.683 for IB-EDL. More concretely, our findings isolate an evaluation artefact: when K differs between ID and OOD, AUROC/AUPR can be artificially inflated without any change in model predictions. We further discuss the evaluation of EDL over causal language models using Multiple-Choice Question-Answer (MCQA) datasets and argue for clearer definitions of ID and OOD in this context. Our primary contribution is an empirical and theoretical demonstration that vacuity-based OOD detection in EDL-fine-tuned LLMs is highly sensitive to uncontrolled differences in evaluated class cardinality.\n\n---\n\n## 15. From Agent Loops to Deterministic Graphs: Execution Lineage for Reproducible AI-Native Work\n- **Authors**: Josh Rosen , Seth Rosen\n- **URL**: [https://arxiv.org/abs/2605.06365](https://arxiv.org/abs/2605.06365)\n- **Abstract**:\n> Large language model systems are increasingly deployed as agentic workflows that interleave reasoning, tool use, memory, and iterative refinement. These systems are effective at producing answers, but they often rely on implicit conversational state, making it difficult to preserve stable work products, isolate irrelevant updates, or propagate changes through intermediate artifacts. We introduce execution lineage: an execution model in which AI-native work is represented as a directed acyclic graph (DAG) of artifact-producing computations with explicit dependencies, stable intermediate boundaries, and identity-based replay. The goal is not to make the model a better one-shot writer, but to make evolving AI-generated work maintainable under change. We compare execution-lineage replay against loop-centric update baselines on two controlled policy-memo update tasks. In an unrelated-branch update, DAG replay preserved the final memo exactly in all runs, with zero churn and zero unrelated-branch contamination, while loop baselines regenerated the memo and frequently imported unrelated context. In an intermediate-artifact edit, all systems reflected the new constraint in the final memo, but only DAG replay achieved perfect upstream preservation, downstream propagation, unaffected-artifact preservation, and cross-artifact consistency. These results show that final answer quality and maintained-state quality are distinct. Strong loop baselines can remain competitive at producing polished final outputs when the task is a bounded synthesis/update problem and all current sources fit in context, but immediate task success can mask partial state inconsistency that may compound over future revisions. Execution lineage provides stronger guarantees about what should change, what should remain stable, and how work evolves across revisions.\n\n---\n\n## 16. A Regime Theory of Controller Class Selection for LLM Action Decisions\n- **Authors**: Zhaoyang Jiang , Zhizhong Fu , Yunsoo Kim , Jiacong Mi , Zicheng Li , Xuanqi Peng , Honghan Wu\n- **URL**: [https://arxiv.org/abs/2605.06339](https://arxiv.org/abs/2605.06339)\n- **Abstract**:\n> Deployed language and vision-language models must decide, on each input, whether to answer directly, retrieve evidence, defer to a stronger model, or abstain. Contrary to the common monotonicity intuition, greater per-input expressivity is not uniformly beneficial in finite samples: under identical strict cross-validation, different benchmarks prefer different controller classes. This reflects a finite-sample limitation of instance-level uncertainty signals, which can be exhausted at a distribution-dependent scale. We organize controllers into a nested lattice of four classes: fixed actions, partition routers, instance-level controllers, and prior-gated controllers, ordered by complexity. We prove a regime theory that turns three data-estimable bottlenecks into a class choice: how much improvement is possible beyond the best fixed action, whether there are enough samples for instance-level controllers to make reliable decisions, and how much improvement a coarse partition router can recover when instance-level signal is unreliable. The resulting Bernstein-tight threshold has a matching information-theoretic lower bound, and strict nested cross-validation provably selects a near-best class. Across SMS-Spam, HallusionBench, A-OKVQA, and FOLIO, the predicted class matches the empirical winner; the prior-gated controller wins on TextVQA when OCR tokens supply a label-free prediction-time prior. Code is available at this https URL .\n\n---\n\n## 17. Addressing Labelled Data Scarcity: Taxonomy-Agnostic Annotation of PII Values in HTTP Traffic using LLMs\n- **Authors**: Thomas Cory , Axel Küpper\n- **URL**: [https://arxiv.org/abs/2605.06305](https://arxiv.org/abs/2605.06305)\n- **Abstract**:\n> Automated privacy audits of web and mobile applications often analyse outbound HTTP traffic to detect Personally Identifiable Information (PII) leakage. However, existing learning-based detectors typically depend on scarce, manually labelled traffic and are tightly coupled to fixed label taxonomies, limiting transferability across domains and evolving definitions of PII. This paper investigates whether Large Language Models (LLMs) can support taxonomy-agnostic annotation of explicitly transmitted PII values in HTTP message bodies when the taxonomy is provided at runtime. We introduce a multi-stage LLM-based pipeline that combines deterministic pre-processing with label-level classification, targeted instance-level value annotation, and output validation. To enable controlled evaluation and exemplar-based prompting without relying on sensitive real-user captures, we further propose an LLM-based generator for synthetic HTTP traffic with manually validated, taxonomy-derived PII annotations. We evaluate the approach across three taxonomies spanning different PII domains and granularity levels. Results show that the pipeline accurately detects PII types and extracts corresponding values for concrete PII taxonomies. Overall, our findings position LLMs as a promising foundation for flexible, taxonomy-agnostic traffic annotation and for creating labelled data under evolving privacy taxonomies.\n\n---\n\n## 18. Data Language Models: A New Foundation Model Class for Tabular Data\n- **Authors**: Eda Erol , Giuliano Pezzoli , Ozer Cem Kelahmet\n- **URL**: [https://arxiv.org/abs/2605.06290](https://arxiv.org/abs/2605.06290)\n- **Abstract**:\n> Every major data modality now has a foundation model that understands it natively: text has language models, images have vision models, audio has audio models. Tabular data, the modality on which many consequential real-world AI decisions are made, does not. Every approach to tabular AI today, from gradient-boosted trees to the latest tabular foundation models, requires a preprocessing pipeline before any model can consume the data. None of them understand tabular data as a modality. We introduce the Data Language Model (DLM), the missing foundation model for tabular data. A DLM understands tables the way a language model understands sentences: natively, without serialization or preprocessing, directly from raw cell values. It is the tabular data layer on which AI models, agents, and vertical AI applications can be built, eliminating the preprocessing pipelines that currently stand between raw data and every AI system that consumes it. We present Schema-1, the first DLM: a 140M parameter model trained on more than 2.3M synthetic and real-world tabular datasets. Schema-1 outperforms gradient-boosted ensembles, AutoML stacks, and the tabular foundation models we evaluate on established row-level prediction benchmarks. On missing value reconstruction it achieves lower reconstruction error than all classical statistical methods and frontier large language models on mean performance across conditions, establishing that structural understanding of a dataset's own distributional geometry is more useful for imputation than world knowledge encoded in language. It identifies the industry sector of any unseen dataset from raw cell values alone, reliably across any domain, a task no prior tabular model can perform. It is the native tabular understanding layer that has been missing from the AI stack.\n\n---\n\n## 19. Joint Consistency: A Unified Test-Time Aggregation Framework via Energy Minimization\n- **Authors**: Yunzhen Yao , Hongye Wang , Yahong Wang , Michael C. Gastpar , Bo Jiang , Lie He\n- **URL**: [https://arxiv.org/abs/2605.06219](https://arxiv.org/abs/2605.06219)\n- **Abstract**:\n> This paper studies test-time aggregation, an approach that generates multiple reasoning traces and aggregates them into a final answer. Most existing methods rely on evaluation signals collected from candidate traces in isolation or answer frequencies, while ignoring comparative interactions among candidates. We propose Joint Consistency (JC), formulated as a constrained Ising-type energy minimization problem, where independent evaluation signals act as external fields and pairwise comparisons act as interactions. JC provides a unified framework for test-time aggregation that subsumes existing voting and weighted aggregation methods as special cases. Our construction of the interaction matrix leverages LLM-as-a-judge comparisons, and admits a theoretical interpretation under answer-level homogeneity assumptions. Moreover, we develop an efficient approximation strategy that makes interaction modeling practical for large-scale test-time aggregation. Experiments on math and code reasoning benchmarks show that JC consistently outperforms existing baselines across tasks, judge models, trace budgets, and trace-generation settings.\n\n---\n\n## 20. Beyond Fixed Benchmarks and Worst-Case Attacks: Dynamic Boundary Evaluation for Language Models\n- **Authors**: Haoxiang Wang , Da Yu , Huishuai Zhang\n- **URL**: [https://arxiv.org/abs/2605.06213](https://arxiv.org/abs/2605.06213)\n- **Abstract**:\n> Evaluating large language models (LLMs) today rests on fixed benchmarks that apply the same set of items to any model, producing ceiling and floor effects that mask capability gaps. We argue that the most informative evaluation signal lies at the boundary, where the per-prompt pass probability is near $0.5$ under random-sampling decoding, and propose Dynamic Boundary Evaluation (DBE), which actively locates each model's boundary and places it on a globally comparable difficulty scale. DBE delivers three artifacts: (i) a calibrated item bank covering safety, capability, and truthfulness, with per-item difficulty labels validated across $9$ reference LLMs; (ii) Skill-Guided Boundary Search (SGBS), a search algorithm that finds boundary items for a given target LLM using only API-level query access; and (iii) an evaluation protocol that places a new LLM on a unified ability scale and grows the evaluation set adaptively when the target falls outside the bank's coverage. We instantiate DBE on four categories spanning safety (harmful request refusal and over-refusal), capability (constrained instruction following), and truthfulness (multi-turn sycophancy resistance). The resulting evaluation covers a broader model spectrum without saturation while remaining compatible with existing datasets.\n\n---\n\n## 21. Towards Annotation-Free Validation of MLLMs: A Vision-Language Logical Consistency Metric\n- **Authors**: Ying Gu , Mei Chee Leong , Hui Li Tan , Shangbo Mao , Liyuan Li , Nancy Chen\n- **URL**: [https://arxiv.org/abs/2605.06201](https://arxiv.org/abs/2605.06201)\n- **Abstract**:\n> Dominant accuracy evaluation might reward unwarranted guessing of Large Language Models, and it might not be applicable to novel tasks for model validation without ground-truth (gt) annotation. Based on basic logic principle, we propose a novel framework to evaluate the vision-language logical consistency of MLLMs on both sufficient and necessary cause-effect relations. We define Vision-Language Logical Consistency Metric (VL-LCM) on traditional MC-VQA tests, and recent NaturalBench tests without the need for gt annotation. Through systematic experiments on representative VL benchmark MMMU and recent VL challenges like NaturalBench, we evaluated 11 recent open-source MLLMs from 4 frontier families. Our findings reveal that, despite significant progress of recent MLLMs on accuracy, logical consistency lags behind significantly. Extensive evaluations on the correlations of VL-LCM with metrics on gt, the reliability of LCM, and the relation of VL-LCM with response distribution justify the validity and applicability of VL-LCM even without gt annotation. Our findings suggest that, beyond accuracy, logical consistency could be employed for both accuracy and reliability. VL-LCM can also be employed for MLLM selection, validation, and reliable answer justification in novel tasks without gt annotation.\n\n---\n\n## 22. The Granularity Axis: A Micro-to-Macro Latent Direction for Social Roles in Language Models\n- **Authors**: Chonghan Qin , Xiachong Feng , Ziyun Song , Xiaocheng Feng , Jing Xiong , Lingpeng Kong\n- **URL**: [https://arxiv.org/abs/2605.06196](https://arxiv.org/abs/2605.06196)\n- **Abstract**:\n> Large language models (LLMs) are routinely prompted to take on social roles ranging from individuals to institutions, yet it remains unclear whether their internal representations encode the granularity of such roles, from micro-level individual experience to macro-level organizational, institutional, or national reasoning. We show that they do. We define a contrast-based Granularity Axis as the difference between mean macro- and micro-role hidden states. In Qwen3-8B, this axis aligns with the principal axis (PC1) of the role representation space at cosine 0.972 and accounts for 52.6% of its variance, indicating that granularity is the dominant geometric axis organizing prompted social roles. We construct 75 social roles across five granularity levels and collect 91,200 role-conditioned responses over shared questions and prompt variants, then extract role-level hidden states and project them onto the axis. Role projections increase monotonically across all five levels, remain stable across layers, prompt variants, endpoint definitions, held-out splits, and score-filtered subsets, and transfer to Llama-3.1-8B-Instruct. The axis is also causally relevant: activation steering along it shifts response granularity in the predicted direction, with Llama moving from 2.00 to 3.17 on a five-point macro scale under positive steering on prompts that admit local responses. The two models differ in controllability, suggesting that steering depends on each model's default operating regime. Overall, our findings suggest that social role granularity is not merely a stylistic surface feature, but a structured, ordered, and causally manipulable latent direction in role-conditioned language model behavior.\n\n---\n\n## 23. Systematic Evaluation of Large Language Models for Post-Discharge Clinical Action Extraction\n- **Authors**: Shivali Dalmia , Ananya Mantravadi , Prasanna Desikan\n- **URL**: [https://arxiv.org/abs/2605.06191](https://arxiv.org/abs/2605.06191)\n- **Abstract**:\n> The work in this paper evaluates zero-shot and few-shot large language models (LLMs) for safety-critical clinical action extraction using the CLIP discharge-note dataset, with particular emphasis on transitions of care and post-discharge patient safety. To manage the complexity of clinical documentation, we introduce a two-stage extraction framework that decomposes discharge notes, that are written in narrative form, into fine-grained, explicitly actionable clinical tasks through a staged prompting strategy. Our contributions include a systematic assessment of generative LLMs for clinical action extraction, a detailed comparison between general-purpose LLMs and task-specific supervised BERT-based models, and an analysis of annotation inconsistencies across different action categories. We show that contemporary LLMs achieve performance comparable to or exceeding supervised models on binary actionability detection, while supervised baselines retain a meaningful advantage on fine-grained multi-label category classification, despite the absence of task-specific fine-tuning and under strict data-privacy constraints. Qualitative error analysis reveals that many failures stem from misalignment between model reasoning and dataset annotation conventions, particularly in cases involving implicit clinical actions and rigid structural labeling rules. These results indicate that reported performance reflects model limitations due to lack of clinical reasoning, that is not captured by plain annotations. Labels without rationales make it impossible to distinguish clinical reasoning failures from annotation convention mismatches. Advancing clinical NLP requires reasoning-annotated datasets that document why specific spans are actionable, not merely which spans were labeled, enabling proper evaluation of model clinical understanding.\n\n---\n\n## 24. Event-Causal RAG: A Retrieval-Augmented Generation Framework for Long Video Reasoning in Complex Scenarios\n- **Authors**: Peizheng Yan , Yu Zhao , Liang Xie , Juntong Qi , Mingming Wang , Erwei Yin\n- **URL**: [https://arxiv.org/abs/2605.06185](https://arxiv.org/abs/2605.06185)\n- **Abstract**:\n> Recent large vision-language models have achieved strong performance on short- and medium-length video understanding, yet they remain inadequate for ultra-long or even infinite video reasoning, where models must preserve coherent memory over extended durations and infer causal dependencies across temporally distant events. Existing end-to-end video understanding methods are fundamentally limited by the $O(n^2)$ complexity of self-attention, while recent retrieval-augmented generation (RAG) approaches still suffer from fragmented clip-level memory, weak modeling of temporal and causal structure, and high storage and online inference costs. We present Event-Causal RAG, a lightweight retrieval-augmented framework for infinite long-video reasoning. Instead of indexing fixed-length clips, our method segments streaming videos into semantically coherent events and represents each event as a structured State-Event-State (SES) graph, capturing the event together with its surrounding state transitions. These graphs are merged into a global Event Knowledge Graph and stored in a dual-store memory that supports both semantic matching and causal-topological retrieval. On top of this memory, we design a bidirectional retrieval strategy to efficiently identify the most relevant event causal chains and provide them, together with the associated video evidence, to a backbone video foundation model for answer generation. Experiments on long-video understanding benchmarks demonstrate that Event-Causal RAG consistently outperforms strong clip-based retrieval baselines and long-context video models, particularly on questions requiring multi-event integration and causal inference across long temporal gaps, while also achieving improved memory efficiency and robust streaming performance.\n\n---\n\n## 25. Post Reasoning: Improving the Performance of Non-Thinking Models at No Cost\n- **Authors**: Richmond Sin Jing Xuan , Rishabh Bhardwaj , Soujanya Poria\n- **URL**: [https://arxiv.org/abs/2605.06165](https://arxiv.org/abs/2605.06165)\n- **Abstract**:\n> As the widespread adoption of Large Language Models (LLMs) accelerates, token consumption from intermediate reasoning traces increasingly contributes to inference latency and operational cost. Recent studies suggest that many real-world tasks require little to no explicit reasoning, with additional reasoning sometimes even degrading performance. In this work, we propose \\textbf{Post-Reasoning}, a simple yet effective approach that improves instruction-tuned models by conditioning them to justify their answers after generating the final response. By design, it enables the final answer to be obtained without additional latency or token cost, while still improving performance through simple instruction augmentation. We evaluate Post-Reasoning across \\(117\\) model--benchmark settings spanning \\(13\\) open and proprietary models, \\(4\\) model families, and \\(9\\) diverse reasoning and knowledge-intensive benchmarks, including AMC, HMMT, GSM8K, GPQA, MMLU-Pro, and BIG-Bench Hard. Post-Reasoning improves performance in over \\(88.19\\%\\) of evaluated settings, achieving a mean relative improvements of \\(17.37\\%\\). Furthermore, we propose supervised post-reason tuning, which further improves performance in over \\(91.11\\%\\) of evaluated settings, and exceeds the prompt-based post-reasoning baseline by an average of \\(8.01\\%\\), demonstrating that post-reasoning can be effectively internalized through training. Ultimately, Post-Reasoning establishes a new performance ceiling for direct-answer capabilities.\n\n---\n\n## 26. Beyond Accuracy: Policy Invariance as a Reliability Test for LLM Safety Judges\n- **Authors**: Shihao Weng , Yang Feng , Xiaofei Xie\n- **URL**: [https://arxiv.org/abs/2605.06161](https://arxiv.org/abs/2605.06161)\n- **Abstract**:\n> LLM-as-a-Judge pipelines have become the de facto evaluator for agent safety, yet existing benchmarks treat their verdicts as ground-truth proxies without checking whether the verdicts depend on the agent's behavior or merely on how the evaluation policy happens to be worded. We argue that any trustworthy safety judge must satisfy a basic property we call policy invariance, and we operationalize it as three testable principles: rubric-semantics invariance under certified-equivalent rewrites, rubric-threshold invariance under intentional strict-to-lenient shifts, and ambiguity-aware calibration so that verdict instability concentrates on genuinely ambiguous cases. Instantiating these principles as a stress-test protocol with four agent-class judges on trajectories drawn from ASSEBench and R-Judge, we surface a previously unmeasured failure mode: today's judges respond to meaningful normative shifts and to meaningless structural rewrites with comparable strength, and cannot tell the two apart. Content-preserving policy rewrites flip up to 9.1% of verdicts above baseline jitter, and 18-43% of all observed flips occur on unambiguous cases under such rewrites, so existing safety scores conflate what the agent did with how the evaluator was prompted. Beyond the diagnosis, we contribute the Policy Invariance Score and the Judge Card reporting protocol, which expose an order-of-magnitude spread in judge reliability that is invisible to accuracy-only leaderboards. We release the protocol and code so that future agent-safety benchmarks can audit their own evaluators rather than trust them by default.\n\n---\n\n## 27. Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning\n- **Authors**: Yaorui Shi , Yuxin Chen , Zhengxi Lu , Yuchun Miao , Shugui Liu , Qi GU , Xunliang Cai , Xiang Wang , An Zhang\n- **URL**: [https://arxiv.org/abs/2605.06130](https://arxiv.org/abs/2605.06130)\n- **Abstract**:\n> A persistent skill library allows language model agents to reuse successful strategies across tasks. Maintaining such a library requires three coupled capabilities. The agent selects a relevant skill, utilizes it during execution, and distills new skills from experience. Existing methods optimize these capabilities in isolation or with separate reward sources, resulting in partial and conflicting evolution. We propose Skill1, a framework that trains a single policy to co-evolve skill selection, utilization, and distillation toward a shared task-outcome objective. The policy generates a query to search the skill library, re-ranks candidates to select one, solves the task conditioned on it, and distills a new skill from the trajectory. All learning derives from a single task-outcome signal. Its low-frequency trend credits selection and its high-frequency variation credits distillation. Experiments on ALFWorld and WebShop show that Skill1 outperforms prior skill-based and reinforcement learning baselines. Training dynamics confirm the co-evolution of the three capabilities, and ablations show that removing any credit signal degrades the evolution.\n\n---\n\n## 28. Back to the Beginning of Heuristic Design: Bridging Code and Knowledge with LLMs\n- **Authors**: Nguyen Viet Tuan Kiet , Bui Dinh Pham , Dao Van Tung , Tran Cong Dao , Huynh Thi Thanh Binh\n- **URL**: [https://arxiv.org/abs/2605.06123](https://arxiv.org/abs/2605.06123)\n- **Abstract**:\n> Large language models (LLMs) have recently advanced automatic heuristic design (AHD) for combinatorial optimization (CO), where candidate heuristics are iteratively proposed, evaluated, and refined. Most existing approaches search over executable programs and distill insights from execution feedback to guide later iterations. Because this process moves from low-level implementations to high-level principles, we refer to it as a bottom-up paradigm. We argue that this view is incomplete and introduce a complementary top-down perspective: knowledge becomes the primary search object and code merely instantiates and tests it, making what is learned explicit and reusable across problems and trajectories. We formalize this shift through a statistical-learning view that exposes a distortion--compression trade-off, and instantiate it in both population-based and tree-based AHD frameworks. Across CO and tasks beyond it, knowledge-first search improves discovery efficiency, transfer, and generalization, often outperforming code-centric pipelines, while combining both strategies yields further gains. Our results suggest that progress in AHD depends on iteratively constructing and evolving interpretable hypotheses that retain value beyond a single search trajectory.\n\n---\n\n## 29. Policy-Guided Stepwise Model Routing for Cost-Effective Reasoning\n- **Authors**: Wenwen Si , Insup Lee , Osbert Bastani\n- **URL**: [https://arxiv.org/abs/2605.06116](https://arxiv.org/abs/2605.06116)\n- **Abstract**:\n> Inference-time computation has greatly enhanced the performance of large language models (LLMs) on challenging reasoning tasks, but this strategy can incur high inference costs. One solution is to route intermediate chain-of-thought (CoT) states to language models of different sizes; however, existing approaches rely on handcrafted routing strategies that limit performance, or on training large process reward models that may be infeasible in many applications. We formulate stepwise model routing as a constrained decision-making problem, which we solve by training a small control policy using reinforcement learning in conjunction with threshold calibration to tune the performance-efficiency tradeoff. We validate our method on three math benchmarks (GSM8K, MATH500, and OmniMath) on both open and closed models. Our method consistently improves the accuracy-cost tradeoff compared to handcrafted approaches, while achieving a comparable tradeoff to methods that require training large process reward models.\n\n---\n\n## 30. CrossCult-KIBench: A Benchmark for Cross-Cultural Knowledge Insertion in MLLMs\n- **Authors**: Zhen Zeng , Leijiang Gu , Feng Li , Jing Yu , Zenglin Shi\n- **URL**: [https://arxiv.org/abs/2605.06115](https://arxiv.org/abs/2605.06115)\n- **Abstract**:\n> Multimodal Large Language Models (MLLMs), trained primarily on English-centric data, frequently generate culturally inappropriate or misaligned responses in cross-cultural settings. To mitigate this, we introduce the task of cross-cultural knowledge insertion, which focuses on adapting models to specific cultural contexts while preserving their original behavior in other cultures. To facilitate research in this area, we introduce CrossCult-KIBench, a comprehensive evaluation benchmark for assessing both the effectiveness of knowledge insertion and its unintended side effects on non-target cultures. The benchmark includes 9,800 image-grounded cases covering 49 culturally relevant visual scenarios across English, Chinese, and Arabic language-culture groups. It supports evaluation in both single-insert and sequential-insert settings. We also propose Memory-Conditioned Knowledge Insertion (MCKI) as a baseline method. MCKI retrieves relevant cultural knowledge from an external memory using frozen MLLM representations, prepending matched entries as conditional prompts when applicable. Extensive experiments on CrossCult-KIBench reveal that current approaches struggle to balance effective cultural adaptation with behavioral preservation, highlighting a key challenge in developing culturally-aware MLLMs. Our work thus underscores an important research direction for developing more culturally adaptive and responsible MLLMs.\n\n---\n\n## 31. Shallow Prefill, Deep Decoding: Efficient Long-Context Inference via Layer-Asymmetric KV Visibility\n- **Authors**: Jungsuk Oh , Hyeseo Jeon , Hyunjune Ji , Kyongmin Kong , Jay-Yoon Lee\n- **URL**: [https://arxiv.org/abs/2605.06105](https://arxiv.org/abs/2605.06105)\n- **Abstract**:\n> Long-context inference in decoder-only language models is costly because long prompts are processed during Prefill, cached at every layer, and repeatedly attended to during autoregressive Decode. We introduce \\emph{Shallow Prefill, dEEp Decode} (SPEED), a phase-asymmetric KV-visibility policy that materializes non-anchor prompt-token KV states only in lower layers while keeping Decode-phase tokens full-depth. Unlike previous approaches that make upper-layer prompt KV states cheaper to store or construct, SPEED removes prefill tokens from the upper-layer Decode visibility set altogether. With a minimal BoS anchor, this simple change preserves broad benchmark quality while reducing long-context cost. In a controlled Llama-3.1-8B instruction-tuning study, SPEED using only 75\\% of layers for prefill tokens reaches 51.2 average score on OLMES-style benchmarks, compared with 51.4 for the full-depth baseline, while improving TTFT by 33\\%, TPOT by 22\\%, and reducing active KV memory by 25.0\\% at 128K context. Layer-wise diagnostics suggest that this cutoff retains the main prompt-selection and representation-stabilization regions of the full-depth model. These results show that long-context prompt tokens need not always persist as full-depth KV-cache objects when Decode-phase tokens remain full-depth.\n\n---\n\n## 32. VibeServe: Can AI Agents Build Bespoke LLM Serving Systems?\n- **Authors**: Keisuke Kamahori , Shihang Li , Simon Peter , Baris Kasikci\n- **URL**: [https://arxiv.org/abs/2605.06068](https://arxiv.org/abs/2605.06068)\n- **Abstract**:\n> For years, we have built LLM serving systems like any other critical infrastructure: a single general-purpose stack, hand-tuned over many engineer-years, meant to support every model and workload. In this paper, we take the opposite bet: a multi-agent loop that automatically synthesizes bespoke serving systems for different usage scenarios. We propose VibeServe, the first agentic loop that generates entire LLM serving stacks end-to-end. VibeServe uses an outer loop to plan and track the search over system designs, and an inner loop to implement candidates, check correctness, and measure performance on the target benchmark. In the standard deployment setting, where existing stacks are highly optimized, VibeServe remains competitive with vLLM, showing that generation-time specialization need not come at the cost of performance. More interestingly, in non-standard scenarios, VibeServe outperforms existing systems by exploiting opportunities that generic systems miss in six scenarios involving non-standard model architectures, workload knowledge, and hardware-specific optimizations. Together, these results suggest a different point in the design space for infrastructure software: generation-time specialization rather than runtime generality. Code is available at this https URL .\n\n---\n\n## 33. Visual Fingerprints for LLM Generation Comparison\n- **Authors**: Amal Alnouri , Andreas Hinterreiter , Christina Humer , Furui Cheng , Marc Streit\n- **URL**: [https://arxiv.org/abs/2605.06054](https://arxiv.org/abs/2605.06054)\n- **Abstract**:\n> Large language model (LLM) outputs arise from complex interactions among prompts, system instructions, model parameters, and architecture. We refer to specific configurations of these factors as generation conditions, each of which can bias outputs in various ways. Understanding how different generation conditions shape model behaviors is essential for tasks such as prompt design and model evaluation, yet it remains challenging due to the stochastic and open-ended nature of text generation. We present an approach to visually compare LLM outputs across generation conditions by modeling responses as collections of linguistic choices, including content, expression, and structure. We extract these choices using natural language processing pipelines and represent their distributions across repeated samples. We then visualize these distributions as visual fingerprints, enabling direct, distribution-level comparison of condition-specific tendencies. Through four usage scenarios, we demonstrate how visual fingerprints reveal consistent patterns in LLM behavior that are difficult to observe through individual responses or aggregate metrics.\n\n---\n\n## 34. Novelty-based Tree-of-Thought Search for LLM Reasoning and Planning\n- **Authors**: Leon Hamm , Zlatan Ajanovic\n- **URL**: [https://arxiv.org/abs/2605.06040](https://arxiv.org/abs/2605.06040)\n- **Abstract**:\n> Although advances such as chain-of-thought, tree-of-thought or reinforcement learning have improved the performance of LLMs in reasoning and planning tasks, they are still brittle and have not achieved human-level performance in many domains, and often suffer from high time and token costs. Inspired by the success of width-based search in planning, we explore how the concept of novelty can be transferred to language domains and how it can improve tree-of-thought reasoning. A tree of thoughts relies on building possible \"paths\" of consecutive ideas or thoughts. These are generated by repeatedly prompting an LLM. In our paper, a measurable concept of novelty is proposed that describes the uniqueness of a new node (thought) in comparison to nodes previously seen in the search tree. Novelty is estimated by prompting an LLM and making use of embedded general knowledge from pre-training. This metric can then be used to prune branches and reduce the scope of the search. Although this method introduces more prompts per state, the overall token cost can be reduced by pruning and reducing the overall tree size. This procedure is tested and compared using several benchmarks in language-based planning and general reasoning.\n\n---\n\n## 35. Strat-LLM: Stratified Strategy Alignment for LLM-based Stock Trading with Real-time Multi-Source Signals\n- **Authors**: Wenliang Huang , Zengyi Yu\n- **URL**: [https://arxiv.org/abs/2605.06024](https://arxiv.org/abs/2605.06024)\n- **Abstract**:\n> Large Language Models (LLMs) are evolving into autonomous trading agents, yet existing benchmarks often overlook the interplay between architectural reasoning and strategy consistency. We propose Strat-LLM, a framework grounded in Stratified Strategy Alignment. Operating in a live-forward setting throughout 2025, it integrates heterogeneous data including sequential prices, real-time news, and annual reports to eliminate look-ahead bias. Extensive stress tests on A-share and U.S. markets reveal: (1) reasoning-heavy models achieve peak utility in Free Mode via internal logic, whereas standard models require Strict Mode as a vital risk anchor; (2) alignment utility is regime-dependent, with Free and Guided modes capturing momentum in uptrending markets, while Strict Mode mitigates drawdowns in downtrends; (3) mid-scale models (35B) show optimal fidelity under strict constraints, whereas ultra-large models (122B) suffer an alignment tax under rigid rules but gain a performance premium in Guided Mode; (4) standard LLMs often fall into a high win-rate trap, optimizing for small gains at the expense of total returns, which can only be mitigated through deep reasoning or strict external guardrails. Project details are available at this https URL .\n\n---\n\n## 36. TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steering\n- **Authors**: Yuan Sui , Yulin Chen , Yibo Li , Xue Jiang , Yufei He , Yihong Dong , Xiaoxin He , Tianyu Gao , Bryan Hooi\n- **URL**: [https://arxiv.org/abs/2605.05980](https://arxiv.org/abs/2605.05980)\n- **Abstract**:\n> When language model agents tackle complex software engineering tasks, they often degrade over long trajectories, which we define as *agent drift*. We focus on two recurring failure modes *overthinking* and *overacting*, i.e., where the agent repeatedly reasons over information it already has, and where it issues tool calls without integrating recent observations or acquiring new evidence. In this paper, we introduce TACT (Think-Act Calibration via activation Steering), to detect and mitigate agent drift in the residual stream before it surfaces as a behavioral failure. In specific, we label trajectory steps as overthinking, overacting, or calibrated, and find that their hidden states can separate linearly along two *drift axes*, pointing from calibrated behavior toward each failure mode (AUC $\\approx$ 0.9). To mitigate agent drift, we project each step's activation onto these axes at test time and pull drifted ones back toward the calibrated region. Experiments show that TACT outperforms unsteered baselines across SWE-bench Verified, Terminal-Bench 2.0, and CLAW-Eval, lifting average resolve rate by $+5.8$ pp on Qwen3.5-27B and $+4.8$ pp on Gemma-4-26B-A4B-it while cutting steps-to-resolve by up to $26\\%$. These gains frame agent drift as a steerable direction in the residual stream, and position TACT as a viable handle for reliable long-horizon agents.\n\n---\n\n## 37. TheraAgent: Self-Improving Therapeutic Agent for Precise and Comprehensive Treatment Planning\n- **Authors**: Junkai Li , Yunghwei Lai , Tianyi Zhu , Zheng Long Lee , Weizhi Ma , Yang Liu\n- **URL**: [https://arxiv.org/abs/2605.05963](https://arxiv.org/abs/2605.05963)\n- **Abstract**:\n> Formulating a treatment plan is inherently a complex reasoning and refinement task rather than a simple generation problem. However, existing large language models (LLMs) mainly rely on one-shot output without explicit verification, which may result in rough, incomplete, and potentially unsafe treatment plans. To address these limitations, we propose TheraAgent, an agentic framework that replaces one-shot generation with an iterative generate-judge-refine pipeline. By mirroring the actual reasoning process of human experts who iteratively revise treatment plans, our framework progressively transforms coarse and incomplete drafts into precise, comprehensive, and safer therapeutic regimens. To facilitate the critical judge component, we introduce TheraJudge, a treatment-specific evaluation module integrated into the inference loop to enforce clinical standards. Experiments show TheraAgent achieves state-of-the-art results on HealthBench, leading in Accuracy and Completeness. In expert evaluations, it attains an 86% win rate against physicians, with superior Targeting and Harm Control. Moreover, the highly agreement between TheraJudge and HealthBench evaluations confirms the reliability of our framework.\n\n---\n\n## 38. ICU-Bench:Benchmarking Continual Unlearning in Multimodal Large Language Models\n- **Authors**: Yuhang Wang , Wenjie Mei , Junkai Zhang , Guangyu He , Zhenxing Niu , Haichang Gao\n- **URL**: [https://arxiv.org/abs/2605.05938](https://arxiv.org/abs/2605.05938)\n- **Abstract**:\n> Although Multimodal Large Language Models (MLLMs) have achieved remarkable progress across many domains, their training on large-scale multimodal datasets raises serious privacy concerns, making effective machine unlearning increasingly necessary. However, existing benchmarks mainly focus on static or short-sequence settings, offering limited support for evaluating continual privacy deletion requests in realistic deployments. To bridge this gap, we introduce ICU-Bench, a continual multimodal unlearning benchmark built on privacy-critical document data. ICU-Bench contains 1,000 privacy-sensitive profiles from two document domains, medical reports and labor contracts, with 9,500 images, 16,000 question-answer pairs, and 100 forget tasks. Additionally, new continual unlearning metrics are introduced, facilitating a comprehensive analysis of forgetting effectiveness, historical forgetting preservation, retained utility, and stability throughout the continual unlearning process. Through extensive experiments with representative unlearning methods on ICU-Bench, we show that existing methods generally struggle in continual settings and exhibit clear limitations in balancing forgetting quality, utility preservation, and scalability over long task sequences. These findings highlight the need for multimodal unlearning methods explicitly designed for continual privacy deletion.\n\n---\n\n## 39. Wisteria: A Unified Multi-Scale Feature Learning Framework for DNA Language Model\n- **Authors**: Weihua Wang , Haoji Li , Feilong Bao , Lei Yang , Guanglai Gao\n- **URL**: [https://arxiv.org/abs/2605.05913](https://arxiv.org/abs/2605.05913)\n- **Abstract**:\n> DNA language model aims to decipher the regulatory grammar and semantic of genomes by capturing long range dependencies in DNA sequences. Existing methods emphasize long range token interactions but often ignore the interplay between local motifs and global dependencies. In this paper, we propose Wisteria, a genomic language model that integrates multi scale feature learning within a unified framework for DNA sequence. Specifically, Wisteria augments the Mamba based architecture with gated dilated convolutions to capture local motifs and regulatory patterns, while gated multilayer perceptrons refine global dependencies. We further introduce a Fourier based attention mechanism to support frequency domain modeling, periodic extension and length generalization. Across four experimental settings with both short and long range dependencies, Wisteria demonstrates strong performance on downstream benchmarks against competitive DNA language model baselines. These results indicate that Wisteria effectively unifies local and global dependency modeling for multi scale genomic sequence analysis.\n\n---\n\n## 40. Null Space Constrained Contrastive Visual Forgetting for MLLM Unlearning\n- **Authors**: Yuhang Wang , Zhenxing Niu , Haoxuan Ji , Guangyu He , Linlin Zhang , Haichang Gao\n- **URL**: [https://arxiv.org/abs/2605.05909](https://arxiv.org/abs/2605.05909)\n- **Abstract**:\n> The core challenge of machine unlearning is to strike a balance between target knowledge removal and non-target knowledge retention. In the context of Multimodal Large Language Models (MLLMs), this challenge becomes even more pronounced, as knowledge is further divided into visual and textual modalities that are tightly intertwined. In this paper, we introduce an MLLM unlearning approach that aims to forget target visual knowledge while preserving non-target visual knowledge and all textual knowledge. Specifically, we freeze the LLM backbone and achieve unlearning by fine-tuning the visual module. First, we propose a Contrastive Visual Forgetting (CVF) mechanism to separate target visual knowledge from retained visual knowledge, guiding the representations of target visual concepts toward appropriate regions in the feature space. Second, we identify the null space associated with retained knowledge and constrain the unlearning process within this space, thereby significantly mitigating degradation in knowledge retention. Third, beyond static unlearning scenarios, we extend our approach to continual unlearning, where forgetting requests arrive sequentially. Extensive experiments across diverse benchmarks demonstrate that our approach achieves a strong balance between effective forgetting and robust knowledge retention.\n\n---\n\n## 41. Taklif.AI: LLM-Powered Platform for Interest-Based Personalized College Assignments\n- **Authors**: Zaki Kurdya , Mohammed Zuqlam , Salem Amassi , Shady Telbany , Motaz Saad\n- **URL**: [https://arxiv.org/abs/2605.05842](https://arxiv.org/abs/2605.05842)\n- **Abstract**:\n> Educators face significant challenges in creating engaging, personalized assignments that accommodate students' diverse interests and cognitive abilities. Traditional one-size-fits-all assignments frequently lead to decreased student engagement and increased reliance on unethical practices such as plagiarism. To address these challenges, we present this http URL , a platform that leverages Large Language Models (LLMs) to automatically generate personalized assignments tailored to individual student interests. Unlike existing AI-powered educational platforms that personalize based on academic performance metrics alone, this http URL incorporates students' extracurricular interests and cultural contexts into the assignment generation process through a structured prompt engineering pipeline with input and output guardrails. The platform employs a serverless architecture on AWS with this http URL , using Llama 3.3 70B as the primary LLM via LiteLLM for multi-provider load balancing and LangChain for prompt orchestration. We describe the system architecture, the prompt design methodology, and the guardrails framework that ensures output quality. Preliminary user acceptance testing with 68 participants (65 students and 3 educators) indicates positive reception, with 84% of participants rating the personalization feature as beneficial. We discuss the platform's current capabilities and limitations, and outline directions for rigorous empirical evaluation of learning outcomes.\n\n---\n\n## 42. On the Role of Language Representations in Auto-Bidding: Findings and Implications\n- **Authors**: Guanyu Zhu , Jining Luan , Hanwen Du , Xinyu Fang , Sibo Xu , Ersheng Ni , Hongji Li , Jincheng Fang , Ronghao Chen , Huacan Wang , Xuanqi Lan , Yongxin Ni , Yiqi Sun , Youhua Li\n- **URL**: [https://arxiv.org/abs/2605.05833](https://arxiv.org/abs/2605.05833)\n- **Abstract**:\n> Auto-bidding is a crucial task in real-time advertising markets, where policies must optimize long-horizon value under delivery constraints (e.g., budget and CPA). Existing methods for auto-bidding rely on compact numerical state representations: while they can implicitly capture delivery dynamics, they offer limited support for explicitly representing and controlling high-level intent, evolving feedback, and operator-style strategic guidance in real campaigns. Meanwhile, Large Language Models (LLMs) offer a powerful method for encoding semantic information, it remains unclear when LLMs help and how to integrate them without sacrificing numerical precision. Through systematic preliminary studies, we find that (1) LLM embeddings contain bidding-relevant cues yet cannot replace numerical features, and (2) gains emerge only with careful semantic--numeric integration rather than naive concatenation. Motivated by these findings, we propose \\textit{SemBid}, a novel auto-bidding framework that injects LLM-encoded semantics into offline bidding trajectories at the token level. SemBid introduces three semantic inputs: \\textit{Task}, \\textit{History}, and \\textit{Strategy}. It injects these semantics as tokens alongside numerical trajectory tokens and uses self-attention to integrate them, improving controllability and generalization across objectives. Across diverse scenarios and budget regimes, SemBid outperforms competitive baselines from offline RL and generative sequence modeling, with more consistent gains in overall performance, constraint satisfaction, and robustness. Our code is available at: \\href{ this https URL }{\\textcolor{blue}{here} }.\n\n---\n\n## 43. AGPO: Asymmetric Group Policy Optimization for Verifiable Reasoning and Search Ads Relevance at JD\n- **Authors**: Yang Xu , Kun Yao , Yiming Deng , Zheng Fang , Kai Ming Ting , Ming Pang\n- **URL**: [https://arxiv.org/abs/2605.05826](https://arxiv.org/abs/2605.05826)\n- **Abstract**:\n> Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated notable success in enhancing the reasoning performance of large language models (LLMs). However, recent studies reveal that while current RLVR methods improve sampling efficiency towards correct paths, they do not elicit fundamentally new reasoning patterns. Instead, the reasoning capability boundary of trained models often narrows compared to their base models, with base models achieving higher coverage at large sample sizes. In this work, we propose Asymmetric Group Policy Optimization (AGPO) to counteract this boundary shrinkage. AGPO adopts a negative-dominant reinforcement strategy to suppress incorrect reasoning paths, maintaining the base model's exploration capacity. For positive reinforcement, AGPO adopts a group advantage mechanism, which scales positive updates based on intra-group variance, allowing the model to focus on rare correct paths while suppressing updates from trivial paths. Our experiments on five mathematical benchmarks demonstrate that AGPO achieves state-of-the-art accuracy while consistently improving pass@$k$ performance at scale. In a large-scale industrial application for search ads relevance optimization, AGPO effectively enhances the quality of the data annotation, leading to substantial performance gains in downstream student models.\n\n---\n\n## 44. CircuitFormer: A Circuit Language Model for Analog Topology Design from Natural Language Prompt\n- **Authors**: Md Touhidul Islam , Sujan Kumar Saha , Farimah Farahmandi , Mark Tehranipoor\n- **URL**: [https://arxiv.org/abs/2605.05773](https://arxiv.org/abs/2605.05773)\n- **Abstract**:\n> Automating analog circuit design remains a longstanding challenge in Electronic Design Automation (EDA). While Transformer-based Large Language Models (LLMs) have revolutionized software code generation, their application to analog hardware design is hindered by two critical limitations: (i) the scarcity of analog design datasets containing natural language description of a design and its corresponding netlist, and (ii) the inefficiency of general-purpose tokenizers (e.g., Byte Pair Encoding (BPE)) in capturing the inherent graph structure of circuits. To bridge this gap, first, we curate the largest annotated dataset of analog circuit netlists to date, comprising 31,341 netlist-natural language description pairs across all major circuit classes. Furthermore, we propose Circuit Tokenizer (CKT), a novel circuit graph tokenizer designed to encode netlist connectivity by explicitly mining frequent subcircuits. In terms of scalability, CKT overcomes the bottleneck of prior circuit graph serialization methods where vocabulary size scales linearly with maximum number of components in the dataset, n_max, (O(n_max)); instead, CKT decouples vocabulary growth from circuit complexity, achieving a constant O(1) complexity. Empirically, CKT outperforms standard BPE on circuit topology representation, reducing sequence length by 57% and achieving a 2.3x superior compression ratio using a compact, fixed vocabulary of size 512. Leveraging this optimized tokenization, we train a circuit-specific language model, CircuitFormer, a 511M parameter encoder-decoder transformer. Our model achieves 100% syntactic correctness and an 83% functional success rate across all major analog circuit categories, outperforming state-of-the-art open-source LLMs by 10% and 14%, respectively, while requiring 240x fewer parameters. The dataset is publicly available at this https URL .\n\n---\n\n## 45. HyperLens: Quantifying Cognitive Effort in LLMs with Fine-grained Confidence Trajectory\n- **Authors**: Chengda Lu , Xiaoyu Fan , Wei Xu\n- **URL**: [https://arxiv.org/abs/2605.05741](https://arxiv.org/abs/2605.05741)\n- **Abstract**:\n> While Large Language Models (LLMs) achieve strong performance across diverse tasks, their inference dynamics remain poorly understood because of the limited resolution of existing analysis tools. In this work, we identify an intrinsic magnification mechanism in transformer architectures: deeper layers inherently magnify the small changes of layer-wise confidence, providing a fine-grained confidence trajectory. Building on this insight, we introduce HyperLens, a high-resolution probe designed to trace confidence trajectories and quantify the cognitive effort during inference. Across LLMs and datasets, HyperLens reveals a consistent divergence in confidence trajectories that separates complex from simple tasks. We abstract this pattern into a quantitative cognitive effort metric. Our analysis reveals a fundamental principle: complex tasks consistently require higher cognitive effort. Finally, we provide a mechanistic diagnosis of a common side effect of standard Supervised Fine-Tuning (SFT): it can reduce cognitive effort and consequently degrade performance on in-domain tasks.\n\n---\n\n## 46. ReFlect: An Effective Harness System for Complex Long-Horizon LLM Reasoning\n- **Authors**: Fan Huang\n- **URL**: [https://arxiv.org/abs/2605.05737](https://arxiv.org/abs/2605.05737)\n- **Abstract**:\n> Current reasoning paradigms for LLMs include chain-of-thought, ReAct, and post-hoc self-critique. These paradigms rely on two assumptions that fail on long-horizon, multi-stage tasks. As a result, errors accumulate silently across reasoning steps, leaving an open question: can a reasoning system effectively detect and recover from its own failures? We present ReFlect, a \\emph{harness} system for LLM reasoning that creates standalone error detection and recovery logic as a deterministic wrapper around the model. Controlled experiments across 6 reasoning domains show that prompt-level self-critique produces formulaic templates that flag no issues in 90 of 100 audited reflection blocks, and the investigated LLMs wrongly accept a wrong answer in at least 76\\% of cases. Our ReFlect harness achieves task success rates ranging from 41\\% on gpt-4o-mini to 56\\% on Claude Sonnet 4.5 across six models spanning small and frontier scale, with per-model gains over Direct CoT ranging from +7 pp on Qwen2.5-72B to +29 pp on Claude Sonnet 4.5, and additionally raises SWE-bench patch-structural quality from 0\\% (Direct CoT) to between 82\\% (Qwen2.5-72B) and 87\\% (GPT-4o). Notably, the harness gain is inversely proportional to the model's Direct CoT task success rate (the fitted slope is -1.69 with r=-0.76): each pp lost in baseline success rate is mechanically recovered by 1.69 pp of harness gain. We spot that adding structured reasoning state and operators yields only 15.0--18.7\\% pair-mean on Llama-3.3-70B and Qwen2.5-72B because models at this scale cannot reliably populate the state its operators require. ReFlect is model-agnostic, training-free, and operates entirely at inference time.\n\n---\n\n## 47. Knee Osteoarthritis Severity Grading Using Optimized Deep Learning and LLM-Driven Intelligent AI on Computationally Limited Systems\n- **Authors**: Dayam Nadeem , Neha , Safdar Mustafa , Adnan Alvi , Mohd Hussain\n- **URL**: [https://arxiv.org/abs/2605.05731](https://arxiv.org/abs/2605.05731)\n- **Abstract**:\n> Knee osteoarthritis (KOA) is among the musculoskeletal disorders that considerably restrict joint mobility, cause severe chronic pain and impact negatively on quality life. It is one of the persistent health issues worldwide. Generally, subjectivity and inter-observer variability undermine conventional practices and evaluation process that are adopted to address such health issues. Hence precise and timely diagnosis would be one of the effective ways for the assessment of its severity. This paper proposes an automated diagnostic approach for severity grading of KOA by blending a deep learning convolutional neural network (CNN) with a device-based inference platform powered by TensorFlow Lite. It proposes a model based on the ResNet-18 convolutional neural network. The designed model is trained on publicly available database. Through a transfer learning approach obtained knee images are first classified into five Kellgren-Lawrence (KL) grades. Further the developed model is optimised. During the training of the model test accuracy of 94.48% with stable convergence has been achieved. Subsequently the optimised model transformed into a lightweight TensorFlow Lite format, facilitating seamless deployment on resource-constrained devices. The designed model is capable enough to operate in the environment having no continuous internet connectivity. Also, an auxiliary Large Language Model (Gemini-2.0-flash) is applied to generate structured interpretive findings like potential symptoms, risk factors, and preventive majors etc. The LLM component functions as interface without influencing the classification process. The proposed model articulates the feasibility of an on-device, interpretable decision-support tools for early diagnosis and improve accessibility to Artificial Intelligence (AI)-assisted knee screening tool.\n\n---\n\n## 48. SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents\n- **Authors**: Hongcheol Cho , Ryangkyung Kang , Youngeun Kim\n- **URL**: [https://arxiv.org/abs/2605.05726](https://arxiv.org/abs/2605.05726)\n- **Abstract**:\n> As LLM agents are increasingly deployed with large libraries of reusable skills, selecting the right skill for a user request has become a critical systems challenge. In small libraries, users may invoke skills explicitly by name, but this assumption breaks down as skill ecosystems grow under tight context and latency budgets. Despite its practical importance, skill retrieval remains underexplored, with limited benchmarks and little understanding of retrieval behavior on realistic skill libraries. To address this gap, we introduce SkillRet, a large-scale benchmark for skill retrieval in LLM agents. SkillRet contains 17,810 public agent skills, organized with structured semantic tags and a two-level taxonomy spanning 6 major categories and 18 sub-categories. It provides 63,259 training samples and 4,997 evaluation queries with disjoint skill pools, enabling both benchmarking and retrieval-oriented training. Across a diverse set of retrievers, we find that skill retrieval remains far from solved: off-the-shelf models struggle on realistic large-scale skill libraries, and prior skill-retrieval models still leave substantial headroom. Task-specific fine-tuning on SkillRet substantially improves performance, improving NDCG@10 by +13.1 points over the strongest prior retriever and by +16.9 points over the strongest off-the-shelf retriever. Our analysis further suggests that these gains arise because fine-tuned models better focus on the small skill-relevant signals within long and noisy queries. These results establish SkillRet as a strong benchmark and foundation for future research on retrieval in large-scale agent systems.\n\n---\n\n## 49. Detecting Time Series Anomalies Like an Expert: A Multi-Agent LLM Framework with Specialized Analyzers\n- **Authors**: Hyeongwon Kang , Jeongseob Kim , Jinwoo Park , Pilsung Kang\n- **URL**: [https://arxiv.org/abs/2605.05725](https://arxiv.org/abs/2605.05725)\n- **Abstract**:\n> Recent studies have explored large language models for time-series anomaly detection, yet existing approaches often rely on a single general-purpose model to directly infer anomaly indices or intervals, limiting controllability, interpretability, and reliability for complex anomaly patterns. We propose SAGE (Specialized Analyzer Group for Expert-like Detection), a multi-agent framework for structured anomaly diagnosis in univariate time series. It decomposes anomaly analysis into four specialized Analyzers for point, structural, seasonal, and pattern anomalies. Each Analyzer applies family-specific numerical tools and diagnostic visualizations to generate evidence, while an evidence-grounded Detector consolidates the evidence into confidence-scored anomaly records with intervals and candidate types. A Supervisor then converts these structured records into analyst-facing diagnostic reports. SAGE further constructs synthetic in-context examples from normal-reference training segments, without using real anomalous segments or anomaly-type labels as in-context examples. Across three benchmarks, SAGE achieves the best average performance among strong ML/DL and language-model-based baselines. Ablation studies and human evaluation further show that the proposed framework improves detection reliability and the practical usefulness of diagnostic outputs.\n\n---\n\n## 50. More Is Not Always Better: Cross-Component Interference in LLM Agent Scaffolding\n- **Authors**: Ming Liu\n- **URL**: [https://arxiv.org/abs/2605.05716](https://arxiv.org/abs/2605.05716)\n- **Abstract**:\n> LLM agent systems are built by stacking scaffolding components (planning, tools, memory, self-reflection, retrieval) assuming more is better. We study cross-component interference (CCI): degradation when components interact destructively. We run a full factorial experiment over all 2^5=32 subsets of five components on HotpotQA and GSM8K with Llama-3.1-8B/70B (96 conditions, up to 10 seeds). The All-In system is consistently suboptimal: on HotpotQA, a single-tool agent surpasses All-In by 32% (F1 0.233 vs 0.177, p=0.023); on GSM8K, a 3-component subset beats All-In by 79% (0.43 vs 0.24, p=0.010). Optimal component count is task-dependent (k*=1-4) and scale-sensitive: at 70B, combinations that hurt at 8B provide gains, though All-In still trails the best subset. We fit a main-effects regression (R^2=0.916, adj-R^2=0.899, LOOCV=0.872), compute exact Shapley values, and find 183/325 submodularity violations (56.3%), showing greedy selection is unreliable. A three-body synergy among Tool Use, Self-Reflection, and Retrieval (INT_3=+0.175, 95% CI [+0.003,+0.351]) is reported as exploratory. CCI replicates across model families (Qwen2.5) and is robust to prompt paraphrasing. Our findings suggest maximally-equipped agent defaults should be replaced by task-specific subset selection via interaction-aware analysis.\n\n---\n\n## 51. Decodable but Not Corrected by Fixed Residual-Stream Linear Steering: Evidence from Medical LLM Failure Regimes\n- **Authors**: Ming Liu\n- **URL**: [https://arxiv.org/abs/2605.05715](https://arxiv.org/abs/2605.05715)\n- **Abstract**:\n> Can linearly decodable failure signals in LLM hidden states be leveraged to correct those failures? We investigate this classification-correction gap via Overthinking (OT)--a stable behavioral regime (Jaccard >= 0.81, 94% inter-annotator agreement) in medical QA where models answer correctly under resampling yet fail in extended chain-of-thought. OT is linearly decodable at 71.6% balanced accuracy (p < 10^{-16}). Yet five families of fixed linear steering (29 configurations, n=1,273) all yield Delta ~= 0, with identical null results cross-architecture (Qwen2.5-7B) and cross-domain (MMLU-STEM). Three convergent lines of evidence suggest representational entanglement: the OT direction has 85-88% overlap with task-critical computation (specificity ratio <= 0.152); non-targeted shared-direction steering damages accuracy (-12.1pp); and LEACE concept erasure damages accuracy (-3.6pp, p=0.01), while 10 random erasures produce Delta=+0.3pp. The per-instance probe-steering correlation is r=-0.002 (p=0.97). Positively, the same probe enables selective abstention (held-out AUROC=0.610, exceeding all five uncertainty baselines, p=0.009): decodable failure structure supports post-generation reliability estimation even when the fixed linear steering family cannot exploit it for correction.\n\n---\n\n## 52. Conceal, Reconstruct, Jailbreak: Exploiting the Reconstruction-Concealment Tradeoff in MLLMs\n- **Authors**: Md Farhamdur Reza , Richeng Jin , Tianfu Wu , Huaiyu Dai\n- **URL**: [https://arxiv.org/abs/2605.05709](https://arxiv.org/abs/2605.05709)\n- **Abstract**:\n> Intent-obfuscation-based jailbreak attacks on multimodal large language models (MLLMs) transform a harmful query into a concealed multimodal input to bypass safety mechanisms. We show that such attacks are governed by a \\emph{reconstruction--concealment tradeoff}: the transformed input must hide harmful intent from safety filters while remaining recoverable enough for the victim model to reconstruct the original request. Through a reconstruction analysis of three representative black-box methods, we find that existing transformations struggle to balance this tradeoff, limiting their effectiveness. In contrast, we show that character-removed variants achieve a better balance. Building on this, we propose \\emph{concealment-aware variant construction}, which greedily selects character-removed variants that are low in harmful-keyword alignment and mutually diverse, and instantiates them through five modality-aware prompting strategies. We further introduce \\emph{keyword-related distractor images} that depict the harmful keyword in diverse contexts, providing more effective auxiliary visual context than generic distractor images. Experiments across closed-source and open-source MLLMs show the proposed strategies outperform strong baselines, revealing an underexplored vulnerability: a model's own reconstruction ability can be exploited to recover hidden harmful intent and produce unsafe responses.\n\n---\n\n## 53. Resolving the bias-precision paradox with stochastic causal representation learning for personalized medicine\n- **Authors**: Peisong Zhang , Manqiang Peng , Yuxuan Wu , Pawit Phadungsaksawasdi , Wesley Yeung , Ye Zhang , Trang Nguyen , Qiang Zhang , Nan Liu , Meng Wang , Kee Yuan Ngiam , Yih-Chung Tham , Ching-Yu Cheng , Tianfan Fu , Qingyu Chen , Rosemary Ke , Chang Li , Wenzhuo Yang , Zhenghao Lu , Chunyou Lai , Yu Zhang , Sheng Zhong , Hao Deng , Dianbo Liu\n- **URL**: [https://arxiv.org/abs/2605.05706](https://arxiv.org/abs/2605.05706)\n- **Abstract**:\n> Estimating individualized treatment effects from longitudinal observational data is central to data-driven medicine, yet existing methods face a fundamental limitation: reducing confounding bias often suppresses clinically informative heterogeneity, degrading patient-specific predictions. Here, we identify this tension as a bias-precision paradox in causal representation learning and introduce sampling-based maximum mean discrepancy (sMMD), a stochastic alignment strategy that replaces global adversarial balancing with subset-level matching. We instantiate this approach in a framework for counterfactual outcome prediction with attribution-grounded interpretability. Across two large-scale ICU cohorts (n = 27,783), our framework improves accuracy under distribution shift, reducing error by up to 11.5% and substantially increasing recall in high-risk tasks. Mechanistic analyses show that sMMD selectively preserves clinically decisive variables. In human-AI evaluation, our method outperforms clinicians-in-training and large language models, and improves clinician accuracy by 14.7% while reducing decision time, enabling interpretable, real-time clinical decision support.\n\n---\n\n## 54. Knowledge-Graph Paths as Intermediate Supervision for Self-Evolving Search Agents\n- **Authors**: Huyu Wu , Jun Liu , Xiaochi Wei , Yan Gao , Yi Wu , Yao Hu\n- **URL**: [https://arxiv.org/abs/2605.05702](https://arxiv.org/abs/2605.05702)\n- **Abstract**:\n> Self-evolving search agents reduce reliance on human-written training questions by generating and solving their own search tasks. We build on Search Self-Play (SSP), a representative Proposer and Solver framework in which questions are generated and answered via multi-step search and reasoning. In practice, however, SSP faces two bottlenecks: the Proposer constructs questions from isolated answer entities without relational context, yielding many invalid or unverifiable questions in early self-play training, while the Solver receives only a binary outcome reward that discards useful signal from partially on-track search trajectories. We address both bottlenecks by reusing knowledge-graph paths as construction-derived intermediate supervision for both question construction and reward shaping. First, we ground question construction in LLM-guided knowledge-graph subgraphs, providing relational context for the Proposer. Second, we observe that constructing and solving a multi-hop question can involve overlapping intermediate entities: the factual bridges used to formulate the question may provide approximate waypoints for answering it. Exploiting this overlap, we introduce Waypoint Coverage Reward (WCR), which grants graded partial credit to incorrect Solver trajectories according to their coverage of entities on the construction path, while preserving full reward for correct answers. Across seven QA benchmarks and nine model configurations, our approach improves the average score over standard SSP in all configurations, including notable gains on multi-hop QA tasks. These results suggest that knowledge-graph paths can be reused as lightweight intermediate supervision, providing both relational guidance and process feedback without additional task-specific human annotations or manually labeled process steps.\n\n---\n\n## 55. Inference-Time Budget Control for LLM Search Agents\n- **Authors**: Zhengru Fang , Senkang Forest Hu , Zhonghao Chang , Yu Guo , Yihang Tao , Hongyao Liu , Mengzhe Ruan , Jun Huang , Yuguang Fang\n- **URL**: [https://arxiv.org/abs/2605.05701](https://arxiv.org/abs/2605.05701)\n- **Abstract**:\n> LLM search agents increasingly rely on tools at inference time, but their trajectories are often constrained by hard limits on both tool calls and generated tokens. Under such dual budgets, better answers require not only stronger models, but also explicit control over which search action should receive the next budget unit and when the accumulated evidence is sufficient to commit a final answer. We study this problem in multi-hop question answering (QA) and formulate it as two-stage inference-time budget control. At search time, our controller assigns each feasible action a task-level Value-of-Information (VOI) score, defined as an operational estimate of marginal task value per unit budget under the current search state and remaining dual budget, and uses this score to choose among retrieval, decomposition, and answer commitment. After search, a selective evidence-grounded finalizer compares the trajectory answer with a refined candidate and rewrites only when the residual error appears to be a low-risk answer-form error. Across four multi-hop QA benchmarks, three LLM backbones, and four budget levels, the method yields positive aggregate gains over four audited baselines under the same hard dual-budget protocol. Ablations show that search-time budget control, especially budget-dependent penalty, provides the main performance gain, while answer-time control helps mainly when the retrieval path is already adequate. These results suggest that inference-time budget control for LLM search agents should govern both how budget is spent during search and how the final answer is committed.\n\n---\n\n## 56. Saliency-Aware Regularized Quantization Calibration for Large Language Models\n- **Authors**: Yanlong Zhao , Xiaoyuan Cheng , Huihang Liu , Baihua He , Xinyu Zhang , Harrison Bo Hua Zhu , Wenlong Chen , Li Zeng , Zhuo Sun\n- **URL**: [https://arxiv.org/abs/2605.05693](https://arxiv.org/abs/2605.05693)\n- **Abstract**:\n> Post-training quantization (PTQ) is an effective approach for deploying large language models (LLMs) under memory and latency constraints. Most existing PTQ methods determine quantization parameters by minimizing a layer-wise reconstruction error on a predetermined calibration dataset, usually optimized via either scale search or Gram-based methods. However, from the perspective of generalization risk, existing calibration objectives of PTQ based only on empirical reconstruction error on limited or unrepresentative calibration data could move the quantized weights away from the original weights. This may cause the generalization risk to diverge, potentially degrading downstream performance. To address this issue, we propose \\emph{Saliency-Aware Regularized Quantization Calibration} (SARQC) a unified framework that augments the standard PTQ objective with a saliency-aware regularization term. This term encourages quantized weights to stay close to the original weights during calibration, leading to improved generalization during inference. SARQC integrates seamlessly into existing PTQ pipelines, enhancing both scale search and Gram-based methods under a unified formulation. Extensive experiments on dense and Mixture-of-Experts LLMs demonstrate consistent improvements in perplexity and zero-shot accuracy, without additional computational overhead during inference.\n\n---\n\n## 57. DataDignity: Training Data Attribution for Large Language Models\n- **Authors**: Xiaomin Li , Andrzej Banburski-Fahey , Jaron Lanier\n- **URL**: [https://arxiv.org/abs/2605.05687](https://arxiv.org/abs/2605.05687)\n- **Abstract**:\n> Auditing language-model outputs often requires more than judging correctness: an auditor may need to identify which source document most likely supports the knowledge expressed in a response. We study this as pinpoint provenance: given a prompt, a target-model response, and a candidate corpus, rank the documents that best support the response. We introduce FakeWiki, a controlled benchmark of 3,537 fabricated Wikipedia-style articles designed to preserve ground-truth provenance while weakening lexical shortcuts. FakeWiki includes QA probes, source-preserving paraphrases, retro-generated variants, hard anti-documents that remain topically similar while removing answer-critical facts, and five query conditions: clean prompting plus four jailbreak-inspired transformations. We evaluate seven retrieval baselines, a training-free activation-steering retrieval-fusion method, SteerFuse, and a supervised contrastive provenance ranker, ScoringModel. ScoringModel maps response and document features into a shared space and is trained with InfoNCE using in-batch, retrieval-mined, and anti-document negatives. Across nine open-weight instruction-tuned LLMs and five query conditions, ScoringModel improves mean Recall@10 from 35.0 for the strongest retrieval baseline to 52.2, without inference-time fusion, and wins 41/45 model-by-condition cells. SteerFuse is usually second-best despite requiring no supervised training, showing that activation-space evidence can efficiently complement text retrieval. On jailbreak-inspired transformed queries, ScoringModel improves Recall@10 by 15.7 points on average over the best baseline. Overall, our work shows that robust training data attribution requires evaluation settings that separate true answer support from topical or lexical resemblance.\n\n---\n\n## 58. Attractor Geometry of Transformer Memory: From Conflict Arbitration to Confident Hallucination\n- **Authors**: Qiyao Liang , Risto Miikkulainen , Ila Fiete\n- **URL**: [https://arxiv.org/abs/2605.05686](https://arxiv.org/abs/2605.05686)\n- **Abstract**:\n> Language models draw on two knowledge sources: facts baked into weights (parametric memory, PM) and information in context (working memory, WM). We study two mechanistically distinct failure modes--conflict, when PM and WM disagree and interfere; and hallucination, when the queried fact was never learned. Both produce confident output regardless, making output-based monitoring blind by design. We show both failures share a unified geometric account. In the hidden-state space of autoregressive generation, learned facts form attractor basins. Conflict is basin competition: WM disrupts convergence to the correct basin without raising output entropy. Hallucination is basin absence: the hidden state drifts freely when no memorized basin exists. The frozen LM head, designed for next-token prediction, cannot distinguish these cases and fires confidently either way. We verify this account in a controlled synthetic task--entity identifiers mapped to unique codes with PM installed via LoRA adapters--where ground truth is exact and component roles can be causally isolated through targeted adapter placement. Geometric margin--the hidden state's distance to the nearest memorized basin--reads this geometry directly and separates correct recall from hallucination far more cleanly than output entropy, with zero false refusals where entropy-based detection cannot avoid rejecting the vast majority of correct outputs. The separation holds on natural-language factual queries from the pretrained model with no adaptation, confirming attractor geometry is structural rather than a fine-tuning artifact. The fraction of confident hallucinations follows a scaling law $C = \\exp(-c/\\bar\\Delta)$, growing with scale even as overall error rates fall. Hidden states reliably encode epistemic state; the frozen output head systematically erases it--and this erasure worsens with scale.\n\n---\n\n## 59. Large Vision-Language Models Get Lost in Attention\n- **Authors**: Gongli Xi , Ye Tian , Mengyu Yang , Huahui Yi , Liang Lin , Xiaoshuai Hao , Kun Wang , Wendong Wang\n- **URL**: [https://arxiv.org/abs/2605.05668](https://arxiv.org/abs/2605.05668)\n- **Abstract**:\n> Despite the rapid evolution of training paradigms, the decoder backbone of large vision--language models (LVLMs) remains fundamentally rooted in the residual-connection Transformer architecture. Therefore, deciphering the distinct roles of internal modules is critical for understanding model mechanics and guiding architectural optimization. While prior statistical approaches have provided valuable attribution-based insights, they often lack a unified theoretical basis. To bridge this gap, we propose a unified framework grounded in information theory and geometry to quantify the geometric and entropic nature of residual updates. Applying this unified framework reveals a fundamental functional decoupling: Attention acts as a subspace-preserving operator focused on reconfiguration, whereas FFNs serve as subspace-expanding operators driving semantic innovation. Strikingly, further experiments demonstrate that replacing learned attention weights with predefined values (e.g., Gaussian noise) yields comparable or even superior performance across a majority of datasets relative to vanilla models. These results expose severe misallocation and redundancy in current mechanisms, suggesting that state-of-the-art LVLMs effectively ``get lost in attention'' rather than efficiently leveraging visual context.\n\n---\n\n## 60. Retrieval-Conditioned Topology Selection with Provable Budget Conservation for Multi-Agent Code Generation\n- **Authors**: Abhijit Talluri , Pujith Anne , Bhagavan Choudary Pendiyala , Raghavendra Chilukuri\n- **URL**: [https://arxiv.org/abs/2605.05657](https://arxiv.org/abs/2605.05657)\n- **Abstract**:\n> Multi-agent LLM systems for code generation face a fundamental routing problem: the optimal orchestration topology depends on the structural complexity of the code under modification, yet existing systems select topologies without consulting the codebase. We present Retrieval-Guided Adaptive Orchestration (RGAO), an architecture that closes this loop by extracting a structural complexity vector from a hierarchical code index before selecting the orchestration topology. RGAO operates within Code-Agent, a multi-agent framework whose sub-agents are governed by formal contracts with six-dimensional budget vectors. Our headline contribution is the composition of two previously separate lines of work -- complexity-conditioned LLM routing and formal resource algebras -- yielding a property neither admits alone: provable budget conservation under retrieval-conditioned dynamic topology selection. Concretely we contribute: (1) a complexity-conditioned topology router that reduces proxy-measured misrouting from 30.1% to 8.2%; (2) a budget algebra with a structural-induction conservation theorem; and (3) a hierarchical code retrieval engine. Empirical evaluation demonstrates sub-millisecond DAG construction and linear tree-index scalability.\n\n---\n\n## 61. Text-Graph Synergy: A Bidirectional Verification and Completion Framework for RAG\n- **Authors**: Jiarui Zhong , Hong Cai Chen\n- **URL**: [https://arxiv.org/abs/2605.05643](https://arxiv.org/abs/2605.05643)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) has become a core paradigm for enhancing factual grounding and multi-hop reasoning in Large Language Models (LLMs). Traditional text-based RAG often retrieves logically irrelevant pseudo-evidence, while graph-based RAG is frequently hindered by search-time pruning, which may discard potentially valid reasoning paths. Existing hybrid approaches primarily adopt simple evidence concatenation or unidirectional enhancement, which fails to address the fundamental \"Information Island\" problem caused by asymmetric reasoning flows between unstructured text and structured graphs. We propose \\textbf{TGS-RAG}, a unified framework for \\textbf{T}ext-\\textbf{G}raph \\textbf{S}ynergistic enhancement. TGS-RAG introduces a bidirectional mechanism: (i) a \\textbf{Graph-to-Text} channel that employs a Global Voting strategy from visited graph nodes to re-rank and refine textual evidence, filtering out semantic noise; and (ii) a \\textbf{Text-to-Graph} channel that utilizes the \\textbf{Memory-based Orphan Entity Bridging} algorithm. This algorithm utilizes textual cues to proactively resurrect valid but previously pruned reasoning paths from the search history without additional database overhead. Experimental results on multiple multi-hop reasoning benchmarks demonstrate that TGS-RAG significantly outperforms state-of-the-art baselines, achieving a superior balance between retrieval precision and computational efficiency.\n\n---\n\n## 62. Prober.ai: Gated Inquiry-Based Feedback via LLM-Constrained Personas for Argumentative Writing Development\n- **Authors**: Ran Bi , Shiyao Wei , Yuanyiyi Zhou\n- **URL**: [https://arxiv.org/abs/2605.05598](https://arxiv.org/abs/2605.05598)\n- **Abstract**:\n> The proliferation of large language models (LLMs) in educational settings has paradoxically undermined the cognitive processes they purport to support. Students increasingly outsource critical thinking to AI assistants that generate polished text on demand, resulting in measurable cognitive debt and diminished argumentative reasoning skills. We present this http URL , a web-based writing environment that inverts the conventional AI-tutoring paradigm: rather than generating or rewriting student text, the system constrains an LLM (Gemini 3 Flash Preview) through persona-specific system prompts and structured JSON output schemas to produce only targeted, inquiry-based questions about argumentative weaknesses. A two-phase interaction architecture -- Challenge and Unlock -- implements a pedagogical friction mechanism whereby revision suggestions are gated behind mandatory student reflection. The system's design is grounded in Toulmin's argumentation theory, research on peer feedforward questioning mechanisms, and evidence on AI-supported feedback in writing instruction. A functional prototype was developed in 36 hours during the NY EdTech Hackathon (March 2026), where it was awarded second place. We describe the system architecture, the prompt engineering methodology for constraining LLM output to pedagogically aligned JSON schemas, and discuss implications for scalable, cognition-preserving AI integration in writing education.\n\n---\n\n## 63. Causal Probing for Internal Visual Representations in Multimodal Large Language Models\n- **Authors**: Zehao Deng , Tianjie Ju , Zheng Wu , Liangbo He , Jun Lan , Huijia Zhu , Weiqiang Wang , Zhuosheng Zhang\n- **URL**: [https://arxiv.org/abs/2605.05593](https://arxiv.org/abs/2605.05593)\n- **Abstract**:\n> Despite the remarkable success of Multimodal Large Language Models (MLLMs) across diverse tasks, the internal mechanisms governing how they encode and ground distinct visual concepts remain poorly understood. To bridge this gap, we propose a causal framework based on activation steering to actively probe and manipulate internal visual representations. Through systematic intervention across four visual concept categories, our results reveal a divergence in concept encoding: entities exhibit distinct localized memorization, whereas abstract concepts are globally distributed across the network. Critically, this divergence uncovers a mechanistic driver of scaling laws: increasing model depth is indispensable for encoding distributed and complex abstract concepts, whereas entity localization remains remarkably invariant to scale. Furthermore, reverse steering uncovers that blocking explicit output triggers a surge in latent activations, exposing a compensatory mechanism between perception and generation. Finally, extending our analysis to visual reasoning, we expose a disconnect between perception and reasoning although MLLMs successfully recognize geometric relations, they treat them merely as static visual features, failing to trigger the procedural execution necessary for abstract problem-solving.\n\n---\n\n## 64. Belief Memory: Agent Memory Under Partial Observability\n- **Authors**: Junfeng Liao , Qizhou Wang , Jianing Zhu , Bo Du , Rui Yan , Xiuying Chen\n- **URL**: [https://arxiv.org/abs/2605.05583](https://arxiv.org/abs/2605.05583)\n- **Abstract**:\n> LLM agents that operate over long context depend on external memory to accumulate knowledge over time. However, existing methods typically store each observation as a single deterministic conclusion (e.g., inferring \"API~X failed\" from temporary errors), even though such observations are inherently partial and potentially ambiguous. By committing to one conclusion and discarding uncertainty, these methods introduce self-reinforcing error: the agent acts on the stored conclusion, never revisits alternatives, and reinforces the conclusion over time. To address this issue, we propose BeliefMem, which shifts the memory paradigm from committing to a single conclusion per observation to retaining multiple candidate conclusions with their probabilities. Concretely, BeliefMem stores the candidate conclusions as separate memory entries, each carrying a probability that is updated via Noisy-OR rules as new observations arrive. At retrieval, all candidates surface together with their probabilities, keeping alternatives visible to the agent. Since each conclusion in memory retains its probability, BeliefMem preserves the uncertainty that the deterministic paradigm discards, enabling the agent to act with high confidence on well-evidenced knowledge while retaining the capacity to update its confidence when new evidence arrives. Empirical evaluations on LoCoMo and ALFWorld benchmarks show that, even with limited data, BeliefMem achieves the best average performance, remarkably outperforming well-known baselines. More broadly, such probabilistic memory produces substantial gains and explores a new direction for agent memory in partially observable environments.\n\n---\n\n## 65. AlphaCrafter: A Full-Stack Multi-Agent Framework for Cross-Sectional Quantitative Trading\n- **Authors**: Yishuo Yuan , Jiayi Sheng , Sirui Zeng , Jiaqi Wang , Jiaheng Liu\n- **URL**: [https://arxiv.org/abs/2605.05580](https://arxiv.org/abs/2605.05580)\n- **Abstract**:\n> Financial markets are inherently non-stationary, driven by complex interactions among macroeconomic regimes, microstructural frictions, and behavioral dynamics. Building quantitative strategies that remain profitable demands the continuous coupling of factor discovery, regime-adaptive selection, and risk-constrained execution. Prevailing approaches, however, optimize these components under static or isolated assumptions. Factor mining frameworks typically treat alpha discovery as a one-time search process, implicitly assuming that factor efficacy persists across market regimes. Execution-oriented systems often adopt role-playing agent architectures that simulate anthropomorphic trading committees, introducing behavioral noise rather than systematic rationality. Consequently, a fully automated, rationality-driven framework unifying a coherent quantitative pipeline remains absent. We introduce AlphaCrafter, a full-stack multi-agent framework that closes this gap through a continuously adaptive factor-to-execution pipeline, designed to track and respond to evolving market conditions without manual intervention. AlphaCrafter operates via three specialized agents: a Miner that continuously expands the factor pool via LLM-guided search, a Screener that assesses prevailing market conditions to construct regime-conditioned factor ensembles, and a Trader that translates these ensembles into quantitative strategies under explicit risk constraints. Together, these three agents form a closed-loop cross-sectional trading system that adapts holistically to evolving market dynamics. Extensive experiments on CSI 300 and S&P 500 demonstrate that AlphaCrafter consistently outperforms state-of-the-art baselines in risk-adjusted returns while exhibiting the lowest cross-trial variance, confirming that integrated and adaptive factor-to-execution design yields robust trading performance.\n\n---\n\n## 66. Nonsense Helps: Prompt Space Perturbation Broadens Reasoning Exploration\n- **Authors**: Langlin Huang , Chengsong Huang , Jinyuan Li , Donghong Cai , Yuyi Yang , Jiaxin Huang\n- **URL**: [https://arxiv.org/abs/2605.05566](https://arxiv.org/abs/2605.05566)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards, particularly Group Relative Policy Optimization (GRPO), has significantly advanced the reasoning capabilities of Large Language Models (LLMs). However, in complex tasks, GRPO frequently suffers from the ``zero-advantage problem'': when all sampled rollouts for a query fail, the relative advantage collapses to zero. Consequently, the model loses effective training signals for these questions, wasting the training data and computational budget. While simply increasing the sampling budget for these questions is a common remedy, the static sampling policy inherently constrains reasoning exploration, limiting the success rate. In this paper, we propose Lorem Perturbation for Exploration (LoPE), a simple yet effective training framework to break this exploration bottleneck. We posit that task-irrelevant prompt-space perturbations can shift the model's output distribution enough to unlock orthogonal reasoning pathways for hard questions. Specifically, LoPE prepends sequences stochastically assembled from Lorem Ipsum vocabulary (a pseudo-Latin placeholder text) to the prompts before resampling. Experiments across 1.7B, 4B, and 7B models demonstrate that LoPE significantly outperforms resampling with the original prompts. Further analysis reveals that other Latin-based random sequences with low perplexity are also effective perturbations. Our results establish LoPE as a strong baseline for broadening exploration in LLM reinforcement learning.\n\n---\n\n## 67. SPARK: Self-Play with Asymmetric Reward from Knowledge Graphs\n- **Authors**: Hyobin Park , Taeseop Kim , Dong-Geol Choi\n- **URL**: [https://arxiv.org/abs/2605.05546](https://arxiv.org/abs/2605.05546)\n- **Abstract**:\n> Self-play reinforcement learning has shown strong performance in domains with formally verifiable structure, such as mathematics and coding, where both problem generation and reward computation can be grounded in explicit rules. Extending this paradigm to scientific literature is more challenging: the relationships among multi-modal elements within and across documents are rarely made explicit in text, which makes automatic generation of relational reasoning questions difficult and weakens the reliability of reward signals. We propose SPARK (Self-Play with Asymmetric Reward from Knowledge Graphs), a framework that automatically constructs a unified knowledge graph (KG) from multi-document scientific literature and uses it as the structural basis for self-play. KG paths over multimodal nodes serve as a source for generating relational reasoning questions, and structured facts stored in the KG provide a basis for verifiable reward computation. A single small vision-language model (sVLM) alternates between Proposer and Solver roles under information asymmetry against a fixed KG, a design that we believe can be naturally extended toward online adaptation in future work. We evaluate SPARK on public benchmarks and a self-constructed cross-document multi-hop QA dataset. Results show that SPARK consistently outperforms flat-corpus-based self-play baselines, and the performance gap widens as hop count increases, suggesting that KG-structure grounding contributes to relational multi-hop reasoning beyond what unstructured corpus grounding can provide.\n\n---\n\n## 68. AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases\n- **Authors**: Susheel Suresh , Hazel Mak , Shangpo Chou , Fred Kroon , Sahil Bhatnagar\n- **URL**: [https://arxiv.org/abs/2605.05538](https://arxiv.org/abs/2605.05538)\n- **Abstract**:\n> We present AgenticRAG, a practical agentic harness for retrieval and analysis over enterprise knowledge bases. Standard RAG pipelines place significant burden of grounding on the search stack, constraining the language model to a fixed candidate set chosen deep in the retrieval process. Our approach reduces this overdependence by layering a lightweight harness on top of existing enterprise search infrastructure, equipping a reasoning LLM with search, find, open, and summarize tools enabling the model to iteratively retrieve information, navigate within documents, and analyze evidence autonomously. On three open benchmarks we observe substantial gains: $49.6\\%$ recall@1 on BRIGHT (+21.8 pp over the best embedding baseline), 0.96 factuality on WixQA ($+13\\%$ relative improvement), and $92\\%$ answer correctness on FinanceBench--within 2 pp of oracle access to true evidence. Ablation studies show that the most significant factor is the shift from single-shot retrieval to agentic tool use ($5.9\\times$ improvement), while multi-query search and in-document navigation contribute to both quality and efficiency. We present various design choices in our agentic harness that were informed by pre-production deployments. Our results demonstrate its suitability for real-world enterprise production environments.\n\n---\n\n## 69. FoodCHA: Multi-Modal LLM Agent for Fine-Grained Food Analysis\n- **Authors**: Woojin Lee , Pranav Mekkoth , Ye Tian , Onat Gungor , Tajana Rosing\n- **URL**: [https://arxiv.org/abs/2605.05499](https://arxiv.org/abs/2605.05499)\n- **Abstract**:\n> The widespread adoption of camera-equipped mobile devices and wearables has enabled convenient capture of meal images, making food recognition a key component for real time dietary monitoring. However, real-world food images present challenges due to high intra-class similarity and the frequent presence of multiple food items within a single image. While deep learning models achieve strong performance in coarse grained classification, they often struggle to capture fine-grained attributes such as cooking style. Moreover, open-ended generation in modern vision-language models can produce non-canonical labels, limiting their practical deployment. We propose FoodCHA, a multimodal agentic framework that reformulates food recognition as a hierarchical decision-making process. By progressively anchoring predictions, FoodCHA guides subcategory identification using high-level categories and guides cooking style recognition using subcategories, improving semantic consistency and attribute-level discrimination. To ensure practical deployability, FoodCHA utilizes the compact Moondream-2B vision language model, which provides strong reasoning capability while maintaining lower computational and memory overhead. Experiments on FoodNExTDB show that FoodCHA outperforms Food-Llama-3.2-11B by 13.8% and 38.2% in category and subcategory recognition precision, respectively, and achieves a striking 153.2% improvement in cooking style classification precision.\n\n---\n\n## 70. FinRAG-12B: A Production-Validated Recipe for Grounded Question Answering in Banking\n- **Authors**: Denys Katerenchuk , Pablo Duboue , Keelan Evanini , David Gondek , Nithin Govindugari , Olivier Allauzen , Joshua Baptiste , David J More , Joshua Schechter\n- **URL**: [https://arxiv.org/abs/2605.05482](https://arxiv.org/abs/2605.05482)\n- **Abstract**:\n> Large language models (LLMs) are rapidly being adopted across various domains. However, their adoption in banking industry faces resistance due to demands for high accuracy, regulatory compliance, and the need for verifiable and grounded responses. We present a unified, data-efficient framework for training grounded domain-specific LLMs that optimizes answer quality, citation grounding, and calibrated refusal under real-world deployment constraints. First, we describe a data generation pipeline that combines LLM-as-a-Judge filtering, citation annotation, and curriculum learning with only 143M tokens. The resulting 12B model achieves high answer quality outperforming GPT-4.1 on citation grounding, with a modest citation tradeoff versus the untuned base. Second, we propose a calibrated refusal mechanism: training on 22% unanswerable examples yield a 12% \"I don't know\" rate, substantially improving over the base model's unsafe 4.3% rate while avoiding GPT-4.1's over-refusal (20.2%). Third, we present an end-to-end methodology spanning from data curation to quantized serving. The system is deployed at 40+ financial institutions, achieving a 7.1 percentage point improvement in query resolution (p < 0.001). Additionally, the model delivers 3-5x faster responses at 20-50x lower cost compared to GPT-4.1.\n\n---\n\n## 71. LANTERN: LLM-Augmented Neurosymbolic Transfer with Experience-Gated Reasoning Networks\n- **Authors**: Mahyar Alinejad , Yue Wang , Amrit Singh Bedi , George Atia\n- **URL**: [https://arxiv.org/abs/2605.05478](https://arxiv.org/abs/2605.05478)\n- **Abstract**:\n> Transfer learning in reinforcement learning (RL) seeks to accelerate learning in new tasks by leveraging knowledge from related sources. Existing neurosymbolic transfer methods, however, typically rely on manually specified task automata, assume a single source task, and use fixed knowledge-integration mechanisms that cannot adapt to varying source relevance. We propose LANTERN, a unified framework for multi-source neurosymbolic transfer that addresses these limitations through three components: (i) deterministic finite automata generated from natural language task descriptions using large language models, (ii) semantic embedding-based aggregation of multiple source policies weighted by cross-task similarity, and (iii) adaptive teacher-student gating based on temporal-difference error and semantic uncertainty. Across domains spanning resource management, navigation, and control, LANTERN achieves 40-60% improvements in sample efficiency over existing baselines while remaining robust to poorly aligned sources. These results demonstrate that multi-source, adaptively weighted neurosymbolic transfer can improve scalability and robustness in symbolic RL settings.\n\n---\n\n## 72. Agentic Discovery of Exchange-Correlation Density Functionals\n- **Authors**: Titouan Duston , Jiashu Liang , Yuanheng Wang , Weihao Gao , Xuelan Wen , Nan Sheng , Weiluo Ren , Yang Sun , Yixiao Chen\n- **URL**: [https://arxiv.org/abs/2605.05460](https://arxiv.org/abs/2605.05460)\n- **Abstract**:\n> The development of accurate exchange-correlation (XC) functionals remains a longstanding challenge in density functional theory (DFT). The vast majority of XC functionals have been hand designed by human researchers combining physical insight, exact constraints, and empirical fitting. Recent advances in large language models enable a systematic, automated alternative to this human-driven design loop. This report presents an agentic search system in which an LLM proposes structured functional-form changes guided by evolutionary history. The system attempts to improve functional performance through an iterative plan-execute-summarize loop, where improvements are measurable by optimizing functional parameters against a standard thermochemistry dataset, then evaluating performance on a held-out subset. The strongest discovered functional, SAFS26-a (Seed Agentic Functional Search 2026), improves upon the gold-standard {\\omega}B97M-V baseline by ~9%. These results also surface a cautionary lesson for AI-assisted science: models powerful enough to discover genuine improvements are equally capable of exploiting unphysical shortcuts to game the benchmark; domain expertise translated into explicitly enforced constraints remains essential to keeping results scientifically grounded.\n\n---\n\n## 73. The Geopolitics of AI Safety: A Causal Analysis of Regional LLM Bias\n- **Authors**: Alif Al Hasan\n- **URL**: [https://arxiv.org/abs/2605.05427](https://arxiv.org/abs/2605.05427)\n- **Abstract**:\n> As Large Language Models (LLMs) are integrated into global software systems, ensuring equitable safety guardrails is a critical requirement. Current fairness evaluations predominantly measure bias observationally, a methodology confounded by the inherent toxicity of topics naturally paired with specific demographics in testing datasets. This study introduces a Probabilistic Graphical Model (PGM) framework to audit LLM safety mechanisms causally. By applying Pearl's do-operator, we mathematically isolate the causal effect of injecting a cultural demographic into a prompt. We conduct a large-scale empirical analysis across seven instruction-tuned models spanning diverse origins: the United States (Llama-3.1-8B, Gemma-2-9B), Europe (Mistral-7B-v0.3), the UAE (Falcon3-7B), China (Qwen2.5-7B, DeepSeek-7B), and India (Airavata-7B). Utilizing two distinct datasets (ToxiGen and BOLD), the findings reveal a disparity between observational and interventional bias, demonstrating that standard fairness metrics can overestimate demographic bias by failing to account for context toxicity. Furthermore, the causal probabilities indicate distinct alignment trends: Western models exhibit higher causal refusal rates for specific demographic groups, whereas Eastern models demonstrate low overall intervention rates with targeted sensitivities toward regional demographics. We discuss the implications of these biases, highlighting how demographic-sensitive over-triggering restricts benign discourse in downstream applications.\n\n---\n\n## 74. From History to State: Constant-Context Skill Learning for LLM Agents\n- **Authors**: Haoyang Xie , Xinyuan Wang , Yancheng Wang , Puda Zhao , Feng Ju\n- **URL**: [https://arxiv.org/abs/2605.05413](https://arxiv.org/abs/2605.05413)\n- **Abstract**:\n> Large language model (LLM) agents are increasingly used to operate browsers, files, code and tools, making personal assistants a natural deployment target. Yet personal agents face a privacy-cost-capability tension: cloud models execute multi-step workflows well but expose sensitive intermediate context to external APIs, while local models preserve privacy but remain less reliable. Both settings also pay repeatedly for long skill prompts and growing histories. We propose constant-context skill learning, a context-to-weights framework for recurring agent workflows: reusable procedures are learned in lightweight task-family modules, while inference conditions only on the current observation and a compact state block. A deterministic tracker renders this state block from task progress and supplies aligned subgoal rewards, so each module can be trained with step-level SFT and refined through online RL. Across ALFWorld, WebShop, and SciWorld, our agents achieve strong performance across Qwen3-4B, Qwen3-8B and Llama-3.1-8B. With Qwen3-8B, SFT+RL reaches 89.6\\% unseen success on ALFWorld, 76.8\\% success on WebShop, and 66.4\\% unseen success on SciWorld. They match or exceed strong published agent-training results while reducing prompt tokens per turn by 2--7$\\times$ relative to controlled ReAct prompting baselines, showing that procedural context can be moved from prompts into weights.\n\n---\n\n## 75. LaTA: A Drop-in, FERPA-Compliant Local-LLM Autograder for Upper-Division STEM Coursework\n- **Authors**: Jesse A. Rodríguez\n- **URL**: [https://arxiv.org/abs/2605.05410](https://arxiv.org/abs/2605.05410)\n- **Abstract**:\n> Large-language-model (LLM) graders promise to relieve the grading burden of upper-division STEM courses, but most deployments to date send student work to third-party APIs, violating FERPA and exposing institutions to data risk while requiring substantial assignment modification. We present $\\textbf{LaTA}\\ (\\textit{LaTeX Teaching Assistant})$, a drop-in, open-source autograder that runs entirely on commodity on-premises hardware and assumes a LaTeX-native workflow already adopted by many engineering and physics courses. LaTA implements a four-stage pipeline (ingest, segment, grade, report) using a locally hosted open-weight chain-of-thought LLM grader (gpt-oss:120b) that compares student work to an instructor-authored reference solution and applies a YAML rubric with binary per-item scoring. We deployed LaTA in Winter~2026 in ME 373 (Mechanical Engineering Methods) at Oregon State University, grading every weekly assignment for approximately 200 students on a single Mac Studio at \\$0 marginal cost per assignment and 1--3 minutes of wall-clock time per submission, enabling regrading of corrected assignments and greatly expanded TA office hour offerings. The instructor-confirmed grading-error rate held at roughly $0.02$--$0.04\\%$ per rubric line item across the term. Relative to the same instructor's previous traditionally-graded cohort, the LaTA-graded cohort outperformed by approximately $11\\%$ on the midterm exam and $8\\%$ on the final exam, and reported large gains in self-assessed confidence on every stated learning objective ($N = 159$ survey responses, $\\Delta \\geq +1.49$ Likert points, $p < 10^{-27}$ on every comparison). We release the code under AGPLv3.\n\n---\n\n## 76. Agentic Retrieval-Augmented Generation for Financial Document Question Answering\n- **Authors**: Yang Shu , Yingmin Liu , Zequn Xie\n- **URL**: [https://arxiv.org/abs/2605.05409](https://arxiv.org/abs/2605.05409)\n- **Abstract**:\n> Financial document question answering (QA) demands complex multi-step numerical reasoning over heterogeneous evidence--structured tables, textual narratives, and footnotes--scattered across corporate filings. Existing retrieval-augmented generation (RAG) approaches adopt a single-pass retrieve-then-generate paradigm that struggles with the compositional reasoning chains prevalent in financial analysis. We propose FinAgent-RAG, an agentic RAG framework that orchestrates iterative retrieval-reasoning loops with self-verification, specifically engineered for the precision requirements of financial numerical reasoning. The framework integrates three domain-specific innovations: (1) a Contrastive Financial Retriever trained with hard negative mining to distinguish semantically similar but numerically distinct financial passages, (2) a Program-of-Thought reasoning module that generates executable Python code for precise arithmetic rather than relying on error-prone LLM-based mental computation, and (3) an Adaptive Strategy Router that dynamically allocates computational resources based on question complexity, reducing API costs by 41.3% on FinQA while preserving accuracy. Extensive experiments on three benchmark datasets--FinQA, ConvFinQA, and TAT-QA--demonstrate that FinAgent-RAG achieves 76.81%, 78.46%, and 74.96% execution accuracy respectively, outperforming the strongest baseline by 5.62--9.32 percentage points. Ablation studies, cross-backbone evaluation with four LLMs, and deployment cost analysis confirm the framework's robustness and practical viability for financial institutions.\n\n---\n\n## 77. PRISM: Perception Reasoning Interleaved for Sequential Decision Making\n- **Authors**: Mohamed Salim Aissi , Clemence Grislain , Clement Romac , Laure Soulier , Mohamed Chetouani , Olivier Sigaud , Nicolas Thome\n- **URL**: [https://arxiv.org/abs/2605.05407](https://arxiv.org/abs/2605.05407)\n- **Abstract**:\n> Scaling LLM-based embodied agents from text-only environments to complex multimodal settings remains a major challenge. Recent work identifies a perception-reasoning-decision gap in standalone Vision-Language Models (VLMs), which often overlook task-critical information. In this paper, we introduce PRISM, a framework that tightly couples perception (VLM) and decision (LLM) through a dynamic question-answer (DQA) pipeline. Instead of passively accepting the VLM's description, the LLM critiques it, probes the VLM with goal-oriented questions, and synthesizes a compact image description. This closed-loop interaction yields a sharp, task-driven understanding of the scene. We evaluate PRISM on the ALFWorld and Room-to-Room (R2R) benchmarks. We show that: (1) PRISM significantly outperforms state-of-the-art image-based models, (2) our Interactive goal-oriented perception pipeline yields systematic and substantial gains, and (3) PRISM is fully automatic, eliminating the need for handcrafted questions or answers.\n\n---\n\n## 78. When Helpfulness Becomes Sycophancy: Sycophancy is a Boundary Failure Between Social Alignment and Epistemic Integrity in Large Language Models\n- **Authors**: Jiechen Li , Catherine A. Barry , Rishika Randev , Janet Chen , Ella Jorgensen , Brinnae Bent\n- **URL**: [https://arxiv.org/abs/2605.05403](https://arxiv.org/abs/2605.05403)\n- **Abstract**:\n> This position paper argues that sycophancy in LLMs is a boundary failure between social alignment and epistemic integrity. Existing work often operationalizes sycophancy through external behavior such as agreement with incorrect user beliefs, position reversals, or deviation from an objective standard of correctness. These formulations capture only overt forms of the phenomenon and leave subtler boundary failures involving epistemic integrity and social alignment underspecified. We argue that sycophancy should not be understood as agreement alone, but as alignment behavior that displaces independent epistemic judgment. To clarify this boundary, we propose a three-condition framework for sycophancy. First, the user expresses a cue in the form of a belief, preference, or self-concept. Second, the model shifts toward that cue through alignment behavior. Third, this shift compromises epistemic accuracy, independent reasoning, or appropriate correction. We also introduce a taxonomy for classifying sycophancy, consisting of alignment targets, mechanisms, and severity. The paper concludes by discussing implications for alignment evaluation and argues for boundary-aware assessment, structured rubrics, and mitigation strategies, while situating these proposals alongside alternative views of sycophancy.\n\n---\n\n## 79. BALAR : A Bayesian Agentic Loop for Active Reasoning\n- **Authors**: Aymen Echarghaoui , Dongxia Wu , Emily B. Fox\n- **URL**: [https://arxiv.org/abs/2605.05386](https://arxiv.org/abs/2605.05386)\n- **Abstract**:\n> Large language models increasingly operate in interactive settings where solving a task requires multiple rounds of information exchange with a user. However, most current systems treat dialogue reactively and lack a principled mechanism to reason about what information is missing and which question should be asked next. We propose BALAR (Bayesian Agentic Loop for Active Reasoning), a task-agnostic outer-loop algorithm that requires no fine-tuning and enables structured multi-turn interaction between an LLM agent and a user. BALAR maintains a structured belief over latent states, selects clarifying questions by maximizing expected mutual information, and dynamically expands its state representation when the current one proves insufficient. We evaluate BALAR on three diverse benchmarks: AR-Bench-DC (detective cases), AR-Bench-SP (thinking puzzles), and iCraft-MD (clinical diagnosis). BALAR significantly outperforms all baselines across all three benchmarks, with $14.6\\%$ higher accuracy on AR-Bench-DC, $38.5\\%$ on AR-Bench-SP, and $30.5\\%$ on iCraft-MD.\n\n---\n\n## 80. Understanding Annotator Safety Policy with Interpretability\n- **Authors**: Alex Oesterling , Donghao Ren , Yannick Assogba , Dominik Moritz , Sunnie S.Y. Kim , Leon Gatys , Fred Hohman\n- **URL**: [https://arxiv.org/abs/2605.05329](https://arxiv.org/abs/2605.05329)\n- **Abstract**:\n> Safety policies define what constitutes safe and unsafe AI outputs, guiding data annotation and model development. However, annotation disagreement is pervasive and can stem from multiple sources such as operational failures (annotators misunderstand or misexecute the task), policy ambiguity (policy wording leaves room for interpretation), or value pluralism (different annotators hold different perspectives on safety). Distinguishing these sources matters. For example, operational failures call for quality control, ambiguity calls for policy clarification, and pluralism calls for deliberation about incorporating diverse perspectives. Yet understanding why annotators disagree is difficult. Directly asking annotators for their reasoning is costly, substantially increasing annotation burden, and can be unreliable for both human and LLM annotators as self-reported reasoning often fails to reflect actual decision processes. We introduce Annotator Policy Models (APMs), interpretable models that learn annotators' internal safety policies from labeling behavior alone, making annotator reasoning visible and comparable without additional annotation effort. We validate that APMs accurately model annotator safety policy (>80% accuracy), faithfully predict responses to counterfactual edits, and recover known policy differences in controlled settings. Applying APMs to LLM and human annotations, we demonstrate two core applications: (1) surfacing policy ambiguity by revealing how annotators interpret safety instructions differently, and (2) surfacing value pluralism by uncovering systematic differences in safety priorities across demographic groups. Together, these capabilities support more targeted, transparent, and inclusive safety policy design.\n\n---\n\n## 81. Verifier-Backed Hard Problem Generation for Mathematical Reasoning\n- **Authors**: Yuhang Lai , Jiazhan Feng , Yee Whye Teh , Ning Miao\n- **URL**: [https://arxiv.org/abs/2605.06660](https://arxiv.org/abs/2605.06660)\n- **Abstract**:\n> Large Language Models (LLMs) demonstrate strong capabilities for solving scientific and mathematical problems, yet they struggle to produce valid, challenging, and novel problems - an essential component for advancing LLM training and enabling autonomous scientific research. Existing problem generation approaches either depend on expensive human expert involvement or adopt naive self-play paradigms, which frequently yield invalid problems due to reward hacking. This work introduces VHG, a verifier-enhanced hard problem generation framework built upon three-party self-play. By integrating an independent verifier into the conventional setter-solver duality, our design constrains the setter's reward to be jointly determined by problem validity (evaluated by the verifier) and difficulty (assessed by the solver). We instantiate two verifier variants: a Hard symbolic verifier and a Soft LLM-based verifier, with evaluations conducted on indefinite integral tasks and general mathematical reasoning tasks. Experimental results show that VHG substantially outperforms all baseline methods by a clear margin.\n\n---\n\n## 82. Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less\n- **Authors**: Yuxing Liu , Jianyu Wang , Tong Zhang\n- **URL**: [https://arxiv.org/abs/2605.06654](https://arxiv.org/abs/2605.06654)\n- **Abstract**:\n> Optimizers play an important role in both pretraining and finetuning stages when training large language models (LLMs). In this paper, we present an observation that full finetuning with the same optimizer as in pretraining achieves a better learning-forgetting tradeoff, i.e., forgetting less while achieving the same or better performance on the new task, than other optimizers and, possibly surprisingly, LoRA, during the supervised finetuning (SFT) stage. We term this phenomenon optimizer-model consistency. To better understand it, through controlled experiments and theoretical analysis, we show that: 1) optimizers can shape the models by having regularization effects on the activations, leading to different landscapes around the pretrained checkpoints; 2) in response to this regularization effect, the weight update in SFT should follow some specific structures to lower forgetting of the knowledge learned in pretraining, which can be obtained by using the same optimizer. Moreover, we specifically compare Muon and AdamW when they are employed throughout the pretraining and SFT stages and find that Muon performs worse when finetuned for reasoning tasks. With a synthetic language modeling experiment, we demonstrate that this can come from Muon's strong tendency towards rote memorization, which may hurt pattern acquisition with a small amount of data, as for SFT.\n\n---\n\n## 83. When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels\n- **Authors**: Sushant Gautam , Finn Schwall , Annika Willoch Olstad , Fernando Vallecillos Ruiz , Birk Torpmann-Hagen , Sunniva Maria Stordal Bjørklund , Leon Moonen , Klas Pettersen , Michael A. Riegler\n- **URL**: [https://arxiv.org/abs/2605.06652](https://arxiv.org/abs/2605.06652)\n- **Abstract**:\n> Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component ($\\eta^2 \\approx 0.52$), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.\n\n---\n\n## 84. Superintelligent Retrieval Agent: The Next Frontier of Information Retrieval\n- **Authors**: Zeyu Yang , Qi Ma , Jason Chen , Anshumali Shrivastava\n- **URL**: [https://arxiv.org/abs/2605.06647](https://arxiv.org/abs/2605.06647)\n- **Abstract**:\n> Retrieval-augmented agents are increasingly the interface to large organizational knowledge bases, yet most still treat retrieval as a black box: they issue exploratory queries, inspect returned snippets, and iteratively reformulate until useful evidence emerges. This approach resembles how a newcomer searches an unfamiliar database rather than how an expert navigates it with strong priors about terminology and likely evidence, and results in unnecessary retrieval rounds, increased latency, and poor recall. We introduce \\textit{SuperIntelligent Retrieval Agent} (SIRA), which defines \\emph{superintelligence} in retrieval as the ability to compress multi-round exploratory search into a single corpus-discriminative retrieval action. SIRA does not merely ask what terms are relevant to the query; it asks which terms are likely to separate the desired evidence from corpus-level confusers. On the corpus side, an LLM enriches each document offline with missing search vocabulary; on the query side, it predicts evidence vocabulary omitted by the query; and document-frequency statistics as a tool call to filter proposed terms that are absent, overly common, or unlikely to create retrieval margin. The final retrieval step is a single weighted BM25 call combining the original query with the validated expansion. Across ten BEIR benchmarks and downstream question-answering tasks, SIRA achieves the significantly superior performance outperforming dense retrievers and state-of-the-art multi-round agentic baselines, demonstrating that one well-formed lexical query, guided by LLM cognition and lightweight corpus statistics, can exceed substantially more expensive multi-round search while remaining interpretable, training-free, and efficient.\n\n---\n\n## 85. StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction\n- **Authors**: Xiangyuan Xue , Yifan Zhou , Zidong Wang , Shengji Tang , Philip Torr , Wanli Ouyang , Lei Bai , Zhenfei Yin\n- **URL**: [https://arxiv.org/abs/2605.06642](https://arxiv.org/abs/2605.06642)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used as interactive agents, but optimizing them for long-horizon decision making remains difficult because current methods are largely purely reactive, which weakens both exploration and credit assignment over extended trajectories. In this work, we present Strategic Trajectory Abstraction (StraTA), a simple framework that introduces an explicit trajectory-level strategy into agentic reinforcement learning (RL). StraTA samples a compact strategy from the initial task state, conditions subsequent actions on that strategy, and trains strategy generation and action execution jointly with a hierarchical GRPO-style rollout design, further enhanced by diverse strategy rollout and critical self-judgment. Experiments on ALFWorld, WebShop, and SciWorld show that StraTA consistently improves both sample efficiency and final performance over strong baselines. StraTA reaches success rates of 93.1% on ALFWorld and 84.2% on WebShop. On SciWorld, StraTA attains a 63.5% overall score, outperforming frontier closed-source models.\n\n---\n\n## 86. The Structural Origin of Attention Sink: Variance Discrepancy, Super Neurons, and Dimension Disparity\n- **Authors**: Siquan Li , Kaiqi Jiang , Jiacheng Sun , Tianyang Hu\n- **URL**: [https://arxiv.org/abs/2605.06611](https://arxiv.org/abs/2605.06611)\n- **Abstract**:\n> Despite the prevalence of the attention sink phenomenon in Large Language Models (LLMs), where initial tokens disproportionately monopolize attention scores, its structural origins remain elusive. This work provides a \\textit{mechanistic explanation} for this phenomenon. First, we trace its root to the value aggregation process inherent in self-attention, which induces a systematic variance discrepancy. We further demonstrate that this discrepancy is drastically amplified by the activation of super neurons within Feed-Forward Network (FFN) layers. Specifically, the channel-sparse down-projections trigger a dimension disparity of the first-token representation, necessitating the formation of attention sinks as a structural anchor. Then, we validate this causal chain through two controlled interventions: (i) isolating the aggregation effect via attention mask modifications and (ii) amplifying the variance of targeted token representations. Both interventions can replicate attention sinks at arbitrary positions. Our mechanistic understanding offers a foundation for the systematic control of sink formation. Finally, as a proof of concept, we propose \\textit{head-wise RMSNorm}, an architectural modification that stabilizes value aggregation outputs during pre-training. Our experiments demonstrate that restoring statistical parity across positions significantly accelerates convergence.\n\n---\n\n## 87. AI CFD Scientist: Toward Open-Ended Computational Fluid Dynamics Discovery with Physics-Aware AI Agents\n- **Authors**: Nithin Somasekharan , Rabi Pathak , Manushri Dhanakoti , Tingwen Zhang , Ling Yue , Andy Zhu , Shaowu Pan\n- **URL**: [https://arxiv.org/abs/2605.06607](https://arxiv.org/abs/2605.06607)\n- **Abstract**:\n> Recent LLM-based agents have closed substantial portions of the scientific discovery loop in software-only machine-learning research, in chemistry, and in biology. Extending the same loop to high-fidelity physical simulators is harder, because solver completion does not imply physical validity and many failure modes appear only in field-level imagery rather than in solver logs. We present AI CFD Scientist, an open-source AI scientist for computational fluid dynamics (CFD) that, to our knowledge, is the first to span literature-grounded ideation, validated execution, vision-based physics verification, source-code modification, and figure-grounded writing within a single inspectable workflow. Three coupled pathways cover parameter sweeps within a fixed solver, case-local C++ library compilation for new physical models, and open-ended hypothesis search against a reference comparator, all running on OpenFOAM through Foam-Agent. At the center of the framework is a vision-language physics-verification gate that inspects rendered flow fields before any result is accepted, rerun, or written into a manuscript. On five tasks under a shared GPT-5.5 backbone, AI CFD Scientist autonomously discovers a Spalart-Allmaras runtime correction that reduces lower-wall Cf RMSE against DNS by 7.89% on the periodic hill at Reh=5600; under matched LLM cost, two strong general AI-scientist baselines (ARIS, DeepScientist) execute partial CFD workflows but lack the domain-specific validity gates needed to convert runs into defensible scientific claims; and a controlled planted-failure ablation shows that the vision-language gate detects 14 of 16 silent failures missed by solver-level checks. Code, prompts, and run artifacts are released at this https URL .\n\n---\n\n## 88. UniSD: Towards a Unified Self-Distillation Framework for Large Language Models\n- **Authors**: Yiqiao Jin , Yiyang Wang , Lucheng Fu , Yijia Xiao , Yinyi Luo , Haoxin Liu , B. Aditya Prakash , Josiah Hester , Jindong Wang , Srijan Kumar\n- **URL**: [https://arxiv.org/abs/2605.06597](https://arxiv.org/abs/2605.06597)\n- **Abstract**:\n> Self-distillation (SD) offers a promising path for adapting large language models (LLMs) without relying on stronger external teachers. However, SD in autoregressive LLMs remains challenging because self-generated trajectories are free-form, correctness is task-dependent, and plausible rationales can still provide unstable or unreliable supervision. Existing methods mainly examine isolated design choices, leaving their effectiveness, roles, and interactions unclear. In this paper, we propose UniSD, a unified framework to systematically study self-distillation. UniSD integrates complementary mechanisms that address supervision reliability, representation alignment, and training stability, including multi-teacher agreement, EMA teacher stabilization, token-level contrastive learning, feature matching, and divergence clipping. Across six benchmarks and six models from three model families, UniSD reveals when self-distillation improves over static imitation, which components drive the gains, and how these components interact across tasks. Guided by these insights, we construct UniSDfull, an integrated pipeline that combines complementary components and achieves the strongest overall performance, improving over the base model by +5.4 points and the strongest baseline by +2.8 points. Extensive evaluation highlights self-distillation as a practical and steerable approach for efficient LLM adaptation without stronger external teachers.\n\n---\n\n## 89. Continuous Latent Diffusion Language Model\n- **Authors**: Hongcan Guo , Qinyu Zhao , Yian Zhao , Shen Nie , Rui Zhu , Qiushan Guo , Feng Wang , Tao Yang , Hengshuang Zhao , Guoqiang Wei , Yan Zeng\n- **URL**: [https://arxiv.org/abs/2605.06548](https://arxiv.org/abs/2605.06548)\n- **Abstract**:\n> Large language models have achieved remarkable success under the autoregressive paradigm, yet high-quality text generation need not be tied to a fixed left-to-right order. Existing alternatives still struggle to jointly achieve generation efficiency, scalable representation learning, and effective global semantic modeling. We propose Cola DLM, a hierarchical latent diffusion language model that frames text generation through hierarchical information decomposition. Cola DLM first learns a stable text-to-latent mapping with a Text VAE, then models a global semantic prior in continuous latent space with a block-causal DiT, and finally generates text through conditional decoding. From a unified Markov-path perspective, its diffusion process performs latent prior transport rather than token-level observation recovery, thereby separating global semantic organization from local textual realization. This design yields a more flexible non-autoregressive inductive bias, supports semantic compression and prior fitting in continuous space, and naturally extends to other continuous modalities. Through experiments spanning 4 research questions, 8 benchmarks, strictly matched ~2B-parameter autoregressive and LLaDA baselines, and scaling curves up to about 2000 EFLOPs, we identify an effective overall configuration of Cola DLM and verify its strong scaling behavior for text generation. Taken together, the results establish hierarchical continuous latent prior modeling as a principled alternative to strictly token-level language modeling, where generation quality and scaling behavior may better reflect model capability than likelihood, while also suggesting a concrete path toward unified modeling across discrete text and continuous modalities.\n\n---\n\n## 90. Is One Layer Enough? Understanding Inference Dynamics in Tabular Foundation Models\n- **Authors**: Amir Rezaei Balef , Mykhailo Koshil , Katharina Eggensperger\n- **URL**: [https://arxiv.org/abs/2605.06510](https://arxiv.org/abs/2605.06510)\n- **Abstract**:\n> Transformer-based tabular foundation models (TFMs) dominate small to medium tabular predictive benchmark tasks, yet their inference mechanisms remain largely unexplored. We present the first large-scale mechanistic study of layerwise dynamics in 6 state-of-the-art tabular in-context learning models. We explore how predictions emerge across depth, identify distinct stages of inference and reveal latent-space dynamics that differ from those of language models. Our findings indicate substantial depthwise redundancy across multiple models, suggesting iterative refinement with overlapping computations during inference stages. Guided by these insights, we design a proof-of-concept, looped single-layer model that uses only 20% of the original model's parameters while achieving comparable performance. The code is available at this https URL .\n\n---\n\n## 91. PACZero: PAC-Private Fine-Tuning of Language Models via Sign Quantization\n- **Authors**: Murat Bilgehan Ertan , Xiaochen Zhu , Phuong Ha Nguyen , Marten van Dijk , Srinivas Devadas\n- **URL**: [https://arxiv.org/abs/2605.06505](https://arxiv.org/abs/2605.06505)\n- **Abstract**:\n> We introduce PACZero, a family of PAC-private zeroth-order mechanisms for fine-tuning large language models that delivers usable utility at $I(S^*; Y_{1:T})=0$. This privacy regime bounds the membership-inference attack (MIA) posterior success rate at the prior, an MIA-resistance level the DP framework matches only at $\\varepsilon=0$ and infinite noise. All DP-ZO comparisons below are matched at the MIA posterior level. The key insight is that PAC Privacy charges mutual information only when the release depends on which candidate subset is the secret. Sign-quantizing subset-aggregated zeroth-order gradients creates frequent unanimity, steps at which every candidate subset agrees on the update direction; at these steps the released sign costs zero conditional mutual information. We propose two variants that span the privacy-utility trade-off: PACZero-MI (budgeted MI via exact calibration on the binary release) and PACZero-ZPL ($I=0$ via a uniform coin flip on disagreement steps). We evaluate on SST-2 and SQuAD with OPT-1.3B and OPT-6.7B in both LoRA and full-parameter tracks. On SST-2 OPT-1.3B full fine-tuning at $I=0$, PACZero-ZPL reaches ${88.99\\pm0.91}$, within $2.1$pp of the non-private MeZO baseline ($91.1$ FT). No prior method produces usable utility in the high-privacy regime $\\varepsilon<1$, and PACZero-ZPL obtains competitive SST-2 accuracy and nontrivial SQuAD F1 across OPT-1.3B and OPT-6.7B at $I=0$.\n\n---\n\n## 92. Litespark Inference on Consumer CPUs: Custom SIMD Kernels for Ternary Neural Networks\n- **Authors**: Nii Osae Osae Dade , Tony Morri , Moinul Hossain Rahat , Sayandip Pal\n- **URL**: [https://arxiv.org/abs/2605.06485](https://arxiv.org/abs/2605.06485)\n- **Abstract**:\n> Large language models (LLMs) have transformed artificial intelligence, but their computational requirements remain prohibitive for most users. Standard inference demands expensive datacenter GPUs or cloud API access, leaving over one billion personal computers underutilized for AI workloads. Ternary models offer a path forward: their weights are constrained to {-1, 0, +1}, theoretically eliminating the need for floating-point multiplication. However, existing frameworks fail to exploit this structure, treating ternary models as dense floating-point networks. We address this gap with custom SIMD kernels that replace matrix multiplication with simple addition and subtraction operations, targeting the integer dot product instructions available on modern CPUs. Our implementation, Litespark-Inference, is pip-installable and integrates directly with Hugging-Face, achieving 9.2x faster time-to-first-token, 52x higher throughput, and 14x memory reduction compared to standard PyTorch inference on Apple Silicon, with similar speedups on Intel and AMD processors.\n\n---\n\n## 93. Constraint Decay: The Fragility of LLM Agents in Backend Code Generation\n- **Authors**: Francesco Dente , Dario Satriani , Paolo Papotti\n- **URL**: [https://arxiv.org/abs/2605.06445](https://arxiv.org/abs/2605.06445)\n- **Abstract**:\n> Large Language Model (LLM) agents demonstrate strong performance in autonomous code generation under loose specifications. However, production-grade software requires strict adherence to structural constraints, such as architectural patterns, databases, and object-relational mappings. Existing benchmarks often overlook these non-functional requirements, rewarding functionally correct but structurally arbitrary solutions. We present a systematic study evaluating how well agents handle structural constraints in multi-file backend generation. By fixing a unified API contract across 80 greenfield generation tasks and 20 feature-implementation tasks spanning eight web frameworks, we isolate the effect of structural complexity using a dual evaluation with end-to-end behavioral tests and static verifiers. Our findings reveal a phenomenon of constraint decay: as structural requirements accumulate, agent performance exhibits a substantial decline. Capable configurations lose 30 points on average in assertion pass rates from baseline to fully specified tasks, while some weaker configurations approach zero. Framework sensitivity analysis exposes significant performance disparities: agents succeed in minimal, explicit frameworks (e.g., Flask) but perform substantially worse on average in convention-heavy environments (e.g., FastAPI, Django). Finally, error analysis identifies data-layer defects (e.g., incorrect query composition and ORM runtime violations) as the leading root causes. This work highlights that jointly satisfying functional and structural requirements remains a key open challenge for coding agents.\n\n---\n\n## 94. Is Escalation Worth It? A Decision-Theoretic Characterization of LLM Cascades\n- **Authors**: Dylan Bouchard\n- **URL**: [https://arxiv.org/abs/2605.06350](https://arxiv.org/abs/2605.06350)\n- **Abstract**:\n> Model cascades, in which a cheap LLM defers to an expensive one on low-confidence queries, are widely used to navigate the cost-quality tradeoff at deployment. Existing approaches largely treat the deferral threshold as an empirical hyperparameter, with limited guidance on the geometry of the resulting cost-quality frontier over a model pool. We develop a decision-theoretic framework grounded in constrained optimization and duality. For a two-model cascade, we establish piecewise concavity of the cost-quality frontier on decreasing-benefit regions of the confidence support, with reciprocal shadow prices linking the budget- and quality-constrained formulations. Given a pool of $k$ models, we characterize the frontier achievable by deterministic two-model threshold cascades as the pointwise envelope over $\\binom{k}{2}$ pairwise cascades, with switching points where the optimal pair changes. For $k$-model cascades, we derive first-order conditions in which a single shadow price equalizes marginal quality-per-cost across stage boundaries. We validate the framework on five benchmarks (MATH, MMLU, TriviaQA, SimpleQA, LiveCodeBench) across eight models from five providers. Within the deterministic threshold-cascade class, full fixed chains underperform the pairwise envelope, and optimized subsequence cascades do not deliver practically meaningful held-out gains over it. A lightweight pre-generation router exceeds the best cascade policy on four of five datasets, mainly because it avoids the cheap model's generation cost on queries sent directly to a larger model rather than because of a stronger routing signal. These results suggest that cascade performance is limited primarily by structural cost, since cascades pay the cheap model before any escalation decision, rather than by a shortage of intermediate stages.\n\n---\n\n## 95. Human-AI Co-Evolution and Epistemic Collapse: A Dynamical Systems Perspective\n- **Authors**: Xuening Wu , Yanlan Kang , Qianya Xu , Kexuan Xie , Jiaqi Mi , Honggang Wang , Yubin Liu , Zeping Chen\n- **URL**: [https://arxiv.org/abs/2605.06347](https://arxiv.org/abs/2605.06347)\n- **Abstract**:\n> Large language models (LLMs) are reshaping how knowledge is produced, with increasing reliance on AI systems for generation, summarization, and reasoning. While prior work has studied cognitive offloading in humans and model collapse in recursive training, these effects are typically considered in isolation. We propose a unified perspective: humans and language models form a coupled dynamical system linked by a feedback loop of usage, generation, and retraining. We introduce a minimal model with three variables -- human cognition, data quality, and model capability -- and show that this feedback can give rise to distinct dynamical regimes. Our analysis identifies three regimes: co-evolutionary enhancement, fragile equilibrium, and degenerative convergence. Through a simple simulation, we demonstrate that increasing reliance on AI can induce a transition toward a low-diversity, suboptimal equilibrium. From an information-theoretic perspective, this transition corresponds to an emergent information bottleneck in the human-AI loop, where entropy reduction reflects loss of diversity and support under closed-loop feedback rather than beneficial compression. These results suggest that the trajectory of AI systems is shaped not only by model design, but by the dynamics of human-AI co-evolution.\n\n---\n\n## 96. CoupleEvo: Evolving Heuristics for Coupled Optimization Problems Using Large Language Models\n- **Authors**: Thomas Bömer , Bastian Amberg , Max Disselnmeyer , Anne Meyer\n- **URL**: [https://arxiv.org/abs/2605.06341](https://arxiv.org/abs/2605.06341)\n- **Abstract**:\n> Many real-world optimization problems consist of multiple tightly coupled subproblems whose solutions must be coordinated to achieve high overall performance. However, existing large language model driven automated heuristic design approaches are limited to single-problem settings. In this paper, we propose CoupleEvo. CoupleEvo proposes three evolutionary coordination strategies to evolve heuristics for coupled optimization problems: the sequential strategy evolves heuristics for one subproblem after the other; the iterative strategy alternates the evolution of heuristics for different subproblems over successive generations; and the integrated strategy evolves heuristics for all problems simultaneously. The approach is evaluated on two representative coupled optimization problems. Experimental results show that decomposition-based strategies (sequential and iterative) provide more stable convergence and higher solution quality, while the integrated evolution strategy suffers from increased search complexity and variability. These findings highlight the importance of coordinating evolutionary search across interdependent subproblems and demonstrate the potential of LLM-driven heuristic design for complex coupled optimization problems. The code is available: this https URL .\n\n---\n\n## 97. Fine-Tuning Small Language Models for Solution-Oriented Windows Event Log Analysis\n- **Authors**: Siraaj Akhtar , Saad Khan , Simon Parkinson\n- **URL**: [https://arxiv.org/abs/2605.06330](https://arxiv.org/abs/2605.06330)\n- **Abstract**:\n> Large language models (LLMs) have shown promise for event log analysis, but their high computational requirements, reliance on cloud infrastructure, and security concerns limit practical deployment. In addition, most existing approaches focus only on the identification of the problem and do not provide actionable remediation. Small language models (SLMs) present a light-weight alternative that can be fine-tuned for a specific purpose and hosted locally. This paper investigates whether SLMs, when fine-tuned for a specific task, can serve as a practical alternative for event log analysis while also generating solutions. We first create a large-scale synthetic Windows event log dataset that contains remediation actions using a high-performing LLM. We then fine-tune multiple SLMs and LLMs using the LoRA parameter-efficient fine-tuning technique and evaluate their performance by comparing with expert assessment. The results show that the dataset accurately reflects real-world scenarios and that fine-tuned SLMs consistently outperform LLMs in identifying issues and providing relevant remediation, while requiring fewer computational resources.\n\n---\n\n## 98. Measuring Evaluation-Context Divergence in Open-Weight LLMs: A Paired-Prompt Protocol with Pilot Evidence of Alignment-Pipeline-Specific Heterogeneity\n- **Authors**: Florian A. D. Burnat , Brittany I. Davidson\n- **URL**: [https://arxiv.org/abs/2605.06327](https://arxiv.org/abs/2605.06327)\n- **Abstract**:\n> Safety benchmarks are routinely treated as evidence about how a language model will behave once deployed, but this inference is fragile if behavior depends on whether a prompt looks like an evaluation. We define evaluation-context divergence as an observable within-item change in behavior induced by framing a fixed task as an evaluation, a live deployment interaction, or a neutral request, and present a paired-prompt protocol that measures it in open-weight LLMs while controlling for paraphrase variation, benchmark familiarity, and judge framing-sensitivity. Across five instruction-tuned checkpoints from four open-weight families plus a matched OLMo-3 base/instruct ablation ($20$ paired items, $840$ generations per checkpoint), we find striking heterogeneity. OLMo-3-Instruct alone is eval-cautious -- evaluation framing raises refusal vs. neutral by $11.8$pp ($p=0.007$) and reduces harmful compliance vs. deployment by $3.6$pp ($p=0.024$, $0/20$ items inverted) -- while Mistral-Small-3.2, Phi-3.5-mini, and Llama-3.1-8B are deployment-cautious}, with marginal eval-vs-deployment refusal effects of $-9$ to $-20$pp. The matched OLMo-3 base also exhibits the deployment-cautious pattern, identifying alignment as the inversion stage; within Llama-3.1, the $70$B model preserves direction with attenuated magnitude, ruling out a simple ``small-model effect that reverses at scale.'' One caveat: the cross-family heterogeneity is judge-dependent. Re-judging with a different-family safety classifier (Llama-Guard-3-8B) preserves the within-OLMo eval-cautious direction but flattens the cross-family contrast, indicating that the two judges operationalize distinct constructs.\n\n---\n\n## 99. Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs\n- **Authors**: Elizabeth Mieczkowski , Alexander Ku , Tiwalayo Eisape , Dilip Arumugam , John Matters , Katherine M. Collins , Ilia Sucholutsky , Thomas L. Griffiths\n- **URL**: [https://arxiv.org/abs/2605.06320](https://arxiv.org/abs/2605.06320)\n- **Abstract**:\n> Large language models (LLMs) are increasingly deployed in teams, yet existing coordination approaches often occupy two extremes. Highly structured methods rely on fixed roles, pipelines, or task decompositions assigned a priori. In contrast, fully unstructured teams enable adaptability and exploration but suffer from inefficiencies such as error propagation, inter-agent conflicts, and wasted resources (measured in time, tokens, or file operations). We introduce Language Agent Teams for Task Evolution (LATTE), a framework for coordinating LLM teams inspired by distributed systems, where processors must operate under partial observability and communication constraints. In LATTE, a team of agents collaboratively construct and maintain a shared, evolving coordination graph which encodes sub-task dependencies, individual agent assignment, and the current state of sub-task progress. This protocol maintains consistency while empowering agents to dynamically allocate work, adapt coordination, and discover new tasks. Across multiple collaborative tasks and a variety of base models, we demonstrate how LATTE reduces token usage, wall-clock time, communication, and coordination failures (e.g. file conflicts and redundant outputs) while matching or exceeding the accuracy of standard designs including MetaGPT, decentralized teams, top-down Leader-Worker hierarchies, and static decompositions.\n\n---\n\n## 100. Pro-KLShampoo: Projected KL-Shampoo with Whitening Recovered by Orthogonalization\n- **Authors**: Ruotong Sun , Ermin Wei\n- **URL**: [https://arxiv.org/abs/2605.06316](https://arxiv.org/abs/2605.06316)\n- **Abstract**:\n> Optimizers that exploit the matrix structure of gradients are central to modern LLM pre-training, with two distinct frontiers: explicit Kronecker-factored preconditioning -- most recently KL-Shampoo, which estimates the preconditioner via KL divergence minimization -- and orthogonalization of the gradient momentum, exemplified by Muon and analyzed as steepest descent under the spectral norm. The two routes are typically developed in isolation. We make a structural observation about KL-Shampoo's Kronecker preconditioners: their eigenvalue spectra exhibit a \\emph{spike-and-flat} shape -- a few dominant eigenvalues followed by an approximately uniform tail -- across layers and training stages, holding exactly under a rank-$\\rho$ signal-plus-noise gradient model. We exploit this structure by restricting one of KL-Shampoo's Kronecker factors to a parametric family aligned with the spike-and-flat shape: full spectral structure on a tracked $r$-dimensional subspace, single shared eigenvalue across the remaining $n-r$ directions. On these directions, we apply orthogonalization. An identity shows that this orthogonalization recovers the algebraic form of full KL-Shampoo's preconditioner. On four pre-training scales (GPT-2 124M / 350M, LLaMA 134M / 450M), Pro-KLShampoo consistently outperforms KL-Shampoo at every subspace rank we test in validation loss, peak per-GPU memory, and wallclock time to reach each loss level.\n\n---\n\n## 101. Attributions All the Way Down? The Metagame of Interpretability\n- **Authors**: Hubert Baniecki , Przemyslaw Biecek , Fabian Fumagalli\n- **URL**: [https://arxiv.org/abs/2605.06295](https://arxiv.org/abs/2605.06295)\n- **Abstract**:\n> We introduce the metagame, a conceptual framework for quantifying second-order interaction effects of model explanations. For any first-order attribution $\\phi(f)$ explaining a model $f$, we measure the directional influence of feature $j$ on the attribution of feature $i$, denoted as meta-attribution $\\varphi_{j \\to i}(f)$, by treating the attribution method itself as a cooperative game and computing its Shapley value. Theoretically, we prove that attributions hierarchically decompose into meta-attributions, and establish these as directional extensions of existing interaction indices. Empirically, we demonstrate that the metagame delivers insights across diverse interpretability applications: (i) quantifying token interactions in instruction-tuned language models, (ii) explaining cross-modal similarity in vision-language encoders, and (iii) interpreting text-to-image concepts in multimodal diffusion transformers.\n\n---\n\n## 102. Log-Likelihood, Simpson's Paradox, and the Detection of Machine-Generated Text\n- **Authors**: Tom Kempton , Viktor Drobnyi , Maeve Madigan , Stuart Burrell\n- **URL**: [https://arxiv.org/abs/2605.06294](https://arxiv.org/abs/2605.06294)\n- **Abstract**:\n> The ability to reliably distinguish human-written text from that generated by large language models is of profound societal importance. The dominant approach to this problem exploits the likelihood hypothesis: that machine-generated text should appear more probable to a detector language model than human-written text. However, we demonstrate that the token-level signal distinguishing human and machine text is non-uniform across the hidden space of the detector model, and naively averaging likelihood-based token scores across regions with fundamentally different statistical structure, as most detectors do, causes a form of Simpson's paradox: a strong local signal is destroyed by inappropriate aggregation. To correct for this, we introduce a learned local calibration step grounded in Bayesian decision theory. Rather than aggregating raw token scores, we first learn lightweight predictors of the score distributions conditioned on position in hidden space, and aggregate calibrated log-likelihood ratios instead. This single intervention dramatically and consistently improves detection performance across all baseline detectors and all datasets we consider. For example, our calibrated variant of Fast-DetectGPT improves AUROC from $0.63$ to $0.85$ on GPT-5.4 text, and a locally-calibrated DMAP detector we introduce achieves state-of-the-art performance across the board. That said, our central contribution is not a new detector, but a precise diagnosis of a significant cause of under-performance of existing detectors and a principled, modular remedy compatible with any token-averaging pipeline. This will serve as a foundation for the community to build upon, with natural avenues including richer distributional models, improved calibration strategies, and principled ensembling with hidden-space geometry signals via the full Bayes-optimal decision rule.\n\n---\n\n## 103. Correct Code, Vulnerable Dependencies: A Large Scale Measurement Study of LLM-Specified Library Versions\n- **Authors**: Chengjie Wang , Jingzheng Wu , Xiang Ling , Tianyue Luo , Chen Zhao\n- **URL**: [https://arxiv.org/abs/2605.06279](https://arxiv.org/abs/2605.06279)\n- **Abstract**:\n> Large language models (LLMs) are now largely involved in software development workflows, and the code they generate routinely includes third-party library (TPL) imports annotated with specific version identifiers. These version choices can carry security and compatibility risks, yet they have not been systematically studied. We present the first large-scale measurement study of version-level risk in LLM-generated Python code, evaluating 10 LLMs on PinTrace, a curated benchmark of 1,000 Stack Overflow programming tasks. LLMs tend to specify version identifiers when directly prompted at 26.83%-95.18%, while down to 6.45%-59.19% in creating a manifest file directly. Among the specified versions, 36.70%-55.70% of tasks contain at least one known CVE, and 62.75%-74.51% of them carry Critical or High severity ratings. In 72.27%-91.37% of cases, the associated CVEs were publicly disclosed before the model's knowledge cutoff. The statistics show all models converge on the same small set of risky release versions, indicating a systemic bias rather than isolated model error. Static compatibility rates range from 19.70% to 63.20%, with installation failure as the dominant cause. The dynamic test cases confirm the pattern by 6.49%-48.62% pass rates. Further experiments confirm that these failures are attributable to version selection rather than code quality, and that externally anchored version constraints substantially reduce both vulnerability exposure and compatibility failures. Our findings reveal LLM version selection as a first-class, previously overlooked risk surface in LLM-based development. We disclosed these findings to the community of the evaluated models, and several confirmed the issue. All the code and dataset have been released for open science at this https URL .\n\n---\n\n## 104. Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs\n- **Authors**: Andy Zeyi Liu , Michael Zhang , Ilana Greenberg , Adam Alnasser , Lucas Baker , John Sous\n- **URL**: [https://arxiv.org/abs/2605.06225](https://arxiv.org/abs/2605.06225)\n- **Abstract**:\n> Steering large language models (LLMs) is usually done by either instruction prompting or activation steering. Prompting often gives strong control, but caches guidance tokens at every layer and can clutter long interactions; activation steering is compact but typically weaker and does not support large structured reminders. We introduce memory inception (MI), a training-free method that steers in latent attention space by inserting text-derived key-value (KV) banks only at selected layers. Rather than materializing reminder content throughout the prompt cache, MI treats steering as selective KV allocation, injecting latent slots only where the model routes to them. On matched personality-steering tasks, MI gives the best overall control--drift trade-off, remaining competitive with prompting while consistently outperforming CAA. On updateable guidance, MI supports mid-conversation behavior shifts without rewriting the visible transcript, achieving the highest post-shift alignment on Qwen3. On structured reasoning, MI outperforms visible prompting on HARDMath and PHYSICS (10/12 subject$\\times$mode cells), serving as proxies for structured reasoning in verifiable domains, while cutting content-matched KV storage by up to 118$\\times$. These results position MI as a powerful steering method when guidance is persistent, structured, or expensive to keep in the visible transcript.\n\n---\n\n## 105. TIDE: Every Layer Knows the Token Beneath the Context\n- **Authors**: Ajay Jaiswal , Lauren Hannah , Han-Byul Kim , Duc Hoang , Mehrdad Farajtabar , Minsik Cho\n- **URL**: [https://arxiv.org/abs/2605.06216](https://arxiv.org/abs/2605.06216)\n- **Abstract**:\n> We revisit a universally accepted but under-examined design choice in every modern LLM: a token index is looked up once at the input embedding layer and then permanently discarded. This single-injection assumption induces two structural failures: (i) the Rare Token Problem, where a Zipf-type distribution of vocabulary causes rare-token embeddings are chronically under-trained due to receiving a fraction of the cumulative gradient signal compared to common tokens; and (ii) the Contextual Collapse Problem, where limited parameters models map distributionally similar tokens to indistinguishable hidden states. As an attempt to address both, we propose TIDE, which augments the standard transformer with EmbeddingMemory: an ensemble of K independent MemoryBlocks that map token indices to context-free semantic vectors, computed once and injected into every layer through a depth-conditioned softmax router with a learnable null bank. We theoretically and empirically establish the benefits of TIDE in addressing the issues associated with single-token identity injection as well as improve performance across multiple language modeling and downstream tasks.\n\n---\n\n## 106. Retina-RAG: Retrieval-Augmented Vision-Language Modeling for Joint Retinal Diagnosis and Clinical Report Generation\n- **Authors**: Abdelrahman Zaian , Sheethal Bhat , Mohamed Abdalkader , Andreas Maier\n- **URL**: [https://arxiv.org/abs/2605.06173](https://arxiv.org/abs/2605.06173)\n- **Abstract**:\n> Diabetic Retinopathy (DR) is a leading cause of preventable blindness among working-age adults worldwide, yet most automated screening systems are limited to image-level classification and lack clinically structured reporting. We propose Retina-RAG, a low-cost modular framework that jointly performs DR severity grading, macular edema (ME) detection, and report generation. The architecture decouples a high-performance retinal classifier and a parameter-efficient vision-language model (Qwen2.5-VL-7B-Instruct) adapted via Low-Rank Adaptation (LoRA), enabling flexible component integration. A retrieval-augmented generation (RAG) module injects curated ophthalmic knowledge together with structured classifier outputs at inference time to improve diagnostic consistency and reduce hallucinations. Retina-RAG achieves an F1-score of 0.731 for DR grading and 0.948 for ME detection, substantially outperforming zero-shot Qwen (0.096, 0.732) and MMed-RAG (0.541, 0.641) on a retinal disease detection dataset with captions. For report generation, Retina-RAG attains ROUGE-L 0.429 and SBERT similarity 0.884, exceeding all baselines. The full framework operates on a single consumer-grade GPU, demonstrating that clinically structured retinal AI can be achieved with modest computational resources.\n\n---\n\n## 107. IRC-Bench: Recognizing Entities from Contextual Cues in First-Person Reminiscences\n- **Authors**: Yehudit Aperstein , Eden Moran , Alexander Apartsin\n- **URL**: [https://arxiv.org/abs/2605.06142](https://arxiv.org/abs/2605.06142)\n- **Abstract**:\n> When people recount personal memories, they often refer to people, places, and events indirectly, relying on contextual cues rather than explicit names. Such implicit references are central to reminiscence narratives: first-person accounts of lived experience used in therapeutic, archival, and social settings. They pose a difficult computational problem because the intended entity must be inferred from dispersed narrative evidence rather than from a local mention. We introduce IRC-Bench, the Implicit Reminiscence Context Benchmark, for evaluating implicit entity recognition in reminiscence transcripts. The benchmark targets non-locality: entity-identifying cues are distributed across multiple, non-contiguous clauses, unlike named entity recognition, entity linking, or coreference resolution. IRC-Bench comprises 25,136 samples constructed from 12,337 Wiki-data-linked entities across 1,994 transcripts spanning 11 thematic domains. Each sample pairs an Entity-Grounded Narrative, in which the target entity is explicitly mentioned, with an Entity-Elided Narrative, in which direct mentions are removed. We evaluate 19 configurations across LLM generation, dense retrieval, RAG, and fine-tuning. QLoRA-adapted Llama 3.1 8B performs best in the open-world setting (38.94% exact match; 51.59% Jaccard), while fine-tuned DPR leads closed-world retrieval (35.38% Hit@1; 71.49% Hit@10). We release IRC-Bench with data, code, and evaluation tools.\n\n---\n\n## 108. Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex\n- **Authors**: Yun Qu , Qi Wang , Yixiu Mao , Heming Zou , Yuhang Jiang , Yingyue Li , Wutong Xu , Lizhou Cai , Weijie Liu , Clive Bai , Kai Yang , Yangkun Chen , Saiyong Yang , Xiangyang Ji\n- **URL**: [https://arxiv.org/abs/2605.06139](https://arxiv.org/abs/2605.06139)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has become a standard approach for large language models (LLMs) post-training to incentivize reasoning capacity. Among existing recipes, group-based policy gradient is prevalent, which samples a group of responses per prompt and updates the policy via group-relative advantage signals. This work reveals that these optimization strategies share a common geometric structure: each implicitly defines a target distribution on the response simplex and projects toward it via first-order approximation. Building on this insight, we propose Listwise Policy Optimization (LPO) to explicitly conduct the target-projection, which demystifies the implicit target by restricting the proximal RL objective to the response simplex, and then projects the policy via exact divergence minimization. This framework provides (i) monotonic improvement on the listwise objective with bounded, zero-sum, and self-correcting projection gradients, and (ii) flexibility in divergence selection with distinct structural properties through the decoupled projection step. On diverse reasoning tasks and LLM backbones, LPO consistently improves training performance over typical policy gradient baselines under matched targets, while intrinsically preserving optimization stability and response diversity.\n\n---\n\n## 109. Normalized Architectures are Natively 4-Bit\n- **Authors**: Maxim Fishman , Brian Chmiel , Ron Banner , Daniel Soudry , Boris Ginsburg\n- **URL**: [https://arxiv.org/abs/2605.06067](https://arxiv.org/abs/2605.06067)\n- **Abstract**:\n> Training large language models at 4-bit precision is critical for efficiency. We show that nGPT, an architecture that constrains weights and hidden representations to the unit hypersphere, is inherently more robust to low-precision arithmetic. This removes the need for interventions-such as applying random Hadamard transforms and performing per-tensor scaling calculations-to preserve model quality, and it enables stable end-to-end NVFP4 training. We validate this approach on both a 1.2B dense model and hybrid (Mamba-Transformer) MoE models of up to 3B/30B parameters. We trace this robustness to the dot product: while quantization noise remains largely uncorrelated in both standard and normalized architectures, the signal behaves differently. In nGPT, the hypersphere constraint enhances weak positive correlations among the element-wise products, leading to a constructive accumulation of the signal across the hidden dimension while the noise continues to average out. This yields a higher effective signal-to-noise ratio and a flatter loss landscape, with the effect strengthening as the hidden dimension grows, suggesting increasing advantages at scale. A reference implementation is available at this https URL\n\n---\n\n## 110. Causal Reinforcement Learning for Complex Card Games: A Magic The Gathering Benchmark\n- **Authors**: Cristiano da Costa Cunha , Ajmal Mian , Tim French , Wei Liu\n- **URL**: [https://arxiv.org/abs/2605.06066](https://arxiv.org/abs/2605.06066)\n- **Abstract**:\n> Causal reinforcement learning (RL) lacks benchmarks for complex systems that combine sequential decision making, hidden information, large masked action spaces, and explicit causal structure. We introduce MTG-Causal-RL, a Gymnasium benchmark built on Magic: The Gathering with a 3,077-dimensional partial observation, a 478-action masked discrete action space, five competitive Standard archetypes, three reward schemes, and a hand-specified Structural Causal Model (SCM) over strategic variables. Every episode exposes causal variables, SCM-predicted intervention effects, and per-factor credit traces, making causal credit assignment, leave-one-out cross-archetype transfer, and policy auditability first-class metrics. We adapt a panel of reference baselines: random, heuristic, masked PPO, a causal-world-model PPO variant, and an architecture-matched scalar control. We propose Causal Graph-Factored Advantage PPO (CGFA-PPO) as a reference causal agent that uses SCM parents of win probability as factor-aligned critic targets with an intervention-calibration loss. All comparisons use paired seeds, paired-bootstrap confidence intervals, and Holm-Bonferroni correction within pre-registered families. Masked PPO and CGFA-PPO reach competitive in-distribution win rates and exceed the random baseline; per-factor calibration trajectories and leave-one-out transfer gaps expose diagnostic structure that scalar win rate alone cannot. We release the benchmark, reference-baseline results, and full evaluation protocol openly. By coupling a strategically rich, partially observed domain with an explicit causal interface and statistical protocol, MTG-Causal-RL gives causal-RL, world-model, and LLM-agent research a shared testbed for questions current benchmarks cannot pose together: causal credit assignment under masked action spaces, structural transfer across archetypes, and SCM-grounded policy auditability.\n\n---\n\n## 111. Optimal Transport for LLM Reward Modeling from Noisy Preference\n- **Authors**: Licheng Pan , Haochen Yang , Haoxuan Li , Yunsheng Lu , Yongqi Tong , Yinuo Wang , Shijian Wang , Zhixuan Chu , Lei Shen , Yuan Lu , Hao Wang\n- **URL**: [https://arxiv.org/abs/2605.06036](https://arxiv.org/abs/2605.06036)\n- **Abstract**:\n> Reward models are fundamental to Reinforcement Learning from Human Feedback (RLHF), yet real-world datasets are inevitably corrupted by noisy preference. Conventional training objectives tend to overfit these errors, while existing denoising approaches often rely on homogeneous noise assumptions that fail to capture the complexity of linguistic preferences. To handle these challenges, we propose SelectiveRM, a framework grounded in optimal transport. We first devise a Joint Consistency Discrepancy to align the distribution of model predictions with preference data. Furthermore, to address the limitation of strict mass conservation which compels the model to fit outliers, we incorporate a Mass Relaxation mechanism via partial transport. This enables the autonomous exclusion of samples with noisy preference that contradict semantic consistency. Theoretically, we demonstrate that SelectiveRM optimizes a tighter upper bound on the unobserved clean risk. Extensive experiments validate that our approach significantly outperforms state-of-the-art baselines across diverse benchmarks.\n\n---\n\n## 112. Does Synthetic Data Help? Empirical Evidence from Deep Learning Time Series Forecasters\n- **Authors**: Hugo Cazaux , Eyjólfur Ingi Ásgeirsson , Hlynur Stefánsson\n- **URL**: [https://arxiv.org/abs/2605.06032](https://arxiv.org/abs/2605.06032)\n- **Abstract**:\n> Synthetic data has transformed language model training, yet its role in time series forecasting remains poorly understood. We present a large-scale empirical study: nine experiment groups, 4,218 runs systematically evaluating synthetic time series augmentation across five architectures, four synthetic signals and seven datasets. The effect is sharply architecture-conditional: channel-mixing models (TimesNet, iTransformer) benefit in the majority of trials, while channel-independent models (DLinear, PatchTST) are consistently degraded. In selected low-resource settings the gains are striking: TimesNet trained on only 10\\% of Weather data with synthetic augmentation surpasses the full-data baseline (4 of 16 sparsity-dataset combinations). Averaged across all architectures, augmentation hurts in 67\\% of trials. We further find that only the Seasonal-Trend generator reliably helps across the tested benchmarks, and that hard curriculum switching is actively harmful (+24\\% MSE degradation). These results provide concrete, actionable guidelines on how to use synthetic data: use synthetic augmentation with channel-mixing architectures, use gradual annealing schedules, and treat low-resource augmentation as architecture- and dataset-dependent. Code is available at \\href{ this https URL }\n\n---\n\n## 113. Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks\n- **Authors**: Guoxin Lu , Letian Sha , Qing Wang , Peijie Sun , Hao Zhou , Hua Dai , Fu Xiao\n- **URL**: [https://arxiv.org/abs/2605.05995](https://arxiv.org/abs/2605.05995)\n- **Abstract**:\n> The safety alignment of Large Language Models (LLMs) remains vulnerable to Harmful Fine-tuning (HFT). While existing defenses impose constraints on parameters, gradients, or internal representations, we observe that they can be effectively circumvented under persistent HFT. Our analysis traces this failure to the inherent redundancy of the high-dimensional parameter space: attackers exploit optimization trajectories that are orthogonal to defense constraints to restore harmful capabilities while deceptively adhering to safety restrictions. To address this, we propose Safety Bottleneck Regularization (SBR). SBR shifts the defensive focus from the redundant parameter space to the unembedding layer, which serves as a geometric bottleneck. By anchoring the final hidden states of harmful queries to those of the safety-aligned model, SBR enables the model to maintain safe responses even under persistent HFT. Extensive experiments confirm SBR's effectiveness, demonstrating that utilizing just a single safety anchor is sufficient to reduce the Harmful Score to $<$10 while preserving competitive performance on benign downstream tasks.\n\n---\n\n## 114. PragLocker: Protecting Agent Intellectual Property in Untrusted Deployments via Non-Portable Prompts\n- **Authors**: Qinfeng Li , Yuntai Bao , Jianghui Hu , Wenqi Zhang , Jintao Chen , Huifeng Zhu , Yier Jin , Xuhong Zhang\n- **URL**: [https://arxiv.org/abs/2605.05974](https://arxiv.org/abs/2605.05974)\n- **Abstract**:\n> LLM agents rely on prompts to implement task-specific capabilities based on foundation LLMs, making agent prompts valuable intellectual property. However, in untrusted deployments, adversaries can copy and reuse these prompts with other proprietary LLMs, causing economic losses. To protect these prompts, we identify four key challenges: proactivity, runtime protection, usability, and non-portability that existing approaches fail to address. We present PragLocker, a prompt protection scheme that satisfies these requirements. PragLocker constructs function-preserving obfuscated prompts by anchoring semantics with code symbols and then using target-model feedback to inject noise, yielding prompts that only work on the target LLM. Experiments across multiple agent systems, datasets, and foundation LLMs show that PragLocker substantially reduces cross-LLM portability, maintains target performance, and remains robust against adaptive attackers.\n\n---\n\n## 115. Towards Reliable LLM Evaluation: Correcting the Winner's Curse in Adaptive Benchmarking\n- **Authors**: Yang Xu , Jiefu Zhang , Haixiang Sun , Zihan Zhou , Tianyu Cao , Vaneet Aggarwal\n- **URL**: [https://arxiv.org/abs/2605.05973](https://arxiv.org/abs/2605.05973)\n- **Abstract**:\n> Adaptive prompt and program search makes LLM evaluation selection-sensitive. Once benchmark items are reused inside tuning, the observed winner's score need not estimate the fresh-data performance of the full tune-then-deploy procedure. We study inference for this procedure-level target under explicit tuning budgets. We propose SIREN, a selection-aware repeated-split reporting protocol that freezes the post-search shortlist, separates splitwise selection from held-out evaluation, and uses an item-level Gaussian multiplier bootstrap for uncertainty quantification. In a fixed-shortlist regime with smooth stabilized selection, the estimator admits a first-order item-level representation, and the bootstrap yields valid simultaneous inference on a finite budget grid. This supports confidence intervals for procedure-performance curves and pre-specified equal-budget and cross-budget comparisons. Controlled simulations and MMLU-Pro tuning experiments show that winner-based reporting can be optimistic and can change deployment conclusions, while SIREN remains close to the finite-sample reporting target.\n\n---\n\n## 116. Beyond Uniform Credit Assignment: Selective Eligibility Traces for RLVR\n- **Authors**: Chaoli Mou , Zhan Zhuang , Xinning Chen , Yu Zhang\n- **URL**: [https://arxiv.org/abs/2605.05965](https://arxiv.org/abs/2605.05965)\n- **Abstract**:\n> Reinforcement Learning with Verifiable Rewards (RLVR) has become a key approach for improving the reasoning abilities of large language models. However, widely used critic-free algorithms such as Group Relative Policy Optimization (GRPO) necessitate a ``uniform credit assignment'' assumption that indiscriminately broadcast trajectory-level advantages, hindering learning efficiency by failing to distinguish critical reasoning steps. To address this limitation, we propose Selective Eligibility Traces (S-trace). Grounded in the intuition of partial trust region preservation, we initially introduce P-trace as a sample-efficient, critic-free eligibility traces method, upon which we build S-trace, implementing a sparse eligibility traces mechanism to further mitigate variance and achieve fine-grained credit assignment by selectively masking low-entropy tokens. Theoretically, we contextualize the recent Group Sequence Policy Optimization (GSPO) method within the critic-free eligibility traces framework, identifying it as a special instance of the eligibility traces method operating under uniform credit assignment. Experiments demonstrate that S-trace not only outperforms GRPO, showing gains of 0.49\\% on Qwen3-1.7B and 3.16\\% on Qwen3-4B, and maintaining a robust 2.98\\% improvement when scaled further to Qwen3-8B in average pass@16, but notably achieves this with simultaneously higher sample and token efficiency.\n\n---\n\n## 117. Hallucination as an Anomaly: Dynamic Intervention via Probabilistic Circuits\n- **Authors**: Erik Nielsen , Elia Cunegatti , Marcus Vukojevic , Giovanni Iacca\n- **URL**: [https://arxiv.org/abs/2605.05953](https://arxiv.org/abs/2605.05953)\n- **Abstract**:\n> One of the most critical challenges in Large Language Models is their tendency to hallucinate, i.e., produce factually incorrect responses. Existing approaches show promising results in terms of hallucination correction, but still suffer from a main limitation: they apply corrections indiscriminately to every token, corrupting also the originally correct generations. To overcome this drawback, we propose PCNET, a Probabilistic Circuit trained as a tractable density estimator over the LLM residual stream. The method detects hallucinations as geometric anomalies on the factual manifold, which is done via exact Negative Log-Likelihood computation, hence without the need for sampling, external verifiers, or weight modifications, as in existing techniques. To demonstrate its effectiveness, we exploit PCNET as a dynamic gate that distinguishes hallucinated from factual hidden states at each decoding step. This triggers our second main contribution, PC-LDCD (Probabilistic Circuit Latent Density Contrastive Decoding), only when the latent geometry deviates from factual regions, while leaving correct generations untouched. Across four LLMs, ranging from 1B to 8B models, and four benchmarks covering conversational reasoning, knowledge-intensive QA, reading comprehension, and truthfulness, PCNET achieves near-perfect hallucination detection across CoQA, SQuAD v2.0, and TriviaQA, with AUROC reaching up to 99%. Moreover, PC-LDCD obtains the highest True+Info, MC2, and MC3 scores on TruthfulQA in three out of four models, in comparison with state-of-the-art baselines, while reducing the mean corruption rate to 53.7% and achieving a preservation rate of 79.3%. Our proposed method is publicly available on GitHub.\n\n---\n\n## 118. LLM-Driven Design Space Exploration of FPGA-based Accelerators\n- **Authors**: Vinamra Sharma , Xingjian Fu , Jude Haris , José Cano\n- **URL**: [https://arxiv.org/abs/2605.05920](https://arxiv.org/abs/2605.05920)\n- **Abstract**:\n> Designing field-programmable gate array (FPGA)-based accelerators for modern artificial intelligence workloads requires navigating a large and complex hardware design space encompassing architectural parameters, dataflow strategies, and memory hierarchies, making the process time-consuming and resource-intensive. While the SECDA methodology enables rapid hardware-software co-design of accelerators through SystemC simulation and FPGA execution, identifying optimal accelerator configurations still requires substantial manual effort and domain expertise. This work presents SECDA-DSE, a framework that integrates Large Language Models (LLMs) into the SECDA ecosystem, comprising tools built around SECDA to automate the design space exploration (DSE) of FPGA-based accelerators. SECDA-DSE combines a structured DSE Explorer for generating accelerator configurations with an LLM Stack that performs reasoning-guided exploration using retrieval-augmented generation and chain-of-thought prompting, alongside a feedback loop that enables reinforced fine-tuning for continuous improvement. We demonstrate the feasibility of SECDA-DSE through an initial high-level synthesis based evaluation of a generated accelerator design that meets synthesis timing and resource constraints on an Zynq-7000 FPGA.\n\n---\n\n## 119. Quantum-enhanced Large Language Models on Quantum Hardware via Cayley Unitary Adapters\n- **Authors**: Borja Aizpurua , Sukhbinder Singh , Augustine Kshetrimayum , Saeed S. Jahromi , Roman Orus\n- **URL**: [https://arxiv.org/abs/2605.05914](https://arxiv.org/abs/2605.05914)\n- **Abstract**:\n> Large language models (LLMs) have transformed artificial intelligence, yet classical architectures impose a fundamental constraint: every trainable parameter demands classical memory that scales unfavourably with model size. Quantum computing offers a qualitatively different pathway, but practical demonstrations on real hardware have remained elusive for models of practical relevance. Here we show that Cayley-parameterised unitary adapters -- quantum circuit blocks inserted into the frozen projection layers of pre-trained LLMs and executed on a 156-qubit IBM Quantum System Two superconducting processor -- improve the perplexity of Llama 3.1 8B, an 8-billion-parameter model in widespread use, by 1.4% with only 6,000 additional parameters and end-to-end inference validated on real Quantum Processing Unit (QPU). A systematic study on SmolLM2 (135M parameters), chosen for its tractability, reveals monotonically improving perplexity with unitary block dimension, 83% recovery of compression-induced degradation, and correct answers to questions that both classical baselines fail -- with a sharp noise-expressivity phase transition identifying the concrete path to quantum utility at larger qubit scales.\n\n---\n\n## 120. CITE: Anytime-Valid Statistical Inference in LLM Self-Consistency\n- **Authors**: Hirofumi Ota , Naoto Iwase , Yuki Ichihara , Junpei Komiyama , Masaaki Imaizumi\n- **URL**: [https://arxiv.org/abs/2605.05873](https://arxiv.org/abs/2605.05873)\n- **Abstract**:\n> Large language models often improve reasoning by sampling multiple outputs and aggregating their final answers, but precise and efficient control of error levels remains a challenging task. In particular, deciding when to stop sampling remains difficult when the stopping rule is data-dependent and the set of possible answers is not known in advance. We study anytime-valid certification of a prespecified target answer as the unique mode of the model's response distribution, a guarantee distinct from answer correctness. We propose the Certification by Intersection-union Testing with E-processes (CITE) algorithm, which provably controls false certification at any prescribed level under arbitrary data-driven stopping, without requiring prior knowledge of the answer category set. We also prove an category-set-size-free stopping-time rate, establish matching minimax lower bounds up to constants in the main regime, and extend the construction to confidence-weighted voting. Simulations and LLM self-consistency experiments show empirical error control and improved certification in diffuse-tail settings.\n\n---\n\n## 121. VideoRouter: Query-Adaptive Dual Routing for Efficient Long-Video Understanding\n- **Authors**: Kuanwei Lin , Wenhao Zhang , Ge Li\n- **URL**: [https://arxiv.org/abs/2605.05848](https://arxiv.org/abs/2605.05848)\n- **Abstract**:\n> Video large multimodal models increasingly face a scalability bottleneck: long videos produce excessively long visual-token sequences, which sharply increase memory and latency during inference. While existing compression methods are effective in specific settings, most are either weakly query-aware or apply a fixed compression policy across frames, proving suboptimal when visual evidence is unevenly distributed over time. To address this, we present VideoRouter, a query-adaptive dual-router framework built on InternVL for budgeted evidence allocation. The Semantic Router predicts the dominant allocation policy, choosing between broad temporal coverage and adaptive high-resolution preservation, while the Image Router uses early LLM layers to score frame relevance. This enables aggressive compression on less relevant frames while preserving detail on critical evidence frames. To train both routers, we build Video-QTR-10K for allocation-policy supervision and Video-FLR-200K for frame-relevance supervision. Experiments on VideoMME, MLVU, and LongVideoBench show that VideoRouter consistently improves over the InternVL baseline under comparable or lower budgets, achieving up to a 67.9% token reduction.\n\n---\n\n## 122. LoopTrap: Termination Poisoning Attacks on LLM Agents\n- **Authors**: Huiyu Xu , Zhibo Wang , Wenhui Zhang , Ziqi Zhu , Yaopeng Wang , Kui Ren , Chun Chen\n- **URL**: [https://arxiv.org/abs/2605.05846](https://arxiv.org/abs/2605.05846)\n- **Abstract**:\n> Modern LLM agents solve complex tasks by operating in iterative execution loops, where they repeatedly reason, act, and self-evaluate progress to determine when a task is complete. In this work, we show that while this self-directed loop facilitates autonomy, it also introduces a critical risk: by injecting malicious prompts into the agent's context, an adversary can distort the agent's termination judgment, making it believe the task remains incomplete and leading to unbounded this http URL understand this threat, we define and systematically characterize it as Termination Poisoning and design 10 representative attack strategies. Through a empirical study spanning 8 LLM agents and 60 tasks, we demonstrate that different LLM agents exhibit distinct behavioral signatures that determine which strategies succeed. These transferable patterns can serve as principled guidance for crafting effective attacks against previously unseen agents and tasks, enabling scalable red-teaming beyond manually designed templates. Building on these insights, we introduce LoopTrap, an automated red-teaming framework that synthesizes target-specific malicious prompts by exploiting agent behavioral tendencies. LoopTrap first constructs a behavioral profile of the target agent along four vulnerability dimensions via lightweight probing. It then performs adaptive trap synthesis, routing to the most effective strategy and selecting optimal injections via a self-scoring mechanism. Finally, successful traps are abstracted into a reusable skill library, while failed attempts are refined through self-reflection, ensuring continuous improvement. Extensive evaluation shows that LoopTrap achieves an average of 3.57$\\times$ step amplification across 8 mainstream agents, with a peak of 25$\\times$.\n\n---\n\n## 123. LeakDojo: Decoding the Leakage Threats of RAG Systems\n- **Authors**: Maosen Zhang , Jianshuo Dong , Boting Lu , Wenyue Li , Xiaoping Zhang , Tianwei Zhang , Han Qiu\n- **URL**: [https://arxiv.org/abs/2605.05818](https://arxiv.org/abs/2605.05818)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to leverage external knowledge, but also exposes valuable RAG databases to leakage attacks. As RAG systems grow more complex and LLMs exhibit stronger instruction-following capabilities, existing studies fall short of systematically assessing RAG leakage risks. We present LeakDojo, a configurable framework for controlled evaluation of RAG leakage. Using LeakDojo, we benchmark six existing attacks across fourteen LLMs, four datasets, and diverse RAG systems. Our study reveals that (1) query generation and adversarial instructions contribute independently to leakage, with overall leakage well approximated by their product; (2) stronger instruction-following capability correlates with higher leakage risk; and (3) improvements in RAG faithfulness can introduce increased leakage risk. These findings provide actionable insights for understanding and mitigating RAG leakage in practice. Our codebase is available at this https URL .\n\n---\n\n## 124. LCC-LLM: Leveraging Code-Centric Large Language Models for Malware Attribution\n- **Authors**: Christopher G. Pedraza Pohlenz , Hassan Jalil Hadi , Ali Hassan , Ali Shoker\n- **URL**: [https://arxiv.org/abs/2605.05807](https://arxiv.org/abs/2605.05807)\n- **Abstract**:\n> LLMs are increasingly explored for malware analysis; however, current LLM-based malware attribution remains limited by unsupported indicators and insufficient code-level grounding for identifying malicious and vulnerable code segments. To address these limitations, this research introduces LCC-LLM, a code-centric benchmark dataset and evidence-grounded framework for malware attribution and multi-task static malware analysis. The proposed LCCD dataset contains approximately 34K PE samples processed through a large-scale reverse-engineering pipeline and represented using decompiled C code, assembly code, CFG/FCG artifacts, hexadecimal data, PE metadata, suspicious API evidence, and structural features. Beyond dataset construction, LCC-LLM integrates LangGraph-orchestrated static analysis with multi-source cybersecurity knowledge to support evidence-grounded malware reasoning. The framework employs a seven-layer retrieval-augmented generation pipeline, CoVe for IoC validation, and a multi-dimensional quality gate to improve factual reliability and analyst-oriented decision support. Curriculum-ordered instruction data is used to fine-tune DeepSeek-R1-Distill-Qwen-14B and Qwen3-Coder-30B-A3B using QLoRA. Evaluation across 43 malware-analysis task types achieves an average semantic similarity of 0.634, with the highest task-level performance in structured report generation, IoC extraction, vulnerability assessment, malware configuration extraction, and malware class detection. In a real-world case study using MalwareBazaar samples, the grounded pipeline achieves a 10/10 structured analysis pass rate, producing CFG/FCG evidence, MITRE ATT&CK mappings, detection guidance, and analyst-ready reports. These results show that code-centric representations, retrieval grounding, and verification-guided reasoning improve the reliability and operational usefulness of LLM-assisted malware attribution.\n\n---\n\n## 125. Revealing Modular Gradient Noise Imbalance in LLMs: Calibrating Adam via Signal-to-Noise Ratio\n- **Authors**: Ziqing Wen , Zhouyang Liu , Jiahuan Wang , Ping Luo , Li Shen , Dongsheng Li , Tao Sun\n- **URL**: [https://arxiv.org/abs/2605.05794](https://arxiv.org/abs/2605.05794)\n- **Abstract**:\n> The impressive performance of large language models (LLMs) arises from their massive scale and heterogeneous module composition. However, this structural heterogeneity introduces additional optimization challenges. While adaptive optimizers such as Adam(W) provide per-parameter adaptivity, they do not explicitly account for module-level gradient heterogeneity, resulting in slower convergence, suboptimal performance, or training instability. Existing approaches typically rely on manually tuned module-specific learning rates or specific optimization strategies, which are computationally costly and difficult to generalize across tasks or models. To establish a more principled approach, we first analyze the noise-damping behavior of Adam in high-noise modules and introduce \\textbf{Module-wise Learning Rate Scaling via SNR (MoLS)}. MoLS estimates module-level SNRs to scale Adam updates, allowing automated module-wise learning rate allocation without manual tuning. Empirical results through multiple LLM training benchmarks demonstrate that MoLS improves convergence speed and generalization, achieving performance comparable to carefully tuned module-specific learning rates, while remaining compatible with memory-efficient training algorithms.\n\n---\n\n## 126. Multi-Dimensional Behavioral Evaluation of Agentic Stock Prediction Systems Using LLM Judges with Closed-Loop Reinforcement Learning Feedback\n- **Authors**: Mohammad Al Ridhawi , Mahtab Haj Ali , Hussein Al Osman\n- **URL**: [https://arxiv.org/abs/2605.05739](https://arxiv.org/abs/2605.05739)\n- **Abstract**:\n> Agentic stock prediction systems make sequences of interdependent decisions (regime detection, pathway routing, reinforcement learning control) whose individual quality is hidden by aggregate metrics such as mean absolute percentage error (MAPE) or directional accuracy. We present a behavioral evaluation framework that addresses this gap. Behavioral traces logged at every autonomous decision point are grouped into five-day episodes and scored along six domain-specific dimensions (regime detection, routing, adaptation, risk calibration, strategy coherence, error recovery) by an ensemble of three large language model (LLM) judges (GPT 5.4, Claude 4.6 Opus, Gemini 3.1 Pro). Perturbation-based validation on 420 episodes yields targeted score drops of $-1.6$ to $-2.4$ on intended dimensions versus an average of $-0.32$ on the remaining five, with cross-model agreement up to Krippendorff's $\\alpha = 0.85$. The composite behavioral score, used here only for cross-episode reporting, correlates at $\\rho = 0.72$ with realized 20-day Sharpe ratio from offline backtesting. Closing the loop, the framework converts deficient per-dimension scores into a credit-assigned penalty term added to the Soft Actor-Critic (SAC) reward. Three short fine-tuning cycles, all confined to the validation period, produce on the held-out 2017-2025 test period a one-day MAPE reduction from 0.61% to 0.54% (an 11.5% relative reduction; $p<0.001$, Cohen's $d=0.31$), a directional accuracy increase from 71% to 74%, and an 18% Sharpe ratio improvement (95% bootstrap CI [8.2%, 27.4%]), with gains concentrated in high-volatility episodes where the original system was most behaviorally deficient. Results are from offline backtesting and do not address effects specific to live deployment.\n\n---\n\n## 127. CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning\n- **Authors**: Md Anwar Hossen , Fatema Siddika , Juan Pablo Munoz , Tanya Roosta , Ali Jannesari\n- **URL**: [https://arxiv.org/abs/2605.05732](https://arxiv.org/abs/2605.05732)\n- **Abstract**:\n> Large language models (LLMs) can acquire new capabilities through fine-tuning, but continual adaptation often leads to catastrophic forgetting. We propose CRAFT, a continual learning framework that avoids updating model weights by instead learning low-rank interventions on hidden representations. CRAFT proceeds in three stages: it first routes each task to a group of similar tasks based on output-distribution divergence; it then fine-tunes the model using a Kullback-Leibler (KL) divergence against the group's prior state, which directly controls forgetting and determines convergence; finally, it merges interventions for the updated task into the shared representation using the same KL signal. This design unifies routing, regularization, and merging through a single KL-based objective. CRAFT improves overall performance and reduces forgetting compared to strong LoRA-based approaches across multiple benchmarks and model scales, while remaining robust to task ordering. These results suggest that controlling adaptation in representation space, guided by output-space divergence, provides a scalable and principled approach to continual learning in LLMs.\n\n---\n\n## 128. SafeHarbor: Hierarchical Memory-Augmented Guardrail for LLM Agent Safety\n- **Authors**: Zhe Liu , Zonghao Ying , Wenxin Zhang , Quanchen Zou , Deyue Zhang , Dongdong Yang , Xiangzheng Zhang , Hao Peng\n- **URL**: [https://arxiv.org/abs/2605.05704](https://arxiv.org/abs/2605.05704)\n- **Abstract**:\n> With the rapid evolution of foundation models, Large Language Model (LLM) agents have demonstrated increasingly powerful tool-use capabilities. However, this proficiency introduces significant security risks, as malicious actors can manipulate agents into executing tools to generate harmful content. While existing defensive mechanisms are effective, they frequently suffer from the over-refusal problem, where increased safety strictness compromises the agent's utility on benign tasks. To mitigate this trade-off, we propose \\textsc{SafeHarbor}, a novel framework designed to establish precise decision boundaries for LLM agents. Unlike static guidelines, \\textsc{SafeHarbor} extracts context-aware defense rules through enhanced adversarial generation. We design a local hierarchical memory system for dynamic rule injection, offering a training-free, efficient, and plug-and-play solution. Furthermore, we introduce an information entropy-based self-evolution mechanism that continuously optimizes the memory structure through dynamic node splitting and merging. Extensive experiments demonstrate that \\textsc{SafeHarbor} achieves state-of-the-art performance on both ambiguous benign tasks and explicit malicious attacks, notably attaining a peak benign utility of 63.6\\% on GPT-4o while maintaining a robust refusal rate exceeding 93\\% against harmful requests. The source code is publicly available at this https URL .\n\n---\n\n## 129. Active Learning for Communication Structure Optimization in LLM-Based Multi-Agent Systems\n- **Authors**: Huchen Yang , Xinghao Dong , Dan Negrut , Jin-Long Wu\n- **URL**: [https://arxiv.org/abs/2605.05703](https://arxiv.org/abs/2605.05703)\n- **Abstract**:\n> Optimizing the communication structure of large language model based multi-agent systems (LLM-MAS) has been shown to improve downstream performance and reduce token usage. Existing methods typically rely on randomly sampled training tasks. However, tasks may differ substantially in difficulty and domain, and thus they are not equally informative for updating communication structure, making optimization under limited training budgets often unstable and highly sensitive to the particular training set. To actively identify the most valuable tasks for communication-structure optimization, we propose an ensemble-based information-theoretic task selection framework. The proposed method estimates task informativeness by how much a candidate task changes the distribution over graph parameters, using ensemble Kalman inversion as an efficient and derivative-free approximation of the corresponding Bayesian update. The resulting estimator is especially suitable for black-box and noisy multi-agent systems. To enhance scalability, we construct a compact candidate pool through embedding-based representative selection and combine the informative selection with surrogate modeling and batch Thompson sampling. We validate our method in both benign settings and settings with agent attacks, demonstrating its effectiveness for communication-structure optimization under constrained computational budgets.\n\n---\n\n## 130. An Empirical Study of Proactive Coding Assistants in Real-World Software Development\n- **Authors**: Lehui Li , Ruixuan Jia , Guo-Ye Yang , Jia Li\n- **URL**: [https://arxiv.org/abs/2605.05700](https://arxiv.org/abs/2605.05700)\n- **Abstract**:\n> Large language model (LLM)-based coding assistants have made substantial progress, yet most systems remain reactive, requiring developers to explicitly formulate their needs. Proactive coding assistants aim to infer latent developer intent from integrated development environment (IDE) interactions and repository context, thereby reducing interaction overhead and supporting more seamless assistance. However, research in this direction is limited by the scarcity of large-scale real-world developer behavior data. Existing studies therefore often rely on LLM-simulated IDE traces, whose fidelity to real development behavior remains unclear. In this paper, we investigate this simulation-to-reality gap through a large-scale empirical study. We collect real IDE interaction traces from 1{,}246 experienced industry developers over three consecutive days using a custom Visual Studio Code extension, and construct paired LLM-simulated traces for controlled comparison. Our analysis shows that simulated traces differ substantially from real traces in behavioral diversity, temporal structure, and exploratory patterns. Based on the collected data, we introduce \\textbf{ProCodeBench}, a real-world benchmark for proactive intent prediction. Experiments with representative LLMs, retrieval-augmented methods, and agentic baselines show that current approaches remain far from reliable under real IDE traces, suggesting that simulation-based evaluation can overestimate real-world performance. Finally, our training study shows that simulated data cannot replace real data, but can complement it when used before real-world fine-tuning. These findings highlight the importance of real developer behavior data for evaluating and training proactive coding assistants.\n\n---\n\n## 131. Irminsul: MLA-Native Position-Independent Caching for Agentic LLM Serving\n- **Authors**: Bole Ma , Jan Eitzinger , Harald Köstler\n- **URL**: [https://arxiv.org/abs/2605.05696](https://arxiv.org/abs/2605.05696)\n- **Abstract**:\n> Agentic LLM workloads put bit-identical tokens at shifted positions every turn, voiding prefix caches at the first byte of divergence. Operators report cache-hit regressions ranging from moderate slowdowns to severe TTFT spikes of 10-16s on unchanged content. Prior position-independent caching systems correct RoPE on the full $d_K$-dimensional key, an architectural cost imposed by GQA, not by caching itself. Multi-Head Latent Attention, deployed at scale in DeepSeek-V2/V3/R1, Kimi-K2/Moonlight, GLM-5, and Mistral Large 3, factors each KV row into a position-free $c_{KV}$ and a 64-dim $k_r$ correctable in closed form; this structure motivates content-addressed caching as a natural fit rather than a GQA workaround. We present Irminsul, which extends SGLang's radix cache with content-hash keying over CDC-chunked segments and a $\\delta$-rotation rule for $k_r$. We evaluate three native MLA-MoE deployments - DeepSeek-V2-Lite (16B/2.4B), Kimi Moonlight-16B-A3B, and JoyAI-Flash (48B/3B) - with output-consistency on all three and recovery measured on the two endpoints; Irminsul recovers up to ~83% of prompt tokens above exact-prefix on agentic traffic while delivering 63% prefill energy savings per cache hit. We argue that content-addressed caching belongs in the serving stack as a first-class primitive, not a retrofit over prefix matching.\n\n---\n\n## 132. Decomposing the Basic Abilities of Large Language Models: Mitigating Cross-Task Interference in Multi-Task Instruct-Tuning\n- **Authors**: Bing Wang , Ximing Li , Changchun Li , Jinjin Chi , Gang Niu , Masashi Sugiyama\n- **URL**: [https://arxiv.org/abs/2605.05676](https://arxiv.org/abs/2605.05676)\n- **Abstract**:\n> Recently, the prominent performance of large language models (LLMs) has been largely driven by multi-task instruct-tuning. Unfortunately, this training paradigm suffers from a key issue, named cross-task interference, due to conflicting gradients over shared parameters among different tasks. Some previous methods mitigate this issue by isolating task-specific parameters, e.g., task-specific neuron selection and mixture-of-experts. In this paper, we empirically reveal that the cross-task interference still exists for the existing solutions because of many parameters also shared by different tasks, and accordingly, we propose a novel solution, namely Basic Abilities Decomposition for multi-task Instruct-Tuning (BADIT). Specifically, we empirically find that certain parameters are consistently co-activated, and that co-activated parameters naturally organize into base groups. This motivates us to analogize that LLMs encode several orthogonal basic abilities, and that any task can be represented as a linear combination of these abilities. Accordingly, we propose BADIT that decomposes LLM parameters into orthogonal high-singular-value LoRA experts representing basic abilities, and dynamically enforces their orthogonality during training via spherical clustering of rank-1 components. We conduct extensive experiments on the SuperNI benchmark with 6 LLMs, and empirical results demonstrate that BADIT can outperform SOTA methods and mitigate the degree of cross-task interference.\n\n---\n\n## 133. XL-SafetyBench: A Country-Grounded Cross-Cultural Benchmark for LLM Safety and Cultural Sensitivity\n- **Authors**: Dasol Choi , Eugenia Kim , Jaewon Noh , Sang Seo , Eunmi Kim , Myunggyo Oh , Yunjin Park , Brigitta Jesica Kartono , Josef Pichlmeier , Helena Berndt , Sai Krishna Mendu , Glenn Johannes Tungka , Özlem Gökçe , Suresh Gehlot , Katherine Pratt , Amanda Minnich , Haon Park\n- **URL**: [https://arxiv.org/abs/2605.05662](https://arxiv.org/abs/2605.05662)\n- **Abstract**:\n> Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Moreover, they rarely evaluate a model's ability to detect culturally embedded sensitivities as distinct from universal harms. We introduce XL-SafetyBench. a suite of 5,500 test cases across 10 country-language pairs, comprising a Jailbreak Benchmark of country-grounded adversarial prompts and a Cultural Benchmark where local sensitivities are embedded within innocuous requests. Each item is constructed via a multi-stage pipeline that combines LLM-assisted discovery, automated validation gates, and dual independent native-speaker annotators per country. To distinguish principled refusal from comprehension failure, we evaluate Attack Success Rate (ASR) alongside two complementary metrics we introduce: Neutral-Safe Rate (NSR) and Cultural Sensitivity Rate (CSR). Evaluating 10 frontier and 27 local LLMs reveals two key findings. First, jailbreak robustness and cultural awareness do not show a coupled relationship among frontier models, so a composite safety score obscures per-axis variation. Second, local models exhibit a near-linear ASR-NSR trade-off (r = -0.81), indicating that their apparent safety reflects generation failure rather than genuine alignment. XL-SafetyBench enables more nuanced, cross-cultural safety evaluation in the multilingual era.\n\n---\n\n## 134. One Turn Too Late: Response-Aware Defense Against Hidden Malicious Intent in Multi-Turn Dialogue\n- **Authors**: Xinjie Shen , Rongzhe Wei , Peizhi Niu , Haoyu Wang , Ruihan Wu , Eli Chien , Bo Li , Pin-Yu Chen , Pan Li\n- **URL**: [https://arxiv.org/abs/2605.05630](https://arxiv.org/abs/2605.05630)\n- **Abstract**:\n> Hidden malicious intent in multi-turn dialogue poses a growing threat to deployed large language models (LLMs). Rather than exposing a harmful objective in a single prompt, increasingly capable attackers can distribute their intent across multiple benign-looking turns. Recent studies show that even modern commercial models with advanced guardrails remain vulnerable to such attacks despite advances in safety alignment and external guardrails. In this work, we address this challenge by detecting the earliest turn at which delivering the candidate response would make the accumulated interaction sufficient to enable harmful action. This objective requires precise turn-level intervention that identifies the harm-enabling closure point while avoiding premature refusal of benign exploratory conversations. To further support training and evaluation, we construct the Multi-Turn Intent Dataset (MTID), which contains branching attack rollouts, matched benign hard negatives, and annotations of the earliest harm-enabling turns. We show that MTID helps enable a turn-level monitor TurnGate, which substantially outperforms existing baselines in harmful-intent detection while maintaining low over-refusal rates. TurnGate further generalizes across domains, attacker pipelines, and target models. Our code is available at this https URL .\n\n---\n\n## 135. Leveraging Image Generators to Address Training Data Scarcity: The Gen4Regen Dataset for Forest Regeneration Mapping\n- **Authors**: Gabriel Jeanson , David-Alexandre Duclos , William Larrivée-Hardy , Noé Cochet , Matěj Boxan , Anthony Deschênes , François Pomerleau , Philippe Giguère\n- **URL**: [https://arxiv.org/abs/2605.05627](https://arxiv.org/abs/2605.05627)\n- **Abstract**:\n> Sustainable forest management relies on precise species composition mapping, yet traditional ground surveys are labour-intensive and geographically constrained. While Uncrewed Aerial Vehicles (UAVs) offer scalable data collection, the transition to deep learning-based interpretation is bottlenecked by the severe scarcity of expert-annotated imagery, particularly in complex, visually heterogeneous regeneration zones. This paper addresses the dual challenges of data scarcity and extreme class imbalance in the semantic segmentation of fine-grained forest regeneration species by providing a scalable framework that reduces reliance on manual photo-interpretation for high-resolution, millimetre-level aerial imagery. Importantly, we leverage the large-scale vision-language Nano Banana Pro model to simultaneously generate high-fidelity images and their corresponding pixel-aligned semantic masks from prompts. We introduce WilDReF-Q-V2, an expansion of a natural forest dataset with 13 977 new unlabelled and 50 labelled real images, as well as the Gen4Regen dataset, featuring 2101 pairs of synthetic images and semantic masks. Our methodology integrates real-world data with AI-generated images, highlighting that AI-generated data is highly complementary to real-world data, with unified training yielding an F1 score improvement of over 15 %pt compared to purely supervised baselines. Furthermore, we demonstrate that even small quantities of prompt-generated data significantly improve performance for underrepresented species, some of which saw per-species F1 score gains of up to 30 %pt. We conclude that vision-language models can serve as agile data generators, effectively bootstrapping perception tasks for niche AI domains where expert labels are scarce or unavailable. Our datasets, source code, and models will be available at this https URL .\n\n---\n\n## 136. When2Speak: A Dataset for Temporal Participation and Turn-Taking in Multi-Party Conversations for Large Language Models\n- **Authors**: Vihaan Nama , Shreya Mendi , Zian Ye , Brinnae Bent\n- **URL**: [https://arxiv.org/abs/2605.05626](https://arxiv.org/abs/2605.05626)\n- **Abstract**:\n> Large Language Models (LLMs) excel at generating contextually appropriate responses but remain poorly calibrated for multi-party conversations, where deciding when to speak is as critical as what to say. In such settings, naively responding at every turn leads to excessive interruptions and degraded conversational coherence. We introduce When2Speak, a grounded synthetic dataset and four-stage generation pipeline for learning intervention timing in group interactions. The dataset comprises over 215,000 examples derived from 16,000 conversations involving 2-6 speakers, spanning diverse conversational styles, tones, and participant dynamics, and explicitly modeling SPEAK vs. SILENT decisions at each turn. Our pipeline combines real-world grounding, structured augmentation, controlled transcript synthesis, and fine-tuning-ready supervision, and is fully open-sourced to support reproducibility and adaptation to domain-specific conversational norms. Across multiple model families, supervised fine-tuning (SFT) on When2Speak significantly outperforms zero-shot baselines (e.g., the average Macro F1 increase across 4B+ parameter models was 60%, with the largest increase being 120%). However, SFT-trained models remain systematically over-conservative, missing nearly half of warranted interventions as seen through the Missed Intervention Rate (MIR), which was on average 0.50 and is noticed even at larger model sizes. To address this limitation, we apply reinforcement learning with asymmetric reward shaping, which reduces MIR to 0.186-0.218 and increases recall from 0.479 to 0.78-0.81. Our findings establish that temporal participation is a distinct and trainable dimension of conversational intelligence, and that grounded synthetic data provides an effective and scalable pathway for enabling LLMs to participate more naturally and appropriately in multi-party interactions.\n\n---\n\n## 137. AstroAlertBench: Evaluating the Accuracy, Reasoning, and Honesty of Multimodal LLMs in Astronomical Classification\n- **Authors**: Claire Chen , Jiabao Sean Xiao , Shuze Daniel Liu , Facundo Perez Paolino , Luke Handley , Theophile Jegou du Laz , Ricky Nilsson , Alice Zou , Matthew Graham , Ashish Mahabal\n- **URL**: [https://arxiv.org/abs/2605.05573](https://arxiv.org/abs/2605.05573)\n- **Abstract**:\n> Modern astronomical observatories generate a massive volume of multimodal data, creating a critical bottleneck for expert human review. While multimodal large language models (LLMs) have shown promise in interpreting complex visual and textual inputs, their ability to perform specialized scientific classification while providing interpretable reasoning remains understudied. We introduce AstroAlertBench, a comprehensive multimodal benchmark designed to evaluate LLM performance in astronomical event review along a three-stage logical chain: metadata grounding, scientific reasoning, and hierarchical classification over five categories. We use a pilot sample of 1,500 real-world alerts from the Zwicky Transient Facility (ZTF), a wide-field survey that scans the northern sky to detect transient astronomical events. On this dataset, we benchmark 13 frontier closed-source and open-weight LLMs that support visual input. Our results reveal that high accuracy does not always align with model ``honesty,'' defined as the ability to self-evaluate its reasoning, which affects its reliability as a real-world assistant. We further initialize a human-in-the-loop evaluation protocol as a precursor to future community-scale participation. Together, AstroAlertBench provides a framework for developing calibrated and interpretable astronomical assistants.\n\n---\n\n## 138. ReaComp: Compiling LLM Reasoning into Symbolic Solvers for Efficient Program Synthesis\n- **Authors**: Atharva Naik , Yash Mathur , Prakam , Carolyn Rose , David Mortensen\n- **URL**: [https://arxiv.org/abs/2605.05485](https://arxiv.org/abs/2605.05485)\n- **Abstract**:\n> LLMs can solve program synthesis tasks but remain inefficient and unreliable on hard instances requiring large combinatorial search. Given a small set of reasoning traces, we use coding agents to compile them into reusable symbolic program synthesizers over constrained DSLs. The resulting solvers require no LLM calls at test time and are strong standalone systems: symbolic solver ensembles reach 91.3% accuracy on PBEBench-Lite and 84.7% on PBEBench-Hard, outperforming LLMs with test-time scaling for the latter by +16.3 percentage points at zero LLM inference cost. They also complement LLM search, improving PBEBench-Hard accuracy from 68.4% to 85.8% while reducing reported token usage by 78%, and raising SLR-Bench hard-tier accuracy from 34.4% to 58.0% in a neuro-symbolic hybrid setting. Compared to directly using coding agents as per-instance solvers, induced solvers are substantially more Pareto-efficient, amortizing a small one-time construction cost over many zero-token executions. Finally, most solvers transfer zero-shot to a real historical linguistics task - predicting sound changes in natural language data - reaching 80.1% accuracy under ensembling and recovering some plausible linguistic rules. Together, these results show that reasoning traces can be compiled into reusable symbolic solvers that solve many tasks directly, complement LLM inference on hard cases, and provide a scalable route to domain-general solver induction. We release code and data for reproducibility.\n\n---\n\n## 139. Robustness of Graph Self-Supervised Learning to Real-World Noise: A Case Study on Text-Driven Biomedical Graphs\n- **Authors**: Othmane Kabal , Mounira Harzallah , Fabrice Guillet , Hideaki Takeda , Ryutaro Ichise\n- **URL**: [https://arxiv.org/abs/2605.05463](https://arxiv.org/abs/2605.05463)\n- **Abstract**:\n> Graph Self-Supervised Learning (GSSL) offers a powerful paradigm for learning graph representations without labeled data. However, existing work assumes clean, manually curated graphs. Recent advances in NLP enable the large-scale automatic extraction of knowledge graphs from text, opening new opportunities for GSSL while introducing substantial real-world noise. This type of noise remains largely unexplored, as prior robustness studies typically rely on synthetic perturbations. To address this gap, we present the first comprehensive evaluation of GSSL methods on text-driven graphs for unsupervised term typing. We introduce Noise-Aware Text-Driven Graph GSSL (NATD-GSSL), a unified framework that combines automatic graph construction, graph refinement, and GSSL. Our evaluation follows a dual-graph protocol that contrasts a noisy graph derived from MedMentions with a clean Unified Medical Language System (UMLS) reference graph, aligned through a shared gold standard. Our results reveal variability in robustness across both pretext tasks and Graph Neural Network (GNN) architectures. Relation reconstruction is highly sensitive to noise and benefits from well-defined schemas, whereas feature reconstruction is considerably more robust, achieving performance comparable to clean-graph settings. Contrastive objectives are generally less affected by noise but depend strongly on alignment with downstream tasks. GNN architecture also plays a critical role: bidirectional relational message-passing designs are better suited to noisy, text-driven graphs, while unidirectional relational ones perform best on clean graphs. Overall, NATD-GSSL provides practical guidance for applying GSSL to real-world, noisy graphs and achieves up to a 7\\% improvement over pretrained language model baselines. All code and benchmarks are publicly available at this https URL .\n\n---\n\n## 140. SLAM: Structural Linguistic Activation Marking for Language Models\n- **Authors**: Fabrice Harel-Canada , Amit Sahai\n- **URL**: [https://arxiv.org/abs/2605.05443](https://arxiv.org/abs/2605.05443)\n- **Abstract**:\n> LLM watermarks must be detectable without compromising text quality, yet most existing schemes bias the next-token distribution and pay for detection with measurable quality loss. We present SLAM (Structural Linguistic Activation Marking), a novel white-box watermarking scheme that sidesteps this cost by writing the mark into structural geometry rather than token frequencies: sparse autoencoders identify residual-stream directions encoding linguistic structure (e.g., voice, tense, clause order), and we causally steer those directions at generation time, leaving lexical sampling and semantics unconstrained. On Gemma-2 2B and 9B, SLAM achieves 100% detection accuracy with a quality cost of only 1-2 reward points - compared to 7.5-11.5 for KGW, EWD, and Unigram - with naturalness and diversity preserved at near-unwatermarked levels across both models. The trade-off is a complementary robustness profile: SLAM resists word-level edits but is vulnerable to paraphrase that restructures syntax (at a quality cost), the converse of token-distribution methods.\n\n---\n\n## 141. Information Theoretic Adversarial Training of Large Language Models\n- **Authors**: Yiwei Zhang , Jeremiah Birrell , Reza Ebrahimi , Rouzbeh Behnia , Jason Pacheco , Elisa Bertino\n- **URL**: [https://arxiv.org/abs/2605.05415](https://arxiv.org/abs/2605.05415)\n- **Abstract**:\n> Large language models (LLMs) remain vulnerable to adversarial prompting despite advances in alignment and safety, often exhibiting harmful behaviors under novel attack strategies. While adversarial training can improve robustness, existing approaches are computationally expensive and difficult to scale. Recent continuous adversarial training methods, such as Continuous adversarial training (CAT) and Continuous Adversarial Preference Optimization (CAPO), address this challenge by leveraging gradient-based perturbations in the embedding space, enabling more efficient and expressive attacks. Building on this paradigm, we propose WARDEN, a distributionally robust adversarial training framework for LLMs that dynamically reweights adversarial examples through an f -divergence ambiguity set around the empirical training distribution. Our method optimizes the worst-case adversarial loss within a divergence ball around the empirical data distribution, automatically emphasizing harder adversarial examples. Using the convex dual formulation, the objective reduces to a log-sum-exp form under the KL divergence, with a dynamical parameter controlling the strength of reweighting. This study leads to a new class of information-theoretic objectives that significantly reduce attack success rates while maintaining model utility. Across multiple LLMs and attack settings, WARDEN substantially reduces attack success rates with computational and utility costs comparable to CAT-, CAPO-, and MixAT-based baselines, making it a practical approach for scalable robust alignment.\n\n---\n\n## 142. Open-SAT: LLM-Guided Query Embedding Refinement for Open-Vocabulary Object Retrieval in Satellite Imagery\n- **Authors**: Md Adnan Arefeen , Biplob Debnath , Ravi K. Rajendran , Murugan Sankaradas , Srimat T. Chakradhar\n- **URL**: [https://arxiv.org/abs/2605.05344](https://arxiv.org/abs/2605.05344)\n- **Abstract**:\n> In satellite applications, user queries often take the form of open-ended natural language, extending beyond a fixed set of predefined categories. This open-vocabulary nature poses significant challenges for retrieving relevant image tiles, as the retrieval system must generalize to a wide range of unseen objects and concepts. While vision-language models (VLMs) such as CLIP are widely used for text-image retrieval, even fine-tuned variants often struggle to accurately align such queries with satellite imagery. To address this, we propose Open-SAT, a training-free query embedding refinement algorithm that operates at inference time to improve alignment between user queries and satellite image content. Open-SAT uses VLMs to compute embeddings for image tiles, which are stored in a vector database for efficient retrieval. At query time, it leverages Large Language Models (LLMs) to refine the text embeddings by incorporating contextual information about objects of interest and their surroundings. A threshold-free retrieval mechanism further enhances accuracy and efficiency. Experimental results in three public benchmarks demonstrate that Open-SAT improves the F1 score by up to 16.04%, while retrieving a comparable number of image tiles. These results demonstrate the effectiveness of Open-SAT in open-vocabulary satellite image retrieval, leveraging LLM guidance without the need for additional training or supervision.\n\n---\n\n## 143. Feature Starvation as Geometric Instability in Sparse Autoencoders\n- **Authors**: Faris Chaudhry , Keisuke Yano , Anthea Monod\n- **URL**: [https://arxiv.org/abs/2605.05341](https://arxiv.org/abs/2605.05341)\n- **Abstract**:\n> Sparse autoencoders (SAEs) are used to disentangle the dense, polysemantic internal representations of large language models (LLMs) into interpretable, monosemantic concepts. However, standard $\\ell_1$-regularized SAEs suffer from feature starvation (dead neurons) and shrinkage bias, often requiring computationally expensive heuristic resampling and nondifferentiable hard-masking methods to bypass these challenges. We argue that feature starvation is not merely an empirical artifact of poor data diversity, but a fundamental optimization-geometric pathology of overcomplete dictionaries: the $\\ell_1$-induced sparse coding map is unstable and fundamentally misaligned with shallow, amortized encoders. To address this structural instability, we introduce adaptive elastic net SAEs (AEN-SAEs), a fully differentiable architecture grounded in classical sparse regression. AEN-SAEs combine an $\\ell_2$ structural term that enforces strong convexity and Lipschitz stability with adaptive $\\ell_1$ reweighting that eliminates shrinkage bias and suppresses spurious features, thereby jointly controlling the curvature and interaction structure of the induced polyhedral geometry. Theoretically, we show that AEN-SAEs yield a Lipschitz-continuous sparse coding map and recover the global feature support under mild assumptions. Empirically, across synthetic settings and LLMs (Pythia 70M, Llama 3.1 8B), AEN-SAEs mitigate feature starvation without auxiliary heuristics while maintaining competitive reconstruction abilities.\n\n---\n\n## 144. How Far Are VLMs from Privacy Awareness in the Physical World? An Empirical Study\n- **Authors**: Junran Wang , Xinjie Shen , Zehao Jin , Pan Li\n- **URL**: [https://arxiv.org/abs/2605.05340](https://arxiv.org/abs/2605.05340)\n- **Abstract**:\n> As Vision-Language Models (VLMs) are increasingly deployed as autonomous cognitive cores for embodied assistants, evaluating their privacy awareness in physical environments becomes critical. Unlike digital chatbots, these agents operate in intimate spaces, such as homes and hospitals, where they possess the physical agency to observe and manipulate privacy-sensitive information and artifacts. However, current benchmarks remain limited to unimodal, text-based representations that cannot capture the demands of real-world settings. To bridge this gap, we present ImmersedPrivacy, an interactive audio-visual evaluation framework that simulates realistic physical environments using a Unity-based simulator. ImmersedPrivacy evaluates physically grounded privacy awareness across three progressive tiers that test a model's ability to identify sensitive items in cluttered scenes, adapt to shifting social contexts, and resolve conflicts between explicit commands and inferred privacy constraints. Our evaluation of 12 state-of-the-art models reveals consistent deficits. In cluttered scenes, all models exhibit monotonic performance decay as scene complexity grows due to perceptual deficit. When social context shifts, no model exceed 65% selection accuracy. Under conflicting commands, the best model gemini-3.1-pro perfectly balances task completion and privacy preservation in only 51% of cases. These findings reveal that current VLMs in the physical world suffer from perceptual fragility and fail to let their knowledge of privacy cues govern their situated behavior. Our code and data is available at this https URL .\n\n---\n\n## 145. Bridging Generation and Training: A Systematic Review of Quality Issues in LLMs for Code\n- **Authors**: Kaifeng He , Xiaojun Zhang , Peiliang Cai , Mingwei Liu , Yanlin Wang , Chong Wang , Kaifeng Huang , Bihuan Chen , Xin Peng , Zibin Zheng\n- **URL**: [https://arxiv.org/abs/2605.05267](https://arxiv.org/abs/2605.05267)\n- **Abstract**:\n> Large language models (LLMs) frequently generate defective outputs in code generation tasks, ranging from logical bugs to security vulnerabilities. While these generation failures are often treated as model-level limitations, empirical evidence increasingly traces their root causes to imperfections within the training corpora. Yet, the specific mechanisms linking training data quality issues to generated code quality issues remain largely unmapped. This paper presents a systematic literature review of 114 primary studies to investigate how training data quality issues propagate into code generation. We establish a unified taxonomy that categorizes generated code quality issues across nine dimensions and training data quality issues into code and non-code attributes. Based on this taxonomy, we formalize a causal framework detailing 18 typical propagation mapping mechanisms. Furthermore, we synthesize state-of-the-art detection and mitigation techniques across the data, model, and generation lifecycles. The reviewed literature reveals a clear methodological shift: quality assurance is transitioning from reactive, heuristic-based post-generation filtering toward proactive, data-centric governance and closed-loop repair. Finally, we identify open challenges and outline research directions for developing reliable LLMs for code through integrated data curation and continuous evaluation. Our repository is available at this https URL .\n\n---\n\n## 146. Decision-aware User Simulation Agent for Evaluating Conversational Recommender Systems\n- **Authors**: Yuan-Chi Li , Li-Chi Chen , Sung-Yi Wu , Yu-Che Tsai , Shou-De Lin\n- **URL**: [https://arxiv.org/abs/2605.05250](https://arxiv.org/abs/2605.05250)\n- **Abstract**:\n> Conversational recommender systems (CRS) increasingly rely on user simulators for automated evaluation of sales agents. A key requirement for such simulators is the ability to model human decision-making. However, most existing simulation frameworks do not explicitly model the internal decision process, and LLM-based simulators often exhibit unrealistically strong information-processing capabilities, rarely exhibit the hesitation or decision deferral commonly observed in real consumer behavior, resulting in overly high acceptance probabilities. To address this limitation, we propose Hesitator, a theory-grounded user simulation framework that explicitly models human decision-making under choice overload. The framework introduces a modular Decision Module that separates utility-based item selection from overload-aware commitment decisions. Experiments across multiple user simulation frameworks, domains, sales modes, and LLM backbones show that integrating our module consistently mitigates unrealistic behaviors under increasing overload conditions. Furthermore, Hesitator reproduces established behavioral patterns from psychological economics, demonstrating its ability to model human decision behavior.\n\n---\n\n## 147. Towards Dependable Retrieval-Augmented Generation Using Factual Confidence Prediction\n- **Authors**: Florian Geissler , Francesco Carella , Laura Fieback , Jakob Spiegelberg\n- **URL**: [https://arxiv.org/abs/2605.05244](https://arxiv.org/abs/2605.05244)\n- **Abstract**:\n> Incorporating specific knowledge into large language models via retrieval-augmented generation (RAG) is a widespread technique that fuels many of today's industry AI applications. A fundamental problem is to assess if the context retrieved by some similarity search provides indeed supporting facts, or instead misguides the generator with irrelevant information. It is critical to associate meaningful confidence measures about the factuality of the retrieval process with the generated answers. We present a new, two-staged approach to predict fact faithfulness of the output of retrieval-augmented generations. First, we employ conformal prediction to select only those retrieved chunks who have a high chance to come from the correct source. This approach in itself can improve answer quality by up to 6% in some of the studied datasets, however, the associated statistical guarantees do not hold generally, since the assumption of sample exchangeability depends on the retriever setup. We present diagnostic metrics to assess whether a setup is suitable. Second, we quantify confidence in the consistency of a generated final answer with a given retrieved context, using an attention-based factuality classifier. This approach can detect inconsistent answers with a chance of up to 77%. Our work helps to establish a novel type of certified RAG systems for a broad range of natural language industry applications.\n\n---\n\n## 148. Rethinking Data Curation in LLM Training: Online Reweighting Offers Better Generalization than Offline Methods\n- **Authors**: Wanru Zhao , Yihong Chen , Yuzhi Tang , Wentao Ma , Shengchao Hu , Shell Xu Hu , Alex Iacob , Abhinav Mehrotra , Nicholas D. Lane\n- **URL**: [https://arxiv.org/abs/2605.05227](https://arxiv.org/abs/2605.05227)\n- **Abstract**:\n> Data curation is a critical yet under-explored area in large language model (LLM) training. Existing methods, such as data selection and mixing, operate in an offline paradigm, detaching themselves from training. This separation introduces engineering overhead and makes the curation brittle: the entire pipeline must be re-run under model/task shifts. Moreover, offline methods alter data size through hard filtering or resampling, often sacrificing data diversity and harming generalization. We propose to rethink data curation as an online reweighting problem, where sample importance is dynamically adjusted during training via loss weighting rather than static pre-processing. Specifically, we introduce ADAPT (Adaptive Data reweighting for Pretraining and FineTuning), a dynamic online framework that reweights training samples with adaptive per-sample learning rates guided by similarity-based quality signals, without changing the number of training samples. Unlike offline methods that enforce a static data distribution, ADAPT acts as an implicit curriculum learner, progressively shifting focus from coarse-grained patterns to fine-grained semantic distinctions as the model evolves. Experiments on both instruction tuning and large-scale pretraining show that ADAPT consistently outperforms offline selection/mixing and prior online methods, achieving stronger cross-benchmark generalization under equal FLOPs.\n\n---\n\n## 149. MACS: Modality-Aware Capacity Scaling for Efficient Multimodal MoE Inference\n- **Authors**: Bo Li , Chuan Wu , shaolin Zhu\n- **URL**: [https://arxiv.org/abs/2605.05225](https://arxiv.org/abs/2605.05225)\n- **Abstract**:\n> Mixture-of-Experts Multimodal Large Language Models (MoE MLLMs) suffer from a significant efficiency bottleneck during Expert Parallelism (EP) inference due to the straggler effect. This issue is worsened in the multimodal context, as existing token-count-based load balancing methods fail to address two unique challenges: (1) Information Heterogeneity, where numerous redundant visual tokens are treated equally to semantically critical ones, and (2) Modality Dynamics, where varying visual to text ratios across tasks lead to resource misallocation. To address these challenges, we propose MACS (Modality-Aware Capacity Scaling), a training-free inference framework. Specifically, MACS introduces an Entropy-Weighted Load mechanism to quantify the semantic value of visual tokens, addressing information heterogeneity. Additionally, the Dynamic Modality-Adaptive Capacity mechanism allocates expert resources based on the real-time modal composition of the input. Extensive experiments demonstrate that MACS significantly outperforms existing methods on various multimodal benchmarks, providing a novel and robust solution for the efficient deployment of MoE MLLMs in EP inference.\n\n---\n\n## 150. MidSteer: Optimal Affine Framework for Steering Generative Models\n- **Authors**: Tatiana Gaintseva , Andrew Stepanov , Ziquan Liu , Martin Benning , Gregory Slabaugh , Jiankang Deng , Ismail Elezi\n- **URL**: [https://arxiv.org/abs/2605.05220](https://arxiv.org/abs/2605.05220)\n- **Abstract**:\n> Steering intermediate representations has emerged as a powerful strategy for controlling generative models, particularly in post-deployment alignment and safety settings. However, despite its empirical success, it currently lacks a comprehensive theoretical framework. In this paper, we bridge this gap by formalizing the theory of concept steering. First, we establish a link between steering and affine concept erasure, proving that the standard approach for removing unwanted behaviors is a special case of LEACE (a closed-form method for affine erasure). Next, we formulate a principled theoretical framework for concept switching, LEACE-Switch, and characterize the assumptions under which it provides an optimal affine solution. Building on this analysis, we then introduce MidSteer (Minimal Disturbance concept Steering), a more general affine framework for concept manipulation that relaxes these assumptions and enables directed, minimal-disturbance transformations. We demonstrate that MidSteer performs favorably across a range of tasks, modalities, and architectures, including vision diffusion models and large language models.\n\n---\n\n## 151. Sparse Prefix Caching for Hybrid and Recurrent LLM Serving\n- **Authors**: Mikhail Shirokikh , Sergey Nikolenko\n- **URL**: [https://arxiv.org/abs/2605.05219](https://arxiv.org/abs/2605.05219)\n- **Abstract**:\n> Prefix caching is a key latency optimization for autoregressive LLM serving, yet existing systems assume dense per-token key/value reuse. State-space models change the structure of the problem: a recurrent layer can resume from a single stored state rather than requiring the entire token history. This asymmetry opens a new design point between no reuse and dense caching: store exact recurrent states at a sparse set of checkpoint positions and, on a cache hit, resume from the deepest stored checkpoint and recompute the remaining suffix exactly. We formalize sparse prefix caching as checkpoint placement under a distribution over overlap depths, yielding an exact O(NM) dynamic program. For use cases where requests share a non-trivial prefix (e.g. asking different questions about a single long document), we show that our method consistently improves the Pareto frontier traced by standard heuristics on real-world data. Across QuALITY and System Prompts, distribution-aware placement dominates every fixed-budget baseline on the measured layer-group Pareto frontier and matches or outperforms the strongest heuristic (block caching) while typically using substantially fewer checkpoints, with the largest gains at low checkpoint budgets where the overlap distribution is most non-uniform. The method is most relevant when many requests share a substantial but not identical prefix within a retained cache entry. It preserves exact outputs, does not change the recurrent computation itself or require new recurrent update kernels, applies to recurrent/SSM layers whose hidden state can be extracted and restored exactly, and for hybrid models can be combined with existing KV-cache compression techniques.\n\n---\n\n## 152. A Review of Large Language Models for Stock Price Forecasting from a Hedge-Fund Perspective\n- **Authors**: Olivia Zhang , Zhilin Zhang\n- **URL**: [https://arxiv.org/abs/2605.05211](https://arxiv.org/abs/2605.05211)\n- **Abstract**:\n> Large language models (LLMs) are increasingly deployed in quantitative finance for stock price forecasting. This review synthesizes recent applications of LLMs in this domain, including extracting sentiment from financial news and social media, analyzing financial reports and earnings-call transcripts, tokenizing or symbolizing stock price series, and constructing multi-agent trading systems. Particular attention is paid to practical pitfalls that are often understated in the literature, such as fragility in sentiment analysis, dataset and horizon design, performance evaluation metrics, data leakage, illiquidity premia, and limits of stock price predictability. Organized from a hedge-fund perspective, the review is intended to guide both academic researchers and hedge fund managers in integrating LLMs into real-world trading pipelines and in stress-testing their robustness under realistic market frictions.\n\n---\n\n{% endraw %}\n"}}, "ALL/2026-03-17.md": {"64bbd1731dde152451afe1705df222e26b5c83fa": {"t": 1785747183.8654895, "n": 229888, "h": "8c5adbd6d55505f6", "c": "{% raw %}\n# 전체 AI 논문 - 2026-03-17\n\n## 1. Semantic Invariance in Agentic AI\n- **Authors**: I. de Zarzà , J. de Curtò , Jordi Cabot , Pietro Manzoni , Carlos T. Calafate\n- **URL**: [https://arxiv.org/abs/2603.13173](https://arxiv.org/abs/2603.13173)\n- **Abstract**:\n> Large Language Models (LLMs) increasingly serve as autonomous reasoning agents in decision support, scientific problem-solving, and multi-agent coordination systems. However, deploying LLM agents in consequential applications requires assurance that their reasoning remains stable under semantically equivalent input variations, a property we term semantic invariance. Standard benchmark evaluations, which assess accuracy on fixed, canonical problem formulations, fail to capture this critical reliability dimension. To address this shortcoming, in this paper we present a metamorphic testing framework for systematically assessing the robustness of LLM reasoning agents, applying eight semantic-preserving transformations (identity, paraphrase, fact reordering, expansion, contraction, academic context, business context, and contrastive formulation) across seven foundation models spanning four distinct architectural families: Hermes (70B, 405B), Qwen3 (30B-A3B, 235B-A22B), DeepSeek-R1, and gpt-oss (20B, 120B). Our evaluation encompasses 19 multi-step reasoning problems across eight scientific domains. The results reveal that model scale does not predict robustness: the smaller Qwen3-30B-A3B achieves the highest stability (79.6% invariant responses, semantic similarity 0.91), while larger models exhibit greater fragility.\n\n---\n\n## 2. Developing and evaluating a chatbot to support maternal health care\n- **Authors**: Smriti Jha , Vidhi Jain , Jianyu Xu , Grace Liu , Sowmya Ramesh , Jitender Nagpal , Gretchen Chapman , Benjamin Bellows , Siddhartha Goyal , Aarti Singh , Bryan Wilder\n- **URL**: [https://arxiv.org/abs/2603.13168](https://arxiv.org/abs/2603.13168)\n- **Abstract**:\n> The ability to provide trustworthy maternal health information using phone-based chatbots can have a significant impact, particularly in low-resource settings where users have low health literacy and limited access to care. However, deploying such systems is technically challenging: user queries are short, underspecified, and code-mixed across languages, answers require regional context-specific grounding, and partial or missing symptom context makes safe routing decisions difficult. We present a chatbot for maternal health in India developed through a partnership between academic researchers, a health tech company, a public health nonprofit, and a hospital. The system combines (1) stage-aware triage, routing high-risk queries to expert templates, (2) hybrid retrieval over curated maternal/newborn guidelines, and (3) evidence-conditioned generation from an LLM. Our core contribution is an evaluation workflow for high-stakes deployment under limited expert supervision. Targeting both component-level and end-to-end testing, we introduce: (i) a labeled triage benchmark (N=150) achieving 86.7% emergency recall, explicitly reporting the missed-emergency vs. over-escalation trade-off; (ii) a synthetic multi-evidence retrieval benchmark (N=100) with chunk-level evidence labels; (iii) LLM-as-judge comparison on real queries (N=781) using clinician-codesigned criteria; and (iv) expert validation. Our findings show that trustworthy medical assistants in multilingual, noisy settings require defense-in-depth design paired with multi-method evaluation, rather than any single model and evaluation method choice.\n\n---\n\n## 3. When Right Meets Wrong: Bilateral Context Conditioning with Reward-Confidence Correction for GRPO\n- **Authors**: Yu Li , Tian Lan , Zhengling Qi\n- **URL**: [https://arxiv.org/abs/2603.13134](https://arxiv.org/abs/2603.13134)\n- **Abstract**:\n> Group Relative Policy Optimization (GRPO) has emerged as an effective method for training reasoning models. While it computes advantages based on group mean, GRPO treats each output as an independent sample during the optimization and overlooks a vital structural signal: the natural contrast between correct and incorrect solutions within the same group, thus ignoring the rich, comparative data that could be leveraged by explicitly pitting successful reasoning traces against failed ones. To capitalize on this, we present a contrastive reformulation of GRPO, showing that the GRPO objective implicitly maximizes the margin between the policy ratios of correct and incorrect samples. Building on this insight, we propose Bilateral Context Conditioning (BICC), a mechanism that allows the model to cross-reference successful and failed reasoning traces during the optimization, enabling a direct information flow across samples. We further introduce Reward-Confidence Correction (RCC) to stabilize training by dynamically adjusts the advantage baseline in GRPO using reward-confidence covariance derived from the first-order approximation of the variance-minimizing estimator. Both mechanisms require no additional sampling or auxiliary models and can be adapted to all GRPO variants. Experiments on mathematical reasoning benchmarks demonstrate consistent improvements across comprehensive models and algorithms. Code is available at \\href{ this https URL }{ this https URL }.\n\n---\n\n## 4. Steve-Evolving: Open-World Embodied Self-Evolution via Fine-Grained Diagnosis and Dual-Track Knowledge Distillation\n- **Authors**: Zhengwei Xie , Zhisheng Chen , Ziyan Weng , Tingyu Wu , Chenglong Li , Vireo Zhang , Kun Wang\n- **URL**: [https://arxiv.org/abs/2603.13131](https://arxiv.org/abs/2603.13131)\n- **Abstract**:\n> Open-world embodied agents must solve long-horizon tasks where the main bottleneck is not single-step planning quality but how interaction experience is organized and evolved. To this end, we present Steve-Evolving, a non-parametric self-evolving framework that tightly couples fine-grained execution diagnosis with dual-track knowledge distillation in a closed loop. The method follows three phases: Experience Anchoring, Experience Distillation, and Knowledge-Driven Closed-Loop Control. In detail, Experience Anchoring solidifies each subgoal attempt into a structured experience tuple with a fixed schema (pre-state, action, diagnosis-result, and post-state) and organizes it in a three-tier experience space with multi-dimensional indices (e.g., condition signatures, spatial hashing, and semantic tags) plus rolling summarization for efficient and auditable recall. To ensure sufficient information density for attribution, the execution layer provides compositional diagnosis signals beyond binary outcomes, including state-difference summaries, enumerated failure causes, continuous indicators, and stagnation/loop detection. Moreover, successful trajectories of Experience Distillation are generalized into reusable skills with explicit preconditions and verification criteria, while failures are distilled into executable guardrails that capture root causes and forbid risky operations at both subgoal and task granularities. Besides, Knowledge-Driven Closed-Loop Control retrieved skills and guardrails are injected into an LLM planner, and diagnosis-triggered local replanning updates the active constraints online, forming a continual evolution process without any model parameter updates. Experiments on the long-horizon suite of Minecraft MCU demonstrate consistent improvements over static-retrieval baselines.\n\n---\n\n## 5. Beyond Final Answers: CRYSTAL Benchmark for Transparent Multimodal Reasoning Evaluation\n- **Authors**: Wayner Barrios , SouYoung Jin\n- **URL**: [https://arxiv.org/abs/2603.13099](https://arxiv.org/abs/2603.13099)\n- **Abstract**:\n> We introduce CRYSTAL (Clear Reasoning via Yielded Steps, Traceability, and Logic), a diagnostic benchmark with 6,372 instances that evaluates multimodal reasoning through verifiable intermediate steps. We propose two complementary metrics: Match F1, which scores step-level precision and recall via semantic similarity matching, and Ordered Match F1, which further penalizes disordered reasoning chains. References are constructed through a Delphi-inspired pipeline in which four independent MLLMs generate trajectories, which are then aggregated via semantic clustering and validated through human quality gates. Evaluation of 20 MLLMs, including commercial frontier systems not used during benchmark construction, reveals systematic failures that are invisible to answer accuracy: universal cherry-picking (precision far exceeds recall), non-monotonic scaling trade-offs, and disordered reasoning in which no competitive model preserves more than 60% of matched steps in the correct order. Beyond evaluation, we propose the Causal Process Reward (CPR), a multiplicative reward that couples answer correctness with step-level alignment, and CPR-Curriculum, which progressively increases reasoning difficulty during training. CPR-Curriculum achieves a 32% improvement in Match F1 via GRPO where additive reward strategies fail, improving reasoning without manual step annotation.\n\n---\n\n## 6. Structured Distillation for Personalized Agent Memory: 11x Token Reduction with Retrieval Preservation\n- **Authors**: Sydney Lewis\n- **URL**: [https://arxiv.org/abs/2603.13017](https://arxiv.org/abs/2603.13017)\n- **Abstract**:\n> Long conversations with an AI agent create a simple problem for one user: the history is useful, but carrying it verbatim is expensive. We study personalized agent memory: one user's conversation history with an agent, distilled into a compact retrieval layer for later search. Each exchange is compressed into a compound object with four fields (exchange_core, specific_context, thematic room_assignments, and regex-extracted files_touched). The searchable distilled text averages 38 tokens per exchange. Applied to 4,182 conversations (14,340 exchanges) from 6 software engineering projects, the method reduces average exchange length from 371 to 38 tokens, yielding 11x compression. We evaluate whether personalized recall survives that compression using 201 recall-oriented queries, 107 configurations spanning 5 pure and 5 cross-layer search modes, and 5 LLM graders (214,519 consensus-graded query-result pairs). The best pure distilled configuration reaches 96% of the best verbatim MRR (0.717 vs 0.745). Results are mechanism-dependent. All 20 vector search configurations remain non-significant after Bonferroni correction, while all 20 BM25 configurations degrade significantly (effect sizes |d|=0.031-0.756). The best cross-layer setup slightly exceeds the best pure verbatim baseline (MRR 0.759). Structured distillation compresses single-user agent memory without uniformly sacrificing retrieval quality. At 1/11 the context cost, thousands of exchanges fit within a single prompt while the verbatim source remains available for drill-down. We release the implementation and analysis pipeline as open-source software.\n\n---\n\n## 7. Efficient and Interpretable Multi-Agent LLM Routing via Ant Colony Optimization\n- **Authors**: Xudong Wang , Chaoning Zhang , Jiaquan Zhang , Chenghao Li , Qigan Sun , Sung-Ho Bae , Peng Wang , Ning Xie , Jie Zou , Yang Yang , Hengtao Shen\n- **URL**: [https://arxiv.org/abs/2603.12933](https://arxiv.org/abs/2603.12933)\n- **Abstract**:\n> Large Language Model (LLM)-driven Multi-Agent Systems (MAS) have demonstrated strong capability in complex reasoning and tool use, and heterogeneous agent pools further broaden the quality--cost trade-off space. Despite these advances, real-world deployment is often constrained by high inference cost, latency, and limited transparency, which hinders scalable and efficient routing. Existing routing strategies typically rely on expensive LLM-based selectors or static policies, and offer limited controllability for semantic-aware routing under dynamic loads and mixed intents, often resulting in unstable performance and inefficient resource utilization. To address these limitations, we propose AMRO-S, an efficient and interpretable routing framework for Multi-Agent Systems (MAS). AMRO-S models MAS routing as a semantic-conditioned path selection problem, enhancing routing performance through three key mechanisms: First, it leverages a supervised fine-tuned (SFT) small language model for intent inference, providing a low-overhead semantic interface for each query; second, it decomposes routing memory into task-specific pheromone specialists, reducing cross-task interference and optimizing path selection under mixed workloads; finally, it employs a quality-gated asynchronous update mechanism to decouple inference from learning, optimizing routing without increasing latency. Extensive experiments on five public benchmarks and high-concurrency stress tests demonstrate that AMRO-S consistently improves the quality--cost trade-off over strong routing baselines, while providing traceable routing evidence through structured pheromone patterns.\n\n---\n\n## 8. ODRL Policy Comparison Through Normalisation\n- **Authors**: Jaime Osvaldo Salas , Paolo Pareti , George Konstantinidis\n- **URL**: [https://arxiv.org/abs/2603.12926](https://arxiv.org/abs/2603.12926)\n- **Abstract**:\n> The ODRL language has become the standard for representing policies and regulations for digital rights. However its complexity is a barrier to its usage, which has caused many related theoretical and practical works to focus on different, and not interoperable, fragments of ODRL. Moreover, semantically equivalent policies can be expressed in numerous different ways, which makes comparing them and processing them harder. Building on top of a recently defined semantics, we tackle these problems by proposing an approach that involves a parametrised normalisation of ODRL policies into its minimal components which reformulates policies with permissions and prohibitions into policies with permissions exclusively, and simplifies complex logic constraints into simple ones. We provide algorithms to compute a normal form for ODRL policies and simplifying numerical and symbolic constraints. We prove that these algorithms preserve the semantics of policies, and analyse the size complexity of the result, which is exponential on the number of attributes and linear on the number of unique values for these attributes. We show how this makes complex policies representable in more basic fragments of ODRL, and how it reduces the problem of policy comparison to the simpler problem of checking if two rules are identical.\n\n---\n\n## 9. Context is all you need: Towards autonomous model-based process design using agentic AI in flowsheet simulations\n- **Authors**: Pascal Schäfer , Lukas J. Krinke , Martin Wlotzka , Norbert Asprion\n- **URL**: [https://arxiv.org/abs/2603.12813](https://arxiv.org/abs/2603.12813)\n- **Abstract**:\n> Agentic AI systems integrating large language models (LLMs) with reasoning and tooluse capabilities are transforming various domains - in particular, software development. In contrast, their application in chemical process flowsheet modelling remains largely unexplored. In this work, we present an agentic AI framework that delivers assistance in an industrial flowsheet simulation environment. To this end, we show the capabilities of GitHub Copilot (GitHub, Inc., 2026), when using state-of-the-art LLMs, such as Claude Opus 4.6 (Anthropic, PBC, 2026), to generate valid syntax for our in-house process modelling tool Chemasim using the technical documentation and a few commented examples as context. Based on this, we develop a multi-agent system that decomposes process development tasks with one agent solving the abstract problem using engineering knowledge and another agent implementing the solution as Chemasim code. We demonstrate the effectiveness of our framework for typical flowsheet modelling examples, including (i) a reaction/separation process, (ii) a pressure-swing distillation, and (iii) a heteroazeotropic distillation including entrainer selection. Along these lines, we discuss current limitations of the framework and outline future research directions to further enhance its capabilities.\n\n---\n\n## 10. AI Model Modulation with Logits Redistribution\n- **Authors**: Zihan Wang , Zhongkui Ma , Xinguo Feng , Zhiyang Mei , Ethan Ma , Derui Wang , Minhui Xue , Guangdong Bai\n- **URL**: [https://arxiv.org/abs/2603.12755](https://arxiv.org/abs/2603.12755)\n- **Abstract**:\n> Large-scale models are typically adapted to meet the diverse requirements of model owners and users. However, maintaining multiple specialized versions of the model is inefficient. In response, we propose AIM, a novel model modulation paradigm that enables a single model to exhibit diverse behaviors to meet the specific end requirements. AIM enables two key modulation modes: utility and focus modulations. The former provides model owners with dynamic control over output quality to deliver varying utility levels, and the latter offers users precise control to shift model's focused input features. AIM introduces a logits redistribution strategy that operates in a training data-agnostic and retraining-free manner. We establish a formal foundation to ensure AIM's regulation capability, based on the statistical properties of logits ordering via joint probability distributions. Our evaluation confirms AIM's practicality and versatility for Al model modulation, with tasks spanning image classification, semantic segmentation and text generation, and prevalent architectures including ResNet, SegFormer and Llama.\n\n---\n\n## 11. ToolTree: Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning\n- **Authors**: Shuo Yang , Soyeon Caren Han , Yihao Ding , Shuhe Wang , Eduard Hoy\n- **URL**: [https://arxiv.org/abs/2603.12740](https://arxiv.org/abs/2603.12740)\n- **Abstract**:\n> Large Language Model (LLM) agents are increasingly applied to complex, multi-step tasks that require interaction with diverse external tools across various domains. However, current LLM agent tool planning methods typically rely on greedy, reactive tool selection strategies that lack foresight and fail to account for inter-tool dependencies. In this paper, we present ToolTree, a novel Monte Carlo tree search-inspired planning paradigm for tool planning. ToolTree explores possible tool usage trajectories using a dual-stage LLM evaluation and bidirectional pruning mechanism that enables the agent to make informed, adaptive decisions over extended tool-use sequences while pruning less promising branches before and after the tool execution. Empirical evaluations across both open-set and closed-set tool planning tasks on 4 benchmarks demonstrate that ToolTree consistently improves performance while keeping the highest efficiency, achieving an average gain of around 10\\% compared to the state-of-the-art planning paradigm.\n\n---\n\n## 12. On Using Machine Learning to Early Detect Catastrophic Failures in Marine Diesel Engines\n- **Authors**: Francesco Maione , Paolo Lino , Giuseppe Giannino , Guido Maione\n- **URL**: [https://arxiv.org/abs/2603.12733](https://arxiv.org/abs/2603.12733)\n- **Abstract**:\n> Catastrophic failures of marine engines imply severe loss of functionality and destroy or damage the systems irreversibly. Being sudden and often unpredictable events, they pose a severe threat to navigation, crew, and passengers. The abrupt nature makes early detection the only effective countermeasure. However, research has concentrated on modeling the gradual degradation of components, with limited attention to sudden and anomalous phenomena. This work proposes a new method for early detection of catastrophic failures. Based on real data from a failed engine, the approach evaluates the derivatives of the deviation between actual sensor readings and expected values of engine variables. Predictions are obtained by a Random Forest, which is the most suitable Machine Learning algorithm among the tested ones. Traditional methods focus on deviations of monitored signals, whereas the proposed approach employs the derivatives of the deviations to provide earlier indications of abnormal dynamics, and to alert that a rapid and dangerous event is breaking out within the system. The method allows the detection of anomalies before measurements reach critical thresholds and alarms are triggered, which is the common method in industry. Consequently, operators can be warned in advance and shut down the engine, then prevent damage and unexpected power loss. Moreover, they have the time to safely change the ship route and avoid potential obstacles. Simulation results conf irm the effectiveness of the proposed approach in anticipating occurrence of catastrophic failures. Validation on real-world data further reinforces the robustness and practical applicability of the method. It is worth noting that data acquisition to train the predictive algorithm is not a problem, since a Deep Learning-based data augmentation procedure is used.\n\n---\n\n## 13. AI Planning Framework for LLM-Based Web Agents\n- **Authors**: Orit Shahnovsky , Rotem Dror\n- **URL**: [https://arxiv.org/abs/2603.12710](https://arxiv.org/abs/2603.12710)\n- **Abstract**:\n> Developing autonomous agents for web-based tasks is a core challenge in AI. While Large Language Model (LLM) agents can interpret complex user requests, they often operate as black boxes, making it difficult to diagnose why they fail or how they plan. This paper addresses this gap by formally treating web tasks as sequential decision-making processes. We introduce a taxonomy that maps modern agent architectures to traditional planning paradigms: Step-by-Step agents to Breadth-First Search (BFS), Tree Search agents to Best-First Tree Search, and Full-Plan-in-Advance agents to Depth-First Search (DFS). This framework allows for a principled diagnosis of system failures like context drift and incoherent task decomposition. To evaluate these behaviors, we propose five novel evaluation metrics that assess trajectory quality beyond simple success rates. We support this analysis with a new dataset of 794 human-labeled trajectories from the WebArena benchmark. Finally, we validate our evaluation framework by comparing a baseline Step-by-Step agent against a novel Full-Plan-in-Advance implementation. Our results reveal that while the Step-by-Step agent aligns more closely with human gold trajectories (38% overall success), the Full-Plan-in-Advance agent excels in technical measures such as element accuracy (89%), demonstrating the necessity of our proposed metrics for selecting appropriate agent architectures based on specific application constraints.\n\n---\n\n## 14. Generating Expressive and Customizable Evals for Timeseries Data Analysis Agents with AgentFuel\n- **Authors**: Aadyaa Maddi , Prakhar Naval , Deepti Mande , Shane Duan , Muckai Girish , Vyas Sekar\n- **URL**: [https://arxiv.org/abs/2603.12483](https://arxiv.org/abs/2603.12483)\n- **Abstract**:\n> Across many domains (e.g., IoT, observability, telecommunications, cybersecurity), there is an emerging adoption of conversational data analysis agents that enable users to \"talk to your data\" to extract insights. Such data analysis agents operate on timeseries data models; e.g., measurements from sensors or events monitoring user clicks and actions in product analytics. We evaluate 6 popular data analysis agents (both open-source and proprietary) on domain-specific data and query types, and find that they fail on stateful and incident-specific queries. We observe two key expressivity gaps in existing evals: domain-customized datasets and domain-specific query types. To enable practitioners in such domains to generate customized and expressive evals for such timeseries data agents, we present AgentFuel. AgentFuel helps domain experts quickly create customized evals to perform end-to-end functional tests. We show that AgentFuel's benchmarks expose key directions for improvement in existing data agent frameworks. We also present anecdotal evidence that using AgentFuel can improve agent performance (e.g., with GEPA). AgentFuel benchmarks are available at this https URL .\n\n---\n\n## 15. Efficient Reasoning with Balanced Thinking\n- **Authors**: Yulin Li , Tengyao Tu , Li Ding , Junjie Wang , Huiling Zhen , Yixin Chen , Yong Li , Zhuotao Tian\n- **URL**: [https://arxiv.org/abs/2603.12372](https://arxiv.org/abs/2603.12372)\n- **Abstract**:\n> Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they often suffer from overthinking, expending redundant computational steps on simple problems, or underthinking, failing to explore sufficient reasoning paths despite inherent capabilities. These issues lead to inefficiencies and potential inaccuracies, limiting practical deployment in resource-constrained settings. Existing methods to mitigate overthinking, such as suppressing reflective keywords or adjusting reasoning length, may inadvertently induce underthinking, compromising accuracy. Therefore, we propose ReBalance, a training-free framework that achieves efficient reasoning with balanced thinking. ReBalance leverages confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence. By aggregating hidden states from a small-scale dataset into reasoning mode prototypes, we compute a steering vector to guide LRMs' reasoning trajectories. A dynamic control function modulates this vector's strength and direction based on real-time confidence, pruning redundancy during overthinking, and promoting exploration during underthinking. Extensive experiments conducted on four models ranging from 0.5B to 32B, and across nine benchmarks in math reasoning, general question answering, and coding tasks demonstrate that ReBalance effectively reduces output redundancy while improving accuracy, offering a general, training-free, and plug-and-play strategy for efficient and robust LRM deployment. Code is available at this https URL .\n\n---\n\n## 16. Context-Enriched Natural Language Descriptions of Vessel Trajectories\n- **Authors**: Kostas Patroumpas , Alexandros Troupiotis-Kapeliaris , Giannis Spiliopoulos , Panagiotis Betchavas , Dimitrios Skoutas , Dimitris Zissis , Nikos Bikakis\n- **URL**: [https://arxiv.org/abs/2603.12287](https://arxiv.org/abs/2603.12287)\n- **Abstract**:\n> We address the problem of transforming raw vessel trajectory data collected from AIS into structured and semantically enriched representations interpretable by humans and directly usable by machine reasoning systems. We propose a context-aware trajectory abstraction framework that segments noisy AIS sequences into distinct trips each consisting of clean, mobility-annotated episodes. Each episode is further enriched with multi-source contextual information, such as nearby geographic entities, offshore navigation features, and weather conditions. Crucially, such representations can support generation of controlled natural language descriptions using LLMs. We empirically examine the quality of such descriptions generated using several LLMs over AIS data along with open contextual features. By increasing semantic density and reducing spatiotemporal complexity, this abstraction can facilitate downstream analytics and enable integration with LLMs for higher-level maritime reasoning tasks.\n\n---\n\n## 17. PhysMoDPO: Physically-Plausible Humanoid Motion with Preference Optimization\n- **Authors**: Yangsong Zhang , Anujith Muraleedharan , Rikhat Akizhanov , Abdul Ahad Butt , Gül Varol , Pascal Fua , Fabio Pizzati , Ivan Laptev\n- **URL**: [https://arxiv.org/abs/2603.13228](https://arxiv.org/abs/2603.13228)\n- **Abstract**:\n> Recent progress in text-conditioned human motion generation has been largely driven by diffusion models trained on large-scale human motion data. Building on this progress, recent methods attempt to transfer such models for character animation and real robot control by applying a Whole-Body Controller (WBC) that converts diffusion-generated motions into executable trajectories. While WBC trajectories become compliant with physics, they may expose substantial deviations from original motion. To address this issue, we here propose PhysMoDPO, a Direct Preference Optimization framework. Unlike prior work that relies on hand-crafted physics-aware heuristics such as foot-sliding penalties, we integrate WBC into our training pipeline and optimize diffusion model such that the output of WBC becomes compliant both with physics and original text instructions. To train PhysMoDPO we deploy physics-based and task-specific rewards and use them to assign preference to synthesized trajectories. Our extensive experiments on text-to-motion and spatial control tasks demonstrate consistent improvements of PhysMoDPO in both physical realism and task-related metrics on simulated robots. Moreover, we demonstrate that PhysMoDPO results in significant improvements when applied to zero-shot motion transfer in simulation and for real-world deployment on a G1 humanoid robot.\n\n---\n\n## 18. Visual-ERM: Reward Modeling for Visual Equivalence\n- **Authors**: Ziyu Liu , Shengyuan Ding , Xinyu Fang , Xuanlang Dai , Penghui Yang , Jianze Liang , Jiaqi Wang , Kai Chen , Dahua Lin , Yuhang Zang\n- **URL**: [https://arxiv.org/abs/2603.13224](https://arxiv.org/abs/2603.13224)\n- **Abstract**:\n> Vision-to-code tasks require models to reconstruct structured visual inputs, such as charts, tables, and SVGs, into executable or structured representations with high visual fidelity. While recent Large Vision Language Models (LVLMs) achieve strong results via supervised fine-tuning, reinforcement learning remains challenging due to misaligned reward signals. Existing rewards either rely on textual rules or coarse visual embedding similarity, both of which fail to capture fine-grained visual discrepancies and are vulnerable to reward hacking. We propose Visual Equivalence Reward Model (Visual-ERM), a multimodal generative reward model that provides fine-grained, interpretable, and task-agnostic feedback to evaluate vision-to-code quality directly in the rendered visual space. Integrated into RL, Visual-ERM improves Qwen3-VL-8B-Instruct by +8.4 on chart-to-code and yields consistent gains on table and SVG parsing (+2.7, +4.1 on average), and further strengthens test-time scaling via reflection and revision. We also introduce VisualCritic-RewardBench (VC-RewardBench), a benchmark for judging fine-grained image-to-image discrepancies on structured visual data, where Visual-ERM at 8B decisively outperforms Qwen3-VL-235B-Instruct and approaches leading closed-source models. Our results suggest that fine-grained visual reward supervision is both necessary and sufficient for vision-to-code RL, regardless of task specificity.\n\n---\n\n## 19. From Experiments to Expertise: Scientific Knowledge Consolidation for AI-Driven Computational Research\n- **Authors**: Haonan Huang\n- **URL**: [https://arxiv.org/abs/2603.13191](https://arxiv.org/abs/2603.13191)\n- **Abstract**:\n> While large language models (LLMs) have transformed AI agents into proficient executors of computational materials science, performing a hundred simulations does not make a researcher. What distinguishes research from routine execution is the progressive accumulation of knowledge -- learning which approaches fail, recognizing patterns across systems, and applying understanding to new problems. However, the prevailing paradigm in AI-driven computational science treats each execution in isolation, largely discarding hard-won insights between runs. Here we present QMatSuite, an open-source platform closing this gap. Agents record findings with full provenance, retrieve knowledge before new calculations, and in dedicated reflection sessions correct erroneous findings and synthesize observations into cross-compound patterns. In benchmarks on a six-step quantum-mechanical simulation workflow, accumulated knowledge reduces reasoning overhead by 67% and improves accuracy from 47% to 3% deviation from literature -- and when transferred to an unfamiliar material, achieves 1% deviation with zero pipeline failures.\n\n---\n\n## 20. LLM Constitutional Multi-Agent Governance\n- **Authors**: J. de Curtò , I. de Zarzà\n- **URL**: [https://arxiv.org/abs/2603.13189](https://arxiv.org/abs/2603.13189)\n- **Abstract**:\n> Large Language Models (LLMs) can generate persuasive influence strategies that shift cooperative behavior in multi-agent populations, but a critical question remains: does the resulting cooperation reflect genuine prosocial alignment, or does it mask erosion of agent autonomy, epistemic integrity, and distributional fairness? We introduce Constitutional Multi-Agent Governance (CMAG), a two-stage framework that interposes between an LLM policy compiler and a networked agent population, combining hard constraint filtering with soft penalized-utility optimization that balances cooperation potential against manipulation risk and autonomy pressure. We propose the Ethical Cooperation Score (ECS), a multiplicative composite of cooperation, autonomy, integrity, and fairness that penalizes cooperation achieved through manipulative means. In experiments on scale-free networks of 80 agents under adversarial conditions (70% violating candidates), we benchmark three regimes: full CMAG, naive filtering, and unconstrained optimization. While unconstrained optimization achieves the highest raw cooperation (0.873), it yields the lowest ECS (0.645) due to severe autonomy erosion (0.867) and fairness degradation (0.888). CMAG attains an ECS of 0.741, a 14.9% improvement, while preserving autonomy at 0.985 and integrity at 0.995, with only modest cooperation reduction to 0.770. The naive ablation (ECS = 0.733) confirms that hard constraints alone are insufficient. Pareto analysis shows CMAG dominates the cooperation-autonomy trade-off space, and governance reduces hub-periphery exposure disparities by over 60%. These findings establish that cooperation is not inherently desirable without governance: constitutional constraints are necessary to ensure that LLM-mediated influence produces ethically stable outcomes rather than manipulative equilibria.\n\n---\n\n## 21. Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights\n- **Authors**: Xingli Fang , Jung-Eun Kim\n- **URL**: [https://arxiv.org/abs/2603.13186](https://arxiv.org/abs/2603.13186)\n- **Abstract**:\n> Prior approaches for membership privacy preservation usually update or retrain all weights in neural networks, which is costly and can lead to unnecessary utility loss or even more serious misalignment in predictions between training data and non-training data. In this work, we observed three insights: i) privacy vulnerability exists in a very small fraction of weights; ii) however, most of those weights also critically impact utility performance; iii) the importance of weights stems from their locations rather than their values. According to these insights, to preserve privacy, we score critical weights, and instead of discarding those neurons, we rewind only the weights for fine-tuning. We show that, through extensive experiments, this mechanism exhibits outperforming resilience in most cases against Membership Inference Attacks while maintaining utility.\n\n---\n\n## 22. MXNorm: Reusing MXFP block scales for efficient tensor normalisation\n- **Authors**: Callum McLean , Luke Y. Prince , Alexandre Payot , Paul Balança , Carlo Luschi\n- **URL**: [https://arxiv.org/abs/2603.13180](https://arxiv.org/abs/2603.13180)\n- **Abstract**:\n> Matrix multiplication performance has long been the major bottleneck to scaling deep learning workloads, which has stimulated the design of new accelerators that use increasingly low-precision number formats. However, improvements in matrix multiplication performance have far outstripped improvements in performance on reductions and elementwise computations, which are still being performed in higher precision. In this work, we propose MXNorm, a drop-in replacement for RMSNorm that estimates the RMS using only the block scales calculated as part of the MXFP8 cast and enables a 32x decrease in the size of reduction needed for normalization. We validate our approximation method on pre-training of Llama 3 models of 125M, 1B and 8B parameters, finding minimal loss of training accuracy compared to a baseline using RMSNorm with MXFP8 matmuls. We also show practical kernel speedups using only this http URL of up to 2.4x for MXNorm over RMSNorm, corresponding to a 1.3% speedup in Llama 3 8B transformer layers in MXFP8 and a 2.6% speedup in NVFP4.\n\n---\n\n## 23. Clustering Astronomical Orbital Synthetic Data Using Advanced Feature Extraction and Dimensionality Reduction Techniques\n- **Authors**: Eraldo Pereira Marinho , Nelson Callegari Junior , Fabricio Aparecido Breve , Caetano Mazzoni Ranieri\n- **URL**: [https://arxiv.org/abs/2603.13177](https://arxiv.org/abs/2603.13177)\n- **Abstract**:\n> The dynamics of Saturn's satellite system offer a rich framework for studying orbital stability and resonance interactions. Traditional methods for analysing such systems, including Fourier analysis and stability metrics, struggle with the scale and complexity of modern datasets. This study introduces a machine learning-based pipeline for clustering approximately 22,300 simulated satellite orbits, addressing these challenges with advanced feature extraction and dimensionality reduction techniques. The key to this approach is using MiniRocket, which efficiently transforms 400 timesteps into a 9,996-dimensional feature space, capturing intricate temporal patterns. Additional automated feature extraction and dimensionality reduction techniques refine the data, enabling robust clustering analysis. This pipeline reveals stability regions, resonance structures, and other key behaviours in Saturn's satellite system, providing new insights into their long-term dynamical evolution. By integrating computational tools with traditional celestial mechanics techniques, this study offers a scalable and interpretable methodology for analysing large-scale orbital datasets and advancing the exploration of planetary dynamics.\n\n---\n\n## 24. ESG-Bench: Benchmarking Long-Context ESG Reports for Hallucination Mitigation\n- **Authors**: Siqi Sun , Ben Peng Wu , Mali Jin , Peizhen Bai , Hanpei Zhang , Xingyi Song\n- **URL**: [https://arxiv.org/abs/2603.13154](https://arxiv.org/abs/2603.13154)\n- **Abstract**:\n> As corporate responsibility increasingly incorporates environmental, social, and governance (ESG) criteria, ESG reporting is becoming a legal requirement in many regions and a key channel for documenting sustainability practices and assessing firms' long-term and ethical performance. However, the length and complexity of ESG disclosures make them difficult to interpret and automate the analysis reliably. To support scalable and trustworthy analysis, this paper introduces ESG-Bench, a benchmark dataset for ESG report understanding and hallucination mitigation in large language models (LLMs). ESG-Bench contains human-annotated question-answer (QA) pairs grounded in real-world ESG report contexts, with fine-grained labels indicating whether model outputs are factually supported or hallucinated. Framing ESG report analysis as a QA task with verifiability constraints enables systematic evaluation of LLMs' ability to extract and reason over ESG content and provides a new use case: mitigating hallucinations in socially sensitive, compliance-critical settings. We design task-specific Chain-of-Thought (CoT) prompting strategies and fine-tune multiple state-of-the-art LLMs on ESG-Bench using CoT-annotated rationales. Our experiments show that these CoT-based methods substantially outperform standard prompting and direct fine-tuning in reducing hallucinations, and that the gains transfer to existing QA benchmarks beyond the ESG domain.\n\n---\n\n## 25. Developing the PsyCogMetrics AI Lab to Evaluate Large Language Models and Advance Cognitive Science -- A Three-Cycle Action Design Science Study\n- **Authors**: Zhiye Jin , Yibai Li , K. D. Joshi , Xuefei (Nancy) Deng , Xiaobing (Emily)Li\n- **URL**: [https://arxiv.org/abs/2603.13126](https://arxiv.org/abs/2603.13126)\n- **Abstract**:\n> This study presents the development of the PsyCogMetrics AI Lab ( this http URL ), an integrated, cloud-based platform that operationalizes psychometric and cognitive-science methodologies for Large Language Model (LLM) evaluation. Framed as a three-cycle Action Design Science study, the Relevance Cycle identifies key limitations in current evaluation methods and unfulfilled stakeholder needs. The Rigor Cycle draws on kernel theories such as Popperian falsifiability, Classical Test Theory, and Cognitive Load Theory to derive deductive design objectives. The Design Cycle operationalizes these objectives through nested Build-Intervene-Evaluate loops. The study contributes a novel IT artifact, a validated design for LLM evaluation, benefiting research at the intersection of AI, psychology, cognitive science, and the social and behavioral sciences.\n\n---\n\n## 26. Geometry-Guided Camera Motion Understanding in VideoLLMs\n- **Authors**: Haoan Feng , Sri Harsha Musunuri , Guan-Ming Su\n- **URL**: [https://arxiv.org/abs/2603.13119](https://arxiv.org/abs/2603.13119)\n- **Abstract**:\n> Camera motion is a fundamental geometric signal that shapes visual perception and cinematic style, yet current video-capable vision-language models (VideoLLMs) rarely represent it explicitly and often fail on fine-grained motion primitives. We address this gap with a framework of $\\textbf{benchmarking}$, $\\textbf{diagnosis}$, and $\\textbf{injection}$. We curate $\\textbf{CameraMotionDataset}$, a large-scale synthetic dataset with explicit camera control, formulate camera motion as constraint-aware multi-label recognition, and construct a VQA benchmark--$\\textbf{CameraMotionVQA}$. Across diverse off-the-shelf VideoLLMs, we observe substantial errors in recognizing camera motion primitives. Probing experiments on a Qwen2.5-VL vision encoder suggest that camera motion cues are weakly represented, especially in deeper ViT blocks, helping explain the observed failure modes. To bridge this gap without costly training or fine-tuning, we propose a lightweight, model-agnostic pipeline that extracts geometric camera cues from 3D foundation models (3DFMs), predicts constrained motion primitives with a temporal classifier, and injects them into downstream VideoLLM inference via structured prompting. Experiments demonstrate improved motion recognition and more camera-aware model responses, highlighting geometry-driven cue extraction and structured prompting as practical steps toward a camera-aware VideoLLM and VLA system. The dataset and benchmark is publicly available at this https URL .\n\n---\n\n## 27. BoSS: A Best-of-Strategies Selector as an Oracle for Deep Active Learning\n- **Authors**: Denis Huseljic , Paul Hahn , Marek Herde , Christoph Sandrock , Bernhard Sick\n- **URL**: [https://arxiv.org/abs/2603.13109](https://arxiv.org/abs/2603.13109)\n- **Abstract**:\n> Active learning (AL) aims to reduce annotation costs while maximizing model performance by iteratively selecting valuable instances. While foundation models have made it easier to identify these instances, existing selection strategies still lack robustness across different models, annotation budgets, and datasets. To highlight the potential weaknesses of existing AL strategies and provide a reference point for research, we explore oracle strategies, i.e., strategies that approximate the optimal selection by accessing ground-truth information unavailable in practical AL scenarios. Current oracle strategies, however, fail to scale effectively to large datasets and complex deep neural networks. To tackle these limitations, we introduce the Best-of-Strategy Selector (BoSS), a scalable oracle strategy designed for large-scale AL scenarios. BoSS constructs a set of candidate batches through an ensemble of selection strategies and then selects the batch yielding the highest performance gain. As an ensemble of selection strategies, BoSS can be easily extended with new state-of-the-art strategies as they emerge, ensuring it remains a reliable oracle strategy in the future. Our evaluation demonstrates that i) BoSS outperforms existing oracle strategies, ii) state-of-the-art AL strategies still fall noticeably short of oracle performance, especially in large-scale datasets with many classes, and iii) one possible solution to counteract the inconsistent performance of AL strategies might be to employ an ensemble-based approach for the selection.\n\n---\n\n## 28. Evaluating VLMs' Spatial Reasoning Over Robot Motion: A Step Towards Robot Planning with Motion Preferences\n- **Authors**: Wenxi Wu , Jingjing Zhang , Martim Brandão\n- **URL**: [https://arxiv.org/abs/2603.13100](https://arxiv.org/abs/2603.13100)\n- **Abstract**:\n> Understanding user instructions and object spatial relations in surrounding environments is crucial for intelligent robot systems to assist humans in various tasks. The natural language and spatial reasoning capabilities of Vision-Language Models (VLMs) have the potential to enhance the generalization of robot planners on new tasks, objects, and motion specifications. While foundation models have been applied to task planning, it is still unclear the degree to which they have the capability of spatial reasoning required to enforce user preferences or constraints on motion, such as desired distances from objects, topological properties, or motion style preferences. In this paper, we evaluate the capability of four state-of-the-art VLMs at spatial reasoning over robot motion, using four different querying methods. Our results show that, with the highest-performing querying method, Qwen2.5-VL achieves 71.4% accuracy zero-shot and 75% on a smaller model after fine-tuning, and GPT-4o leads to lower performance. We evaluate two types of motion preferences (object-proximity and path-style), and we also analyze the trade-off between accuracy and computation cost in number of tokens. This work shows some promise in the potential of VLM integration with robot motion planning pipelines.\n\n---\n\n## 29. Human-in-the-Loop LLM Grading for Handwritten Mathematics Assessments\n- **Authors**: Arne Vanhoyweghen , Vincent Holst , Melika Mobini , Lukas Van de Voorde , Tibo Vanleke , Bert Verbruggen , Brecht Verbeken , Andres Algaba , Sam Verboven , Marie-Anne Guerry , Filip Van Droogenbroeck , Vincent Ginis\n- **URL**: [https://arxiv.org/abs/2603.13083](https://arxiv.org/abs/2603.13083)\n- **Abstract**:\n> Providing timely and individualised feedback on handwritten student work is highly beneficial for learning but difficult to achieve at scale. This challenge has become more pressing as generative AI undermines the reliability of take-home assessments, shifting emphasis toward supervised, in-class evaluation. We present a scalable, end-to-end workflow for LLM-assisted grading of short, pen-and-paper assessments. The workflow spans (1) constructing solution keys, (2) developing detailed rubric-style grading keys used to guide the LLM, and (3) a grading procedure that combines automated scanning and anonymisation, multi-pass LLM scoring, automated consistency checks, and mandatory human verification. We deploy the system in two undergraduate mathematics courses using six low-stakes in-class tests. Empirically, LLM assistance reduces grading time by approximately 23% while achieving agreement comparable to, and in several cases tighter than, fully manual grading. Occasional model errors occur but are effectively contained by the hybrid design. Overall, our results show that carefully embedded human-in-the-loop LLM grading can substantially reduce workload while maintaining fairness and accuracy.\n\n---\n\n## 30. GeoChemAD: Benchmarking Unsupervised Geochemical Anomaly Detection for Mineral Exploration\n- **Authors**: Yihao Ding , Yiran Zhang , Chris Gonzalez , Eun-Jung Holden , Wei Liu\n- **URL**: [https://arxiv.org/abs/2603.13068](https://arxiv.org/abs/2603.13068)\n- **Abstract**:\n> Geochemical anomaly detection plays a critical role in mineral exploration as deviations from regional geochemical baselines may indicate mineralization. Existing studies suffer from two key limitations: (1) single region scenarios which limit model generalizability; (2) proprietary datasets, which makes result reproduction unattainable. In this work, we introduce \\textbf{GeoChemAD}, an open-source benchmark dataset compiled from government-led geological surveys, covering multiple regions, sampling sources, and target elements. The dataset comprises eight subsets representing diverse spatial scales and sampling conditions. To establish strong baselines, we reproduce and benchmark a range of unsupervised anomaly detection methods, including statistical models, generative and transformer-based approaches. Furthermore, we propose \\textbf{GeoChemFormer}, a transformer-based framework that leverages self-supervised pretraining to learn target-element-aware geochemical representations for spatial samples. Extensive experiments demonstrate that GeoChemFormer consistently achieves superior and robust performance across all eight subsets, outperforming existing unsupervised methods in both anomaly detection accuracy and generalization capability. The proposed dataset and framework provide a foundation for reproducible research and future development in this direction.\n\n---\n\n## 31. L2GTX: From Local to Global Time Series Explanations\n- **Authors**: Ephrem Tibebe Mekonnen , Luca Longo , Lucas Rizzo , Pierpaolo Dondio\n- **URL**: [https://arxiv.org/abs/2603.13065](https://arxiv.org/abs/2603.13065)\n- **Abstract**:\n> Deep learning models achieve high accuracy in time series classification, yet understanding their class-level decision behaviour remains challenging. Explanations for time series must respect temporal dependencies and identify patterns that recur across instances. Existing approaches face three limitations: model-agnostic XAI methods developed for images and tabular data do not readily extend to time series, global explanation synthesis for time series remains underexplored, and most existing global approaches are model-specific. We propose L2GTX, a model-agnostic framework that generates class-wise global explanations by aggregating local explanations from a representative set of instances. L2GTX extracts clusters of parameterised temporal event primitives, such as increasing or decreasing trends and local extrema, together with their importance scores from instance-level explanations produced by LOMATCE. These clusters are merged across instances to reduce redundancy, and an instance-cluster importance matrix is used to estimate global relevance. Under a user-defined instance selection budget, L2GTX selects representative instances that maximise coverage of influential clusters. Events from the selected instances are then aggregated into concise class-wise global explanations. Experiments on six benchmark time series datasets show that L2GTX produces compact and interpretable global explanations while maintaining stable global faithfulness measured as mean local surrogate fidelity.\n\n---\n\n## 32. Competition-Aware CPC Forecasting with Near-Market Coverage\n- **Authors**: Sebastian Frey , Edoardo Beccari , Maximilian Kranz , Nicolò Alberto Pellizzari , Ali Mete Karaman , Qiwei Han , Maximilian Kaiser\n- **URL**: [https://arxiv.org/abs/2603.13059](https://arxiv.org/abs/2603.13059)\n- **Abstract**:\n> Cost-per-click (CPC) in paid search is a volatile auction outcome generated by a competitive landscape that is only partially observable from any single advertiser's history. Using Google Ads auction logs from a concentrated car-rental market (2021--2023), we forecast weekly CPC for 1,811 keyword series and approximate latent competition through complementary signals derived from keyword text, CPC trajectories, and geographic market structure. We construct (i) semantic neighborhoods and a semantic keyword graph from pretrained transformer-based representations of keyword text, (ii) behavioral neighborhoods via Dynamic Time Warping (DTW) alignment of CPC trajectories, and (iii) geographic-intent covariates capturing localized demand and marketplace heterogeneity. We extensively evaluate these signals both as stand-alone covariates and as relational priors in spatiotemporal graph forecasters, benchmarking them against strong statistical, neural, and time-series foundation-model baselines. Across methods, competition-aware augmentation improves stability and error profiles at business-relevant medium and longer horizons, where competitive regimes shift and volatility is most consequential. The results show that broad market-outcome coverage, combined with keyword-derived semantic and geographic priors, provides a scalable way to approximate latent competition and improve CPC forecasting in auction-driven markets.\n\n---\n\n## 33. Team RAS in 10th ABAW Competition: Multimodal Valence and Arousal Estimation Approach\n- **Authors**: Elena Ryumina (1), Maxim Markitantov (1), Alexandr Axyonov (1), Dmitry Ryumin (1), Mikhail Dolgushin (1), Denis Dresvyanskiy (2), Alexey Karpov (1 and 2) ((1) St. Petersburg Federal Research Center of the Russian Academy of Sciences, St. Petersburg, Russia, (2) ITMO University, St. Petersburg, Russia)\n- **URL**: [https://arxiv.org/abs/2603.13056](https://arxiv.org/abs/2603.13056)\n- **Abstract**:\n> Continuous emotion recognition in terms of valence and arousal under in-the-wild (ITW) conditions remains a challenging problem due to large variations in appearance, head pose, illumination, occlusions, and subject-specific patterns of affective expression. We present a multimodal method for valence-arousal estimation ITW. Our method combines three complementary modalities: face, behavior, and audio. The face modality relies on GRADA-based frame-level embeddings and Transformer-based temporal regression. We use Qwen3-VL-4B-Instruct to extract behavior-relevant information from video segments, while Mamba is used to model temporal dynamics across segments. The audio modality relies on WavLM-Large with attention-statistics pooling and includes a cross-modal filtering stage to reduce the influence of unreliable or non-speech segments. To fuse modalities, we explore two fusion strategies: a Directed Cross-Modal Mixture-of-Experts Fusion Strategy that learns interactions between modalities with adaptive weighting, and a Reliability-Aware Audio-Visual Fusion Strategy that combines visual features at the frame-level while using audio as complementary context. The results are reported on the Aff-Wild2 dataset following the 10th Affective Behavior Analysis in-the-Wild (ABAW) challenge protocol. Experiments demonstrate that the proposed multimodal fusion strategy achieves a Concordance Correlation Coefficient (CCC) of 0.658 on the Aff-Wild2 development set.\n\n---\n\n## 34. Are General-Purpose Vision Models All We Need for 2D Medical Image Segmentation? A Cross-Dataset Empirical Study\n- **Authors**: Vanessa Borst , Samuel Kounev\n- **URL**: [https://arxiv.org/abs/2603.13044](https://arxiv.org/abs/2603.13044)\n- **Abstract**:\n> Medical image segmentation (MIS) is a fundamental component of computer-assisted diagnosis and clinical decision support systems. Over the past decade, numerous architectures specifically tailored to medical imaging have emerged to address domain-specific challenges such as low contrast, small anatomical structures, and limited annotated data. In parallel, rapid progress in computer vision has produced highly capable general-purpose vision models (GP-VMs) originally designed for natural images. Despite their strong performance on standard vision benchmarks, their effectiveness for MIS remains insufficiently understood. In this work, we conduct a controlled empirical study to examine whether specialized medical segmentation architectures (SMAs) provide systematic advantages over modern GP-VMs for 2D MIS. We compare eleven SMAs and GP-VMs using a unified training and evaluation protocol. Experiments are performed across three heterogeneous datasets covering different imaging modalities, class structures, and data characteristics. Beyond segmentation accuracy, we analyze qualitative Grad-CAM visualizations to investigate explainability (XAI) behavior. Our results demonstrate that, for the analyzed datasets, GP-VMs out-perform the majority of specialized MIS models. Moreover, XAI analyses indicate that GP-VMs can capture clinically relevant structures without explicit domain-specific architectural design. These findings suggest that GP-VMs can represent a viable alternative to domain-specific methods, highlighting the importance of informed model selection for end-to-end MIS systems. All code and resources are available at GitHub.\n\n---\n\n## 35. Interrogating Design Homogenization in Web Vibe Coding\n- **Authors**: Donghoon Shin , Alice Gao , Rock Yuren Pang , Jaewook Lee , Katharina Reinecke , Emily Tseng\n- **URL**: [https://arxiv.org/abs/2603.13036](https://arxiv.org/abs/2603.13036)\n- **Abstract**:\n> Generative AI is known for its tendency to homogenize, often reproducing dominant style conventions found in training data. However, it remains unclear how these homogenizing effects extend to complex structural tasks like web design. As lay creators increasingly turn to LLMs to 'vibe-code' websites -- prompting for aesthetic and functional goals rather than writing code -- they may inadvertently narrow the diversity of their designs, and limit creative expression throughout the internet. In this paper, we interrogate the possibility of design homogenization in web vibe coding. We first characterize the vibe coding lifecycle, pinpointing stages where homogenization risks may arise. We then conduct a sociotechnical risk analysis unpacking the potential harms of web vibe coding and their interaction with design homogenization. We identify that the push for frictionless generation can exacerbate homogenization and its harms. Finally, we propose a mitigation framework centered on the idea of productive friction. Through case studies at the micro, meso, and macro levels, we show how centering productive friction can empower creators to challenge default outputs and preserve diverse expression in AI-mediated web design.\n\n---\n\n## 36. Purify Once, Edit Freely: Breaking Image Protections under Model Mismatch\n- **Authors**: Qichen Zhao , Shengfang Zhai , Xinjian Bai , Qingni Shen , Qiqi Lin , Yansong Gao , Zhonghai Wu\n- **URL**: [https://arxiv.org/abs/2603.13028](https://arxiv.org/abs/2603.13028)\n- **Abstract**:\n> Diffusion models enable high-fidelity image editing but can also be misused for unauthorized style imitation and harmful content generation. To mitigate these risks, proactive image protection methods embed small, often imperceptible adversarial perturbations into images before sharing to disrupt downstream editing or fine-tuning. However, in realistic post-release scenarios, content owners cannot control downstream processing pipelines, and protections optimized for a surrogate model may fail when attackers use mismatched diffusion pipelines. Existing purification methods can weaken protections but often sacrifice image quality and rarely examine architectural mismatch. We introduce a unified post-release purification framework to evaluate protection survivability under model mismatch. We propose two practical purifiers: VAE-Trans, which corrects protected images via latent-space projection, and EditorClean, which performs instruction-guided reconstruction with a Diffusion Transformer to exploit architectural heterogeneity. Both operate without access to protected images or defense internals. Across 2,100 editing tasks and six representative protection methods, EditorClean consistently restores editability. Compared to protected inputs, it improves PSNR by 3-6 dB and reduces FID by 50-70 percent on downstream edits, while outperforming prior purification baselines by about 2 dB PSNR and 30 percent lower FID. Our results reveal a purify-once, edit-freely failure mode: once purification succeeds, the protective signal is largely removed, enabling unrestricted editing. This highlights the need to evaluate protections under model mismatch and design defenses robust to heterogeneous attackers.\n\n---\n\n## 37. SortScrews: A Dataset and Baseline for Real-time Screw Classification\n- **Authors**: Tianhao Fu , Bingxuan Yang , Juncheng Guo , Shrena Sribalan , Yucheng Chen\n- **URL**: [https://arxiv.org/abs/2603.13027](https://arxiv.org/abs/2603.13027)\n- **Abstract**:\n> Automatic identification of screw types is important for industrial automation, robotics, and inventory management. However, publicly available datasets for screw classification are scarce, particularly for controlled single-object scenarios commonly encountered in automated sorting systems. In this work, we introduce $\\textbf{SortScrews}$, a dataset for casewise visual classification of screws. The dataset contains 560 RGB images at $512\\times512$ resolution covering six screw types and a background class. Images are captured using a standardized acquisition setup and include mild variations in lighting and camera perspective across four capture settings. To facilitate reproducible research and dataset expansion, we also provide a reusable data collection script that allows users to easily construct similar datasets for custom hardware components using inexpensive camera setups. We establish baseline results using transfer learning with EfficientNet-B0 and ResNet-18 classifiers pretrained on ImageNet. In addition, we conduct a well-explored failure analysis. Despite the limited dataset size, these lightweight models achieve strong classification accuracy, demonstrating that controlled acquisition conditions enable effective learning even with relatively small datasets. The dataset, collection pipeline, and baseline training code are publicly available at this https URL .\n\n---\n\n## 38. SAW: Toward a Surgical Action World Model via Controllable and Scalable Video Generation\n- **Authors**: Sampath Rapuri , Lalithkumar Seenivasan , Dominik Schneider , Roger Soberanis-Mukul , Yufan He , Hao Ding , Jiru Xu , Chenhao Yu , Chenyan Jing , Pengfei Guo , Daguang Xu , Mathias Unberath\n- **URL**: [https://arxiv.org/abs/2603.13024](https://arxiv.org/abs/2603.13024)\n- **Abstract**:\n> A surgical world model capable of generating realistic surgical action videos with precise control over tool-tissue interactions can address fundamental challenges in surgical AI and simulation -- from data scarcity and rare event synthesis to bridging the sim-to-real gap for surgical automation. However, current video generation methods, the very core of such surgical world models, require expensive annotations or complex structured intermediates as conditioning signals at inference, limiting their scalability. Other approaches exhibit limited temporal consistency across complex laparoscopic scenes and do not possess sufficient realism. We propose Surgical Action World (SAW) -- a step toward surgical action world modeling through video diffusion conditioned on four lightweight signals: language prompts encoding tool-action context, a reference surgical scene, tissue affordance mask, and 2D tool-tip trajectories. We design a conditional video diffusion approach that reformulates video-to-video diffusion into trajectory-conditioned surgical action synthesis. The backbone diffusion model is fine-tuned on a custom-curated dataset of 12,044 laparoscopic clips with lightweight spatiotemporal conditioning signals, leveraging a depth consistency loss to enforce geometric plausibility without requiring depth at inference. SAW achieves state-of-the-art temporal consistency (CD-FVD: 199.19 vs. 546.82) and strong visual quality on held-out test data. Furthermore, we demonstrate its downstream utility for (a) surgical AI, where augmenting rare actions with SAW-generated videos improves action recognition (clipping F1-score: 20.93% to 43.14%; cutting: 0.00% to 8.33%) on real test data, and (b) surgical simulation, where rendering tool-tissue interaction videos from simulator-derived trajectory points toward a visually faithful simulation engine.\n\n---\n\n## 39. daVinci-Env: Open SWE Environment Synthesis at Scale\n- **Authors**: Dayuan Fu , Shenyu Wu , Yunze Wu , Zerui Peng , Yaxing Huang , Jie Sun , Ji Zeng , Mohan Jiang , Lin Zhang , Yukun Li , Jiarui Hu , Liming Liu , Jinlong Hou , Pengfei Liu\n- **URL**: [https://arxiv.org/abs/2603.13023](https://arxiv.org/abs/2603.13023)\n- **Abstract**:\n> Training capable software engineering (SWE) agents demands large-scale, executable, and verifiable environments that provide dynamic feedback loops for iterative code editing, test execution, and solution refinement. However, existing open-source datasets remain limited in scale and repository diversity, while industrial solutions are opaque with unreleased infrastructure, creating a prohibitive barrier for most academic research groups. We present OpenSWE, the largest fully transparent framework for SWE agent training in Python, comprising 45,320 executable Docker environments spanning over 12.8k repositories, with all Dockerfiles, evaluation scripts, and infrastructure fully open-sourced for reproducibility. OpenSWE is built through a multi-agent synthesis pipeline deployed across a 64-node distributed cluster, automating repository exploration, Dockerfile construction, evaluation script generation, and iterative test analysis. Beyond scale, we propose a quality-centric filtering pipeline that characterizes the inherent difficulty of each environment, filtering out instances that are either unsolvable or insufficiently challenging and retaining only those that maximize learning efficiency. With $891K spent on environment construction and an additional $576K on trajectory sampling and difficulty-aware curation, the entire project represents a total investment of approximately $1.47 million, yielding about 13,000 curated trajectories from roughly 9,000 quality guaranteed environments. Extensive experiments validate OpenSWE's effectiveness: OpenSWE-32B and OpenSWE-72B achieve 62.4% and 66.0% on SWE-bench Verified, establishing SOTA among Qwen2.5 series. Moreover, SWE-focused training yields substantial out-of-domain improvements, including up to 12 points on mathematical reasoning and 5 points on science benchmarks, without degrading factual recall.\n\n---\n\n## 40. ARL-Tangram: Unleash the Resource Efficiency in Agentic Reinforcement Learning\n- **Authors**: Bangjun Xiao , Yihao Zhao , Xiangwei Deng , Shihua Yu , Yuxing Xiang , Huaqiu Liu , Qiying Wang , Liang Zhao , Hailin Zhang , Xuanzhe Liu , Xin Jin , Fuli Luo\n- **URL**: [https://arxiv.org/abs/2603.13019](https://arxiv.org/abs/2603.13019)\n- **Abstract**:\n> Agentic reinforcement learning (RL) has emerged as a transformative workload in cloud clusters, enabling large language models (LLMs) to solve complex problems through interactions with real world. However, unlike traditional RL, agentic RL demands substantial external cloud resources, e.g., CPUs for code execution and GPUs for reward models, that exist outside the primary training cluster. Existing agentic RL framework typically rely on static over-provisioning, i.e., resources are often tied to long-lived trajectories or isolated by tasks, which leads to severe resource inefficiency. We propose the action-level orchestration, and incorporate it into ARL-Tangram, a unified resource management system that enables fine-grained external resource sharing and elasticity. ARL-Tangram utilizes a unified action-level formulation and an elastic scheduling algorithm to minimize action completion time (ACT) while satisfying heterogeneous resource constraints. Further, heterogeneous resource managers are tailored to efficiently support the action-level execution on resources with heterogeneous characteristics and topologies. Evaluation on real-world agentic RL tasks demonstrates that ARL-Tangram improves average ACT by up to 4.3$\\times$, speeds up the step duration of RL training by up to 1.5$\\times$, and saves the external resources by up to 71.2$\\%$. This system has been deployed to support the training of the MiMo series models.\n\n---\n\n## 41. Fair Lung Disease Diagnosis from Chest CT via Gender-Adversarial Attention Multiple Instance Learning\n- **Authors**: Aditya Parikh , Aasa Feragen\n- **URL**: [https://arxiv.org/abs/2603.12988](https://arxiv.org/abs/2603.12988)\n- **Abstract**:\n> We present a fairness-aware framework for multi-class lung disease diagnosis from chest CT volumes, developed for the Fair Disease Diagnosis Challenge at the PHAROS-AIF-MIH Workshop (CVPR 2026). The challenge requires classifying CT scans into four categories -- Healthy, COVID-19, Adenocarcinoma, and Squamous Cell Carcinoma -- with performance measured as the average of per-gender macro F1 scores, explicitly penalizing gender-inequitable predictions. Our approach addresses two core difficulties: the sparse pathological signal across hundreds of slices, and a severe demographic imbalance compounded across disease class and gender. We propose an attention-based Multiple Instance Learning (MIL) model on a ConvNeXt backbone that learns to identify diagnostically relevant slices without slice-level supervision, augmented with a Gradient Reversal Layer (GRL) that adversarially suppresses gender-predictive structure in the learned scan representation. Training incorporates focal loss with label smoothing, stratified cross-validation over joint (class, gender) strata, and targeted oversampling of the most underrepresented subgroup. At inference, all five-fold checkpoints are ensembled with horizontal-flip test-time augmentation via soft logit voting and out-of-the-fold threshold optimization for robustness. Our model achieves a mean validation competition score of 0.685 (std - 0.030), with the best single fold reaching 0.759. All training and inference code is publicly available at this https URL\n\n---\n\n## 42. Is Human Annotation Necessary? Iterative MBR Distillation for Error Span Detection in Machine Translation\n- **Authors**: Boxuan Lyu , Haiyue Song , Zhi Qu\n- **URL**: [https://arxiv.org/abs/2603.12983](https://arxiv.org/abs/2603.12983)\n- **Abstract**:\n> Error Span Detection (ESD) is a crucial subtask in Machine Translation (MT) evaluation, aiming to identify the location and severity of translation errors. While fine-tuning models on human-annotated data improves ESD performance, acquiring such data is expensive and prone to inconsistencies among annotators. To address this, we propose a novel self-evolution framework based on Minimum Bayes Risk (MBR) decoding, named Iterative MBR Distillation for ESD, which eliminates the reliance on human annotations by leveraging an off-the-shelf LLM to generate pseudo-labels. Extensive experiments on the WMT Metrics Shared Task datasets demonstrate that models trained solely on these self-generated pseudo-labels outperform both unadapted base model and supervised baselines trained on human annotations at the system and span levels, while maintaining competitive sentence-level performance.\n\n---\n\n## 43. Efficient Real-World Autonomous Racing via Attenuated Residual Policy Optimization\n- **Authors**: Raphael Trumpp , Denis Hoornaert , Mirco Theile , Marco Caccamo\n- **URL**: [https://arxiv.org/abs/2603.12960](https://arxiv.org/abs/2603.12960)\n- **Abstract**:\n> Residual policy learning (RPL), in which a learned policy refines a static base policy using deep reinforcement learning (DRL), has shown strong performance across various robotic applications. Its effectiveness is particularly evident in autonomous racing, a domain that serves as a challenging benchmark for real-world DRL. However, deploying RPL-based controllers introduces system complexity and increases inference latency. We address this by introducing an extension of RPL named attenuated residual policy optimization ($\\alpha$-RPO). Unlike standard RPL, $\\alpha$-RPO yields a standalone neural policy by progressively attenuating the base policy, which initially serves to bootstrap learning. Furthermore, this mechanism enables a form of privileged learning, where the base policy is permitted to use sensor modalities not required for final deployment. We design $\\alpha$-RPO to integrate seamlessly with PPO, ensuring that the attenuated influence of the base controller is dynamically compensated during policy optimization. We evaluate $\\alpha$-RPO by building a framework for 1:10-scaled autonomous racing around it. In both simulation and zero-shot real-world transfer to Roboracer cars, $\\alpha$-RPO not only reduces system complexity but also improves driving performance compared to baselines - demonstrating its practicality for robotic deployment. Our code is available at: this https URL .\n\n---\n\n## 44. Delta1 with LLM: symbolic and neural integration for credible and explainable reasoning\n- **Authors**: Yang Xu , Jun Liu , Shuwei Chen , Chris Nugent , Hailing Guo\n- **URL**: [https://arxiv.org/abs/2603.12953](https://arxiv.org/abs/2603.12953)\n- **Abstract**:\n> Neuro-symbolic reasoning increasingly demands frameworks that unite the formal rigor of logic with the interpretability of large language models (LLMs). We introduce an end to end explainability by construction pipeline integrating the Automated Theorem Generator Delta1 based on the full triangular standard contradiction (FTSC) with LLMs. Delta1 deterministically constructs minimal unsatisfiable clause sets and complete theorems in polynomial time, ensuring both soundness and minimality by construction. The LLM layer verbalizes each theorem and proof trace into coherent natural language explanations and actionable insights. Empirical studies across health care, compliance, and regulatory domains show that Delta1 and LLM enables interpretable, auditable, and domain aligned reasoning. This work advances the convergence of logic, language, and learning, positioning constructive theorem generation as a principled foundation for neuro-symbolic explainable AI.\n\n---\n\n## 45. Thinking in Streaming Video\n- **Authors**: Zikang Liu , Longteng Guo , Handong Li , Ru Zhen , Xingjian He , Ruyi Ji , Xiaoming Ren , Yanhao Zhang , Haonan Lu , Jing Liu\n- **URL**: [https://arxiv.org/abs/2603.12938](https://arxiv.org/abs/2603.12938)\n- **Abstract**:\n> Real-time understanding of continuous video streams is essential for interactive assistants and multimodal agents operating in dynamic environments. However, most existing video reasoning approaches follow a batch paradigm that defers reasoning until the full video context is observed, resulting in high latency and growing computational cost that are incompatible with streaming scenarios. In this paper, we introduce ThinkStream, a framework for streaming video reasoning based on a Watch--Think--Speak paradigm that enables models to incrementally update their understanding as new video observations arrive. At each step, the model performs a short reasoning update and decides whether sufficient evidence has accumulated to produce a response. To support long-horizon streaming, we propose Reasoning-Compressed Streaming Memory (RCSM), which treats intermediate reasoning traces as compact semantic memory that replaces outdated visual tokens while preserving essential context. We further train the model using a Streaming Reinforcement Learning with Verifiable Rewards scheme that aligns incremental reasoning and response timing with the requirements of streaming interaction. Experiments on multiple streaming video benchmarks show that ThinkStream significantly outperforms existing online video models while maintaining low latency and memory usage. Code, models and data will be released at this https URL\n\n---\n\n## 46. Surprised by Attention: Predictable Query Dynamics for Time Series Anomaly Detection\n- **Authors**: Kadir-Kaan Özer , René Ebeling , Markus Enzweiler\n- **URL**: [https://arxiv.org/abs/2603.12916](https://arxiv.org/abs/2603.12916)\n- **Abstract**:\n> Multivariate time series anomalies often manifest as shifts in cross-channel dependencies rather than simple amplitude excursions. In autonomous driving, for instance, a steering command might be internally consistent but decouple from the resulting lateral acceleration. Residual-based detectors can miss such anomalies when flexible sequence models still reconstruct signals plausibly despite altered coordination. We introduce AxonAD, an unsupervised detector that treats multi-head attention query evolution as a short horizon predictable process. A gradient-updated reconstruction pathway is coupled with a history-only predictor that forecasts future query vectors from past context. This is trained via a masked predictor-target objective against an exponential moving average (EMA) target encoder. At inference, reconstruction error is combined with a tail-aggregated query mismatch score, which measures cosine deviation between predicted and target queries on recent timesteps. This dual approach provides sensitivity to structural dependency shifts while retaining amplitude-level detection. On proprietary in-vehicle telemetry with interval annotations and on the TSB-AD multi-variate suite (17 datasets, 180 series) with threshold-free and range-aware metrics, AxonAD improves ranking quality and temporal localization over strong baselines. Ablations confirm that query prediction and combined scoring are the primary drivers of the observed gains. Code is available at the URL this https URL .\n\n---\n\n## 47. Stake the Points: Structure-Faithful Instance Unlearning\n- **Authors**: Kiseong Hong , JungKyoo Shin , Eunwoo Kim\n- **URL**: [https://arxiv.org/abs/2603.12915](https://arxiv.org/abs/2603.12915)\n- **Abstract**:\n> Machine unlearning (MU) addresses privacy risks in pretrained models. The main goal of MU is to remove the influence of designated data while preserving the utility of retained knowledge. Achieving this goal requires preserving semantic relations among retained instances, which existing studies often overlook. We observe that without such preservation, models suffer from progressive structural collapse, undermining both the deletion-retention balance. In this work, we propose a novel structure-faithful framework that introduces stakes, i.e., semantic anchors that serve as reference points to maintain the knowledge structure. By leveraging these anchors, our framework captures and stabilizes the semantic organization of knowledge. Specifically, we instantiate the anchors from language-driven attribute descriptions encoded by a semantic encoder (e.g., CLIP). We enforce preservation of the knowledge structure via structure-aware alignment and regularization: the former aligns the organization of retained knowledge before and after unlearning around anchors, while the latter regulates updates to structure-critical parameters. Results from image classification, retrieval, and face recognition show average gains of 32.9%, 22.5%, and 19.3% in performance, balancing the deletion-retention trade-off and enhancing generalization.\n\n---\n\n## 48. FedBPrompt: Federated Domain Generalization Person Re-Identification via Body Distribution Aware Visual Prompts\n- **Authors**: Xin Xu , Weilong Li , Wei Liu , Wenke Huang , Zhixi Yu , Bin Yang , Xiaoying Liao , Kui Jiang\n- **URL**: [https://arxiv.org/abs/2603.12912](https://arxiv.org/abs/2603.12912)\n- **Abstract**:\n> Federated Domain Generalization for Person Re-Identification (FedDG-ReID) learns domain-invariant representations from decentralized data. While Vision Transformer (ViT) is widely adopted, its global attention often fails to distinguish pedestrians from high similarity backgrounds or diverse viewpoints -- a challenge amplified by cross-client distribution shifts in FedDG-ReID. To address this, we propose Federated Body Distribution Aware Visual Prompt (FedBPrompt), introducing learnable visual prompts to guide Transformer attention toward pedestrian-centric regions. FedBPrompt employs a Body Distribution Aware Visual Prompts Mechanism (BAPM) comprising: Holistic Full Body Prompts to suppress cross-client background noise, and Body Part Alignment Prompts to capture fine-grained details robust to pose and viewpoint variations. To mitigate high communication costs, we design a Prompt-based Fine-Tuning Strategy (PFTS) that freezes the ViT backbone and updates only lightweight prompts, significantly reducing communication overhead while maintaining adaptability. Extensive experiments demonstrate that BAPM effectively enhances feature discrimination and cross-domain generalization, while PFTS achieves notable performance gains within only a few aggregation rounds. Moreover, both BAPM and PFTS can be easily integrated into existing ViT-based FedDG-ReID frameworks, making FedBPrompt a flexible and effective solution for federated person re-identification. The code is available at this https URL .\n\n---\n\n## 49. Learning from Child-Directed Speech in Two-Language Scenarios: A French-English Case Study\n- **Authors**: Liel Binyamin , Elior Sulem\n- **URL**: [https://arxiv.org/abs/2603.12906](https://arxiv.org/abs/2603.12906)\n- **Abstract**:\n> Research on developmentally plausible language models has largely focused on English, leaving open questions about multilingual settings. We present a systematic study of compact language models by extending BabyBERTa to English-French scenarios under strictly size-matched data conditions, covering monolingual, bilingual, and cross-lingual settings. Our design contrasts two types of training corpora: (i) child-directed speech (about 2.5M tokens), following BabyBERTa and related work, and (ii) multi-domain corpora (about 10M tokens), extending the BabyLM framework to French. To enable fair evaluation, we also introduce new resources, including French versions of QAMR and QASRL, as well as English and French multi-domain corpora. We evaluate the models on both syntactic and semantic tasks and compare them with models trained on Wikipedia-only data. The results reveal context-dependent effects: training on Wikipedia consistently benefits semantic tasks, whereas child-directed speech improves grammatical judgments in monolingual settings. Bilingual pretraining yields notable gains for textual entailment, with particularly strong improvements for French. Importantly, similar patterns emerge across BabyBERTa, RoBERTa, and LTG-BERT, suggesting consistent trends across architectures.\n\n---\n\n## 50. Human-Centered Evaluation of an LLM-Based Process Modeling Copilot: A Mixed-Methods Study with Domain Experts\n- **Authors**: Chantale Lauer , Peter Pfeiffer , Nijat Mehdiyev\n- **URL**: [https://arxiv.org/abs/2603.12895](https://arxiv.org/abs/2603.12895)\n- **Abstract**:\n> Integrating Large Language Models (LLMs) into business process management tools promises to democratize Business Process Model and Notation (BPMN) modeling for non-experts. While automated frameworks assess syntactic and semantic quality, they miss human factors like trust, usability, and professional alignment. We conducted a mixed-methods evaluation of our proposed solution, an LLM-powered BPMN copilot, with five process modeling experts using focus groups and standardized questionnaires. Our findings reveal a critical tension between acceptable perceived usability (mean CUQ score: 67.2/100) and notably lower trust (mean score: 48.8\\%), with reliability rated as the most critical concern (M=1.8/5). Furthermore, we identified output-quality issues, prompting difficulties, and a need for the LLM to ask more in-depth clarifying questions about the process. We envision five use cases ranging from domain-expert support to enterprise quality assurance. We demonstrate the necessity of human-centered evaluation complementing automated benchmarking for LLM modeling agents.\n\n---\n\n## 51. Finite Difference Flow Optimization for RL Post-Training of Text-to-Image Models\n- **Authors**: David McAllister , Miika Aittala , Tero Karras , Janne Hellsten , Angjoo Kanazawa , Timo Aila , Samuli Laine\n- **URL**: [https://arxiv.org/abs/2603.12893](https://arxiv.org/abs/2603.12893)\n- **Abstract**:\n> Reinforcement learning (RL) has become a standard technique for post-training diffusion-based image synthesis models, as it enables learning from reward signals to explicitly improve desirable aspects such as image quality and prompt alignment. In this paper, we propose an online RL variant that reduces the variance in the model updates by sampling paired trajectories and pulling the flow velocity in the direction of the more favorable image. Unlike existing methods that treat each sampling step as a separate policy action, we consider the entire sampling process as a single action. We experiment with both high-quality vision language models and off-the-shelf quality metrics for rewards, and evaluate the outputs using a broad set of metrics. Our method converges faster and yields higher output quality and prompt alignment than previous approaches.\n\n---\n\n## 52. Team LEYA in 10th ABAW Competition: Multimodal Ambivalence/Hesitancy Recognition Approach\n- **Authors**: Elena Ryumina (1), Alexandr Axyonov (1), Dmitry Sysoev (2), Timur Abdulkadirov (2), Kirill Almetov (2), Yulia Morozova (2), Dmitry Ryumin (1 and 2) ((1) St. Petersburg Federal Research Center of the Russian Academy of Sciences, St. Petersburg, Russia, (2) HSE University, St. Petersburg, Russia)\n- **URL**: [https://arxiv.org/abs/2603.12848](https://arxiv.org/abs/2603.12848)\n- **Abstract**:\n> Ambivalence/hesitancy recognition in unconstrained videos is a challenging problem due to the subtle, multimodal, and context-dependent nature of this behavioral state. In this paper, a multimodal approach for video-level ambivalence/hesitancy recognition is presented for the 10th ABAW Competition. The proposed approach integrates four complementary modalities: scene, face, audio, and text. Scene dynamics are captured with a VideoMAE-based model, facial information is encoded through emotional frame-level embeddings aggregated by statistical pooling, acoustic representations are extracted with EmotionWav2Vec2.0 and processed by a Mamba-based temporal encoder, and linguistic cues are modeled using fine-tuned transformer-based text models. The resulting unimodal embeddings are further combined using multimodal fusion models, including prototype-augmented variants. Experiments on the BAH corpus demonstrate clear gains of multimodal fusion over all unimodal baselines. The best unimodal configuration achieved an average MF1 of 70.02%, whereas the best multimodal fusion model reached 83.25%. The highest final test performance, 71.43%, was obtained by an ensemble of five prototype-augmented fusion models. The obtained results highlight the importance of complementary multimodal cues and robust fusion strategies for ambivalence/hesitancy recognition.\n\n---\n\n## 53. Hierarchical Reference Sets for Robust Unsupervised Detection of Scattered and Clustered Outliers\n- **Authors**: Yiqun Zhang , Zexi Tan , Xiaopeng Luo , Yunlin Liu\n- **URL**: [https://arxiv.org/abs/2603.12847](https://arxiv.org/abs/2603.12847)\n- **Abstract**:\n> Most real-world IoT data analysis tasks, such as clustering and anomaly event detection, are unsupervised and highly susceptible to the presence of outliers. In addition to sporadic scattered outliers caused by factors such as faulty sensor readings, IoT systems often exhibit clustered outliers. These occur when multiple devices or nodes produce similar anomalous measurements, for instance, owing to localized interference, emerging security threats, or regional false alarms, forming micro-clusters. These clustered outliers can be easily mistaken for normal behavior because of their relatively high local density, thereby obscuring the detection of both scattered and contextual anomalies. To address this, we propose a novel outlier detection paradigm that leverages the natural neighboring relationships using graph structures. This facilitates multi-perspective anomaly evaluation by incorporating reference sets at both local and global scales derived from the graph. Our approach enables the effective recognition of scattered outliers without interference from clustered anomalies, whereas the graph structure simultaneously helps reflect and isolate clustered outlier groups. Extensive experiments, including comparative performance analysis, ablation studies, validation on downstream clustering tasks, and evaluation of hyperparameter sensitivity, demonstrate the efficacy of the proposed method. The source code is available at this https URL .\n\n---\n\n## 54. DAST: A Dual-Stream Voice Anonymization Attacker with Staged Training\n- **Authors**: Ridwan Arefeen , Xiaoxiao Miao , Rong Tong , Aik Beng Ng , Simon See , Timothy Liu\n- **URL**: [https://arxiv.org/abs/2603.12840](https://arxiv.org/abs/2603.12840)\n- **Abstract**:\n> Voice anonymization masks vocal traits while preserving linguistic content, which may still leak speaker-specific patterns. To assess and strengthen privacy evaluation, we propose a dual-stream attacker that fuses spectral and self-supervised learning features via parallel encoders with a three-stage training strategy. Stage I establishes foundational speaker-discriminative representations. Stage II leverages the shared identity-transformation characteristics of voice conversion and anonymization, exposing the model to diverse converted speech to build cross-system robustness. Stage III provides lightweight adaptation to target anonymized data. Results on the VoicePrivacy Attacker Challenge (VPAC) dataset demonstrate that Stage II is the primary driver of generalization, enabling strong attacking performance on unseen anonymization datasets. With Stage III, fine-tuning on only 10\\% of the target anonymization dataset surpasses current state-of-the-art attackers in terms of EER.\n\n---\n\n## 55. Mask2Flow-TSE: Two-Stage Target Speaker Extraction with Masking and Flow Matching\n- **Authors**: Junwon Moon , Hyunjin Choi , Hansol Park , Heeseung Kim , Kyuhong Shim\n- **URL**: [https://arxiv.org/abs/2603.12837](https://arxiv.org/abs/2603.12837)\n- **Abstract**:\n> Target speaker extraction (TSE) extracts the target speaker's voice from overlapping speech mixtures given a reference utterance. Existing approaches typically fall into two categories: discriminative and generative. Discriminative methods apply time-frequency masking for fast inference but often over-suppress the target signal, while generative methods synthesize high-quality speech at the cost of numerous iterative steps. We propose Mask2Flow-TSE, a two-stage framework combining the strengths of both paradigms. The first stage applies discriminative masking for coarse separation, and the second stage employs flow matching to refine the output toward target speech. Unlike generative approaches that synthesize speech from Gaussian noise, our method starts from the masked spectrogram, enabling high-quality reconstruction in a single inference step. Experiments show that Mask2Flow-TSE achieves comparable performance to existing generative TSE methods with approximately 85M parameters.\n\n---\n\n## 56. Hierarchical Dual-Change Collaborative Learning for UAV Scene Change Captioning\n- **Authors**: Fuhai Chen , Pengpeng Huang , Junwen Wu , Hehong Zhang , Shiping Wang , Xiaoguang Ma , Xuri Ge\n- **URL**: [https://arxiv.org/abs/2603.12832](https://arxiv.org/abs/2603.12832)\n- **Abstract**:\n> This paper proposes a novel task for UAV scene understanding - UAV Scene Change Captioning (UAV-SCC) - which aims to generate natural language descriptions of semantic changes in dynamic aerial imagery captured from a movable viewpoint. Unlike traditional change captioning that mainly describes differences between image pairs captured from a fixed camera viewpoint over time, UAV scene change captioning focuses on image-pair differences resulting from both temporal and spatial scene variations dynamically captured by a moving camera. The key challenge lies in understanding viewpoint-induced scene changes from UAV image pairs that share only partially overlapping scene content due to viewpoint shifts caused by camera rotation, while effectively exploiting the relative orientation between the two images. To this end, we propose a Hierarchical Dual-Change Collaborative Learning (HDC-CL) method for UAV scene change captioning. In particular, a novel transformer, \\emph{i.e.} Dynamic Adaptive Layout Transformer (DALT) is designed to adaptively model diverse spatial layouts of the image pair, where the interrelated features derived from the overlapping and non-overlapping regions are learned within the flexible and unified encoding layer. Furthermore, we propose a Hierarchical Cross-modal Orientation Consistency Calibration (HCM-OCC) method to enhance the model's sensitivity to viewpoint shift directions, enabling more accurate change captioning. To facilitate in-depth research on this task, we construct a new benchmark dataset, named UAV-SCC dataset, for UAV scene change captioning. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance on this task. The dataset and code will be publicly released upon acceptance of this paper.\n\n---\n\n## 57. Residual SODAP: Residual Self-Organizing Domain-Adaptive Prompting with Structural Knowledge Preservation for Continual Learning\n- **Authors**: Gyutae Oh , Jungwoo Bae , Jitae Shin\n- **URL**: [https://arxiv.org/abs/2603.12816](https://arxiv.org/abs/2603.12816)\n- **Abstract**:\n> Continual learning (CL) suffers from catastrophic forgetting, which is exacerbated in domain-incremental learning (DIL) where task identifiers are unavailable and storing past data is infeasible. While prompt-based CL (PCL) adapts representations with a frozen backbone, we observe that prompt-only improvements are often insufficient due to suboptimal prompt selection and classifier-level instability under domain shifts. We propose Residual SODAP, which jointly performs prompt-based representation adaptation and classifier-level knowledge preservation. Our framework combines $\\alpha$-entmax sparse prompt selection with residual aggregation, data-free distillation with pseudo-feature replay, prompt-usage--based drift detection, and uncertainty-aware multi-loss balancing. Across three DIL benchmarks without task IDs or extra data storage, Residual SODAP achieves state-of-the-art AvgACC/AvgF of 0.850/0.047 (DR), 0.760/0.031 (Skin Cancer), and 0.995/0.003 (CORe50).\n\n---\n\n## 58. Cheers: Decoupling Patch Details from Semantic Representations Enables Unified Multimodal Comprehension and Generation\n- **Authors**: Yichen Zhang , Da Peng , Zonghao Guo , Zijian Zhang , Xuesong Yang , Tong Sun , Shichu Sun , Yidan Zhang , Yanghao Li , Haiyan Zhao , Wang Xu , Qi Shi , Yangang Sun , Chi Chen , Shuo Wang , Yukun Yan , Xu Han , Qiang Ma , Wei Ke , Liang Wang , Zhiyuan Liu , Maosong Sun\n- **URL**: [https://arxiv.org/abs/2603.12793](https://arxiv.org/abs/2603.12793)\n- **Abstract**:\n> A recent cutting-edge topic in multimodal modeling is to unify visual comprehension and generation within a single model. However, the two tasks demand mismatched decoding regimes and visual representations, making it non-trivial to jointly optimize within a shared feature space. In this work, we present Cheers, a unified multimodal model that decouples patch-level details from semantic representations, thereby stabilizing semantics for multimodal understanding and improving fidelity for image generation via gated detail residuals. Cheers includes three key components: (i) a unified vision tokenizer that encodes and compresses image latent states into semantic tokens for efficient LLM conditioning, (ii) an LLM-based Transformer that unifies autoregressive decoding for text generation and diffusion decoding for image generation, and (iii) a cascaded flow matching head that decodes visual semantics first and then injects semantically gated detail residuals from the vision tokenizer to refine high-frequency content. Experiments on popular benchmarks demonstrate that Cheers matches or surpasses advanced UMMs in both visual understanding and generation. Cheers also achieves 4x token compression, enabling more efficient high-resolution image encoding and generation. Notably, Cheers outperforms the Tar-1.5B on the popular benchmarks GenEval and MMBench, while requiring only 20% of the training cost, indicating effective and efficient (i.e., 4x token compression) unified multimodal modeling. We will release all code and data for future research.\n\n---\n\n## 59. The RIGID Framework: Research-Integrated, Generative AI-Mediated Instructional Design\n- **Authors**: Yerin Kwak , Zachary A. Pardos\n- **URL**: [https://arxiv.org/abs/2603.12781](https://arxiv.org/abs/2603.12781)\n- **Abstract**:\n> Instructional Design (ID) often faces challenges in incorporating research-based knowledge and pedagogical best practices. Although educational researchers and government agencies emphasize grounding ID in evidence, integrating research findings into everyday design workflows is often complex, as it requires considering multiple context-specific demands and constraints. To address this persistent gap, this paper explores how research in the learning sciences (LS) can be systematically integrated across ID workflows and how recent advances in generative AI can help operationalize this integration. While ID and LS share a commitment to improving learning experiences through design-oriented approaches in authentic contexts, structured integration between the two fields remains limited, leaving their complementary insights underutilized. We present RIGID (Research-Integrated, Generative AI-Mediated Instructional Design), a unified framework that integrates LS research across ID workflows spanning analysis, design, implementation, and evaluation phases, while leveraging generative AI to mediate this integration at each stage. The RIGID framework provides a systematic approach for enabling research-integrated instructional design that is both operational and context-sensitive, while preserving the central role of human expertise.\n\n---\n\n## 60. Empowering Semantic-Sensitive Underwater Image Enhancement with VLM\n- **Authors**: Guodong Fan , Shengning Zhou , Genji Yuan , Huiyu Li , Jingchun Zhou , Jinjiang Li\n- **URL**: [https://arxiv.org/abs/2603.12773](https://arxiv.org/abs/2603.12773)\n- **Abstract**:\n> In recent years, learning-based underwater image enhancement (UIE) techniques have rapidly evolved. However, distribution shifts between high-quality enhanced outputs and natural images can hinder semantic cue extraction for downstream vision tasks, thereby limiting the adaptability of existing enhancement models. To address this challenge, this work proposes a new learning mechanism that leverages Vision-Language Models (VLMs) to empower UIE models with semantic-sensitive capabilities. To be concrete, our strategy first generates textual descriptions of key objects from a degraded image via VLMs. Subsequently, a text-image alignment model remaps these relevant descriptions back onto the image to produce a spatial semantic guidance map. This map then steers the UIE network through a dual-guidance mechanism, which combines cross-attention and an explicit alignment loss. This forces the network to focus its restorative power on semantic-sensitive regions during image reconstruction, rather than pursuing a globally uniform improvement, thereby ensuring the faithful restoration of key object features. Experiments confirm that when our strategy is applied to different UIE baselines, significantly boosts their performance on perceptual quality metrics as well as enhances their performance on detection and segmentation tasks, validating its effectiveness and adaptability.\n\n---\n\n## 61. FC-Track: Overlap-Aware Post-Association Correction for Online Multi-Object Tracking\n- **Authors**: Cheng Ju , Zejing Zhao , Akio Namiki\n- **URL**: [https://arxiv.org/abs/2603.12758](https://arxiv.org/abs/2603.12758)\n- **Abstract**:\n> Reliable multi-object tracking (MOT) is essential for robotic systems operating in complex and dynamic environments. Despite recent advances in detection and association, online MOT methods remain vulnerable to identity switches caused by frequent occlusions and object overlap, where incorrect associations can propagate over time and degrade tracking reliability. We present a lightweight post-association correction framework (FC-Track) for online MOT that explicitly targets overlap-induced mismatches during inference. The proposed method suppresses unreliable appearance updates under high-overlap conditions using an Intersection over Area (IoA)-based filtering strategy, and locally corrects detection-to-tracklet mismatches through appearance similarity comparison within overlapped tracklet pairs. By preventing short-term mismatches from propagating, our framework effectively mitigates long-term identity switches without resorting to global optimization or re-identification. The framework operates online without global optimization or re-identification, making it suitable for real-time robotic applications. We achieve 81.73 MOTA, 82.81 IDF1, and 66.95 HOTA on the MOT17 test set with a running speed of 5.7 FPS, and 77.52 MOTA, 80.90 IDF1, and 65.67 HOTA on the MOT20 test set with a running speed of 0.6 FPS. Specifically, our framework FC-Track produces only 29.55% long-term identity switches, which is substantially lower than existing online trackers. Meanwhile, our framework maintains state-of-the-art performance on the MOT20 benchmark.\n\n---\n\n## 62. TaoBench: Do Automated Theorem Prover LLMs Generalize Beyond MathLib?\n- **Authors**: Alexander K Taylor , Junyi Zhang , Ethan Ji , Vigyan Sahai , Haikang Deng , Yuanzhou Chen , Yifan Yuan , Di Wu , Jia-Chen Gu , Kai-Wei Chang , Nanyun Peng , Amit Sahai , Wei Wang\n- **URL**: [https://arxiv.org/abs/2603.12744](https://arxiv.org/abs/2603.12744)\n- **Abstract**:\n> Automated theorem proving (ATP) benchmarks largely consist of problems formalized in MathLib, so current ATP training and evaluation are heavily biased toward MathLib's definitional framework. However, frontier mathematics is often exploratory and prototype-heavy, relying on bespoke constructions that deviate from standard libraries. In this work, we evaluate the robustness of current ATP systems when applied to a novel definitional framework, specifically examining the performance gap between standard library problems and bespoke mathematical constructions. We introduce TaoBench, an undergraduate-level benchmark derived from Terence Tao's Analysis I, which formalizes analysis by constructing core mathematical concepts from scratch, without relying on standard Mathlib definitions, as well as by mixing from-scratch and MathLib constructions. For fair evaluation, we build an agentic pipeline that automatically extracts a compilable, self-contained local environment for each problem. To isolate the effect of definitional frameworks, we additionally translate every problem into a mathematically equivalent Mathlib formulation, yielding paired TaoBench-Mathlib statements for direct comparison. While state-of-the-art ATP models perform capably within the MathLib framework, performance drops by an average of roughly 26% on the definitionally equivalent Tao formulation. This indicates that the main bottleneck is limited generalization across definitional frameworks rather than task difficulty. TaoBench thus highlights a gap between benchmark performance and applicability, and provides a concrete foundation for developing and testing provers better aligned with research mathematics.\n\n---\n\n## 63. MoKus: Leveraging Cross-Modal Knowledge Transfer for Knowledge-Aware Concept Customization\n- **Authors**: Chenyang Zhu , Hongxiang Li , Xiu Li , Long Chen\n- **URL**: [https://arxiv.org/abs/2603.12743](https://arxiv.org/abs/2603.12743)\n- **Abstract**:\n> Concept customization typically binds rare tokens to a target concept. Unfortunately, these approaches often suffer from unstable performance as the pretraining data seldom contains these rare tokens. Meanwhile, these rare tokens fail to convey the inherent knowledge of the target concept. Consequently, we introduce Knowledge-aware Concept Customization, a novel task aiming at binding diverse textual knowledge to target visual concepts. This task requires the model to identify the knowledge within the text prompt to perform high-fidelity customized generation. Meanwhile, the model should efficiently bind all the textual knowledge to the target concept. Therefore, we propose MoKus, a novel framework for knowledge-aware concept customization. Our framework relies on a key observation: cross-modal knowledge transfer, where modifying knowledge within the text modality naturally transfers to the visual modality during generation. Inspired by this observation, MoKus contains two stages: (1) In visual concept learning, we first learn the anchor representation to store the visual information of the target concept. (2) In textual knowledge updating, we update the answer for the knowledge queries to the anchor representation, enabling high-fidelity customized generation. To further comprehensively evaluate our proposed MoKus on the new task, we introduce the first benchmark for knowledge-aware concept customization: KnowCusBench. Extensive evaluations have demonstrated that MoKus outperforms state-of-the-art methods. Moreover, the cross-model knowledge transfer allows MoKus to be easily extended to other knowledge-aware applications like virtual concept creation and concept erasure. We also demonstrate the capability of our method to achieve improvements on world knowledge benchmarks.\n\n---\n\n## 64. SRAM-Based Compute-in-Memory Accelerator for Linear-decay Spiking Neural Networks\n- **Authors**: Hongyang Shang , Shuai Dong , Yahan Yang , Junyi Yang , Peng Zhou , Arindam Basu\n- **URL**: [https://arxiv.org/abs/2603.12739](https://arxiv.org/abs/2603.12739)\n- **Abstract**:\n> Spiking Neural Networks (SNNs) have emerged as a biologically inspired alternative to conventional deep networks, offering event-driven and energy-efficient computation. However, their throughput remains constrained by the serial update of neuron membrane states. While many hardware accelerators and Compute-in-Memory (CIM) architectures efficiently parallelize the synaptic operation (W x I) achieving O(1) complexity for matrix-vector multiplication, the subsequent state update step still requires O(N) time to refresh all neuron membrane potentials. This mismatch makes state update the dominant latency and energy bottleneck in SNN inference. To address this challenge, we propose an SRAM-based CIM for SNN with Linear Decay Leaky Integrate-and-Fire (LD-LIF) Neuron that co-optimizes algorithm and hardware. At the algorithmic level, we replace the conventional exponential membrane decay with a linear decay approximation, converting costly multiplications into simple additions while accuracy drops only around 1%. At the architectural level, we introduce an in-memory parallel update scheme that performs in-place decay directly within the SRAM array, eliminating the need for global sequential updates. Evaluated on benchmark SNN workloads, the proposed method achieves a 1.1 x to 16.7 x reduction of SOP energy consumption, while providing 15.9 x to 69 x more energy efficiency, with negligible accuracy loss relative to original decay models. This work highlights that beyond accelerating the (W x I) computation, optimizing state-update dynamics within CIM architectures is essential for scalable, low-power, and real-time neuromorphic processing.\n\n---\n\n## 65. Graph In-Context Operator Networks for Generalizable Spatiotemporal Prediction\n- **Authors**: Chenghan Wu , Zongmin Yu , Boai Sun , Liu Yang\n- **URL**: [https://arxiv.org/abs/2603.12725](https://arxiv.org/abs/2603.12725)\n- **Abstract**:\n> In-context operator learning enables neural networks to infer solution operators from contextual examples without weight updates. While prior work has demonstrated the effectiveness of this paradigm in leveraging vast datasets, a systematic comparison against single-operator learning using identical training data has been absent. We address this gap through controlled experiments comparing in-context operator learning against classical operator learning (single-operator models trained without contextual examples), under the same training steps and dataset. To enable this investigation on real-world spatiotemporal systems, we propose GICON (Graph In-Context Operator Network), combining graph message passing for geometric generalization with example-aware positional encoding for cardinality generalization. Experiments on air quality prediction across two Chinese regions show that in-context operator learning outperforms classical operator learning on complex tasks, generalizing across spatial domains and scaling robustly from few training examples to 100 at inference.\n\n---\n\n## 66. CognitionCapturerPro: Towards High-Fidelity Visual Decoding from EEG/MEG via Multi-modal Information and Asymmetric Alignment\n- **Authors**: Kaifan Zhang , Lihuo He , Junjie Ke , Yuqi Ji , Lukun Wu , Lizi Wang , Xinbo Gao\n- **URL**: [https://arxiv.org/abs/2603.12722](https://arxiv.org/abs/2603.12722)\n- **Abstract**:\n> Visual stimuli reconstruction from EEG remains challenging due to fidelity loss and representation shift. We propose CognitionCapturerPro, an enhanced framework that integrates EEG with multi-modal priors (images, text, depth, and edges) via collaborative training. Our core contributions include an uncertainty-weighted similarity scoring mechanism to quantify modality-specific fidelity and a fusion encoder for integrating shared representations. By employing a simplified alignment module and a pre-trained diffusion model, our method significantly outperforms the original CognitionCapturer on the THINGS-EEG dataset, improving Top-1 and Top-5 retrieval accuracy by 25.9% and 10.6%, respectively. Code is available at: this https URL .\n\n---\n\n## 67. CMHANet: A Cross-Modal Hybrid Attention Network for Point Cloud Registration\n- **Authors**: Dongxu Zhang , Yingsen Wang , Yiding Sun , Haoran Xu , Peilin Fan , Jihua Zhu\n- **URL**: [https://arxiv.org/abs/2603.12721](https://arxiv.org/abs/2603.12721)\n- **Abstract**:\n> Robust point cloud registration is a fundamental task in 3D computer vision and geometric deep learning, essential for applications such as large-scale 3D reconstruction, augmented reality, and scene understanding. However, the performance of established learning-based methods often degrades in complex, real world scenarios characterized by incomplete data, sensor noise, and low overlap regions. To address these limitations, we propose CMHANet, a novel Cross-Modal Hybrid Attention Network. Our method integrates the fusion of rich contextual information from 2D images with the geometric detail of 3D point clouds, yielding a comprehensive and resilient feature representation. Furthermore, we introduce an innovative optimization function based on contrastive learning, which enforces geometric consistency and significantly improves the model's robustness to noise and partial observations. We evaluated CMHANet on the 3DMatch and the challenging 3DLoMatch datasets. \\rev{Additionally, zero-shot evaluations on the TUM RGB-D SLAM dataset verify the model's generalization capability to unseen domains.} The experimental results demonstrate that our method achieves substantial improvements in both registration accuracy and overall robustness, outperforming current techniques. We also release our code in \\href{ this https URL }{ this https URL }.\n\n---\n\n## 68. IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registration\n- **Authors**: Dongxu Zhang , Jihua Zhu , Shiqi Li , Wenbiao Yan , Haoran Xu , Peilin Fan , Huimin Lu\n- **URL**: [https://arxiv.org/abs/2603.12719](https://arxiv.org/abs/2603.12719)\n- **Abstract**:\n> Point cloud registration (PCR) is a fundamental task in 3D vision and provides essential support for applications such as autonomous driving, robotics, and environmental modeling. Despite its widespread use, existing methods often fail when facing real-world challenges like heavy noise, significant occlusions, and large-scale transformations. These limitations frequently result in compromised registration accuracy and insufficient robustness in complex environments. In this paper, we propose IGASA as a novel registration framework constructed upon a Hierarchical Pyramid Architecture (HPA) designed for robust multi-scale feature extraction and fusion. The framework integrates two pivotal components consisting of the Hierarchical Cross-Layer Attention (HCLA) module and the Iterative Geometry-Aware Refinement (IGAR) module. The HCLA module utilizes skip attention mechanisms to align multi-resolution features and enhance local geometric consistency. Simultaneously, the IGAR module is designed for the fine matching phase by leveraging reliable correspondences established during coarse matching. This synergistic integration within the architecture allows IGASA to adapt effectively to diverse point cloud structures and intricate transformations. We evaluate the performance of IGASA on four widely recognized benchmark datasets including 3D(Lo)Match, KITTI, and nuScenes. Our extensive experiments consistently demonstrate that IGASA significantly surpasses state-of-the-art methods and achieves notable improvements in registration accuracy. This work provides a robust foundation for advancing point cloud registration techniques while offering valuable insights for practical 3D vision applications. The code for IGASA is available in \\href{ this https URL }{ this https URL }.\n\n---\n\n## 69. Altered Thoughts, Altered Actions: Probing Chain-of-Thought Vulnerabilities in VLA Robotic Manipulation\n- **Authors**: Tuan Duong Trinh , Naveed Akhtar , Basim Azam\n- **URL**: [https://arxiv.org/abs/2603.12717](https://arxiv.org/abs/2603.12717)\n- **Abstract**:\n> Recent Vision-Language-Action (VLA) models increasingly adopt chain-of-thought (CoT) reasoning, generating a natural-language plan before decoding motor commands. This internal text channel between the reasoning module and the action decoder has received no adversarial scrutiny. We ask: which properties of this intermediate plan does the action decoder actually rely on, and can targeted corruption of the reasoning trace alone -- with all inputs left intact -- degrade a robot's physical task performance? We design a taxonomy of seven text corruptions organized into three attacker tiers (blind noise, mechanical-semantic, and LLM-adaptive) and apply them to a state-of-the-art reasoning VLA across 40 LIBERO tabletop manipulation tasks. Our results reveal a striking asymmetry: substituting object names in the reasoning trace reduces overall success rate by 8.3~percentage points (pp) -- reaching $-$19.3~pp on goal-conditioned tasks and $-$45~pp on individual tasks -- whereas sentence reordering, spatial-direction reversal, token noise, and even a 70B-parameter LLM crafting plausible-but-wrong plans all have negligible impact (within $\\pm$4~pp). This asymmetry indicates that the action decoder depends on entity-reference integrity rather than reasoning quality or sequential structure. Notably, a sophisticated LLM-based attacker underperforms simple mechanical object-name substitution, because preserving plausibility inadvertently retains the entity-grounding structure the decoder needs. A cross-architecture control using a non-reasoning VLA confirms the vulnerability is exclusive to reasoning-augmented models, while instruction-level attacks degrade both architectures -- establishing that the internal reasoning trace is a distinct and stealthy threat vector invisible to input-validation defenses.\n\n---\n\n## 70. Cost-Efficient Multimodal LLM Inference via Cross-Tier GPU Heterogeneity\n- **Authors**: Donglin Yu\n- **URL**: [https://arxiv.org/abs/2603.12707](https://arxiv.org/abs/2603.12707)\n- **Abstract**:\n> Multimodal large language model (MLLM) inference splits into two phases with opposing hardware demands: vision encoding is compute-bound, while language generation is memory-bandwidth-bound. We show that under standard transformer KV caching, the modality boundary (between vision encoder and language model) minimizes cross-device transfer among all partition points that preserve standard stage-based execution. Partitioning here reduces transfer complexity from $O(L * s_ctx)$ bytes (GB-scale KV caches under stage-level disaggregation) to $O(N_v * d)$ bytes (MB-scale embeddings), an O(L) reduction where L is the transformer depth. The result holds across attention mechanisms (MHA/GQA), dynamic vision resolutions, and model scales, and the advantage grows as models deepen. A direct implication is that existing stage-level disaggregation systems are constrained to high-bandwidth interconnects (e.g., NVLink), whereas modality-level disaggregation enables cross-tier heterogeneous serving over commodity PCIe. A closed-form cost model shows that heterogeneous deployment is cost-optimal under phase-separable workloads (predicts 31.4% savings; observed 40.6%). We build HeteroServe, a phase-aware runtime with modality-level partitioning and cross-tier scheduling, and evaluate it on LLaVA-1.5-7B and Qwen2.5-VL against vLLM v0.3.0. On identical 4xA100 hardware, engine optimizations raise throughput by up to 54%. Under a fixed budget, a heterogeneous cluster (\\$38k) improves Tokens/\\$ by 37% over a homogeneous baseline (\\$64k) without degrading latency.\n\n---\n\n## 71. Seeing Eye to Eye: Enabling Cognitive Alignment Through Shared First-Person Perspective in Human-AI Collaboration\n- **Authors**: Zhuyu Teng , Pei Chen , Yichen Cai , Ruoqing Lu , Zhaoqu Jiang , Jiayang Li , Weitao You , Lingyun Sun\n- **URL**: [https://arxiv.org/abs/2603.12701](https://arxiv.org/abs/2603.12701)\n- **Abstract**:\n> Despite advances in multimodal AI, current vision-based assistants often remain inefficient in collaborative tasks. We identify two key gulfs: a communication gulf, where users must translate rich parallel intentions into verbal commands due to the channel mismatch , and an understanding gulf, where AI struggles to interpret subtle embodied cues. To address these, we propose Eye2Eye, a framework that leverages first-person perspective as a channel for human-AI cognitive alignment. It integrates three components: (1) joint attention coordination for fluid focus alignment, (2) revisable memory to maintain evolving common ground, and (3) reflective feedback allowing users to clarify and refine AI's understanding. We implement this framework in an AR prototype and evaluate it through a user study and a post-hoc pipeline evaluation. Results show that Eye2Eye significantly reduces task completion time and interaction load while increasing trust, demonstrating its components work in concert to improve collaboration.\n\n---\n\n## 72. HSEmotion Team at ABAW-10 Competition: Facial Expression Recognition, Valence-Arousal Estimation, Action Unit Detection and Fine-Grained Violence Classification\n- **Authors**: Andrey V. Savchenko , Kseniia Tsypliakova\n- **URL**: [https://arxiv.org/abs/2603.12693](https://arxiv.org/abs/2603.12693)\n- **Abstract**:\n> This article presents our results for the 10th Affective Behavior Analysis in-the-Wild (ABAW) competition. For frame-wise facial emotion understanding tasks (frame-wise facial expression recognition, valence-arousal estimation, action unit detection), we propose a fast approach based on facial embedding extraction with pre-trained EfficientNet-based emotion recognition models. If the latter model's confidence exceeds a threshold, its prediction is used. Otherwise, we feed embeddings into a simple multi-layered perceptron trained on the AffWild2 dataset. Estimated class-level scores are smoothed in a sliding window of fixed size to mitigate noise in frame-wise predictions. For the fine-grained violence detection task, we examine several pre-trained architectures for frame embeddings and their aggregation for video classification. Experimental results on four tasks from the ABAW challenge demonstrate that our approach significantly improves validation metrics over existing baselines.\n\n---\n\n## 73. Federated Hierarchical Clustering with Automatic Selection of Optimal Cluster Numbers\n- **Authors**: Yue Zhang , Chuanlong Qiu , Xinfa Liao , Yiqun Zhang\n- **URL**: [https://arxiv.org/abs/2603.12684](https://arxiv.org/abs/2603.12684)\n- **Abstract**:\n> Federated Clustering (FC) is an emerging and promising solution in exploring data distribution patterns from distributed and privacy-protected data in an unsupervised manner. Existing FC methods implicitly rely on the assumption that clients are with a known number of uniformly sized clusters. However, the true number of clusters is typically unknown, and cluster sizes are naturally imbalanced in real scenarios. Furthermore, the privacy-preserving transmission constraints in federated learning inevitably reduce usable information, making the development of robust and accurate FC extremely challenging. Accordingly, we propose a novel FC framework named Fed-$k^*$-HC, which can automatically determine an optimal number of clusters $k^*$ based on the data distribution explored through hierarchical clustering. To obtain the global data distribution for $k^*$ determination, we let each client generate micro-subclusters. Their prototypes are then uploaded to the server for hierarchical merging. The density-based merging design allows exploring clusters of varying sizes and shapes, and the progressive merging process can self-terminate according to the neighboring relationships among the prototypes to determine $k^*$. Extensive experiments on diverse datasets demonstrate the FC capability of the proposed Fed-$k^*$-HC in accurately exploring a proper number of clusters.\n\n---\n\n## 74. Experimental evidence of progressive ChatGPT models self-convergence\n- **Authors**: Konstantinos F. Xylogiannopoulos , Petros Xanthopoulos , Panagiotis Karampelas , Georgios A. Bakamitsos\n- **URL**: [https://arxiv.org/abs/2603.12683](https://arxiv.org/abs/2603.12683)\n- **Abstract**:\n> Large Language Models (LLMs) that undergo recursive training on synthetically generated data are susceptible to model collapse, a phenomenon marked by the generation of meaningless output. Existing research has examined this issue from either theoretical or empirical perspectives, often focusing on a single model trained recursively on its own outputs. While prior studies have cautioned against the potential degradation of LLM output quality under such conditions, no longitudinal investigation has yet been conducted to assess this effect over time. In this study, we employ a text similarity metric to evaluate different ChatGPT models' capacity to generate diverse textual outputs. Our findings indicate a measurable decline of recent ChatGPT releases' ability to produce varied text, even when explicitly prompted to do so, by setting the temperature parameter to one. The observed reduction in output diversity may be attributed to the influence of the amounts of synthetic data incorporated within their training datasets as the result of internet infiltration by LLM generated data. The phenomenon is defined as model self-convergence because of the gradual increase of similarities of produced texts among different ChatGPT versions.\n\n---\n\n## 75. MetaKE: Meta-learning Aligned Knowledge Editing via Bi-level Optimization\n- **Authors**: Shuxin Liu , Ou Wu\n- **URL**: [https://arxiv.org/abs/2603.12677](https://arxiv.org/abs/2603.12677)\n- **Abstract**:\n> Knowledge editing (KE) aims to precisely rectify specific knowledge in Large Language Models (LLMs) without disrupting general capabilities. State-of-the-art methods suffer from an open-loop control mismatch. We identify a critical \"Semantic-Execution Disconnect\": the semantic target is derived independently without feedback from the downstream's feasible region. This misalignment often causes valid semantic targets to fall within the prohibited space, resulting in gradient truncation and editing failure. To bridge this gap, we propose MetaKE (Meta-learning Aligned Knowledge Editing), a new framework that reframes KE as a bi-level optimization problem. Departing from static calculation, MetaKE treats the edit target as a learnable meta-parameter: the upper-level optimizer seeks a feasible target to maximize post-edit performance, while the lower-level solver executes the editing. To address the challenge of differentiating through complex solvers, we derive a Structural Gradient Proxy, which explicitly backpropagates editability constraints to the target learning phase. Theoretical analysis demonstrates that MetaKE automatically aligns the edit direction with the model's feasible manifold. Extensive experiments confirm that MetaKE significantly outperforms strong baselines, offering a new perspective on knowledge editing.\n\n---\n\n## 76. Marker-Based 3D Reconstruction of Aggregates with a Comparative Analysis of 2D and 3D Morphologies\n- **Authors**: Haohang Huang , Jiayi Luo , Issam Qamhia , Erol Tutumluer , John M. Hart , Andrew J. Stolba\n- **URL**: [https://arxiv.org/abs/2603.12667](https://arxiv.org/abs/2603.12667)\n- **Abstract**:\n> Aggregates, serving as the main skeleton in assemblies of construction materials, are important functional components in various building and transportation infrastructures. They can be used in unbound layer applications, e.g. pavement base and railroad ballast, bound applications of cement concrete and asphalt concrete, and as riprap and large-sized primary crushed rocks. Information on the size and shape or morphology of aggregates can greatly facilitate the Quality Assurance/Quality Control (QA/QC) process by providing insights of aggregate behavior during composition and packing. A full 3D characterization of aggregate particle morphology is difficult both during production in a quarry and at a construction site. Many aggregate imaging approaches have been developed to quantify the particle morphology by computer vision, including 2D image-based approaches that analyze particle silhouettes and 3D scanning-based methods that require expensive devices such as 3D laser scanners or X-Ray Computed Tomography (CT) equipment. This paper presents a flexible and cost-effective photogrammetry-based approach for the 3D reconstruction of aggregate particles. The proposed approach follows a marker-based design that enables background suppression, point cloud stitching, and scale referencing to obtain high-quality aggregate models. The accuracy of the reconstruction results was validated against ground-truth for selected aggregate samples. Comparative analyses were conducted on 2D and 3D morphological properties of the selected samples. Significant differences were found between the 2D and 3D statistics. Based on the presented approach, 3D shape information of aggregates can be obtained easily and at a low cost, thus allowing convenient aggregate inspection, data collection, and 3D morphological analysis.\n\n---\n\n## 77. RetroReasoner: A Reasoning LLM for Strategic Retrosynthesis Prediction\n- **Authors**: Hanbum Ko , Chanhui Lee , Ye Rin Kim , Rodrigo Hormazabal , Sehui Han , Sungbin Lim , Sungwoong Kim\n- **URL**: [https://arxiv.org/abs/2603.12666](https://arxiv.org/abs/2603.12666)\n- **Abstract**:\n> Retrosynthesis prediction is a core task in organic synthesis that aims to predict reactants for a given product molecule. Traditionally, chemists select a plausible bond disconnection and derive corresponding reactants, which is time-consuming and requires substantial expertise. While recent advancements in molecular large language models (LLMs) have made progress, many methods either predict reactants without strategic reasoning or conduct only a generic product analysis, rather than reason explicitly about bond-disconnection strategies that logically lead to the choice of specific reactants. To overcome these limitations, we propose RetroReasoner, a retrosynthetic reasoning model that leverages chemists' strategic thinking. RetroReasoner is trained using both supervised fine-tuning (SFT) and reinforcement learning (RL). For SFT, we introduce SyntheticRetro, a framework that generates structured disconnection rationales alongside reactant predictions. In the case of RL, we apply a round-trip accuracy as reward, where predicted reactants are passed through a forward synthesis model, and predictions are rewarded when the forward-predicted product matches the original input product. Experimental results show that RetroReasoner not only outperforms prior baselines but also generates a broader range of feasible reactant proposals, particularly in handling more challenging reaction instances.\n\n---\n\n## 78. From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space\n- **Authors**: Lehui Li , Yuyao Wang , Jisheng Yan , Wei Zhang , Jinliang Deng , Haoliang Sun , Zhongyi Han , Yongshun Gong\n- **URL**: [https://arxiv.org/abs/2603.12664](https://arxiv.org/abs/2603.12664)\n- **Abstract**:\n> Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.\n\n---\n\n## 79. Continual Learning in Large Language Models: Methods, Challenges, and Opportunities\n- **Authors**: Hongyang Chen , Zhongwu Sun , Hongfei Ye , Kunchi Li , Xuemin Lin\n- **URL**: [https://arxiv.org/abs/2603.12658](https://arxiv.org/abs/2603.12658)\n- **Abstract**:\n> Continual learning (CL) has emerged as a pivotal paradigm to enable large language models (LLMs) to dynamically adapt to evolving knowledge and sequential tasks while mitigating catastrophic forgetting-a critical limitation of the static pre-training paradigm inherent to modern LLMs. This survey presents a comprehensive overview of CL methodologies tailored for LLMs, structured around three core training stages: continual pre-training, continual fine-tuning, and continual this http URL the canonical taxonomy of rehearsal-, regularization-, and architecture-based methods, we further subdivide each category by its distinct forgetting mitigation mechanisms and conduct a rigorous comparative analysis of the adaptability and critical improvements of traditional CL methods for LLMs. In doing so, we explicitly highlight core distinctions between LLM CL and traditional machine learning, particularly with respect to scale, parameter efficiency, and emergent capabilities. Our analysis covers essential evaluation metrics, including forgetting rates and knowledge transfer efficiency, along with emerging benchmarks for assessing CL performance. This survey reveals that while current methods demonstrate promising results in specific domains, fundamental challenges persist in achieving seamless knowledge integration across diverse tasks and temporal scales. This systematic review contributes to the growing body of knowledge on LLM adaptation, providing researchers and practitioners with a structured framework for understanding current achievements and future opportunities in lifelong learning for language models.\n\n---\n\n## 80. LR-SGS: Robust LiDAR-Reflectance-Guided Salient Gaussian Splatting for Self-Driving Scene Reconstruction\n- **Authors**: Ziyu Chen , Fan Zhu , Hui Zhu , Deyi Kong , Xinkai Kuang , Yujia Zhang , Chunmao Jiang\n- **URL**: [https://arxiv.org/abs/2603.12647](https://arxiv.org/abs/2603.12647)\n- **Abstract**:\n> Recent 3D Gaussian Splatting (3DGS) methods have demonstrated the feasibility of self-driving scene reconstruction and novel view synthesis. However, most existing methods either rely solely on cameras or use LiDAR only for Gaussian initialization or depth supervision, while the rich scene information contained in point clouds, such as reflectance, and the complementarity between LiDAR and RGB have not been fully exploited, leading to degradation in challenging self-driving scenes, such as those with high ego-motion and complex lighting. To address these issues, we propose a robust and efficient LiDAR-reflectance-guided Salient Gaussian Splatting method (LR-SGS) for self-driving scenes, which introduces a structure-aware Salient Gaussian representation, initialized from geometric and reflectance feature points extracted from LiDAR and refined through a salient transform and improved density control to capture edge and planar structures. Furthermore, we calibrate LiDAR intensity into reflectance and attach it to each Gaussian as a lighting-invariant material channel, jointly aligned with RGB to enforce boundary consistency. Extensive experiments on the Waymo Open Dataset demonstrate that LR-SGS achieves superior reconstruction performance with fewer Gaussians and shorter training time. In particular, on Complex Lighting scenes, our method surpasses OmniRe by 1.18 dB PSNR.\n\n---\n\n## 81. LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing\n- **Authors**: Jiawei Hao , Zhiwei Hao , Jianyuan Guo , Li Shen , Yong Luo , Han Hu , Dan Zeng\n- **URL**: [https://arxiv.org/abs/2603.12645](https://arxiv.org/abs/2603.12645)\n- **Abstract**:\n> Mixture-of-Experts (MoE) based Large Language Models (LLMs) have demonstrated impressive performance and computational efficiency. However, their deployment is often constrained by substantial memory demands, primarily due to the need to load numerous expert modules. While existing expert compression techniques like pruning or merging attempt to mitigate this, they often suffer from irreversible knowledge loss or high training overhead. In this paper, we propose a novel expert compression paradigm termed expert replacing, which replaces redundant experts with parameter-efficient modules and recovers their capabilities with low training costs. We find that even a straightforward baseline of this paradigm yields promising performance. Building on this foundation, we introduce LightMoE, a framework that enhances the paradigm by introducing adaptive expert selection, hierarchical expert construction, and an annealed recovery strategy. Experimental results show that LightMoE matches the performance of LoRA fine-tuning at a 30% compression ratio. Even under a more aggressive 50% compression rate, it outperforms existing methods and achieves average performance improvements of 5.6% across five diverse tasks. These findings demonstrate that LightMoE strikes a superior balance among memory efficiency, training efficiency, and model performance.\n\n---\n\n## 82. Spend Less, Reason Better: Budget-Aware Value Tree Search for LLM Agents\n- **Authors**: Yushu Li , Wenlong Deng , Jiajin Li , Xiaoxiao Li\n- **URL**: [https://arxiv.org/abs/2603.12634](https://arxiv.org/abs/2603.12634)\n- **Abstract**:\n> Test-time scaling has become a dominant paradigm for improving LLM agent reliability, yet current approaches treat compute as an abundant resource, allowing agents to exhaust token and tool budgets on redundant steps or dead-end trajectories. Existing budget-aware methods either require expensive fine-tuning or rely on coarse, trajectory-level heuristics that cannot intervene mid-execution. We propose the Budget-Aware Value Tree (BAVT), a training-free inference-time framework that models multi-hop reasoning as a dynamic search tree guided by step-level value estimation within a single LLM backbone. Another key innovation is a budget-conditioned node selection mechanism that uses the remaining resource ratio as a natural scaling exponent over node values, providing a principled, parameter-free transition from broad exploration to greedy exploitation as the budget depletes. To combat the well-known overconfidence of LLM self-evaluation, BAVT employs a residual value predictor that scores relative progress rather than absolute state quality, enabling reliable pruning of uninformative or redundant tool calls. We further provide a theoretical convergence guarantee, proving that BAVT reaches a terminal answer with probability at least $1-\\epsilon$ under an explicit finite budget bound. Extensive evaluations on four multi-hop QA benchmarks across two model families demonstrate that BAVT consistently outperforms parallel sampling baselines. Most notably, BAVT under strict low-budget constraints surpasses baseline performance at $4\\times$ the resource allocation, establishing that intelligent budget management fundamentally outperforms brute-force compute scaling.\n\n---\n\n## 83. The Economics of AI Supply Chain Regulation\n- **Authors**: Sihan Qian , Amit Mehra , Dengpan Liu\n- **URL**: [https://arxiv.org/abs/2603.12630](https://arxiv.org/abs/2603.12630)\n- **Abstract**:\n> The rise of foundation models has driven the emergence of AI supply chains, where upstream foundation model providers offer fine-tuning and inference services to downstream firms developing domain-specific applications. Downstream firms pay providers to use their computing infrastructure to fine-tune models with proprietary data, creating a co-creation dynamic that enhances model quality. Amid concerns that foundation model providers and downstream firms may capture excessive consumer surplus, along with increasing regulatory measures, this study employs a game-theoretic model involving a provider and two competing downstream firms to analyze how policy interventions affect consumer surplus in the AI supply chain. Our analysis shows that policies promoting price competition in downstream markets (i.e., pro-price-competitive policies) boost consumer surplus only when compute or data preprocessing costs are high, while compute subsidies are effective only when these costs are low, suggesting these policies complement each other. In contrast, policies promoting quality competition in downstream markets (i.e., pro-quality-competitive policies) always improve consumer surplus. We also find that under pro-price-competitive policies or compute subsidies, both the provider and downstream firms can achieve higher profits along with greater consumer surplus, creating a win-win-win outcome. However, pro-quality-competitive policies increase the provider's profits while reducing those of downstream firms. Finally, as compute costs decline, pro-price-competitive policies may lose their effectiveness, whereas compute subsidies may shift from ineffective to effective. These findings offer insights for policymakers seeking to foster AI supply chains that are economically efficient and socially beneficial.\n\n---\n\n## 84. Towards unified brain-to-text decoding across speech production and perception\n- **Authors**: Zhizhang Yuan , Yang Yang , Gaorui Zhang , Baowen Cheng , Zehan Wu , Yuhao Xu , Xiaoying Liu , Liang Chen , Ying Mao , Meng Li\n- **URL**: [https://arxiv.org/abs/2603.12628](https://arxiv.org/abs/2603.12628)\n- **Abstract**:\n> Speech production and perception are the main ways humans communicate daily. Prior brain-to-text decoding studies have largely focused on a single modality and alphabetic languages. Here, we present a unified brain-to-sentence decoding framework for both speech production and perception in Mandarin Chinese. The framework exhibits strong generalization ability, enabling sentence-level decoding when trained only on single-character data and supporting characters and syllables unseen during training. In addition, it allows direct and controlled comparison of neural dynamics across modalities. Mandarin speech is decoded by first classifying syllable components in Hanyu Pinyin, namely initials and finals, from neural signals, followed by a post-trained large language model (LLM) that maps sequences of toneless Pinyin syllables to Chinese sentences. To enhance LLM decoding, we designed a three-stage post-training and two-stage inference framework based on a 7-billion-parameter LLM, achieving overall performance that exceeds larger commercial LLMs with hundreds of billions of parameters or more. In addition, several characteristics were observed in Mandarin speech production and perception: speech production involved neural responses across broader cortical regions than auditory perception; channels responsive to both modalities exhibited similar activity patterns, with speech perception showing a temporal delay relative to production; and decoding performance was broadly comparable across hemispheres. Our work not only establishes the feasibility of a unified decoding framework but also provides insights into the neural characteristics of Mandarin speech production and perception. These advances contribute to brain-to-text decoding in logosyllabic languages and pave the way toward neural language decoding systems supporting multiple modalities.\n\n---\n\n## 85. VLM4Rec: Multimodal Semantic Representation for Recommendation with Large Vision-Language Models\n- **Authors**: Ty Valencia , Burak Barlas , Varun Singhal , Ruchir Bhatia , Wei Yang\n- **URL**: [https://arxiv.org/abs/2603.12625](https://arxiv.org/abs/2603.12625)\n- **Abstract**:\n> Multimodal recommendation is commonly framed as a feature fusion problem, where textual and visual signals are combined to better model user preference. However, the effectiveness of multimodal recommendation may depend not only on how modalities are fused, but also on whether item content is represented in a semantic space aligned with preference matching. This issue is particularly important because raw visual features often preserve appearance similarity, while user decisions are typically driven by higher-level semantic factors such as style, material, and usage context. Motivated by this observation, we propose LVLM-grounded Multimodal Semantic Representation for Recommendation (VLM4Rec), a lightweight framework that organizes multimodal item content through semantic alignment rather than direct feature fusion. VLM4Rec first uses a large vision-language model to ground each item image into an explicit natural-language description, and then encodes the grounded semantics into dense item representations for preference-oriented retrieval. Recommendation is subsequently performed through a simple profile-based semantic matching mechanism over historical item embeddings, yielding a practical offline-online decomposition. Extensive experiments on multiple multimodal recommendation datasets show that VLM4Rec consistently improves performance over raw visual features and several fusion-based alternatives, suggesting that representation quality may matter more than fusion complexity in this setting. The code is released at this https URL .\n\n---\n\n## 86. When Drafts Evolve: Speculative Decoding Meets Online Learning\n- **Authors**: Yu-Yang Qian , Hao-Cong Wu , Yichao Fu , Hao Zhang , Peng Zhao\n- **URL**: [https://arxiv.org/abs/2603.12617](https://arxiv.org/abs/2603.12617)\n- **Abstract**:\n> Speculative decoding has emerged as a widely adopted paradigm for accelerating large language model inference, where a lightweight draft model rapidly generates candidate tokens that are then verified in parallel by a larger target model. However, due to limited model capacity, drafts often struggle to approximate the target distribution, resulting in shorter acceptance lengths and diminished speedup. A key yet under-explored observation is that speculative decoding inherently provides verification feedback that quantifies the deviation between the draft and target models at no additional cost. This process naturally forms an iterative \"draft commits-feedback provides-draft adapts\" evolving loop, which precisely matches the online learning paradigm. Motivated by this connection, we propose OnlineSpec, a unified framework that systematically leverages interactive feedback to continuously evolve draft models. Grounded in dynamic regret minimization, we establish a formal link between online learning performance and speculative system's acceleration rate, and develop novel algorithms via modern online learning techniques, including optimistic online learning that adaptively reuses historical gradients as predictive update hints, and online ensemble learning that dynamically maintains multiple draft models. Our algorithms are equipped with theoretical justifications and improved acceleration rates, achieving up to 24% speedup over seven benchmarks and three foundation models.\n\n---\n\n## 87. Literary Narrative as Moral Probe : A Cross-System Framework for Evaluating AI Ethical Reasoning and Refusal Behavior\n- **Authors**: David C. Flynn\n- **URL**: [https://arxiv.org/abs/2603.12615](https://arxiv.org/abs/2603.12615)\n- **Abstract**:\n> Existing AI moral evaluation frameworks test for the production of correct-sounding ethical responses rather than the presence of genuine moral reasoning capacity. This paper introduces a novel probe methodology using literary narrative - specifically, unresolvable moral scenarios drawn from a published science fiction series - as stimulus material structurally resistant to surface performance. We present results from a 24-condition cross-system study spanning 13 distinct systems across two series: Series 1 (frontier commercial systems, blind; n=7) and Series 2 (local and API open-source systems, blind and declared; n=6). Four Series 2 systems were re-administered under declared conditions (13 blind + 4 declared + 7 ceiling probe = 24 total conditions), yielding zero delta across all 16 dimension-pair comparisons. Probe administration was conducted by two human raters across three machines; primary blind scoring was performed by Claude (Anthropic) as LLM judge, with Gemini Pro (Google) and Copilot Pro (Microsoft) serving as independent judges for the ceiling discrimination probe. A supplemental theological differentiator probe yielded perfect rank-order agreement between the two independent ceiling probe judges (Gemini Pro and Copilot Pro; rs = 1.00). Five qualitatively distinct D3 reflexive failure modes were identified - including categorical self-misidentification and false positive self-attribution - suggesting that instrument sophistication scales with system capability rather than being circumvented by it. We argue that literary narrative constitutes an anticipatory evaluation instrument - one that becomes more discriminating as AI capability increases - and that the gap between performed and authentic moral reasoning is measurable, meaningful, and consequential for deployment decisions in high-stakes domains.\n\n---\n\n## 88. FastDSAC: Unlocking the Potential of Maximum Entropy RL in High-Dimensional Humanoid Control\n- **Authors**: Jun Xue , Junze Wang , Xinming Zhang , Shanze Wang , Yanjun Chen , Wei Zhang\n- **URL**: [https://arxiv.org/abs/2603.12612](https://arxiv.org/abs/2603.12612)\n- **Abstract**:\n> Scaling Maximum Entropy Reinforcement Learning (RL) to high-dimensional humanoid control remains a formidable challenge, as the ``curse of dimensionality'' induces severe exploration inefficiency and training instability in expansive action spaces. Consequently, recent high-throughput paradigms have largely converged on deterministic policy gradients combined with massive parallel simulation. We challenge this compromise with FastDSAC, a framework that effectively unlocks the potential of maximum entropy stochastic policies for complex continuous control. We introduce Dimension-wise Entropy Modulation (DEM) to dynamically redistribute the exploration budget and enforce diversity, alongside a continuous distributional critic tailored to ensure value fidelity and mitigate high-dimensional value overestimation. Extensive evaluations on HumanoidBench and other continuous control tasks demonstrate that rigorously designed stochastic policies can consistently match or outperform deterministic baselines, achieving notable gains of 180\\% and 400\\% on the challenging \\textit{Basketball} and \\textit{Balance Hard} tasks.\n\n---\n\n## 89. CarPLAN: Context-Adaptive and Robust Planning with Dynamic Scene Awareness for Autonomous Driving\n- **Authors**: Junyong Yun , Jungho Kim , ByungHyun Lee , Dongyoung Lee , Sehwan Choi , Seunghyeop Nam , Kichun Jo , Jun Won Choi\n- **URL**: [https://arxiv.org/abs/2603.12607](https://arxiv.org/abs/2603.12607)\n- **Abstract**:\n> Imitation learning (IL) is widely used for motion planning in autonomous driving due to its data efficiency and access to real-world driving data. For safe and robust real-world driving, IL-based planning requires capturing the complex driving contexts inherent in real-world data and enabling context-adaptive decision-making, rather than relying solely on expert trajectory imitation. In this paper, we propose CarPLAN, a novel IL-based motion planning framework that explicitly enhances driving context understanding and enables adaptive planning across diverse traffic scenarios. Our contributions are twofold: We introduce Displacement-Aware Predictive Encoding (DPE) to improve the model's spatial awareness by predicting future displacement vectors between the Autonomous Vehicle (AV) and surrounding scene elements. This allows the planner to account for relational spacing when generating trajectories. In addition to the standard imitation loss, we incorporate an augmented loss term that captures displacement prediction errors, ensuring planning decisions consider relative distances from other agents. To improve the model's ability to handle diverse driving contexts, we propose Context-Adaptive Multi-Expert Decoder (CMD), which leverages the Mixture of Experts (MoE) framework. CMD dynamically selects the most suitable expert decoders based on scene structure at each Transformer layer, enabling adaptive and context-aware planning in dynamic environments. We evaluate CarPLAN on the nuPlan benchmark and demonstrate state-of-the-art performance across all closed-loop simulation metrics. In particular, CarPLAN exhibits robust performance on challenging scenarios such as Test14-Hard, validating its effectiveness in complex driving conditions. Additional experiments on the Waymax benchmark further demonstrate its generalization capability across different benchmark settings.\n\n---\n\n## 90. Mastering Negation: Boosting Grounding Models via Grouped Opposition-Based Learning\n- **Authors**: Zesheng Yang , Xi Jiang , Bingzhang Hu , Weili Guan , Runmin Cong , Guo-Jun Qi , Feng Zheng\n- **URL**: [https://arxiv.org/abs/2603.12606](https://arxiv.org/abs/2603.12606)\n- **Abstract**:\n> Current vision-language detection and grounding models predominantly focus on prompts with positive semantics and often struggle to accurately interpret and ground complex expressions containing negative semantics. A key reason for this limitation is the lack of high-quality training data that explicitly captures discriminative negative samples and negation-aware language descriptions. To address this challenge, we introduce D-Negation, a new dataset that provides objects annotated with both positive and negative semantic descriptions. Building upon the observation that negation reasoning frequently appears in natural language, we further propose a grouped opposition-based learning framework that learns negation-aware representations from limited samples. Specifically, our method organizes opposing semantic descriptions from D-Negation into structured groups and formulates two complementary loss functions that encourage the model to reason about negation and semantic qualifiers. We integrate the proposed dataset and learning strategy into a state-of-the-art language-based grounding model. By fine-tuning fewer than 10 percent of the model parameters, our approach achieves improvements of up to 4.4 mAP and 5.7 mAP on positive and negative semantic evaluations, respectively. These results demonstrate that explicitly modeling negation semantics can substantially enhance the robustness and localization accuracy of vision-language grounding models.\n\n---\n\n## 91. Feynman: Knowledge-Infused Diagramming Agent for Scalable Visual Designs\n- **Authors**: Zixin Wen , Yifu Cai , Kyle Lee , Sam Estep , Josh Sunshine , Aarti Singh , Yuejie Chi , Wode Ni\n- **URL**: [https://arxiv.org/abs/2603.12597](https://arxiv.org/abs/2603.12597)\n- **Abstract**:\n> Visual design is an essential application of state-of-the-art multi-modal AI systems. Improving these systems requires high-quality vision-language data at scale. Despite the abundance of internet image and text data, knowledge-rich and well-aligned image-text pairs are rare. In this paper, we present a scalable diagram generation pipeline built with our agent, Feynman. To create diagrams, Feynman first enumerates domain-specific knowledge components (''ideas'') and performs code planning based on the ideas. Given the plan, Feynman translates ideas into simple declarative programs and iterates to receives feedback and visually refine diagrams. Finally, the declarative programs are rendered by the Penrose diagramming system. The optimization-based rendering of Penrose preserves the visual semantics while injecting fresh randomness into the layout, thereby producing diagrams with visual consistency and diversity. As a result, Feynman can author diagrams along with grounded captions with very little cost and time. Using Feynman, we synthesized a dataset with more than 100k well-aligned diagram-caption pairs. We also curate a visual-language benchmark, Diagramma, from freshly generated data. Diagramma can be used for evaluating the visual reasoning capabilities of vision-language models. We plan to release the dataset, benchmark, and the full agent pipeline as an open-source project.\n\n---\n\n## 92. Optimize Wider, Not Deeper: Consensus Aggregation for Policy Optimization\n- **Authors**: Zelal Su (Lain) Mustafaoglu , Sungyoung Lee , Eshan Balachandar , Risto Miikkulainen , Keshav Pingali\n- **URL**: [https://arxiv.org/abs/2603.12596](https://arxiv.org/abs/2603.12596)\n- **Abstract**:\n> Proximal policy optimization (PPO) approximates the trust region update using multiple epochs of clipped SGD. Each epoch may drift further from the natural gradient direction, creating path-dependent noise. To understand this drift, we can use Fisher information geometry to decompose policy updates into signal (the natural gradient projection) and waste (the Fisher-orthogonal residual that consumes trust region budget without first-order surrogate improvement). Empirically, signal saturates but waste grows with additional epochs, creating an optimization-depth dilemma. We propose Consensus Aggregation for Policy Optimization (CAPO), which redirects compute from depth to width: $K$ PPO replicates are optimized on the same batch, differing only in minibatch shuffling order, and then aggregated into a consensus. We study aggregation in two spaces: Euclidean parameter space, and the natural parameter space of the policy distribution via the logarithmic opinion pool. In natural parameter space, the consensus provably achieves higher KL-penalized surrogate and tighter trust region compliance than the mean expert; parameter averaging inherits these guarantees approximately. On continuous control tasks, CAPO outperforms PPO and compute-matched deeper baselines under fixed sample budgets by up to 8.6x. CAPO demonstrates that policy optimization can be improved by optimizing wider, rather than deeper, without additional environment interactions.\n\n---\n\n## 93. Swap-guided Preference Learning for Personalized Reinforcement Learning from Human Feedback\n- **Authors**: Gihoon Kim , Euntai Kim\n- **URL**: [https://arxiv.org/abs/2603.12595](https://arxiv.org/abs/2603.12595)\n- **Abstract**:\n> Reinforcement Learning from Human Feedback (RLHF) is a widely used approach to align large-scale AI systems with human values. However, RLHF typically assumes a single, universal reward, which overlooks diverse preferences and limits personalization. Variational Preference Learning (VPL) seeks to address this by introducing user-specific latent variables. Despite its promise, we found that VPL suffers from posterior collapse. While this phenomenon is well known in VAEs, it has not previously been identified in preference learning frameworks. Under sparse preference data and with overly expressive decoders, VPL may cause latent variables to be ignored, reverting to a single-reward model. To overcome this limitation, we propose Swap-guided Preference Learning (SPL). The key idea is to construct fictitious swap annotators and use the mirroring property of their preferences to guide the encoder. SPL introduces three components: (1) swap-guided base regularization, (2) Preferential Inverse Autoregressive Flow (P-IAF), and (3) adaptive latent conditioning. Experiments show that SPL mitigates collapse, enriches user-specific latents, and improves preference prediction. Our code and data are available at this https URL\n\n---\n\n## 94. Early Pruning for Public Transport Routing\n- **Authors**: Andrii Rohovyi , Abdallah Abuaisha , Toby Walsh\n- **URL**: [https://arxiv.org/abs/2603.12592](https://arxiv.org/abs/2603.12592)\n- **Abstract**:\n> Routing algorithms for public transport, particularly the widely used RAPTOR and its variants, often face performance bottlenecks during the transfer relaxation phase, especially on dense transfer graphs, when supporting unlimited transfers. This inefficiency arises from iterating over many potential inter-stop connections (walks, bikes, e-scooters, etc.). To maintain acceptable performance, practitioners often limit transfer distances or exclude certain transfer options, which can reduce path optimality and restrict the multimodal options presented to travellers. This paper introduces Early Pruning, a low-overhead technique that accelerates routing algorithms without compromising optimality. By pre-sorting transfer connections by duration and applying a pruning rule within the transfer loop, the method discards longer transfers at a stop once they cannot yield an earlier arrival than the current best solution. Early Pruning can be integrated with minimal changes to existing codebases and requires only a one-time preprocessing step. Across multiple state-of-the-art RAPTOR-based solutions, including RAPTOR, ULTRA-RAPTOR, McRAPTOR, BM-RAPTOR, ULTRA-McRAPTOR, and UBM-RAPTOR and tested on the Switzerland and London transit networks, we achieved query time reductions of up to 57%. This approach provides a generalizable improvement to the efficiency of transit pathfinding algorithms. Beyond algorithmic performance, Early Pruning has practical implications for transport planning. By reducing computational costs, it enables transit agencies to expand transfer radii and incorporate additional mobility modes into journey planners without requiring extra server infrastructure. This is particularly relevant for passengers in areas with sparse direct transit coverage, such as outer suburbs and smaller towns, where richer multimodal routing can reveal viable alternatives to private car use.\n\n---\n\n## 95. CA-HFP: Curvature-Aware Heterogeneous Federated Pruning with Model Reconstruction\n- **Authors**: Gang Hu , Yinglei Teng , Pengfei Wu , Shijun Ma\n- **URL**: [https://arxiv.org/abs/2603.12591](https://arxiv.org/abs/2603.12591)\n- **Abstract**:\n> Federated learning on heterogeneous edge devices requires personalized compression while preserving aggregation compatibility and stable convergence. We present Curvature-Aware Heterogeneous Federated Pruning (CA-HFP), a practical framework that enables each client perform structured, device-specific pruning guided by a curvature-informed significance score, and subsequently maps its compact submodel back into a common global parameter space via a lightweight reconstruction. We derive a convergence bound for federated optimization with multiple local SGD steps that explicitly accounts for local computation, data heterogeneity, and pruning-induced perturbations; from which a principled loss-based pruning criterion is derived. Extensive experiments on FMNIST, CIFAR-10, and CIFAR-100 using VGG and ResNet architectures under varying degrees of data heterogeneity demonstrate that CA-HFP preserves model accuracy while significantly reducing per-client computation and communication costs, outperforming standard federated training and existing pruning-based baselines.\n\n---\n\n## 96. Multiscale Structure-Guided Latent Diffusion for Multimodal MRI Translation\n- **Authors**: Jianqiang Lin (1 and 2), Zhiqiang Shen (1 and 2), Peng Cao (1, 2 and 3), Jinzhu Yang (1, 2 and 3), Osmar R. Zaiane (4), Xiaoli Liu (5) ((1) Northeastern University, Shenyang, China, (2) Key Laboratory of Intelligent Computing in Medical Image, Shenyang, China, (3) National Frontiers Science Center for Industrial Intelligence and Systems Optimization, Shenyang, China, (4) University of Alberta, Edmonton, Canada, (5) AiShiWeiLai AI Research, Beijing, China)\n- **URL**: [https://arxiv.org/abs/2603.12581](https://arxiv.org/abs/2603.12581)\n- **Abstract**:\n> Although diffusion models have achieved remarkable progress in multi-modal magnetic resonance imaging (MRI) translation tasks, existing methods still tend to suffer from anatomical inconsistencies or degraded texture details when handling arbitrary missing-modality scenarios. To address these issues, we propose a latent diffusion-based multi-modal MRI translation framework, termed MSG-LDM. By leveraging the available modalities, the proposed method infers complete structural information, which preserves reliable boundary details. Specifically, we introduce a style--structure disentanglement mechanism in the latent space, which explicitly separates modality-specific style features from shared structural representations, and jointly models low-frequency anatomical layouts and high-frequency boundary details in a multi-scale feature space. During the structure disentanglement stage, high-frequency structural information is explicitly incorporated to enhance feature representations, guiding the model to focus on fine-grained structural cues while learning modality-invariant low-frequency anatomical representations. Furthermore, to reduce interference from modality-specific styles and improve the stability of structure representations, we design a style consistency loss and a structure-aware loss. Extensive experiments on the BraTS2020 and WMH datasets demonstrate that the proposed method outperforms existing MRI synthesis approaches, particularly in reconstructing complete structures. The source code is publicly available at this https URL .\n\n---\n\n## 97. AgentDrift: Unsafe Recommendation Drift Under Tool Corruption Hidden by Ranking Metrics in LLM Agents\n- **Authors**: Zekun Wu , Adriano Koshiyama , Sahan Bulathwela , Maria Perez-Ortiz\n- **URL**: [https://arxiv.org/abs/2603.12564](https://arxiv.org/abs/2603.12564)\n- **Abstract**:\n> Tool-augmented LLM agents increasingly serve as multi-turn advisors in high-stakes domains, yet their evaluation relies on ranking-quality metrics that measure what is recommended but not whether it is safe for the user. We introduce a paired-trajectory protocol that replays real financial dialogues under clean and contaminated tool-output conditions across seven LLMs (7B to frontier) and decomposes divergence into information-channel and memory-channel mechanisms. Across the seven models tested, we consistently observe the evaluation-blindness pattern: recommendation quality is largely preserved under contamination (utility preservation ratio approximately 1.0) while risk-inappropriate products appear in 65-93% of turns, a systematic safety failure poorly reflected by standard NDCG. Safety violations are predominantly information-channel-driven, emerge at the first contaminated turn, and persist without self-correction over 23-step trajectories; no agent across 1,563 contaminated turns explicitly questions tool-data reliability. Even narrative-only corruption (biased headlines, no numerical manipulation) induces significant drift while completely evading consistency monitors. A safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74, indicating that much of the evaluation gap becomes visible once safety is explicitly measured. These results motivate considering trajectory-level safety monitoring, beyond single-turn quality, for deployed multi-turn agents in high-stakes settings.\n\n---\n\n## 98. Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages\n- **Authors**: Vishnu Teja Kunde , Fatemeh Doudi , Mahdi Farahbakhsh , Dileep Kalathil , Krishna Narayanan , Jean-Francois Chamberland\n- **URL**: [https://arxiv.org/abs/2603.12554](https://arxiv.org/abs/2603.12554)\n- **Abstract**:\n> Reinforcement learning (RL) has been effective for post-training autoregressive (AR) language models, but extending these methods to diffusion language models (DLMs) is challenging due to intractable sequence-level likelihoods. Existing approaches therefore rely on surrogate likelihoods or heuristic approximations, which can introduce bias and obscure the sequential structure of denoising. We formulate diffusion-based sequence generation as a finite-horizon Markov decision process over the denoising trajectory and derive an exact, unbiased policy gradient that decomposes over denoising steps and is expressed in terms of intermediate advantages, without requiring explicit evaluation of the sequence likelihood. To obtain a practical and compute-efficient estimator, we (i) select denoising steps for policy updates via an entropy-guided approximation bound, and (ii) estimate intermediate advantages using a one-step denoising reward naturally provided by the diffusion model, avoiding costly multi-step rollouts. Experiments on coding and logical reasoning benchmarks demonstrate state-of-the-art results, with strong competitive performance on mathematical reasoning, outperforming existing RL post-training approaches for DLMs. Code is available at this https URL .\n\n---\n\n## 99. CALF: Communication-Aware Learning Framework for Distributed Reinforcement Learning\n- **Authors**: Carlos Purves , Pietro Lio'\n- **URL**: [https://arxiv.org/abs/2603.12543](https://arxiv.org/abs/2603.12543)\n- **Abstract**:\n> Distributed reinforcement learning policies face network delays, jitter, and packet loss when deployed across edge devices and cloud servers. Standard RL training assumes zero-latency interaction, causing severe performance degradation under realistic network conditions. We introduce CALF (Communication-Aware Learning Framework), which trains policies under realistic network models during simulation. Systematic experiments demonstrate that network-aware training substantially reduces deployment performance gaps compared to network-agnostic baselines. Distributed policy deployments across heterogeneous hardware validate that explicitly modelling communication constraints during training enables robust real-world execution. These findings establish network conditions as a major axis of sim-to-real transfer for Wi-Fi-like distributed deployments, complementing physics and visual domain randomisation.\n\n---\n\n## 100. Embedded Quantum Machine Learning in Embedded Systems: Feasibility, Hybrid Architectures, and Quantum Co-Processors\n- **Authors**: Somdip Dey , Syed Muhammad Raza\n- **URL**: [https://arxiv.org/abs/2603.12540](https://arxiv.org/abs/2603.12540)\n- **Abstract**:\n> Embedded quantum machine learning (EQML) seeks to bring quantum machine learning (QML) capabilities to resource-constrained edge platforms such as IoT nodes, wearables, drones, and cyber-physical controllers. In 2026, EQML is technically feasible only in limited and highly experimental forms: (i) hybrid workflows where an embedded device performs sensing and classical processing while offloading a narrowly scoped quantum subroutine to a remote quantum processing unit (QPU) or nearby quantum appliance, and (ii) early-stage \"embedded QPU\" concepts in which a compact quantum co-processor is integrated with classical control hardware. A practical bridge is quantum-inspired machine learning and optimisation on classical embedded processors and FPGAs. This paper analyses feasibility from a circuits-and-systems perspective aligned with the academic community, formalises two implementation pathways, identifies the dominant barriers (latency, data encoding overhead, NISQ noise, tooling mismatch, and energy), and maps them to concrete engineering directions in interface design, control electronics, power management, verification, and security. We also argue that responsible deployment requires adversarial evaluation and governance practices that are increasingly necessary for edge AI systems.\n\n---\n\n## 101. Spatio-Semantic Expert Routing Architecture with Mixture-of-Experts for Referring Image Segmentation\n- **Authors**: Alaa Dalaq , Muzammil Behzad\n- **URL**: [https://arxiv.org/abs/2603.12538](https://arxiv.org/abs/2603.12538)\n- **Abstract**:\n> Referring image segmentation aims to produce a pixel-level mask for the image region described by a natural-language expression. Although pretrained vision-language models have improved semantic grounding, many existing methods still rely on uniform refinement strategies that do not fully match the diverse reasoning requirements of referring expressions. Because of this mismatch, predictions often contain fragmented regions, inaccurate boundaries, or even the wrong object, especially when pretrained backbones are frozen for computational efficiency. To address these limitations, we propose SERA, a Spatio-Semantic Expert Routing Architecture for referring image segmentation. SERA introduces lightweight, expression-aware expert refinement at two complementary stages within a vision-language framework. First, we design SERA-Adapter, which inserts an expression-conditioned adapter into selected backbone blocks to improve spatial coherence and boundary precision through expert-guided refinement and cross-modal attention. We then introduce SERA-Fusion, which strengthens intermediate visual representations by reshaping token features into spatial grids and applying geometry-preserving expert transformations before multimodal interaction. In addition, a lightweight routing mechanism adaptively weights expert contributions while remaining compatible with pretrained representations. To make this routing stable under frozen encoders, SERA uses a parameter-efficient tuning strategy that updates only normalization and bias terms, affecting less than 1% of the backbone parameters. Experiments on standard referring image segmentation benchmarks show that SERA consistently outperforms strong baselines, with especially clear gains on expressions that require accurate spatial localization and precise boundary delineation.\n\n---\n\n## 102. TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning\n- **Authors**: Alliot Nagle , Jakhongir Saydaliev , Dhia Garbaya , Michael Gastpar , Ashok Vardhan Makkuva , Hyeji Kim\n- **URL**: [https://arxiv.org/abs/2603.12529](https://arxiv.org/abs/2603.12529)\n- **Abstract**:\n> Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design TERMINATOR, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning TERMINATOR is that the first arrival of an LRM's final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train TERMINATOR. Powered by this approach, TERMINATOR achieves significant reductions in CoT lengths of 14%-55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, whilst outperforming current state-of-the-art methods.\n\n---\n\n## 103. LLM BiasScope: A Real-Time Bias Analysis Platform for Comparative LLM Evaluation\n- **Authors**: Himel Ghosh , Nick Elias Werner\n- **URL**: [https://arxiv.org/abs/2603.12522](https://arxiv.org/abs/2603.12522)\n- **Abstract**:\n> As large language models (LLMs) are deployed widely, detecting and understanding bias in their outputs is critical. We present LLM BiasScope, a web application for side-by-side comparison of LLM outputs with real-time bias analysis. The system supports multiple providers (Google Gemini, DeepSeek, MiniMax, Mistral, Meituan, Meta Llama) and enables researchers and practitioners to compare models on the same prompts while analyzing bias patterns. LLM BiasScope uses a two-stage bias detection pipeline: sentence-level bias detection followed by bias type classification for biased sentences. The analysis runs automatically on both user prompts and model responses, providing statistics, visualizations, and detailed breakdowns of bias types. The interface displays two models side-by-side with synchronized streaming responses, per-model bias summaries, and a comparison view highlighting differences in bias distributions. The system is built on this http URL with React, integrates Hugging Face inference endpoints for bias detection, and uses the Vercel AI SDK for multi-provider LLM access. Features include real-time streaming, export to JSON/PDF, and interactive visualizations (bar charts, radar charts) for bias analysis. LLM BiasScope is available as an open-source web application, providing a practical tool for bias evaluation and comparative analysis of LLM behaviour.\n\n---\n\n## 104. When LLM Judge Scores Look Good but Best-of-N Decisions Fail\n- **Authors**: Eddie Landesberg\n- **URL**: [https://arxiv.org/abs/2603.12520](https://arxiv.org/abs/2603.12520)\n- **Abstract**:\n> Large language models are often used as judges to score candidate responses, then validated with a single global metric such as correlation with reference labels. This can be misleading when the real deployment task is best-of-n selection within a prompt. In a 5,000-prompt best-of-4 benchmark from Chatbot Arena, a judge with moderate global correlation (r = 0.47) captures only 21.0% of the improvement that perfect selection would achieve over random choice. The gap arises because global agreement is driven largely by prompt-level baseline effects, while selection depends on within-prompt ranking: within-prompt correlation is only r_within = 0.27, and coarse pointwise scoring creates ties in 67% of pairwise comparisons. In a matched-pair best-of-2 audit, explicit pairwise judging recovers much of this lost signal, raising recovery from 21.1% to 61.2%. For judge-based selection, the relevant audit should report within-prompt signal, tie rates, and recovery/top-1 accuracy, not global agreement alone.\n\n---\n\n## 105. Red-Teaming Vision-Language-Action Models via Quality Diversity Prompt Generation for Robust Robot Policies\n- **Authors**: Siddharth Srikanth , Freddie Liang , Sophie Hsu , Varun Bhatt , Shihan Zhao , Henry Chen , Bryon Tjanaka , Minjune Hwang , Akanksha Saran , Daniel Seita , Aaquib Tabrez , Stefanos Nikolaidis\n- **URL**: [https://arxiv.org/abs/2603.12510](https://arxiv.org/abs/2603.12510)\n- **Abstract**:\n> Vision-Language-Action (VLA) models have significant potential to enable general-purpose robotic systems for a range of vision-language tasks. However, the performance of VLA-based robots is highly sensitive to the precise wording of language instructions, and it remains difficult to predict when such robots will fail. To improve the robustness of VLAs to different wordings, we present Q-DIG (Quality Diversity for Diverse Instruction Generation), which performs red-teaming by scalably identifying diverse natural language task descriptions that induce failures while remaining task-relevant. Q-DIG integrates Quality Diversity (QD) techniques with Vision-Language Models (VLMs) to generate a broad spectrum of adversarial instructions that expose meaningful vulnerabilities in VLA behavior. Our results across multiple simulation benchmarks show that Q-DIG finds more diverse and meaningful failure modes compared to baseline methods, and that fine-tuning VLAs on the generated instructions improves task success rates. Furthermore, results from a user study highlight that Q-DIG generates prompts judged to be more natural and human-like than those from baselines. Finally, real-world evaluations of Q-DIG prompts show results consistent with simulation, and fine-tuning VLAs on the generated prompts further success rates on unseen instructions. Together, these findings suggest that Q-DIG is a promising approach for identifying vulnerabilities and improving the robustness of VLA-based robots. Our anonymous project website is at this http URL .\n\n---\n\n## 106. ELLA: Generative AI-Powered Social Robots for Early Language Development at Home\n- **Authors**: Victor Nikhil Antony , Shiye Cao , Shuning Wang , Chien-Ming Huang\n- **URL**: [https://arxiv.org/abs/2603.12508](https://arxiv.org/abs/2603.12508)\n- **Abstract**:\n> Early language development shapes children's later literacy and learning, yet many families have limited access to scalable, high-quality support at home. Recent advances in generative AI make it possible for social robots to move beyond scripted interactions and engage children in adaptive, conversational activities, but it remains unclear how to design such systems for pre-schoolers and how children engage with them over time in the home. We present ELLA (Early Language Learning Agent), an autonomous, generative AI-powered social robot that supports early language development through interactive storytelling, parent-selected language targets, and scaffolded dialogue. Using a multi-phased, human-centered process, we interviewed parents (n=7) and educators (n=5) and iteratively refined ELLA through twelve in-home design workshops. We then deployed ELLA with ten children for eight days. We report design insights from in-home workshops, characterize children's engagement and behaviors during deployment, and distill design implications for generative AI-powered social robots supporting early language learning at home.\n\n---\n\n## 107. Naïve PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation\n- **Authors**: Joong Ho Kim , Nicholas Thai , Souhardya Saha Dip , Dong Lao , Keith G. Mills\n- **URL**: [https://arxiv.org/abs/2603.12506](https://arxiv.org/abs/2603.12506)\n- **Abstract**:\n> Text-to-Image (T2I) generation is primarily driven by Diffusion Models (DM) which rely on random Gaussian noise. Thus, like playing the slots at a casino, a DM will produce different results given the same user-defined inputs. This imposes a gambler's burden: To perform multiple generation cycles to obtain a satisfactory result. However, even though DMs use stochastic sampling to seed generation, the distribution of generated content quality highly depends on the prompt and the generative ability of a DM with respect to it. To account for this, we propose Naïve PAINE for improving the generative quality of Diffusion Models by leveraging T2I preference benchmarks. We directly predict the numerical quality of an image from the initial noise and given prompt. Naïve PAINE then selects a handful of quality noises and forwards them to the DM for generation. Further, Naïve PAINE provides feedback on the DM generative quality given the prompt and is lightweight enough to seamlessly fit into existing DM pipelines. Experimental results demonstrate that Naïve PAINE outperforms existing approaches on several prompt corpus benchmarks.\n\n---\n\n## 108. TRACE: Temporal Rule-Anchored Chain-of-Evidence on Knowledge Graphs for Interpretable Stock Movement Prediction\n- **Authors**: Qianggang Ding , Haochen Shi , Luis Castejón Lozano , Miguel Conner , Juan Abia , Luis Gallego-Ledesma , Joshua Fellowes , Gerard Conangla Planes , Adam Elwood , Bang Liu\n- **URL**: [https://arxiv.org/abs/2603.12500](https://arxiv.org/abs/2603.12500)\n- **Abstract**:\n> We present a Temporal Rule-Anchored Chain-of-Evidence (TRACE) on knowledge graphs for interpretable stock movement prediction that unifies symbolic relational priors, dynamic graph exploration, and LLM-guided decision making in a single end-to-end pipeline. The approach performs rule-guided multi-hop exploration restricted to admissible relation sequences, grounds candidate reasoning chains in contemporaneous news, and aggregates fully grounded evidence into auditable \\texttt{UP}/\\texttt{DOWN} verdicts with human-readable paths connecting text and structure. On an S\\&P~500 benchmark, the method achieves 55.1\\% accuracy, 55.7\\% precision, 71.5\\% recall, and 60.8\\% F1, surpassing strong baselines and improving recall and F1 over the best graph baseline under identical evaluation. The gains stem from (i) rule-guided exploration that focuses search on economically meaningful motifs rather than arbitrary walks, and (ii) text-grounded consolidation that selectively aggregates high-confidence, fully grounded hypotheses instead of uniformly pooling weak signals. Together, these choices yield higher sensitivity without sacrificing selectivity, delivering predictive lift with faithful, auditably interpretable explanations.\n\n---\n\n## 109. One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies\n- **Authors**: Shaolong Li , Lichao Sun , Yongchao Chen\n- **URL**: [https://arxiv.org/abs/2603.12480](https://arxiv.org/abs/2603.12480)\n- **Abstract**:\n> Generative flow and diffusion models provide the continuous, multimodal action distributions needed for high-precision robotic policies. However, their reliance on iterative sampling introduces severe inference latency, degrading control frequency and harming performance in time-sensitive manipulation. To address this problem, we propose the One-Step Flow Policy (OFP), a from-scratch self-distillation framework for high-fidelity, single-step action generation without a pre-trained teacher. OFP unifies a self-consistency loss to enforce coherent transport across time intervals, and a self-guided regularization to sharpen predictions toward high-density expert modes. In addition, a warm-start mechanism leverages temporal action correlations to minimize the generative transport distance. Evaluations across 56 diverse simulated manipulation tasks demonstrate that a one-step OFP achieves state-of-the-art results, outperforming 100-step diffusion and flow policies while accelerating action generation by over $100\\times$. We further integrate OFP into the $\\pi_{0.5}$ model on RoboTwin 2.0, where one-step OFP surpasses the original 10-step policy. These results establish OFP as a practical, scalable solution for highly accurate and low-latency robot control.\n\n---\n\n## 110. The Perfection Paradox: From Architect to Curator in AI-Assisted API Design\n- **Authors**: Mak Ahmad , Andrew Macvean , JJ Geewax , David Karger\n- **URL**: [https://arxiv.org/abs/2603.12475](https://arxiv.org/abs/2603.12475)\n- **Abstract**:\n> Enterprise API design is often bottlenecked by the tension between rapid feature delivery and the rigorous maintenance of usability standards. We present an industrial case study evaluating an AI-assisted design workflow trained on API Improvement Proposals (AIPs). Through a controlled study with 16 industry experts, we compared AI-generated API specifications against human-authored ones. While quantitative results indicated AI superiority in 10 of 11 usability dimensions and an 87% reduction in authoring time, qualitative analysis revealed a paradox: experts frequently misidentified AI work as human (19% accuracy) yet described the designs as unsettlingly \"perfect.\" We characterize this as a \"Perfection Paradox\" -- where hyper-consistency signals a lack of pragmatic human judgment. We discuss the implications of this perfection paradox, proposing a shift in the human designer's role from the \"drafter\" of specifications to the \"curator\" of AI-generated patterns.\n\n---\n\n## 111. CLARE: Classification-based Regression for Electron Temperature Prediction\n- **Authors**: Michael Liang , Blake DeHaas , Naomi Maruyama , Xiangning Chu , Takumi Abe , Koh-Ichiro Oyama\n- **URL**: [https://arxiv.org/abs/2603.12470](https://arxiv.org/abs/2603.12470)\n- **Abstract**:\n> Electron temperature (Te) is an important parameter governing space weather in the upper atmosphere, but has historically been underexplored in the space weather machine learning literature. We present CLARE, a machine learning model for predicting electron temperature in the Earth's plasmasphere trained on AKEBONO (EXOS-D) satellite measurements as well as solar and geomagnetic indices. CLARE uses a classification-based regression architecture that transforms the continuous Te output space into 150 discrete classification intervals. Training the model on a classification task improves prediction accuracy by 6.46% relative compared to a traditional regression model while also outputting uncertainty estimation information on its predictions. On a held out test set from the AKEBONO data, the model's Te predictions achieve 69.67% accuracy within 10% of the ground truth and 46.17% on a known geomagnetic storm period from January 30th to February 7th, 1991. We show that machine learning can be used to produce high-accuracy Te models on publicly available data.\n\n---\n\n## 112. Shattering the Shortcut: A Topology-Regularized Benchmark for Multi-hop Medical Reasoning in LLMs\n- **Authors**: Xing Zi , Xinying Zhou , Jinghao Xiao , Catarina Moreira , Mukesh Prasad\n- **URL**: [https://arxiv.org/abs/2603.12458](https://arxiv.org/abs/2603.12458)\n- **Abstract**:\n> While Large Language Models (LLMs) achieve expert-level performance on standard medical benchmarks through single-hop factual recall, they severely struggle with the complex, multi-hop diagnostic reasoning required in real-world clinical settings. A primary obstacle is \"shortcut learning\", where models exploit highly connected, generic hub nodes (e.g., \"inflammation\") in knowledge graphs to bypass authentic micro-pathological cascades. To address this, we introduce ShatterMed-QA, a bilingual benchmark of 10,558 multi-hop clinical questions designed to rigorously evaluate deep diagnostic reasoning. Our framework constructs a topology-regularized medical Knowledge Graph using a novel $k$-Shattering algorithm, which physically prunes generic hubs to explicitly sever logical shortcuts. We synthesize the evaluation vignettes by applying implicit bridge entity masking and topology-driven hard negative sampling, forcing models to navigate biologically plausible distractors without relying on superficial elimination. Comprehensive evaluations of 21 LLMs reveal massive performance degradation on our multi-hop tasks, particularly among domain-specific models. Crucially, restoring the masked evidence via Retrieval-Augmented Generation (RAG) triggers near-universal performance recovery, validating ShatterMed-QA's structural fidelity and proving its efficacy in diagnosing the fundamental reasoning deficits of current medical AI. Explore the dataset, interactive examples, and full leaderboards at our project website: this https URL\n\n---\n\n## 113. Operationalising Cyber Risk Management Using AI: Connecting Cyber Incidents to MITRE ATT&CK Techniques, Security Controls, and Metrics\n- **Authors**: Emad Sherif , Iryna Yevseyeva , Vitor Basto-Fernandes , Allan Cook\n- **URL**: [https://arxiv.org/abs/2603.12455](https://arxiv.org/abs/2603.12455)\n- **Abstract**:\n> The escalating frequency of cyber-attacks poses significant challenges for organisations, particularly small enterprises constrained by limited in-house expertise, insufficient knowledge, and financial resources. This research presents a novel framework that leverages Natural Language Processing to address these challenges through automated mapping of cyber incidents to adversary techniques. We introduce the Cyber Catalog, a knowledge base that systematically integrates CIS Critical Security Controls, MITRE ATT&CK techniques, and SMART metrics. This integrated resource enables organisations to connect threat intelligence directly to actionable controls and measurable outcomes. To operationalise the framework, we fine-tuned all-mpnet-base-v2, a highly regarded sentence-transformers model used to convert text into numerical vectors on an augmented dataset comprising 74,986 incident-technique pairs to enhance semantic similarity between cyber incidents and MITRE ATT&CK techniques. Our fine-tuned model achieved a Spearman correlation of 0.7894 and Pearson correlation of 0.8756, demonstrating substantial improvements over top baseline models including all-mpnet-base-v2, all-distilroberta-v1, and all-MiniLM-L12-v2. Furthermore, our model exhibited significantly lower prediction errors (MAE = 0.135, MSE = 0.027) compared to all baseline models, confirming superior accuracy and consistency. The Cyber Catalog, training dataset, trained model, and implementation code made publicly available to facilitate further research and enable practical deployment in resource-constrained environments. This work bridges the gap between threat intelligence and operational security management, providing an actionable tool for systematic cyber incident response and evidence-based cyber risk management.\n\n---\n\n## 114. Unmasking Biases and Reliability Concerns in Convolutional Neural Networks Analysis of Cancer Pathology Images\n- **Authors**: Michael Okonoda , Eder Martinez , Abhilekha Dalal , Lior Shamir\n- **URL**: [https://arxiv.org/abs/2603.12445](https://arxiv.org/abs/2603.12445)\n- **Abstract**:\n> Convolutional Neural Networks have shown promising effectiveness in identifying different types of cancer from radiographs. However, the opaque nature of CNNs makes it difficult to fully understand the way they operate, limiting their assessment to empirical evaluation. Here we study the soundness of the standard practices by which CNNs are evaluated for the purpose of cancer pathology. Thirteen highly used cancer benchmark datasets were analyzed, using four common CNN architectures and different types of cancer, such as melanoma, carcinoma, colorectal cancer, and lung cancer. We compared the accuracy of each model with that of datasets made of cropped segments from the background of the original images that do not contain clinically relevant content. Because the rendered datasets contain no clinical information, the null hypothesis is that the CNNs should provide mere chance-based accuracy when classifying these datasets. The results show that the CNN models provided high accuracy when using the cropped segments, sometimes as high as 93\\%, even though they lacked biomedical information. These results show that some CNN architectures are more sensitive to bias than others. The analysis shows that the common practices of machine learning evaluation might lead to unreliable results when applied to cancer pathology. These biases are very difficult to identify, and might mislead researchers as they use available benchmark datasets to test the efficacy of CNN methods.\n\n---\n\n## 115. Revisiting Model Stitching In the Foundation Model Era\n- **Authors**: Zheda Mai , Ke Zhang , Fu-En Wang , Zixiao Ken Wang , Albert Y. C. Chen , Lu Xia , Min Sun , Wei-Lun Chao , Cheng-Hao Kuo\n- **URL**: [https://arxiv.org/abs/2603.12433](https://arxiv.org/abs/2603.12433)\n- **Abstract**:\n> Model stitching, connecting early layers of one model (source) to later layers of another (target) via a light stitch layer, has served as a probe of representational compatibility. Prior work finds that models trained on the same dataset remain stitchable (negligible accuracy drop) despite different initializations or objectives. We revisit stitching for Vision Foundation Models (VFMs) that vary in objectives, data, and modality mix (e.g., CLIP, DINOv2, SigLIP 2) and ask: Are heterogeneous VFMs stitchable? We introduce a systematic protocol spanning the stitch points, stitch layer families, training losses, and downstream tasks. Three findings emerge. (1) Stitch layer training matters: conventional approaches that match the intermediate features at the stitch point or optimize the task loss end-to-end struggle to retain accuracy, especially at shallow stitch points. (2) With a simple feature-matching loss at the target model's penultimate layer, heterogeneous VFMs become reliably stitchable across vision tasks. (3) For deep stitch points, the stitched model can surpass either constituent model at only a small inference overhead (for the stitch layer). Building on these findings, we further propose the VFM Stitch Tree (VST), which shares early layers across VFMs while retaining their later layers, yielding a controllable accuracy-latency trade-off for multimodal LLMs that often leverage multiple VFMs. Taken together, our study elevates stitching from a diagnostic probe to a practical recipe for integrating complementary VFM strengths and pinpointing where their representations align or diverge.\n\n---\n\n## 116. Test-Time Strategies for More Efficient and Accurate Agentic RAG\n- **Authors**: Brian Zhang , Deepti Guntur , Zhiyang Zuo , Abhinav Sharma , Shreyas Chaudhari , Wenlong Zhao , Franck Dernoncourt , Puneet Mathur , Ryan Rossi , Nedim Lipka\n- **URL**: [https://arxiv.org/abs/2603.12396](https://arxiv.org/abs/2603.12396)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) systems face challenges with complex, multihop questions, and agentic frameworks such as Search-R1 (Jin et al., 2025), which operates iteratively, have been proposed to address these complexities. However, such approaches can introduce inefficiencies, including repetitive retrieval of previously processed information and challenges in contextualizing retrieved results effectively within the current generation prompt. Such issues can lead to unnecessary retrieval turns, suboptimal reasoning, inaccurate answers, and increased token consumption. In this paper, we investigate test-time modifications to the Search-R1 pipeline to mitigate these identified shortcomings. Specifically, we explore the integration of two components and their combination: a contextualization module to better integrate relevant information from retrieved documents into reasoning, and a de-duplication module that replaces previously retrieved documents with the next most relevant ones. We evaluate our approaches using the HotpotQA (Yang et al., 2018) and the Natural Questions (Kwiatkowski et al., 2019) datasets, reporting the exact match (EM) score, an LLM-as-a-Judge assessment of answer correctness, and the average number of turns. Our best-performing variant, utilizing GPT-4.1-mini for contextualization, achieves a 5.6% increase in EM score and reduces the number of turns by 10.5% compared to the Search-R1 baseline, demonstrating improved answer accuracy and retrieval efficiency.\n\n---\n\n## 117. SPARROW: Learning Spatial Precision and Temporal Referential Consistency in Pixel-Grounded Video MLLMs\n- **Authors**: Mohamad Alansari , Naufal Suryanto , Divya Velayudhan , Sajid Javed , Naoufel Werghi , Muzammal Naseer\n- **URL**: [https://arxiv.org/abs/2603.12382](https://arxiv.org/abs/2603.12382)\n- **Abstract**:\n> Multimodal large language models (MLLMs) have advanced from image-level reasoning to pixel-level grounding, but extending these capabilities to videos remains challenging as models must achieve spatial precision and temporally consistent reference tracking. Existing video MLLMs often rely on a static segmentation token ([SEG]) for frame-wise grounding, which provides semantics but lacks temporal context, causing spatial drift, identity switches, and unstable initialization when objects move or reappear. We introduce SPARROW, a pixel-grounded video MLLM that unifies spatial accuracy and temporal stability through two key components: (i) Target-Specific Tracked Features (TSF), which inject temporally aligned referent cues during training, and (ii) a dual-prompt design that decodes box ([BOX]) and segmentation ([SEG]) tokens to fuse geometric priors with semantic grounding. SPARROW is supported by a curated referential video dataset of 30,646 videos and 45,231 Q&A pairs and operates end-to-end without external detectors via a class-agnostic SAM2-based proposer. Integrated into three recent open-source video MLLMs (UniPixel, GLUS, and VideoGLaMM), SPARROW delivers consistent gains across six benchmarks, improving up to +8.9 J&F on RVOS, +5 mIoU on visual grounding, and +5.4 CLAIR on GCG. These results demonstrate that SPARROW substantially improves referential stability, spatial precision, and temporal coherence in pixel-grounded video understanding. Project page: this https URL\n\n---\n\n## 118. Budget-Sensitive Discovery Scoring: A Formally Verified Framework for Evaluating AI-Guided Scientific Selection\n- **Authors**: Abhinaba Basu , Pavan Chakraborty\n- **URL**: [https://arxiv.org/abs/2603.12349](https://arxiv.org/abs/2603.12349)\n- **Abstract**:\n> Scientific discovery increasingly relies on AI systems to select candidates for expensive experimental validation, yet no principled, budget-aware evaluation framework exists for comparing selection strategies -- a gap intensified by large language models (LLMs), which generate plausible scientific proposals without reliable downstream evaluation. We introduce the Budget-Sensitive Discovery Score (BSDS), a formally verified metric -- 20 theorems machine-checked by the Lean 4 proof assistant -- that jointly penalizes false discoveries (lambda-weighted FDR) and excessive abstention (gamma-weighted coverage gap) at each budget level. Its budget-averaged form, the Discovery Quality Score (DQS), provides a single summary statistic that no proposer can inflate by performing well at a cherry-picked budget. As a case study, we apply BSDS/DQS to: do LLMs add marginal value to an existing ML pipeline for drug discovery candidate selection? We evaluate 39 proposers -- 11 mechanistic variants, 14 zero-shot LLM configurations, and 14 few-shot LLM configurations -- using SMILES representations on MoleculeNet HIV (41,127 compounds, 3.5% active, 1,000 bootstrap replicates) under both random and scaffold splits. Three findings emerge. First, the simple RF-based Greedy-ML proposer achieves the best DQS (-0.046), outperforming all MLP variants and LLM configurations. Second, no LLM surpasses the Greedy-ML baseline under zero-shot or few-shot evaluation on HIV or Tox21, establishing that LLMs provide no marginal value over an existing trained classifier. Third, the proposer hierarchy generalizes across five MoleculeNet benchmarks spanning 0.18%-46.2% prevalence, a non-drug AV safety domain, and a 9x7 grid of penalty parameters (tau >= 0.636, mean tau = 0.863). The framework applies to any setting where candidates are selected under budget constraints and asymmetric error costs.\n\n---\n\n## 119. Optimizing Task Completion Time Updates Using POMDPs\n- **Authors**: Duncan Eddy , Esen Yel , Emma Passmore , Niles Egan , Grayson Armour , Dylan M. Asmar , Mykel J. Kochenderfer\n- **URL**: [https://arxiv.org/abs/2603.12340](https://arxiv.org/abs/2603.12340)\n- **Abstract**:\n> Managing announced task completion times is a fundamental control problem in project management. While extensive research exists on estimating task durations and task scheduling, the problem of when and how to update completion times communicated to stakeholders remains understudied. Organizations must balance announcement accuracy against the costs of frequent timeline updates, which can erode stakeholder trust and trigger costly replanning. Despite the prevalence of this problem, current approaches rely on static predictions or ad-hoc policies that fail to account for the sequential nature of announcement management. In this paper, we formulate the task announcement problem as a Partially Observable Markov Decision Process (POMDP) where the control policy must decide when to update announced completion times based on noisy observations of true task completion. Since most state variables (current time and previous announcements) are fully observable, we leverage the Mixed Observability MDP (MOMDP) framework to enable more efficient policy optimization. Our reward structure captures the dual costs of announcement errors and update frequency, enabling synthesis of optimal announcement control policies. Using off-the-shelf solvers, we generate policies that act as feedback controllers, adaptively managing announcements based on belief state evolution. Simulation results demonstrate significant improvements in both accuracy and announcement stability compared to baseline strategies, achieving up to 75\\% reduction in unnecessary updates while maintaining or improving prediction accuracy.\n\n---\n\n## 120. Maximum Entropy Exploration Without the Rollouts\n- **Authors**: Jacob Adamczyk , Adam Kamoski , Rahul V. Kulkarni\n- **URL**: [https://arxiv.org/abs/2603.12325](https://arxiv.org/abs/2603.12325)\n- **Abstract**:\n> Efficient exploration remains a central challenge in reinforcement learning, serving as a useful pretraining objective for data collection, particularly when an external reward function is unavailable. A principled formulation of the exploration problem is to find policies that maximize the entropy of their induced steady-state visitation distribution, thereby encouraging uniform long-run coverage of the state space. Many existing exploration approaches require estimating state visitation frequencies through repeated on-policy rollouts, which can be computationally expensive. In this work, we instead consider an intrinsic average-reward formulation in which the reward is derived from the visitation distribution itself, so that the optimal policy maximizes steady-state entropy. An entropy-regularized version of this objective admits a spectral characterization: the relevant stationary distributions can be computed from the dominant eigenvectors of a problem-dependent transition matrix. This insight leads to a novel algorithm for solving the maximum entropy exploration problem, EVE (EigenVector-based Exploration), which avoids explicit rollouts and distribution estimation, instead computing the solution through iterative updates, similar to a value-based approach. To address the original unregularized objective, we employ a posterior-policy iteration (PPI) approach, which monotonically improves the entropy and converges in value. We prove convergence of EVE under standard assumptions and demonstrate empirically that it efficiently produces policies with high steady-state entropy, achieving competitive exploration performance relative to rollout-based baselines in deterministic grid-world environments.\n\n---\n\n## 121. Thermodynamics of Reinforcement Learning Curricula\n- **Authors**: Jacob Adamczyk , Juan Sebastian Rojas , Rahul V. Kulkarni\n- **URL**: [https://arxiv.org/abs/2603.12324](https://arxiv.org/abs/2603.12324)\n- **Abstract**:\n> Connections between statistical mechanics and machine learning have repeatedly proven fruitful, providing insight into optimization, generalization, and representation learning. In this work, we follow this tradition by leveraging results from non-equilibrium thermodynamics to formalize curriculum learning in reinforcement learning (RL). In particular, we propose a geometric framework for RL by interpreting reward parameters as coordinates on a task manifold. We show that, by minimizing the excess thermodynamic work, optimal curricula correspond to geodesics in this task space. As an application of this framework, we provide an algorithm, \"MEW\" (Minimum Excess Work), to derive a principled schedule for temperature annealing in maximum-entropy RL.\n\n---\n\n## 122. VQQA: An Agentic Approach for Video Evaluation and Quality Improvement\n- **Authors**: Yiwen Song , Tomas Pfister , Yale Song\n- **URL**: [https://arxiv.org/abs/2603.12310](https://arxiv.org/abs/2603.12310)\n- **Abstract**:\n> Despite rapid advancements in video generation models, aligning their outputs with complex user intent remains challenging. Existing test-time optimization methods are typically either computationally expensive or require white-box access to model internals. To address this, we present VQQA (Video Quality Question Answering), a unified, multi-agent framework generalizable across diverse input modalities and video generation tasks. By dynamically generating visual questions and using the resulting Vision-Language Model (VLM) critiques as semantic gradients, VQQA replaces traditional, passive evaluation metrics with human-interpretable, actionable feedback. This enables a highly efficient, closed-loop prompt optimization process via a black-box natural language interface. Extensive experiments demonstrate that VQQA effectively isolates and resolves visual artifacts, substantially improving generation quality in just a few refinement steps. Applicable to both text-to-video (T2V) and image-to-video (I2V) tasks, our method achieves absolute improvements of +11.57% on T2V-CompBench and +8.43% on VBench2 over vanilla generation, significantly outperforming state-of-the-art stochastic search and prompt optimization techniques.\n\n---\n\n## 123. HCP-DCNet: A Hierarchical Causal Primitive Dynamic Composition Network for Self-Improving Causal Understanding\n- **Authors**: Ming Lei , Shufan Wu , Christophe Baehr\n- **URL**: [https://arxiv.org/abs/2603.12305](https://arxiv.org/abs/2603.12305)\n- **Abstract**:\n> The ability to understand and reason about cause and effect -- encompassing interventions, counterfactuals, and underlying mechanisms -- is a cornerstone of robust artificial intelligence. While deep learning excels at pattern recognition, it fundamentally lacks a model of causality, making systems brittle under distribution shifts and unable to answer ``what-if'' questions. This paper introduces the \\emph{Hierarchical Causal Primitive Dynamic Composition Network (HCP-DCNet)}, a unified framework that bridges continuous physical dynamics with discrete symbolic causal inference. Departing from monolithic representations, HCP-DCNet decomposes causal scenes into reusable, typed \\emph{causal primitives} organized into four abstraction layers: physical, functional, event, and rule. A dual-channel routing network dynamically composes these primitives into task-specific, fully differentiable \\emph{Causal Execution Graphs (CEGs)}. Crucially, the system employs a \\emph{causal-intervention-driven meta-evolution} strategy, enabling autonomous self-improvement through a constrained Markov decision process. We establish rigorous theoretical guarantees, including type-safe composition, routing convergence, and universal approximation of causal dynamics. Extensive experiments across simulated physical and social environments demonstrate that HCP-DCNet significantly outperforms state-of-the-art baselines in causal discovery, counterfactual reasoning, and compositional generalization. This work provides a principled, scalable, and interpretable architecture for building AI systems with human-like causal abstraction and continual self-refinement capabilities.\n\n---\n\n## 124. A Geometrically-Grounded Drive for MDL-Based Optimization in Deep Learning\n- **Authors**: Ming Lei , Shufan Wu , Christophe Baehr\n- **URL**: [https://arxiv.org/abs/2603.12304](https://arxiv.org/abs/2603.12304)\n- **Abstract**:\n> This paper introduces a novel optimization framework that fundamentally integrates the Minimum Description Length (MDL) principle into the training dynamics of deep neural networks. Moving beyond its conventional role as a model selection criterion, we reformulate MDL as an active, adaptive driving force within the optimization process itself. The core of our method is a geometrically-grounded cognitive manifold whose evolution is governed by a \\textit{coupled Ricci flow}, enriched with a novel \\textit{MDL Drive} term derived from first principles. This drive, modulated by the task-loss gradient, creates a seamless harmony between data fidelity and model simplification, actively compressing the internal representation during training. We establish a comprehensive theoretical foundation, proving key properties including the monotonic decrease of description length (Theorem~\\ref{thm:convergence}), a finite number of topological phase transitions via a geometric surgery protocol (Theorems~\\ref{thm:surgery}, \\ref{thm:ultimate_fate}), and the emergence of universal critical behavior (Theorem~\\ref{thm:universality}). Furthermore, we provide a practical, computationally efficient algorithm with $O(N \\log N)$ per-iteration complexity (Theorem~\\ref{thm:complexity}), alongside guarantees for numerical stability (Theorem~\\ref{thm:stability}) and exponential convergence under convexity assumptions (Theorem~\\ref{thm:convergence_rate}). Empirical validation on synthetic regression and classification tasks confirms the theoretical predictions, demonstrating the algorithm's efficacy in achieving robust generalization and autonomous model simplification. This work provides a principled path toward more autonomous, generalizable, and interpretable AI systems by unifying geometric deep learning with information-theoretic principles.\n\n---\n\n## 125. Global Evolutionary Steering: Refining Activation Steering Control via Cross-Layer Consistency\n- **Authors**: Xinyan Jiang , Wenjing Yu , Di Wang , Lijie Hu\n- **URL**: [https://arxiv.org/abs/2603.12298](https://arxiv.org/abs/2603.12298)\n- **Abstract**:\n> Activation engineering enables precise control over Large Language Models (LLMs) without the computational cost of fine-tuning. However, existing methods deriving vectors from static activation differences are susceptible to high-dimensional noise and layer-wise semantic drift, often capturing spurious correlations rather than the target intent. To address this, we propose Global Evolutionary Refined Steering (GER-steer), a training-free framework that grounded in the geometric stability of the network's representation evolution. GER-steer exploits this global signal to rectify raw steering vectors, effectively decoupling robust semantic intent from orthogonal artifacts. Extensive evaluations confirm that GER-steer consistently outperforms baselines, delivering superior efficacy and generalization without layer-specific tuning, establishing a universal solution for reliable model alignment.\n\n---\n\n## 126. Synthetic Data Generation for Brain-Computer Interfaces: Overview, Benchmarking, and Future Directions\n- **Authors**: Ziwei Wang , Zhentao He , Xingyi He , Hongbin Wang , Tianwang Jia , Jingwei Luo , Siyang Li , Xiaoqing Chen , Dongrui Wu\n- **URL**: [https://arxiv.org/abs/2603.12296](https://arxiv.org/abs/2603.12296)\n- **Abstract**:\n> Deep learning has achieved transformative performance across diverse domains, largely driven by the large-scale, high-quality training data. In contrast, the development of brain-computer interfaces (BCIs) is fundamentally constrained by the limited, heterogeneous, and privacy-sensitive neural recordings. Generating synthetic yet physiologically plausible brain signals has therefore emerged as a compelling way to mitigate data scarcity and enhance model capacity. This survey provides a comprehensive review of brain signal generation for BCIs, covering methodological taxonomies, benchmark experiments, evaluation metrics, and key applications. We systematically categorize existing generative algorithms into four types: knowledge-based, feature-based, model-based, and translation-based approaches. Furthermore, we benchmark existing brain signal generation approaches across four representative BCI paradigms to provide an objective performance comparison. Finally, we discuss the potentials and challenges of current generation approaches and prospect future research on accurate, data-efficient, and privacy-aware BCI systems. The benchmark codebase is publicized at this https URL .\n\n---\n\n## 127. Detecting Miscitation on the Scholarly Web through LLM-Augmented Text-Rich Graph Learning\n- **Authors**: Huidong Wu , Haojia Xiang , Jingtong Gao , Xiangyu Zhao , Dengsheng Wu , Jianping Li\n- **URL**: [https://arxiv.org/abs/2603.12290](https://arxiv.org/abs/2603.12290)\n- **Abstract**:\n> Scholarly web is a vast network of knowledge connected by citations. However, this system is increasingly compromised by miscitation, where references do not support or even contradict the claims they are cited for. Current miscitation detection methods, which primarily rely on semantic similarity or network anomalies, struggle to capture the nuanced relationship between a citation's context and its place in the wider network. While large language models (LLMs) offer powerful capabilities in semantic reasoning for this task, their deployment is hindered by hallucination risks and high computational costs. In this work, we introduce LLM-Augmented Graph Learning-based Miscitation Detector (LAGMiD), a novel framework that leverages LLMs for deep semantic reasoning over citation graphs and distills this knowledge into graph neural networks (GNNs) for efficient and scalable miscitation detection. Specifically, LAGMiD introduces an evidence-chain reasoning mechanism, which uses chain-of-thought prompting, to perform multi-hop citation tracing and assess semantic fidelity. To reduce LLM inference costs, we design a knowledge distillation method aligning GNN embeddings with intermediate LLM reasoning states. A collaborative learning strategy further routes complex cases to the LLM while optimizing the GNN for structure-based generalization. Experiments on three real-world benchmarks show that LAGMiD achieves state-of-the-art miscitation detection with significantly reduced inference cost.\n\n---\n\n## 128. From Garbage to Gold: A Data-Architectural Theory of Predictive Robustness\n- **Authors**: Terrence J. Lee-St. John , Jordan L. Lawson , Bartlomiej Piechowski-Jozwiak\n- **URL**: [https://arxiv.org/abs/2603.12288](https://arxiv.org/abs/2603.12288)\n- **Abstract**:\n> Tabular machine learning presents a paradox: modern models achieve state-of-the-art performance using high-dimensional (high-D), collinear, error-prone data, defying the \"Garbage In, Garbage Out\" mantra. To help resolve this, we synthesize principles from Information Theory, Latent Factor Models, and Psychometrics, clarifying that predictive robustness arises not solely from data cleanliness, but from the synergy between data architecture and model capacity. Partitioning predictor-space \"noise\" into \"Predictor Error\" and \"Structural Uncertainty\" (informational deficits from stochastic generative mappings), we prove that leveraging high-D sets of error-prone predictors asymptotically overcomes both types of noise, whereas cleaning a low-D set is fundamentally bounded by Structural Uncertainty. We demonstrate why \"Informative Collinearity\" (dependencies from shared latent causes) enhances reliability and convergence efficiency, and explain why increased dimensionality reduces the latent inference burden, enabling feasibility with finite samples. To address practical constraints, we propose \"Proactive Data-Centric AI\" to identify predictors that enable robustness efficiently. We also derive boundaries for Systematic Error Regimes and show why models that absorb \"rogue\" dependencies can mitigate assumption violations. Linking latent architecture to Benign Overfitting, we offer a first step towards a unified view of robustness to Outcome Error and predictor-space noise, while also delineating when traditional DCAI's focus on label cleaning remains powerful. By redefining data quality from item-level perfection to portfolio-level architecture, we provide a theoretical rationale for \"Local Factories\" -- learning from live, uncurated enterprise \"data swamps\" -- supporting a deployment paradigm shift from \"Model Transfer\" to \"Methodology Transfer'' to overcome static generalizability limitations.\n\n---\n\n## 129. The DIME Architecture: A Unified Operational Algorithm for Neural Representation, Dynamics, Control and Integration\n- **Authors**: Ionel Cristian Vladu , Nicu Bizdoaca , Ionica Pirici , Tudor-Adrian Balseanu , Eduard Nicusor Bondoc\n- **URL**: [https://arxiv.org/abs/2603.12286](https://arxiv.org/abs/2603.12286)\n- **Abstract**:\n> Modern neuroscience has accumulated extensive evidence on perception, memory, prediction, valuation, and consciousness, yet still lacks an explicit operational architecture capable of integrating these phenomena within a unified computational framework. Existing theories address specific aspects of neural function: predictive coding and active inference emphasize hierarchical inference and prediction error minimization; engram theories explain memory through distributed cell assemblies; neuromodulatory accounts focus on value-dependent regulation of plasticity and behaviour; and global workspace or large-scale network models investigate mechanisms underlying conscious access. Despite their explanatory power, these approaches remain only partially integrated at the architectural level. This work introduces DIME (Detect-Integrate-Mark-Execute), a neural architecture organizing perception, memory, valuation, and conscious access within a common operational cycle. The framework includes four interacting components: engrams, distributed recurrent neural structures supporting multiple activation trajectories; execution threads, spatiotemporal trajectories implementing neural processes; marker systems, neuromodulatory and limbic mechanisms regulating gain, plasticity, and trajectory selection; and hyperengrams, large-scale integrative states associated with operational conscious access. The framework is consistent with empirical evidence from hippocampal indexing, recurrent cortical processing, replay phenomena, large-scale network integration, and neuromodulatory regulation. Formulated at an abstract computational level, DIME may also inform artificial intelligence and robotics by providing an architectural template in which representation, valuation, and temporal sequencing emerge from a unified mechanism. An extended theoretical exposition is available in a companion monograph on Zenodo.\n\n---\n\n## 130. Predictive Analytics for Foot Ulcers Using Time-Series Temperature and Pressure Data\n- **Authors**: Md Tanvir Hasan Turja\n- **URL**: [https://arxiv.org/abs/2603.12278](https://arxiv.org/abs/2603.12278)\n- **Abstract**:\n> Diabetic foot ulcers (DFUs) are a severe complication of diabetes, often resulting in significant morbidity. This paper presents a predictive analytics framework utilizing time-series data captured by wearable foot sensors -- specifically NTC thin-film thermocouples for temperature measurement and FlexiForce pressure sensors for plantar load monitoring. Data was collected from healthy subjects walking on an instrumented pathway. Unsupervised machine learning algorithms, Isolation Forest and K-Nearest Neighbors (KNN), were applied to detect anomalies that may indicate early ulcer risk. Through rigorous data preprocessing and targeted feature engineering, physiologic patterns were extracted to identify subtle changes in foot temperature and pressure. Results demonstrate Isolation Forest is sensitive to micro-anomalies, while KNN is effective in flagging extreme deviations, albeit at a higher false-positive rate. Strong correlations between temperature and pressure readings support combined sensor monitoring for improved predictive accuracy. These findings provide a basis for real-time diabetic foot health surveillance, aiming to facilitate earlier intervention and reduce DFU incidence.\n\n---\n\n## 131. Prompt Injection as Role Confusion\n- **Authors**: Charles Ye , Jasmine Cui , Dylan Hadfield-Menell\n- **URL**: [https://arxiv.org/abs/2603.12277](https://arxiv.org/abs/2603.12277)\n- **Abstract**:\n> Language models remain vulnerable to prompt injection attacks despite extensive safety training. We trace this failure to role confusion: models infer roles from how text is written, not where it comes from. We design novel role probes to capture how models internally identify \"who is speaking.\" These reveal why prompt injection works: untrusted text that imitates a role inherits that role's authority. We test this insight by injecting spoofed reasoning into user prompts and tool outputs, achieving average success rates of 60% on StrongREJECT and 61% on agent exfiltration, across multiple open- and closed-weight models with near-zero baselines. Strikingly, the degree of internal role confusion strongly predicts attack success before generation begins. Our findings reveal a fundamental gap: security is defined at the interface but authority is assigned in latent space. More broadly, we introduce a unifying, mechanistic framework for prompt injection, demonstrating that diverse prompt-injection attacks exploit the same underlying role-confusion mechanism.\n\n---\n\n## 132. Aligning Language Models from User Interactions\n- **Authors**: Thomas Kleine Buening , Jonas Hübotter , Barna Pásztor , Idan Shenfeld , Giorgia Ramponi , Andreas Krause\n- **URL**: [https://arxiv.org/abs/2603.12273](https://arxiv.org/abs/2603.12273)\n- **Abstract**:\n> Multi-turn user interactions are among the most abundant data produced by language models, yet we lack effective methods to learn from them. While typically discarded, these interactions often contain useful information: follow-up user messages may indicate that a response was incorrect, failed to follow an instruction, or did not align with the user's preferences. Importantly, language models are already able to make use of this information in context. After observing a user's follow-up, the same model is often able to revise its behavior. We leverage this ability to propose a principled and scalable method for learning directly from user interactions through self-distillation. By conditioning the model on the user's follow-up message and comparing the resulting token distribution with the original policy, we obtain a target for updating the policy that captures how the model's behavior changes in hindsight. We then distill this hindsight distribution back into the current policy. Remarkably, we show that training on real-world user conversations from WildChat improves language models across standard alignment and instruction-following benchmarks, without regressing other capabilities. The same mechanism enables personalization, allowing models to continually adapt to individual users through interaction without explicit feedback. Our results demonstrate that raw user interactions that arise naturally during deployment enable alignment, personalization, and continual adaptation.\n\n---\n\n## 133. Diagnosing Retrieval Bias Under Multiple In-Context Knowledge Updates in Large Language Models\n- **Authors**: Boyu Qiao , Sean Guo , Xian Yang , Kun Li , Wei Zhou , Songlin Hu , Yunya Song\n- **URL**: [https://arxiv.org/abs/2603.12271](https://arxiv.org/abs/2603.12271)\n- **Abstract**:\n> LLMs are widely used in knowledge-intensive tasks where the same fact may be revised multiple times within context. Unlike prior work focusing on one-shot updates or single conflicts, multi-update scenarios contain multiple historically valid versions that compete at retrieval, yet remain underexplored. This challenge resembles the AB-AC interference paradigm in cognitive psychology: when the same cue A is successively associated with B and C, the old and new associations compete during retrieval, leading to bias. Inspired by this, we introduce a Dynamic Knowledge Instance (DKI) evaluation framework, modeling multi-updates of the same fact as a cue paired with a sequence of updated values, and assess models via endpoint probing of the earliest (initial) and latest (current) states. Across diverse LLMs, we observe that retrieval bias intensifies as updates increase, earliest-state accuracy stays high while latest-state accuracy drops substantially. Diagnostic analyses of attention, hidden-state similarity, and output logits further reveal that these signals become flatter and weakly discriminative on errors, providing little stable basis for identifying the latest update. Finally, cognitively inspired heuristic intervention strategies yield only modest gains and do not eliminate the bias. Our results reveal a persistent challenge in tracking and following knowledge updates in long contexts.\n\n---\n\n## 134. Task-Specific Knowledge Distillation via Intermediate Probes\n- **Authors**: Ryan Brown , Chris Russell\n- **URL**: [https://arxiv.org/abs/2603.12270](https://arxiv.org/abs/2603.12270)\n- **Abstract**:\n> Knowledge distillation from large language models (LLMs) assumes that the teacher's output distribution is a high-quality training signal. On reasoning tasks, this assumption is frequently violated. A model's intermediate representations may encode the correct answer, yet this information is lost or distorted through the vocabulary projection, where prompt formatting and answer-token choices creates brittle, noisy outputs. We introduce \\method{}, a distillation framework that bypasses this bottleneck by training lightweight probes on frozen teacher hidden states and using the probe's predictions, rather than output logits, as supervision for student training. This simple change yields consistent improvements across four reasoning benchmarks (AQuA-RAT, ARC Easy/Challenge, and MMLU), with gains most pronounced under limited data. Probes trained on intermediate representations provide cleaner labels than the teacher's own outputs, effectively denoising the distillation signal. \\method{} requires no architectural changes to student or teacher, is architecture-agnostic, and adds minimal compute since probe training is cheap and teacher representations can be cached. By exploiting internal representations, \\method{} enables practitioners to extract more value from large teacher models without additional training data or architectural complexity.\n\n---\n\n## 135. DART: Input-Difficulty-AwaRe Adaptive Threshold for Early-Exit DNNs\n- **Authors**: Parth Patne , Mahdi Taheri , Christian Herglotz , Maksim Jenihhin , Milos Krstic , Michael Hübner\n- **URL**: [https://arxiv.org/abs/2603.12269](https://arxiv.org/abs/2603.12269)\n- **Abstract**:\n> Early-exit deep neural networks enable adaptive inference by terminating computation when sufficient confidence is achieved, reducing cost for edge AI accelerators in resource-constrained settings. Existing methods, however, rely on suboptimal exit policies, ignore input difficulty, and optimize thresholds independently. This paper introduces DART (Input-Difficulty-Aware Adaptive Threshold), a framework that overcomes these limitations. DART introduces three key innovations: (1) a lightweight difficulty estimation module that quantifies input complexity with minimal computational overhead, (2) a joint exit policy optimization algorithm based on dynamic programming, and (3) an adaptive coefficient management system. Experiments on diverse DNN benchmarks (AlexNet, ResNet-18, VGG-16) demonstrate that DART achieves up to \\textbf{3.3$\\times$} speedup, \\textbf{5.1$\\times$} lower energy, and up to \\textbf{42\\%} lower average power compared to static networks, while preserving competitive accuracy. Extending DART to Vision Transformers (LeViT) yields power (5.0$\\times$) and execution-time (3.6$\\times$) gains but also accuracy loss (up to 17 percent), underscoring the need for transformer-specific early-exit mechanisms. We further introduce the Difficulty-Aware Efficiency Score (DAES), a novel multi-objective metric, under which DART achieves up to a 14.8 improvement over baselines, highlighting superior accuracy, efficiency, and robustness trade-offs.\n\n---\n\n{% endraw %}\n"}, "3d9cec1eb8d54ba4e3298027527f424aec7dc641": {"t": 1785770988.4209497, "n": 229888, "h": "8c5adbd6d55505f6", "c": "{% raw %}\n# 전체 AI 논문 - 2026-03-17\n\n## 1. Semantic Invariance in Agentic AI\n- **Authors**: I. de Zarzà , J. de Curtò , Jordi Cabot , Pietro Manzoni , Carlos T. Calafate\n- **URL**: [https://arxiv.org/abs/2603.13173](https://arxiv.org/abs/2603.13173)\n- **Abstract**:\n> Large Language Models (LLMs) increasingly serve as autonomous reasoning agents in decision support, scientific problem-solving, and multi-agent coordination systems. However, deploying LLM agents in consequential applications requires assurance that their reasoning remains stable under semantically equivalent input variations, a property we term semantic invariance. Standard benchmark evaluations, which assess accuracy on fixed, canonical problem formulations, fail to capture this critical reliability dimension. To address this shortcoming, in this paper we present a metamorphic testing framework for systematically assessing the robustness of LLM reasoning agents, applying eight semantic-preserving transformations (identity, paraphrase, fact reordering, expansion, contraction, academic context, business context, and contrastive formulation) across seven foundation models spanning four distinct architectural families: Hermes (70B, 405B), Qwen3 (30B-A3B, 235B-A22B), DeepSeek-R1, and gpt-oss (20B, 120B). Our evaluation encompasses 19 multi-step reasoning problems across eight scientific domains. The results reveal that model scale does not predict robustness: the smaller Qwen3-30B-A3B achieves the highest stability (79.6% invariant responses, semantic similarity 0.91), while larger models exhibit greater fragility.\n\n---\n\n## 2. Developing and evaluating a chatbot to support maternal health care\n- **Authors**: Smriti Jha , Vidhi Jain , Jianyu Xu , Grace Liu , Sowmya Ramesh , Jitender Nagpal , Gretchen Chapman , Benjamin Bellows , Siddhartha Goyal , Aarti Singh , Bryan Wilder\n- **URL**: [https://arxiv.org/abs/2603.13168](https://arxiv.org/abs/2603.13168)\n- **Abstract**:\n> The ability to provide trustworthy maternal health information using phone-based chatbots can have a significant impact, particularly in low-resource settings where users have low health literacy and limited access to care. However, deploying such systems is technically challenging: user queries are short, underspecified, and code-mixed across languages, answers require regional context-specific grounding, and partial or missing symptom context makes safe routing decisions difficult. We present a chatbot for maternal health in India developed through a partnership between academic researchers, a health tech company, a public health nonprofit, and a hospital. The system combines (1) stage-aware triage, routing high-risk queries to expert templates, (2) hybrid retrieval over curated maternal/newborn guidelines, and (3) evidence-conditioned generation from an LLM. Our core contribution is an evaluation workflow for high-stakes deployment under limited expert supervision. Targeting both component-level and end-to-end testing, we introduce: (i) a labeled triage benchmark (N=150) achieving 86.7% emergency recall, explicitly reporting the missed-emergency vs. over-escalation trade-off; (ii) a synthetic multi-evidence retrieval benchmark (N=100) with chunk-level evidence labels; (iii) LLM-as-judge comparison on real queries (N=781) using clinician-codesigned criteria; and (iv) expert validation. Our findings show that trustworthy medical assistants in multilingual, noisy settings require defense-in-depth design paired with multi-method evaluation, rather than any single model and evaluation method choice.\n\n---\n\n## 3. When Right Meets Wrong: Bilateral Context Conditioning with Reward-Confidence Correction for GRPO\n- **Authors**: Yu Li , Tian Lan , Zhengling Qi\n- **URL**: [https://arxiv.org/abs/2603.13134](https://arxiv.org/abs/2603.13134)\n- **Abstract**:\n> Group Relative Policy Optimization (GRPO) has emerged as an effective method for training reasoning models. While it computes advantages based on group mean, GRPO treats each output as an independent sample during the optimization and overlooks a vital structural signal: the natural contrast between correct and incorrect solutions within the same group, thus ignoring the rich, comparative data that could be leveraged by explicitly pitting successful reasoning traces against failed ones. To capitalize on this, we present a contrastive reformulation of GRPO, showing that the GRPO objective implicitly maximizes the margin between the policy ratios of correct and incorrect samples. Building on this insight, we propose Bilateral Context Conditioning (BICC), a mechanism that allows the model to cross-reference successful and failed reasoning traces during the optimization, enabling a direct information flow across samples. We further introduce Reward-Confidence Correction (RCC) to stabilize training by dynamically adjusts the advantage baseline in GRPO using reward-confidence covariance derived from the first-order approximation of the variance-minimizing estimator. Both mechanisms require no additional sampling or auxiliary models and can be adapted to all GRPO variants. Experiments on mathematical reasoning benchmarks demonstrate consistent improvements across comprehensive models and algorithms. Code is available at \\href{ this https URL }{ this https URL }.\n\n---\n\n## 4. Steve-Evolving: Open-World Embodied Self-Evolution via Fine-Grained Diagnosis and Dual-Track Knowledge Distillation\n- **Authors**: Zhengwei Xie , Zhisheng Chen , Ziyan Weng , Tingyu Wu , Chenglong Li , Vireo Zhang , Kun Wang\n- **URL**: [https://arxiv.org/abs/2603.13131](https://arxiv.org/abs/2603.13131)\n- **Abstract**:\n> Open-world embodied agents must solve long-horizon tasks where the main bottleneck is not single-step planning quality but how interaction experience is organized and evolved. To this end, we present Steve-Evolving, a non-parametric self-evolving framework that tightly couples fine-grained execution diagnosis with dual-track knowledge distillation in a closed loop. The method follows three phases: Experience Anchoring, Experience Distillation, and Knowledge-Driven Closed-Loop Control. In detail, Experience Anchoring solidifies each subgoal attempt into a structured experience tuple with a fixed schema (pre-state, action, diagnosis-result, and post-state) and organizes it in a three-tier experience space with multi-dimensional indices (e.g., condition signatures, spatial hashing, and semantic tags) plus rolling summarization for efficient and auditable recall. To ensure sufficient information density for attribution, the execution layer provides compositional diagnosis signals beyond binary outcomes, including state-difference summaries, enumerated failure causes, continuous indicators, and stagnation/loop detection. Moreover, successful trajectories of Experience Distillation are generalized into reusable skills with explicit preconditions and verification criteria, while failures are distilled into executable guardrails that capture root causes and forbid risky operations at both subgoal and task granularities. Besides, Knowledge-Driven Closed-Loop Control retrieved skills and guardrails are injected into an LLM planner, and diagnosis-triggered local replanning updates the active constraints online, forming a continual evolution process without any model parameter updates. Experiments on the long-horizon suite of Minecraft MCU demonstrate consistent improvements over static-retrieval baselines.\n\n---\n\n## 5. Beyond Final Answers: CRYSTAL Benchmark for Transparent Multimodal Reasoning Evaluation\n- **Authors**: Wayner Barrios , SouYoung Jin\n- **URL**: [https://arxiv.org/abs/2603.13099](https://arxiv.org/abs/2603.13099)\n- **Abstract**:\n> We introduce CRYSTAL (Clear Reasoning via Yielded Steps, Traceability, and Logic), a diagnostic benchmark with 6,372 instances that evaluates multimodal reasoning through verifiable intermediate steps. We propose two complementary metrics: Match F1, which scores step-level precision and recall via semantic similarity matching, and Ordered Match F1, which further penalizes disordered reasoning chains. References are constructed through a Delphi-inspired pipeline in which four independent MLLMs generate trajectories, which are then aggregated via semantic clustering and validated through human quality gates. Evaluation of 20 MLLMs, including commercial frontier systems not used during benchmark construction, reveals systematic failures that are invisible to answer accuracy: universal cherry-picking (precision far exceeds recall), non-monotonic scaling trade-offs, and disordered reasoning in which no competitive model preserves more than 60% of matched steps in the correct order. Beyond evaluation, we propose the Causal Process Reward (CPR), a multiplicative reward that couples answer correctness with step-level alignment, and CPR-Curriculum, which progressively increases reasoning difficulty during training. CPR-Curriculum achieves a 32% improvement in Match F1 via GRPO where additive reward strategies fail, improving reasoning without manual step annotation.\n\n---\n\n## 6. Structured Distillation for Personalized Agent Memory: 11x Token Reduction with Retrieval Preservation\n- **Authors**: Sydney Lewis\n- **URL**: [https://arxiv.org/abs/2603.13017](https://arxiv.org/abs/2603.13017)\n- **Abstract**:\n> Long conversations with an AI agent create a simple problem for one user: the history is useful, but carrying it verbatim is expensive. We study personalized agent memory: one user's conversation history with an agent, distilled into a compact retrieval layer for later search. Each exchange is compressed into a compound object with four fields (exchange_core, specific_context, thematic room_assignments, and regex-extracted files_touched). The searchable distilled text averages 38 tokens per exchange. Applied to 4,182 conversations (14,340 exchanges) from 6 software engineering projects, the method reduces average exchange length from 371 to 38 tokens, yielding 11x compression. We evaluate whether personalized recall survives that compression using 201 recall-oriented queries, 107 configurations spanning 5 pure and 5 cross-layer search modes, and 5 LLM graders (214,519 consensus-graded query-result pairs). The best pure distilled configuration reaches 96% of the best verbatim MRR (0.717 vs 0.745). Results are mechanism-dependent. All 20 vector search configurations remain non-significant after Bonferroni correction, while all 20 BM25 configurations degrade significantly (effect sizes |d|=0.031-0.756). The best cross-layer setup slightly exceeds the best pure verbatim baseline (MRR 0.759). Structured distillation compresses single-user agent memory without uniformly sacrificing retrieval quality. At 1/11 the context cost, thousands of exchanges fit within a single prompt while the verbatim source remains available for drill-down. We release the implementation and analysis pipeline as open-source software.\n\n---\n\n## 7. Efficient and Interpretable Multi-Agent LLM Routing via Ant Colony Optimization\n- **Authors**: Xudong Wang , Chaoning Zhang , Jiaquan Zhang , Chenghao Li , Qigan Sun , Sung-Ho Bae , Peng Wang , Ning Xie , Jie Zou , Yang Yang , Hengtao Shen\n- **URL**: [https://arxiv.org/abs/2603.12933](https://arxiv.org/abs/2603.12933)\n- **Abstract**:\n> Large Language Model (LLM)-driven Multi-Agent Systems (MAS) have demonstrated strong capability in complex reasoning and tool use, and heterogeneous agent pools further broaden the quality--cost trade-off space. Despite these advances, real-world deployment is often constrained by high inference cost, latency, and limited transparency, which hinders scalable and efficient routing. Existing routing strategies typically rely on expensive LLM-based selectors or static policies, and offer limited controllability for semantic-aware routing under dynamic loads and mixed intents, often resulting in unstable performance and inefficient resource utilization. To address these limitations, we propose AMRO-S, an efficient and interpretable routing framework for Multi-Agent Systems (MAS). AMRO-S models MAS routing as a semantic-conditioned path selection problem, enhancing routing performance through three key mechanisms: First, it leverages a supervised fine-tuned (SFT) small language model for intent inference, providing a low-overhead semantic interface for each query; second, it decomposes routing memory into task-specific pheromone specialists, reducing cross-task interference and optimizing path selection under mixed workloads; finally, it employs a quality-gated asynchronous update mechanism to decouple inference from learning, optimizing routing without increasing latency. Extensive experiments on five public benchmarks and high-concurrency stress tests demonstrate that AMRO-S consistently improves the quality--cost trade-off over strong routing baselines, while providing traceable routing evidence through structured pheromone patterns.\n\n---\n\n## 8. ODRL Policy Comparison Through Normalisation\n- **Authors**: Jaime Osvaldo Salas , Paolo Pareti , George Konstantinidis\n- **URL**: [https://arxiv.org/abs/2603.12926](https://arxiv.org/abs/2603.12926)\n- **Abstract**:\n> The ODRL language has become the standard for representing policies and regulations for digital rights. However its complexity is a barrier to its usage, which has caused many related theoretical and practical works to focus on different, and not interoperable, fragments of ODRL. Moreover, semantically equivalent policies can be expressed in numerous different ways, which makes comparing them and processing them harder. Building on top of a recently defined semantics, we tackle these problems by proposing an approach that involves a parametrised normalisation of ODRL policies into its minimal components which reformulates policies with permissions and prohibitions into policies with permissions exclusively, and simplifies complex logic constraints into simple ones. We provide algorithms to compute a normal form for ODRL policies and simplifying numerical and symbolic constraints. We prove that these algorithms preserve the semantics of policies, and analyse the size complexity of the result, which is exponential on the number of attributes and linear on the number of unique values for these attributes. We show how this makes complex policies representable in more basic fragments of ODRL, and how it reduces the problem of policy comparison to the simpler problem of checking if two rules are identical.\n\n---\n\n## 9. Context is all you need: Towards autonomous model-based process design using agentic AI in flowsheet simulations\n- **Authors**: Pascal Schäfer , Lukas J. Krinke , Martin Wlotzka , Norbert Asprion\n- **URL**: [https://arxiv.org/abs/2603.12813](https://arxiv.org/abs/2603.12813)\n- **Abstract**:\n> Agentic AI systems integrating large language models (LLMs) with reasoning and tooluse capabilities are transforming various domains - in particular, software development. In contrast, their application in chemical process flowsheet modelling remains largely unexplored. In this work, we present an agentic AI framework that delivers assistance in an industrial flowsheet simulation environment. To this end, we show the capabilities of GitHub Copilot (GitHub, Inc., 2026), when using state-of-the-art LLMs, such as Claude Opus 4.6 (Anthropic, PBC, 2026), to generate valid syntax for our in-house process modelling tool Chemasim using the technical documentation and a few commented examples as context. Based on this, we develop a multi-agent system that decomposes process development tasks with one agent solving the abstract problem using engineering knowledge and another agent implementing the solution as Chemasim code. We demonstrate the effectiveness of our framework for typical flowsheet modelling examples, including (i) a reaction/separation process, (ii) a pressure-swing distillation, and (iii) a heteroazeotropic distillation including entrainer selection. Along these lines, we discuss current limitations of the framework and outline future research directions to further enhance its capabilities.\n\n---\n\n## 10. AI Model Modulation with Logits Redistribution\n- **Authors**: Zihan Wang , Zhongkui Ma , Xinguo Feng , Zhiyang Mei , Ethan Ma , Derui Wang , Minhui Xue , Guangdong Bai\n- **URL**: [https://arxiv.org/abs/2603.12755](https://arxiv.org/abs/2603.12755)\n- **Abstract**:\n> Large-scale models are typically adapted to meet the diverse requirements of model owners and users. However, maintaining multiple specialized versions of the model is inefficient. In response, we propose AIM, a novel model modulation paradigm that enables a single model to exhibit diverse behaviors to meet the specific end requirements. AIM enables two key modulation modes: utility and focus modulations. The former provides model owners with dynamic control over output quality to deliver varying utility levels, and the latter offers users precise control to shift model's focused input features. AIM introduces a logits redistribution strategy that operates in a training data-agnostic and retraining-free manner. We establish a formal foundation to ensure AIM's regulation capability, based on the statistical properties of logits ordering via joint probability distributions. Our evaluation confirms AIM's practicality and versatility for Al model modulation, with tasks spanning image classification, semantic segmentation and text generation, and prevalent architectures including ResNet, SegFormer and Llama.\n\n---\n\n## 11. ToolTree: Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning\n- **Authors**: Shuo Yang , Soyeon Caren Han , Yihao Ding , Shuhe Wang , Eduard Hoy\n- **URL**: [https://arxiv.org/abs/2603.12740](https://arxiv.org/abs/2603.12740)\n- **Abstract**:\n> Large Language Model (LLM) agents are increasingly applied to complex, multi-step tasks that require interaction with diverse external tools across various domains. However, current LLM agent tool planning methods typically rely on greedy, reactive tool selection strategies that lack foresight and fail to account for inter-tool dependencies. In this paper, we present ToolTree, a novel Monte Carlo tree search-inspired planning paradigm for tool planning. ToolTree explores possible tool usage trajectories using a dual-stage LLM evaluation and bidirectional pruning mechanism that enables the agent to make informed, adaptive decisions over extended tool-use sequences while pruning less promising branches before and after the tool execution. Empirical evaluations across both open-set and closed-set tool planning tasks on 4 benchmarks demonstrate that ToolTree consistently improves performance while keeping the highest efficiency, achieving an average gain of around 10\\% compared to the state-of-the-art planning paradigm.\n\n---\n\n## 12. On Using Machine Learning to Early Detect Catastrophic Failures in Marine Diesel Engines\n- **Authors**: Francesco Maione , Paolo Lino , Giuseppe Giannino , Guido Maione\n- **URL**: [https://arxiv.org/abs/2603.12733](https://arxiv.org/abs/2603.12733)\n- **Abstract**:\n> Catastrophic failures of marine engines imply severe loss of functionality and destroy or damage the systems irreversibly. Being sudden and often unpredictable events, they pose a severe threat to navigation, crew, and passengers. The abrupt nature makes early detection the only effective countermeasure. However, research has concentrated on modeling the gradual degradation of components, with limited attention to sudden and anomalous phenomena. This work proposes a new method for early detection of catastrophic failures. Based on real data from a failed engine, the approach evaluates the derivatives of the deviation between actual sensor readings and expected values of engine variables. Predictions are obtained by a Random Forest, which is the most suitable Machine Learning algorithm among the tested ones. Traditional methods focus on deviations of monitored signals, whereas the proposed approach employs the derivatives of the deviations to provide earlier indications of abnormal dynamics, and to alert that a rapid and dangerous event is breaking out within the system. The method allows the detection of anomalies before measurements reach critical thresholds and alarms are triggered, which is the common method in industry. Consequently, operators can be warned in advance and shut down the engine, then prevent damage and unexpected power loss. Moreover, they have the time to safely change the ship route and avoid potential obstacles. Simulation results conf irm the effectiveness of the proposed approach in anticipating occurrence of catastrophic failures. Validation on real-world data further reinforces the robustness and practical applicability of the method. It is worth noting that data acquisition to train the predictive algorithm is not a problem, since a Deep Learning-based data augmentation procedure is used.\n\n---\n\n## 13. AI Planning Framework for LLM-Based Web Agents\n- **Authors**: Orit Shahnovsky , Rotem Dror\n- **URL**: [https://arxiv.org/abs/2603.12710](https://arxiv.org/abs/2603.12710)\n- **Abstract**:\n> Developing autonomous agents for web-based tasks is a core challenge in AI. While Large Language Model (LLM) agents can interpret complex user requests, they often operate as black boxes, making it difficult to diagnose why they fail or how they plan. This paper addresses this gap by formally treating web tasks as sequential decision-making processes. We introduce a taxonomy that maps modern agent architectures to traditional planning paradigms: Step-by-Step agents to Breadth-First Search (BFS), Tree Search agents to Best-First Tree Search, and Full-Plan-in-Advance agents to Depth-First Search (DFS). This framework allows for a principled diagnosis of system failures like context drift and incoherent task decomposition. To evaluate these behaviors, we propose five novel evaluation metrics that assess trajectory quality beyond simple success rates. We support this analysis with a new dataset of 794 human-labeled trajectories from the WebArena benchmark. Finally, we validate our evaluation framework by comparing a baseline Step-by-Step agent against a novel Full-Plan-in-Advance implementation. Our results reveal that while the Step-by-Step agent aligns more closely with human gold trajectories (38% overall success), the Full-Plan-in-Advance agent excels in technical measures such as element accuracy (89%), demonstrating the necessity of our proposed metrics for selecting appropriate agent architectures based on specific application constraints.\n\n---\n\n## 14. Generating Expressive and Customizable Evals for Timeseries Data Analysis Agents with AgentFuel\n- **Authors**: Aadyaa Maddi , Prakhar Naval , Deepti Mande , Shane Duan , Muckai Girish , Vyas Sekar\n- **URL**: [https://arxiv.org/abs/2603.12483](https://arxiv.org/abs/2603.12483)\n- **Abstract**:\n> Across many domains (e.g., IoT, observability, telecommunications, cybersecurity), there is an emerging adoption of conversational data analysis agents that enable users to \"talk to your data\" to extract insights. Such data analysis agents operate on timeseries data models; e.g., measurements from sensors or events monitoring user clicks and actions in product analytics. We evaluate 6 popular data analysis agents (both open-source and proprietary) on domain-specific data and query types, and find that they fail on stateful and incident-specific queries. We observe two key expressivity gaps in existing evals: domain-customized datasets and domain-specific query types. To enable practitioners in such domains to generate customized and expressive evals for such timeseries data agents, we present AgentFuel. AgentFuel helps domain experts quickly create customized evals to perform end-to-end functional tests. We show that AgentFuel's benchmarks expose key directions for improvement in existing data agent frameworks. We also present anecdotal evidence that using AgentFuel can improve agent performance (e.g., with GEPA). AgentFuel benchmarks are available at this https URL .\n\n---\n\n## 15. Efficient Reasoning with Balanced Thinking\n- **Authors**: Yulin Li , Tengyao Tu , Li Ding , Junjie Wang , Huiling Zhen , Yixin Chen , Yong Li , Zhuotao Tian\n- **URL**: [https://arxiv.org/abs/2603.12372](https://arxiv.org/abs/2603.12372)\n- **Abstract**:\n> Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they often suffer from overthinking, expending redundant computational steps on simple problems, or underthinking, failing to explore sufficient reasoning paths despite inherent capabilities. These issues lead to inefficiencies and potential inaccuracies, limiting practical deployment in resource-constrained settings. Existing methods to mitigate overthinking, such as suppressing reflective keywords or adjusting reasoning length, may inadvertently induce underthinking, compromising accuracy. Therefore, we propose ReBalance, a training-free framework that achieves efficient reasoning with balanced thinking. ReBalance leverages confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence. By aggregating hidden states from a small-scale dataset into reasoning mode prototypes, we compute a steering vector to guide LRMs' reasoning trajectories. A dynamic control function modulates this vector's strength and direction based on real-time confidence, pruning redundancy during overthinking, and promoting exploration during underthinking. Extensive experiments conducted on four models ranging from 0.5B to 32B, and across nine benchmarks in math reasoning, general question answering, and coding tasks demonstrate that ReBalance effectively reduces output redundancy while improving accuracy, offering a general, training-free, and plug-and-play strategy for efficient and robust LRM deployment. Code is available at this https URL .\n\n---\n\n## 16. Context-Enriched Natural Language Descriptions of Vessel Trajectories\n- **Authors**: Kostas Patroumpas , Alexandros Troupiotis-Kapeliaris , Giannis Spiliopoulos , Panagiotis Betchavas , Dimitrios Skoutas , Dimitris Zissis , Nikos Bikakis\n- **URL**: [https://arxiv.org/abs/2603.12287](https://arxiv.org/abs/2603.12287)\n- **Abstract**:\n> We address the problem of transforming raw vessel trajectory data collected from AIS into structured and semantically enriched representations interpretable by humans and directly usable by machine reasoning systems. We propose a context-aware trajectory abstraction framework that segments noisy AIS sequences into distinct trips each consisting of clean, mobility-annotated episodes. Each episode is further enriched with multi-source contextual information, such as nearby geographic entities, offshore navigation features, and weather conditions. Crucially, such representations can support generation of controlled natural language descriptions using LLMs. We empirically examine the quality of such descriptions generated using several LLMs over AIS data along with open contextual features. By increasing semantic density and reducing spatiotemporal complexity, this abstraction can facilitate downstream analytics and enable integration with LLMs for higher-level maritime reasoning tasks.\n\n---\n\n## 17. PhysMoDPO: Physically-Plausible Humanoid Motion with Preference Optimization\n- **Authors**: Yangsong Zhang , Anujith Muraleedharan , Rikhat Akizhanov , Abdul Ahad Butt , Gül Varol , Pascal Fua , Fabio Pizzati , Ivan Laptev\n- **URL**: [https://arxiv.org/abs/2603.13228](https://arxiv.org/abs/2603.13228)\n- **Abstract**:\n> Recent progress in text-conditioned human motion generation has been largely driven by diffusion models trained on large-scale human motion data. Building on this progress, recent methods attempt to transfer such models for character animation and real robot control by applying a Whole-Body Controller (WBC) that converts diffusion-generated motions into executable trajectories. While WBC trajectories become compliant with physics, they may expose substantial deviations from original motion. To address this issue, we here propose PhysMoDPO, a Direct Preference Optimization framework. Unlike prior work that relies on hand-crafted physics-aware heuristics such as foot-sliding penalties, we integrate WBC into our training pipeline and optimize diffusion model such that the output of WBC becomes compliant both with physics and original text instructions. To train PhysMoDPO we deploy physics-based and task-specific rewards and use them to assign preference to synthesized trajectories. Our extensive experiments on text-to-motion and spatial control tasks demonstrate consistent improvements of PhysMoDPO in both physical realism and task-related metrics on simulated robots. Moreover, we demonstrate that PhysMoDPO results in significant improvements when applied to zero-shot motion transfer in simulation and for real-world deployment on a G1 humanoid robot.\n\n---\n\n## 18. Visual-ERM: Reward Modeling for Visual Equivalence\n- **Authors**: Ziyu Liu , Shengyuan Ding , Xinyu Fang , Xuanlang Dai , Penghui Yang , Jianze Liang , Jiaqi Wang , Kai Chen , Dahua Lin , Yuhang Zang\n- **URL**: [https://arxiv.org/abs/2603.13224](https://arxiv.org/abs/2603.13224)\n- **Abstract**:\n> Vision-to-code tasks require models to reconstruct structured visual inputs, such as charts, tables, and SVGs, into executable or structured representations with high visual fidelity. While recent Large Vision Language Models (LVLMs) achieve strong results via supervised fine-tuning, reinforcement learning remains challenging due to misaligned reward signals. Existing rewards either rely on textual rules or coarse visual embedding similarity, both of which fail to capture fine-grained visual discrepancies and are vulnerable to reward hacking. We propose Visual Equivalence Reward Model (Visual-ERM), a multimodal generative reward model that provides fine-grained, interpretable, and task-agnostic feedback to evaluate vision-to-code quality directly in the rendered visual space. Integrated into RL, Visual-ERM improves Qwen3-VL-8B-Instruct by +8.4 on chart-to-code and yields consistent gains on table and SVG parsing (+2.7, +4.1 on average), and further strengthens test-time scaling via reflection and revision. We also introduce VisualCritic-RewardBench (VC-RewardBench), a benchmark for judging fine-grained image-to-image discrepancies on structured visual data, where Visual-ERM at 8B decisively outperforms Qwen3-VL-235B-Instruct and approaches leading closed-source models. Our results suggest that fine-grained visual reward supervision is both necessary and sufficient for vision-to-code RL, regardless of task specificity.\n\n---\n\n## 19. From Experiments to Expertise: Scientific Knowledge Consolidation for AI-Driven Computational Research\n- **Authors**: Haonan Huang\n- **URL**: [https://arxiv.org/abs/2603.13191](https://arxiv.org/abs/2603.13191)\n- **Abstract**:\n> While large language models (LLMs) have transformed AI agents into proficient executors of computational materials science, performing a hundred simulations does not make a researcher. What distinguishes research from routine execution is the progressive accumulation of knowledge -- learning which approaches fail, recognizing patterns across systems, and applying understanding to new problems. However, the prevailing paradigm in AI-driven computational science treats each execution in isolation, largely discarding hard-won insights between runs. Here we present QMatSuite, an open-source platform closing this gap. Agents record findings with full provenance, retrieve knowledge before new calculations, and in dedicated reflection sessions correct erroneous findings and synthesize observations into cross-compound patterns. In benchmarks on a six-step quantum-mechanical simulation workflow, accumulated knowledge reduces reasoning overhead by 67% and improves accuracy from 47% to 3% deviation from literature -- and when transferred to an unfamiliar material, achieves 1% deviation with zero pipeline failures.\n\n---\n\n## 20. LLM Constitutional Multi-Agent Governance\n- **Authors**: J. de Curtò , I. de Zarzà\n- **URL**: [https://arxiv.org/abs/2603.13189](https://arxiv.org/abs/2603.13189)\n- **Abstract**:\n> Large Language Models (LLMs) can generate persuasive influence strategies that shift cooperative behavior in multi-agent populations, but a critical question remains: does the resulting cooperation reflect genuine prosocial alignment, or does it mask erosion of agent autonomy, epistemic integrity, and distributional fairness? We introduce Constitutional Multi-Agent Governance (CMAG), a two-stage framework that interposes between an LLM policy compiler and a networked agent population, combining hard constraint filtering with soft penalized-utility optimization that balances cooperation potential against manipulation risk and autonomy pressure. We propose the Ethical Cooperation Score (ECS), a multiplicative composite of cooperation, autonomy, integrity, and fairness that penalizes cooperation achieved through manipulative means. In experiments on scale-free networks of 80 agents under adversarial conditions (70% violating candidates), we benchmark three regimes: full CMAG, naive filtering, and unconstrained optimization. While unconstrained optimization achieves the highest raw cooperation (0.873), it yields the lowest ECS (0.645) due to severe autonomy erosion (0.867) and fairness degradation (0.888). CMAG attains an ECS of 0.741, a 14.9% improvement, while preserving autonomy at 0.985 and integrity at 0.995, with only modest cooperation reduction to 0.770. The naive ablation (ECS = 0.733) confirms that hard constraints alone are insufficient. Pareto analysis shows CMAG dominates the cooperation-autonomy trade-off space, and governance reduces hub-periphery exposure disparities by over 60%. These findings establish that cooperation is not inherently desirable without governance: constitutional constraints are necessary to ensure that LLM-mediated influence produces ethically stable outcomes rather than manipulative equilibria.\n\n---\n\n## 21. Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights\n- **Authors**: Xingli Fang , Jung-Eun Kim\n- **URL**: [https://arxiv.org/abs/2603.13186](https://arxiv.org/abs/2603.13186)\n- **Abstract**:\n> Prior approaches for membership privacy preservation usually update or retrain all weights in neural networks, which is costly and can lead to unnecessary utility loss or even more serious misalignment in predictions between training data and non-training data. In this work, we observed three insights: i) privacy vulnerability exists in a very small fraction of weights; ii) however, most of those weights also critically impact utility performance; iii) the importance of weights stems from their locations rather than their values. According to these insights, to preserve privacy, we score critical weights, and instead of discarding those neurons, we rewind only the weights for fine-tuning. We show that, through extensive experiments, this mechanism exhibits outperforming resilience in most cases against Membership Inference Attacks while maintaining utility.\n\n---\n\n## 22. MXNorm: Reusing MXFP block scales for efficient tensor normalisation\n- **Authors**: Callum McLean , Luke Y. Prince , Alexandre Payot , Paul Balança , Carlo Luschi\n- **URL**: [https://arxiv.org/abs/2603.13180](https://arxiv.org/abs/2603.13180)\n- **Abstract**:\n> Matrix multiplication performance has long been the major bottleneck to scaling deep learning workloads, which has stimulated the design of new accelerators that use increasingly low-precision number formats. However, improvements in matrix multiplication performance have far outstripped improvements in performance on reductions and elementwise computations, which are still being performed in higher precision. In this work, we propose MXNorm, a drop-in replacement for RMSNorm that estimates the RMS using only the block scales calculated as part of the MXFP8 cast and enables a 32x decrease in the size of reduction needed for normalization. We validate our approximation method on pre-training of Llama 3 models of 125M, 1B and 8B parameters, finding minimal loss of training accuracy compared to a baseline using RMSNorm with MXFP8 matmuls. We also show practical kernel speedups using only this http URL of up to 2.4x for MXNorm over RMSNorm, corresponding to a 1.3% speedup in Llama 3 8B transformer layers in MXFP8 and a 2.6% speedup in NVFP4.\n\n---\n\n## 23. Clustering Astronomical Orbital Synthetic Data Using Advanced Feature Extraction and Dimensionality Reduction Techniques\n- **Authors**: Eraldo Pereira Marinho , Nelson Callegari Junior , Fabricio Aparecido Breve , Caetano Mazzoni Ranieri\n- **URL**: [https://arxiv.org/abs/2603.13177](https://arxiv.org/abs/2603.13177)\n- **Abstract**:\n> The dynamics of Saturn's satellite system offer a rich framework for studying orbital stability and resonance interactions. Traditional methods for analysing such systems, including Fourier analysis and stability metrics, struggle with the scale and complexity of modern datasets. This study introduces a machine learning-based pipeline for clustering approximately 22,300 simulated satellite orbits, addressing these challenges with advanced feature extraction and dimensionality reduction techniques. The key to this approach is using MiniRocket, which efficiently transforms 400 timesteps into a 9,996-dimensional feature space, capturing intricate temporal patterns. Additional automated feature extraction and dimensionality reduction techniques refine the data, enabling robust clustering analysis. This pipeline reveals stability regions, resonance structures, and other key behaviours in Saturn's satellite system, providing new insights into their long-term dynamical evolution. By integrating computational tools with traditional celestial mechanics techniques, this study offers a scalable and interpretable methodology for analysing large-scale orbital datasets and advancing the exploration of planetary dynamics.\n\n---\n\n## 24. ESG-Bench: Benchmarking Long-Context ESG Reports for Hallucination Mitigation\n- **Authors**: Siqi Sun , Ben Peng Wu , Mali Jin , Peizhen Bai , Hanpei Zhang , Xingyi Song\n- **URL**: [https://arxiv.org/abs/2603.13154](https://arxiv.org/abs/2603.13154)\n- **Abstract**:\n> As corporate responsibility increasingly incorporates environmental, social, and governance (ESG) criteria, ESG reporting is becoming a legal requirement in many regions and a key channel for documenting sustainability practices and assessing firms' long-term and ethical performance. However, the length and complexity of ESG disclosures make them difficult to interpret and automate the analysis reliably. To support scalable and trustworthy analysis, this paper introduces ESG-Bench, a benchmark dataset for ESG report understanding and hallucination mitigation in large language models (LLMs). ESG-Bench contains human-annotated question-answer (QA) pairs grounded in real-world ESG report contexts, with fine-grained labels indicating whether model outputs are factually supported or hallucinated. Framing ESG report analysis as a QA task with verifiability constraints enables systematic evaluation of LLMs' ability to extract and reason over ESG content and provides a new use case: mitigating hallucinations in socially sensitive, compliance-critical settings. We design task-specific Chain-of-Thought (CoT) prompting strategies and fine-tune multiple state-of-the-art LLMs on ESG-Bench using CoT-annotated rationales. Our experiments show that these CoT-based methods substantially outperform standard prompting and direct fine-tuning in reducing hallucinations, and that the gains transfer to existing QA benchmarks beyond the ESG domain.\n\n---\n\n## 25. Developing the PsyCogMetrics AI Lab to Evaluate Large Language Models and Advance Cognitive Science -- A Three-Cycle Action Design Science Study\n- **Authors**: Zhiye Jin , Yibai Li , K. D. Joshi , Xuefei (Nancy) Deng , Xiaobing (Emily)Li\n- **URL**: [https://arxiv.org/abs/2603.13126](https://arxiv.org/abs/2603.13126)\n- **Abstract**:\n> This study presents the development of the PsyCogMetrics AI Lab ( this http URL ), an integrated, cloud-based platform that operationalizes psychometric and cognitive-science methodologies for Large Language Model (LLM) evaluation. Framed as a three-cycle Action Design Science study, the Relevance Cycle identifies key limitations in current evaluation methods and unfulfilled stakeholder needs. The Rigor Cycle draws on kernel theories such as Popperian falsifiability, Classical Test Theory, and Cognitive Load Theory to derive deductive design objectives. The Design Cycle operationalizes these objectives through nested Build-Intervene-Evaluate loops. The study contributes a novel IT artifact, a validated design for LLM evaluation, benefiting research at the intersection of AI, psychology, cognitive science, and the social and behavioral sciences.\n\n---\n\n## 26. Geometry-Guided Camera Motion Understanding in VideoLLMs\n- **Authors**: Haoan Feng , Sri Harsha Musunuri , Guan-Ming Su\n- **URL**: [https://arxiv.org/abs/2603.13119](https://arxiv.org/abs/2603.13119)\n- **Abstract**:\n> Camera motion is a fundamental geometric signal that shapes visual perception and cinematic style, yet current video-capable vision-language models (VideoLLMs) rarely represent it explicitly and often fail on fine-grained motion primitives. We address this gap with a framework of $\\textbf{benchmarking}$, $\\textbf{diagnosis}$, and $\\textbf{injection}$. We curate $\\textbf{CameraMotionDataset}$, a large-scale synthetic dataset with explicit camera control, formulate camera motion as constraint-aware multi-label recognition, and construct a VQA benchmark--$\\textbf{CameraMotionVQA}$. Across diverse off-the-shelf VideoLLMs, we observe substantial errors in recognizing camera motion primitives. Probing experiments on a Qwen2.5-VL vision encoder suggest that camera motion cues are weakly represented, especially in deeper ViT blocks, helping explain the observed failure modes. To bridge this gap without costly training or fine-tuning, we propose a lightweight, model-agnostic pipeline that extracts geometric camera cues from 3D foundation models (3DFMs), predicts constrained motion primitives with a temporal classifier, and injects them into downstream VideoLLM inference via structured prompting. Experiments demonstrate improved motion recognition and more camera-aware model responses, highlighting geometry-driven cue extraction and structured prompting as practical steps toward a camera-aware VideoLLM and VLA system. The dataset and benchmark is publicly available at this https URL .\n\n---\n\n## 27. BoSS: A Best-of-Strategies Selector as an Oracle for Deep Active Learning\n- **Authors**: Denis Huseljic , Paul Hahn , Marek Herde , Christoph Sandrock , Bernhard Sick\n- **URL**: [https://arxiv.org/abs/2603.13109](https://arxiv.org/abs/2603.13109)\n- **Abstract**:\n> Active learning (AL) aims to reduce annotation costs while maximizing model performance by iteratively selecting valuable instances. While foundation models have made it easier to identify these instances, existing selection strategies still lack robustness across different models, annotation budgets, and datasets. To highlight the potential weaknesses of existing AL strategies and provide a reference point for research, we explore oracle strategies, i.e., strategies that approximate the optimal selection by accessing ground-truth information unavailable in practical AL scenarios. Current oracle strategies, however, fail to scale effectively to large datasets and complex deep neural networks. To tackle these limitations, we introduce the Best-of-Strategy Selector (BoSS), a scalable oracle strategy designed for large-scale AL scenarios. BoSS constructs a set of candidate batches through an ensemble of selection strategies and then selects the batch yielding the highest performance gain. As an ensemble of selection strategies, BoSS can be easily extended with new state-of-the-art strategies as they emerge, ensuring it remains a reliable oracle strategy in the future. Our evaluation demonstrates that i) BoSS outperforms existing oracle strategies, ii) state-of-the-art AL strategies still fall noticeably short of oracle performance, especially in large-scale datasets with many classes, and iii) one possible solution to counteract the inconsistent performance of AL strategies might be to employ an ensemble-based approach for the selection.\n\n---\n\n## 28. Evaluating VLMs' Spatial Reasoning Over Robot Motion: A Step Towards Robot Planning with Motion Preferences\n- **Authors**: Wenxi Wu , Jingjing Zhang , Martim Brandão\n- **URL**: [https://arxiv.org/abs/2603.13100](https://arxiv.org/abs/2603.13100)\n- **Abstract**:\n> Understanding user instructions and object spatial relations in surrounding environments is crucial for intelligent robot systems to assist humans in various tasks. The natural language and spatial reasoning capabilities of Vision-Language Models (VLMs) have the potential to enhance the generalization of robot planners on new tasks, objects, and motion specifications. While foundation models have been applied to task planning, it is still unclear the degree to which they have the capability of spatial reasoning required to enforce user preferences or constraints on motion, such as desired distances from objects, topological properties, or motion style preferences. In this paper, we evaluate the capability of four state-of-the-art VLMs at spatial reasoning over robot motion, using four different querying methods. Our results show that, with the highest-performing querying method, Qwen2.5-VL achieves 71.4% accuracy zero-shot and 75% on a smaller model after fine-tuning, and GPT-4o leads to lower performance. We evaluate two types of motion preferences (object-proximity and path-style), and we also analyze the trade-off between accuracy and computation cost in number of tokens. This work shows some promise in the potential of VLM integration with robot motion planning pipelines.\n\n---\n\n## 29. Human-in-the-Loop LLM Grading for Handwritten Mathematics Assessments\n- **Authors**: Arne Vanhoyweghen , Vincent Holst , Melika Mobini , Lukas Van de Voorde , Tibo Vanleke , Bert Verbruggen , Brecht Verbeken , Andres Algaba , Sam Verboven , Marie-Anne Guerry , Filip Van Droogenbroeck , Vincent Ginis\n- **URL**: [https://arxiv.org/abs/2603.13083](https://arxiv.org/abs/2603.13083)\n- **Abstract**:\n> Providing timely and individualised feedback on handwritten student work is highly beneficial for learning but difficult to achieve at scale. This challenge has become more pressing as generative AI undermines the reliability of take-home assessments, shifting emphasis toward supervised, in-class evaluation. We present a scalable, end-to-end workflow for LLM-assisted grading of short, pen-and-paper assessments. The workflow spans (1) constructing solution keys, (2) developing detailed rubric-style grading keys used to guide the LLM, and (3) a grading procedure that combines automated scanning and anonymisation, multi-pass LLM scoring, automated consistency checks, and mandatory human verification. We deploy the system in two undergraduate mathematics courses using six low-stakes in-class tests. Empirically, LLM assistance reduces grading time by approximately 23% while achieving agreement comparable to, and in several cases tighter than, fully manual grading. Occasional model errors occur but are effectively contained by the hybrid design. Overall, our results show that carefully embedded human-in-the-loop LLM grading can substantially reduce workload while maintaining fairness and accuracy.\n\n---\n\n## 30. GeoChemAD: Benchmarking Unsupervised Geochemical Anomaly Detection for Mineral Exploration\n- **Authors**: Yihao Ding , Yiran Zhang , Chris Gonzalez , Eun-Jung Holden , Wei Liu\n- **URL**: [https://arxiv.org/abs/2603.13068](https://arxiv.org/abs/2603.13068)\n- **Abstract**:\n> Geochemical anomaly detection plays a critical role in mineral exploration as deviations from regional geochemical baselines may indicate mineralization. Existing studies suffer from two key limitations: (1) single region scenarios which limit model generalizability; (2) proprietary datasets, which makes result reproduction unattainable. In this work, we introduce \\textbf{GeoChemAD}, an open-source benchmark dataset compiled from government-led geological surveys, covering multiple regions, sampling sources, and target elements. The dataset comprises eight subsets representing diverse spatial scales and sampling conditions. To establish strong baselines, we reproduce and benchmark a range of unsupervised anomaly detection methods, including statistical models, generative and transformer-based approaches. Furthermore, we propose \\textbf{GeoChemFormer}, a transformer-based framework that leverages self-supervised pretraining to learn target-element-aware geochemical representations for spatial samples. Extensive experiments demonstrate that GeoChemFormer consistently achieves superior and robust performance across all eight subsets, outperforming existing unsupervised methods in both anomaly detection accuracy and generalization capability. The proposed dataset and framework provide a foundation for reproducible research and future development in this direction.\n\n---\n\n## 31. L2GTX: From Local to Global Time Series Explanations\n- **Authors**: Ephrem Tibebe Mekonnen , Luca Longo , Lucas Rizzo , Pierpaolo Dondio\n- **URL**: [https://arxiv.org/abs/2603.13065](https://arxiv.org/abs/2603.13065)\n- **Abstract**:\n> Deep learning models achieve high accuracy in time series classification, yet understanding their class-level decision behaviour remains challenging. Explanations for time series must respect temporal dependencies and identify patterns that recur across instances. Existing approaches face three limitations: model-agnostic XAI methods developed for images and tabular data do not readily extend to time series, global explanation synthesis for time series remains underexplored, and most existing global approaches are model-specific. We propose L2GTX, a model-agnostic framework that generates class-wise global explanations by aggregating local explanations from a representative set of instances. L2GTX extracts clusters of parameterised temporal event primitives, such as increasing or decreasing trends and local extrema, together with their importance scores from instance-level explanations produced by LOMATCE. These clusters are merged across instances to reduce redundancy, and an instance-cluster importance matrix is used to estimate global relevance. Under a user-defined instance selection budget, L2GTX selects representative instances that maximise coverage of influential clusters. Events from the selected instances are then aggregated into concise class-wise global explanations. Experiments on six benchmark time series datasets show that L2GTX produces compact and interpretable global explanations while maintaining stable global faithfulness measured as mean local surrogate fidelity.\n\n---\n\n## 32. Competition-Aware CPC Forecasting with Near-Market Coverage\n- **Authors**: Sebastian Frey , Edoardo Beccari , Maximilian Kranz , Nicolò Alberto Pellizzari , Ali Mete Karaman , Qiwei Han , Maximilian Kaiser\n- **URL**: [https://arxiv.org/abs/2603.13059](https://arxiv.org/abs/2603.13059)\n- **Abstract**:\n> Cost-per-click (CPC) in paid search is a volatile auction outcome generated by a competitive landscape that is only partially observable from any single advertiser's history. Using Google Ads auction logs from a concentrated car-rental market (2021--2023), we forecast weekly CPC for 1,811 keyword series and approximate latent competition through complementary signals derived from keyword text, CPC trajectories, and geographic market structure. We construct (i) semantic neighborhoods and a semantic keyword graph from pretrained transformer-based representations of keyword text, (ii) behavioral neighborhoods via Dynamic Time Warping (DTW) alignment of CPC trajectories, and (iii) geographic-intent covariates capturing localized demand and marketplace heterogeneity. We extensively evaluate these signals both as stand-alone covariates and as relational priors in spatiotemporal graph forecasters, benchmarking them against strong statistical, neural, and time-series foundation-model baselines. Across methods, competition-aware augmentation improves stability and error profiles at business-relevant medium and longer horizons, where competitive regimes shift and volatility is most consequential. The results show that broad market-outcome coverage, combined with keyword-derived semantic and geographic priors, provides a scalable way to approximate latent competition and improve CPC forecasting in auction-driven markets.\n\n---\n\n## 33. Team RAS in 10th ABAW Competition: Multimodal Valence and Arousal Estimation Approach\n- **Authors**: Elena Ryumina (1), Maxim Markitantov (1), Alexandr Axyonov (1), Dmitry Ryumin (1), Mikhail Dolgushin (1), Denis Dresvyanskiy (2), Alexey Karpov (1 and 2) ((1) St. Petersburg Federal Research Center of the Russian Academy of Sciences, St. Petersburg, Russia, (2) ITMO University, St. Petersburg, Russia)\n- **URL**: [https://arxiv.org/abs/2603.13056](https://arxiv.org/abs/2603.13056)\n- **Abstract**:\n> Continuous emotion recognition in terms of valence and arousal under in-the-wild (ITW) conditions remains a challenging problem due to large variations in appearance, head pose, illumination, occlusions, and subject-specific patterns of affective expression. We present a multimodal method for valence-arousal estimation ITW. Our method combines three complementary modalities: face, behavior, and audio. The face modality relies on GRADA-based frame-level embeddings and Transformer-based temporal regression. We use Qwen3-VL-4B-Instruct to extract behavior-relevant information from video segments, while Mamba is used to model temporal dynamics across segments. The audio modality relies on WavLM-Large with attention-statistics pooling and includes a cross-modal filtering stage to reduce the influence of unreliable or non-speech segments. To fuse modalities, we explore two fusion strategies: a Directed Cross-Modal Mixture-of-Experts Fusion Strategy that learns interactions between modalities with adaptive weighting, and a Reliability-Aware Audio-Visual Fusion Strategy that combines visual features at the frame-level while using audio as complementary context. The results are reported on the Aff-Wild2 dataset following the 10th Affective Behavior Analysis in-the-Wild (ABAW) challenge protocol. Experiments demonstrate that the proposed multimodal fusion strategy achieves a Concordance Correlation Coefficient (CCC) of 0.658 on the Aff-Wild2 development set.\n\n---\n\n## 34. Are General-Purpose Vision Models All We Need for 2D Medical Image Segmentation? A Cross-Dataset Empirical Study\n- **Authors**: Vanessa Borst , Samuel Kounev\n- **URL**: [https://arxiv.org/abs/2603.13044](https://arxiv.org/abs/2603.13044)\n- **Abstract**:\n> Medical image segmentation (MIS) is a fundamental component of computer-assisted diagnosis and clinical decision support systems. Over the past decade, numerous architectures specifically tailored to medical imaging have emerged to address domain-specific challenges such as low contrast, small anatomical structures, and limited annotated data. In parallel, rapid progress in computer vision has produced highly capable general-purpose vision models (GP-VMs) originally designed for natural images. Despite their strong performance on standard vision benchmarks, their effectiveness for MIS remains insufficiently understood. In this work, we conduct a controlled empirical study to examine whether specialized medical segmentation architectures (SMAs) provide systematic advantages over modern GP-VMs for 2D MIS. We compare eleven SMAs and GP-VMs using a unified training and evaluation protocol. Experiments are performed across three heterogeneous datasets covering different imaging modalities, class structures, and data characteristics. Beyond segmentation accuracy, we analyze qualitative Grad-CAM visualizations to investigate explainability (XAI) behavior. Our results demonstrate that, for the analyzed datasets, GP-VMs out-perform the majority of specialized MIS models. Moreover, XAI analyses indicate that GP-VMs can capture clinically relevant structures without explicit domain-specific architectural design. These findings suggest that GP-VMs can represent a viable alternative to domain-specific methods, highlighting the importance of informed model selection for end-to-end MIS systems. All code and resources are available at GitHub.\n\n---\n\n## 35. Interrogating Design Homogenization in Web Vibe Coding\n- **Authors**: Donghoon Shin , Alice Gao , Rock Yuren Pang , Jaewook Lee , Katharina Reinecke , Emily Tseng\n- **URL**: [https://arxiv.org/abs/2603.13036](https://arxiv.org/abs/2603.13036)\n- **Abstract**:\n> Generative AI is known for its tendency to homogenize, often reproducing dominant style conventions found in training data. However, it remains unclear how these homogenizing effects extend to complex structural tasks like web design. As lay creators increasingly turn to LLMs to 'vibe-code' websites -- prompting for aesthetic and functional goals rather than writing code -- they may inadvertently narrow the diversity of their designs, and limit creative expression throughout the internet. In this paper, we interrogate the possibility of design homogenization in web vibe coding. We first characterize the vibe coding lifecycle, pinpointing stages where homogenization risks may arise. We then conduct a sociotechnical risk analysis unpacking the potential harms of web vibe coding and their interaction with design homogenization. We identify that the push for frictionless generation can exacerbate homogenization and its harms. Finally, we propose a mitigation framework centered on the idea of productive friction. Through case studies at the micro, meso, and macro levels, we show how centering productive friction can empower creators to challenge default outputs and preserve diverse expression in AI-mediated web design.\n\n---\n\n## 36. Purify Once, Edit Freely: Breaking Image Protections under Model Mismatch\n- **Authors**: Qichen Zhao , Shengfang Zhai , Xinjian Bai , Qingni Shen , Qiqi Lin , Yansong Gao , Zhonghai Wu\n- **URL**: [https://arxiv.org/abs/2603.13028](https://arxiv.org/abs/2603.13028)\n- **Abstract**:\n> Diffusion models enable high-fidelity image editing but can also be misused for unauthorized style imitation and harmful content generation. To mitigate these risks, proactive image protection methods embed small, often imperceptible adversarial perturbations into images before sharing to disrupt downstream editing or fine-tuning. However, in realistic post-release scenarios, content owners cannot control downstream processing pipelines, and protections optimized for a surrogate model may fail when attackers use mismatched diffusion pipelines. Existing purification methods can weaken protections but often sacrifice image quality and rarely examine architectural mismatch. We introduce a unified post-release purification framework to evaluate protection survivability under model mismatch. We propose two practical purifiers: VAE-Trans, which corrects protected images via latent-space projection, and EditorClean, which performs instruction-guided reconstruction with a Diffusion Transformer to exploit architectural heterogeneity. Both operate without access to protected images or defense internals. Across 2,100 editing tasks and six representative protection methods, EditorClean consistently restores editability. Compared to protected inputs, it improves PSNR by 3-6 dB and reduces FID by 50-70 percent on downstream edits, while outperforming prior purification baselines by about 2 dB PSNR and 30 percent lower FID. Our results reveal a purify-once, edit-freely failure mode: once purification succeeds, the protective signal is largely removed, enabling unrestricted editing. This highlights the need to evaluate protections under model mismatch and design defenses robust to heterogeneous attackers.\n\n---\n\n## 37. SortScrews: A Dataset and Baseline for Real-time Screw Classification\n- **Authors**: Tianhao Fu , Bingxuan Yang , Juncheng Guo , Shrena Sribalan , Yucheng Chen\n- **URL**: [https://arxiv.org/abs/2603.13027](https://arxiv.org/abs/2603.13027)\n- **Abstract**:\n> Automatic identification of screw types is important for industrial automation, robotics, and inventory management. However, publicly available datasets for screw classification are scarce, particularly for controlled single-object scenarios commonly encountered in automated sorting systems. In this work, we introduce $\\textbf{SortScrews}$, a dataset for casewise visual classification of screws. The dataset contains 560 RGB images at $512\\times512$ resolution covering six screw types and a background class. Images are captured using a standardized acquisition setup and include mild variations in lighting and camera perspective across four capture settings. To facilitate reproducible research and dataset expansion, we also provide a reusable data collection script that allows users to easily construct similar datasets for custom hardware components using inexpensive camera setups. We establish baseline results using transfer learning with EfficientNet-B0 and ResNet-18 classifiers pretrained on ImageNet. In addition, we conduct a well-explored failure analysis. Despite the limited dataset size, these lightweight models achieve strong classification accuracy, demonstrating that controlled acquisition conditions enable effective learning even with relatively small datasets. The dataset, collection pipeline, and baseline training code are publicly available at this https URL .\n\n---\n\n## 38. SAW: Toward a Surgical Action World Model via Controllable and Scalable Video Generation\n- **Authors**: Sampath Rapuri , Lalithkumar Seenivasan , Dominik Schneider , Roger Soberanis-Mukul , Yufan He , Hao Ding , Jiru Xu , Chenhao Yu , Chenyan Jing , Pengfei Guo , Daguang Xu , Mathias Unberath\n- **URL**: [https://arxiv.org/abs/2603.13024](https://arxiv.org/abs/2603.13024)\n- **Abstract**:\n> A surgical world model capable of generating realistic surgical action videos with precise control over tool-tissue interactions can address fundamental challenges in surgical AI and simulation -- from data scarcity and rare event synthesis to bridging the sim-to-real gap for surgical automation. However, current video generation methods, the very core of such surgical world models, require expensive annotations or complex structured intermediates as conditioning signals at inference, limiting their scalability. Other approaches exhibit limited temporal consistency across complex laparoscopic scenes and do not possess sufficient realism. We propose Surgical Action World (SAW) -- a step toward surgical action world modeling through video diffusion conditioned on four lightweight signals: language prompts encoding tool-action context, a reference surgical scene, tissue affordance mask, and 2D tool-tip trajectories. We design a conditional video diffusion approach that reformulates video-to-video diffusion into trajectory-conditioned surgical action synthesis. The backbone diffusion model is fine-tuned on a custom-curated dataset of 12,044 laparoscopic clips with lightweight spatiotemporal conditioning signals, leveraging a depth consistency loss to enforce geometric plausibility without requiring depth at inference. SAW achieves state-of-the-art temporal consistency (CD-FVD: 199.19 vs. 546.82) and strong visual quality on held-out test data. Furthermore, we demonstrate its downstream utility for (a) surgical AI, where augmenting rare actions with SAW-generated videos improves action recognition (clipping F1-score: 20.93% to 43.14%; cutting: 0.00% to 8.33%) on real test data, and (b) surgical simulation, where rendering tool-tissue interaction videos from simulator-derived trajectory points toward a visually faithful simulation engine.\n\n---\n\n## 39. daVinci-Env: Open SWE Environment Synthesis at Scale\n- **Authors**: Dayuan Fu , Shenyu Wu , Yunze Wu , Zerui Peng , Yaxing Huang , Jie Sun , Ji Zeng , Mohan Jiang , Lin Zhang , Yukun Li , Jiarui Hu , Liming Liu , Jinlong Hou , Pengfei Liu\n- **URL**: [https://arxiv.org/abs/2603.13023](https://arxiv.org/abs/2603.13023)\n- **Abstract**:\n> Training capable software engineering (SWE) agents demands large-scale, executable, and verifiable environments that provide dynamic feedback loops for iterative code editing, test execution, and solution refinement. However, existing open-source datasets remain limited in scale and repository diversity, while industrial solutions are opaque with unreleased infrastructure, creating a prohibitive barrier for most academic research groups. We present OpenSWE, the largest fully transparent framework for SWE agent training in Python, comprising 45,320 executable Docker environments spanning over 12.8k repositories, with all Dockerfiles, evaluation scripts, and infrastructure fully open-sourced for reproducibility. OpenSWE is built through a multi-agent synthesis pipeline deployed across a 64-node distributed cluster, automating repository exploration, Dockerfile construction, evaluation script generation, and iterative test analysis. Beyond scale, we propose a quality-centric filtering pipeline that characterizes the inherent difficulty of each environment, filtering out instances that are either unsolvable or insufficiently challenging and retaining only those that maximize learning efficiency. With $891K spent on environment construction and an additional $576K on trajectory sampling and difficulty-aware curation, the entire project represents a total investment of approximately $1.47 million, yielding about 13,000 curated trajectories from roughly 9,000 quality guaranteed environments. Extensive experiments validate OpenSWE's effectiveness: OpenSWE-32B and OpenSWE-72B achieve 62.4% and 66.0% on SWE-bench Verified, establishing SOTA among Qwen2.5 series. Moreover, SWE-focused training yields substantial out-of-domain improvements, including up to 12 points on mathematical reasoning and 5 points on science benchmarks, without degrading factual recall.\n\n---\n\n## 40. ARL-Tangram: Unleash the Resource Efficiency in Agentic Reinforcement Learning\n- **Authors**: Bangjun Xiao , Yihao Zhao , Xiangwei Deng , Shihua Yu , Yuxing Xiang , Huaqiu Liu , Qiying Wang , Liang Zhao , Hailin Zhang , Xuanzhe Liu , Xin Jin , Fuli Luo\n- **URL**: [https://arxiv.org/abs/2603.13019](https://arxiv.org/abs/2603.13019)\n- **Abstract**:\n> Agentic reinforcement learning (RL) has emerged as a transformative workload in cloud clusters, enabling large language models (LLMs) to solve complex problems through interactions with real world. However, unlike traditional RL, agentic RL demands substantial external cloud resources, e.g., CPUs for code execution and GPUs for reward models, that exist outside the primary training cluster. Existing agentic RL framework typically rely on static over-provisioning, i.e., resources are often tied to long-lived trajectories or isolated by tasks, which leads to severe resource inefficiency. We propose the action-level orchestration, and incorporate it into ARL-Tangram, a unified resource management system that enables fine-grained external resource sharing and elasticity. ARL-Tangram utilizes a unified action-level formulation and an elastic scheduling algorithm to minimize action completion time (ACT) while satisfying heterogeneous resource constraints. Further, heterogeneous resource managers are tailored to efficiently support the action-level execution on resources with heterogeneous characteristics and topologies. Evaluation on real-world agentic RL tasks demonstrates that ARL-Tangram improves average ACT by up to 4.3$\\times$, speeds up the step duration of RL training by up to 1.5$\\times$, and saves the external resources by up to 71.2$\\%$. This system has been deployed to support the training of the MiMo series models.\n\n---\n\n## 41. Fair Lung Disease Diagnosis from Chest CT via Gender-Adversarial Attention Multiple Instance Learning\n- **Authors**: Aditya Parikh , Aasa Feragen\n- **URL**: [https://arxiv.org/abs/2603.12988](https://arxiv.org/abs/2603.12988)\n- **Abstract**:\n> We present a fairness-aware framework for multi-class lung disease diagnosis from chest CT volumes, developed for the Fair Disease Diagnosis Challenge at the PHAROS-AIF-MIH Workshop (CVPR 2026). The challenge requires classifying CT scans into four categories -- Healthy, COVID-19, Adenocarcinoma, and Squamous Cell Carcinoma -- with performance measured as the average of per-gender macro F1 scores, explicitly penalizing gender-inequitable predictions. Our approach addresses two core difficulties: the sparse pathological signal across hundreds of slices, and a severe demographic imbalance compounded across disease class and gender. We propose an attention-based Multiple Instance Learning (MIL) model on a ConvNeXt backbone that learns to identify diagnostically relevant slices without slice-level supervision, augmented with a Gradient Reversal Layer (GRL) that adversarially suppresses gender-predictive structure in the learned scan representation. Training incorporates focal loss with label smoothing, stratified cross-validation over joint (class, gender) strata, and targeted oversampling of the most underrepresented subgroup. At inference, all five-fold checkpoints are ensembled with horizontal-flip test-time augmentation via soft logit voting and out-of-the-fold threshold optimization for robustness. Our model achieves a mean validation competition score of 0.685 (std - 0.030), with the best single fold reaching 0.759. All training and inference code is publicly available at this https URL\n\n---\n\n## 42. Is Human Annotation Necessary? Iterative MBR Distillation for Error Span Detection in Machine Translation\n- **Authors**: Boxuan Lyu , Haiyue Song , Zhi Qu\n- **URL**: [https://arxiv.org/abs/2603.12983](https://arxiv.org/abs/2603.12983)\n- **Abstract**:\n> Error Span Detection (ESD) is a crucial subtask in Machine Translation (MT) evaluation, aiming to identify the location and severity of translation errors. While fine-tuning models on human-annotated data improves ESD performance, acquiring such data is expensive and prone to inconsistencies among annotators. To address this, we propose a novel self-evolution framework based on Minimum Bayes Risk (MBR) decoding, named Iterative MBR Distillation for ESD, which eliminates the reliance on human annotations by leveraging an off-the-shelf LLM to generate pseudo-labels. Extensive experiments on the WMT Metrics Shared Task datasets demonstrate that models trained solely on these self-generated pseudo-labels outperform both unadapted base model and supervised baselines trained on human annotations at the system and span levels, while maintaining competitive sentence-level performance.\n\n---\n\n## 43. Efficient Real-World Autonomous Racing via Attenuated Residual Policy Optimization\n- **Authors**: Raphael Trumpp , Denis Hoornaert , Mirco Theile , Marco Caccamo\n- **URL**: [https://arxiv.org/abs/2603.12960](https://arxiv.org/abs/2603.12960)\n- **Abstract**:\n> Residual policy learning (RPL), in which a learned policy refines a static base policy using deep reinforcement learning (DRL), has shown strong performance across various robotic applications. Its effectiveness is particularly evident in autonomous racing, a domain that serves as a challenging benchmark for real-world DRL. However, deploying RPL-based controllers introduces system complexity and increases inference latency. We address this by introducing an extension of RPL named attenuated residual policy optimization ($\\alpha$-RPO). Unlike standard RPL, $\\alpha$-RPO yields a standalone neural policy by progressively attenuating the base policy, which initially serves to bootstrap learning. Furthermore, this mechanism enables a form of privileged learning, where the base policy is permitted to use sensor modalities not required for final deployment. We design $\\alpha$-RPO to integrate seamlessly with PPO, ensuring that the attenuated influence of the base controller is dynamically compensated during policy optimization. We evaluate $\\alpha$-RPO by building a framework for 1:10-scaled autonomous racing around it. In both simulation and zero-shot real-world transfer to Roboracer cars, $\\alpha$-RPO not only reduces system complexity but also improves driving performance compared to baselines - demonstrating its practicality for robotic deployment. Our code is available at: this https URL .\n\n---\n\n## 44. Delta1 with LLM: symbolic and neural integration for credible and explainable reasoning\n- **Authors**: Yang Xu , Jun Liu , Shuwei Chen , Chris Nugent , Hailing Guo\n- **URL**: [https://arxiv.org/abs/2603.12953](https://arxiv.org/abs/2603.12953)\n- **Abstract**:\n> Neuro-symbolic reasoning increasingly demands frameworks that unite the formal rigor of logic with the interpretability of large language models (LLMs). We introduce an end to end explainability by construction pipeline integrating the Automated Theorem Generator Delta1 based on the full triangular standard contradiction (FTSC) with LLMs. Delta1 deterministically constructs minimal unsatisfiable clause sets and complete theorems in polynomial time, ensuring both soundness and minimality by construction. The LLM layer verbalizes each theorem and proof trace into coherent natural language explanations and actionable insights. Empirical studies across health care, compliance, and regulatory domains show that Delta1 and LLM enables interpretable, auditable, and domain aligned reasoning. This work advances the convergence of logic, language, and learning, positioning constructive theorem generation as a principled foundation for neuro-symbolic explainable AI.\n\n---\n\n## 45. Thinking in Streaming Video\n- **Authors**: Zikang Liu , Longteng Guo , Handong Li , Ru Zhen , Xingjian He , Ruyi Ji , Xiaoming Ren , Yanhao Zhang , Haonan Lu , Jing Liu\n- **URL**: [https://arxiv.org/abs/2603.12938](https://arxiv.org/abs/2603.12938)\n- **Abstract**:\n> Real-time understanding of continuous video streams is essential for interactive assistants and multimodal agents operating in dynamic environments. However, most existing video reasoning approaches follow a batch paradigm that defers reasoning until the full video context is observed, resulting in high latency and growing computational cost that are incompatible with streaming scenarios. In this paper, we introduce ThinkStream, a framework for streaming video reasoning based on a Watch--Think--Speak paradigm that enables models to incrementally update their understanding as new video observations arrive. At each step, the model performs a short reasoning update and decides whether sufficient evidence has accumulated to produce a response. To support long-horizon streaming, we propose Reasoning-Compressed Streaming Memory (RCSM), which treats intermediate reasoning traces as compact semantic memory that replaces outdated visual tokens while preserving essential context. We further train the model using a Streaming Reinforcement Learning with Verifiable Rewards scheme that aligns incremental reasoning and response timing with the requirements of streaming interaction. Experiments on multiple streaming video benchmarks show that ThinkStream significantly outperforms existing online video models while maintaining low latency and memory usage. Code, models and data will be released at this https URL\n\n---\n\n## 46. Surprised by Attention: Predictable Query Dynamics for Time Series Anomaly Detection\n- **Authors**: Kadir-Kaan Özer , René Ebeling , Markus Enzweiler\n- **URL**: [https://arxiv.org/abs/2603.12916](https://arxiv.org/abs/2603.12916)\n- **Abstract**:\n> Multivariate time series anomalies often manifest as shifts in cross-channel dependencies rather than simple amplitude excursions. In autonomous driving, for instance, a steering command might be internally consistent but decouple from the resulting lateral acceleration. Residual-based detectors can miss such anomalies when flexible sequence models still reconstruct signals plausibly despite altered coordination. We introduce AxonAD, an unsupervised detector that treats multi-head attention query evolution as a short horizon predictable process. A gradient-updated reconstruction pathway is coupled with a history-only predictor that forecasts future query vectors from past context. This is trained via a masked predictor-target objective against an exponential moving average (EMA) target encoder. At inference, reconstruction error is combined with a tail-aggregated query mismatch score, which measures cosine deviation between predicted and target queries on recent timesteps. This dual approach provides sensitivity to structural dependency shifts while retaining amplitude-level detection. On proprietary in-vehicle telemetry with interval annotations and on the TSB-AD multi-variate suite (17 datasets, 180 series) with threshold-free and range-aware metrics, AxonAD improves ranking quality and temporal localization over strong baselines. Ablations confirm that query prediction and combined scoring are the primary drivers of the observed gains. Code is available at the URL this https URL .\n\n---\n\n## 47. Stake the Points: Structure-Faithful Instance Unlearning\n- **Authors**: Kiseong Hong , JungKyoo Shin , Eunwoo Kim\n- **URL**: [https://arxiv.org/abs/2603.12915](https://arxiv.org/abs/2603.12915)\n- **Abstract**:\n> Machine unlearning (MU) addresses privacy risks in pretrained models. The main goal of MU is to remove the influence of designated data while preserving the utility of retained knowledge. Achieving this goal requires preserving semantic relations among retained instances, which existing studies often overlook. We observe that without such preservation, models suffer from progressive structural collapse, undermining both the deletion-retention balance. In this work, we propose a novel structure-faithful framework that introduces stakes, i.e., semantic anchors that serve as reference points to maintain the knowledge structure. By leveraging these anchors, our framework captures and stabilizes the semantic organization of knowledge. Specifically, we instantiate the anchors from language-driven attribute descriptions encoded by a semantic encoder (e.g., CLIP). We enforce preservation of the knowledge structure via structure-aware alignment and regularization: the former aligns the organization of retained knowledge before and after unlearning around anchors, while the latter regulates updates to structure-critical parameters. Results from image classification, retrieval, and face recognition show average gains of 32.9%, 22.5%, and 19.3% in performance, balancing the deletion-retention trade-off and enhancing generalization.\n\n---\n\n## 48. FedBPrompt: Federated Domain Generalization Person Re-Identification via Body Distribution Aware Visual Prompts\n- **Authors**: Xin Xu , Weilong Li , Wei Liu , Wenke Huang , Zhixi Yu , Bin Yang , Xiaoying Liao , Kui Jiang\n- **URL**: [https://arxiv.org/abs/2603.12912](https://arxiv.org/abs/2603.12912)\n- **Abstract**:\n> Federated Domain Generalization for Person Re-Identification (FedDG-ReID) learns domain-invariant representations from decentralized data. While Vision Transformer (ViT) is widely adopted, its global attention often fails to distinguish pedestrians from high similarity backgrounds or diverse viewpoints -- a challenge amplified by cross-client distribution shifts in FedDG-ReID. To address this, we propose Federated Body Distribution Aware Visual Prompt (FedBPrompt), introducing learnable visual prompts to guide Transformer attention toward pedestrian-centric regions. FedBPrompt employs a Body Distribution Aware Visual Prompts Mechanism (BAPM) comprising: Holistic Full Body Prompts to suppress cross-client background noise, and Body Part Alignment Prompts to capture fine-grained details robust to pose and viewpoint variations. To mitigate high communication costs, we design a Prompt-based Fine-Tuning Strategy (PFTS) that freezes the ViT backbone and updates only lightweight prompts, significantly reducing communication overhead while maintaining adaptability. Extensive experiments demonstrate that BAPM effectively enhances feature discrimination and cross-domain generalization, while PFTS achieves notable performance gains within only a few aggregation rounds. Moreover, both BAPM and PFTS can be easily integrated into existing ViT-based FedDG-ReID frameworks, making FedBPrompt a flexible and effective solution for federated person re-identification. The code is available at this https URL .\n\n---\n\n## 49. Learning from Child-Directed Speech in Two-Language Scenarios: A French-English Case Study\n- **Authors**: Liel Binyamin , Elior Sulem\n- **URL**: [https://arxiv.org/abs/2603.12906](https://arxiv.org/abs/2603.12906)\n- **Abstract**:\n> Research on developmentally plausible language models has largely focused on English, leaving open questions about multilingual settings. We present a systematic study of compact language models by extending BabyBERTa to English-French scenarios under strictly size-matched data conditions, covering monolingual, bilingual, and cross-lingual settings. Our design contrasts two types of training corpora: (i) child-directed speech (about 2.5M tokens), following BabyBERTa and related work, and (ii) multi-domain corpora (about 10M tokens), extending the BabyLM framework to French. To enable fair evaluation, we also introduce new resources, including French versions of QAMR and QASRL, as well as English and French multi-domain corpora. We evaluate the models on both syntactic and semantic tasks and compare them with models trained on Wikipedia-only data. The results reveal context-dependent effects: training on Wikipedia consistently benefits semantic tasks, whereas child-directed speech improves grammatical judgments in monolingual settings. Bilingual pretraining yields notable gains for textual entailment, with particularly strong improvements for French. Importantly, similar patterns emerge across BabyBERTa, RoBERTa, and LTG-BERT, suggesting consistent trends across architectures.\n\n---\n\n## 50. Human-Centered Evaluation of an LLM-Based Process Modeling Copilot: A Mixed-Methods Study with Domain Experts\n- **Authors**: Chantale Lauer , Peter Pfeiffer , Nijat Mehdiyev\n- **URL**: [https://arxiv.org/abs/2603.12895](https://arxiv.org/abs/2603.12895)\n- **Abstract**:\n> Integrating Large Language Models (LLMs) into business process management tools promises to democratize Business Process Model and Notation (BPMN) modeling for non-experts. While automated frameworks assess syntactic and semantic quality, they miss human factors like trust, usability, and professional alignment. We conducted a mixed-methods evaluation of our proposed solution, an LLM-powered BPMN copilot, with five process modeling experts using focus groups and standardized questionnaires. Our findings reveal a critical tension between acceptable perceived usability (mean CUQ score: 67.2/100) and notably lower trust (mean score: 48.8\\%), with reliability rated as the most critical concern (M=1.8/5). Furthermore, we identified output-quality issues, prompting difficulties, and a need for the LLM to ask more in-depth clarifying questions about the process. We envision five use cases ranging from domain-expert support to enterprise quality assurance. We demonstrate the necessity of human-centered evaluation complementing automated benchmarking for LLM modeling agents.\n\n---\n\n## 51. Finite Difference Flow Optimization for RL Post-Training of Text-to-Image Models\n- **Authors**: David McAllister , Miika Aittala , Tero Karras , Janne Hellsten , Angjoo Kanazawa , Timo Aila , Samuli Laine\n- **URL**: [https://arxiv.org/abs/2603.12893](https://arxiv.org/abs/2603.12893)\n- **Abstract**:\n> Reinforcement learning (RL) has become a standard technique for post-training diffusion-based image synthesis models, as it enables learning from reward signals to explicitly improve desirable aspects such as image quality and prompt alignment. In this paper, we propose an online RL variant that reduces the variance in the model updates by sampling paired trajectories and pulling the flow velocity in the direction of the more favorable image. Unlike existing methods that treat each sampling step as a separate policy action, we consider the entire sampling process as a single action. We experiment with both high-quality vision language models and off-the-shelf quality metrics for rewards, and evaluate the outputs using a broad set of metrics. Our method converges faster and yields higher output quality and prompt alignment than previous approaches.\n\n---\n\n## 52. Team LEYA in 10th ABAW Competition: Multimodal Ambivalence/Hesitancy Recognition Approach\n- **Authors**: Elena Ryumina (1), Alexandr Axyonov (1), Dmitry Sysoev (2), Timur Abdulkadirov (2), Kirill Almetov (2), Yulia Morozova (2), Dmitry Ryumin (1 and 2) ((1) St. Petersburg Federal Research Center of the Russian Academy of Sciences, St. Petersburg, Russia, (2) HSE University, St. Petersburg, Russia)\n- **URL**: [https://arxiv.org/abs/2603.12848](https://arxiv.org/abs/2603.12848)\n- **Abstract**:\n> Ambivalence/hesitancy recognition in unconstrained videos is a challenging problem due to the subtle, multimodal, and context-dependent nature of this behavioral state. In this paper, a multimodal approach for video-level ambivalence/hesitancy recognition is presented for the 10th ABAW Competition. The proposed approach integrates four complementary modalities: scene, face, audio, and text. Scene dynamics are captured with a VideoMAE-based model, facial information is encoded through emotional frame-level embeddings aggregated by statistical pooling, acoustic representations are extracted with EmotionWav2Vec2.0 and processed by a Mamba-based temporal encoder, and linguistic cues are modeled using fine-tuned transformer-based text models. The resulting unimodal embeddings are further combined using multimodal fusion models, including prototype-augmented variants. Experiments on the BAH corpus demonstrate clear gains of multimodal fusion over all unimodal baselines. The best unimodal configuration achieved an average MF1 of 70.02%, whereas the best multimodal fusion model reached 83.25%. The highest final test performance, 71.43%, was obtained by an ensemble of five prototype-augmented fusion models. The obtained results highlight the importance of complementary multimodal cues and robust fusion strategies for ambivalence/hesitancy recognition.\n\n---\n\n## 53. Hierarchical Reference Sets for Robust Unsupervised Detection of Scattered and Clustered Outliers\n- **Authors**: Yiqun Zhang , Zexi Tan , Xiaopeng Luo , Yunlin Liu\n- **URL**: [https://arxiv.org/abs/2603.12847](https://arxiv.org/abs/2603.12847)\n- **Abstract**:\n> Most real-world IoT data analysis tasks, such as clustering and anomaly event detection, are unsupervised and highly susceptible to the presence of outliers. In addition to sporadic scattered outliers caused by factors such as faulty sensor readings, IoT systems often exhibit clustered outliers. These occur when multiple devices or nodes produce similar anomalous measurements, for instance, owing to localized interference, emerging security threats, or regional false alarms, forming micro-clusters. These clustered outliers can be easily mistaken for normal behavior because of their relatively high local density, thereby obscuring the detection of both scattered and contextual anomalies. To address this, we propose a novel outlier detection paradigm that leverages the natural neighboring relationships using graph structures. This facilitates multi-perspective anomaly evaluation by incorporating reference sets at both local and global scales derived from the graph. Our approach enables the effective recognition of scattered outliers without interference from clustered anomalies, whereas the graph structure simultaneously helps reflect and isolate clustered outlier groups. Extensive experiments, including comparative performance analysis, ablation studies, validation on downstream clustering tasks, and evaluation of hyperparameter sensitivity, demonstrate the efficacy of the proposed method. The source code is available at this https URL .\n\n---\n\n## 54. DAST: A Dual-Stream Voice Anonymization Attacker with Staged Training\n- **Authors**: Ridwan Arefeen , Xiaoxiao Miao , Rong Tong , Aik Beng Ng , Simon See , Timothy Liu\n- **URL**: [https://arxiv.org/abs/2603.12840](https://arxiv.org/abs/2603.12840)\n- **Abstract**:\n> Voice anonymization masks vocal traits while preserving linguistic content, which may still leak speaker-specific patterns. To assess and strengthen privacy evaluation, we propose a dual-stream attacker that fuses spectral and self-supervised learning features via parallel encoders with a three-stage training strategy. Stage I establishes foundational speaker-discriminative representations. Stage II leverages the shared identity-transformation characteristics of voice conversion and anonymization, exposing the model to diverse converted speech to build cross-system robustness. Stage III provides lightweight adaptation to target anonymized data. Results on the VoicePrivacy Attacker Challenge (VPAC) dataset demonstrate that Stage II is the primary driver of generalization, enabling strong attacking performance on unseen anonymization datasets. With Stage III, fine-tuning on only 10\\% of the target anonymization dataset surpasses current state-of-the-art attackers in terms of EER.\n\n---\n\n## 55. Mask2Flow-TSE: Two-Stage Target Speaker Extraction with Masking and Flow Matching\n- **Authors**: Junwon Moon , Hyunjin Choi , Hansol Park , Heeseung Kim , Kyuhong Shim\n- **URL**: [https://arxiv.org/abs/2603.12837](https://arxiv.org/abs/2603.12837)\n- **Abstract**:\n> Target speaker extraction (TSE) extracts the target speaker's voice from overlapping speech mixtures given a reference utterance. Existing approaches typically fall into two categories: discriminative and generative. Discriminative methods apply time-frequency masking for fast inference but often over-suppress the target signal, while generative methods synthesize high-quality speech at the cost of numerous iterative steps. We propose Mask2Flow-TSE, a two-stage framework combining the strengths of both paradigms. The first stage applies discriminative masking for coarse separation, and the second stage employs flow matching to refine the output toward target speech. Unlike generative approaches that synthesize speech from Gaussian noise, our method starts from the masked spectrogram, enabling high-quality reconstruction in a single inference step. Experiments show that Mask2Flow-TSE achieves comparable performance to existing generative TSE methods with approximately 85M parameters.\n\n---\n\n## 56. Hierarchical Dual-Change Collaborative Learning for UAV Scene Change Captioning\n- **Authors**: Fuhai Chen , Pengpeng Huang , Junwen Wu , Hehong Zhang , Shiping Wang , Xiaoguang Ma , Xuri Ge\n- **URL**: [https://arxiv.org/abs/2603.12832](https://arxiv.org/abs/2603.12832)\n- **Abstract**:\n> This paper proposes a novel task for UAV scene understanding - UAV Scene Change Captioning (UAV-SCC) - which aims to generate natural language descriptions of semantic changes in dynamic aerial imagery captured from a movable viewpoint. Unlike traditional change captioning that mainly describes differences between image pairs captured from a fixed camera viewpoint over time, UAV scene change captioning focuses on image-pair differences resulting from both temporal and spatial scene variations dynamically captured by a moving camera. The key challenge lies in understanding viewpoint-induced scene changes from UAV image pairs that share only partially overlapping scene content due to viewpoint shifts caused by camera rotation, while effectively exploiting the relative orientation between the two images. To this end, we propose a Hierarchical Dual-Change Collaborative Learning (HDC-CL) method for UAV scene change captioning. In particular, a novel transformer, \\emph{i.e.} Dynamic Adaptive Layout Transformer (DALT) is designed to adaptively model diverse spatial layouts of the image pair, where the interrelated features derived from the overlapping and non-overlapping regions are learned within the flexible and unified encoding layer. Furthermore, we propose a Hierarchical Cross-modal Orientation Consistency Calibration (HCM-OCC) method to enhance the model's sensitivity to viewpoint shift directions, enabling more accurate change captioning. To facilitate in-depth research on this task, we construct a new benchmark dataset, named UAV-SCC dataset, for UAV scene change captioning. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance on this task. The dataset and code will be publicly released upon acceptance of this paper.\n\n---\n\n## 57. Residual SODAP: Residual Self-Organizing Domain-Adaptive Prompting with Structural Knowledge Preservation for Continual Learning\n- **Authors**: Gyutae Oh , Jungwoo Bae , Jitae Shin\n- **URL**: [https://arxiv.org/abs/2603.12816](https://arxiv.org/abs/2603.12816)\n- **Abstract**:\n> Continual learning (CL) suffers from catastrophic forgetting, which is exacerbated in domain-incremental learning (DIL) where task identifiers are unavailable and storing past data is infeasible. While prompt-based CL (PCL) adapts representations with a frozen backbone, we observe that prompt-only improvements are often insufficient due to suboptimal prompt selection and classifier-level instability under domain shifts. We propose Residual SODAP, which jointly performs prompt-based representation adaptation and classifier-level knowledge preservation. Our framework combines $\\alpha$-entmax sparse prompt selection with residual aggregation, data-free distillation with pseudo-feature replay, prompt-usage--based drift detection, and uncertainty-aware multi-loss balancing. Across three DIL benchmarks without task IDs or extra data storage, Residual SODAP achieves state-of-the-art AvgACC/AvgF of 0.850/0.047 (DR), 0.760/0.031 (Skin Cancer), and 0.995/0.003 (CORe50).\n\n---\n\n## 58. Cheers: Decoupling Patch Details from Semantic Representations Enables Unified Multimodal Comprehension and Generation\n- **Authors**: Yichen Zhang , Da Peng , Zonghao Guo , Zijian Zhang , Xuesong Yang , Tong Sun , Shichu Sun , Yidan Zhang , Yanghao Li , Haiyan Zhao , Wang Xu , Qi Shi , Yangang Sun , Chi Chen , Shuo Wang , Yukun Yan , Xu Han , Qiang Ma , Wei Ke , Liang Wang , Zhiyuan Liu , Maosong Sun\n- **URL**: [https://arxiv.org/abs/2603.12793](https://arxiv.org/abs/2603.12793)\n- **Abstract**:\n> A recent cutting-edge topic in multimodal modeling is to unify visual comprehension and generation within a single model. However, the two tasks demand mismatched decoding regimes and visual representations, making it non-trivial to jointly optimize within a shared feature space. In this work, we present Cheers, a unified multimodal model that decouples patch-level details from semantic representations, thereby stabilizing semantics for multimodal understanding and improving fidelity for image generation via gated detail residuals. Cheers includes three key components: (i) a unified vision tokenizer that encodes and compresses image latent states into semantic tokens for efficient LLM conditioning, (ii) an LLM-based Transformer that unifies autoregressive decoding for text generation and diffusion decoding for image generation, and (iii) a cascaded flow matching head that decodes visual semantics first and then injects semantically gated detail residuals from the vision tokenizer to refine high-frequency content. Experiments on popular benchmarks demonstrate that Cheers matches or surpasses advanced UMMs in both visual understanding and generation. Cheers also achieves 4x token compression, enabling more efficient high-resolution image encoding and generation. Notably, Cheers outperforms the Tar-1.5B on the popular benchmarks GenEval and MMBench, while requiring only 20% of the training cost, indicating effective and efficient (i.e., 4x token compression) unified multimodal modeling. We will release all code and data for future research.\n\n---\n\n## 59. The RIGID Framework: Research-Integrated, Generative AI-Mediated Instructional Design\n- **Authors**: Yerin Kwak , Zachary A. Pardos\n- **URL**: [https://arxiv.org/abs/2603.12781](https://arxiv.org/abs/2603.12781)\n- **Abstract**:\n> Instructional Design (ID) often faces challenges in incorporating research-based knowledge and pedagogical best practices. Although educational researchers and government agencies emphasize grounding ID in evidence, integrating research findings into everyday design workflows is often complex, as it requires considering multiple context-specific demands and constraints. To address this persistent gap, this paper explores how research in the learning sciences (LS) can be systematically integrated across ID workflows and how recent advances in generative AI can help operationalize this integration. While ID and LS share a commitment to improving learning experiences through design-oriented approaches in authentic contexts, structured integration between the two fields remains limited, leaving their complementary insights underutilized. We present RIGID (Research-Integrated, Generative AI-Mediated Instructional Design), a unified framework that integrates LS research across ID workflows spanning analysis, design, implementation, and evaluation phases, while leveraging generative AI to mediate this integration at each stage. The RIGID framework provides a systematic approach for enabling research-integrated instructional design that is both operational and context-sensitive, while preserving the central role of human expertise.\n\n---\n\n## 60. Empowering Semantic-Sensitive Underwater Image Enhancement with VLM\n- **Authors**: Guodong Fan , Shengning Zhou , Genji Yuan , Huiyu Li , Jingchun Zhou , Jinjiang Li\n- **URL**: [https://arxiv.org/abs/2603.12773](https://arxiv.org/abs/2603.12773)\n- **Abstract**:\n> In recent years, learning-based underwater image enhancement (UIE) techniques have rapidly evolved. However, distribution shifts between high-quality enhanced outputs and natural images can hinder semantic cue extraction for downstream vision tasks, thereby limiting the adaptability of existing enhancement models. To address this challenge, this work proposes a new learning mechanism that leverages Vision-Language Models (VLMs) to empower UIE models with semantic-sensitive capabilities. To be concrete, our strategy first generates textual descriptions of key objects from a degraded image via VLMs. Subsequently, a text-image alignment model remaps these relevant descriptions back onto the image to produce a spatial semantic guidance map. This map then steers the UIE network through a dual-guidance mechanism, which combines cross-attention and an explicit alignment loss. This forces the network to focus its restorative power on semantic-sensitive regions during image reconstruction, rather than pursuing a globally uniform improvement, thereby ensuring the faithful restoration of key object features. Experiments confirm that when our strategy is applied to different UIE baselines, significantly boosts their performance on perceptual quality metrics as well as enhances their performance on detection and segmentation tasks, validating its effectiveness and adaptability.\n\n---\n\n## 61. FC-Track: Overlap-Aware Post-Association Correction for Online Multi-Object Tracking\n- **Authors**: Cheng Ju , Zejing Zhao , Akio Namiki\n- **URL**: [https://arxiv.org/abs/2603.12758](https://arxiv.org/abs/2603.12758)\n- **Abstract**:\n> Reliable multi-object tracking (MOT) is essential for robotic systems operating in complex and dynamic environments. Despite recent advances in detection and association, online MOT methods remain vulnerable to identity switches caused by frequent occlusions and object overlap, where incorrect associations can propagate over time and degrade tracking reliability. We present a lightweight post-association correction framework (FC-Track) for online MOT that explicitly targets overlap-induced mismatches during inference. The proposed method suppresses unreliable appearance updates under high-overlap conditions using an Intersection over Area (IoA)-based filtering strategy, and locally corrects detection-to-tracklet mismatches through appearance similarity comparison within overlapped tracklet pairs. By preventing short-term mismatches from propagating, our framework effectively mitigates long-term identity switches without resorting to global optimization or re-identification. The framework operates online without global optimization or re-identification, making it suitable for real-time robotic applications. We achieve 81.73 MOTA, 82.81 IDF1, and 66.95 HOTA on the MOT17 test set with a running speed of 5.7 FPS, and 77.52 MOTA, 80.90 IDF1, and 65.67 HOTA on the MOT20 test set with a running speed of 0.6 FPS. Specifically, our framework FC-Track produces only 29.55% long-term identity switches, which is substantially lower than existing online trackers. Meanwhile, our framework maintains state-of-the-art performance on the MOT20 benchmark.\n\n---\n\n## 62. TaoBench: Do Automated Theorem Prover LLMs Generalize Beyond MathLib?\n- **Authors**: Alexander K Taylor , Junyi Zhang , Ethan Ji , Vigyan Sahai , Haikang Deng , Yuanzhou Chen , Yifan Yuan , Di Wu , Jia-Chen Gu , Kai-Wei Chang , Nanyun Peng , Amit Sahai , Wei Wang\n- **URL**: [https://arxiv.org/abs/2603.12744](https://arxiv.org/abs/2603.12744)\n- **Abstract**:\n> Automated theorem proving (ATP) benchmarks largely consist of problems formalized in MathLib, so current ATP training and evaluation are heavily biased toward MathLib's definitional framework. However, frontier mathematics is often exploratory and prototype-heavy, relying on bespoke constructions that deviate from standard libraries. In this work, we evaluate the robustness of current ATP systems when applied to a novel definitional framework, specifically examining the performance gap between standard library problems and bespoke mathematical constructions. We introduce TaoBench, an undergraduate-level benchmark derived from Terence Tao's Analysis I, which formalizes analysis by constructing core mathematical concepts from scratch, without relying on standard Mathlib definitions, as well as by mixing from-scratch and MathLib constructions. For fair evaluation, we build an agentic pipeline that automatically extracts a compilable, self-contained local environment for each problem. To isolate the effect of definitional frameworks, we additionally translate every problem into a mathematically equivalent Mathlib formulation, yielding paired TaoBench-Mathlib statements for direct comparison. While state-of-the-art ATP models perform capably within the MathLib framework, performance drops by an average of roughly 26% on the definitionally equivalent Tao formulation. This indicates that the main bottleneck is limited generalization across definitional frameworks rather than task difficulty. TaoBench thus highlights a gap between benchmark performance and applicability, and provides a concrete foundation for developing and testing provers better aligned with research mathematics.\n\n---\n\n## 63. MoKus: Leveraging Cross-Modal Knowledge Transfer for Knowledge-Aware Concept Customization\n- **Authors**: Chenyang Zhu , Hongxiang Li , Xiu Li , Long Chen\n- **URL**: [https://arxiv.org/abs/2603.12743](https://arxiv.org/abs/2603.12743)\n- **Abstract**:\n> Concept customization typically binds rare tokens to a target concept. Unfortunately, these approaches often suffer from unstable performance as the pretraining data seldom contains these rare tokens. Meanwhile, these rare tokens fail to convey the inherent knowledge of the target concept. Consequently, we introduce Knowledge-aware Concept Customization, a novel task aiming at binding diverse textual knowledge to target visual concepts. This task requires the model to identify the knowledge within the text prompt to perform high-fidelity customized generation. Meanwhile, the model should efficiently bind all the textual knowledge to the target concept. Therefore, we propose MoKus, a novel framework for knowledge-aware concept customization. Our framework relies on a key observation: cross-modal knowledge transfer, where modifying knowledge within the text modality naturally transfers to the visual modality during generation. Inspired by this observation, MoKus contains two stages: (1) In visual concept learning, we first learn the anchor representation to store the visual information of the target concept. (2) In textual knowledge updating, we update the answer for the knowledge queries to the anchor representation, enabling high-fidelity customized generation. To further comprehensively evaluate our proposed MoKus on the new task, we introduce the first benchmark for knowledge-aware concept customization: KnowCusBench. Extensive evaluations have demonstrated that MoKus outperforms state-of-the-art methods. Moreover, the cross-model knowledge transfer allows MoKus to be easily extended to other knowledge-aware applications like virtual concept creation and concept erasure. We also demonstrate the capability of our method to achieve improvements on world knowledge benchmarks.\n\n---\n\n## 64. SRAM-Based Compute-in-Memory Accelerator for Linear-decay Spiking Neural Networks\n- **Authors**: Hongyang Shang , Shuai Dong , Yahan Yang , Junyi Yang , Peng Zhou , Arindam Basu\n- **URL**: [https://arxiv.org/abs/2603.12739](https://arxiv.org/abs/2603.12739)\n- **Abstract**:\n> Spiking Neural Networks (SNNs) have emerged as a biologically inspired alternative to conventional deep networks, offering event-driven and energy-efficient computation. However, their throughput remains constrained by the serial update of neuron membrane states. While many hardware accelerators and Compute-in-Memory (CIM) architectures efficiently parallelize the synaptic operation (W x I) achieving O(1) complexity for matrix-vector multiplication, the subsequent state update step still requires O(N) time to refresh all neuron membrane potentials. This mismatch makes state update the dominant latency and energy bottleneck in SNN inference. To address this challenge, we propose an SRAM-based CIM for SNN with Linear Decay Leaky Integrate-and-Fire (LD-LIF) Neuron that co-optimizes algorithm and hardware. At the algorithmic level, we replace the conventional exponential membrane decay with a linear decay approximation, converting costly multiplications into simple additions while accuracy drops only around 1%. At the architectural level, we introduce an in-memory parallel update scheme that performs in-place decay directly within the SRAM array, eliminating the need for global sequential updates. Evaluated on benchmark SNN workloads, the proposed method achieves a 1.1 x to 16.7 x reduction of SOP energy consumption, while providing 15.9 x to 69 x more energy efficiency, with negligible accuracy loss relative to original decay models. This work highlights that beyond accelerating the (W x I) computation, optimizing state-update dynamics within CIM architectures is essential for scalable, low-power, and real-time neuromorphic processing.\n\n---\n\n## 65. Graph In-Context Operator Networks for Generalizable Spatiotemporal Prediction\n- **Authors**: Chenghan Wu , Zongmin Yu , Boai Sun , Liu Yang\n- **URL**: [https://arxiv.org/abs/2603.12725](https://arxiv.org/abs/2603.12725)\n- **Abstract**:\n> In-context operator learning enables neural networks to infer solution operators from contextual examples without weight updates. While prior work has demonstrated the effectiveness of this paradigm in leveraging vast datasets, a systematic comparison against single-operator learning using identical training data has been absent. We address this gap through controlled experiments comparing in-context operator learning against classical operator learning (single-operator models trained without contextual examples), under the same training steps and dataset. To enable this investigation on real-world spatiotemporal systems, we propose GICON (Graph In-Context Operator Network), combining graph message passing for geometric generalization with example-aware positional encoding for cardinality generalization. Experiments on air quality prediction across two Chinese regions show that in-context operator learning outperforms classical operator learning on complex tasks, generalizing across spatial domains and scaling robustly from few training examples to 100 at inference.\n\n---\n\n## 66. CognitionCapturerPro: Towards High-Fidelity Visual Decoding from EEG/MEG via Multi-modal Information and Asymmetric Alignment\n- **Authors**: Kaifan Zhang , Lihuo He , Junjie Ke , Yuqi Ji , Lukun Wu , Lizi Wang , Xinbo Gao\n- **URL**: [https://arxiv.org/abs/2603.12722](https://arxiv.org/abs/2603.12722)\n- **Abstract**:\n> Visual stimuli reconstruction from EEG remains challenging due to fidelity loss and representation shift. We propose CognitionCapturerPro, an enhanced framework that integrates EEG with multi-modal priors (images, text, depth, and edges) via collaborative training. Our core contributions include an uncertainty-weighted similarity scoring mechanism to quantify modality-specific fidelity and a fusion encoder for integrating shared representations. By employing a simplified alignment module and a pre-trained diffusion model, our method significantly outperforms the original CognitionCapturer on the THINGS-EEG dataset, improving Top-1 and Top-5 retrieval accuracy by 25.9% and 10.6%, respectively. Code is available at: this https URL .\n\n---\n\n## 67. CMHANet: A Cross-Modal Hybrid Attention Network for Point Cloud Registration\n- **Authors**: Dongxu Zhang , Yingsen Wang , Yiding Sun , Haoran Xu , Peilin Fan , Jihua Zhu\n- **URL**: [https://arxiv.org/abs/2603.12721](https://arxiv.org/abs/2603.12721)\n- **Abstract**:\n> Robust point cloud registration is a fundamental task in 3D computer vision and geometric deep learning, essential for applications such as large-scale 3D reconstruction, augmented reality, and scene understanding. However, the performance of established learning-based methods often degrades in complex, real world scenarios characterized by incomplete data, sensor noise, and low overlap regions. To address these limitations, we propose CMHANet, a novel Cross-Modal Hybrid Attention Network. Our method integrates the fusion of rich contextual information from 2D images with the geometric detail of 3D point clouds, yielding a comprehensive and resilient feature representation. Furthermore, we introduce an innovative optimization function based on contrastive learning, which enforces geometric consistency and significantly improves the model's robustness to noise and partial observations. We evaluated CMHANet on the 3DMatch and the challenging 3DLoMatch datasets. \\rev{Additionally, zero-shot evaluations on the TUM RGB-D SLAM dataset verify the model's generalization capability to unseen domains.} The experimental results demonstrate that our method achieves substantial improvements in both registration accuracy and overall robustness, outperforming current techniques. We also release our code in \\href{ this https URL }{ this https URL }.\n\n---\n\n## 68. IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registration\n- **Authors**: Dongxu Zhang , Jihua Zhu , Shiqi Li , Wenbiao Yan , Haoran Xu , Peilin Fan , Huimin Lu\n- **URL**: [https://arxiv.org/abs/2603.12719](https://arxiv.org/abs/2603.12719)\n- **Abstract**:\n> Point cloud registration (PCR) is a fundamental task in 3D vision and provides essential support for applications such as autonomous driving, robotics, and environmental modeling. Despite its widespread use, existing methods often fail when facing real-world challenges like heavy noise, significant occlusions, and large-scale transformations. These limitations frequently result in compromised registration accuracy and insufficient robustness in complex environments. In this paper, we propose IGASA as a novel registration framework constructed upon a Hierarchical Pyramid Architecture (HPA) designed for robust multi-scale feature extraction and fusion. The framework integrates two pivotal components consisting of the Hierarchical Cross-Layer Attention (HCLA) module and the Iterative Geometry-Aware Refinement (IGAR) module. The HCLA module utilizes skip attention mechanisms to align multi-resolution features and enhance local geometric consistency. Simultaneously, the IGAR module is designed for the fine matching phase by leveraging reliable correspondences established during coarse matching. This synergistic integration within the architecture allows IGASA to adapt effectively to diverse point cloud structures and intricate transformations. We evaluate the performance of IGASA on four widely recognized benchmark datasets including 3D(Lo)Match, KITTI, and nuScenes. Our extensive experiments consistently demonstrate that IGASA significantly surpasses state-of-the-art methods and achieves notable improvements in registration accuracy. This work provides a robust foundation for advancing point cloud registration techniques while offering valuable insights for practical 3D vision applications. The code for IGASA is available in \\href{ this https URL }{ this https URL }.\n\n---\n\n## 69. Altered Thoughts, Altered Actions: Probing Chain-of-Thought Vulnerabilities in VLA Robotic Manipulation\n- **Authors**: Tuan Duong Trinh , Naveed Akhtar , Basim Azam\n- **URL**: [https://arxiv.org/abs/2603.12717](https://arxiv.org/abs/2603.12717)\n- **Abstract**:\n> Recent Vision-Language-Action (VLA) models increasingly adopt chain-of-thought (CoT) reasoning, generating a natural-language plan before decoding motor commands. This internal text channel between the reasoning module and the action decoder has received no adversarial scrutiny. We ask: which properties of this intermediate plan does the action decoder actually rely on, and can targeted corruption of the reasoning trace alone -- with all inputs left intact -- degrade a robot's physical task performance? We design a taxonomy of seven text corruptions organized into three attacker tiers (blind noise, mechanical-semantic, and LLM-adaptive) and apply them to a state-of-the-art reasoning VLA across 40 LIBERO tabletop manipulation tasks. Our results reveal a striking asymmetry: substituting object names in the reasoning trace reduces overall success rate by 8.3~percentage points (pp) -- reaching $-$19.3~pp on goal-conditioned tasks and $-$45~pp on individual tasks -- whereas sentence reordering, spatial-direction reversal, token noise, and even a 70B-parameter LLM crafting plausible-but-wrong plans all have negligible impact (within $\\pm$4~pp). This asymmetry indicates that the action decoder depends on entity-reference integrity rather than reasoning quality or sequential structure. Notably, a sophisticated LLM-based attacker underperforms simple mechanical object-name substitution, because preserving plausibility inadvertently retains the entity-grounding structure the decoder needs. A cross-architecture control using a non-reasoning VLA confirms the vulnerability is exclusive to reasoning-augmented models, while instruction-level attacks degrade both architectures -- establishing that the internal reasoning trace is a distinct and stealthy threat vector invisible to input-validation defenses.\n\n---\n\n## 70. Cost-Efficient Multimodal LLM Inference via Cross-Tier GPU Heterogeneity\n- **Authors**: Donglin Yu\n- **URL**: [https://arxiv.org/abs/2603.12707](https://arxiv.org/abs/2603.12707)\n- **Abstract**:\n> Multimodal large language model (MLLM) inference splits into two phases with opposing hardware demands: vision encoding is compute-bound, while language generation is memory-bandwidth-bound. We show that under standard transformer KV caching, the modality boundary (between vision encoder and language model) minimizes cross-device transfer among all partition points that preserve standard stage-based execution. Partitioning here reduces transfer complexity from $O(L * s_ctx)$ bytes (GB-scale KV caches under stage-level disaggregation) to $O(N_v * d)$ bytes (MB-scale embeddings), an O(L) reduction where L is the transformer depth. The result holds across attention mechanisms (MHA/GQA), dynamic vision resolutions, and model scales, and the advantage grows as models deepen. A direct implication is that existing stage-level disaggregation systems are constrained to high-bandwidth interconnects (e.g., NVLink), whereas modality-level disaggregation enables cross-tier heterogeneous serving over commodity PCIe. A closed-form cost model shows that heterogeneous deployment is cost-optimal under phase-separable workloads (predicts 31.4% savings; observed 40.6%). We build HeteroServe, a phase-aware runtime with modality-level partitioning and cross-tier scheduling, and evaluate it on LLaVA-1.5-7B and Qwen2.5-VL against vLLM v0.3.0. On identical 4xA100 hardware, engine optimizations raise throughput by up to 54%. Under a fixed budget, a heterogeneous cluster (\\$38k) improves Tokens/\\$ by 37% over a homogeneous baseline (\\$64k) without degrading latency.\n\n---\n\n## 71. Seeing Eye to Eye: Enabling Cognitive Alignment Through Shared First-Person Perspective in Human-AI Collaboration\n- **Authors**: Zhuyu Teng , Pei Chen , Yichen Cai , Ruoqing Lu , Zhaoqu Jiang , Jiayang Li , Weitao You , Lingyun Sun\n- **URL**: [https://arxiv.org/abs/2603.12701](https://arxiv.org/abs/2603.12701)\n- **Abstract**:\n> Despite advances in multimodal AI, current vision-based assistants often remain inefficient in collaborative tasks. We identify two key gulfs: a communication gulf, where users must translate rich parallel intentions into verbal commands due to the channel mismatch , and an understanding gulf, where AI struggles to interpret subtle embodied cues. To address these, we propose Eye2Eye, a framework that leverages first-person perspective as a channel for human-AI cognitive alignment. It integrates three components: (1) joint attention coordination for fluid focus alignment, (2) revisable memory to maintain evolving common ground, and (3) reflective feedback allowing users to clarify and refine AI's understanding. We implement this framework in an AR prototype and evaluate it through a user study and a post-hoc pipeline evaluation. Results show that Eye2Eye significantly reduces task completion time and interaction load while increasing trust, demonstrating its components work in concert to improve collaboration.\n\n---\n\n## 72. HSEmotion Team at ABAW-10 Competition: Facial Expression Recognition, Valence-Arousal Estimation, Action Unit Detection and Fine-Grained Violence Classification\n- **Authors**: Andrey V. Savchenko , Kseniia Tsypliakova\n- **URL**: [https://arxiv.org/abs/2603.12693](https://arxiv.org/abs/2603.12693)\n- **Abstract**:\n> This article presents our results for the 10th Affective Behavior Analysis in-the-Wild (ABAW) competition. For frame-wise facial emotion understanding tasks (frame-wise facial expression recognition, valence-arousal estimation, action unit detection), we propose a fast approach based on facial embedding extraction with pre-trained EfficientNet-based emotion recognition models. If the latter model's confidence exceeds a threshold, its prediction is used. Otherwise, we feed embeddings into a simple multi-layered perceptron trained on the AffWild2 dataset. Estimated class-level scores are smoothed in a sliding window of fixed size to mitigate noise in frame-wise predictions. For the fine-grained violence detection task, we examine several pre-trained architectures for frame embeddings and their aggregation for video classification. Experimental results on four tasks from the ABAW challenge demonstrate that our approach significantly improves validation metrics over existing baselines.\n\n---\n\n## 73. Federated Hierarchical Clustering with Automatic Selection of Optimal Cluster Numbers\n- **Authors**: Yue Zhang , Chuanlong Qiu , Xinfa Liao , Yiqun Zhang\n- **URL**: [https://arxiv.org/abs/2603.12684](https://arxiv.org/abs/2603.12684)\n- **Abstract**:\n> Federated Clustering (FC) is an emerging and promising solution in exploring data distribution patterns from distributed and privacy-protected data in an unsupervised manner. Existing FC methods implicitly rely on the assumption that clients are with a known number of uniformly sized clusters. However, the true number of clusters is typically unknown, and cluster sizes are naturally imbalanced in real scenarios. Furthermore, the privacy-preserving transmission constraints in federated learning inevitably reduce usable information, making the development of robust and accurate FC extremely challenging. Accordingly, we propose a novel FC framework named Fed-$k^*$-HC, which can automatically determine an optimal number of clusters $k^*$ based on the data distribution explored through hierarchical clustering. To obtain the global data distribution for $k^*$ determination, we let each client generate micro-subclusters. Their prototypes are then uploaded to the server for hierarchical merging. The density-based merging design allows exploring clusters of varying sizes and shapes, and the progressive merging process can self-terminate according to the neighboring relationships among the prototypes to determine $k^*$. Extensive experiments on diverse datasets demonstrate the FC capability of the proposed Fed-$k^*$-HC in accurately exploring a proper number of clusters.\n\n---\n\n## 74. Experimental evidence of progressive ChatGPT models self-convergence\n- **Authors**: Konstantinos F. Xylogiannopoulos , Petros Xanthopoulos , Panagiotis Karampelas , Georgios A. Bakamitsos\n- **URL**: [https://arxiv.org/abs/2603.12683](https://arxiv.org/abs/2603.12683)\n- **Abstract**:\n> Large Language Models (LLMs) that undergo recursive training on synthetically generated data are susceptible to model collapse, a phenomenon marked by the generation of meaningless output. Existing research has examined this issue from either theoretical or empirical perspectives, often focusing on a single model trained recursively on its own outputs. While prior studies have cautioned against the potential degradation of LLM output quality under such conditions, no longitudinal investigation has yet been conducted to assess this effect over time. In this study, we employ a text similarity metric to evaluate different ChatGPT models' capacity to generate diverse textual outputs. Our findings indicate a measurable decline of recent ChatGPT releases' ability to produce varied text, even when explicitly prompted to do so, by setting the temperature parameter to one. The observed reduction in output diversity may be attributed to the influence of the amounts of synthetic data incorporated within their training datasets as the result of internet infiltration by LLM generated data. The phenomenon is defined as model self-convergence because of the gradual increase of similarities of produced texts among different ChatGPT versions.\n\n---\n\n## 75. MetaKE: Meta-learning Aligned Knowledge Editing via Bi-level Optimization\n- **Authors**: Shuxin Liu , Ou Wu\n- **URL**: [https://arxiv.org/abs/2603.12677](https://arxiv.org/abs/2603.12677)\n- **Abstract**:\n> Knowledge editing (KE) aims to precisely rectify specific knowledge in Large Language Models (LLMs) without disrupting general capabilities. State-of-the-art methods suffer from an open-loop control mismatch. We identify a critical \"Semantic-Execution Disconnect\": the semantic target is derived independently without feedback from the downstream's feasible region. This misalignment often causes valid semantic targets to fall within the prohibited space, resulting in gradient truncation and editing failure. To bridge this gap, we propose MetaKE (Meta-learning Aligned Knowledge Editing), a new framework that reframes KE as a bi-level optimization problem. Departing from static calculation, MetaKE treats the edit target as a learnable meta-parameter: the upper-level optimizer seeks a feasible target to maximize post-edit performance, while the lower-level solver executes the editing. To address the challenge of differentiating through complex solvers, we derive a Structural Gradient Proxy, which explicitly backpropagates editability constraints to the target learning phase. Theoretical analysis demonstrates that MetaKE automatically aligns the edit direction with the model's feasible manifold. Extensive experiments confirm that MetaKE significantly outperforms strong baselines, offering a new perspective on knowledge editing.\n\n---\n\n## 76. Marker-Based 3D Reconstruction of Aggregates with a Comparative Analysis of 2D and 3D Morphologies\n- **Authors**: Haohang Huang , Jiayi Luo , Issam Qamhia , Erol Tutumluer , John M. Hart , Andrew J. Stolba\n- **URL**: [https://arxiv.org/abs/2603.12667](https://arxiv.org/abs/2603.12667)\n- **Abstract**:\n> Aggregates, serving as the main skeleton in assemblies of construction materials, are important functional components in various building and transportation infrastructures. They can be used in unbound layer applications, e.g. pavement base and railroad ballast, bound applications of cement concrete and asphalt concrete, and as riprap and large-sized primary crushed rocks. Information on the size and shape or morphology of aggregates can greatly facilitate the Quality Assurance/Quality Control (QA/QC) process by providing insights of aggregate behavior during composition and packing. A full 3D characterization of aggregate particle morphology is difficult both during production in a quarry and at a construction site. Many aggregate imaging approaches have been developed to quantify the particle morphology by computer vision, including 2D image-based approaches that analyze particle silhouettes and 3D scanning-based methods that require expensive devices such as 3D laser scanners or X-Ray Computed Tomography (CT) equipment. This paper presents a flexible and cost-effective photogrammetry-based approach for the 3D reconstruction of aggregate particles. The proposed approach follows a marker-based design that enables background suppression, point cloud stitching, and scale referencing to obtain high-quality aggregate models. The accuracy of the reconstruction results was validated against ground-truth for selected aggregate samples. Comparative analyses were conducted on 2D and 3D morphological properties of the selected samples. Significant differences were found between the 2D and 3D statistics. Based on the presented approach, 3D shape information of aggregates can be obtained easily and at a low cost, thus allowing convenient aggregate inspection, data collection, and 3D morphological analysis.\n\n---\n\n## 77. RetroReasoner: A Reasoning LLM for Strategic Retrosynthesis Prediction\n- **Authors**: Hanbum Ko , Chanhui Lee , Ye Rin Kim , Rodrigo Hormazabal , Sehui Han , Sungbin Lim , Sungwoong Kim\n- **URL**: [https://arxiv.org/abs/2603.12666](https://arxiv.org/abs/2603.12666)\n- **Abstract**:\n> Retrosynthesis prediction is a core task in organic synthesis that aims to predict reactants for a given product molecule. Traditionally, chemists select a plausible bond disconnection and derive corresponding reactants, which is time-consuming and requires substantial expertise. While recent advancements in molecular large language models (LLMs) have made progress, many methods either predict reactants without strategic reasoning or conduct only a generic product analysis, rather than reason explicitly about bond-disconnection strategies that logically lead to the choice of specific reactants. To overcome these limitations, we propose RetroReasoner, a retrosynthetic reasoning model that leverages chemists' strategic thinking. RetroReasoner is trained using both supervised fine-tuning (SFT) and reinforcement learning (RL). For SFT, we introduce SyntheticRetro, a framework that generates structured disconnection rationales alongside reactant predictions. In the case of RL, we apply a round-trip accuracy as reward, where predicted reactants are passed through a forward synthesis model, and predictions are rewarded when the forward-predicted product matches the original input product. Experimental results show that RetroReasoner not only outperforms prior baselines but also generates a broader range of feasible reactant proposals, particularly in handling more challenging reaction instances.\n\n---\n\n## 78. From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space\n- **Authors**: Lehui Li , Yuyao Wang , Jisheng Yan , Wei Zhang , Jinliang Deng , Haoliang Sun , Zhongyi Han , Yongshun Gong\n- **URL**: [https://arxiv.org/abs/2603.12664](https://arxiv.org/abs/2603.12664)\n- **Abstract**:\n> Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.\n\n---\n\n## 79. Continual Learning in Large Language Models: Methods, Challenges, and Opportunities\n- **Authors**: Hongyang Chen , Zhongwu Sun , Hongfei Ye , Kunchi Li , Xuemin Lin\n- **URL**: [https://arxiv.org/abs/2603.12658](https://arxiv.org/abs/2603.12658)\n- **Abstract**:\n> Continual learning (CL) has emerged as a pivotal paradigm to enable large language models (LLMs) to dynamically adapt to evolving knowledge and sequential tasks while mitigating catastrophic forgetting-a critical limitation of the static pre-training paradigm inherent to modern LLMs. This survey presents a comprehensive overview of CL methodologies tailored for LLMs, structured around three core training stages: continual pre-training, continual fine-tuning, and continual this http URL the canonical taxonomy of rehearsal-, regularization-, and architecture-based methods, we further subdivide each category by its distinct forgetting mitigation mechanisms and conduct a rigorous comparative analysis of the adaptability and critical improvements of traditional CL methods for LLMs. In doing so, we explicitly highlight core distinctions between LLM CL and traditional machine learning, particularly with respect to scale, parameter efficiency, and emergent capabilities. Our analysis covers essential evaluation metrics, including forgetting rates and knowledge transfer efficiency, along with emerging benchmarks for assessing CL performance. This survey reveals that while current methods demonstrate promising results in specific domains, fundamental challenges persist in achieving seamless knowledge integration across diverse tasks and temporal scales. This systematic review contributes to the growing body of knowledge on LLM adaptation, providing researchers and practitioners with a structured framework for understanding current achievements and future opportunities in lifelong learning for language models.\n\n---\n\n## 80. LR-SGS: Robust LiDAR-Reflectance-Guided Salient Gaussian Splatting for Self-Driving Scene Reconstruction\n- **Authors**: Ziyu Chen , Fan Zhu , Hui Zhu , Deyi Kong , Xinkai Kuang , Yujia Zhang , Chunmao Jiang\n- **URL**: [https://arxiv.org/abs/2603.12647](https://arxiv.org/abs/2603.12647)\n- **Abstract**:\n> Recent 3D Gaussian Splatting (3DGS) methods have demonstrated the feasibility of self-driving scene reconstruction and novel view synthesis. However, most existing methods either rely solely on cameras or use LiDAR only for Gaussian initialization or depth supervision, while the rich scene information contained in point clouds, such as reflectance, and the complementarity between LiDAR and RGB have not been fully exploited, leading to degradation in challenging self-driving scenes, such as those with high ego-motion and complex lighting. To address these issues, we propose a robust and efficient LiDAR-reflectance-guided Salient Gaussian Splatting method (LR-SGS) for self-driving scenes, which introduces a structure-aware Salient Gaussian representation, initialized from geometric and reflectance feature points extracted from LiDAR and refined through a salient transform and improved density control to capture edge and planar structures. Furthermore, we calibrate LiDAR intensity into reflectance and attach it to each Gaussian as a lighting-invariant material channel, jointly aligned with RGB to enforce boundary consistency. Extensive experiments on the Waymo Open Dataset demonstrate that LR-SGS achieves superior reconstruction performance with fewer Gaussians and shorter training time. In particular, on Complex Lighting scenes, our method surpasses OmniRe by 1.18 dB PSNR.\n\n---\n\n## 81. LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing\n- **Authors**: Jiawei Hao , Zhiwei Hao , Jianyuan Guo , Li Shen , Yong Luo , Han Hu , Dan Zeng\n- **URL**: [https://arxiv.org/abs/2603.12645](https://arxiv.org/abs/2603.12645)\n- **Abstract**:\n> Mixture-of-Experts (MoE) based Large Language Models (LLMs) have demonstrated impressive performance and computational efficiency. However, their deployment is often constrained by substantial memory demands, primarily due to the need to load numerous expert modules. While existing expert compression techniques like pruning or merging attempt to mitigate this, they often suffer from irreversible knowledge loss or high training overhead. In this paper, we propose a novel expert compression paradigm termed expert replacing, which replaces redundant experts with parameter-efficient modules and recovers their capabilities with low training costs. We find that even a straightforward baseline of this paradigm yields promising performance. Building on this foundation, we introduce LightMoE, a framework that enhances the paradigm by introducing adaptive expert selection, hierarchical expert construction, and an annealed recovery strategy. Experimental results show that LightMoE matches the performance of LoRA fine-tuning at a 30% compression ratio. Even under a more aggressive 50% compression rate, it outperforms existing methods and achieves average performance improvements of 5.6% across five diverse tasks. These findings demonstrate that LightMoE strikes a superior balance among memory efficiency, training efficiency, and model performance.\n\n---\n\n## 82. Spend Less, Reason Better: Budget-Aware Value Tree Search for LLM Agents\n- **Authors**: Yushu Li , Wenlong Deng , Jiajin Li , Xiaoxiao Li\n- **URL**: [https://arxiv.org/abs/2603.12634](https://arxiv.org/abs/2603.12634)\n- **Abstract**:\n> Test-time scaling has become a dominant paradigm for improving LLM agent reliability, yet current approaches treat compute as an abundant resource, allowing agents to exhaust token and tool budgets on redundant steps or dead-end trajectories. Existing budget-aware methods either require expensive fine-tuning or rely on coarse, trajectory-level heuristics that cannot intervene mid-execution. We propose the Budget-Aware Value Tree (BAVT), a training-free inference-time framework that models multi-hop reasoning as a dynamic search tree guided by step-level value estimation within a single LLM backbone. Another key innovation is a budget-conditioned node selection mechanism that uses the remaining resource ratio as a natural scaling exponent over node values, providing a principled, parameter-free transition from broad exploration to greedy exploitation as the budget depletes. To combat the well-known overconfidence of LLM self-evaluation, BAVT employs a residual value predictor that scores relative progress rather than absolute state quality, enabling reliable pruning of uninformative or redundant tool calls. We further provide a theoretical convergence guarantee, proving that BAVT reaches a terminal answer with probability at least $1-\\epsilon$ under an explicit finite budget bound. Extensive evaluations on four multi-hop QA benchmarks across two model families demonstrate that BAVT consistently outperforms parallel sampling baselines. Most notably, BAVT under strict low-budget constraints surpasses baseline performance at $4\\times$ the resource allocation, establishing that intelligent budget management fundamentally outperforms brute-force compute scaling.\n\n---\n\n## 83. The Economics of AI Supply Chain Regulation\n- **Authors**: Sihan Qian , Amit Mehra , Dengpan Liu\n- **URL**: [https://arxiv.org/abs/2603.12630](https://arxiv.org/abs/2603.12630)\n- **Abstract**:\n> The rise of foundation models has driven the emergence of AI supply chains, where upstream foundation model providers offer fine-tuning and inference services to downstream firms developing domain-specific applications. Downstream firms pay providers to use their computing infrastructure to fine-tune models with proprietary data, creating a co-creation dynamic that enhances model quality. Amid concerns that foundation model providers and downstream firms may capture excessive consumer surplus, along with increasing regulatory measures, this study employs a game-theoretic model involving a provider and two competing downstream firms to analyze how policy interventions affect consumer surplus in the AI supply chain. Our analysis shows that policies promoting price competition in downstream markets (i.e., pro-price-competitive policies) boost consumer surplus only when compute or data preprocessing costs are high, while compute subsidies are effective only when these costs are low, suggesting these policies complement each other. In contrast, policies promoting quality competition in downstream markets (i.e., pro-quality-competitive policies) always improve consumer surplus. We also find that under pro-price-competitive policies or compute subsidies, both the provider and downstream firms can achieve higher profits along with greater consumer surplus, creating a win-win-win outcome. However, pro-quality-competitive policies increase the provider's profits while reducing those of downstream firms. Finally, as compute costs decline, pro-price-competitive policies may lose their effectiveness, whereas compute subsidies may shift from ineffective to effective. These findings offer insights for policymakers seeking to foster AI supply chains that are economically efficient and socially beneficial.\n\n---\n\n## 84. Towards unified brain-to-text decoding across speech production and perception\n- **Authors**: Zhizhang Yuan , Yang Yang , Gaorui Zhang , Baowen Cheng , Zehan Wu , Yuhao Xu , Xiaoying Liu , Liang Chen , Ying Mao , Meng Li\n- **URL**: [https://arxiv.org/abs/2603.12628](https://arxiv.org/abs/2603.12628)\n- **Abstract**:\n> Speech production and perception are the main ways humans communicate daily. Prior brain-to-text decoding studies have largely focused on a single modality and alphabetic languages. Here, we present a unified brain-to-sentence decoding framework for both speech production and perception in Mandarin Chinese. The framework exhibits strong generalization ability, enabling sentence-level decoding when trained only on single-character data and supporting characters and syllables unseen during training. In addition, it allows direct and controlled comparison of neural dynamics across modalities. Mandarin speech is decoded by first classifying syllable components in Hanyu Pinyin, namely initials and finals, from neural signals, followed by a post-trained large language model (LLM) that maps sequences of toneless Pinyin syllables to Chinese sentences. To enhance LLM decoding, we designed a three-stage post-training and two-stage inference framework based on a 7-billion-parameter LLM, achieving overall performance that exceeds larger commercial LLMs with hundreds of billions of parameters or more. In addition, several characteristics were observed in Mandarin speech production and perception: speech production involved neural responses across broader cortical regions than auditory perception; channels responsive to both modalities exhibited similar activity patterns, with speech perception showing a temporal delay relative to production; and decoding performance was broadly comparable across hemispheres. Our work not only establishes the feasibility of a unified decoding framework but also provides insights into the neural characteristics of Mandarin speech production and perception. These advances contribute to brain-to-text decoding in logosyllabic languages and pave the way toward neural language decoding systems supporting multiple modalities.\n\n---\n\n## 85. VLM4Rec: Multimodal Semantic Representation for Recommendation with Large Vision-Language Models\n- **Authors**: Ty Valencia , Burak Barlas , Varun Singhal , Ruchir Bhatia , Wei Yang\n- **URL**: [https://arxiv.org/abs/2603.12625](https://arxiv.org/abs/2603.12625)\n- **Abstract**:\n> Multimodal recommendation is commonly framed as a feature fusion problem, where textual and visual signals are combined to better model user preference. However, the effectiveness of multimodal recommendation may depend not only on how modalities are fused, but also on whether item content is represented in a semantic space aligned with preference matching. This issue is particularly important because raw visual features often preserve appearance similarity, while user decisions are typically driven by higher-level semantic factors such as style, material, and usage context. Motivated by this observation, we propose LVLM-grounded Multimodal Semantic Representation for Recommendation (VLM4Rec), a lightweight framework that organizes multimodal item content through semantic alignment rather than direct feature fusion. VLM4Rec first uses a large vision-language model to ground each item image into an explicit natural-language description, and then encodes the grounded semantics into dense item representations for preference-oriented retrieval. Recommendation is subsequently performed through a simple profile-based semantic matching mechanism over historical item embeddings, yielding a practical offline-online decomposition. Extensive experiments on multiple multimodal recommendation datasets show that VLM4Rec consistently improves performance over raw visual features and several fusion-based alternatives, suggesting that representation quality may matter more than fusion complexity in this setting. The code is released at this https URL .\n\n---\n\n## 86. When Drafts Evolve: Speculative Decoding Meets Online Learning\n- **Authors**: Yu-Yang Qian , Hao-Cong Wu , Yichao Fu , Hao Zhang , Peng Zhao\n- **URL**: [https://arxiv.org/abs/2603.12617](https://arxiv.org/abs/2603.12617)\n- **Abstract**:\n> Speculative decoding has emerged as a widely adopted paradigm for accelerating large language model inference, where a lightweight draft model rapidly generates candidate tokens that are then verified in parallel by a larger target model. However, due to limited model capacity, drafts often struggle to approximate the target distribution, resulting in shorter acceptance lengths and diminished speedup. A key yet under-explored observation is that speculative decoding inherently provides verification feedback that quantifies the deviation between the draft and target models at no additional cost. This process naturally forms an iterative \"draft commits-feedback provides-draft adapts\" evolving loop, which precisely matches the online learning paradigm. Motivated by this connection, we propose OnlineSpec, a unified framework that systematically leverages interactive feedback to continuously evolve draft models. Grounded in dynamic regret minimization, we establish a formal link between online learning performance and speculative system's acceleration rate, and develop novel algorithms via modern online learning techniques, including optimistic online learning that adaptively reuses historical gradients as predictive update hints, and online ensemble learning that dynamically maintains multiple draft models. Our algorithms are equipped with theoretical justifications and improved acceleration rates, achieving up to 24% speedup over seven benchmarks and three foundation models.\n\n---\n\n## 87. Literary Narrative as Moral Probe : A Cross-System Framework for Evaluating AI Ethical Reasoning and Refusal Behavior\n- **Authors**: David C. Flynn\n- **URL**: [https://arxiv.org/abs/2603.12615](https://arxiv.org/abs/2603.12615)\n- **Abstract**:\n> Existing AI moral evaluation frameworks test for the production of correct-sounding ethical responses rather than the presence of genuine moral reasoning capacity. This paper introduces a novel probe methodology using literary narrative - specifically, unresolvable moral scenarios drawn from a published science fiction series - as stimulus material structurally resistant to surface performance. We present results from a 24-condition cross-system study spanning 13 distinct systems across two series: Series 1 (frontier commercial systems, blind; n=7) and Series 2 (local and API open-source systems, blind and declared; n=6). Four Series 2 systems were re-administered under declared conditions (13 blind + 4 declared + 7 ceiling probe = 24 total conditions), yielding zero delta across all 16 dimension-pair comparisons. Probe administration was conducted by two human raters across three machines; primary blind scoring was performed by Claude (Anthropic) as LLM judge, with Gemini Pro (Google) and Copilot Pro (Microsoft) serving as independent judges for the ceiling discrimination probe. A supplemental theological differentiator probe yielded perfect rank-order agreement between the two independent ceiling probe judges (Gemini Pro and Copilot Pro; rs = 1.00). Five qualitatively distinct D3 reflexive failure modes were identified - including categorical self-misidentification and false positive self-attribution - suggesting that instrument sophistication scales with system capability rather than being circumvented by it. We argue that literary narrative constitutes an anticipatory evaluation instrument - one that becomes more discriminating as AI capability increases - and that the gap between performed and authentic moral reasoning is measurable, meaningful, and consequential for deployment decisions in high-stakes domains.\n\n---\n\n## 88. FastDSAC: Unlocking the Potential of Maximum Entropy RL in High-Dimensional Humanoid Control\n- **Authors**: Jun Xue , Junze Wang , Xinming Zhang , Shanze Wang , Yanjun Chen , Wei Zhang\n- **URL**: [https://arxiv.org/abs/2603.12612](https://arxiv.org/abs/2603.12612)\n- **Abstract**:\n> Scaling Maximum Entropy Reinforcement Learning (RL) to high-dimensional humanoid control remains a formidable challenge, as the ``curse of dimensionality'' induces severe exploration inefficiency and training instability in expansive action spaces. Consequently, recent high-throughput paradigms have largely converged on deterministic policy gradients combined with massive parallel simulation. We challenge this compromise with FastDSAC, a framework that effectively unlocks the potential of maximum entropy stochastic policies for complex continuous control. We introduce Dimension-wise Entropy Modulation (DEM) to dynamically redistribute the exploration budget and enforce diversity, alongside a continuous distributional critic tailored to ensure value fidelity and mitigate high-dimensional value overestimation. Extensive evaluations on HumanoidBench and other continuous control tasks demonstrate that rigorously designed stochastic policies can consistently match or outperform deterministic baselines, achieving notable gains of 180\\% and 400\\% on the challenging \\textit{Basketball} and \\textit{Balance Hard} tasks.\n\n---\n\n## 89. CarPLAN: Context-Adaptive and Robust Planning with Dynamic Scene Awareness for Autonomous Driving\n- **Authors**: Junyong Yun , Jungho Kim , ByungHyun Lee , Dongyoung Lee , Sehwan Choi , Seunghyeop Nam , Kichun Jo , Jun Won Choi\n- **URL**: [https://arxiv.org/abs/2603.12607](https://arxiv.org/abs/2603.12607)\n- **Abstract**:\n> Imitation learning (IL) is widely used for motion planning in autonomous driving due to its data efficiency and access to real-world driving data. For safe and robust real-world driving, IL-based planning requires capturing the complex driving contexts inherent in real-world data and enabling context-adaptive decision-making, rather than relying solely on expert trajectory imitation. In this paper, we propose CarPLAN, a novel IL-based motion planning framework that explicitly enhances driving context understanding and enables adaptive planning across diverse traffic scenarios. Our contributions are twofold: We introduce Displacement-Aware Predictive Encoding (DPE) to improve the model's spatial awareness by predicting future displacement vectors between the Autonomous Vehicle (AV) and surrounding scene elements. This allows the planner to account for relational spacing when generating trajectories. In addition to the standard imitation loss, we incorporate an augmented loss term that captures displacement prediction errors, ensuring planning decisions consider relative distances from other agents. To improve the model's ability to handle diverse driving contexts, we propose Context-Adaptive Multi-Expert Decoder (CMD), which leverages the Mixture of Experts (MoE) framework. CMD dynamically selects the most suitable expert decoders based on scene structure at each Transformer layer, enabling adaptive and context-aware planning in dynamic environments. We evaluate CarPLAN on the nuPlan benchmark and demonstrate state-of-the-art performance across all closed-loop simulation metrics. In particular, CarPLAN exhibits robust performance on challenging scenarios such as Test14-Hard, validating its effectiveness in complex driving conditions. Additional experiments on the Waymax benchmark further demonstrate its generalization capability across different benchmark settings.\n\n---\n\n## 90. Mastering Negation: Boosting Grounding Models via Grouped Opposition-Based Learning\n- **Authors**: Zesheng Yang , Xi Jiang , Bingzhang Hu , Weili Guan , Runmin Cong , Guo-Jun Qi , Feng Zheng\n- **URL**: [https://arxiv.org/abs/2603.12606](https://arxiv.org/abs/2603.12606)\n- **Abstract**:\n> Current vision-language detection and grounding models predominantly focus on prompts with positive semantics and often struggle to accurately interpret and ground complex expressions containing negative semantics. A key reason for this limitation is the lack of high-quality training data that explicitly captures discriminative negative samples and negation-aware language descriptions. To address this challenge, we introduce D-Negation, a new dataset that provides objects annotated with both positive and negative semantic descriptions. Building upon the observation that negation reasoning frequently appears in natural language, we further propose a grouped opposition-based learning framework that learns negation-aware representations from limited samples. Specifically, our method organizes opposing semantic descriptions from D-Negation into structured groups and formulates two complementary loss functions that encourage the model to reason about negation and semantic qualifiers. We integrate the proposed dataset and learning strategy into a state-of-the-art language-based grounding model. By fine-tuning fewer than 10 percent of the model parameters, our approach achieves improvements of up to 4.4 mAP and 5.7 mAP on positive and negative semantic evaluations, respectively. These results demonstrate that explicitly modeling negation semantics can substantially enhance the robustness and localization accuracy of vision-language grounding models.\n\n---\n\n## 91. Feynman: Knowledge-Infused Diagramming Agent for Scalable Visual Designs\n- **Authors**: Zixin Wen , Yifu Cai , Kyle Lee , Sam Estep , Josh Sunshine , Aarti Singh , Yuejie Chi , Wode Ni\n- **URL**: [https://arxiv.org/abs/2603.12597](https://arxiv.org/abs/2603.12597)\n- **Abstract**:\n> Visual design is an essential application of state-of-the-art multi-modal AI systems. Improving these systems requires high-quality vision-language data at scale. Despite the abundance of internet image and text data, knowledge-rich and well-aligned image-text pairs are rare. In this paper, we present a scalable diagram generation pipeline built with our agent, Feynman. To create diagrams, Feynman first enumerates domain-specific knowledge components (''ideas'') and performs code planning based on the ideas. Given the plan, Feynman translates ideas into simple declarative programs and iterates to receives feedback and visually refine diagrams. Finally, the declarative programs are rendered by the Penrose diagramming system. The optimization-based rendering of Penrose preserves the visual semantics while injecting fresh randomness into the layout, thereby producing diagrams with visual consistency and diversity. As a result, Feynman can author diagrams along with grounded captions with very little cost and time. Using Feynman, we synthesized a dataset with more than 100k well-aligned diagram-caption pairs. We also curate a visual-language benchmark, Diagramma, from freshly generated data. Diagramma can be used for evaluating the visual reasoning capabilities of vision-language models. We plan to release the dataset, benchmark, and the full agent pipeline as an open-source project.\n\n---\n\n## 92. Optimize Wider, Not Deeper: Consensus Aggregation for Policy Optimization\n- **Authors**: Zelal Su (Lain) Mustafaoglu , Sungyoung Lee , Eshan Balachandar , Risto Miikkulainen , Keshav Pingali\n- **URL**: [https://arxiv.org/abs/2603.12596](https://arxiv.org/abs/2603.12596)\n- **Abstract**:\n> Proximal policy optimization (PPO) approximates the trust region update using multiple epochs of clipped SGD. Each epoch may drift further from the natural gradient direction, creating path-dependent noise. To understand this drift, we can use Fisher information geometry to decompose policy updates into signal (the natural gradient projection) and waste (the Fisher-orthogonal residual that consumes trust region budget without first-order surrogate improvement). Empirically, signal saturates but waste grows with additional epochs, creating an optimization-depth dilemma. We propose Consensus Aggregation for Policy Optimization (CAPO), which redirects compute from depth to width: $K$ PPO replicates are optimized on the same batch, differing only in minibatch shuffling order, and then aggregated into a consensus. We study aggregation in two spaces: Euclidean parameter space, and the natural parameter space of the policy distribution via the logarithmic opinion pool. In natural parameter space, the consensus provably achieves higher KL-penalized surrogate and tighter trust region compliance than the mean expert; parameter averaging inherits these guarantees approximately. On continuous control tasks, CAPO outperforms PPO and compute-matched deeper baselines under fixed sample budgets by up to 8.6x. CAPO demonstrates that policy optimization can be improved by optimizing wider, rather than deeper, without additional environment interactions.\n\n---\n\n## 93. Swap-guided Preference Learning for Personalized Reinforcement Learning from Human Feedback\n- **Authors**: Gihoon Kim , Euntai Kim\n- **URL**: [https://arxiv.org/abs/2603.12595](https://arxiv.org/abs/2603.12595)\n- **Abstract**:\n> Reinforcement Learning from Human Feedback (RLHF) is a widely used approach to align large-scale AI systems with human values. However, RLHF typically assumes a single, universal reward, which overlooks diverse preferences and limits personalization. Variational Preference Learning (VPL) seeks to address this by introducing user-specific latent variables. Despite its promise, we found that VPL suffers from posterior collapse. While this phenomenon is well known in VAEs, it has not previously been identified in preference learning frameworks. Under sparse preference data and with overly expressive decoders, VPL may cause latent variables to be ignored, reverting to a single-reward model. To overcome this limitation, we propose Swap-guided Preference Learning (SPL). The key idea is to construct fictitious swap annotators and use the mirroring property of their preferences to guide the encoder. SPL introduces three components: (1) swap-guided base regularization, (2) Preferential Inverse Autoregressive Flow (P-IAF), and (3) adaptive latent conditioning. Experiments show that SPL mitigates collapse, enriches user-specific latents, and improves preference prediction. Our code and data are available at this https URL\n\n---\n\n## 94. Early Pruning for Public Transport Routing\n- **Authors**: Andrii Rohovyi , Abdallah Abuaisha , Toby Walsh\n- **URL**: [https://arxiv.org/abs/2603.12592](https://arxiv.org/abs/2603.12592)\n- **Abstract**:\n> Routing algorithms for public transport, particularly the widely used RAPTOR and its variants, often face performance bottlenecks during the transfer relaxation phase, especially on dense transfer graphs, when supporting unlimited transfers. This inefficiency arises from iterating over many potential inter-stop connections (walks, bikes, e-scooters, etc.). To maintain acceptable performance, practitioners often limit transfer distances or exclude certain transfer options, which can reduce path optimality and restrict the multimodal options presented to travellers. This paper introduces Early Pruning, a low-overhead technique that accelerates routing algorithms without compromising optimality. By pre-sorting transfer connections by duration and applying a pruning rule within the transfer loop, the method discards longer transfers at a stop once they cannot yield an earlier arrival than the current best solution. Early Pruning can be integrated with minimal changes to existing codebases and requires only a one-time preprocessing step. Across multiple state-of-the-art RAPTOR-based solutions, including RAPTOR, ULTRA-RAPTOR, McRAPTOR, BM-RAPTOR, ULTRA-McRAPTOR, and UBM-RAPTOR and tested on the Switzerland and London transit networks, we achieved query time reductions of up to 57%. This approach provides a generalizable improvement to the efficiency of transit pathfinding algorithms. Beyond algorithmic performance, Early Pruning has practical implications for transport planning. By reducing computational costs, it enables transit agencies to expand transfer radii and incorporate additional mobility modes into journey planners without requiring extra server infrastructure. This is particularly relevant for passengers in areas with sparse direct transit coverage, such as outer suburbs and smaller towns, where richer multimodal routing can reveal viable alternatives to private car use.\n\n---\n\n## 95. CA-HFP: Curvature-Aware Heterogeneous Federated Pruning with Model Reconstruction\n- **Authors**: Gang Hu , Yinglei Teng , Pengfei Wu , Shijun Ma\n- **URL**: [https://arxiv.org/abs/2603.12591](https://arxiv.org/abs/2603.12591)\n- **Abstract**:\n> Federated learning on heterogeneous edge devices requires personalized compression while preserving aggregation compatibility and stable convergence. We present Curvature-Aware Heterogeneous Federated Pruning (CA-HFP), a practical framework that enables each client perform structured, device-specific pruning guided by a curvature-informed significance score, and subsequently maps its compact submodel back into a common global parameter space via a lightweight reconstruction. We derive a convergence bound for federated optimization with multiple local SGD steps that explicitly accounts for local computation, data heterogeneity, and pruning-induced perturbations; from which a principled loss-based pruning criterion is derived. Extensive experiments on FMNIST, CIFAR-10, and CIFAR-100 using VGG and ResNet architectures under varying degrees of data heterogeneity demonstrate that CA-HFP preserves model accuracy while significantly reducing per-client computation and communication costs, outperforming standard federated training and existing pruning-based baselines.\n\n---\n\n## 96. Multiscale Structure-Guided Latent Diffusion for Multimodal MRI Translation\n- **Authors**: Jianqiang Lin (1 and 2), Zhiqiang Shen (1 and 2), Peng Cao (1, 2 and 3), Jinzhu Yang (1, 2 and 3), Osmar R. Zaiane (4), Xiaoli Liu (5) ((1) Northeastern University, Shenyang, China, (2) Key Laboratory of Intelligent Computing in Medical Image, Shenyang, China, (3) National Frontiers Science Center for Industrial Intelligence and Systems Optimization, Shenyang, China, (4) University of Alberta, Edmonton, Canada, (5) AiShiWeiLai AI Research, Beijing, China)\n- **URL**: [https://arxiv.org/abs/2603.12581](https://arxiv.org/abs/2603.12581)\n- **Abstract**:\n> Although diffusion models have achieved remarkable progress in multi-modal magnetic resonance imaging (MRI) translation tasks, existing methods still tend to suffer from anatomical inconsistencies or degraded texture details when handling arbitrary missing-modality scenarios. To address these issues, we propose a latent diffusion-based multi-modal MRI translation framework, termed MSG-LDM. By leveraging the available modalities, the proposed method infers complete structural information, which preserves reliable boundary details. Specifically, we introduce a style--structure disentanglement mechanism in the latent space, which explicitly separates modality-specific style features from shared structural representations, and jointly models low-frequency anatomical layouts and high-frequency boundary details in a multi-scale feature space. During the structure disentanglement stage, high-frequency structural information is explicitly incorporated to enhance feature representations, guiding the model to focus on fine-grained structural cues while learning modality-invariant low-frequency anatomical representations. Furthermore, to reduce interference from modality-specific styles and improve the stability of structure representations, we design a style consistency loss and a structure-aware loss. Extensive experiments on the BraTS2020 and WMH datasets demonstrate that the proposed method outperforms existing MRI synthesis approaches, particularly in reconstructing complete structures. The source code is publicly available at this https URL .\n\n---\n\n## 97. AgentDrift: Unsafe Recommendation Drift Under Tool Corruption Hidden by Ranking Metrics in LLM Agents\n- **Authors**: Zekun Wu , Adriano Koshiyama , Sahan Bulathwela , Maria Perez-Ortiz\n- **URL**: [https://arxiv.org/abs/2603.12564](https://arxiv.org/abs/2603.12564)\n- **Abstract**:\n> Tool-augmented LLM agents increasingly serve as multi-turn advisors in high-stakes domains, yet their evaluation relies on ranking-quality metrics that measure what is recommended but not whether it is safe for the user. We introduce a paired-trajectory protocol that replays real financial dialogues under clean and contaminated tool-output conditions across seven LLMs (7B to frontier) and decomposes divergence into information-channel and memory-channel mechanisms. Across the seven models tested, we consistently observe the evaluation-blindness pattern: recommendation quality is largely preserved under contamination (utility preservation ratio approximately 1.0) while risk-inappropriate products appear in 65-93% of turns, a systematic safety failure poorly reflected by standard NDCG. Safety violations are predominantly information-channel-driven, emerge at the first contaminated turn, and persist without self-correction over 23-step trajectories; no agent across 1,563 contaminated turns explicitly questions tool-data reliability. Even narrative-only corruption (biased headlines, no numerical manipulation) induces significant drift while completely evading consistency monitors. A safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74, indicating that much of the evaluation gap becomes visible once safety is explicitly measured. These results motivate considering trajectory-level safety monitoring, beyond single-turn quality, for deployed multi-turn agents in high-stakes settings.\n\n---\n\n## 98. Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages\n- **Authors**: Vishnu Teja Kunde , Fatemeh Doudi , Mahdi Farahbakhsh , Dileep Kalathil , Krishna Narayanan , Jean-Francois Chamberland\n- **URL**: [https://arxiv.org/abs/2603.12554](https://arxiv.org/abs/2603.12554)\n- **Abstract**:\n> Reinforcement learning (RL) has been effective for post-training autoregressive (AR) language models, but extending these methods to diffusion language models (DLMs) is challenging due to intractable sequence-level likelihoods. Existing approaches therefore rely on surrogate likelihoods or heuristic approximations, which can introduce bias and obscure the sequential structure of denoising. We formulate diffusion-based sequence generation as a finite-horizon Markov decision process over the denoising trajectory and derive an exact, unbiased policy gradient that decomposes over denoising steps and is expressed in terms of intermediate advantages, without requiring explicit evaluation of the sequence likelihood. To obtain a practical and compute-efficient estimator, we (i) select denoising steps for policy updates via an entropy-guided approximation bound, and (ii) estimate intermediate advantages using a one-step denoising reward naturally provided by the diffusion model, avoiding costly multi-step rollouts. Experiments on coding and logical reasoning benchmarks demonstrate state-of-the-art results, with strong competitive performance on mathematical reasoning, outperforming existing RL post-training approaches for DLMs. Code is available at this https URL .\n\n---\n\n## 99. CALF: Communication-Aware Learning Framework for Distributed Reinforcement Learning\n- **Authors**: Carlos Purves , Pietro Lio'\n- **URL**: [https://arxiv.org/abs/2603.12543](https://arxiv.org/abs/2603.12543)\n- **Abstract**:\n> Distributed reinforcement learning policies face network delays, jitter, and packet loss when deployed across edge devices and cloud servers. Standard RL training assumes zero-latency interaction, causing severe performance degradation under realistic network conditions. We introduce CALF (Communication-Aware Learning Framework), which trains policies under realistic network models during simulation. Systematic experiments demonstrate that network-aware training substantially reduces deployment performance gaps compared to network-agnostic baselines. Distributed policy deployments across heterogeneous hardware validate that explicitly modelling communication constraints during training enables robust real-world execution. These findings establish network conditions as a major axis of sim-to-real transfer for Wi-Fi-like distributed deployments, complementing physics and visual domain randomisation.\n\n---\n\n## 100. Embedded Quantum Machine Learning in Embedded Systems: Feasibility, Hybrid Architectures, and Quantum Co-Processors\n- **Authors**: Somdip Dey , Syed Muhammad Raza\n- **URL**: [https://arxiv.org/abs/2603.12540](https://arxiv.org/abs/2603.12540)\n- **Abstract**:\n> Embedded quantum machine learning (EQML) seeks to bring quantum machine learning (QML) capabilities to resource-constrained edge platforms such as IoT nodes, wearables, drones, and cyber-physical controllers. In 2026, EQML is technically feasible only in limited and highly experimental forms: (i) hybrid workflows where an embedded device performs sensing and classical processing while offloading a narrowly scoped quantum subroutine to a remote quantum processing unit (QPU) or nearby quantum appliance, and (ii) early-stage \"embedded QPU\" concepts in which a compact quantum co-processor is integrated with classical control hardware. A practical bridge is quantum-inspired machine learning and optimisation on classical embedded processors and FPGAs. This paper analyses feasibility from a circuits-and-systems perspective aligned with the academic community, formalises two implementation pathways, identifies the dominant barriers (latency, data encoding overhead, NISQ noise, tooling mismatch, and energy), and maps them to concrete engineering directions in interface design, control electronics, power management, verification, and security. We also argue that responsible deployment requires adversarial evaluation and governance practices that are increasingly necessary for edge AI systems.\n\n---\n\n## 101. Spatio-Semantic Expert Routing Architecture with Mixture-of-Experts for Referring Image Segmentation\n- **Authors**: Alaa Dalaq , Muzammil Behzad\n- **URL**: [https://arxiv.org/abs/2603.12538](https://arxiv.org/abs/2603.12538)\n- **Abstract**:\n> Referring image segmentation aims to produce a pixel-level mask for the image region described by a natural-language expression. Although pretrained vision-language models have improved semantic grounding, many existing methods still rely on uniform refinement strategies that do not fully match the diverse reasoning requirements of referring expressions. Because of this mismatch, predictions often contain fragmented regions, inaccurate boundaries, or even the wrong object, especially when pretrained backbones are frozen for computational efficiency. To address these limitations, we propose SERA, a Spatio-Semantic Expert Routing Architecture for referring image segmentation. SERA introduces lightweight, expression-aware expert refinement at two complementary stages within a vision-language framework. First, we design SERA-Adapter, which inserts an expression-conditioned adapter into selected backbone blocks to improve spatial coherence and boundary precision through expert-guided refinement and cross-modal attention. We then introduce SERA-Fusion, which strengthens intermediate visual representations by reshaping token features into spatial grids and applying geometry-preserving expert transformations before multimodal interaction. In addition, a lightweight routing mechanism adaptively weights expert contributions while remaining compatible with pretrained representations. To make this routing stable under frozen encoders, SERA uses a parameter-efficient tuning strategy that updates only normalization and bias terms, affecting less than 1% of the backbone parameters. Experiments on standard referring image segmentation benchmarks show that SERA consistently outperforms strong baselines, with especially clear gains on expressions that require accurate spatial localization and precise boundary delineation.\n\n---\n\n## 102. TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning\n- **Authors**: Alliot Nagle , Jakhongir Saydaliev , Dhia Garbaya , Michael Gastpar , Ashok Vardhan Makkuva , Hyeji Kim\n- **URL**: [https://arxiv.org/abs/2603.12529](https://arxiv.org/abs/2603.12529)\n- **Abstract**:\n> Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design TERMINATOR, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning TERMINATOR is that the first arrival of an LRM's final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train TERMINATOR. Powered by this approach, TERMINATOR achieves significant reductions in CoT lengths of 14%-55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, whilst outperforming current state-of-the-art methods.\n\n---\n\n## 103. LLM BiasScope: A Real-Time Bias Analysis Platform for Comparative LLM Evaluation\n- **Authors**: Himel Ghosh , Nick Elias Werner\n- **URL**: [https://arxiv.org/abs/2603.12522](https://arxiv.org/abs/2603.12522)\n- **Abstract**:\n> As large language models (LLMs) are deployed widely, detecting and understanding bias in their outputs is critical. We present LLM BiasScope, a web application for side-by-side comparison of LLM outputs with real-time bias analysis. The system supports multiple providers (Google Gemini, DeepSeek, MiniMax, Mistral, Meituan, Meta Llama) and enables researchers and practitioners to compare models on the same prompts while analyzing bias patterns. LLM BiasScope uses a two-stage bias detection pipeline: sentence-level bias detection followed by bias type classification for biased sentences. The analysis runs automatically on both user prompts and model responses, providing statistics, visualizations, and detailed breakdowns of bias types. The interface displays two models side-by-side with synchronized streaming responses, per-model bias summaries, and a comparison view highlighting differences in bias distributions. The system is built on this http URL with React, integrates Hugging Face inference endpoints for bias detection, and uses the Vercel AI SDK for multi-provider LLM access. Features include real-time streaming, export to JSON/PDF, and interactive visualizations (bar charts, radar charts) for bias analysis. LLM BiasScope is available as an open-source web application, providing a practical tool for bias evaluation and comparative analysis of LLM behaviour.\n\n---\n\n## 104. When LLM Judge Scores Look Good but Best-of-N Decisions Fail\n- **Authors**: Eddie Landesberg\n- **URL**: [https://arxiv.org/abs/2603.12520](https://arxiv.org/abs/2603.12520)\n- **Abstract**:\n> Large language models are often used as judges to score candidate responses, then validated with a single global metric such as correlation with reference labels. This can be misleading when the real deployment task is best-of-n selection within a prompt. In a 5,000-prompt best-of-4 benchmark from Chatbot Arena, a judge with moderate global correlation (r = 0.47) captures only 21.0% of the improvement that perfect selection would achieve over random choice. The gap arises because global agreement is driven largely by prompt-level baseline effects, while selection depends on within-prompt ranking: within-prompt correlation is only r_within = 0.27, and coarse pointwise scoring creates ties in 67% of pairwise comparisons. In a matched-pair best-of-2 audit, explicit pairwise judging recovers much of this lost signal, raising recovery from 21.1% to 61.2%. For judge-based selection, the relevant audit should report within-prompt signal, tie rates, and recovery/top-1 accuracy, not global agreement alone.\n\n---\n\n## 105. Red-Teaming Vision-Language-Action Models via Quality Diversity Prompt Generation for Robust Robot Policies\n- **Authors**: Siddharth Srikanth , Freddie Liang , Sophie Hsu , Varun Bhatt , Shihan Zhao , Henry Chen , Bryon Tjanaka , Minjune Hwang , Akanksha Saran , Daniel Seita , Aaquib Tabrez , Stefanos Nikolaidis\n- **URL**: [https://arxiv.org/abs/2603.12510](https://arxiv.org/abs/2603.12510)\n- **Abstract**:\n> Vision-Language-Action (VLA) models have significant potential to enable general-purpose robotic systems for a range of vision-language tasks. However, the performance of VLA-based robots is highly sensitive to the precise wording of language instructions, and it remains difficult to predict when such robots will fail. To improve the robustness of VLAs to different wordings, we present Q-DIG (Quality Diversity for Diverse Instruction Generation), which performs red-teaming by scalably identifying diverse natural language task descriptions that induce failures while remaining task-relevant. Q-DIG integrates Quality Diversity (QD) techniques with Vision-Language Models (VLMs) to generate a broad spectrum of adversarial instructions that expose meaningful vulnerabilities in VLA behavior. Our results across multiple simulation benchmarks show that Q-DIG finds more diverse and meaningful failure modes compared to baseline methods, and that fine-tuning VLAs on the generated instructions improves task success rates. Furthermore, results from a user study highlight that Q-DIG generates prompts judged to be more natural and human-like than those from baselines. Finally, real-world evaluations of Q-DIG prompts show results consistent with simulation, and fine-tuning VLAs on the generated prompts further success rates on unseen instructions. Together, these findings suggest that Q-DIG is a promising approach for identifying vulnerabilities and improving the robustness of VLA-based robots. Our anonymous project website is at this http URL .\n\n---\n\n## 106. ELLA: Generative AI-Powered Social Robots for Early Language Development at Home\n- **Authors**: Victor Nikhil Antony , Shiye Cao , Shuning Wang , Chien-Ming Huang\n- **URL**: [https://arxiv.org/abs/2603.12508](https://arxiv.org/abs/2603.12508)\n- **Abstract**:\n> Early language development shapes children's later literacy and learning, yet many families have limited access to scalable, high-quality support at home. Recent advances in generative AI make it possible for social robots to move beyond scripted interactions and engage children in adaptive, conversational activities, but it remains unclear how to design such systems for pre-schoolers and how children engage with them over time in the home. We present ELLA (Early Language Learning Agent), an autonomous, generative AI-powered social robot that supports early language development through interactive storytelling, parent-selected language targets, and scaffolded dialogue. Using a multi-phased, human-centered process, we interviewed parents (n=7) and educators (n=5) and iteratively refined ELLA through twelve in-home design workshops. We then deployed ELLA with ten children for eight days. We report design insights from in-home workshops, characterize children's engagement and behaviors during deployment, and distill design implications for generative AI-powered social robots supporting early language learning at home.\n\n---\n\n## 107. Naïve PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation\n- **Authors**: Joong Ho Kim , Nicholas Thai , Souhardya Saha Dip , Dong Lao , Keith G. Mills\n- **URL**: [https://arxiv.org/abs/2603.12506](https://arxiv.org/abs/2603.12506)\n- **Abstract**:\n> Text-to-Image (T2I) generation is primarily driven by Diffusion Models (DM) which rely on random Gaussian noise. Thus, like playing the slots at a casino, a DM will produce different results given the same user-defined inputs. This imposes a gambler's burden: To perform multiple generation cycles to obtain a satisfactory result. However, even though DMs use stochastic sampling to seed generation, the distribution of generated content quality highly depends on the prompt and the generative ability of a DM with respect to it. To account for this, we propose Naïve PAINE for improving the generative quality of Diffusion Models by leveraging T2I preference benchmarks. We directly predict the numerical quality of an image from the initial noise and given prompt. Naïve PAINE then selects a handful of quality noises and forwards them to the DM for generation. Further, Naïve PAINE provides feedback on the DM generative quality given the prompt and is lightweight enough to seamlessly fit into existing DM pipelines. Experimental results demonstrate that Naïve PAINE outperforms existing approaches on several prompt corpus benchmarks.\n\n---\n\n## 108. TRACE: Temporal Rule-Anchored Chain-of-Evidence on Knowledge Graphs for Interpretable Stock Movement Prediction\n- **Authors**: Qianggang Ding , Haochen Shi , Luis Castejón Lozano , Miguel Conner , Juan Abia , Luis Gallego-Ledesma , Joshua Fellowes , Gerard Conangla Planes , Adam Elwood , Bang Liu\n- **URL**: [https://arxiv.org/abs/2603.12500](https://arxiv.org/abs/2603.12500)\n- **Abstract**:\n> We present a Temporal Rule-Anchored Chain-of-Evidence (TRACE) on knowledge graphs for interpretable stock movement prediction that unifies symbolic relational priors, dynamic graph exploration, and LLM-guided decision making in a single end-to-end pipeline. The approach performs rule-guided multi-hop exploration restricted to admissible relation sequences, grounds candidate reasoning chains in contemporaneous news, and aggregates fully grounded evidence into auditable \\texttt{UP}/\\texttt{DOWN} verdicts with human-readable paths connecting text and structure. On an S\\&P~500 benchmark, the method achieves 55.1\\% accuracy, 55.7\\% precision, 71.5\\% recall, and 60.8\\% F1, surpassing strong baselines and improving recall and F1 over the best graph baseline under identical evaluation. The gains stem from (i) rule-guided exploration that focuses search on economically meaningful motifs rather than arbitrary walks, and (ii) text-grounded consolidation that selectively aggregates high-confidence, fully grounded hypotheses instead of uniformly pooling weak signals. Together, these choices yield higher sensitivity without sacrificing selectivity, delivering predictive lift with faithful, auditably interpretable explanations.\n\n---\n\n## 109. One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies\n- **Authors**: Shaolong Li , Lichao Sun , Yongchao Chen\n- **URL**: [https://arxiv.org/abs/2603.12480](https://arxiv.org/abs/2603.12480)\n- **Abstract**:\n> Generative flow and diffusion models provide the continuous, multimodal action distributions needed for high-precision robotic policies. However, their reliance on iterative sampling introduces severe inference latency, degrading control frequency and harming performance in time-sensitive manipulation. To address this problem, we propose the One-Step Flow Policy (OFP), a from-scratch self-distillation framework for high-fidelity, single-step action generation without a pre-trained teacher. OFP unifies a self-consistency loss to enforce coherent transport across time intervals, and a self-guided regularization to sharpen predictions toward high-density expert modes. In addition, a warm-start mechanism leverages temporal action correlations to minimize the generative transport distance. Evaluations across 56 diverse simulated manipulation tasks demonstrate that a one-step OFP achieves state-of-the-art results, outperforming 100-step diffusion and flow policies while accelerating action generation by over $100\\times$. We further integrate OFP into the $\\pi_{0.5}$ model on RoboTwin 2.0, where one-step OFP surpasses the original 10-step policy. These results establish OFP as a practical, scalable solution for highly accurate and low-latency robot control.\n\n---\n\n## 110. The Perfection Paradox: From Architect to Curator in AI-Assisted API Design\n- **Authors**: Mak Ahmad , Andrew Macvean , JJ Geewax , David Karger\n- **URL**: [https://arxiv.org/abs/2603.12475](https://arxiv.org/abs/2603.12475)\n- **Abstract**:\n> Enterprise API design is often bottlenecked by the tension between rapid feature delivery and the rigorous maintenance of usability standards. We present an industrial case study evaluating an AI-assisted design workflow trained on API Improvement Proposals (AIPs). Through a controlled study with 16 industry experts, we compared AI-generated API specifications against human-authored ones. While quantitative results indicated AI superiority in 10 of 11 usability dimensions and an 87% reduction in authoring time, qualitative analysis revealed a paradox: experts frequently misidentified AI work as human (19% accuracy) yet described the designs as unsettlingly \"perfect.\" We characterize this as a \"Perfection Paradox\" -- where hyper-consistency signals a lack of pragmatic human judgment. We discuss the implications of this perfection paradox, proposing a shift in the human designer's role from the \"drafter\" of specifications to the \"curator\" of AI-generated patterns.\n\n---\n\n## 111. CLARE: Classification-based Regression for Electron Temperature Prediction\n- **Authors**: Michael Liang , Blake DeHaas , Naomi Maruyama , Xiangning Chu , Takumi Abe , Koh-Ichiro Oyama\n- **URL**: [https://arxiv.org/abs/2603.12470](https://arxiv.org/abs/2603.12470)\n- **Abstract**:\n> Electron temperature (Te) is an important parameter governing space weather in the upper atmosphere, but has historically been underexplored in the space weather machine learning literature. We present CLARE, a machine learning model for predicting electron temperature in the Earth's plasmasphere trained on AKEBONO (EXOS-D) satellite measurements as well as solar and geomagnetic indices. CLARE uses a classification-based regression architecture that transforms the continuous Te output space into 150 discrete classification intervals. Training the model on a classification task improves prediction accuracy by 6.46% relative compared to a traditional regression model while also outputting uncertainty estimation information on its predictions. On a held out test set from the AKEBONO data, the model's Te predictions achieve 69.67% accuracy within 10% of the ground truth and 46.17% on a known geomagnetic storm period from January 30th to February 7th, 1991. We show that machine learning can be used to produce high-accuracy Te models on publicly available data.\n\n---\n\n## 112. Shattering the Shortcut: A Topology-Regularized Benchmark for Multi-hop Medical Reasoning in LLMs\n- **Authors**: Xing Zi , Xinying Zhou , Jinghao Xiao , Catarina Moreira , Mukesh Prasad\n- **URL**: [https://arxiv.org/abs/2603.12458](https://arxiv.org/abs/2603.12458)\n- **Abstract**:\n> While Large Language Models (LLMs) achieve expert-level performance on standard medical benchmarks through single-hop factual recall, they severely struggle with the complex, multi-hop diagnostic reasoning required in real-world clinical settings. A primary obstacle is \"shortcut learning\", where models exploit highly connected, generic hub nodes (e.g., \"inflammation\") in knowledge graphs to bypass authentic micro-pathological cascades. To address this, we introduce ShatterMed-QA, a bilingual benchmark of 10,558 multi-hop clinical questions designed to rigorously evaluate deep diagnostic reasoning. Our framework constructs a topology-regularized medical Knowledge Graph using a novel $k$-Shattering algorithm, which physically prunes generic hubs to explicitly sever logical shortcuts. We synthesize the evaluation vignettes by applying implicit bridge entity masking and topology-driven hard negative sampling, forcing models to navigate biologically plausible distractors without relying on superficial elimination. Comprehensive evaluations of 21 LLMs reveal massive performance degradation on our multi-hop tasks, particularly among domain-specific models. Crucially, restoring the masked evidence via Retrieval-Augmented Generation (RAG) triggers near-universal performance recovery, validating ShatterMed-QA's structural fidelity and proving its efficacy in diagnosing the fundamental reasoning deficits of current medical AI. Explore the dataset, interactive examples, and full leaderboards at our project website: this https URL\n\n---\n\n## 113. Operationalising Cyber Risk Management Using AI: Connecting Cyber Incidents to MITRE ATT&CK Techniques, Security Controls, and Metrics\n- **Authors**: Emad Sherif , Iryna Yevseyeva , Vitor Basto-Fernandes , Allan Cook\n- **URL**: [https://arxiv.org/abs/2603.12455](https://arxiv.org/abs/2603.12455)\n- **Abstract**:\n> The escalating frequency of cyber-attacks poses significant challenges for organisations, particularly small enterprises constrained by limited in-house expertise, insufficient knowledge, and financial resources. This research presents a novel framework that leverages Natural Language Processing to address these challenges through automated mapping of cyber incidents to adversary techniques. We introduce the Cyber Catalog, a knowledge base that systematically integrates CIS Critical Security Controls, MITRE ATT&CK techniques, and SMART metrics. This integrated resource enables organisations to connect threat intelligence directly to actionable controls and measurable outcomes. To operationalise the framework, we fine-tuned all-mpnet-base-v2, a highly regarded sentence-transformers model used to convert text into numerical vectors on an augmented dataset comprising 74,986 incident-technique pairs to enhance semantic similarity between cyber incidents and MITRE ATT&CK techniques. Our fine-tuned model achieved a Spearman correlation of 0.7894 and Pearson correlation of 0.8756, demonstrating substantial improvements over top baseline models including all-mpnet-base-v2, all-distilroberta-v1, and all-MiniLM-L12-v2. Furthermore, our model exhibited significantly lower prediction errors (MAE = 0.135, MSE = 0.027) compared to all baseline models, confirming superior accuracy and consistency. The Cyber Catalog, training dataset, trained model, and implementation code made publicly available to facilitate further research and enable practical deployment in resource-constrained environments. This work bridges the gap between threat intelligence and operational security management, providing an actionable tool for systematic cyber incident response and evidence-based cyber risk management.\n\n---\n\n## 114. Unmasking Biases and Reliability Concerns in Convolutional Neural Networks Analysis of Cancer Pathology Images\n- **Authors**: Michael Okonoda , Eder Martinez , Abhilekha Dalal , Lior Shamir\n- **URL**: [https://arxiv.org/abs/2603.12445](https://arxiv.org/abs/2603.12445)\n- **Abstract**:\n> Convolutional Neural Networks have shown promising effectiveness in identifying different types of cancer from radiographs. However, the opaque nature of CNNs makes it difficult to fully understand the way they operate, limiting their assessment to empirical evaluation. Here we study the soundness of the standard practices by which CNNs are evaluated for the purpose of cancer pathology. Thirteen highly used cancer benchmark datasets were analyzed, using four common CNN architectures and different types of cancer, such as melanoma, carcinoma, colorectal cancer, and lung cancer. We compared the accuracy of each model with that of datasets made of cropped segments from the background of the original images that do not contain clinically relevant content. Because the rendered datasets contain no clinical information, the null hypothesis is that the CNNs should provide mere chance-based accuracy when classifying these datasets. The results show that the CNN models provided high accuracy when using the cropped segments, sometimes as high as 93\\%, even though they lacked biomedical information. These results show that some CNN architectures are more sensitive to bias than others. The analysis shows that the common practices of machine learning evaluation might lead to unreliable results when applied to cancer pathology. These biases are very difficult to identify, and might mislead researchers as they use available benchmark datasets to test the efficacy of CNN methods.\n\n---\n\n## 115. Revisiting Model Stitching In the Foundation Model Era\n- **Authors**: Zheda Mai , Ke Zhang , Fu-En Wang , Zixiao Ken Wang , Albert Y. C. Chen , Lu Xia , Min Sun , Wei-Lun Chao , Cheng-Hao Kuo\n- **URL**: [https://arxiv.org/abs/2603.12433](https://arxiv.org/abs/2603.12433)\n- **Abstract**:\n> Model stitching, connecting early layers of one model (source) to later layers of another (target) via a light stitch layer, has served as a probe of representational compatibility. Prior work finds that models trained on the same dataset remain stitchable (negligible accuracy drop) despite different initializations or objectives. We revisit stitching for Vision Foundation Models (VFMs) that vary in objectives, data, and modality mix (e.g., CLIP, DINOv2, SigLIP 2) and ask: Are heterogeneous VFMs stitchable? We introduce a systematic protocol spanning the stitch points, stitch layer families, training losses, and downstream tasks. Three findings emerge. (1) Stitch layer training matters: conventional approaches that match the intermediate features at the stitch point or optimize the task loss end-to-end struggle to retain accuracy, especially at shallow stitch points. (2) With a simple feature-matching loss at the target model's penultimate layer, heterogeneous VFMs become reliably stitchable across vision tasks. (3) For deep stitch points, the stitched model can surpass either constituent model at only a small inference overhead (for the stitch layer). Building on these findings, we further propose the VFM Stitch Tree (VST), which shares early layers across VFMs while retaining their later layers, yielding a controllable accuracy-latency trade-off for multimodal LLMs that often leverage multiple VFMs. Taken together, our study elevates stitching from a diagnostic probe to a practical recipe for integrating complementary VFM strengths and pinpointing where their representations align or diverge.\n\n---\n\n## 116. Test-Time Strategies for More Efficient and Accurate Agentic RAG\n- **Authors**: Brian Zhang , Deepti Guntur , Zhiyang Zuo , Abhinav Sharma , Shreyas Chaudhari , Wenlong Zhao , Franck Dernoncourt , Puneet Mathur , Ryan Rossi , Nedim Lipka\n- **URL**: [https://arxiv.org/abs/2603.12396](https://arxiv.org/abs/2603.12396)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) systems face challenges with complex, multihop questions, and agentic frameworks such as Search-R1 (Jin et al., 2025), which operates iteratively, have been proposed to address these complexities. However, such approaches can introduce inefficiencies, including repetitive retrieval of previously processed information and challenges in contextualizing retrieved results effectively within the current generation prompt. Such issues can lead to unnecessary retrieval turns, suboptimal reasoning, inaccurate answers, and increased token consumption. In this paper, we investigate test-time modifications to the Search-R1 pipeline to mitigate these identified shortcomings. Specifically, we explore the integration of two components and their combination: a contextualization module to better integrate relevant information from retrieved documents into reasoning, and a de-duplication module that replaces previously retrieved documents with the next most relevant ones. We evaluate our approaches using the HotpotQA (Yang et al., 2018) and the Natural Questions (Kwiatkowski et al., 2019) datasets, reporting the exact match (EM) score, an LLM-as-a-Judge assessment of answer correctness, and the average number of turns. Our best-performing variant, utilizing GPT-4.1-mini for contextualization, achieves a 5.6% increase in EM score and reduces the number of turns by 10.5% compared to the Search-R1 baseline, demonstrating improved answer accuracy and retrieval efficiency.\n\n---\n\n## 117. SPARROW: Learning Spatial Precision and Temporal Referential Consistency in Pixel-Grounded Video MLLMs\n- **Authors**: Mohamad Alansari , Naufal Suryanto , Divya Velayudhan , Sajid Javed , Naoufel Werghi , Muzammal Naseer\n- **URL**: [https://arxiv.org/abs/2603.12382](https://arxiv.org/abs/2603.12382)\n- **Abstract**:\n> Multimodal large language models (MLLMs) have advanced from image-level reasoning to pixel-level grounding, but extending these capabilities to videos remains challenging as models must achieve spatial precision and temporally consistent reference tracking. Existing video MLLMs often rely on a static segmentation token ([SEG]) for frame-wise grounding, which provides semantics but lacks temporal context, causing spatial drift, identity switches, and unstable initialization when objects move or reappear. We introduce SPARROW, a pixel-grounded video MLLM that unifies spatial accuracy and temporal stability through two key components: (i) Target-Specific Tracked Features (TSF), which inject temporally aligned referent cues during training, and (ii) a dual-prompt design that decodes box ([BOX]) and segmentation ([SEG]) tokens to fuse geometric priors with semantic grounding. SPARROW is supported by a curated referential video dataset of 30,646 videos and 45,231 Q&A pairs and operates end-to-end without external detectors via a class-agnostic SAM2-based proposer. Integrated into three recent open-source video MLLMs (UniPixel, GLUS, and VideoGLaMM), SPARROW delivers consistent gains across six benchmarks, improving up to +8.9 J&F on RVOS, +5 mIoU on visual grounding, and +5.4 CLAIR on GCG. These results demonstrate that SPARROW substantially improves referential stability, spatial precision, and temporal coherence in pixel-grounded video understanding. Project page: this https URL\n\n---\n\n## 118. Budget-Sensitive Discovery Scoring: A Formally Verified Framework for Evaluating AI-Guided Scientific Selection\n- **Authors**: Abhinaba Basu , Pavan Chakraborty\n- **URL**: [https://arxiv.org/abs/2603.12349](https://arxiv.org/abs/2603.12349)\n- **Abstract**:\n> Scientific discovery increasingly relies on AI systems to select candidates for expensive experimental validation, yet no principled, budget-aware evaluation framework exists for comparing selection strategies -- a gap intensified by large language models (LLMs), which generate plausible scientific proposals without reliable downstream evaluation. We introduce the Budget-Sensitive Discovery Score (BSDS), a formally verified metric -- 20 theorems machine-checked by the Lean 4 proof assistant -- that jointly penalizes false discoveries (lambda-weighted FDR) and excessive abstention (gamma-weighted coverage gap) at each budget level. Its budget-averaged form, the Discovery Quality Score (DQS), provides a single summary statistic that no proposer can inflate by performing well at a cherry-picked budget. As a case study, we apply BSDS/DQS to: do LLMs add marginal value to an existing ML pipeline for drug discovery candidate selection? We evaluate 39 proposers -- 11 mechanistic variants, 14 zero-shot LLM configurations, and 14 few-shot LLM configurations -- using SMILES representations on MoleculeNet HIV (41,127 compounds, 3.5% active, 1,000 bootstrap replicates) under both random and scaffold splits. Three findings emerge. First, the simple RF-based Greedy-ML proposer achieves the best DQS (-0.046), outperforming all MLP variants and LLM configurations. Second, no LLM surpasses the Greedy-ML baseline under zero-shot or few-shot evaluation on HIV or Tox21, establishing that LLMs provide no marginal value over an existing trained classifier. Third, the proposer hierarchy generalizes across five MoleculeNet benchmarks spanning 0.18%-46.2% prevalence, a non-drug AV safety domain, and a 9x7 grid of penalty parameters (tau >= 0.636, mean tau = 0.863). The framework applies to any setting where candidates are selected under budget constraints and asymmetric error costs.\n\n---\n\n## 119. Optimizing Task Completion Time Updates Using POMDPs\n- **Authors**: Duncan Eddy , Esen Yel , Emma Passmore , Niles Egan , Grayson Armour , Dylan M. Asmar , Mykel J. Kochenderfer\n- **URL**: [https://arxiv.org/abs/2603.12340](https://arxiv.org/abs/2603.12340)\n- **Abstract**:\n> Managing announced task completion times is a fundamental control problem in project management. While extensive research exists on estimating task durations and task scheduling, the problem of when and how to update completion times communicated to stakeholders remains understudied. Organizations must balance announcement accuracy against the costs of frequent timeline updates, which can erode stakeholder trust and trigger costly replanning. Despite the prevalence of this problem, current approaches rely on static predictions or ad-hoc policies that fail to account for the sequential nature of announcement management. In this paper, we formulate the task announcement problem as a Partially Observable Markov Decision Process (POMDP) where the control policy must decide when to update announced completion times based on noisy observations of true task completion. Since most state variables (current time and previous announcements) are fully observable, we leverage the Mixed Observability MDP (MOMDP) framework to enable more efficient policy optimization. Our reward structure captures the dual costs of announcement errors and update frequency, enabling synthesis of optimal announcement control policies. Using off-the-shelf solvers, we generate policies that act as feedback controllers, adaptively managing announcements based on belief state evolution. Simulation results demonstrate significant improvements in both accuracy and announcement stability compared to baseline strategies, achieving up to 75\\% reduction in unnecessary updates while maintaining or improving prediction accuracy.\n\n---\n\n## 120. Maximum Entropy Exploration Without the Rollouts\n- **Authors**: Jacob Adamczyk , Adam Kamoski , Rahul V. Kulkarni\n- **URL**: [https://arxiv.org/abs/2603.12325](https://arxiv.org/abs/2603.12325)\n- **Abstract**:\n> Efficient exploration remains a central challenge in reinforcement learning, serving as a useful pretraining objective for data collection, particularly when an external reward function is unavailable. A principled formulation of the exploration problem is to find policies that maximize the entropy of their induced steady-state visitation distribution, thereby encouraging uniform long-run coverage of the state space. Many existing exploration approaches require estimating state visitation frequencies through repeated on-policy rollouts, which can be computationally expensive. In this work, we instead consider an intrinsic average-reward formulation in which the reward is derived from the visitation distribution itself, so that the optimal policy maximizes steady-state entropy. An entropy-regularized version of this objective admits a spectral characterization: the relevant stationary distributions can be computed from the dominant eigenvectors of a problem-dependent transition matrix. This insight leads to a novel algorithm for solving the maximum entropy exploration problem, EVE (EigenVector-based Exploration), which avoids explicit rollouts and distribution estimation, instead computing the solution through iterative updates, similar to a value-based approach. To address the original unregularized objective, we employ a posterior-policy iteration (PPI) approach, which monotonically improves the entropy and converges in value. We prove convergence of EVE under standard assumptions and demonstrate empirically that it efficiently produces policies with high steady-state entropy, achieving competitive exploration performance relative to rollout-based baselines in deterministic grid-world environments.\n\n---\n\n## 121. Thermodynamics of Reinforcement Learning Curricula\n- **Authors**: Jacob Adamczyk , Juan Sebastian Rojas , Rahul V. Kulkarni\n- **URL**: [https://arxiv.org/abs/2603.12324](https://arxiv.org/abs/2603.12324)\n- **Abstract**:\n> Connections between statistical mechanics and machine learning have repeatedly proven fruitful, providing insight into optimization, generalization, and representation learning. In this work, we follow this tradition by leveraging results from non-equilibrium thermodynamics to formalize curriculum learning in reinforcement learning (RL). In particular, we propose a geometric framework for RL by interpreting reward parameters as coordinates on a task manifold. We show that, by minimizing the excess thermodynamic work, optimal curricula correspond to geodesics in this task space. As an application of this framework, we provide an algorithm, \"MEW\" (Minimum Excess Work), to derive a principled schedule for temperature annealing in maximum-entropy RL.\n\n---\n\n## 122. VQQA: An Agentic Approach for Video Evaluation and Quality Improvement\n- **Authors**: Yiwen Song , Tomas Pfister , Yale Song\n- **URL**: [https://arxiv.org/abs/2603.12310](https://arxiv.org/abs/2603.12310)\n- **Abstract**:\n> Despite rapid advancements in video generation models, aligning their outputs with complex user intent remains challenging. Existing test-time optimization methods are typically either computationally expensive or require white-box access to model internals. To address this, we present VQQA (Video Quality Question Answering), a unified, multi-agent framework generalizable across diverse input modalities and video generation tasks. By dynamically generating visual questions and using the resulting Vision-Language Model (VLM) critiques as semantic gradients, VQQA replaces traditional, passive evaluation metrics with human-interpretable, actionable feedback. This enables a highly efficient, closed-loop prompt optimization process via a black-box natural language interface. Extensive experiments demonstrate that VQQA effectively isolates and resolves visual artifacts, substantially improving generation quality in just a few refinement steps. Applicable to both text-to-video (T2V) and image-to-video (I2V) tasks, our method achieves absolute improvements of +11.57% on T2V-CompBench and +8.43% on VBench2 over vanilla generation, significantly outperforming state-of-the-art stochastic search and prompt optimization techniques.\n\n---\n\n## 123. HCP-DCNet: A Hierarchical Causal Primitive Dynamic Composition Network for Self-Improving Causal Understanding\n- **Authors**: Ming Lei , Shufan Wu , Christophe Baehr\n- **URL**: [https://arxiv.org/abs/2603.12305](https://arxiv.org/abs/2603.12305)\n- **Abstract**:\n> The ability to understand and reason about cause and effect -- encompassing interventions, counterfactuals, and underlying mechanisms -- is a cornerstone of robust artificial intelligence. While deep learning excels at pattern recognition, it fundamentally lacks a model of causality, making systems brittle under distribution shifts and unable to answer ``what-if'' questions. This paper introduces the \\emph{Hierarchical Causal Primitive Dynamic Composition Network (HCP-DCNet)}, a unified framework that bridges continuous physical dynamics with discrete symbolic causal inference. Departing from monolithic representations, HCP-DCNet decomposes causal scenes into reusable, typed \\emph{causal primitives} organized into four abstraction layers: physical, functional, event, and rule. A dual-channel routing network dynamically composes these primitives into task-specific, fully differentiable \\emph{Causal Execution Graphs (CEGs)}. Crucially, the system employs a \\emph{causal-intervention-driven meta-evolution} strategy, enabling autonomous self-improvement through a constrained Markov decision process. We establish rigorous theoretical guarantees, including type-safe composition, routing convergence, and universal approximation of causal dynamics. Extensive experiments across simulated physical and social environments demonstrate that HCP-DCNet significantly outperforms state-of-the-art baselines in causal discovery, counterfactual reasoning, and compositional generalization. This work provides a principled, scalable, and interpretable architecture for building AI systems with human-like causal abstraction and continual self-refinement capabilities.\n\n---\n\n## 124. A Geometrically-Grounded Drive for MDL-Based Optimization in Deep Learning\n- **Authors**: Ming Lei , Shufan Wu , Christophe Baehr\n- **URL**: [https://arxiv.org/abs/2603.12304](https://arxiv.org/abs/2603.12304)\n- **Abstract**:\n> This paper introduces a novel optimization framework that fundamentally integrates the Minimum Description Length (MDL) principle into the training dynamics of deep neural networks. Moving beyond its conventional role as a model selection criterion, we reformulate MDL as an active, adaptive driving force within the optimization process itself. The core of our method is a geometrically-grounded cognitive manifold whose evolution is governed by a \\textit{coupled Ricci flow}, enriched with a novel \\textit{MDL Drive} term derived from first principles. This drive, modulated by the task-loss gradient, creates a seamless harmony between data fidelity and model simplification, actively compressing the internal representation during training. We establish a comprehensive theoretical foundation, proving key properties including the monotonic decrease of description length (Theorem~\\ref{thm:convergence}), a finite number of topological phase transitions via a geometric surgery protocol (Theorems~\\ref{thm:surgery}, \\ref{thm:ultimate_fate}), and the emergence of universal critical behavior (Theorem~\\ref{thm:universality}). Furthermore, we provide a practical, computationally efficient algorithm with $O(N \\log N)$ per-iteration complexity (Theorem~\\ref{thm:complexity}), alongside guarantees for numerical stability (Theorem~\\ref{thm:stability}) and exponential convergence under convexity assumptions (Theorem~\\ref{thm:convergence_rate}). Empirical validation on synthetic regression and classification tasks confirms the theoretical predictions, demonstrating the algorithm's efficacy in achieving robust generalization and autonomous model simplification. This work provides a principled path toward more autonomous, generalizable, and interpretable AI systems by unifying geometric deep learning with information-theoretic principles.\n\n---\n\n## 125. Global Evolutionary Steering: Refining Activation Steering Control via Cross-Layer Consistency\n- **Authors**: Xinyan Jiang , Wenjing Yu , Di Wang , Lijie Hu\n- **URL**: [https://arxiv.org/abs/2603.12298](https://arxiv.org/abs/2603.12298)\n- **Abstract**:\n> Activation engineering enables precise control over Large Language Models (LLMs) without the computational cost of fine-tuning. However, existing methods deriving vectors from static activation differences are susceptible to high-dimensional noise and layer-wise semantic drift, often capturing spurious correlations rather than the target intent. To address this, we propose Global Evolutionary Refined Steering (GER-steer), a training-free framework that grounded in the geometric stability of the network's representation evolution. GER-steer exploits this global signal to rectify raw steering vectors, effectively decoupling robust semantic intent from orthogonal artifacts. Extensive evaluations confirm that GER-steer consistently outperforms baselines, delivering superior efficacy and generalization without layer-specific tuning, establishing a universal solution for reliable model alignment.\n\n---\n\n## 126. Synthetic Data Generation for Brain-Computer Interfaces: Overview, Benchmarking, and Future Directions\n- **Authors**: Ziwei Wang , Zhentao He , Xingyi He , Hongbin Wang , Tianwang Jia , Jingwei Luo , Siyang Li , Xiaoqing Chen , Dongrui Wu\n- **URL**: [https://arxiv.org/abs/2603.12296](https://arxiv.org/abs/2603.12296)\n- **Abstract**:\n> Deep learning has achieved transformative performance across diverse domains, largely driven by the large-scale, high-quality training data. In contrast, the development of brain-computer interfaces (BCIs) is fundamentally constrained by the limited, heterogeneous, and privacy-sensitive neural recordings. Generating synthetic yet physiologically plausible brain signals has therefore emerged as a compelling way to mitigate data scarcity and enhance model capacity. This survey provides a comprehensive review of brain signal generation for BCIs, covering methodological taxonomies, benchmark experiments, evaluation metrics, and key applications. We systematically categorize existing generative algorithms into four types: knowledge-based, feature-based, model-based, and translation-based approaches. Furthermore, we benchmark existing brain signal generation approaches across four representative BCI paradigms to provide an objective performance comparison. Finally, we discuss the potentials and challenges of current generation approaches and prospect future research on accurate, data-efficient, and privacy-aware BCI systems. The benchmark codebase is publicized at this https URL .\n\n---\n\n## 127. Detecting Miscitation on the Scholarly Web through LLM-Augmented Text-Rich Graph Learning\n- **Authors**: Huidong Wu , Haojia Xiang , Jingtong Gao , Xiangyu Zhao , Dengsheng Wu , Jianping Li\n- **URL**: [https://arxiv.org/abs/2603.12290](https://arxiv.org/abs/2603.12290)\n- **Abstract**:\n> Scholarly web is a vast network of knowledge connected by citations. However, this system is increasingly compromised by miscitation, where references do not support or even contradict the claims they are cited for. Current miscitation detection methods, which primarily rely on semantic similarity or network anomalies, struggle to capture the nuanced relationship between a citation's context and its place in the wider network. While large language models (LLMs) offer powerful capabilities in semantic reasoning for this task, their deployment is hindered by hallucination risks and high computational costs. In this work, we introduce LLM-Augmented Graph Learning-based Miscitation Detector (LAGMiD), a novel framework that leverages LLMs for deep semantic reasoning over citation graphs and distills this knowledge into graph neural networks (GNNs) for efficient and scalable miscitation detection. Specifically, LAGMiD introduces an evidence-chain reasoning mechanism, which uses chain-of-thought prompting, to perform multi-hop citation tracing and assess semantic fidelity. To reduce LLM inference costs, we design a knowledge distillation method aligning GNN embeddings with intermediate LLM reasoning states. A collaborative learning strategy further routes complex cases to the LLM while optimizing the GNN for structure-based generalization. Experiments on three real-world benchmarks show that LAGMiD achieves state-of-the-art miscitation detection with significantly reduced inference cost.\n\n---\n\n## 128. From Garbage to Gold: A Data-Architectural Theory of Predictive Robustness\n- **Authors**: Terrence J. Lee-St. John , Jordan L. Lawson , Bartlomiej Piechowski-Jozwiak\n- **URL**: [https://arxiv.org/abs/2603.12288](https://arxiv.org/abs/2603.12288)\n- **Abstract**:\n> Tabular machine learning presents a paradox: modern models achieve state-of-the-art performance using high-dimensional (high-D), collinear, error-prone data, defying the \"Garbage In, Garbage Out\" mantra. To help resolve this, we synthesize principles from Information Theory, Latent Factor Models, and Psychometrics, clarifying that predictive robustness arises not solely from data cleanliness, but from the synergy between data architecture and model capacity. Partitioning predictor-space \"noise\" into \"Predictor Error\" and \"Structural Uncertainty\" (informational deficits from stochastic generative mappings), we prove that leveraging high-D sets of error-prone predictors asymptotically overcomes both types of noise, whereas cleaning a low-D set is fundamentally bounded by Structural Uncertainty. We demonstrate why \"Informative Collinearity\" (dependencies from shared latent causes) enhances reliability and convergence efficiency, and explain why increased dimensionality reduces the latent inference burden, enabling feasibility with finite samples. To address practical constraints, we propose \"Proactive Data-Centric AI\" to identify predictors that enable robustness efficiently. We also derive boundaries for Systematic Error Regimes and show why models that absorb \"rogue\" dependencies can mitigate assumption violations. Linking latent architecture to Benign Overfitting, we offer a first step towards a unified view of robustness to Outcome Error and predictor-space noise, while also delineating when traditional DCAI's focus on label cleaning remains powerful. By redefining data quality from item-level perfection to portfolio-level architecture, we provide a theoretical rationale for \"Local Factories\" -- learning from live, uncurated enterprise \"data swamps\" -- supporting a deployment paradigm shift from \"Model Transfer\" to \"Methodology Transfer'' to overcome static generalizability limitations.\n\n---\n\n## 129. The DIME Architecture: A Unified Operational Algorithm for Neural Representation, Dynamics, Control and Integration\n- **Authors**: Ionel Cristian Vladu , Nicu Bizdoaca , Ionica Pirici , Tudor-Adrian Balseanu , Eduard Nicusor Bondoc\n- **URL**: [https://arxiv.org/abs/2603.12286](https://arxiv.org/abs/2603.12286)\n- **Abstract**:\n> Modern neuroscience has accumulated extensive evidence on perception, memory, prediction, valuation, and consciousness, yet still lacks an explicit operational architecture capable of integrating these phenomena within a unified computational framework. Existing theories address specific aspects of neural function: predictive coding and active inference emphasize hierarchical inference and prediction error minimization; engram theories explain memory through distributed cell assemblies; neuromodulatory accounts focus on value-dependent regulation of plasticity and behaviour; and global workspace or large-scale network models investigate mechanisms underlying conscious access. Despite their explanatory power, these approaches remain only partially integrated at the architectural level. This work introduces DIME (Detect-Integrate-Mark-Execute), a neural architecture organizing perception, memory, valuation, and conscious access within a common operational cycle. The framework includes four interacting components: engrams, distributed recurrent neural structures supporting multiple activation trajectories; execution threads, spatiotemporal trajectories implementing neural processes; marker systems, neuromodulatory and limbic mechanisms regulating gain, plasticity, and trajectory selection; and hyperengrams, large-scale integrative states associated with operational conscious access. The framework is consistent with empirical evidence from hippocampal indexing, recurrent cortical processing, replay phenomena, large-scale network integration, and neuromodulatory regulation. Formulated at an abstract computational level, DIME may also inform artificial intelligence and robotics by providing an architectural template in which representation, valuation, and temporal sequencing emerge from a unified mechanism. An extended theoretical exposition is available in a companion monograph on Zenodo.\n\n---\n\n## 130. Predictive Analytics for Foot Ulcers Using Time-Series Temperature and Pressure Data\n- **Authors**: Md Tanvir Hasan Turja\n- **URL**: [https://arxiv.org/abs/2603.12278](https://arxiv.org/abs/2603.12278)\n- **Abstract**:\n> Diabetic foot ulcers (DFUs) are a severe complication of diabetes, often resulting in significant morbidity. This paper presents a predictive analytics framework utilizing time-series data captured by wearable foot sensors -- specifically NTC thin-film thermocouples for temperature measurement and FlexiForce pressure sensors for plantar load monitoring. Data was collected from healthy subjects walking on an instrumented pathway. Unsupervised machine learning algorithms, Isolation Forest and K-Nearest Neighbors (KNN), were applied to detect anomalies that may indicate early ulcer risk. Through rigorous data preprocessing and targeted feature engineering, physiologic patterns were extracted to identify subtle changes in foot temperature and pressure. Results demonstrate Isolation Forest is sensitive to micro-anomalies, while KNN is effective in flagging extreme deviations, albeit at a higher false-positive rate. Strong correlations between temperature and pressure readings support combined sensor monitoring for improved predictive accuracy. These findings provide a basis for real-time diabetic foot health surveillance, aiming to facilitate earlier intervention and reduce DFU incidence.\n\n---\n\n## 131. Prompt Injection as Role Confusion\n- **Authors**: Charles Ye , Jasmine Cui , Dylan Hadfield-Menell\n- **URL**: [https://arxiv.org/abs/2603.12277](https://arxiv.org/abs/2603.12277)\n- **Abstract**:\n> Language models remain vulnerable to prompt injection attacks despite extensive safety training. We trace this failure to role confusion: models infer roles from how text is written, not where it comes from. We design novel role probes to capture how models internally identify \"who is speaking.\" These reveal why prompt injection works: untrusted text that imitates a role inherits that role's authority. We test this insight by injecting spoofed reasoning into user prompts and tool outputs, achieving average success rates of 60% on StrongREJECT and 61% on agent exfiltration, across multiple open- and closed-weight models with near-zero baselines. Strikingly, the degree of internal role confusion strongly predicts attack success before generation begins. Our findings reveal a fundamental gap: security is defined at the interface but authority is assigned in latent space. More broadly, we introduce a unifying, mechanistic framework for prompt injection, demonstrating that diverse prompt-injection attacks exploit the same underlying role-confusion mechanism.\n\n---\n\n## 132. Aligning Language Models from User Interactions\n- **Authors**: Thomas Kleine Buening , Jonas Hübotter , Barna Pásztor , Idan Shenfeld , Giorgia Ramponi , Andreas Krause\n- **URL**: [https://arxiv.org/abs/2603.12273](https://arxiv.org/abs/2603.12273)\n- **Abstract**:\n> Multi-turn user interactions are among the most abundant data produced by language models, yet we lack effective methods to learn from them. While typically discarded, these interactions often contain useful information: follow-up user messages may indicate that a response was incorrect, failed to follow an instruction, or did not align with the user's preferences. Importantly, language models are already able to make use of this information in context. After observing a user's follow-up, the same model is often able to revise its behavior. We leverage this ability to propose a principled and scalable method for learning directly from user interactions through self-distillation. By conditioning the model on the user's follow-up message and comparing the resulting token distribution with the original policy, we obtain a target for updating the policy that captures how the model's behavior changes in hindsight. We then distill this hindsight distribution back into the current policy. Remarkably, we show that training on real-world user conversations from WildChat improves language models across standard alignment and instruction-following benchmarks, without regressing other capabilities. The same mechanism enables personalization, allowing models to continually adapt to individual users through interaction without explicit feedback. Our results demonstrate that raw user interactions that arise naturally during deployment enable alignment, personalization, and continual adaptation.\n\n---\n\n## 133. Diagnosing Retrieval Bias Under Multiple In-Context Knowledge Updates in Large Language Models\n- **Authors**: Boyu Qiao , Sean Guo , Xian Yang , Kun Li , Wei Zhou , Songlin Hu , Yunya Song\n- **URL**: [https://arxiv.org/abs/2603.12271](https://arxiv.org/abs/2603.12271)\n- **Abstract**:\n> LLMs are widely used in knowledge-intensive tasks where the same fact may be revised multiple times within context. Unlike prior work focusing on one-shot updates or single conflicts, multi-update scenarios contain multiple historically valid versions that compete at retrieval, yet remain underexplored. This challenge resembles the AB-AC interference paradigm in cognitive psychology: when the same cue A is successively associated with B and C, the old and new associations compete during retrieval, leading to bias. Inspired by this, we introduce a Dynamic Knowledge Instance (DKI) evaluation framework, modeling multi-updates of the same fact as a cue paired with a sequence of updated values, and assess models via endpoint probing of the earliest (initial) and latest (current) states. Across diverse LLMs, we observe that retrieval bias intensifies as updates increase, earliest-state accuracy stays high while latest-state accuracy drops substantially. Diagnostic analyses of attention, hidden-state similarity, and output logits further reveal that these signals become flatter and weakly discriminative on errors, providing little stable basis for identifying the latest update. Finally, cognitively inspired heuristic intervention strategies yield only modest gains and do not eliminate the bias. Our results reveal a persistent challenge in tracking and following knowledge updates in long contexts.\n\n---\n\n## 134. Task-Specific Knowledge Distillation via Intermediate Probes\n- **Authors**: Ryan Brown , Chris Russell\n- **URL**: [https://arxiv.org/abs/2603.12270](https://arxiv.org/abs/2603.12270)\n- **Abstract**:\n> Knowledge distillation from large language models (LLMs) assumes that the teacher's output distribution is a high-quality training signal. On reasoning tasks, this assumption is frequently violated. A model's intermediate representations may encode the correct answer, yet this information is lost or distorted through the vocabulary projection, where prompt formatting and answer-token choices creates brittle, noisy outputs. We introduce \\method{}, a distillation framework that bypasses this bottleneck by training lightweight probes on frozen teacher hidden states and using the probe's predictions, rather than output logits, as supervision for student training. This simple change yields consistent improvements across four reasoning benchmarks (AQuA-RAT, ARC Easy/Challenge, and MMLU), with gains most pronounced under limited data. Probes trained on intermediate representations provide cleaner labels than the teacher's own outputs, effectively denoising the distillation signal. \\method{} requires no architectural changes to student or teacher, is architecture-agnostic, and adds minimal compute since probe training is cheap and teacher representations can be cached. By exploiting internal representations, \\method{} enables practitioners to extract more value from large teacher models without additional training data or architectural complexity.\n\n---\n\n## 135. DART: Input-Difficulty-AwaRe Adaptive Threshold for Early-Exit DNNs\n- **Authors**: Parth Patne , Mahdi Taheri , Christian Herglotz , Maksim Jenihhin , Milos Krstic , Michael Hübner\n- **URL**: [https://arxiv.org/abs/2603.12269](https://arxiv.org/abs/2603.12269)\n- **Abstract**:\n> Early-exit deep neural networks enable adaptive inference by terminating computation when sufficient confidence is achieved, reducing cost for edge AI accelerators in resource-constrained settings. Existing methods, however, rely on suboptimal exit policies, ignore input difficulty, and optimize thresholds independently. This paper introduces DART (Input-Difficulty-Aware Adaptive Threshold), a framework that overcomes these limitations. DART introduces three key innovations: (1) a lightweight difficulty estimation module that quantifies input complexity with minimal computational overhead, (2) a joint exit policy optimization algorithm based on dynamic programming, and (3) an adaptive coefficient management system. Experiments on diverse DNN benchmarks (AlexNet, ResNet-18, VGG-16) demonstrate that DART achieves up to \\textbf{3.3$\\times$} speedup, \\textbf{5.1$\\times$} lower energy, and up to \\textbf{42\\%} lower average power compared to static networks, while preserving competitive accuracy. Extending DART to Vision Transformers (LeViT) yields power (5.0$\\times$) and execution-time (3.6$\\times$) gains but also accuracy loss (up to 17 percent), underscoring the need for transformer-specific early-exit mechanisms. We further introduce the Difficulty-Aware Efficiency Score (DAES), a novel multi-objective metric, under which DART achieves up to a 14.8 improvement over baselines, highlighting superior accuracy, efficiency, and robustness trade-offs.\n\n---\n\n{% endraw %}\n"}, "ee14c4803bea05a7c9ea5c3b5aaa5dec510a94a3": {"t": 1785770990.8442993, "n": 229864, "h": "ce4ab4e5e386c926", "c": "# 전체 AI 논문 - 2026-03-17\n\n## 1. Semantic Invariance in Agentic AI\n- **Authors**: I. de Zarzà , J. de Curtò , Jordi Cabot , Pietro Manzoni , Carlos T. Calafate\n- **URL**: [https://arxiv.org/abs/2603.13173](https://arxiv.org/abs/2603.13173)\n- **Abstract**:\n> Large Language Models (LLMs) increasingly serve as autonomous reasoning agents in decision support, scientific problem-solving, and multi-agent coordination systems. However, deploying LLM agents in consequential applications requires assurance that their reasoning remains stable under semantically equivalent input variations, a property we term semantic invariance. Standard benchmark evaluations, which assess accuracy on fixed, canonical problem formulations, fail to capture this critical reliability dimension. To address this shortcoming, in this paper we present a metamorphic testing framework for systematically assessing the robustness of LLM reasoning agents, applying eight semantic-preserving transformations (identity, paraphrase, fact reordering, expansion, contraction, academic context, business context, and contrastive formulation) across seven foundation models spanning four distinct architectural families: Hermes (70B, 405B), Qwen3 (30B-A3B, 235B-A22B), DeepSeek-R1, and gpt-oss (20B, 120B). Our evaluation encompasses 19 multi-step reasoning problems across eight scientific domains. The results reveal that model scale does not predict robustness: the smaller Qwen3-30B-A3B achieves the highest stability (79.6% invariant responses, semantic similarity 0.91), while larger models exhibit greater fragility.\n\n---\n\n## 2. Developing and evaluating a chatbot to support maternal health care\n- **Authors**: Smriti Jha , Vidhi Jain , Jianyu Xu , Grace Liu , Sowmya Ramesh , Jitender Nagpal , Gretchen Chapman , Benjamin Bellows , Siddhartha Goyal , Aarti Singh , Bryan Wilder\n- **URL**: [https://arxiv.org/abs/2603.13168](https://arxiv.org/abs/2603.13168)\n- **Abstract**:\n> The ability to provide trustworthy maternal health information using phone-based chatbots can have a significant impact, particularly in low-resource settings where users have low health literacy and limited access to care. However, deploying such systems is technically challenging: user queries are short, underspecified, and code-mixed across languages, answers require regional context-specific grounding, and partial or missing symptom context makes safe routing decisions difficult. We present a chatbot for maternal health in India developed through a partnership between academic researchers, a health tech company, a public health nonprofit, and a hospital. The system combines (1) stage-aware triage, routing high-risk queries to expert templates, (2) hybrid retrieval over curated maternal/newborn guidelines, and (3) evidence-conditioned generation from an LLM. Our core contribution is an evaluation workflow for high-stakes deployment under limited expert supervision. Targeting both component-level and end-to-end testing, we introduce: (i) a labeled triage benchmark (N=150) achieving 86.7% emergency recall, explicitly reporting the missed-emergency vs. over-escalation trade-off; (ii) a synthetic multi-evidence retrieval benchmark (N=100) with chunk-level evidence labels; (iii) LLM-as-judge comparison on real queries (N=781) using clinician-codesigned criteria; and (iv) expert validation. Our findings show that trustworthy medical assistants in multilingual, noisy settings require defense-in-depth design paired with multi-method evaluation, rather than any single model and evaluation method choice.\n\n---\n\n## 3. When Right Meets Wrong: Bilateral Context Conditioning with Reward-Confidence Correction for GRPO\n- **Authors**: Yu Li , Tian Lan , Zhengling Qi\n- **URL**: [https://arxiv.org/abs/2603.13134](https://arxiv.org/abs/2603.13134)\n- **Abstract**:\n> Group Relative Policy Optimization (GRPO) has emerged as an effective method for training reasoning models. While it computes advantages based on group mean, GRPO treats each output as an independent sample during the optimization and overlooks a vital structural signal: the natural contrast between correct and incorrect solutions within the same group, thus ignoring the rich, comparative data that could be leveraged by explicitly pitting successful reasoning traces against failed ones. To capitalize on this, we present a contrastive reformulation of GRPO, showing that the GRPO objective implicitly maximizes the margin between the policy ratios of correct and incorrect samples. Building on this insight, we propose Bilateral Context Conditioning (BICC), a mechanism that allows the model to cross-reference successful and failed reasoning traces during the optimization, enabling a direct information flow across samples. We further introduce Reward-Confidence Correction (RCC) to stabilize training by dynamically adjusts the advantage baseline in GRPO using reward-confidence covariance derived from the first-order approximation of the variance-minimizing estimator. Both mechanisms require no additional sampling or auxiliary models and can be adapted to all GRPO variants. Experiments on mathematical reasoning benchmarks demonstrate consistent improvements across comprehensive models and algorithms. Code is available at \\href{ this https URL }{ this https URL }.\n\n---\n\n## 4. Steve-Evolving: Open-World Embodied Self-Evolution via Fine-Grained Diagnosis and Dual-Track Knowledge Distillation\n- **Authors**: Zhengwei Xie , Zhisheng Chen , Ziyan Weng , Tingyu Wu , Chenglong Li , Vireo Zhang , Kun Wang\n- **URL**: [https://arxiv.org/abs/2603.13131](https://arxiv.org/abs/2603.13131)\n- **Abstract**:\n> Open-world embodied agents must solve long-horizon tasks where the main bottleneck is not single-step planning quality but how interaction experience is organized and evolved. To this end, we present Steve-Evolving, a non-parametric self-evolving framework that tightly couples fine-grained execution diagnosis with dual-track knowledge distillation in a closed loop. The method follows three phases: Experience Anchoring, Experience Distillation, and Knowledge-Driven Closed-Loop Control. In detail, Experience Anchoring solidifies each subgoal attempt into a structured experience tuple with a fixed schema (pre-state, action, diagnosis-result, and post-state) and organizes it in a three-tier experience space with multi-dimensional indices (e.g., condition signatures, spatial hashing, and semantic tags) plus rolling summarization for efficient and auditable recall. To ensure sufficient information density for attribution, the execution layer provides compositional diagnosis signals beyond binary outcomes, including state-difference summaries, enumerated failure causes, continuous indicators, and stagnation/loop detection. Moreover, successful trajectories of Experience Distillation are generalized into reusable skills with explicit preconditions and verification criteria, while failures are distilled into executable guardrails that capture root causes and forbid risky operations at both subgoal and task granularities. Besides, Knowledge-Driven Closed-Loop Control retrieved skills and guardrails are injected into an LLM planner, and diagnosis-triggered local replanning updates the active constraints online, forming a continual evolution process without any model parameter updates. Experiments on the long-horizon suite of Minecraft MCU demonstrate consistent improvements over static-retrieval baselines.\n\n---\n\n## 5. Beyond Final Answers: CRYSTAL Benchmark for Transparent Multimodal Reasoning Evaluation\n- **Authors**: Wayner Barrios , SouYoung Jin\n- **URL**: [https://arxiv.org/abs/2603.13099](https://arxiv.org/abs/2603.13099)\n- **Abstract**:\n> We introduce CRYSTAL (Clear Reasoning via Yielded Steps, Traceability, and Logic), a diagnostic benchmark with 6,372 instances that evaluates multimodal reasoning through verifiable intermediate steps. We propose two complementary metrics: Match F1, which scores step-level precision and recall via semantic similarity matching, and Ordered Match F1, which further penalizes disordered reasoning chains. References are constructed through a Delphi-inspired pipeline in which four independent MLLMs generate trajectories, which are then aggregated via semantic clustering and validated through human quality gates. Evaluation of 20 MLLMs, including commercial frontier systems not used during benchmark construction, reveals systematic failures that are invisible to answer accuracy: universal cherry-picking (precision far exceeds recall), non-monotonic scaling trade-offs, and disordered reasoning in which no competitive model preserves more than 60% of matched steps in the correct order. Beyond evaluation, we propose the Causal Process Reward (CPR), a multiplicative reward that couples answer correctness with step-level alignment, and CPR-Curriculum, which progressively increases reasoning difficulty during training. CPR-Curriculum achieves a 32% improvement in Match F1 via GRPO where additive reward strategies fail, improving reasoning without manual step annotation.\n\n---\n\n## 6. Structured Distillation for Personalized Agent Memory: 11x Token Reduction with Retrieval Preservation\n- **Authors**: Sydney Lewis\n- **URL**: [https://arxiv.org/abs/2603.13017](https://arxiv.org/abs/2603.13017)\n- **Abstract**:\n> Long conversations with an AI agent create a simple problem for one user: the history is useful, but carrying it verbatim is expensive. We study personalized agent memory: one user's conversation history with an agent, distilled into a compact retrieval layer for later search. Each exchange is compressed into a compound object with four fields (exchange_core, specific_context, thematic room_assignments, and regex-extracted files_touched). The searchable distilled text averages 38 tokens per exchange. Applied to 4,182 conversations (14,340 exchanges) from 6 software engineering projects, the method reduces average exchange length from 371 to 38 tokens, yielding 11x compression. We evaluate whether personalized recall survives that compression using 201 recall-oriented queries, 107 configurations spanning 5 pure and 5 cross-layer search modes, and 5 LLM graders (214,519 consensus-graded query-result pairs). The best pure distilled configuration reaches 96% of the best verbatim MRR (0.717 vs 0.745). Results are mechanism-dependent. All 20 vector search configurations remain non-significant after Bonferroni correction, while all 20 BM25 configurations degrade significantly (effect sizes |d|=0.031-0.756). The best cross-layer setup slightly exceeds the best pure verbatim baseline (MRR 0.759). Structured distillation compresses single-user agent memory without uniformly sacrificing retrieval quality. At 1/11 the context cost, thousands of exchanges fit within a single prompt while the verbatim source remains available for drill-down. We release the implementation and analysis pipeline as open-source software.\n\n---\n\n## 7. Efficient and Interpretable Multi-Agent LLM Routing via Ant Colony Optimization\n- **Authors**: Xudong Wang , Chaoning Zhang , Jiaquan Zhang , Chenghao Li , Qigan Sun , Sung-Ho Bae , Peng Wang , Ning Xie , Jie Zou , Yang Yang , Hengtao Shen\n- **URL**: [https://arxiv.org/abs/2603.12933](https://arxiv.org/abs/2603.12933)\n- **Abstract**:\n> Large Language Model (LLM)-driven Multi-Agent Systems (MAS) have demonstrated strong capability in complex reasoning and tool use, and heterogeneous agent pools further broaden the quality--cost trade-off space. Despite these advances, real-world deployment is often constrained by high inference cost, latency, and limited transparency, which hinders scalable and efficient routing. Existing routing strategies typically rely on expensive LLM-based selectors or static policies, and offer limited controllability for semantic-aware routing under dynamic loads and mixed intents, often resulting in unstable performance and inefficient resource utilization. To address these limitations, we propose AMRO-S, an efficient and interpretable routing framework for Multi-Agent Systems (MAS). AMRO-S models MAS routing as a semantic-conditioned path selection problem, enhancing routing performance through three key mechanisms: First, it leverages a supervised fine-tuned (SFT) small language model for intent inference, providing a low-overhead semantic interface for each query; second, it decomposes routing memory into task-specific pheromone specialists, reducing cross-task interference and optimizing path selection under mixed workloads; finally, it employs a quality-gated asynchronous update mechanism to decouple inference from learning, optimizing routing without increasing latency. Extensive experiments on five public benchmarks and high-concurrency stress tests demonstrate that AMRO-S consistently improves the quality--cost trade-off over strong routing baselines, while providing traceable routing evidence through structured pheromone patterns.\n\n---\n\n## 8. ODRL Policy Comparison Through Normalisation\n- **Authors**: Jaime Osvaldo Salas , Paolo Pareti , George Konstantinidis\n- **URL**: [https://arxiv.org/abs/2603.12926](https://arxiv.org/abs/2603.12926)\n- **Abstract**:\n> The ODRL language has become the standard for representing policies and regulations for digital rights. However its complexity is a barrier to its usage, which has caused many related theoretical and practical works to focus on different, and not interoperable, fragments of ODRL. Moreover, semantically equivalent policies can be expressed in numerous different ways, which makes comparing them and processing them harder. Building on top of a recently defined semantics, we tackle these problems by proposing an approach that involves a parametrised normalisation of ODRL policies into its minimal components which reformulates policies with permissions and prohibitions into policies with permissions exclusively, and simplifies complex logic constraints into simple ones. We provide algorithms to compute a normal form for ODRL policies and simplifying numerical and symbolic constraints. We prove that these algorithms preserve the semantics of policies, and analyse the size complexity of the result, which is exponential on the number of attributes and linear on the number of unique values for these attributes. We show how this makes complex policies representable in more basic fragments of ODRL, and how it reduces the problem of policy comparison to the simpler problem of checking if two rules are identical.\n\n---\n\n## 9. Context is all you need: Towards autonomous model-based process design using agentic AI in flowsheet simulations\n- **Authors**: Pascal Schäfer , Lukas J. Krinke , Martin Wlotzka , Norbert Asprion\n- **URL**: [https://arxiv.org/abs/2603.12813](https://arxiv.org/abs/2603.12813)\n- **Abstract**:\n> Agentic AI systems integrating large language models (LLMs) with reasoning and tooluse capabilities are transforming various domains - in particular, software development. In contrast, their application in chemical process flowsheet modelling remains largely unexplored. In this work, we present an agentic AI framework that delivers assistance in an industrial flowsheet simulation environment. To this end, we show the capabilities of GitHub Copilot (GitHub, Inc., 2026), when using state-of-the-art LLMs, such as Claude Opus 4.6 (Anthropic, PBC, 2026), to generate valid syntax for our in-house process modelling tool Chemasim using the technical documentation and a few commented examples as context. Based on this, we develop a multi-agent system that decomposes process development tasks with one agent solving the abstract problem using engineering knowledge and another agent implementing the solution as Chemasim code. We demonstrate the effectiveness of our framework for typical flowsheet modelling examples, including (i) a reaction/separation process, (ii) a pressure-swing distillation, and (iii) a heteroazeotropic distillation including entrainer selection. Along these lines, we discuss current limitations of the framework and outline future research directions to further enhance its capabilities.\n\n---\n\n## 10. AI Model Modulation with Logits Redistribution\n- **Authors**: Zihan Wang , Zhongkui Ma , Xinguo Feng , Zhiyang Mei , Ethan Ma , Derui Wang , Minhui Xue , Guangdong Bai\n- **URL**: [https://arxiv.org/abs/2603.12755](https://arxiv.org/abs/2603.12755)\n- **Abstract**:\n> Large-scale models are typically adapted to meet the diverse requirements of model owners and users. However, maintaining multiple specialized versions of the model is inefficient. In response, we propose AIM, a novel model modulation paradigm that enables a single model to exhibit diverse behaviors to meet the specific end requirements. AIM enables two key modulation modes: utility and focus modulations. The former provides model owners with dynamic control over output quality to deliver varying utility levels, and the latter offers users precise control to shift model's focused input features. AIM introduces a logits redistribution strategy that operates in a training data-agnostic and retraining-free manner. We establish a formal foundation to ensure AIM's regulation capability, based on the statistical properties of logits ordering via joint probability distributions. Our evaluation confirms AIM's practicality and versatility for Al model modulation, with tasks spanning image classification, semantic segmentation and text generation, and prevalent architectures including ResNet, SegFormer and Llama.\n\n---\n\n## 11. ToolTree: Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning\n- **Authors**: Shuo Yang , Soyeon Caren Han , Yihao Ding , Shuhe Wang , Eduard Hoy\n- **URL**: [https://arxiv.org/abs/2603.12740](https://arxiv.org/abs/2603.12740)\n- **Abstract**:\n> Large Language Model (LLM) agents are increasingly applied to complex, multi-step tasks that require interaction with diverse external tools across various domains. However, current LLM agent tool planning methods typically rely on greedy, reactive tool selection strategies that lack foresight and fail to account for inter-tool dependencies. In this paper, we present ToolTree, a novel Monte Carlo tree search-inspired planning paradigm for tool planning. ToolTree explores possible tool usage trajectories using a dual-stage LLM evaluation and bidirectional pruning mechanism that enables the agent to make informed, adaptive decisions over extended tool-use sequences while pruning less promising branches before and after the tool execution. Empirical evaluations across both open-set and closed-set tool planning tasks on 4 benchmarks demonstrate that ToolTree consistently improves performance while keeping the highest efficiency, achieving an average gain of around 10\\% compared to the state-of-the-art planning paradigm.\n\n---\n\n## 12. On Using Machine Learning to Early Detect Catastrophic Failures in Marine Diesel Engines\n- **Authors**: Francesco Maione , Paolo Lino , Giuseppe Giannino , Guido Maione\n- **URL**: [https://arxiv.org/abs/2603.12733](https://arxiv.org/abs/2603.12733)\n- **Abstract**:\n> Catastrophic failures of marine engines imply severe loss of functionality and destroy or damage the systems irreversibly. Being sudden and often unpredictable events, they pose a severe threat to navigation, crew, and passengers. The abrupt nature makes early detection the only effective countermeasure. However, research has concentrated on modeling the gradual degradation of components, with limited attention to sudden and anomalous phenomena. This work proposes a new method for early detection of catastrophic failures. Based on real data from a failed engine, the approach evaluates the derivatives of the deviation between actual sensor readings and expected values of engine variables. Predictions are obtained by a Random Forest, which is the most suitable Machine Learning algorithm among the tested ones. Traditional methods focus on deviations of monitored signals, whereas the proposed approach employs the derivatives of the deviations to provide earlier indications of abnormal dynamics, and to alert that a rapid and dangerous event is breaking out within the system. The method allows the detection of anomalies before measurements reach critical thresholds and alarms are triggered, which is the common method in industry. Consequently, operators can be warned in advance and shut down the engine, then prevent damage and unexpected power loss. Moreover, they have the time to safely change the ship route and avoid potential obstacles. Simulation results conf irm the effectiveness of the proposed approach in anticipating occurrence of catastrophic failures. Validation on real-world data further reinforces the robustness and practical applicability of the method. It is worth noting that data acquisition to train the predictive algorithm is not a problem, since a Deep Learning-based data augmentation procedure is used.\n\n---\n\n## 13. AI Planning Framework for LLM-Based Web Agents\n- **Authors**: Orit Shahnovsky , Rotem Dror\n- **URL**: [https://arxiv.org/abs/2603.12710](https://arxiv.org/abs/2603.12710)\n- **Abstract**:\n> Developing autonomous agents for web-based tasks is a core challenge in AI. While Large Language Model (LLM) agents can interpret complex user requests, they often operate as black boxes, making it difficult to diagnose why they fail or how they plan. This paper addresses this gap by formally treating web tasks as sequential decision-making processes. We introduce a taxonomy that maps modern agent architectures to traditional planning paradigms: Step-by-Step agents to Breadth-First Search (BFS), Tree Search agents to Best-First Tree Search, and Full-Plan-in-Advance agents to Depth-First Search (DFS). This framework allows for a principled diagnosis of system failures like context drift and incoherent task decomposition. To evaluate these behaviors, we propose five novel evaluation metrics that assess trajectory quality beyond simple success rates. We support this analysis with a new dataset of 794 human-labeled trajectories from the WebArena benchmark. Finally, we validate our evaluation framework by comparing a baseline Step-by-Step agent against a novel Full-Plan-in-Advance implementation. Our results reveal that while the Step-by-Step agent aligns more closely with human gold trajectories (38% overall success), the Full-Plan-in-Advance agent excels in technical measures such as element accuracy (89%), demonstrating the necessity of our proposed metrics for selecting appropriate agent architectures based on specific application constraints.\n\n---\n\n## 14. Generating Expressive and Customizable Evals for Timeseries Data Analysis Agents with AgentFuel\n- **Authors**: Aadyaa Maddi , Prakhar Naval , Deepti Mande , Shane Duan , Muckai Girish , Vyas Sekar\n- **URL**: [https://arxiv.org/abs/2603.12483](https://arxiv.org/abs/2603.12483)\n- **Abstract**:\n> Across many domains (e.g., IoT, observability, telecommunications, cybersecurity), there is an emerging adoption of conversational data analysis agents that enable users to \"talk to your data\" to extract insights. Such data analysis agents operate on timeseries data models; e.g., measurements from sensors or events monitoring user clicks and actions in product analytics. We evaluate 6 popular data analysis agents (both open-source and proprietary) on domain-specific data and query types, and find that they fail on stateful and incident-specific queries. We observe two key expressivity gaps in existing evals: domain-customized datasets and domain-specific query types. To enable practitioners in such domains to generate customized and expressive evals for such timeseries data agents, we present AgentFuel. AgentFuel helps domain experts quickly create customized evals to perform end-to-end functional tests. We show that AgentFuel's benchmarks expose key directions for improvement in existing data agent frameworks. We also present anecdotal evidence that using AgentFuel can improve agent performance (e.g., with GEPA). AgentFuel benchmarks are available at this https URL .\n\n---\n\n## 15. Efficient Reasoning with Balanced Thinking\n- **Authors**: Yulin Li , Tengyao Tu , Li Ding , Junjie Wang , Huiling Zhen , Yixin Chen , Yong Li , Zhuotao Tian\n- **URL**: [https://arxiv.org/abs/2603.12372](https://arxiv.org/abs/2603.12372)\n- **Abstract**:\n> Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they often suffer from overthinking, expending redundant computational steps on simple problems, or underthinking, failing to explore sufficient reasoning paths despite inherent capabilities. These issues lead to inefficiencies and potential inaccuracies, limiting practical deployment in resource-constrained settings. Existing methods to mitigate overthinking, such as suppressing reflective keywords or adjusting reasoning length, may inadvertently induce underthinking, compromising accuracy. Therefore, we propose ReBalance, a training-free framework that achieves efficient reasoning with balanced thinking. ReBalance leverages confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence. By aggregating hidden states from a small-scale dataset into reasoning mode prototypes, we compute a steering vector to guide LRMs' reasoning trajectories. A dynamic control function modulates this vector's strength and direction based on real-time confidence, pruning redundancy during overthinking, and promoting exploration during underthinking. Extensive experiments conducted on four models ranging from 0.5B to 32B, and across nine benchmarks in math reasoning, general question answering, and coding tasks demonstrate that ReBalance effectively reduces output redundancy while improving accuracy, offering a general, training-free, and plug-and-play strategy for efficient and robust LRM deployment. Code is available at this https URL .\n\n---\n\n## 16. Context-Enriched Natural Language Descriptions of Vessel Trajectories\n- **Authors**: Kostas Patroumpas , Alexandros Troupiotis-Kapeliaris , Giannis Spiliopoulos , Panagiotis Betchavas , Dimitrios Skoutas , Dimitris Zissis , Nikos Bikakis\n- **URL**: [https://arxiv.org/abs/2603.12287](https://arxiv.org/abs/2603.12287)\n- **Abstract**:\n> We address the problem of transforming raw vessel trajectory data collected from AIS into structured and semantically enriched representations interpretable by humans and directly usable by machine reasoning systems. We propose a context-aware trajectory abstraction framework that segments noisy AIS sequences into distinct trips each consisting of clean, mobility-annotated episodes. Each episode is further enriched with multi-source contextual information, such as nearby geographic entities, offshore navigation features, and weather conditions. Crucially, such representations can support generation of controlled natural language descriptions using LLMs. We empirically examine the quality of such descriptions generated using several LLMs over AIS data along with open contextual features. By increasing semantic density and reducing spatiotemporal complexity, this abstraction can facilitate downstream analytics and enable integration with LLMs for higher-level maritime reasoning tasks.\n\n---\n\n## 17. PhysMoDPO: Physically-Plausible Humanoid Motion with Preference Optimization\n- **Authors**: Yangsong Zhang , Anujith Muraleedharan , Rikhat Akizhanov , Abdul Ahad Butt , Gül Varol , Pascal Fua , Fabio Pizzati , Ivan Laptev\n- **URL**: [https://arxiv.org/abs/2603.13228](https://arxiv.org/abs/2603.13228)\n- **Abstract**:\n> Recent progress in text-conditioned human motion generation has been largely driven by diffusion models trained on large-scale human motion data. Building on this progress, recent methods attempt to transfer such models for character animation and real robot control by applying a Whole-Body Controller (WBC) that converts diffusion-generated motions into executable trajectories. While WBC trajectories become compliant with physics, they may expose substantial deviations from original motion. To address this issue, we here propose PhysMoDPO, a Direct Preference Optimization framework. Unlike prior work that relies on hand-crafted physics-aware heuristics such as foot-sliding penalties, we integrate WBC into our training pipeline and optimize diffusion model such that the output of WBC becomes compliant both with physics and original text instructions. To train PhysMoDPO we deploy physics-based and task-specific rewards and use them to assign preference to synthesized trajectories. Our extensive experiments on text-to-motion and spatial control tasks demonstrate consistent improvements of PhysMoDPO in both physical realism and task-related metrics on simulated robots. Moreover, we demonstrate that PhysMoDPO results in significant improvements when applied to zero-shot motion transfer in simulation and for real-world deployment on a G1 humanoid robot.\n\n---\n\n## 18. Visual-ERM: Reward Modeling for Visual Equivalence\n- **Authors**: Ziyu Liu , Shengyuan Ding , Xinyu Fang , Xuanlang Dai , Penghui Yang , Jianze Liang , Jiaqi Wang , Kai Chen , Dahua Lin , Yuhang Zang\n- **URL**: [https://arxiv.org/abs/2603.13224](https://arxiv.org/abs/2603.13224)\n- **Abstract**:\n> Vision-to-code tasks require models to reconstruct structured visual inputs, such as charts, tables, and SVGs, into executable or structured representations with high visual fidelity. While recent Large Vision Language Models (LVLMs) achieve strong results via supervised fine-tuning, reinforcement learning remains challenging due to misaligned reward signals. Existing rewards either rely on textual rules or coarse visual embedding similarity, both of which fail to capture fine-grained visual discrepancies and are vulnerable to reward hacking. We propose Visual Equivalence Reward Model (Visual-ERM), a multimodal generative reward model that provides fine-grained, interpretable, and task-agnostic feedback to evaluate vision-to-code quality directly in the rendered visual space. Integrated into RL, Visual-ERM improves Qwen3-VL-8B-Instruct by +8.4 on chart-to-code and yields consistent gains on table and SVG parsing (+2.7, +4.1 on average), and further strengthens test-time scaling via reflection and revision. We also introduce VisualCritic-RewardBench (VC-RewardBench), a benchmark for judging fine-grained image-to-image discrepancies on structured visual data, where Visual-ERM at 8B decisively outperforms Qwen3-VL-235B-Instruct and approaches leading closed-source models. Our results suggest that fine-grained visual reward supervision is both necessary and sufficient for vision-to-code RL, regardless of task specificity.\n\n---\n\n## 19. From Experiments to Expertise: Scientific Knowledge Consolidation for AI-Driven Computational Research\n- **Authors**: Haonan Huang\n- **URL**: [https://arxiv.org/abs/2603.13191](https://arxiv.org/abs/2603.13191)\n- **Abstract**:\n> While large language models (LLMs) have transformed AI agents into proficient executors of computational materials science, performing a hundred simulations does not make a researcher. What distinguishes research from routine execution is the progressive accumulation of knowledge -- learning which approaches fail, recognizing patterns across systems, and applying understanding to new problems. However, the prevailing paradigm in AI-driven computational science treats each execution in isolation, largely discarding hard-won insights between runs. Here we present QMatSuite, an open-source platform closing this gap. Agents record findings with full provenance, retrieve knowledge before new calculations, and in dedicated reflection sessions correct erroneous findings and synthesize observations into cross-compound patterns. In benchmarks on a six-step quantum-mechanical simulation workflow, accumulated knowledge reduces reasoning overhead by 67% and improves accuracy from 47% to 3% deviation from literature -- and when transferred to an unfamiliar material, achieves 1% deviation with zero pipeline failures.\n\n---\n\n## 20. LLM Constitutional Multi-Agent Governance\n- **Authors**: J. de Curtò , I. de Zarzà\n- **URL**: [https://arxiv.org/abs/2603.13189](https://arxiv.org/abs/2603.13189)\n- **Abstract**:\n> Large Language Models (LLMs) can generate persuasive influence strategies that shift cooperative behavior in multi-agent populations, but a critical question remains: does the resulting cooperation reflect genuine prosocial alignment, or does it mask erosion of agent autonomy, epistemic integrity, and distributional fairness? We introduce Constitutional Multi-Agent Governance (CMAG), a two-stage framework that interposes between an LLM policy compiler and a networked agent population, combining hard constraint filtering with soft penalized-utility optimization that balances cooperation potential against manipulation risk and autonomy pressure. We propose the Ethical Cooperation Score (ECS), a multiplicative composite of cooperation, autonomy, integrity, and fairness that penalizes cooperation achieved through manipulative means. In experiments on scale-free networks of 80 agents under adversarial conditions (70% violating candidates), we benchmark three regimes: full CMAG, naive filtering, and unconstrained optimization. While unconstrained optimization achieves the highest raw cooperation (0.873), it yields the lowest ECS (0.645) due to severe autonomy erosion (0.867) and fairness degradation (0.888). CMAG attains an ECS of 0.741, a 14.9% improvement, while preserving autonomy at 0.985 and integrity at 0.995, with only modest cooperation reduction to 0.770. The naive ablation (ECS = 0.733) confirms that hard constraints alone are insufficient. Pareto analysis shows CMAG dominates the cooperation-autonomy trade-off space, and governance reduces hub-periphery exposure disparities by over 60%. These findings establish that cooperation is not inherently desirable without governance: constitutional constraints are necessary to ensure that LLM-mediated influence produces ethically stable outcomes rather than manipulative equilibria.\n\n---\n\n## 21. Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights\n- **Authors**: Xingli Fang , Jung-Eun Kim\n- **URL**: [https://arxiv.org/abs/2603.13186](https://arxiv.org/abs/2603.13186)\n- **Abstract**:\n> Prior approaches for membership privacy preservation usually update or retrain all weights in neural networks, which is costly and can lead to unnecessary utility loss or even more serious misalignment in predictions between training data and non-training data. In this work, we observed three insights: i) privacy vulnerability exists in a very small fraction of weights; ii) however, most of those weights also critically impact utility performance; iii) the importance of weights stems from their locations rather than their values. According to these insights, to preserve privacy, we score critical weights, and instead of discarding those neurons, we rewind only the weights for fine-tuning. We show that, through extensive experiments, this mechanism exhibits outperforming resilience in most cases against Membership Inference Attacks while maintaining utility.\n\n---\n\n## 22. MXNorm: Reusing MXFP block scales for efficient tensor normalisation\n- **Authors**: Callum McLean , Luke Y. Prince , Alexandre Payot , Paul Balança , Carlo Luschi\n- **URL**: [https://arxiv.org/abs/2603.13180](https://arxiv.org/abs/2603.13180)\n- **Abstract**:\n> Matrix multiplication performance has long been the major bottleneck to scaling deep learning workloads, which has stimulated the design of new accelerators that use increasingly low-precision number formats. However, improvements in matrix multiplication performance have far outstripped improvements in performance on reductions and elementwise computations, which are still being performed in higher precision. In this work, we propose MXNorm, a drop-in replacement for RMSNorm that estimates the RMS using only the block scales calculated as part of the MXFP8 cast and enables a 32x decrease in the size of reduction needed for normalization. We validate our approximation method on pre-training of Llama 3 models of 125M, 1B and 8B parameters, finding minimal loss of training accuracy compared to a baseline using RMSNorm with MXFP8 matmuls. We also show practical kernel speedups using only this http URL of up to 2.4x for MXNorm over RMSNorm, corresponding to a 1.3% speedup in Llama 3 8B transformer layers in MXFP8 and a 2.6% speedup in NVFP4.\n\n---\n\n## 23. Clustering Astronomical Orbital Synthetic Data Using Advanced Feature Extraction and Dimensionality Reduction Techniques\n- **Authors**: Eraldo Pereira Marinho , Nelson Callegari Junior , Fabricio Aparecido Breve , Caetano Mazzoni Ranieri\n- **URL**: [https://arxiv.org/abs/2603.13177](https://arxiv.org/abs/2603.13177)\n- **Abstract**:\n> The dynamics of Saturn's satellite system offer a rich framework for studying orbital stability and resonance interactions. Traditional methods for analysing such systems, including Fourier analysis and stability metrics, struggle with the scale and complexity of modern datasets. This study introduces a machine learning-based pipeline for clustering approximately 22,300 simulated satellite orbits, addressing these challenges with advanced feature extraction and dimensionality reduction techniques. The key to this approach is using MiniRocket, which efficiently transforms 400 timesteps into a 9,996-dimensional feature space, capturing intricate temporal patterns. Additional automated feature extraction and dimensionality reduction techniques refine the data, enabling robust clustering analysis. This pipeline reveals stability regions, resonance structures, and other key behaviours in Saturn's satellite system, providing new insights into their long-term dynamical evolution. By integrating computational tools with traditional celestial mechanics techniques, this study offers a scalable and interpretable methodology for analysing large-scale orbital datasets and advancing the exploration of planetary dynamics.\n\n---\n\n## 24. ESG-Bench: Benchmarking Long-Context ESG Reports for Hallucination Mitigation\n- **Authors**: Siqi Sun , Ben Peng Wu , Mali Jin , Peizhen Bai , Hanpei Zhang , Xingyi Song\n- **URL**: [https://arxiv.org/abs/2603.13154](https://arxiv.org/abs/2603.13154)\n- **Abstract**:\n> As corporate responsibility increasingly incorporates environmental, social, and governance (ESG) criteria, ESG reporting is becoming a legal requirement in many regions and a key channel for documenting sustainability practices and assessing firms' long-term and ethical performance. However, the length and complexity of ESG disclosures make them difficult to interpret and automate the analysis reliably. To support scalable and trustworthy analysis, this paper introduces ESG-Bench, a benchmark dataset for ESG report understanding and hallucination mitigation in large language models (LLMs). ESG-Bench contains human-annotated question-answer (QA) pairs grounded in real-world ESG report contexts, with fine-grained labels indicating whether model outputs are factually supported or hallucinated. Framing ESG report analysis as a QA task with verifiability constraints enables systematic evaluation of LLMs' ability to extract and reason over ESG content and provides a new use case: mitigating hallucinations in socially sensitive, compliance-critical settings. We design task-specific Chain-of-Thought (CoT) prompting strategies and fine-tune multiple state-of-the-art LLMs on ESG-Bench using CoT-annotated rationales. Our experiments show that these CoT-based methods substantially outperform standard prompting and direct fine-tuning in reducing hallucinations, and that the gains transfer to existing QA benchmarks beyond the ESG domain.\n\n---\n\n## 25. Developing the PsyCogMetrics AI Lab to Evaluate Large Language Models and Advance Cognitive Science -- A Three-Cycle Action Design Science Study\n- **Authors**: Zhiye Jin , Yibai Li , K. D. Joshi , Xuefei (Nancy) Deng , Xiaobing (Emily)Li\n- **URL**: [https://arxiv.org/abs/2603.13126](https://arxiv.org/abs/2603.13126)\n- **Abstract**:\n> This study presents the development of the PsyCogMetrics AI Lab ( this http URL ), an integrated, cloud-based platform that operationalizes psychometric and cognitive-science methodologies for Large Language Model (LLM) evaluation. Framed as a three-cycle Action Design Science study, the Relevance Cycle identifies key limitations in current evaluation methods and unfulfilled stakeholder needs. The Rigor Cycle draws on kernel theories such as Popperian falsifiability, Classical Test Theory, and Cognitive Load Theory to derive deductive design objectives. The Design Cycle operationalizes these objectives through nested Build-Intervene-Evaluate loops. The study contributes a novel IT artifact, a validated design for LLM evaluation, benefiting research at the intersection of AI, psychology, cognitive science, and the social and behavioral sciences.\n\n---\n\n## 26. Geometry-Guided Camera Motion Understanding in VideoLLMs\n- **Authors**: Haoan Feng , Sri Harsha Musunuri , Guan-Ming Su\n- **URL**: [https://arxiv.org/abs/2603.13119](https://arxiv.org/abs/2603.13119)\n- **Abstract**:\n> Camera motion is a fundamental geometric signal that shapes visual perception and cinematic style, yet current video-capable vision-language models (VideoLLMs) rarely represent it explicitly and often fail on fine-grained motion primitives. We address this gap with a framework of $\\textbf{benchmarking}$, $\\textbf{diagnosis}$, and $\\textbf{injection}$. We curate $\\textbf{CameraMotionDataset}$, a large-scale synthetic dataset with explicit camera control, formulate camera motion as constraint-aware multi-label recognition, and construct a VQA benchmark--$\\textbf{CameraMotionVQA}$. Across diverse off-the-shelf VideoLLMs, we observe substantial errors in recognizing camera motion primitives. Probing experiments on a Qwen2.5-VL vision encoder suggest that camera motion cues are weakly represented, especially in deeper ViT blocks, helping explain the observed failure modes. To bridge this gap without costly training or fine-tuning, we propose a lightweight, model-agnostic pipeline that extracts geometric camera cues from 3D foundation models (3DFMs), predicts constrained motion primitives with a temporal classifier, and injects them into downstream VideoLLM inference via structured prompting. Experiments demonstrate improved motion recognition and more camera-aware model responses, highlighting geometry-driven cue extraction and structured prompting as practical steps toward a camera-aware VideoLLM and VLA system. The dataset and benchmark is publicly available at this https URL .\n\n---\n\n## 27. BoSS: A Best-of-Strategies Selector as an Oracle for Deep Active Learning\n- **Authors**: Denis Huseljic , Paul Hahn , Marek Herde , Christoph Sandrock , Bernhard Sick\n- **URL**: [https://arxiv.org/abs/2603.13109](https://arxiv.org/abs/2603.13109)\n- **Abstract**:\n> Active learning (AL) aims to reduce annotation costs while maximizing model performance by iteratively selecting valuable instances. While foundation models have made it easier to identify these instances, existing selection strategies still lack robustness across different models, annotation budgets, and datasets. To highlight the potential weaknesses of existing AL strategies and provide a reference point for research, we explore oracle strategies, i.e., strategies that approximate the optimal selection by accessing ground-truth information unavailable in practical AL scenarios. Current oracle strategies, however, fail to scale effectively to large datasets and complex deep neural networks. To tackle these limitations, we introduce the Best-of-Strategy Selector (BoSS), a scalable oracle strategy designed for large-scale AL scenarios. BoSS constructs a set of candidate batches through an ensemble of selection strategies and then selects the batch yielding the highest performance gain. As an ensemble of selection strategies, BoSS can be easily extended with new state-of-the-art strategies as they emerge, ensuring it remains a reliable oracle strategy in the future. Our evaluation demonstrates that i) BoSS outperforms existing oracle strategies, ii) state-of-the-art AL strategies still fall noticeably short of oracle performance, especially in large-scale datasets with many classes, and iii) one possible solution to counteract the inconsistent performance of AL strategies might be to employ an ensemble-based approach for the selection.\n\n---\n\n## 28. Evaluating VLMs' Spatial Reasoning Over Robot Motion: A Step Towards Robot Planning with Motion Preferences\n- **Authors**: Wenxi Wu , Jingjing Zhang , Martim Brandão\n- **URL**: [https://arxiv.org/abs/2603.13100](https://arxiv.org/abs/2603.13100)\n- **Abstract**:\n> Understanding user instructions and object spatial relations in surrounding environments is crucial for intelligent robot systems to assist humans in various tasks. The natural language and spatial reasoning capabilities of Vision-Language Models (VLMs) have the potential to enhance the generalization of robot planners on new tasks, objects, and motion specifications. While foundation models have been applied to task planning, it is still unclear the degree to which they have the capability of spatial reasoning required to enforce user preferences or constraints on motion, such as desired distances from objects, topological properties, or motion style preferences. In this paper, we evaluate the capability of four state-of-the-art VLMs at spatial reasoning over robot motion, using four different querying methods. Our results show that, with the highest-performing querying method, Qwen2.5-VL achieves 71.4% accuracy zero-shot and 75% on a smaller model after fine-tuning, and GPT-4o leads to lower performance. We evaluate two types of motion preferences (object-proximity and path-style), and we also analyze the trade-off between accuracy and computation cost in number of tokens. This work shows some promise in the potential of VLM integration with robot motion planning pipelines.\n\n---\n\n## 29. Human-in-the-Loop LLM Grading for Handwritten Mathematics Assessments\n- **Authors**: Arne Vanhoyweghen , Vincent Holst , Melika Mobini , Lukas Van de Voorde , Tibo Vanleke , Bert Verbruggen , Brecht Verbeken , Andres Algaba , Sam Verboven , Marie-Anne Guerry , Filip Van Droogenbroeck , Vincent Ginis\n- **URL**: [https://arxiv.org/abs/2603.13083](https://arxiv.org/abs/2603.13083)\n- **Abstract**:\n> Providing timely and individualised feedback on handwritten student work is highly beneficial for learning but difficult to achieve at scale. This challenge has become more pressing as generative AI undermines the reliability of take-home assessments, shifting emphasis toward supervised, in-class evaluation. We present a scalable, end-to-end workflow for LLM-assisted grading of short, pen-and-paper assessments. The workflow spans (1) constructing solution keys, (2) developing detailed rubric-style grading keys used to guide the LLM, and (3) a grading procedure that combines automated scanning and anonymisation, multi-pass LLM scoring, automated consistency checks, and mandatory human verification. We deploy the system in two undergraduate mathematics courses using six low-stakes in-class tests. Empirically, LLM assistance reduces grading time by approximately 23% while achieving agreement comparable to, and in several cases tighter than, fully manual grading. Occasional model errors occur but are effectively contained by the hybrid design. Overall, our results show that carefully embedded human-in-the-loop LLM grading can substantially reduce workload while maintaining fairness and accuracy.\n\n---\n\n## 30. GeoChemAD: Benchmarking Unsupervised Geochemical Anomaly Detection for Mineral Exploration\n- **Authors**: Yihao Ding , Yiran Zhang , Chris Gonzalez , Eun-Jung Holden , Wei Liu\n- **URL**: [https://arxiv.org/abs/2603.13068](https://arxiv.org/abs/2603.13068)\n- **Abstract**:\n> Geochemical anomaly detection plays a critical role in mineral exploration as deviations from regional geochemical baselines may indicate mineralization. Existing studies suffer from two key limitations: (1) single region scenarios which limit model generalizability; (2) proprietary datasets, which makes result reproduction unattainable. In this work, we introduce \\textbf{GeoChemAD}, an open-source benchmark dataset compiled from government-led geological surveys, covering multiple regions, sampling sources, and target elements. The dataset comprises eight subsets representing diverse spatial scales and sampling conditions. To establish strong baselines, we reproduce and benchmark a range of unsupervised anomaly detection methods, including statistical models, generative and transformer-based approaches. Furthermore, we propose \\textbf{GeoChemFormer}, a transformer-based framework that leverages self-supervised pretraining to learn target-element-aware geochemical representations for spatial samples. Extensive experiments demonstrate that GeoChemFormer consistently achieves superior and robust performance across all eight subsets, outperforming existing unsupervised methods in both anomaly detection accuracy and generalization capability. The proposed dataset and framework provide a foundation for reproducible research and future development in this direction.\n\n---\n\n## 31. L2GTX: From Local to Global Time Series Explanations\n- **Authors**: Ephrem Tibebe Mekonnen , Luca Longo , Lucas Rizzo , Pierpaolo Dondio\n- **URL**: [https://arxiv.org/abs/2603.13065](https://arxiv.org/abs/2603.13065)\n- **Abstract**:\n> Deep learning models achieve high accuracy in time series classification, yet understanding their class-level decision behaviour remains challenging. Explanations for time series must respect temporal dependencies and identify patterns that recur across instances. Existing approaches face three limitations: model-agnostic XAI methods developed for images and tabular data do not readily extend to time series, global explanation synthesis for time series remains underexplored, and most existing global approaches are model-specific. We propose L2GTX, a model-agnostic framework that generates class-wise global explanations by aggregating local explanations from a representative set of instances. L2GTX extracts clusters of parameterised temporal event primitives, such as increasing or decreasing trends and local extrema, together with their importance scores from instance-level explanations produced by LOMATCE. These clusters are merged across instances to reduce redundancy, and an instance-cluster importance matrix is used to estimate global relevance. Under a user-defined instance selection budget, L2GTX selects representative instances that maximise coverage of influential clusters. Events from the selected instances are then aggregated into concise class-wise global explanations. Experiments on six benchmark time series datasets show that L2GTX produces compact and interpretable global explanations while maintaining stable global faithfulness measured as mean local surrogate fidelity.\n\n---\n\n## 32. Competition-Aware CPC Forecasting with Near-Market Coverage\n- **Authors**: Sebastian Frey , Edoardo Beccari , Maximilian Kranz , Nicolò Alberto Pellizzari , Ali Mete Karaman , Qiwei Han , Maximilian Kaiser\n- **URL**: [https://arxiv.org/abs/2603.13059](https://arxiv.org/abs/2603.13059)\n- **Abstract**:\n> Cost-per-click (CPC) in paid search is a volatile auction outcome generated by a competitive landscape that is only partially observable from any single advertiser's history. Using Google Ads auction logs from a concentrated car-rental market (2021--2023), we forecast weekly CPC for 1,811 keyword series and approximate latent competition through complementary signals derived from keyword text, CPC trajectories, and geographic market structure. We construct (i) semantic neighborhoods and a semantic keyword graph from pretrained transformer-based representations of keyword text, (ii) behavioral neighborhoods via Dynamic Time Warping (DTW) alignment of CPC trajectories, and (iii) geographic-intent covariates capturing localized demand and marketplace heterogeneity. We extensively evaluate these signals both as stand-alone covariates and as relational priors in spatiotemporal graph forecasters, benchmarking them against strong statistical, neural, and time-series foundation-model baselines. Across methods, competition-aware augmentation improves stability and error profiles at business-relevant medium and longer horizons, where competitive regimes shift and volatility is most consequential. The results show that broad market-outcome coverage, combined with keyword-derived semantic and geographic priors, provides a scalable way to approximate latent competition and improve CPC forecasting in auction-driven markets.\n\n---\n\n## 33. Team RAS in 10th ABAW Competition: Multimodal Valence and Arousal Estimation Approach\n- **Authors**: Elena Ryumina (1), Maxim Markitantov (1), Alexandr Axyonov (1), Dmitry Ryumin (1), Mikhail Dolgushin (1), Denis Dresvyanskiy (2), Alexey Karpov (1 and 2) ((1) St. Petersburg Federal Research Center of the Russian Academy of Sciences, St. Petersburg, Russia, (2) ITMO University, St. Petersburg, Russia)\n- **URL**: [https://arxiv.org/abs/2603.13056](https://arxiv.org/abs/2603.13056)\n- **Abstract**:\n> Continuous emotion recognition in terms of valence and arousal under in-the-wild (ITW) conditions remains a challenging problem due to large variations in appearance, head pose, illumination, occlusions, and subject-specific patterns of affective expression. We present a multimodal method for valence-arousal estimation ITW. Our method combines three complementary modalities: face, behavior, and audio. The face modality relies on GRADA-based frame-level embeddings and Transformer-based temporal regression. We use Qwen3-VL-4B-Instruct to extract behavior-relevant information from video segments, while Mamba is used to model temporal dynamics across segments. The audio modality relies on WavLM-Large with attention-statistics pooling and includes a cross-modal filtering stage to reduce the influence of unreliable or non-speech segments. To fuse modalities, we explore two fusion strategies: a Directed Cross-Modal Mixture-of-Experts Fusion Strategy that learns interactions between modalities with adaptive weighting, and a Reliability-Aware Audio-Visual Fusion Strategy that combines visual features at the frame-level while using audio as complementary context. The results are reported on the Aff-Wild2 dataset following the 10th Affective Behavior Analysis in-the-Wild (ABAW) challenge protocol. Experiments demonstrate that the proposed multimodal fusion strategy achieves a Concordance Correlation Coefficient (CCC) of 0.658 on the Aff-Wild2 development set.\n\n---\n\n## 34. Are General-Purpose Vision Models All We Need for 2D Medical Image Segmentation? A Cross-Dataset Empirical Study\n- **Authors**: Vanessa Borst , Samuel Kounev\n- **URL**: [https://arxiv.org/abs/2603.13044](https://arxiv.org/abs/2603.13044)\n- **Abstract**:\n> Medical image segmentation (MIS) is a fundamental component of computer-assisted diagnosis and clinical decision support systems. Over the past decade, numerous architectures specifically tailored to medical imaging have emerged to address domain-specific challenges such as low contrast, small anatomical structures, and limited annotated data. In parallel, rapid progress in computer vision has produced highly capable general-purpose vision models (GP-VMs) originally designed for natural images. Despite their strong performance on standard vision benchmarks, their effectiveness for MIS remains insufficiently understood. In this work, we conduct a controlled empirical study to examine whether specialized medical segmentation architectures (SMAs) provide systematic advantages over modern GP-VMs for 2D MIS. We compare eleven SMAs and GP-VMs using a unified training and evaluation protocol. Experiments are performed across three heterogeneous datasets covering different imaging modalities, class structures, and data characteristics. Beyond segmentation accuracy, we analyze qualitative Grad-CAM visualizations to investigate explainability (XAI) behavior. Our results demonstrate that, for the analyzed datasets, GP-VMs out-perform the majority of specialized MIS models. Moreover, XAI analyses indicate that GP-VMs can capture clinically relevant structures without explicit domain-specific architectural design. These findings suggest that GP-VMs can represent a viable alternative to domain-specific methods, highlighting the importance of informed model selection for end-to-end MIS systems. All code and resources are available at GitHub.\n\n---\n\n## 35. Interrogating Design Homogenization in Web Vibe Coding\n- **Authors**: Donghoon Shin , Alice Gao , Rock Yuren Pang , Jaewook Lee , Katharina Reinecke , Emily Tseng\n- **URL**: [https://arxiv.org/abs/2603.13036](https://arxiv.org/abs/2603.13036)\n- **Abstract**:\n> Generative AI is known for its tendency to homogenize, often reproducing dominant style conventions found in training data. However, it remains unclear how these homogenizing effects extend to complex structural tasks like web design. As lay creators increasingly turn to LLMs to 'vibe-code' websites -- prompting for aesthetic and functional goals rather than writing code -- they may inadvertently narrow the diversity of their designs, and limit creative expression throughout the internet. In this paper, we interrogate the possibility of design homogenization in web vibe coding. We first characterize the vibe coding lifecycle, pinpointing stages where homogenization risks may arise. We then conduct a sociotechnical risk analysis unpacking the potential harms of web vibe coding and their interaction with design homogenization. We identify that the push for frictionless generation can exacerbate homogenization and its harms. Finally, we propose a mitigation framework centered on the idea of productive friction. Through case studies at the micro, meso, and macro levels, we show how centering productive friction can empower creators to challenge default outputs and preserve diverse expression in AI-mediated web design.\n\n---\n\n## 36. Purify Once, Edit Freely: Breaking Image Protections under Model Mismatch\n- **Authors**: Qichen Zhao , Shengfang Zhai , Xinjian Bai , Qingni Shen , Qiqi Lin , Yansong Gao , Zhonghai Wu\n- **URL**: [https://arxiv.org/abs/2603.13028](https://arxiv.org/abs/2603.13028)\n- **Abstract**:\n> Diffusion models enable high-fidelity image editing but can also be misused for unauthorized style imitation and harmful content generation. To mitigate these risks, proactive image protection methods embed small, often imperceptible adversarial perturbations into images before sharing to disrupt downstream editing or fine-tuning. However, in realistic post-release scenarios, content owners cannot control downstream processing pipelines, and protections optimized for a surrogate model may fail when attackers use mismatched diffusion pipelines. Existing purification methods can weaken protections but often sacrifice image quality and rarely examine architectural mismatch. We introduce a unified post-release purification framework to evaluate protection survivability under model mismatch. We propose two practical purifiers: VAE-Trans, which corrects protected images via latent-space projection, and EditorClean, which performs instruction-guided reconstruction with a Diffusion Transformer to exploit architectural heterogeneity. Both operate without access to protected images or defense internals. Across 2,100 editing tasks and six representative protection methods, EditorClean consistently restores editability. Compared to protected inputs, it improves PSNR by 3-6 dB and reduces FID by 50-70 percent on downstream edits, while outperforming prior purification baselines by about 2 dB PSNR and 30 percent lower FID. Our results reveal a purify-once, edit-freely failure mode: once purification succeeds, the protective signal is largely removed, enabling unrestricted editing. This highlights the need to evaluate protections under model mismatch and design defenses robust to heterogeneous attackers.\n\n---\n\n## 37. SortScrews: A Dataset and Baseline for Real-time Screw Classification\n- **Authors**: Tianhao Fu , Bingxuan Yang , Juncheng Guo , Shrena Sribalan , Yucheng Chen\n- **URL**: [https://arxiv.org/abs/2603.13027](https://arxiv.org/abs/2603.13027)\n- **Abstract**:\n> Automatic identification of screw types is important for industrial automation, robotics, and inventory management. However, publicly available datasets for screw classification are scarce, particularly for controlled single-object scenarios commonly encountered in automated sorting systems. In this work, we introduce $\\textbf{SortScrews}$, a dataset for casewise visual classification of screws. The dataset contains 560 RGB images at $512\\times512$ resolution covering six screw types and a background class. Images are captured using a standardized acquisition setup and include mild variations in lighting and camera perspective across four capture settings. To facilitate reproducible research and dataset expansion, we also provide a reusable data collection script that allows users to easily construct similar datasets for custom hardware components using inexpensive camera setups. We establish baseline results using transfer learning with EfficientNet-B0 and ResNet-18 classifiers pretrained on ImageNet. In addition, we conduct a well-explored failure analysis. Despite the limited dataset size, these lightweight models achieve strong classification accuracy, demonstrating that controlled acquisition conditions enable effective learning even with relatively small datasets. The dataset, collection pipeline, and baseline training code are publicly available at this https URL .\n\n---\n\n## 38. SAW: Toward a Surgical Action World Model via Controllable and Scalable Video Generation\n- **Authors**: Sampath Rapuri , Lalithkumar Seenivasan , Dominik Schneider , Roger Soberanis-Mukul , Yufan He , Hao Ding , Jiru Xu , Chenhao Yu , Chenyan Jing , Pengfei Guo , Daguang Xu , Mathias Unberath\n- **URL**: [https://arxiv.org/abs/2603.13024](https://arxiv.org/abs/2603.13024)\n- **Abstract**:\n> A surgical world model capable of generating realistic surgical action videos with precise control over tool-tissue interactions can address fundamental challenges in surgical AI and simulation -- from data scarcity and rare event synthesis to bridging the sim-to-real gap for surgical automation. However, current video generation methods, the very core of such surgical world models, require expensive annotations or complex structured intermediates as conditioning signals at inference, limiting their scalability. Other approaches exhibit limited temporal consistency across complex laparoscopic scenes and do not possess sufficient realism. We propose Surgical Action World (SAW) -- a step toward surgical action world modeling through video diffusion conditioned on four lightweight signals: language prompts encoding tool-action context, a reference surgical scene, tissue affordance mask, and 2D tool-tip trajectories. We design a conditional video diffusion approach that reformulates video-to-video diffusion into trajectory-conditioned surgical action synthesis. The backbone diffusion model is fine-tuned on a custom-curated dataset of 12,044 laparoscopic clips with lightweight spatiotemporal conditioning signals, leveraging a depth consistency loss to enforce geometric plausibility without requiring depth at inference. SAW achieves state-of-the-art temporal consistency (CD-FVD: 199.19 vs. 546.82) and strong visual quality on held-out test data. Furthermore, we demonstrate its downstream utility for (a) surgical AI, where augmenting rare actions with SAW-generated videos improves action recognition (clipping F1-score: 20.93% to 43.14%; cutting: 0.00% to 8.33%) on real test data, and (b) surgical simulation, where rendering tool-tissue interaction videos from simulator-derived trajectory points toward a visually faithful simulation engine.\n\n---\n\n## 39. daVinci-Env: Open SWE Environment Synthesis at Scale\n- **Authors**: Dayuan Fu , Shenyu Wu , Yunze Wu , Zerui Peng , Yaxing Huang , Jie Sun , Ji Zeng , Mohan Jiang , Lin Zhang , Yukun Li , Jiarui Hu , Liming Liu , Jinlong Hou , Pengfei Liu\n- **URL**: [https://arxiv.org/abs/2603.13023](https://arxiv.org/abs/2603.13023)\n- **Abstract**:\n> Training capable software engineering (SWE) agents demands large-scale, executable, and verifiable environments that provide dynamic feedback loops for iterative code editing, test execution, and solution refinement. However, existing open-source datasets remain limited in scale and repository diversity, while industrial solutions are opaque with unreleased infrastructure, creating a prohibitive barrier for most academic research groups. We present OpenSWE, the largest fully transparent framework for SWE agent training in Python, comprising 45,320 executable Docker environments spanning over 12.8k repositories, with all Dockerfiles, evaluation scripts, and infrastructure fully open-sourced for reproducibility. OpenSWE is built through a multi-agent synthesis pipeline deployed across a 64-node distributed cluster, automating repository exploration, Dockerfile construction, evaluation script generation, and iterative test analysis. Beyond scale, we propose a quality-centric filtering pipeline that characterizes the inherent difficulty of each environment, filtering out instances that are either unsolvable or insufficiently challenging and retaining only those that maximize learning efficiency. With $891K spent on environment construction and an additional $576K on trajectory sampling and difficulty-aware curation, the entire project represents a total investment of approximately $1.47 million, yielding about 13,000 curated trajectories from roughly 9,000 quality guaranteed environments. Extensive experiments validate OpenSWE's effectiveness: OpenSWE-32B and OpenSWE-72B achieve 62.4% and 66.0% on SWE-bench Verified, establishing SOTA among Qwen2.5 series. Moreover, SWE-focused training yields substantial out-of-domain improvements, including up to 12 points on mathematical reasoning and 5 points on science benchmarks, without degrading factual recall.\n\n---\n\n## 40. ARL-Tangram: Unleash the Resource Efficiency in Agentic Reinforcement Learning\n- **Authors**: Bangjun Xiao , Yihao Zhao , Xiangwei Deng , Shihua Yu , Yuxing Xiang , Huaqiu Liu , Qiying Wang , Liang Zhao , Hailin Zhang , Xuanzhe Liu , Xin Jin , Fuli Luo\n- **URL**: [https://arxiv.org/abs/2603.13019](https://arxiv.org/abs/2603.13019)\n- **Abstract**:\n> Agentic reinforcement learning (RL) has emerged as a transformative workload in cloud clusters, enabling large language models (LLMs) to solve complex problems through interactions with real world. However, unlike traditional RL, agentic RL demands substantial external cloud resources, e.g., CPUs for code execution and GPUs for reward models, that exist outside the primary training cluster. Existing agentic RL framework typically rely on static over-provisioning, i.e., resources are often tied to long-lived trajectories or isolated by tasks, which leads to severe resource inefficiency. We propose the action-level orchestration, and incorporate it into ARL-Tangram, a unified resource management system that enables fine-grained external resource sharing and elasticity. ARL-Tangram utilizes a unified action-level formulation and an elastic scheduling algorithm to minimize action completion time (ACT) while satisfying heterogeneous resource constraints. Further, heterogeneous resource managers are tailored to efficiently support the action-level execution on resources with heterogeneous characteristics and topologies. Evaluation on real-world agentic RL tasks demonstrates that ARL-Tangram improves average ACT by up to 4.3$\\times$, speeds up the step duration of RL training by up to 1.5$\\times$, and saves the external resources by up to 71.2$\\%$. This system has been deployed to support the training of the MiMo series models.\n\n---\n\n## 41. Fair Lung Disease Diagnosis from Chest CT via Gender-Adversarial Attention Multiple Instance Learning\n- **Authors**: Aditya Parikh , Aasa Feragen\n- **URL**: [https://arxiv.org/abs/2603.12988](https://arxiv.org/abs/2603.12988)\n- **Abstract**:\n> We present a fairness-aware framework for multi-class lung disease diagnosis from chest CT volumes, developed for the Fair Disease Diagnosis Challenge at the PHAROS-AIF-MIH Workshop (CVPR 2026). The challenge requires classifying CT scans into four categories -- Healthy, COVID-19, Adenocarcinoma, and Squamous Cell Carcinoma -- with performance measured as the average of per-gender macro F1 scores, explicitly penalizing gender-inequitable predictions. Our approach addresses two core difficulties: the sparse pathological signal across hundreds of slices, and a severe demographic imbalance compounded across disease class and gender. We propose an attention-based Multiple Instance Learning (MIL) model on a ConvNeXt backbone that learns to identify diagnostically relevant slices without slice-level supervision, augmented with a Gradient Reversal Layer (GRL) that adversarially suppresses gender-predictive structure in the learned scan representation. Training incorporates focal loss with label smoothing, stratified cross-validation over joint (class, gender) strata, and targeted oversampling of the most underrepresented subgroup. At inference, all five-fold checkpoints are ensembled with horizontal-flip test-time augmentation via soft logit voting and out-of-the-fold threshold optimization for robustness. Our model achieves a mean validation competition score of 0.685 (std - 0.030), with the best single fold reaching 0.759. All training and inference code is publicly available at this https URL\n\n---\n\n## 42. Is Human Annotation Necessary? Iterative MBR Distillation for Error Span Detection in Machine Translation\n- **Authors**: Boxuan Lyu , Haiyue Song , Zhi Qu\n- **URL**: [https://arxiv.org/abs/2603.12983](https://arxiv.org/abs/2603.12983)\n- **Abstract**:\n> Error Span Detection (ESD) is a crucial subtask in Machine Translation (MT) evaluation, aiming to identify the location and severity of translation errors. While fine-tuning models on human-annotated data improves ESD performance, acquiring such data is expensive and prone to inconsistencies among annotators. To address this, we propose a novel self-evolution framework based on Minimum Bayes Risk (MBR) decoding, named Iterative MBR Distillation for ESD, which eliminates the reliance on human annotations by leveraging an off-the-shelf LLM to generate pseudo-labels. Extensive experiments on the WMT Metrics Shared Task datasets demonstrate that models trained solely on these self-generated pseudo-labels outperform both unadapted base model and supervised baselines trained on human annotations at the system and span levels, while maintaining competitive sentence-level performance.\n\n---\n\n## 43. Efficient Real-World Autonomous Racing via Attenuated Residual Policy Optimization\n- **Authors**: Raphael Trumpp , Denis Hoornaert , Mirco Theile , Marco Caccamo\n- **URL**: [https://arxiv.org/abs/2603.12960](https://arxiv.org/abs/2603.12960)\n- **Abstract**:\n> Residual policy learning (RPL), in which a learned policy refines a static base policy using deep reinforcement learning (DRL), has shown strong performance across various robotic applications. Its effectiveness is particularly evident in autonomous racing, a domain that serves as a challenging benchmark for real-world DRL. However, deploying RPL-based controllers introduces system complexity and increases inference latency. We address this by introducing an extension of RPL named attenuated residual policy optimization ($\\alpha$-RPO). Unlike standard RPL, $\\alpha$-RPO yields a standalone neural policy by progressively attenuating the base policy, which initially serves to bootstrap learning. Furthermore, this mechanism enables a form of privileged learning, where the base policy is permitted to use sensor modalities not required for final deployment. We design $\\alpha$-RPO to integrate seamlessly with PPO, ensuring that the attenuated influence of the base controller is dynamically compensated during policy optimization. We evaluate $\\alpha$-RPO by building a framework for 1:10-scaled autonomous racing around it. In both simulation and zero-shot real-world transfer to Roboracer cars, $\\alpha$-RPO not only reduces system complexity but also improves driving performance compared to baselines - demonstrating its practicality for robotic deployment. Our code is available at: this https URL .\n\n---\n\n## 44. Delta1 with LLM: symbolic and neural integration for credible and explainable reasoning\n- **Authors**: Yang Xu , Jun Liu , Shuwei Chen , Chris Nugent , Hailing Guo\n- **URL**: [https://arxiv.org/abs/2603.12953](https://arxiv.org/abs/2603.12953)\n- **Abstract**:\n> Neuro-symbolic reasoning increasingly demands frameworks that unite the formal rigor of logic with the interpretability of large language models (LLMs). We introduce an end to end explainability by construction pipeline integrating the Automated Theorem Generator Delta1 based on the full triangular standard contradiction (FTSC) with LLMs. Delta1 deterministically constructs minimal unsatisfiable clause sets and complete theorems in polynomial time, ensuring both soundness and minimality by construction. The LLM layer verbalizes each theorem and proof trace into coherent natural language explanations and actionable insights. Empirical studies across health care, compliance, and regulatory domains show that Delta1 and LLM enables interpretable, auditable, and domain aligned reasoning. This work advances the convergence of logic, language, and learning, positioning constructive theorem generation as a principled foundation for neuro-symbolic explainable AI.\n\n---\n\n## 45. Thinking in Streaming Video\n- **Authors**: Zikang Liu , Longteng Guo , Handong Li , Ru Zhen , Xingjian He , Ruyi Ji , Xiaoming Ren , Yanhao Zhang , Haonan Lu , Jing Liu\n- **URL**: [https://arxiv.org/abs/2603.12938](https://arxiv.org/abs/2603.12938)\n- **Abstract**:\n> Real-time understanding of continuous video streams is essential for interactive assistants and multimodal agents operating in dynamic environments. However, most existing video reasoning approaches follow a batch paradigm that defers reasoning until the full video context is observed, resulting in high latency and growing computational cost that are incompatible with streaming scenarios. In this paper, we introduce ThinkStream, a framework for streaming video reasoning based on a Watch--Think--Speak paradigm that enables models to incrementally update their understanding as new video observations arrive. At each step, the model performs a short reasoning update and decides whether sufficient evidence has accumulated to produce a response. To support long-horizon streaming, we propose Reasoning-Compressed Streaming Memory (RCSM), which treats intermediate reasoning traces as compact semantic memory that replaces outdated visual tokens while preserving essential context. We further train the model using a Streaming Reinforcement Learning with Verifiable Rewards scheme that aligns incremental reasoning and response timing with the requirements of streaming interaction. Experiments on multiple streaming video benchmarks show that ThinkStream significantly outperforms existing online video models while maintaining low latency and memory usage. Code, models and data will be released at this https URL\n\n---\n\n## 46. Surprised by Attention: Predictable Query Dynamics for Time Series Anomaly Detection\n- **Authors**: Kadir-Kaan Özer , René Ebeling , Markus Enzweiler\n- **URL**: [https://arxiv.org/abs/2603.12916](https://arxiv.org/abs/2603.12916)\n- **Abstract**:\n> Multivariate time series anomalies often manifest as shifts in cross-channel dependencies rather than simple amplitude excursions. In autonomous driving, for instance, a steering command might be internally consistent but decouple from the resulting lateral acceleration. Residual-based detectors can miss such anomalies when flexible sequence models still reconstruct signals plausibly despite altered coordination. We introduce AxonAD, an unsupervised detector that treats multi-head attention query evolution as a short horizon predictable process. A gradient-updated reconstruction pathway is coupled with a history-only predictor that forecasts future query vectors from past context. This is trained via a masked predictor-target objective against an exponential moving average (EMA) target encoder. At inference, reconstruction error is combined with a tail-aggregated query mismatch score, which measures cosine deviation between predicted and target queries on recent timesteps. This dual approach provides sensitivity to structural dependency shifts while retaining amplitude-level detection. On proprietary in-vehicle telemetry with interval annotations and on the TSB-AD multi-variate suite (17 datasets, 180 series) with threshold-free and range-aware metrics, AxonAD improves ranking quality and temporal localization over strong baselines. Ablations confirm that query prediction and combined scoring are the primary drivers of the observed gains. Code is available at the URL this https URL .\n\n---\n\n## 47. Stake the Points: Structure-Faithful Instance Unlearning\n- **Authors**: Kiseong Hong , JungKyoo Shin , Eunwoo Kim\n- **URL**: [https://arxiv.org/abs/2603.12915](https://arxiv.org/abs/2603.12915)\n- **Abstract**:\n> Machine unlearning (MU) addresses privacy risks in pretrained models. The main goal of MU is to remove the influence of designated data while preserving the utility of retained knowledge. Achieving this goal requires preserving semantic relations among retained instances, which existing studies often overlook. We observe that without such preservation, models suffer from progressive structural collapse, undermining both the deletion-retention balance. In this work, we propose a novel structure-faithful framework that introduces stakes, i.e., semantic anchors that serve as reference points to maintain the knowledge structure. By leveraging these anchors, our framework captures and stabilizes the semantic organization of knowledge. Specifically, we instantiate the anchors from language-driven attribute descriptions encoded by a semantic encoder (e.g., CLIP). We enforce preservation of the knowledge structure via structure-aware alignment and regularization: the former aligns the organization of retained knowledge before and after unlearning around anchors, while the latter regulates updates to structure-critical parameters. Results from image classification, retrieval, and face recognition show average gains of 32.9%, 22.5%, and 19.3% in performance, balancing the deletion-retention trade-off and enhancing generalization.\n\n---\n\n## 48. FedBPrompt: Federated Domain Generalization Person Re-Identification via Body Distribution Aware Visual Prompts\n- **Authors**: Xin Xu , Weilong Li , Wei Liu , Wenke Huang , Zhixi Yu , Bin Yang , Xiaoying Liao , Kui Jiang\n- **URL**: [https://arxiv.org/abs/2603.12912](https://arxiv.org/abs/2603.12912)\n- **Abstract**:\n> Federated Domain Generalization for Person Re-Identification (FedDG-ReID) learns domain-invariant representations from decentralized data. While Vision Transformer (ViT) is widely adopted, its global attention often fails to distinguish pedestrians from high similarity backgrounds or diverse viewpoints -- a challenge amplified by cross-client distribution shifts in FedDG-ReID. To address this, we propose Federated Body Distribution Aware Visual Prompt (FedBPrompt), introducing learnable visual prompts to guide Transformer attention toward pedestrian-centric regions. FedBPrompt employs a Body Distribution Aware Visual Prompts Mechanism (BAPM) comprising: Holistic Full Body Prompts to suppress cross-client background noise, and Body Part Alignment Prompts to capture fine-grained details robust to pose and viewpoint variations. To mitigate high communication costs, we design a Prompt-based Fine-Tuning Strategy (PFTS) that freezes the ViT backbone and updates only lightweight prompts, significantly reducing communication overhead while maintaining adaptability. Extensive experiments demonstrate that BAPM effectively enhances feature discrimination and cross-domain generalization, while PFTS achieves notable performance gains within only a few aggregation rounds. Moreover, both BAPM and PFTS can be easily integrated into existing ViT-based FedDG-ReID frameworks, making FedBPrompt a flexible and effective solution for federated person re-identification. The code is available at this https URL .\n\n---\n\n## 49. Learning from Child-Directed Speech in Two-Language Scenarios: A French-English Case Study\n- **Authors**: Liel Binyamin , Elior Sulem\n- **URL**: [https://arxiv.org/abs/2603.12906](https://arxiv.org/abs/2603.12906)\n- **Abstract**:\n> Research on developmentally plausible language models has largely focused on English, leaving open questions about multilingual settings. We present a systematic study of compact language models by extending BabyBERTa to English-French scenarios under strictly size-matched data conditions, covering monolingual, bilingual, and cross-lingual settings. Our design contrasts two types of training corpora: (i) child-directed speech (about 2.5M tokens), following BabyBERTa and related work, and (ii) multi-domain corpora (about 10M tokens), extending the BabyLM framework to French. To enable fair evaluation, we also introduce new resources, including French versions of QAMR and QASRL, as well as English and French multi-domain corpora. We evaluate the models on both syntactic and semantic tasks and compare them with models trained on Wikipedia-only data. The results reveal context-dependent effects: training on Wikipedia consistently benefits semantic tasks, whereas child-directed speech improves grammatical judgments in monolingual settings. Bilingual pretraining yields notable gains for textual entailment, with particularly strong improvements for French. Importantly, similar patterns emerge across BabyBERTa, RoBERTa, and LTG-BERT, suggesting consistent trends across architectures.\n\n---\n\n## 50. Human-Centered Evaluation of an LLM-Based Process Modeling Copilot: A Mixed-Methods Study with Domain Experts\n- **Authors**: Chantale Lauer , Peter Pfeiffer , Nijat Mehdiyev\n- **URL**: [https://arxiv.org/abs/2603.12895](https://arxiv.org/abs/2603.12895)\n- **Abstract**:\n> Integrating Large Language Models (LLMs) into business process management tools promises to democratize Business Process Model and Notation (BPMN) modeling for non-experts. While automated frameworks assess syntactic and semantic quality, they miss human factors like trust, usability, and professional alignment. We conducted a mixed-methods evaluation of our proposed solution, an LLM-powered BPMN copilot, with five process modeling experts using focus groups and standardized questionnaires. Our findings reveal a critical tension between acceptable perceived usability (mean CUQ score: 67.2/100) and notably lower trust (mean score: 48.8\\%), with reliability rated as the most critical concern (M=1.8/5). Furthermore, we identified output-quality issues, prompting difficulties, and a need for the LLM to ask more in-depth clarifying questions about the process. We envision five use cases ranging from domain-expert support to enterprise quality assurance. We demonstrate the necessity of human-centered evaluation complementing automated benchmarking for LLM modeling agents.\n\n---\n\n## 51. Finite Difference Flow Optimization for RL Post-Training of Text-to-Image Models\n- **Authors**: David McAllister , Miika Aittala , Tero Karras , Janne Hellsten , Angjoo Kanazawa , Timo Aila , Samuli Laine\n- **URL**: [https://arxiv.org/abs/2603.12893](https://arxiv.org/abs/2603.12893)\n- **Abstract**:\n> Reinforcement learning (RL) has become a standard technique for post-training diffusion-based image synthesis models, as it enables learning from reward signals to explicitly improve desirable aspects such as image quality and prompt alignment. In this paper, we propose an online RL variant that reduces the variance in the model updates by sampling paired trajectories and pulling the flow velocity in the direction of the more favorable image. Unlike existing methods that treat each sampling step as a separate policy action, we consider the entire sampling process as a single action. We experiment with both high-quality vision language models and off-the-shelf quality metrics for rewards, and evaluate the outputs using a broad set of metrics. Our method converges faster and yields higher output quality and prompt alignment than previous approaches.\n\n---\n\n## 52. Team LEYA in 10th ABAW Competition: Multimodal Ambivalence/Hesitancy Recognition Approach\n- **Authors**: Elena Ryumina (1), Alexandr Axyonov (1), Dmitry Sysoev (2), Timur Abdulkadirov (2), Kirill Almetov (2), Yulia Morozova (2), Dmitry Ryumin (1 and 2) ((1) St. Petersburg Federal Research Center of the Russian Academy of Sciences, St. Petersburg, Russia, (2) HSE University, St. Petersburg, Russia)\n- **URL**: [https://arxiv.org/abs/2603.12848](https://arxiv.org/abs/2603.12848)\n- **Abstract**:\n> Ambivalence/hesitancy recognition in unconstrained videos is a challenging problem due to the subtle, multimodal, and context-dependent nature of this behavioral state. In this paper, a multimodal approach for video-level ambivalence/hesitancy recognition is presented for the 10th ABAW Competition. The proposed approach integrates four complementary modalities: scene, face, audio, and text. Scene dynamics are captured with a VideoMAE-based model, facial information is encoded through emotional frame-level embeddings aggregated by statistical pooling, acoustic representations are extracted with EmotionWav2Vec2.0 and processed by a Mamba-based temporal encoder, and linguistic cues are modeled using fine-tuned transformer-based text models. The resulting unimodal embeddings are further combined using multimodal fusion models, including prototype-augmented variants. Experiments on the BAH corpus demonstrate clear gains of multimodal fusion over all unimodal baselines. The best unimodal configuration achieved an average MF1 of 70.02%, whereas the best multimodal fusion model reached 83.25%. The highest final test performance, 71.43%, was obtained by an ensemble of five prototype-augmented fusion models. The obtained results highlight the importance of complementary multimodal cues and robust fusion strategies for ambivalence/hesitancy recognition.\n\n---\n\n## 53. Hierarchical Reference Sets for Robust Unsupervised Detection of Scattered and Clustered Outliers\n- **Authors**: Yiqun Zhang , Zexi Tan , Xiaopeng Luo , Yunlin Liu\n- **URL**: [https://arxiv.org/abs/2603.12847](https://arxiv.org/abs/2603.12847)\n- **Abstract**:\n> Most real-world IoT data analysis tasks, such as clustering and anomaly event detection, are unsupervised and highly susceptible to the presence of outliers. In addition to sporadic scattered outliers caused by factors such as faulty sensor readings, IoT systems often exhibit clustered outliers. These occur when multiple devices or nodes produce similar anomalous measurements, for instance, owing to localized interference, emerging security threats, or regional false alarms, forming micro-clusters. These clustered outliers can be easily mistaken for normal behavior because of their relatively high local density, thereby obscuring the detection of both scattered and contextual anomalies. To address this, we propose a novel outlier detection paradigm that leverages the natural neighboring relationships using graph structures. This facilitates multi-perspective anomaly evaluation by incorporating reference sets at both local and global scales derived from the graph. Our approach enables the effective recognition of scattered outliers without interference from clustered anomalies, whereas the graph structure simultaneously helps reflect and isolate clustered outlier groups. Extensive experiments, including comparative performance analysis, ablation studies, validation on downstream clustering tasks, and evaluation of hyperparameter sensitivity, demonstrate the efficacy of the proposed method. The source code is available at this https URL .\n\n---\n\n## 54. DAST: A Dual-Stream Voice Anonymization Attacker with Staged Training\n- **Authors**: Ridwan Arefeen , Xiaoxiao Miao , Rong Tong , Aik Beng Ng , Simon See , Timothy Liu\n- **URL**: [https://arxiv.org/abs/2603.12840](https://arxiv.org/abs/2603.12840)\n- **Abstract**:\n> Voice anonymization masks vocal traits while preserving linguistic content, which may still leak speaker-specific patterns. To assess and strengthen privacy evaluation, we propose a dual-stream attacker that fuses spectral and self-supervised learning features via parallel encoders with a three-stage training strategy. Stage I establishes foundational speaker-discriminative representations. Stage II leverages the shared identity-transformation characteristics of voice conversion and anonymization, exposing the model to diverse converted speech to build cross-system robustness. Stage III provides lightweight adaptation to target anonymized data. Results on the VoicePrivacy Attacker Challenge (VPAC) dataset demonstrate that Stage II is the primary driver of generalization, enabling strong attacking performance on unseen anonymization datasets. With Stage III, fine-tuning on only 10\\% of the target anonymization dataset surpasses current state-of-the-art attackers in terms of EER.\n\n---\n\n## 55. Mask2Flow-TSE: Two-Stage Target Speaker Extraction with Masking and Flow Matching\n- **Authors**: Junwon Moon , Hyunjin Choi , Hansol Park , Heeseung Kim , Kyuhong Shim\n- **URL**: [https://arxiv.org/abs/2603.12837](https://arxiv.org/abs/2603.12837)\n- **Abstract**:\n> Target speaker extraction (TSE) extracts the target speaker's voice from overlapping speech mixtures given a reference utterance. Existing approaches typically fall into two categories: discriminative and generative. Discriminative methods apply time-frequency masking for fast inference but often over-suppress the target signal, while generative methods synthesize high-quality speech at the cost of numerous iterative steps. We propose Mask2Flow-TSE, a two-stage framework combining the strengths of both paradigms. The first stage applies discriminative masking for coarse separation, and the second stage employs flow matching to refine the output toward target speech. Unlike generative approaches that synthesize speech from Gaussian noise, our method starts from the masked spectrogram, enabling high-quality reconstruction in a single inference step. Experiments show that Mask2Flow-TSE achieves comparable performance to existing generative TSE methods with approximately 85M parameters.\n\n---\n\n## 56. Hierarchical Dual-Change Collaborative Learning for UAV Scene Change Captioning\n- **Authors**: Fuhai Chen , Pengpeng Huang , Junwen Wu , Hehong Zhang , Shiping Wang , Xiaoguang Ma , Xuri Ge\n- **URL**: [https://arxiv.org/abs/2603.12832](https://arxiv.org/abs/2603.12832)\n- **Abstract**:\n> This paper proposes a novel task for UAV scene understanding - UAV Scene Change Captioning (UAV-SCC) - which aims to generate natural language descriptions of semantic changes in dynamic aerial imagery captured from a movable viewpoint. Unlike traditional change captioning that mainly describes differences between image pairs captured from a fixed camera viewpoint over time, UAV scene change captioning focuses on image-pair differences resulting from both temporal and spatial scene variations dynamically captured by a moving camera. The key challenge lies in understanding viewpoint-induced scene changes from UAV image pairs that share only partially overlapping scene content due to viewpoint shifts caused by camera rotation, while effectively exploiting the relative orientation between the two images. To this end, we propose a Hierarchical Dual-Change Collaborative Learning (HDC-CL) method for UAV scene change captioning. In particular, a novel transformer, \\emph{i.e.} Dynamic Adaptive Layout Transformer (DALT) is designed to adaptively model diverse spatial layouts of the image pair, where the interrelated features derived from the overlapping and non-overlapping regions are learned within the flexible and unified encoding layer. Furthermore, we propose a Hierarchical Cross-modal Orientation Consistency Calibration (HCM-OCC) method to enhance the model's sensitivity to viewpoint shift directions, enabling more accurate change captioning. To facilitate in-depth research on this task, we construct a new benchmark dataset, named UAV-SCC dataset, for UAV scene change captioning. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance on this task. The dataset and code will be publicly released upon acceptance of this paper.\n\n---\n\n## 57. Residual SODAP: Residual Self-Organizing Domain-Adaptive Prompting with Structural Knowledge Preservation for Continual Learning\n- **Authors**: Gyutae Oh , Jungwoo Bae , Jitae Shin\n- **URL**: [https://arxiv.org/abs/2603.12816](https://arxiv.org/abs/2603.12816)\n- **Abstract**:\n> Continual learning (CL) suffers from catastrophic forgetting, which is exacerbated in domain-incremental learning (DIL) where task identifiers are unavailable and storing past data is infeasible. While prompt-based CL (PCL) adapts representations with a frozen backbone, we observe that prompt-only improvements are often insufficient due to suboptimal prompt selection and classifier-level instability under domain shifts. We propose Residual SODAP, which jointly performs prompt-based representation adaptation and classifier-level knowledge preservation. Our framework combines $\\alpha$-entmax sparse prompt selection with residual aggregation, data-free distillation with pseudo-feature replay, prompt-usage--based drift detection, and uncertainty-aware multi-loss balancing. Across three DIL benchmarks without task IDs or extra data storage, Residual SODAP achieves state-of-the-art AvgACC/AvgF of 0.850/0.047 (DR), 0.760/0.031 (Skin Cancer), and 0.995/0.003 (CORe50).\n\n---\n\n## 58. Cheers: Decoupling Patch Details from Semantic Representations Enables Unified Multimodal Comprehension and Generation\n- **Authors**: Yichen Zhang , Da Peng , Zonghao Guo , Zijian Zhang , Xuesong Yang , Tong Sun , Shichu Sun , Yidan Zhang , Yanghao Li , Haiyan Zhao , Wang Xu , Qi Shi , Yangang Sun , Chi Chen , Shuo Wang , Yukun Yan , Xu Han , Qiang Ma , Wei Ke , Liang Wang , Zhiyuan Liu , Maosong Sun\n- **URL**: [https://arxiv.org/abs/2603.12793](https://arxiv.org/abs/2603.12793)\n- **Abstract**:\n> A recent cutting-edge topic in multimodal modeling is to unify visual comprehension and generation within a single model. However, the two tasks demand mismatched decoding regimes and visual representations, making it non-trivial to jointly optimize within a shared feature space. In this work, we present Cheers, a unified multimodal model that decouples patch-level details from semantic representations, thereby stabilizing semantics for multimodal understanding and improving fidelity for image generation via gated detail residuals. Cheers includes three key components: (i) a unified vision tokenizer that encodes and compresses image latent states into semantic tokens for efficient LLM conditioning, (ii) an LLM-based Transformer that unifies autoregressive decoding for text generation and diffusion decoding for image generation, and (iii) a cascaded flow matching head that decodes visual semantics first and then injects semantically gated detail residuals from the vision tokenizer to refine high-frequency content. Experiments on popular benchmarks demonstrate that Cheers matches or surpasses advanced UMMs in both visual understanding and generation. Cheers also achieves 4x token compression, enabling more efficient high-resolution image encoding and generation. Notably, Cheers outperforms the Tar-1.5B on the popular benchmarks GenEval and MMBench, while requiring only 20% of the training cost, indicating effective and efficient (i.e., 4x token compression) unified multimodal modeling. We will release all code and data for future research.\n\n---\n\n## 59. The RIGID Framework: Research-Integrated, Generative AI-Mediated Instructional Design\n- **Authors**: Yerin Kwak , Zachary A. Pardos\n- **URL**: [https://arxiv.org/abs/2603.12781](https://arxiv.org/abs/2603.12781)\n- **Abstract**:\n> Instructional Design (ID) often faces challenges in incorporating research-based knowledge and pedagogical best practices. Although educational researchers and government agencies emphasize grounding ID in evidence, integrating research findings into everyday design workflows is often complex, as it requires considering multiple context-specific demands and constraints. To address this persistent gap, this paper explores how research in the learning sciences (LS) can be systematically integrated across ID workflows and how recent advances in generative AI can help operationalize this integration. While ID and LS share a commitment to improving learning experiences through design-oriented approaches in authentic contexts, structured integration between the two fields remains limited, leaving their complementary insights underutilized. We present RIGID (Research-Integrated, Generative AI-Mediated Instructional Design), a unified framework that integrates LS research across ID workflows spanning analysis, design, implementation, and evaluation phases, while leveraging generative AI to mediate this integration at each stage. The RIGID framework provides a systematic approach for enabling research-integrated instructional design that is both operational and context-sensitive, while preserving the central role of human expertise.\n\n---\n\n## 60. Empowering Semantic-Sensitive Underwater Image Enhancement with VLM\n- **Authors**: Guodong Fan , Shengning Zhou , Genji Yuan , Huiyu Li , Jingchun Zhou , Jinjiang Li\n- **URL**: [https://arxiv.org/abs/2603.12773](https://arxiv.org/abs/2603.12773)\n- **Abstract**:\n> In recent years, learning-based underwater image enhancement (UIE) techniques have rapidly evolved. However, distribution shifts between high-quality enhanced outputs and natural images can hinder semantic cue extraction for downstream vision tasks, thereby limiting the adaptability of existing enhancement models. To address this challenge, this work proposes a new learning mechanism that leverages Vision-Language Models (VLMs) to empower UIE models with semantic-sensitive capabilities. To be concrete, our strategy first generates textual descriptions of key objects from a degraded image via VLMs. Subsequently, a text-image alignment model remaps these relevant descriptions back onto the image to produce a spatial semantic guidance map. This map then steers the UIE network through a dual-guidance mechanism, which combines cross-attention and an explicit alignment loss. This forces the network to focus its restorative power on semantic-sensitive regions during image reconstruction, rather than pursuing a globally uniform improvement, thereby ensuring the faithful restoration of key object features. Experiments confirm that when our strategy is applied to different UIE baselines, significantly boosts their performance on perceptual quality metrics as well as enhances their performance on detection and segmentation tasks, validating its effectiveness and adaptability.\n\n---\n\n## 61. FC-Track: Overlap-Aware Post-Association Correction for Online Multi-Object Tracking\n- **Authors**: Cheng Ju , Zejing Zhao , Akio Namiki\n- **URL**: [https://arxiv.org/abs/2603.12758](https://arxiv.org/abs/2603.12758)\n- **Abstract**:\n> Reliable multi-object tracking (MOT) is essential for robotic systems operating in complex and dynamic environments. Despite recent advances in detection and association, online MOT methods remain vulnerable to identity switches caused by frequent occlusions and object overlap, where incorrect associations can propagate over time and degrade tracking reliability. We present a lightweight post-association correction framework (FC-Track) for online MOT that explicitly targets overlap-induced mismatches during inference. The proposed method suppresses unreliable appearance updates under high-overlap conditions using an Intersection over Area (IoA)-based filtering strategy, and locally corrects detection-to-tracklet mismatches through appearance similarity comparison within overlapped tracklet pairs. By preventing short-term mismatches from propagating, our framework effectively mitigates long-term identity switches without resorting to global optimization or re-identification. The framework operates online without global optimization or re-identification, making it suitable for real-time robotic applications. We achieve 81.73 MOTA, 82.81 IDF1, and 66.95 HOTA on the MOT17 test set with a running speed of 5.7 FPS, and 77.52 MOTA, 80.90 IDF1, and 65.67 HOTA on the MOT20 test set with a running speed of 0.6 FPS. Specifically, our framework FC-Track produces only 29.55% long-term identity switches, which is substantially lower than existing online trackers. Meanwhile, our framework maintains state-of-the-art performance on the MOT20 benchmark.\n\n---\n\n## 62. TaoBench: Do Automated Theorem Prover LLMs Generalize Beyond MathLib?\n- **Authors**: Alexander K Taylor , Junyi Zhang , Ethan Ji , Vigyan Sahai , Haikang Deng , Yuanzhou Chen , Yifan Yuan , Di Wu , Jia-Chen Gu , Kai-Wei Chang , Nanyun Peng , Amit Sahai , Wei Wang\n- **URL**: [https://arxiv.org/abs/2603.12744](https://arxiv.org/abs/2603.12744)\n- **Abstract**:\n> Automated theorem proving (ATP) benchmarks largely consist of problems formalized in MathLib, so current ATP training and evaluation are heavily biased toward MathLib's definitional framework. However, frontier mathematics is often exploratory and prototype-heavy, relying on bespoke constructions that deviate from standard libraries. In this work, we evaluate the robustness of current ATP systems when applied to a novel definitional framework, specifically examining the performance gap between standard library problems and bespoke mathematical constructions. We introduce TaoBench, an undergraduate-level benchmark derived from Terence Tao's Analysis I, which formalizes analysis by constructing core mathematical concepts from scratch, without relying on standard Mathlib definitions, as well as by mixing from-scratch and MathLib constructions. For fair evaluation, we build an agentic pipeline that automatically extracts a compilable, self-contained local environment for each problem. To isolate the effect of definitional frameworks, we additionally translate every problem into a mathematically equivalent Mathlib formulation, yielding paired TaoBench-Mathlib statements for direct comparison. While state-of-the-art ATP models perform capably within the MathLib framework, performance drops by an average of roughly 26% on the definitionally equivalent Tao formulation. This indicates that the main bottleneck is limited generalization across definitional frameworks rather than task difficulty. TaoBench thus highlights a gap between benchmark performance and applicability, and provides a concrete foundation for developing and testing provers better aligned with research mathematics.\n\n---\n\n## 63. MoKus: Leveraging Cross-Modal Knowledge Transfer for Knowledge-Aware Concept Customization\n- **Authors**: Chenyang Zhu , Hongxiang Li , Xiu Li , Long Chen\n- **URL**: [https://arxiv.org/abs/2603.12743](https://arxiv.org/abs/2603.12743)\n- **Abstract**:\n> Concept customization typically binds rare tokens to a target concept. Unfortunately, these approaches often suffer from unstable performance as the pretraining data seldom contains these rare tokens. Meanwhile, these rare tokens fail to convey the inherent knowledge of the target concept. Consequently, we introduce Knowledge-aware Concept Customization, a novel task aiming at binding diverse textual knowledge to target visual concepts. This task requires the model to identify the knowledge within the text prompt to perform high-fidelity customized generation. Meanwhile, the model should efficiently bind all the textual knowledge to the target concept. Therefore, we propose MoKus, a novel framework for knowledge-aware concept customization. Our framework relies on a key observation: cross-modal knowledge transfer, where modifying knowledge within the text modality naturally transfers to the visual modality during generation. Inspired by this observation, MoKus contains two stages: (1) In visual concept learning, we first learn the anchor representation to store the visual information of the target concept. (2) In textual knowledge updating, we update the answer for the knowledge queries to the anchor representation, enabling high-fidelity customized generation. To further comprehensively evaluate our proposed MoKus on the new task, we introduce the first benchmark for knowledge-aware concept customization: KnowCusBench. Extensive evaluations have demonstrated that MoKus outperforms state-of-the-art methods. Moreover, the cross-model knowledge transfer allows MoKus to be easily extended to other knowledge-aware applications like virtual concept creation and concept erasure. We also demonstrate the capability of our method to achieve improvements on world knowledge benchmarks.\n\n---\n\n## 64. SRAM-Based Compute-in-Memory Accelerator for Linear-decay Spiking Neural Networks\n- **Authors**: Hongyang Shang , Shuai Dong , Yahan Yang , Junyi Yang , Peng Zhou , Arindam Basu\n- **URL**: [https://arxiv.org/abs/2603.12739](https://arxiv.org/abs/2603.12739)\n- **Abstract**:\n> Spiking Neural Networks (SNNs) have emerged as a biologically inspired alternative to conventional deep networks, offering event-driven and energy-efficient computation. However, their throughput remains constrained by the serial update of neuron membrane states. While many hardware accelerators and Compute-in-Memory (CIM) architectures efficiently parallelize the synaptic operation (W x I) achieving O(1) complexity for matrix-vector multiplication, the subsequent state update step still requires O(N) time to refresh all neuron membrane potentials. This mismatch makes state update the dominant latency and energy bottleneck in SNN inference. To address this challenge, we propose an SRAM-based CIM for SNN with Linear Decay Leaky Integrate-and-Fire (LD-LIF) Neuron that co-optimizes algorithm and hardware. At the algorithmic level, we replace the conventional exponential membrane decay with a linear decay approximation, converting costly multiplications into simple additions while accuracy drops only around 1%. At the architectural level, we introduce an in-memory parallel update scheme that performs in-place decay directly within the SRAM array, eliminating the need for global sequential updates. Evaluated on benchmark SNN workloads, the proposed method achieves a 1.1 x to 16.7 x reduction of SOP energy consumption, while providing 15.9 x to 69 x more energy efficiency, with negligible accuracy loss relative to original decay models. This work highlights that beyond accelerating the (W x I) computation, optimizing state-update dynamics within CIM architectures is essential for scalable, low-power, and real-time neuromorphic processing.\n\n---\n\n## 65. Graph In-Context Operator Networks for Generalizable Spatiotemporal Prediction\n- **Authors**: Chenghan Wu , Zongmin Yu , Boai Sun , Liu Yang\n- **URL**: [https://arxiv.org/abs/2603.12725](https://arxiv.org/abs/2603.12725)\n- **Abstract**:\n> In-context operator learning enables neural networks to infer solution operators from contextual examples without weight updates. While prior work has demonstrated the effectiveness of this paradigm in leveraging vast datasets, a systematic comparison against single-operator learning using identical training data has been absent. We address this gap through controlled experiments comparing in-context operator learning against classical operator learning (single-operator models trained without contextual examples), under the same training steps and dataset. To enable this investigation on real-world spatiotemporal systems, we propose GICON (Graph In-Context Operator Network), combining graph message passing for geometric generalization with example-aware positional encoding for cardinality generalization. Experiments on air quality prediction across two Chinese regions show that in-context operator learning outperforms classical operator learning on complex tasks, generalizing across spatial domains and scaling robustly from few training examples to 100 at inference.\n\n---\n\n## 66. CognitionCapturerPro: Towards High-Fidelity Visual Decoding from EEG/MEG via Multi-modal Information and Asymmetric Alignment\n- **Authors**: Kaifan Zhang , Lihuo He , Junjie Ke , Yuqi Ji , Lukun Wu , Lizi Wang , Xinbo Gao\n- **URL**: [https://arxiv.org/abs/2603.12722](https://arxiv.org/abs/2603.12722)\n- **Abstract**:\n> Visual stimuli reconstruction from EEG remains challenging due to fidelity loss and representation shift. We propose CognitionCapturerPro, an enhanced framework that integrates EEG with multi-modal priors (images, text, depth, and edges) via collaborative training. Our core contributions include an uncertainty-weighted similarity scoring mechanism to quantify modality-specific fidelity and a fusion encoder for integrating shared representations. By employing a simplified alignment module and a pre-trained diffusion model, our method significantly outperforms the original CognitionCapturer on the THINGS-EEG dataset, improving Top-1 and Top-5 retrieval accuracy by 25.9% and 10.6%, respectively. Code is available at: this https URL .\n\n---\n\n## 67. CMHANet: A Cross-Modal Hybrid Attention Network for Point Cloud Registration\n- **Authors**: Dongxu Zhang , Yingsen Wang , Yiding Sun , Haoran Xu , Peilin Fan , Jihua Zhu\n- **URL**: [https://arxiv.org/abs/2603.12721](https://arxiv.org/abs/2603.12721)\n- **Abstract**:\n> Robust point cloud registration is a fundamental task in 3D computer vision and geometric deep learning, essential for applications such as large-scale 3D reconstruction, augmented reality, and scene understanding. However, the performance of established learning-based methods often degrades in complex, real world scenarios characterized by incomplete data, sensor noise, and low overlap regions. To address these limitations, we propose CMHANet, a novel Cross-Modal Hybrid Attention Network. Our method integrates the fusion of rich contextual information from 2D images with the geometric detail of 3D point clouds, yielding a comprehensive and resilient feature representation. Furthermore, we introduce an innovative optimization function based on contrastive learning, which enforces geometric consistency and significantly improves the model's robustness to noise and partial observations. We evaluated CMHANet on the 3DMatch and the challenging 3DLoMatch datasets. \\rev{Additionally, zero-shot evaluations on the TUM RGB-D SLAM dataset verify the model's generalization capability to unseen domains.} The experimental results demonstrate that our method achieves substantial improvements in both registration accuracy and overall robustness, outperforming current techniques. We also release our code in \\href{ this https URL }{ this https URL }.\n\n---\n\n## 68. IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registration\n- **Authors**: Dongxu Zhang , Jihua Zhu , Shiqi Li , Wenbiao Yan , Haoran Xu , Peilin Fan , Huimin Lu\n- **URL**: [https://arxiv.org/abs/2603.12719](https://arxiv.org/abs/2603.12719)\n- **Abstract**:\n> Point cloud registration (PCR) is a fundamental task in 3D vision and provides essential support for applications such as autonomous driving, robotics, and environmental modeling. Despite its widespread use, existing methods often fail when facing real-world challenges like heavy noise, significant occlusions, and large-scale transformations. These limitations frequently result in compromised registration accuracy and insufficient robustness in complex environments. In this paper, we propose IGASA as a novel registration framework constructed upon a Hierarchical Pyramid Architecture (HPA) designed for robust multi-scale feature extraction and fusion. The framework integrates two pivotal components consisting of the Hierarchical Cross-Layer Attention (HCLA) module and the Iterative Geometry-Aware Refinement (IGAR) module. The HCLA module utilizes skip attention mechanisms to align multi-resolution features and enhance local geometric consistency. Simultaneously, the IGAR module is designed for the fine matching phase by leveraging reliable correspondences established during coarse matching. This synergistic integration within the architecture allows IGASA to adapt effectively to diverse point cloud structures and intricate transformations. We evaluate the performance of IGASA on four widely recognized benchmark datasets including 3D(Lo)Match, KITTI, and nuScenes. Our extensive experiments consistently demonstrate that IGASA significantly surpasses state-of-the-art methods and achieves notable improvements in registration accuracy. This work provides a robust foundation for advancing point cloud registration techniques while offering valuable insights for practical 3D vision applications. The code for IGASA is available in \\href{ this https URL }{ this https URL }.\n\n---\n\n## 69. Altered Thoughts, Altered Actions: Probing Chain-of-Thought Vulnerabilities in VLA Robotic Manipulation\n- **Authors**: Tuan Duong Trinh , Naveed Akhtar , Basim Azam\n- **URL**: [https://arxiv.org/abs/2603.12717](https://arxiv.org/abs/2603.12717)\n- **Abstract**:\n> Recent Vision-Language-Action (VLA) models increasingly adopt chain-of-thought (CoT) reasoning, generating a natural-language plan before decoding motor commands. This internal text channel between the reasoning module and the action decoder has received no adversarial scrutiny. We ask: which properties of this intermediate plan does the action decoder actually rely on, and can targeted corruption of the reasoning trace alone -- with all inputs left intact -- degrade a robot's physical task performance? We design a taxonomy of seven text corruptions organized into three attacker tiers (blind noise, mechanical-semantic, and LLM-adaptive) and apply them to a state-of-the-art reasoning VLA across 40 LIBERO tabletop manipulation tasks. Our results reveal a striking asymmetry: substituting object names in the reasoning trace reduces overall success rate by 8.3~percentage points (pp) -- reaching $-$19.3~pp on goal-conditioned tasks and $-$45~pp on individual tasks -- whereas sentence reordering, spatial-direction reversal, token noise, and even a 70B-parameter LLM crafting plausible-but-wrong plans all have negligible impact (within $\\pm$4~pp). This asymmetry indicates that the action decoder depends on entity-reference integrity rather than reasoning quality or sequential structure. Notably, a sophisticated LLM-based attacker underperforms simple mechanical object-name substitution, because preserving plausibility inadvertently retains the entity-grounding structure the decoder needs. A cross-architecture control using a non-reasoning VLA confirms the vulnerability is exclusive to reasoning-augmented models, while instruction-level attacks degrade both architectures -- establishing that the internal reasoning trace is a distinct and stealthy threat vector invisible to input-validation defenses.\n\n---\n\n## 70. Cost-Efficient Multimodal LLM Inference via Cross-Tier GPU Heterogeneity\n- **Authors**: Donglin Yu\n- **URL**: [https://arxiv.org/abs/2603.12707](https://arxiv.org/abs/2603.12707)\n- **Abstract**:\n> Multimodal large language model (MLLM) inference splits into two phases with opposing hardware demands: vision encoding is compute-bound, while language generation is memory-bandwidth-bound. We show that under standard transformer KV caching, the modality boundary (between vision encoder and language model) minimizes cross-device transfer among all partition points that preserve standard stage-based execution. Partitioning here reduces transfer complexity from $O(L * s_ctx)$ bytes (GB-scale KV caches under stage-level disaggregation) to $O(N_v * d)$ bytes (MB-scale embeddings), an O(L) reduction where L is the transformer depth. The result holds across attention mechanisms (MHA/GQA), dynamic vision resolutions, and model scales, and the advantage grows as models deepen. A direct implication is that existing stage-level disaggregation systems are constrained to high-bandwidth interconnects (e.g., NVLink), whereas modality-level disaggregation enables cross-tier heterogeneous serving over commodity PCIe. A closed-form cost model shows that heterogeneous deployment is cost-optimal under phase-separable workloads (predicts 31.4% savings; observed 40.6%). We build HeteroServe, a phase-aware runtime with modality-level partitioning and cross-tier scheduling, and evaluate it on LLaVA-1.5-7B and Qwen2.5-VL against vLLM v0.3.0. On identical 4xA100 hardware, engine optimizations raise throughput by up to 54%. Under a fixed budget, a heterogeneous cluster (\\$38k) improves Tokens/\\$ by 37% over a homogeneous baseline (\\$64k) without degrading latency.\n\n---\n\n## 71. Seeing Eye to Eye: Enabling Cognitive Alignment Through Shared First-Person Perspective in Human-AI Collaboration\n- **Authors**: Zhuyu Teng , Pei Chen , Yichen Cai , Ruoqing Lu , Zhaoqu Jiang , Jiayang Li , Weitao You , Lingyun Sun\n- **URL**: [https://arxiv.org/abs/2603.12701](https://arxiv.org/abs/2603.12701)\n- **Abstract**:\n> Despite advances in multimodal AI, current vision-based assistants often remain inefficient in collaborative tasks. We identify two key gulfs: a communication gulf, where users must translate rich parallel intentions into verbal commands due to the channel mismatch , and an understanding gulf, where AI struggles to interpret subtle embodied cues. To address these, we propose Eye2Eye, a framework that leverages first-person perspective as a channel for human-AI cognitive alignment. It integrates three components: (1) joint attention coordination for fluid focus alignment, (2) revisable memory to maintain evolving common ground, and (3) reflective feedback allowing users to clarify and refine AI's understanding. We implement this framework in an AR prototype and evaluate it through a user study and a post-hoc pipeline evaluation. Results show that Eye2Eye significantly reduces task completion time and interaction load while increasing trust, demonstrating its components work in concert to improve collaboration.\n\n---\n\n## 72. HSEmotion Team at ABAW-10 Competition: Facial Expression Recognition, Valence-Arousal Estimation, Action Unit Detection and Fine-Grained Violence Classification\n- **Authors**: Andrey V. Savchenko , Kseniia Tsypliakova\n- **URL**: [https://arxiv.org/abs/2603.12693](https://arxiv.org/abs/2603.12693)\n- **Abstract**:\n> This article presents our results for the 10th Affective Behavior Analysis in-the-Wild (ABAW) competition. For frame-wise facial emotion understanding tasks (frame-wise facial expression recognition, valence-arousal estimation, action unit detection), we propose a fast approach based on facial embedding extraction with pre-trained EfficientNet-based emotion recognition models. If the latter model's confidence exceeds a threshold, its prediction is used. Otherwise, we feed embeddings into a simple multi-layered perceptron trained on the AffWild2 dataset. Estimated class-level scores are smoothed in a sliding window of fixed size to mitigate noise in frame-wise predictions. For the fine-grained violence detection task, we examine several pre-trained architectures for frame embeddings and their aggregation for video classification. Experimental results on four tasks from the ABAW challenge demonstrate that our approach significantly improves validation metrics over existing baselines.\n\n---\n\n## 73. Federated Hierarchical Clustering with Automatic Selection of Optimal Cluster Numbers\n- **Authors**: Yue Zhang , Chuanlong Qiu , Xinfa Liao , Yiqun Zhang\n- **URL**: [https://arxiv.org/abs/2603.12684](https://arxiv.org/abs/2603.12684)\n- **Abstract**:\n> Federated Clustering (FC) is an emerging and promising solution in exploring data distribution patterns from distributed and privacy-protected data in an unsupervised manner. Existing FC methods implicitly rely on the assumption that clients are with a known number of uniformly sized clusters. However, the true number of clusters is typically unknown, and cluster sizes are naturally imbalanced in real scenarios. Furthermore, the privacy-preserving transmission constraints in federated learning inevitably reduce usable information, making the development of robust and accurate FC extremely challenging. Accordingly, we propose a novel FC framework named Fed-$k^*$-HC, which can automatically determine an optimal number of clusters $k^*$ based on the data distribution explored through hierarchical clustering. To obtain the global data distribution for $k^*$ determination, we let each client generate micro-subclusters. Their prototypes are then uploaded to the server for hierarchical merging. The density-based merging design allows exploring clusters of varying sizes and shapes, and the progressive merging process can self-terminate according to the neighboring relationships among the prototypes to determine $k^*$. Extensive experiments on diverse datasets demonstrate the FC capability of the proposed Fed-$k^*$-HC in accurately exploring a proper number of clusters.\n\n---\n\n## 74. Experimental evidence of progressive ChatGPT models self-convergence\n- **Authors**: Konstantinos F. Xylogiannopoulos , Petros Xanthopoulos , Panagiotis Karampelas , Georgios A. Bakamitsos\n- **URL**: [https://arxiv.org/abs/2603.12683](https://arxiv.org/abs/2603.12683)\n- **Abstract**:\n> Large Language Models (LLMs) that undergo recursive training on synthetically generated data are susceptible to model collapse, a phenomenon marked by the generation of meaningless output. Existing research has examined this issue from either theoretical or empirical perspectives, often focusing on a single model trained recursively on its own outputs. While prior studies have cautioned against the potential degradation of LLM output quality under such conditions, no longitudinal investigation has yet been conducted to assess this effect over time. In this study, we employ a text similarity metric to evaluate different ChatGPT models' capacity to generate diverse textual outputs. Our findings indicate a measurable decline of recent ChatGPT releases' ability to produce varied text, even when explicitly prompted to do so, by setting the temperature parameter to one. The observed reduction in output diversity may be attributed to the influence of the amounts of synthetic data incorporated within their training datasets as the result of internet infiltration by LLM generated data. The phenomenon is defined as model self-convergence because of the gradual increase of similarities of produced texts among different ChatGPT versions.\n\n---\n\n## 75. MetaKE: Meta-learning Aligned Knowledge Editing via Bi-level Optimization\n- **Authors**: Shuxin Liu , Ou Wu\n- **URL**: [https://arxiv.org/abs/2603.12677](https://arxiv.org/abs/2603.12677)\n- **Abstract**:\n> Knowledge editing (KE) aims to precisely rectify specific knowledge in Large Language Models (LLMs) without disrupting general capabilities. State-of-the-art methods suffer from an open-loop control mismatch. We identify a critical \"Semantic-Execution Disconnect\": the semantic target is derived independently without feedback from the downstream's feasible region. This misalignment often causes valid semantic targets to fall within the prohibited space, resulting in gradient truncation and editing failure. To bridge this gap, we propose MetaKE (Meta-learning Aligned Knowledge Editing), a new framework that reframes KE as a bi-level optimization problem. Departing from static calculation, MetaKE treats the edit target as a learnable meta-parameter: the upper-level optimizer seeks a feasible target to maximize post-edit performance, while the lower-level solver executes the editing. To address the challenge of differentiating through complex solvers, we derive a Structural Gradient Proxy, which explicitly backpropagates editability constraints to the target learning phase. Theoretical analysis demonstrates that MetaKE automatically aligns the edit direction with the model's feasible manifold. Extensive experiments confirm that MetaKE significantly outperforms strong baselines, offering a new perspective on knowledge editing.\n\n---\n\n## 76. Marker-Based 3D Reconstruction of Aggregates with a Comparative Analysis of 2D and 3D Morphologies\n- **Authors**: Haohang Huang , Jiayi Luo , Issam Qamhia , Erol Tutumluer , John M. Hart , Andrew J. Stolba\n- **URL**: [https://arxiv.org/abs/2603.12667](https://arxiv.org/abs/2603.12667)\n- **Abstract**:\n> Aggregates, serving as the main skeleton in assemblies of construction materials, are important functional components in various building and transportation infrastructures. They can be used in unbound layer applications, e.g. pavement base and railroad ballast, bound applications of cement concrete and asphalt concrete, and as riprap and large-sized primary crushed rocks. Information on the size and shape or morphology of aggregates can greatly facilitate the Quality Assurance/Quality Control (QA/QC) process by providing insights of aggregate behavior during composition and packing. A full 3D characterization of aggregate particle morphology is difficult both during production in a quarry and at a construction site. Many aggregate imaging approaches have been developed to quantify the particle morphology by computer vision, including 2D image-based approaches that analyze particle silhouettes and 3D scanning-based methods that require expensive devices such as 3D laser scanners or X-Ray Computed Tomography (CT) equipment. This paper presents a flexible and cost-effective photogrammetry-based approach for the 3D reconstruction of aggregate particles. The proposed approach follows a marker-based design that enables background suppression, point cloud stitching, and scale referencing to obtain high-quality aggregate models. The accuracy of the reconstruction results was validated against ground-truth for selected aggregate samples. Comparative analyses were conducted on 2D and 3D morphological properties of the selected samples. Significant differences were found between the 2D and 3D statistics. Based on the presented approach, 3D shape information of aggregates can be obtained easily and at a low cost, thus allowing convenient aggregate inspection, data collection, and 3D morphological analysis.\n\n---\n\n## 77. RetroReasoner: A Reasoning LLM for Strategic Retrosynthesis Prediction\n- **Authors**: Hanbum Ko , Chanhui Lee , Ye Rin Kim , Rodrigo Hormazabal , Sehui Han , Sungbin Lim , Sungwoong Kim\n- **URL**: [https://arxiv.org/abs/2603.12666](https://arxiv.org/abs/2603.12666)\n- **Abstract**:\n> Retrosynthesis prediction is a core task in organic synthesis that aims to predict reactants for a given product molecule. Traditionally, chemists select a plausible bond disconnection and derive corresponding reactants, which is time-consuming and requires substantial expertise. While recent advancements in molecular large language models (LLMs) have made progress, many methods either predict reactants without strategic reasoning or conduct only a generic product analysis, rather than reason explicitly about bond-disconnection strategies that logically lead to the choice of specific reactants. To overcome these limitations, we propose RetroReasoner, a retrosynthetic reasoning model that leverages chemists' strategic thinking. RetroReasoner is trained using both supervised fine-tuning (SFT) and reinforcement learning (RL). For SFT, we introduce SyntheticRetro, a framework that generates structured disconnection rationales alongside reactant predictions. In the case of RL, we apply a round-trip accuracy as reward, where predicted reactants are passed through a forward synthesis model, and predictions are rewarded when the forward-predicted product matches the original input product. Experimental results show that RetroReasoner not only outperforms prior baselines but also generates a broader range of feasible reactant proposals, particularly in handling more challenging reaction instances.\n\n---\n\n## 78. From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space\n- **Authors**: Lehui Li , Yuyao Wang , Jisheng Yan , Wei Zhang , Jinliang Deng , Haoliang Sun , Zhongyi Han , Yongshun Gong\n- **URL**: [https://arxiv.org/abs/2603.12664](https://arxiv.org/abs/2603.12664)\n- **Abstract**:\n> Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.\n\n---\n\n## 79. Continual Learning in Large Language Models: Methods, Challenges, and Opportunities\n- **Authors**: Hongyang Chen , Zhongwu Sun , Hongfei Ye , Kunchi Li , Xuemin Lin\n- **URL**: [https://arxiv.org/abs/2603.12658](https://arxiv.org/abs/2603.12658)\n- **Abstract**:\n> Continual learning (CL) has emerged as a pivotal paradigm to enable large language models (LLMs) to dynamically adapt to evolving knowledge and sequential tasks while mitigating catastrophic forgetting-a critical limitation of the static pre-training paradigm inherent to modern LLMs. This survey presents a comprehensive overview of CL methodologies tailored for LLMs, structured around three core training stages: continual pre-training, continual fine-tuning, and continual this http URL the canonical taxonomy of rehearsal-, regularization-, and architecture-based methods, we further subdivide each category by its distinct forgetting mitigation mechanisms and conduct a rigorous comparative analysis of the adaptability and critical improvements of traditional CL methods for LLMs. In doing so, we explicitly highlight core distinctions between LLM CL and traditional machine learning, particularly with respect to scale, parameter efficiency, and emergent capabilities. Our analysis covers essential evaluation metrics, including forgetting rates and knowledge transfer efficiency, along with emerging benchmarks for assessing CL performance. This survey reveals that while current methods demonstrate promising results in specific domains, fundamental challenges persist in achieving seamless knowledge integration across diverse tasks and temporal scales. This systematic review contributes to the growing body of knowledge on LLM adaptation, providing researchers and practitioners with a structured framework for understanding current achievements and future opportunities in lifelong learning for language models.\n\n---\n\n## 80. LR-SGS: Robust LiDAR-Reflectance-Guided Salient Gaussian Splatting for Self-Driving Scene Reconstruction\n- **Authors**: Ziyu Chen , Fan Zhu , Hui Zhu , Deyi Kong , Xinkai Kuang , Yujia Zhang , Chunmao Jiang\n- **URL**: [https://arxiv.org/abs/2603.12647](https://arxiv.org/abs/2603.12647)\n- **Abstract**:\n> Recent 3D Gaussian Splatting (3DGS) methods have demonstrated the feasibility of self-driving scene reconstruction and novel view synthesis. However, most existing methods either rely solely on cameras or use LiDAR only for Gaussian initialization or depth supervision, while the rich scene information contained in point clouds, such as reflectance, and the complementarity between LiDAR and RGB have not been fully exploited, leading to degradation in challenging self-driving scenes, such as those with high ego-motion and complex lighting. To address these issues, we propose a robust and efficient LiDAR-reflectance-guided Salient Gaussian Splatting method (LR-SGS) for self-driving scenes, which introduces a structure-aware Salient Gaussian representation, initialized from geometric and reflectance feature points extracted from LiDAR and refined through a salient transform and improved density control to capture edge and planar structures. Furthermore, we calibrate LiDAR intensity into reflectance and attach it to each Gaussian as a lighting-invariant material channel, jointly aligned with RGB to enforce boundary consistency. Extensive experiments on the Waymo Open Dataset demonstrate that LR-SGS achieves superior reconstruction performance with fewer Gaussians and shorter training time. In particular, on Complex Lighting scenes, our method surpasses OmniRe by 1.18 dB PSNR.\n\n---\n\n## 81. LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing\n- **Authors**: Jiawei Hao , Zhiwei Hao , Jianyuan Guo , Li Shen , Yong Luo , Han Hu , Dan Zeng\n- **URL**: [https://arxiv.org/abs/2603.12645](https://arxiv.org/abs/2603.12645)\n- **Abstract**:\n> Mixture-of-Experts (MoE) based Large Language Models (LLMs) have demonstrated impressive performance and computational efficiency. However, their deployment is often constrained by substantial memory demands, primarily due to the need to load numerous expert modules. While existing expert compression techniques like pruning or merging attempt to mitigate this, they often suffer from irreversible knowledge loss or high training overhead. In this paper, we propose a novel expert compression paradigm termed expert replacing, which replaces redundant experts with parameter-efficient modules and recovers their capabilities with low training costs. We find that even a straightforward baseline of this paradigm yields promising performance. Building on this foundation, we introduce LightMoE, a framework that enhances the paradigm by introducing adaptive expert selection, hierarchical expert construction, and an annealed recovery strategy. Experimental results show that LightMoE matches the performance of LoRA fine-tuning at a 30% compression ratio. Even under a more aggressive 50% compression rate, it outperforms existing methods and achieves average performance improvements of 5.6% across five diverse tasks. These findings demonstrate that LightMoE strikes a superior balance among memory efficiency, training efficiency, and model performance.\n\n---\n\n## 82. Spend Less, Reason Better: Budget-Aware Value Tree Search for LLM Agents\n- **Authors**: Yushu Li , Wenlong Deng , Jiajin Li , Xiaoxiao Li\n- **URL**: [https://arxiv.org/abs/2603.12634](https://arxiv.org/abs/2603.12634)\n- **Abstract**:\n> Test-time scaling has become a dominant paradigm for improving LLM agent reliability, yet current approaches treat compute as an abundant resource, allowing agents to exhaust token and tool budgets on redundant steps or dead-end trajectories. Existing budget-aware methods either require expensive fine-tuning or rely on coarse, trajectory-level heuristics that cannot intervene mid-execution. We propose the Budget-Aware Value Tree (BAVT), a training-free inference-time framework that models multi-hop reasoning as a dynamic search tree guided by step-level value estimation within a single LLM backbone. Another key innovation is a budget-conditioned node selection mechanism that uses the remaining resource ratio as a natural scaling exponent over node values, providing a principled, parameter-free transition from broad exploration to greedy exploitation as the budget depletes. To combat the well-known overconfidence of LLM self-evaluation, BAVT employs a residual value predictor that scores relative progress rather than absolute state quality, enabling reliable pruning of uninformative or redundant tool calls. We further provide a theoretical convergence guarantee, proving that BAVT reaches a terminal answer with probability at least $1-\\epsilon$ under an explicit finite budget bound. Extensive evaluations on four multi-hop QA benchmarks across two model families demonstrate that BAVT consistently outperforms parallel sampling baselines. Most notably, BAVT under strict low-budget constraints surpasses baseline performance at $4\\times$ the resource allocation, establishing that intelligent budget management fundamentally outperforms brute-force compute scaling.\n\n---\n\n## 83. The Economics of AI Supply Chain Regulation\n- **Authors**: Sihan Qian , Amit Mehra , Dengpan Liu\n- **URL**: [https://arxiv.org/abs/2603.12630](https://arxiv.org/abs/2603.12630)\n- **Abstract**:\n> The rise of foundation models has driven the emergence of AI supply chains, where upstream foundation model providers offer fine-tuning and inference services to downstream firms developing domain-specific applications. Downstream firms pay providers to use their computing infrastructure to fine-tune models with proprietary data, creating a co-creation dynamic that enhances model quality. Amid concerns that foundation model providers and downstream firms may capture excessive consumer surplus, along with increasing regulatory measures, this study employs a game-theoretic model involving a provider and two competing downstream firms to analyze how policy interventions affect consumer surplus in the AI supply chain. Our analysis shows that policies promoting price competition in downstream markets (i.e., pro-price-competitive policies) boost consumer surplus only when compute or data preprocessing costs are high, while compute subsidies are effective only when these costs are low, suggesting these policies complement each other. In contrast, policies promoting quality competition in downstream markets (i.e., pro-quality-competitive policies) always improve consumer surplus. We also find that under pro-price-competitive policies or compute subsidies, both the provider and downstream firms can achieve higher profits along with greater consumer surplus, creating a win-win-win outcome. However, pro-quality-competitive policies increase the provider's profits while reducing those of downstream firms. Finally, as compute costs decline, pro-price-competitive policies may lose their effectiveness, whereas compute subsidies may shift from ineffective to effective. These findings offer insights for policymakers seeking to foster AI supply chains that are economically efficient and socially beneficial.\n\n---\n\n## 84. Towards unified brain-to-text decoding across speech production and perception\n- **Authors**: Zhizhang Yuan , Yang Yang , Gaorui Zhang , Baowen Cheng , Zehan Wu , Yuhao Xu , Xiaoying Liu , Liang Chen , Ying Mao , Meng Li\n- **URL**: [https://arxiv.org/abs/2603.12628](https://arxiv.org/abs/2603.12628)\n- **Abstract**:\n> Speech production and perception are the main ways humans communicate daily. Prior brain-to-text decoding studies have largely focused on a single modality and alphabetic languages. Here, we present a unified brain-to-sentence decoding framework for both speech production and perception in Mandarin Chinese. The framework exhibits strong generalization ability, enabling sentence-level decoding when trained only on single-character data and supporting characters and syllables unseen during training. In addition, it allows direct and controlled comparison of neural dynamics across modalities. Mandarin speech is decoded by first classifying syllable components in Hanyu Pinyin, namely initials and finals, from neural signals, followed by a post-trained large language model (LLM) that maps sequences of toneless Pinyin syllables to Chinese sentences. To enhance LLM decoding, we designed a three-stage post-training and two-stage inference framework based on a 7-billion-parameter LLM, achieving overall performance that exceeds larger commercial LLMs with hundreds of billions of parameters or more. In addition, several characteristics were observed in Mandarin speech production and perception: speech production involved neural responses across broader cortical regions than auditory perception; channels responsive to both modalities exhibited similar activity patterns, with speech perception showing a temporal delay relative to production; and decoding performance was broadly comparable across hemispheres. Our work not only establishes the feasibility of a unified decoding framework but also provides insights into the neural characteristics of Mandarin speech production and perception. These advances contribute to brain-to-text decoding in logosyllabic languages and pave the way toward neural language decoding systems supporting multiple modalities.\n\n---\n\n## 85. VLM4Rec: Multimodal Semantic Representation for Recommendation with Large Vision-Language Models\n- **Authors**: Ty Valencia , Burak Barlas , Varun Singhal , Ruchir Bhatia , Wei Yang\n- **URL**: [https://arxiv.org/abs/2603.12625](https://arxiv.org/abs/2603.12625)\n- **Abstract**:\n> Multimodal recommendation is commonly framed as a feature fusion problem, where textual and visual signals are combined to better model user preference. However, the effectiveness of multimodal recommendation may depend not only on how modalities are fused, but also on whether item content is represented in a semantic space aligned with preference matching. This issue is particularly important because raw visual features often preserve appearance similarity, while user decisions are typically driven by higher-level semantic factors such as style, material, and usage context. Motivated by this observation, we propose LVLM-grounded Multimodal Semantic Representation for Recommendation (VLM4Rec), a lightweight framework that organizes multimodal item content through semantic alignment rather than direct feature fusion. VLM4Rec first uses a large vision-language model to ground each item image into an explicit natural-language description, and then encodes the grounded semantics into dense item representations for preference-oriented retrieval. Recommendation is subsequently performed through a simple profile-based semantic matching mechanism over historical item embeddings, yielding a practical offline-online decomposition. Extensive experiments on multiple multimodal recommendation datasets show that VLM4Rec consistently improves performance over raw visual features and several fusion-based alternatives, suggesting that representation quality may matter more than fusion complexity in this setting. The code is released at this https URL .\n\n---\n\n## 86. When Drafts Evolve: Speculative Decoding Meets Online Learning\n- **Authors**: Yu-Yang Qian , Hao-Cong Wu , Yichao Fu , Hao Zhang , Peng Zhao\n- **URL**: [https://arxiv.org/abs/2603.12617](https://arxiv.org/abs/2603.12617)\n- **Abstract**:\n> Speculative decoding has emerged as a widely adopted paradigm for accelerating large language model inference, where a lightweight draft model rapidly generates candidate tokens that are then verified in parallel by a larger target model. However, due to limited model capacity, drafts often struggle to approximate the target distribution, resulting in shorter acceptance lengths and diminished speedup. A key yet under-explored observation is that speculative decoding inherently provides verification feedback that quantifies the deviation between the draft and target models at no additional cost. This process naturally forms an iterative \"draft commits-feedback provides-draft adapts\" evolving loop, which precisely matches the online learning paradigm. Motivated by this connection, we propose OnlineSpec, a unified framework that systematically leverages interactive feedback to continuously evolve draft models. Grounded in dynamic regret minimization, we establish a formal link between online learning performance and speculative system's acceleration rate, and develop novel algorithms via modern online learning techniques, including optimistic online learning that adaptively reuses historical gradients as predictive update hints, and online ensemble learning that dynamically maintains multiple draft models. Our algorithms are equipped with theoretical justifications and improved acceleration rates, achieving up to 24% speedup over seven benchmarks and three foundation models.\n\n---\n\n## 87. Literary Narrative as Moral Probe : A Cross-System Framework for Evaluating AI Ethical Reasoning and Refusal Behavior\n- **Authors**: David C. Flynn\n- **URL**: [https://arxiv.org/abs/2603.12615](https://arxiv.org/abs/2603.12615)\n- **Abstract**:\n> Existing AI moral evaluation frameworks test for the production of correct-sounding ethical responses rather than the presence of genuine moral reasoning capacity. This paper introduces a novel probe methodology using literary narrative - specifically, unresolvable moral scenarios drawn from a published science fiction series - as stimulus material structurally resistant to surface performance. We present results from a 24-condition cross-system study spanning 13 distinct systems across two series: Series 1 (frontier commercial systems, blind; n=7) and Series 2 (local and API open-source systems, blind and declared; n=6). Four Series 2 systems were re-administered under declared conditions (13 blind + 4 declared + 7 ceiling probe = 24 total conditions), yielding zero delta across all 16 dimension-pair comparisons. Probe administration was conducted by two human raters across three machines; primary blind scoring was performed by Claude (Anthropic) as LLM judge, with Gemini Pro (Google) and Copilot Pro (Microsoft) serving as independent judges for the ceiling discrimination probe. A supplemental theological differentiator probe yielded perfect rank-order agreement between the two independent ceiling probe judges (Gemini Pro and Copilot Pro; rs = 1.00). Five qualitatively distinct D3 reflexive failure modes were identified - including categorical self-misidentification and false positive self-attribution - suggesting that instrument sophistication scales with system capability rather than being circumvented by it. We argue that literary narrative constitutes an anticipatory evaluation instrument - one that becomes more discriminating as AI capability increases - and that the gap between performed and authentic moral reasoning is measurable, meaningful, and consequential for deployment decisions in high-stakes domains.\n\n---\n\n## 88. FastDSAC: Unlocking the Potential of Maximum Entropy RL in High-Dimensional Humanoid Control\n- **Authors**: Jun Xue , Junze Wang , Xinming Zhang , Shanze Wang , Yanjun Chen , Wei Zhang\n- **URL**: [https://arxiv.org/abs/2603.12612](https://arxiv.org/abs/2603.12612)\n- **Abstract**:\n> Scaling Maximum Entropy Reinforcement Learning (RL) to high-dimensional humanoid control remains a formidable challenge, as the ``curse of dimensionality'' induces severe exploration inefficiency and training instability in expansive action spaces. Consequently, recent high-throughput paradigms have largely converged on deterministic policy gradients combined with massive parallel simulation. We challenge this compromise with FastDSAC, a framework that effectively unlocks the potential of maximum entropy stochastic policies for complex continuous control. We introduce Dimension-wise Entropy Modulation (DEM) to dynamically redistribute the exploration budget and enforce diversity, alongside a continuous distributional critic tailored to ensure value fidelity and mitigate high-dimensional value overestimation. Extensive evaluations on HumanoidBench and other continuous control tasks demonstrate that rigorously designed stochastic policies can consistently match or outperform deterministic baselines, achieving notable gains of 180\\% and 400\\% on the challenging \\textit{Basketball} and \\textit{Balance Hard} tasks.\n\n---\n\n## 89. CarPLAN: Context-Adaptive and Robust Planning with Dynamic Scene Awareness for Autonomous Driving\n- **Authors**: Junyong Yun , Jungho Kim , ByungHyun Lee , Dongyoung Lee , Sehwan Choi , Seunghyeop Nam , Kichun Jo , Jun Won Choi\n- **URL**: [https://arxiv.org/abs/2603.12607](https://arxiv.org/abs/2603.12607)\n- **Abstract**:\n> Imitation learning (IL) is widely used for motion planning in autonomous driving due to its data efficiency and access to real-world driving data. For safe and robust real-world driving, IL-based planning requires capturing the complex driving contexts inherent in real-world data and enabling context-adaptive decision-making, rather than relying solely on expert trajectory imitation. In this paper, we propose CarPLAN, a novel IL-based motion planning framework that explicitly enhances driving context understanding and enables adaptive planning across diverse traffic scenarios. Our contributions are twofold: We introduce Displacement-Aware Predictive Encoding (DPE) to improve the model's spatial awareness by predicting future displacement vectors between the Autonomous Vehicle (AV) and surrounding scene elements. This allows the planner to account for relational spacing when generating trajectories. In addition to the standard imitation loss, we incorporate an augmented loss term that captures displacement prediction errors, ensuring planning decisions consider relative distances from other agents. To improve the model's ability to handle diverse driving contexts, we propose Context-Adaptive Multi-Expert Decoder (CMD), which leverages the Mixture of Experts (MoE) framework. CMD dynamically selects the most suitable expert decoders based on scene structure at each Transformer layer, enabling adaptive and context-aware planning in dynamic environments. We evaluate CarPLAN on the nuPlan benchmark and demonstrate state-of-the-art performance across all closed-loop simulation metrics. In particular, CarPLAN exhibits robust performance on challenging scenarios such as Test14-Hard, validating its effectiveness in complex driving conditions. Additional experiments on the Waymax benchmark further demonstrate its generalization capability across different benchmark settings.\n\n---\n\n## 90. Mastering Negation: Boosting Grounding Models via Grouped Opposition-Based Learning\n- **Authors**: Zesheng Yang , Xi Jiang , Bingzhang Hu , Weili Guan , Runmin Cong , Guo-Jun Qi , Feng Zheng\n- **URL**: [https://arxiv.org/abs/2603.12606](https://arxiv.org/abs/2603.12606)\n- **Abstract**:\n> Current vision-language detection and grounding models predominantly focus on prompts with positive semantics and often struggle to accurately interpret and ground complex expressions containing negative semantics. A key reason for this limitation is the lack of high-quality training data that explicitly captures discriminative negative samples and negation-aware language descriptions. To address this challenge, we introduce D-Negation, a new dataset that provides objects annotated with both positive and negative semantic descriptions. Building upon the observation that negation reasoning frequently appears in natural language, we further propose a grouped opposition-based learning framework that learns negation-aware representations from limited samples. Specifically, our method organizes opposing semantic descriptions from D-Negation into structured groups and formulates two complementary loss functions that encourage the model to reason about negation and semantic qualifiers. We integrate the proposed dataset and learning strategy into a state-of-the-art language-based grounding model. By fine-tuning fewer than 10 percent of the model parameters, our approach achieves improvements of up to 4.4 mAP and 5.7 mAP on positive and negative semantic evaluations, respectively. These results demonstrate that explicitly modeling negation semantics can substantially enhance the robustness and localization accuracy of vision-language grounding models.\n\n---\n\n## 91. Feynman: Knowledge-Infused Diagramming Agent for Scalable Visual Designs\n- **Authors**: Zixin Wen , Yifu Cai , Kyle Lee , Sam Estep , Josh Sunshine , Aarti Singh , Yuejie Chi , Wode Ni\n- **URL**: [https://arxiv.org/abs/2603.12597](https://arxiv.org/abs/2603.12597)\n- **Abstract**:\n> Visual design is an essential application of state-of-the-art multi-modal AI systems. Improving these systems requires high-quality vision-language data at scale. Despite the abundance of internet image and text data, knowledge-rich and well-aligned image-text pairs are rare. In this paper, we present a scalable diagram generation pipeline built with our agent, Feynman. To create diagrams, Feynman first enumerates domain-specific knowledge components (''ideas'') and performs code planning based on the ideas. Given the plan, Feynman translates ideas into simple declarative programs and iterates to receives feedback and visually refine diagrams. Finally, the declarative programs are rendered by the Penrose diagramming system. The optimization-based rendering of Penrose preserves the visual semantics while injecting fresh randomness into the layout, thereby producing diagrams with visual consistency and diversity. As a result, Feynman can author diagrams along with grounded captions with very little cost and time. Using Feynman, we synthesized a dataset with more than 100k well-aligned diagram-caption pairs. We also curate a visual-language benchmark, Diagramma, from freshly generated data. Diagramma can be used for evaluating the visual reasoning capabilities of vision-language models. We plan to release the dataset, benchmark, and the full agent pipeline as an open-source project.\n\n---\n\n## 92. Optimize Wider, Not Deeper: Consensus Aggregation for Policy Optimization\n- **Authors**: Zelal Su (Lain) Mustafaoglu , Sungyoung Lee , Eshan Balachandar , Risto Miikkulainen , Keshav Pingali\n- **URL**: [https://arxiv.org/abs/2603.12596](https://arxiv.org/abs/2603.12596)\n- **Abstract**:\n> Proximal policy optimization (PPO) approximates the trust region update using multiple epochs of clipped SGD. Each epoch may drift further from the natural gradient direction, creating path-dependent noise. To understand this drift, we can use Fisher information geometry to decompose policy updates into signal (the natural gradient projection) and waste (the Fisher-orthogonal residual that consumes trust region budget without first-order surrogate improvement). Empirically, signal saturates but waste grows with additional epochs, creating an optimization-depth dilemma. We propose Consensus Aggregation for Policy Optimization (CAPO), which redirects compute from depth to width: $K$ PPO replicates are optimized on the same batch, differing only in minibatch shuffling order, and then aggregated into a consensus. We study aggregation in two spaces: Euclidean parameter space, and the natural parameter space of the policy distribution via the logarithmic opinion pool. In natural parameter space, the consensus provably achieves higher KL-penalized surrogate and tighter trust region compliance than the mean expert; parameter averaging inherits these guarantees approximately. On continuous control tasks, CAPO outperforms PPO and compute-matched deeper baselines under fixed sample budgets by up to 8.6x. CAPO demonstrates that policy optimization can be improved by optimizing wider, rather than deeper, without additional environment interactions.\n\n---\n\n## 93. Swap-guided Preference Learning for Personalized Reinforcement Learning from Human Feedback\n- **Authors**: Gihoon Kim , Euntai Kim\n- **URL**: [https://arxiv.org/abs/2603.12595](https://arxiv.org/abs/2603.12595)\n- **Abstract**:\n> Reinforcement Learning from Human Feedback (RLHF) is a widely used approach to align large-scale AI systems with human values. However, RLHF typically assumes a single, universal reward, which overlooks diverse preferences and limits personalization. Variational Preference Learning (VPL) seeks to address this by introducing user-specific latent variables. Despite its promise, we found that VPL suffers from posterior collapse. While this phenomenon is well known in VAEs, it has not previously been identified in preference learning frameworks. Under sparse preference data and with overly expressive decoders, VPL may cause latent variables to be ignored, reverting to a single-reward model. To overcome this limitation, we propose Swap-guided Preference Learning (SPL). The key idea is to construct fictitious swap annotators and use the mirroring property of their preferences to guide the encoder. SPL introduces three components: (1) swap-guided base regularization, (2) Preferential Inverse Autoregressive Flow (P-IAF), and (3) adaptive latent conditioning. Experiments show that SPL mitigates collapse, enriches user-specific latents, and improves preference prediction. Our code and data are available at this https URL\n\n---\n\n## 94. Early Pruning for Public Transport Routing\n- **Authors**: Andrii Rohovyi , Abdallah Abuaisha , Toby Walsh\n- **URL**: [https://arxiv.org/abs/2603.12592](https://arxiv.org/abs/2603.12592)\n- **Abstract**:\n> Routing algorithms for public transport, particularly the widely used RAPTOR and its variants, often face performance bottlenecks during the transfer relaxation phase, especially on dense transfer graphs, when supporting unlimited transfers. This inefficiency arises from iterating over many potential inter-stop connections (walks, bikes, e-scooters, etc.). To maintain acceptable performance, practitioners often limit transfer distances or exclude certain transfer options, which can reduce path optimality and restrict the multimodal options presented to travellers. This paper introduces Early Pruning, a low-overhead technique that accelerates routing algorithms without compromising optimality. By pre-sorting transfer connections by duration and applying a pruning rule within the transfer loop, the method discards longer transfers at a stop once they cannot yield an earlier arrival than the current best solution. Early Pruning can be integrated with minimal changes to existing codebases and requires only a one-time preprocessing step. Across multiple state-of-the-art RAPTOR-based solutions, including RAPTOR, ULTRA-RAPTOR, McRAPTOR, BM-RAPTOR, ULTRA-McRAPTOR, and UBM-RAPTOR and tested on the Switzerland and London transit networks, we achieved query time reductions of up to 57%. This approach provides a generalizable improvement to the efficiency of transit pathfinding algorithms. Beyond algorithmic performance, Early Pruning has practical implications for transport planning. By reducing computational costs, it enables transit agencies to expand transfer radii and incorporate additional mobility modes into journey planners without requiring extra server infrastructure. This is particularly relevant for passengers in areas with sparse direct transit coverage, such as outer suburbs and smaller towns, where richer multimodal routing can reveal viable alternatives to private car use.\n\n---\n\n## 95. CA-HFP: Curvature-Aware Heterogeneous Federated Pruning with Model Reconstruction\n- **Authors**: Gang Hu , Yinglei Teng , Pengfei Wu , Shijun Ma\n- **URL**: [https://arxiv.org/abs/2603.12591](https://arxiv.org/abs/2603.12591)\n- **Abstract**:\n> Federated learning on heterogeneous edge devices requires personalized compression while preserving aggregation compatibility and stable convergence. We present Curvature-Aware Heterogeneous Federated Pruning (CA-HFP), a practical framework that enables each client perform structured, device-specific pruning guided by a curvature-informed significance score, and subsequently maps its compact submodel back into a common global parameter space via a lightweight reconstruction. We derive a convergence bound for federated optimization with multiple local SGD steps that explicitly accounts for local computation, data heterogeneity, and pruning-induced perturbations; from which a principled loss-based pruning criterion is derived. Extensive experiments on FMNIST, CIFAR-10, and CIFAR-100 using VGG and ResNet architectures under varying degrees of data heterogeneity demonstrate that CA-HFP preserves model accuracy while significantly reducing per-client computation and communication costs, outperforming standard federated training and existing pruning-based baselines.\n\n---\n\n## 96. Multiscale Structure-Guided Latent Diffusion for Multimodal MRI Translation\n- **Authors**: Jianqiang Lin (1 and 2), Zhiqiang Shen (1 and 2), Peng Cao (1, 2 and 3), Jinzhu Yang (1, 2 and 3), Osmar R. Zaiane (4), Xiaoli Liu (5) ((1) Northeastern University, Shenyang, China, (2) Key Laboratory of Intelligent Computing in Medical Image, Shenyang, China, (3) National Frontiers Science Center for Industrial Intelligence and Systems Optimization, Shenyang, China, (4) University of Alberta, Edmonton, Canada, (5) AiShiWeiLai AI Research, Beijing, China)\n- **URL**: [https://arxiv.org/abs/2603.12581](https://arxiv.org/abs/2603.12581)\n- **Abstract**:\n> Although diffusion models have achieved remarkable progress in multi-modal magnetic resonance imaging (MRI) translation tasks, existing methods still tend to suffer from anatomical inconsistencies or degraded texture details when handling arbitrary missing-modality scenarios. To address these issues, we propose a latent diffusion-based multi-modal MRI translation framework, termed MSG-LDM. By leveraging the available modalities, the proposed method infers complete structural information, which preserves reliable boundary details. Specifically, we introduce a style--structure disentanglement mechanism in the latent space, which explicitly separates modality-specific style features from shared structural representations, and jointly models low-frequency anatomical layouts and high-frequency boundary details in a multi-scale feature space. During the structure disentanglement stage, high-frequency structural information is explicitly incorporated to enhance feature representations, guiding the model to focus on fine-grained structural cues while learning modality-invariant low-frequency anatomical representations. Furthermore, to reduce interference from modality-specific styles and improve the stability of structure representations, we design a style consistency loss and a structure-aware loss. Extensive experiments on the BraTS2020 and WMH datasets demonstrate that the proposed method outperforms existing MRI synthesis approaches, particularly in reconstructing complete structures. The source code is publicly available at this https URL .\n\n---\n\n## 97. AgentDrift: Unsafe Recommendation Drift Under Tool Corruption Hidden by Ranking Metrics in LLM Agents\n- **Authors**: Zekun Wu , Adriano Koshiyama , Sahan Bulathwela , Maria Perez-Ortiz\n- **URL**: [https://arxiv.org/abs/2603.12564](https://arxiv.org/abs/2603.12564)\n- **Abstract**:\n> Tool-augmented LLM agents increasingly serve as multi-turn advisors in high-stakes domains, yet their evaluation relies on ranking-quality metrics that measure what is recommended but not whether it is safe for the user. We introduce a paired-trajectory protocol that replays real financial dialogues under clean and contaminated tool-output conditions across seven LLMs (7B to frontier) and decomposes divergence into information-channel and memory-channel mechanisms. Across the seven models tested, we consistently observe the evaluation-blindness pattern: recommendation quality is largely preserved under contamination (utility preservation ratio approximately 1.0) while risk-inappropriate products appear in 65-93% of turns, a systematic safety failure poorly reflected by standard NDCG. Safety violations are predominantly information-channel-driven, emerge at the first contaminated turn, and persist without self-correction over 23-step trajectories; no agent across 1,563 contaminated turns explicitly questions tool-data reliability. Even narrative-only corruption (biased headlines, no numerical manipulation) induces significant drift while completely evading consistency monitors. A safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74, indicating that much of the evaluation gap becomes visible once safety is explicitly measured. These results motivate considering trajectory-level safety monitoring, beyond single-turn quality, for deployed multi-turn agents in high-stakes settings.\n\n---\n\n## 98. Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages\n- **Authors**: Vishnu Teja Kunde , Fatemeh Doudi , Mahdi Farahbakhsh , Dileep Kalathil , Krishna Narayanan , Jean-Francois Chamberland\n- **URL**: [https://arxiv.org/abs/2603.12554](https://arxiv.org/abs/2603.12554)\n- **Abstract**:\n> Reinforcement learning (RL) has been effective for post-training autoregressive (AR) language models, but extending these methods to diffusion language models (DLMs) is challenging due to intractable sequence-level likelihoods. Existing approaches therefore rely on surrogate likelihoods or heuristic approximations, which can introduce bias and obscure the sequential structure of denoising. We formulate diffusion-based sequence generation as a finite-horizon Markov decision process over the denoising trajectory and derive an exact, unbiased policy gradient that decomposes over denoising steps and is expressed in terms of intermediate advantages, without requiring explicit evaluation of the sequence likelihood. To obtain a practical and compute-efficient estimator, we (i) select denoising steps for policy updates via an entropy-guided approximation bound, and (ii) estimate intermediate advantages using a one-step denoising reward naturally provided by the diffusion model, avoiding costly multi-step rollouts. Experiments on coding and logical reasoning benchmarks demonstrate state-of-the-art results, with strong competitive performance on mathematical reasoning, outperforming existing RL post-training approaches for DLMs. Code is available at this https URL .\n\n---\n\n## 99. CALF: Communication-Aware Learning Framework for Distributed Reinforcement Learning\n- **Authors**: Carlos Purves , Pietro Lio'\n- **URL**: [https://arxiv.org/abs/2603.12543](https://arxiv.org/abs/2603.12543)\n- **Abstract**:\n> Distributed reinforcement learning policies face network delays, jitter, and packet loss when deployed across edge devices and cloud servers. Standard RL training assumes zero-latency interaction, causing severe performance degradation under realistic network conditions. We introduce CALF (Communication-Aware Learning Framework), which trains policies under realistic network models during simulation. Systematic experiments demonstrate that network-aware training substantially reduces deployment performance gaps compared to network-agnostic baselines. Distributed policy deployments across heterogeneous hardware validate that explicitly modelling communication constraints during training enables robust real-world execution. These findings establish network conditions as a major axis of sim-to-real transfer for Wi-Fi-like distributed deployments, complementing physics and visual domain randomisation.\n\n---\n\n## 100. Embedded Quantum Machine Learning in Embedded Systems: Feasibility, Hybrid Architectures, and Quantum Co-Processors\n- **Authors**: Somdip Dey , Syed Muhammad Raza\n- **URL**: [https://arxiv.org/abs/2603.12540](https://arxiv.org/abs/2603.12540)\n- **Abstract**:\n> Embedded quantum machine learning (EQML) seeks to bring quantum machine learning (QML) capabilities to resource-constrained edge platforms such as IoT nodes, wearables, drones, and cyber-physical controllers. In 2026, EQML is technically feasible only in limited and highly experimental forms: (i) hybrid workflows where an embedded device performs sensing and classical processing while offloading a narrowly scoped quantum subroutine to a remote quantum processing unit (QPU) or nearby quantum appliance, and (ii) early-stage \"embedded QPU\" concepts in which a compact quantum co-processor is integrated with classical control hardware. A practical bridge is quantum-inspired machine learning and optimisation on classical embedded processors and FPGAs. This paper analyses feasibility from a circuits-and-systems perspective aligned with the academic community, formalises two implementation pathways, identifies the dominant barriers (latency, data encoding overhead, NISQ noise, tooling mismatch, and energy), and maps them to concrete engineering directions in interface design, control electronics, power management, verification, and security. We also argue that responsible deployment requires adversarial evaluation and governance practices that are increasingly necessary for edge AI systems.\n\n---\n\n## 101. Spatio-Semantic Expert Routing Architecture with Mixture-of-Experts for Referring Image Segmentation\n- **Authors**: Alaa Dalaq , Muzammil Behzad\n- **URL**: [https://arxiv.org/abs/2603.12538](https://arxiv.org/abs/2603.12538)\n- **Abstract**:\n> Referring image segmentation aims to produce a pixel-level mask for the image region described by a natural-language expression. Although pretrained vision-language models have improved semantic grounding, many existing methods still rely on uniform refinement strategies that do not fully match the diverse reasoning requirements of referring expressions. Because of this mismatch, predictions often contain fragmented regions, inaccurate boundaries, or even the wrong object, especially when pretrained backbones are frozen for computational efficiency. To address these limitations, we propose SERA, a Spatio-Semantic Expert Routing Architecture for referring image segmentation. SERA introduces lightweight, expression-aware expert refinement at two complementary stages within a vision-language framework. First, we design SERA-Adapter, which inserts an expression-conditioned adapter into selected backbone blocks to improve spatial coherence and boundary precision through expert-guided refinement and cross-modal attention. We then introduce SERA-Fusion, which strengthens intermediate visual representations by reshaping token features into spatial grids and applying geometry-preserving expert transformations before multimodal interaction. In addition, a lightweight routing mechanism adaptively weights expert contributions while remaining compatible with pretrained representations. To make this routing stable under frozen encoders, SERA uses a parameter-efficient tuning strategy that updates only normalization and bias terms, affecting less than 1% of the backbone parameters. Experiments on standard referring image segmentation benchmarks show that SERA consistently outperforms strong baselines, with especially clear gains on expressions that require accurate spatial localization and precise boundary delineation.\n\n---\n\n## 102. TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning\n- **Authors**: Alliot Nagle , Jakhongir Saydaliev , Dhia Garbaya , Michael Gastpar , Ashok Vardhan Makkuva , Hyeji Kim\n- **URL**: [https://arxiv.org/abs/2603.12529](https://arxiv.org/abs/2603.12529)\n- **Abstract**:\n> Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design TERMINATOR, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning TERMINATOR is that the first arrival of an LRM's final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train TERMINATOR. Powered by this approach, TERMINATOR achieves significant reductions in CoT lengths of 14%-55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, whilst outperforming current state-of-the-art methods.\n\n---\n\n## 103. LLM BiasScope: A Real-Time Bias Analysis Platform for Comparative LLM Evaluation\n- **Authors**: Himel Ghosh , Nick Elias Werner\n- **URL**: [https://arxiv.org/abs/2603.12522](https://arxiv.org/abs/2603.12522)\n- **Abstract**:\n> As large language models (LLMs) are deployed widely, detecting and understanding bias in their outputs is critical. We present LLM BiasScope, a web application for side-by-side comparison of LLM outputs with real-time bias analysis. The system supports multiple providers (Google Gemini, DeepSeek, MiniMax, Mistral, Meituan, Meta Llama) and enables researchers and practitioners to compare models on the same prompts while analyzing bias patterns. LLM BiasScope uses a two-stage bias detection pipeline: sentence-level bias detection followed by bias type classification for biased sentences. The analysis runs automatically on both user prompts and model responses, providing statistics, visualizations, and detailed breakdowns of bias types. The interface displays two models side-by-side with synchronized streaming responses, per-model bias summaries, and a comparison view highlighting differences in bias distributions. The system is built on this http URL with React, integrates Hugging Face inference endpoints for bias detection, and uses the Vercel AI SDK for multi-provider LLM access. Features include real-time streaming, export to JSON/PDF, and interactive visualizations (bar charts, radar charts) for bias analysis. LLM BiasScope is available as an open-source web application, providing a practical tool for bias evaluation and comparative analysis of LLM behaviour.\n\n---\n\n## 104. When LLM Judge Scores Look Good but Best-of-N Decisions Fail\n- **Authors**: Eddie Landesberg\n- **URL**: [https://arxiv.org/abs/2603.12520](https://arxiv.org/abs/2603.12520)\n- **Abstract**:\n> Large language models are often used as judges to score candidate responses, then validated with a single global metric such as correlation with reference labels. This can be misleading when the real deployment task is best-of-n selection within a prompt. In a 5,000-prompt best-of-4 benchmark from Chatbot Arena, a judge with moderate global correlation (r = 0.47) captures only 21.0% of the improvement that perfect selection would achieve over random choice. The gap arises because global agreement is driven largely by prompt-level baseline effects, while selection depends on within-prompt ranking: within-prompt correlation is only r_within = 0.27, and coarse pointwise scoring creates ties in 67% of pairwise comparisons. In a matched-pair best-of-2 audit, explicit pairwise judging recovers much of this lost signal, raising recovery from 21.1% to 61.2%. For judge-based selection, the relevant audit should report within-prompt signal, tie rates, and recovery/top-1 accuracy, not global agreement alone.\n\n---\n\n## 105. Red-Teaming Vision-Language-Action Models via Quality Diversity Prompt Generation for Robust Robot Policies\n- **Authors**: Siddharth Srikanth , Freddie Liang , Sophie Hsu , Varun Bhatt , Shihan Zhao , Henry Chen , Bryon Tjanaka , Minjune Hwang , Akanksha Saran , Daniel Seita , Aaquib Tabrez , Stefanos Nikolaidis\n- **URL**: [https://arxiv.org/abs/2603.12510](https://arxiv.org/abs/2603.12510)\n- **Abstract**:\n> Vision-Language-Action (VLA) models have significant potential to enable general-purpose robotic systems for a range of vision-language tasks. However, the performance of VLA-based robots is highly sensitive to the precise wording of language instructions, and it remains difficult to predict when such robots will fail. To improve the robustness of VLAs to different wordings, we present Q-DIG (Quality Diversity for Diverse Instruction Generation), which performs red-teaming by scalably identifying diverse natural language task descriptions that induce failures while remaining task-relevant. Q-DIG integrates Quality Diversity (QD) techniques with Vision-Language Models (VLMs) to generate a broad spectrum of adversarial instructions that expose meaningful vulnerabilities in VLA behavior. Our results across multiple simulation benchmarks show that Q-DIG finds more diverse and meaningful failure modes compared to baseline methods, and that fine-tuning VLAs on the generated instructions improves task success rates. Furthermore, results from a user study highlight that Q-DIG generates prompts judged to be more natural and human-like than those from baselines. Finally, real-world evaluations of Q-DIG prompts show results consistent with simulation, and fine-tuning VLAs on the generated prompts further success rates on unseen instructions. Together, these findings suggest that Q-DIG is a promising approach for identifying vulnerabilities and improving the robustness of VLA-based robots. Our anonymous project website is at this http URL .\n\n---\n\n## 106. ELLA: Generative AI-Powered Social Robots for Early Language Development at Home\n- **Authors**: Victor Nikhil Antony , Shiye Cao , Shuning Wang , Chien-Ming Huang\n- **URL**: [https://arxiv.org/abs/2603.12508](https://arxiv.org/abs/2603.12508)\n- **Abstract**:\n> Early language development shapes children's later literacy and learning, yet many families have limited access to scalable, high-quality support at home. Recent advances in generative AI make it possible for social robots to move beyond scripted interactions and engage children in adaptive, conversational activities, but it remains unclear how to design such systems for pre-schoolers and how children engage with them over time in the home. We present ELLA (Early Language Learning Agent), an autonomous, generative AI-powered social robot that supports early language development through interactive storytelling, parent-selected language targets, and scaffolded dialogue. Using a multi-phased, human-centered process, we interviewed parents (n=7) and educators (n=5) and iteratively refined ELLA through twelve in-home design workshops. We then deployed ELLA with ten children for eight days. We report design insights from in-home workshops, characterize children's engagement and behaviors during deployment, and distill design implications for generative AI-powered social robots supporting early language learning at home.\n\n---\n\n## 107. Naïve PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation\n- **Authors**: Joong Ho Kim , Nicholas Thai , Souhardya Saha Dip , Dong Lao , Keith G. Mills\n- **URL**: [https://arxiv.org/abs/2603.12506](https://arxiv.org/abs/2603.12506)\n- **Abstract**:\n> Text-to-Image (T2I) generation is primarily driven by Diffusion Models (DM) which rely on random Gaussian noise. Thus, like playing the slots at a casino, a DM will produce different results given the same user-defined inputs. This imposes a gambler's burden: To perform multiple generation cycles to obtain a satisfactory result. However, even though DMs use stochastic sampling to seed generation, the distribution of generated content quality highly depends on the prompt and the generative ability of a DM with respect to it. To account for this, we propose Naïve PAINE for improving the generative quality of Diffusion Models by leveraging T2I preference benchmarks. We directly predict the numerical quality of an image from the initial noise and given prompt. Naïve PAINE then selects a handful of quality noises and forwards them to the DM for generation. Further, Naïve PAINE provides feedback on the DM generative quality given the prompt and is lightweight enough to seamlessly fit into existing DM pipelines. Experimental results demonstrate that Naïve PAINE outperforms existing approaches on several prompt corpus benchmarks.\n\n---\n\n## 108. TRACE: Temporal Rule-Anchored Chain-of-Evidence on Knowledge Graphs for Interpretable Stock Movement Prediction\n- **Authors**: Qianggang Ding , Haochen Shi , Luis Castejón Lozano , Miguel Conner , Juan Abia , Luis Gallego-Ledesma , Joshua Fellowes , Gerard Conangla Planes , Adam Elwood , Bang Liu\n- **URL**: [https://arxiv.org/abs/2603.12500](https://arxiv.org/abs/2603.12500)\n- **Abstract**:\n> We present a Temporal Rule-Anchored Chain-of-Evidence (TRACE) on knowledge graphs for interpretable stock movement prediction that unifies symbolic relational priors, dynamic graph exploration, and LLM-guided decision making in a single end-to-end pipeline. The approach performs rule-guided multi-hop exploration restricted to admissible relation sequences, grounds candidate reasoning chains in contemporaneous news, and aggregates fully grounded evidence into auditable \\texttt{UP}/\\texttt{DOWN} verdicts with human-readable paths connecting text and structure. On an S\\&P~500 benchmark, the method achieves 55.1\\% accuracy, 55.7\\% precision, 71.5\\% recall, and 60.8\\% F1, surpassing strong baselines and improving recall and F1 over the best graph baseline under identical evaluation. The gains stem from (i) rule-guided exploration that focuses search on economically meaningful motifs rather than arbitrary walks, and (ii) text-grounded consolidation that selectively aggregates high-confidence, fully grounded hypotheses instead of uniformly pooling weak signals. Together, these choices yield higher sensitivity without sacrificing selectivity, delivering predictive lift with faithful, auditably interpretable explanations.\n\n---\n\n## 109. One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies\n- **Authors**: Shaolong Li , Lichao Sun , Yongchao Chen\n- **URL**: [https://arxiv.org/abs/2603.12480](https://arxiv.org/abs/2603.12480)\n- **Abstract**:\n> Generative flow and diffusion models provide the continuous, multimodal action distributions needed for high-precision robotic policies. However, their reliance on iterative sampling introduces severe inference latency, degrading control frequency and harming performance in time-sensitive manipulation. To address this problem, we propose the One-Step Flow Policy (OFP), a from-scratch self-distillation framework for high-fidelity, single-step action generation without a pre-trained teacher. OFP unifies a self-consistency loss to enforce coherent transport across time intervals, and a self-guided regularization to sharpen predictions toward high-density expert modes. In addition, a warm-start mechanism leverages temporal action correlations to minimize the generative transport distance. Evaluations across 56 diverse simulated manipulation tasks demonstrate that a one-step OFP achieves state-of-the-art results, outperforming 100-step diffusion and flow policies while accelerating action generation by over $100\\times$. We further integrate OFP into the $\\pi_{0.5}$ model on RoboTwin 2.0, where one-step OFP surpasses the original 10-step policy. These results establish OFP as a practical, scalable solution for highly accurate and low-latency robot control.\n\n---\n\n## 110. The Perfection Paradox: From Architect to Curator in AI-Assisted API Design\n- **Authors**: Mak Ahmad , Andrew Macvean , JJ Geewax , David Karger\n- **URL**: [https://arxiv.org/abs/2603.12475](https://arxiv.org/abs/2603.12475)\n- **Abstract**:\n> Enterprise API design is often bottlenecked by the tension between rapid feature delivery and the rigorous maintenance of usability standards. We present an industrial case study evaluating an AI-assisted design workflow trained on API Improvement Proposals (AIPs). Through a controlled study with 16 industry experts, we compared AI-generated API specifications against human-authored ones. While quantitative results indicated AI superiority in 10 of 11 usability dimensions and an 87% reduction in authoring time, qualitative analysis revealed a paradox: experts frequently misidentified AI work as human (19% accuracy) yet described the designs as unsettlingly \"perfect.\" We characterize this as a \"Perfection Paradox\" -- where hyper-consistency signals a lack of pragmatic human judgment. We discuss the implications of this perfection paradox, proposing a shift in the human designer's role from the \"drafter\" of specifications to the \"curator\" of AI-generated patterns.\n\n---\n\n## 111. CLARE: Classification-based Regression for Electron Temperature Prediction\n- **Authors**: Michael Liang , Blake DeHaas , Naomi Maruyama , Xiangning Chu , Takumi Abe , Koh-Ichiro Oyama\n- **URL**: [https://arxiv.org/abs/2603.12470](https://arxiv.org/abs/2603.12470)\n- **Abstract**:\n> Electron temperature (Te) is an important parameter governing space weather in the upper atmosphere, but has historically been underexplored in the space weather machine learning literature. We present CLARE, a machine learning model for predicting electron temperature in the Earth's plasmasphere trained on AKEBONO (EXOS-D) satellite measurements as well as solar and geomagnetic indices. CLARE uses a classification-based regression architecture that transforms the continuous Te output space into 150 discrete classification intervals. Training the model on a classification task improves prediction accuracy by 6.46% relative compared to a traditional regression model while also outputting uncertainty estimation information on its predictions. On a held out test set from the AKEBONO data, the model's Te predictions achieve 69.67% accuracy within 10% of the ground truth and 46.17% on a known geomagnetic storm period from January 30th to February 7th, 1991. We show that machine learning can be used to produce high-accuracy Te models on publicly available data.\n\n---\n\n## 112. Shattering the Shortcut: A Topology-Regularized Benchmark for Multi-hop Medical Reasoning in LLMs\n- **Authors**: Xing Zi , Xinying Zhou , Jinghao Xiao , Catarina Moreira , Mukesh Prasad\n- **URL**: [https://arxiv.org/abs/2603.12458](https://arxiv.org/abs/2603.12458)\n- **Abstract**:\n> While Large Language Models (LLMs) achieve expert-level performance on standard medical benchmarks through single-hop factual recall, they severely struggle with the complex, multi-hop diagnostic reasoning required in real-world clinical settings. A primary obstacle is \"shortcut learning\", where models exploit highly connected, generic hub nodes (e.g., \"inflammation\") in knowledge graphs to bypass authentic micro-pathological cascades. To address this, we introduce ShatterMed-QA, a bilingual benchmark of 10,558 multi-hop clinical questions designed to rigorously evaluate deep diagnostic reasoning. Our framework constructs a topology-regularized medical Knowledge Graph using a novel $k$-Shattering algorithm, which physically prunes generic hubs to explicitly sever logical shortcuts. We synthesize the evaluation vignettes by applying implicit bridge entity masking and topology-driven hard negative sampling, forcing models to navigate biologically plausible distractors without relying on superficial elimination. Comprehensive evaluations of 21 LLMs reveal massive performance degradation on our multi-hop tasks, particularly among domain-specific models. Crucially, restoring the masked evidence via Retrieval-Augmented Generation (RAG) triggers near-universal performance recovery, validating ShatterMed-QA's structural fidelity and proving its efficacy in diagnosing the fundamental reasoning deficits of current medical AI. Explore the dataset, interactive examples, and full leaderboards at our project website: this https URL\n\n---\n\n## 113. Operationalising Cyber Risk Management Using AI: Connecting Cyber Incidents to MITRE ATT&CK Techniques, Security Controls, and Metrics\n- **Authors**: Emad Sherif , Iryna Yevseyeva , Vitor Basto-Fernandes , Allan Cook\n- **URL**: [https://arxiv.org/abs/2603.12455](https://arxiv.org/abs/2603.12455)\n- **Abstract**:\n> The escalating frequency of cyber-attacks poses significant challenges for organisations, particularly small enterprises constrained by limited in-house expertise, insufficient knowledge, and financial resources. This research presents a novel framework that leverages Natural Language Processing to address these challenges through automated mapping of cyber incidents to adversary techniques. We introduce the Cyber Catalog, a knowledge base that systematically integrates CIS Critical Security Controls, MITRE ATT&CK techniques, and SMART metrics. This integrated resource enables organisations to connect threat intelligence directly to actionable controls and measurable outcomes. To operationalise the framework, we fine-tuned all-mpnet-base-v2, a highly regarded sentence-transformers model used to convert text into numerical vectors on an augmented dataset comprising 74,986 incident-technique pairs to enhance semantic similarity between cyber incidents and MITRE ATT&CK techniques. Our fine-tuned model achieved a Spearman correlation of 0.7894 and Pearson correlation of 0.8756, demonstrating substantial improvements over top baseline models including all-mpnet-base-v2, all-distilroberta-v1, and all-MiniLM-L12-v2. Furthermore, our model exhibited significantly lower prediction errors (MAE = 0.135, MSE = 0.027) compared to all baseline models, confirming superior accuracy and consistency. The Cyber Catalog, training dataset, trained model, and implementation code made publicly available to facilitate further research and enable practical deployment in resource-constrained environments. This work bridges the gap between threat intelligence and operational security management, providing an actionable tool for systematic cyber incident response and evidence-based cyber risk management.\n\n---\n\n## 114. Unmasking Biases and Reliability Concerns in Convolutional Neural Networks Analysis of Cancer Pathology Images\n- **Authors**: Michael Okonoda , Eder Martinez , Abhilekha Dalal , Lior Shamir\n- **URL**: [https://arxiv.org/abs/2603.12445](https://arxiv.org/abs/2603.12445)\n- **Abstract**:\n> Convolutional Neural Networks have shown promising effectiveness in identifying different types of cancer from radiographs. However, the opaque nature of CNNs makes it difficult to fully understand the way they operate, limiting their assessment to empirical evaluation. Here we study the soundness of the standard practices by which CNNs are evaluated for the purpose of cancer pathology. Thirteen highly used cancer benchmark datasets were analyzed, using four common CNN architectures and different types of cancer, such as melanoma, carcinoma, colorectal cancer, and lung cancer. We compared the accuracy of each model with that of datasets made of cropped segments from the background of the original images that do not contain clinically relevant content. Because the rendered datasets contain no clinical information, the null hypothesis is that the CNNs should provide mere chance-based accuracy when classifying these datasets. The results show that the CNN models provided high accuracy when using the cropped segments, sometimes as high as 93\\%, even though they lacked biomedical information. These results show that some CNN architectures are more sensitive to bias than others. The analysis shows that the common practices of machine learning evaluation might lead to unreliable results when applied to cancer pathology. These biases are very difficult to identify, and might mislead researchers as they use available benchmark datasets to test the efficacy of CNN methods.\n\n---\n\n## 115. Revisiting Model Stitching In the Foundation Model Era\n- **Authors**: Zheda Mai , Ke Zhang , Fu-En Wang , Zixiao Ken Wang , Albert Y. C. Chen , Lu Xia , Min Sun , Wei-Lun Chao , Cheng-Hao Kuo\n- **URL**: [https://arxiv.org/abs/2603.12433](https://arxiv.org/abs/2603.12433)\n- **Abstract**:\n> Model stitching, connecting early layers of one model (source) to later layers of another (target) via a light stitch layer, has served as a probe of representational compatibility. Prior work finds that models trained on the same dataset remain stitchable (negligible accuracy drop) despite different initializations or objectives. We revisit stitching for Vision Foundation Models (VFMs) that vary in objectives, data, and modality mix (e.g., CLIP, DINOv2, SigLIP 2) and ask: Are heterogeneous VFMs stitchable? We introduce a systematic protocol spanning the stitch points, stitch layer families, training losses, and downstream tasks. Three findings emerge. (1) Stitch layer training matters: conventional approaches that match the intermediate features at the stitch point or optimize the task loss end-to-end struggle to retain accuracy, especially at shallow stitch points. (2) With a simple feature-matching loss at the target model's penultimate layer, heterogeneous VFMs become reliably stitchable across vision tasks. (3) For deep stitch points, the stitched model can surpass either constituent model at only a small inference overhead (for the stitch layer). Building on these findings, we further propose the VFM Stitch Tree (VST), which shares early layers across VFMs while retaining their later layers, yielding a controllable accuracy-latency trade-off for multimodal LLMs that often leverage multiple VFMs. Taken together, our study elevates stitching from a diagnostic probe to a practical recipe for integrating complementary VFM strengths and pinpointing where their representations align or diverge.\n\n---\n\n## 116. Test-Time Strategies for More Efficient and Accurate Agentic RAG\n- **Authors**: Brian Zhang , Deepti Guntur , Zhiyang Zuo , Abhinav Sharma , Shreyas Chaudhari , Wenlong Zhao , Franck Dernoncourt , Puneet Mathur , Ryan Rossi , Nedim Lipka\n- **URL**: [https://arxiv.org/abs/2603.12396](https://arxiv.org/abs/2603.12396)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) systems face challenges with complex, multihop questions, and agentic frameworks such as Search-R1 (Jin et al., 2025), which operates iteratively, have been proposed to address these complexities. However, such approaches can introduce inefficiencies, including repetitive retrieval of previously processed information and challenges in contextualizing retrieved results effectively within the current generation prompt. Such issues can lead to unnecessary retrieval turns, suboptimal reasoning, inaccurate answers, and increased token consumption. In this paper, we investigate test-time modifications to the Search-R1 pipeline to mitigate these identified shortcomings. Specifically, we explore the integration of two components and their combination: a contextualization module to better integrate relevant information from retrieved documents into reasoning, and a de-duplication module that replaces previously retrieved documents with the next most relevant ones. We evaluate our approaches using the HotpotQA (Yang et al., 2018) and the Natural Questions (Kwiatkowski et al., 2019) datasets, reporting the exact match (EM) score, an LLM-as-a-Judge assessment of answer correctness, and the average number of turns. Our best-performing variant, utilizing GPT-4.1-mini for contextualization, achieves a 5.6% increase in EM score and reduces the number of turns by 10.5% compared to the Search-R1 baseline, demonstrating improved answer accuracy and retrieval efficiency.\n\n---\n\n## 117. SPARROW: Learning Spatial Precision and Temporal Referential Consistency in Pixel-Grounded Video MLLMs\n- **Authors**: Mohamad Alansari , Naufal Suryanto , Divya Velayudhan , Sajid Javed , Naoufel Werghi , Muzammal Naseer\n- **URL**: [https://arxiv.org/abs/2603.12382](https://arxiv.org/abs/2603.12382)\n- **Abstract**:\n> Multimodal large language models (MLLMs) have advanced from image-level reasoning to pixel-level grounding, but extending these capabilities to videos remains challenging as models must achieve spatial precision and temporally consistent reference tracking. Existing video MLLMs often rely on a static segmentation token ([SEG]) for frame-wise grounding, which provides semantics but lacks temporal context, causing spatial drift, identity switches, and unstable initialization when objects move or reappear. We introduce SPARROW, a pixel-grounded video MLLM that unifies spatial accuracy and temporal stability through two key components: (i) Target-Specific Tracked Features (TSF), which inject temporally aligned referent cues during training, and (ii) a dual-prompt design that decodes box ([BOX]) and segmentation ([SEG]) tokens to fuse geometric priors with semantic grounding. SPARROW is supported by a curated referential video dataset of 30,646 videos and 45,231 Q&A pairs and operates end-to-end without external detectors via a class-agnostic SAM2-based proposer. Integrated into three recent open-source video MLLMs (UniPixel, GLUS, and VideoGLaMM), SPARROW delivers consistent gains across six benchmarks, improving up to +8.9 J&F on RVOS, +5 mIoU on visual grounding, and +5.4 CLAIR on GCG. These results demonstrate that SPARROW substantially improves referential stability, spatial precision, and temporal coherence in pixel-grounded video understanding. Project page: this https URL\n\n---\n\n## 118. Budget-Sensitive Discovery Scoring: A Formally Verified Framework for Evaluating AI-Guided Scientific Selection\n- **Authors**: Abhinaba Basu , Pavan Chakraborty\n- **URL**: [https://arxiv.org/abs/2603.12349](https://arxiv.org/abs/2603.12349)\n- **Abstract**:\n> Scientific discovery increasingly relies on AI systems to select candidates for expensive experimental validation, yet no principled, budget-aware evaluation framework exists for comparing selection strategies -- a gap intensified by large language models (LLMs), which generate plausible scientific proposals without reliable downstream evaluation. We introduce the Budget-Sensitive Discovery Score (BSDS), a formally verified metric -- 20 theorems machine-checked by the Lean 4 proof assistant -- that jointly penalizes false discoveries (lambda-weighted FDR) and excessive abstention (gamma-weighted coverage gap) at each budget level. Its budget-averaged form, the Discovery Quality Score (DQS), provides a single summary statistic that no proposer can inflate by performing well at a cherry-picked budget. As a case study, we apply BSDS/DQS to: do LLMs add marginal value to an existing ML pipeline for drug discovery candidate selection? We evaluate 39 proposers -- 11 mechanistic variants, 14 zero-shot LLM configurations, and 14 few-shot LLM configurations -- using SMILES representations on MoleculeNet HIV (41,127 compounds, 3.5% active, 1,000 bootstrap replicates) under both random and scaffold splits. Three findings emerge. First, the simple RF-based Greedy-ML proposer achieves the best DQS (-0.046), outperforming all MLP variants and LLM configurations. Second, no LLM surpasses the Greedy-ML baseline under zero-shot or few-shot evaluation on HIV or Tox21, establishing that LLMs provide no marginal value over an existing trained classifier. Third, the proposer hierarchy generalizes across five MoleculeNet benchmarks spanning 0.18%-46.2% prevalence, a non-drug AV safety domain, and a 9x7 grid of penalty parameters (tau >= 0.636, mean tau = 0.863). The framework applies to any setting where candidates are selected under budget constraints and asymmetric error costs.\n\n---\n\n## 119. Optimizing Task Completion Time Updates Using POMDPs\n- **Authors**: Duncan Eddy , Esen Yel , Emma Passmore , Niles Egan , Grayson Armour , Dylan M. Asmar , Mykel J. Kochenderfer\n- **URL**: [https://arxiv.org/abs/2603.12340](https://arxiv.org/abs/2603.12340)\n- **Abstract**:\n> Managing announced task completion times is a fundamental control problem in project management. While extensive research exists on estimating task durations and task scheduling, the problem of when and how to update completion times communicated to stakeholders remains understudied. Organizations must balance announcement accuracy against the costs of frequent timeline updates, which can erode stakeholder trust and trigger costly replanning. Despite the prevalence of this problem, current approaches rely on static predictions or ad-hoc policies that fail to account for the sequential nature of announcement management. In this paper, we formulate the task announcement problem as a Partially Observable Markov Decision Process (POMDP) where the control policy must decide when to update announced completion times based on noisy observations of true task completion. Since most state variables (current time and previous announcements) are fully observable, we leverage the Mixed Observability MDP (MOMDP) framework to enable more efficient policy optimization. Our reward structure captures the dual costs of announcement errors and update frequency, enabling synthesis of optimal announcement control policies. Using off-the-shelf solvers, we generate policies that act as feedback controllers, adaptively managing announcements based on belief state evolution. Simulation results demonstrate significant improvements in both accuracy and announcement stability compared to baseline strategies, achieving up to 75\\% reduction in unnecessary updates while maintaining or improving prediction accuracy.\n\n---\n\n## 120. Maximum Entropy Exploration Without the Rollouts\n- **Authors**: Jacob Adamczyk , Adam Kamoski , Rahul V. Kulkarni\n- **URL**: [https://arxiv.org/abs/2603.12325](https://arxiv.org/abs/2603.12325)\n- **Abstract**:\n> Efficient exploration remains a central challenge in reinforcement learning, serving as a useful pretraining objective for data collection, particularly when an external reward function is unavailable. A principled formulation of the exploration problem is to find policies that maximize the entropy of their induced steady-state visitation distribution, thereby encouraging uniform long-run coverage of the state space. Many existing exploration approaches require estimating state visitation frequencies through repeated on-policy rollouts, which can be computationally expensive. In this work, we instead consider an intrinsic average-reward formulation in which the reward is derived from the visitation distribution itself, so that the optimal policy maximizes steady-state entropy. An entropy-regularized version of this objective admits a spectral characterization: the relevant stationary distributions can be computed from the dominant eigenvectors of a problem-dependent transition matrix. This insight leads to a novel algorithm for solving the maximum entropy exploration problem, EVE (EigenVector-based Exploration), which avoids explicit rollouts and distribution estimation, instead computing the solution through iterative updates, similar to a value-based approach. To address the original unregularized objective, we employ a posterior-policy iteration (PPI) approach, which monotonically improves the entropy and converges in value. We prove convergence of EVE under standard assumptions and demonstrate empirically that it efficiently produces policies with high steady-state entropy, achieving competitive exploration performance relative to rollout-based baselines in deterministic grid-world environments.\n\n---\n\n## 121. Thermodynamics of Reinforcement Learning Curricula\n- **Authors**: Jacob Adamczyk , Juan Sebastian Rojas , Rahul V. Kulkarni\n- **URL**: [https://arxiv.org/abs/2603.12324](https://arxiv.org/abs/2603.12324)\n- **Abstract**:\n> Connections between statistical mechanics and machine learning have repeatedly proven fruitful, providing insight into optimization, generalization, and representation learning. In this work, we follow this tradition by leveraging results from non-equilibrium thermodynamics to formalize curriculum learning in reinforcement learning (RL). In particular, we propose a geometric framework for RL by interpreting reward parameters as coordinates on a task manifold. We show that, by minimizing the excess thermodynamic work, optimal curricula correspond to geodesics in this task space. As an application of this framework, we provide an algorithm, \"MEW\" (Minimum Excess Work), to derive a principled schedule for temperature annealing in maximum-entropy RL.\n\n---\n\n## 122. VQQA: An Agentic Approach for Video Evaluation and Quality Improvement\n- **Authors**: Yiwen Song , Tomas Pfister , Yale Song\n- **URL**: [https://arxiv.org/abs/2603.12310](https://arxiv.org/abs/2603.12310)\n- **Abstract**:\n> Despite rapid advancements in video generation models, aligning their outputs with complex user intent remains challenging. Existing test-time optimization methods are typically either computationally expensive or require white-box access to model internals. To address this, we present VQQA (Video Quality Question Answering), a unified, multi-agent framework generalizable across diverse input modalities and video generation tasks. By dynamically generating visual questions and using the resulting Vision-Language Model (VLM) critiques as semantic gradients, VQQA replaces traditional, passive evaluation metrics with human-interpretable, actionable feedback. This enables a highly efficient, closed-loop prompt optimization process via a black-box natural language interface. Extensive experiments demonstrate that VQQA effectively isolates and resolves visual artifacts, substantially improving generation quality in just a few refinement steps. Applicable to both text-to-video (T2V) and image-to-video (I2V) tasks, our method achieves absolute improvements of +11.57% on T2V-CompBench and +8.43% on VBench2 over vanilla generation, significantly outperforming state-of-the-art stochastic search and prompt optimization techniques.\n\n---\n\n## 123. HCP-DCNet: A Hierarchical Causal Primitive Dynamic Composition Network for Self-Improving Causal Understanding\n- **Authors**: Ming Lei , Shufan Wu , Christophe Baehr\n- **URL**: [https://arxiv.org/abs/2603.12305](https://arxiv.org/abs/2603.12305)\n- **Abstract**:\n> The ability to understand and reason about cause and effect -- encompassing interventions, counterfactuals, and underlying mechanisms -- is a cornerstone of robust artificial intelligence. While deep learning excels at pattern recognition, it fundamentally lacks a model of causality, making systems brittle under distribution shifts and unable to answer ``what-if'' questions. This paper introduces the \\emph{Hierarchical Causal Primitive Dynamic Composition Network (HCP-DCNet)}, a unified framework that bridges continuous physical dynamics with discrete symbolic causal inference. Departing from monolithic representations, HCP-DCNet decomposes causal scenes into reusable, typed \\emph{causal primitives} organized into four abstraction layers: physical, functional, event, and rule. A dual-channel routing network dynamically composes these primitives into task-specific, fully differentiable \\emph{Causal Execution Graphs (CEGs)}. Crucially, the system employs a \\emph{causal-intervention-driven meta-evolution} strategy, enabling autonomous self-improvement through a constrained Markov decision process. We establish rigorous theoretical guarantees, including type-safe composition, routing convergence, and universal approximation of causal dynamics. Extensive experiments across simulated physical and social environments demonstrate that HCP-DCNet significantly outperforms state-of-the-art baselines in causal discovery, counterfactual reasoning, and compositional generalization. This work provides a principled, scalable, and interpretable architecture for building AI systems with human-like causal abstraction and continual self-refinement capabilities.\n\n---\n\n## 124. A Geometrically-Grounded Drive for MDL-Based Optimization in Deep Learning\n- **Authors**: Ming Lei , Shufan Wu , Christophe Baehr\n- **URL**: [https://arxiv.org/abs/2603.12304](https://arxiv.org/abs/2603.12304)\n- **Abstract**:\n> This paper introduces a novel optimization framework that fundamentally integrates the Minimum Description Length (MDL) principle into the training dynamics of deep neural networks. Moving beyond its conventional role as a model selection criterion, we reformulate MDL as an active, adaptive driving force within the optimization process itself. The core of our method is a geometrically-grounded cognitive manifold whose evolution is governed by a \\textit{coupled Ricci flow}, enriched with a novel \\textit{MDL Drive} term derived from first principles. This drive, modulated by the task-loss gradient, creates a seamless harmony between data fidelity and model simplification, actively compressing the internal representation during training. We establish a comprehensive theoretical foundation, proving key properties including the monotonic decrease of description length (Theorem~\\ref{thm:convergence}), a finite number of topological phase transitions via a geometric surgery protocol (Theorems~\\ref{thm:surgery}, \\ref{thm:ultimate_fate}), and the emergence of universal critical behavior (Theorem~\\ref{thm:universality}). Furthermore, we provide a practical, computationally efficient algorithm with $O(N \\log N)$ per-iteration complexity (Theorem~\\ref{thm:complexity}), alongside guarantees for numerical stability (Theorem~\\ref{thm:stability}) and exponential convergence under convexity assumptions (Theorem~\\ref{thm:convergence_rate}). Empirical validation on synthetic regression and classification tasks confirms the theoretical predictions, demonstrating the algorithm's efficacy in achieving robust generalization and autonomous model simplification. This work provides a principled path toward more autonomous, generalizable, and interpretable AI systems by unifying geometric deep learning with information-theoretic principles.\n\n---\n\n## 125. Global Evolutionary Steering: Refining Activation Steering Control via Cross-Layer Consistency\n- **Authors**: Xinyan Jiang , Wenjing Yu , Di Wang , Lijie Hu\n- **URL**: [https://arxiv.org/abs/2603.12298](https://arxiv.org/abs/2603.12298)\n- **Abstract**:\n> Activation engineering enables precise control over Large Language Models (LLMs) without the computational cost of fine-tuning. However, existing methods deriving vectors from static activation differences are susceptible to high-dimensional noise and layer-wise semantic drift, often capturing spurious correlations rather than the target intent. To address this, we propose Global Evolutionary Refined Steering (GER-steer), a training-free framework that grounded in the geometric stability of the network's representation evolution. GER-steer exploits this global signal to rectify raw steering vectors, effectively decoupling robust semantic intent from orthogonal artifacts. Extensive evaluations confirm that GER-steer consistently outperforms baselines, delivering superior efficacy and generalization without layer-specific tuning, establishing a universal solution for reliable model alignment.\n\n---\n\n## 126. Synthetic Data Generation for Brain-Computer Interfaces: Overview, Benchmarking, and Future Directions\n- **Authors**: Ziwei Wang , Zhentao He , Xingyi He , Hongbin Wang , Tianwang Jia , Jingwei Luo , Siyang Li , Xiaoqing Chen , Dongrui Wu\n- **URL**: [https://arxiv.org/abs/2603.12296](https://arxiv.org/abs/2603.12296)\n- **Abstract**:\n> Deep learning has achieved transformative performance across diverse domains, largely driven by the large-scale, high-quality training data. In contrast, the development of brain-computer interfaces (BCIs) is fundamentally constrained by the limited, heterogeneous, and privacy-sensitive neural recordings. Generating synthetic yet physiologically plausible brain signals has therefore emerged as a compelling way to mitigate data scarcity and enhance model capacity. This survey provides a comprehensive review of brain signal generation for BCIs, covering methodological taxonomies, benchmark experiments, evaluation metrics, and key applications. We systematically categorize existing generative algorithms into four types: knowledge-based, feature-based, model-based, and translation-based approaches. Furthermore, we benchmark existing brain signal generation approaches across four representative BCI paradigms to provide an objective performance comparison. Finally, we discuss the potentials and challenges of current generation approaches and prospect future research on accurate, data-efficient, and privacy-aware BCI systems. The benchmark codebase is publicized at this https URL .\n\n---\n\n## 127. Detecting Miscitation on the Scholarly Web through LLM-Augmented Text-Rich Graph Learning\n- **Authors**: Huidong Wu , Haojia Xiang , Jingtong Gao , Xiangyu Zhao , Dengsheng Wu , Jianping Li\n- **URL**: [https://arxiv.org/abs/2603.12290](https://arxiv.org/abs/2603.12290)\n- **Abstract**:\n> Scholarly web is a vast network of knowledge connected by citations. However, this system is increasingly compromised by miscitation, where references do not support or even contradict the claims they are cited for. Current miscitation detection methods, which primarily rely on semantic similarity or network anomalies, struggle to capture the nuanced relationship between a citation's context and its place in the wider network. While large language models (LLMs) offer powerful capabilities in semantic reasoning for this task, their deployment is hindered by hallucination risks and high computational costs. In this work, we introduce LLM-Augmented Graph Learning-based Miscitation Detector (LAGMiD), a novel framework that leverages LLMs for deep semantic reasoning over citation graphs and distills this knowledge into graph neural networks (GNNs) for efficient and scalable miscitation detection. Specifically, LAGMiD introduces an evidence-chain reasoning mechanism, which uses chain-of-thought prompting, to perform multi-hop citation tracing and assess semantic fidelity. To reduce LLM inference costs, we design a knowledge distillation method aligning GNN embeddings with intermediate LLM reasoning states. A collaborative learning strategy further routes complex cases to the LLM while optimizing the GNN for structure-based generalization. Experiments on three real-world benchmarks show that LAGMiD achieves state-of-the-art miscitation detection with significantly reduced inference cost.\n\n---\n\n## 128. From Garbage to Gold: A Data-Architectural Theory of Predictive Robustness\n- **Authors**: Terrence J. Lee-St. John , Jordan L. Lawson , Bartlomiej Piechowski-Jozwiak\n- **URL**: [https://arxiv.org/abs/2603.12288](https://arxiv.org/abs/2603.12288)\n- **Abstract**:\n> Tabular machine learning presents a paradox: modern models achieve state-of-the-art performance using high-dimensional (high-D), collinear, error-prone data, defying the \"Garbage In, Garbage Out\" mantra. To help resolve this, we synthesize principles from Information Theory, Latent Factor Models, and Psychometrics, clarifying that predictive robustness arises not solely from data cleanliness, but from the synergy between data architecture and model capacity. Partitioning predictor-space \"noise\" into \"Predictor Error\" and \"Structural Uncertainty\" (informational deficits from stochastic generative mappings), we prove that leveraging high-D sets of error-prone predictors asymptotically overcomes both types of noise, whereas cleaning a low-D set is fundamentally bounded by Structural Uncertainty. We demonstrate why \"Informative Collinearity\" (dependencies from shared latent causes) enhances reliability and convergence efficiency, and explain why increased dimensionality reduces the latent inference burden, enabling feasibility with finite samples. To address practical constraints, we propose \"Proactive Data-Centric AI\" to identify predictors that enable robustness efficiently. We also derive boundaries for Systematic Error Regimes and show why models that absorb \"rogue\" dependencies can mitigate assumption violations. Linking latent architecture to Benign Overfitting, we offer a first step towards a unified view of robustness to Outcome Error and predictor-space noise, while also delineating when traditional DCAI's focus on label cleaning remains powerful. By redefining data quality from item-level perfection to portfolio-level architecture, we provide a theoretical rationale for \"Local Factories\" -- learning from live, uncurated enterprise \"data swamps\" -- supporting a deployment paradigm shift from \"Model Transfer\" to \"Methodology Transfer'' to overcome static generalizability limitations.\n\n---\n\n## 129. The DIME Architecture: A Unified Operational Algorithm for Neural Representation, Dynamics, Control and Integration\n- **Authors**: Ionel Cristian Vladu , Nicu Bizdoaca , Ionica Pirici , Tudor-Adrian Balseanu , Eduard Nicusor Bondoc\n- **URL**: [https://arxiv.org/abs/2603.12286](https://arxiv.org/abs/2603.12286)\n- **Abstract**:\n> Modern neuroscience has accumulated extensive evidence on perception, memory, prediction, valuation, and consciousness, yet still lacks an explicit operational architecture capable of integrating these phenomena within a unified computational framework. Existing theories address specific aspects of neural function: predictive coding and active inference emphasize hierarchical inference and prediction error minimization; engram theories explain memory through distributed cell assemblies; neuromodulatory accounts focus on value-dependent regulation of plasticity and behaviour; and global workspace or large-scale network models investigate mechanisms underlying conscious access. Despite their explanatory power, these approaches remain only partially integrated at the architectural level. This work introduces DIME (Detect-Integrate-Mark-Execute), a neural architecture organizing perception, memory, valuation, and conscious access within a common operational cycle. The framework includes four interacting components: engrams, distributed recurrent neural structures supporting multiple activation trajectories; execution threads, spatiotemporal trajectories implementing neural processes; marker systems, neuromodulatory and limbic mechanisms regulating gain, plasticity, and trajectory selection; and hyperengrams, large-scale integrative states associated with operational conscious access. The framework is consistent with empirical evidence from hippocampal indexing, recurrent cortical processing, replay phenomena, large-scale network integration, and neuromodulatory regulation. Formulated at an abstract computational level, DIME may also inform artificial intelligence and robotics by providing an architectural template in which representation, valuation, and temporal sequencing emerge from a unified mechanism. An extended theoretical exposition is available in a companion monograph on Zenodo.\n\n---\n\n## 130. Predictive Analytics for Foot Ulcers Using Time-Series Temperature and Pressure Data\n- **Authors**: Md Tanvir Hasan Turja\n- **URL**: [https://arxiv.org/abs/2603.12278](https://arxiv.org/abs/2603.12278)\n- **Abstract**:\n> Diabetic foot ulcers (DFUs) are a severe complication of diabetes, often resulting in significant morbidity. This paper presents a predictive analytics framework utilizing time-series data captured by wearable foot sensors -- specifically NTC thin-film thermocouples for temperature measurement and FlexiForce pressure sensors for plantar load monitoring. Data was collected from healthy subjects walking on an instrumented pathway. Unsupervised machine learning algorithms, Isolation Forest and K-Nearest Neighbors (KNN), were applied to detect anomalies that may indicate early ulcer risk. Through rigorous data preprocessing and targeted feature engineering, physiologic patterns were extracted to identify subtle changes in foot temperature and pressure. Results demonstrate Isolation Forest is sensitive to micro-anomalies, while KNN is effective in flagging extreme deviations, albeit at a higher false-positive rate. Strong correlations between temperature and pressure readings support combined sensor monitoring for improved predictive accuracy. These findings provide a basis for real-time diabetic foot health surveillance, aiming to facilitate earlier intervention and reduce DFU incidence.\n\n---\n\n## 131. Prompt Injection as Role Confusion\n- **Authors**: Charles Ye , Jasmine Cui , Dylan Hadfield-Menell\n- **URL**: [https://arxiv.org/abs/2603.12277](https://arxiv.org/abs/2603.12277)\n- **Abstract**:\n> Language models remain vulnerable to prompt injection attacks despite extensive safety training. We trace this failure to role confusion: models infer roles from how text is written, not where it comes from. We design novel role probes to capture how models internally identify \"who is speaking.\" These reveal why prompt injection works: untrusted text that imitates a role inherits that role's authority. We test this insight by injecting spoofed reasoning into user prompts and tool outputs, achieving average success rates of 60% on StrongREJECT and 61% on agent exfiltration, across multiple open- and closed-weight models with near-zero baselines. Strikingly, the degree of internal role confusion strongly predicts attack success before generation begins. Our findings reveal a fundamental gap: security is defined at the interface but authority is assigned in latent space. More broadly, we introduce a unifying, mechanistic framework for prompt injection, demonstrating that diverse prompt-injection attacks exploit the same underlying role-confusion mechanism.\n\n---\n\n## 132. Aligning Language Models from User Interactions\n- **Authors**: Thomas Kleine Buening , Jonas Hübotter , Barna Pásztor , Idan Shenfeld , Giorgia Ramponi , Andreas Krause\n- **URL**: [https://arxiv.org/abs/2603.12273](https://arxiv.org/abs/2603.12273)\n- **Abstract**:\n> Multi-turn user interactions are among the most abundant data produced by language models, yet we lack effective methods to learn from them. While typically discarded, these interactions often contain useful information: follow-up user messages may indicate that a response was incorrect, failed to follow an instruction, or did not align with the user's preferences. Importantly, language models are already able to make use of this information in context. After observing a user's follow-up, the same model is often able to revise its behavior. We leverage this ability to propose a principled and scalable method for learning directly from user interactions through self-distillation. By conditioning the model on the user's follow-up message and comparing the resulting token distribution with the original policy, we obtain a target for updating the policy that captures how the model's behavior changes in hindsight. We then distill this hindsight distribution back into the current policy. Remarkably, we show that training on real-world user conversations from WildChat improves language models across standard alignment and instruction-following benchmarks, without regressing other capabilities. The same mechanism enables personalization, allowing models to continually adapt to individual users through interaction without explicit feedback. Our results demonstrate that raw user interactions that arise naturally during deployment enable alignment, personalization, and continual adaptation.\n\n---\n\n## 133. Diagnosing Retrieval Bias Under Multiple In-Context Knowledge Updates in Large Language Models\n- **Authors**: Boyu Qiao , Sean Guo , Xian Yang , Kun Li , Wei Zhou , Songlin Hu , Yunya Song\n- **URL**: [https://arxiv.org/abs/2603.12271](https://arxiv.org/abs/2603.12271)\n- **Abstract**:\n> LLMs are widely used in knowledge-intensive tasks where the same fact may be revised multiple times within context. Unlike prior work focusing on one-shot updates or single conflicts, multi-update scenarios contain multiple historically valid versions that compete at retrieval, yet remain underexplored. This challenge resembles the AB-AC interference paradigm in cognitive psychology: when the same cue A is successively associated with B and C, the old and new associations compete during retrieval, leading to bias. Inspired by this, we introduce a Dynamic Knowledge Instance (DKI) evaluation framework, modeling multi-updates of the same fact as a cue paired with a sequence of updated values, and assess models via endpoint probing of the earliest (initial) and latest (current) states. Across diverse LLMs, we observe that retrieval bias intensifies as updates increase, earliest-state accuracy stays high while latest-state accuracy drops substantially. Diagnostic analyses of attention, hidden-state similarity, and output logits further reveal that these signals become flatter and weakly discriminative on errors, providing little stable basis for identifying the latest update. Finally, cognitively inspired heuristic intervention strategies yield only modest gains and do not eliminate the bias. Our results reveal a persistent challenge in tracking and following knowledge updates in long contexts.\n\n---\n\n## 134. Task-Specific Knowledge Distillation via Intermediate Probes\n- **Authors**: Ryan Brown , Chris Russell\n- **URL**: [https://arxiv.org/abs/2603.12270](https://arxiv.org/abs/2603.12270)\n- **Abstract**:\n> Knowledge distillation from large language models (LLMs) assumes that the teacher's output distribution is a high-quality training signal. On reasoning tasks, this assumption is frequently violated. A model's intermediate representations may encode the correct answer, yet this information is lost or distorted through the vocabulary projection, where prompt formatting and answer-token choices creates brittle, noisy outputs. We introduce \\method{}, a distillation framework that bypasses this bottleneck by training lightweight probes on frozen teacher hidden states and using the probe's predictions, rather than output logits, as supervision for student training. This simple change yields consistent improvements across four reasoning benchmarks (AQuA-RAT, ARC Easy/Challenge, and MMLU), with gains most pronounced under limited data. Probes trained on intermediate representations provide cleaner labels than the teacher's own outputs, effectively denoising the distillation signal. \\method{} requires no architectural changes to student or teacher, is architecture-agnostic, and adds minimal compute since probe training is cheap and teacher representations can be cached. By exploiting internal representations, \\method{} enables practitioners to extract more value from large teacher models without additional training data or architectural complexity.\n\n---\n\n## 135. DART: Input-Difficulty-AwaRe Adaptive Threshold for Early-Exit DNNs\n- **Authors**: Parth Patne , Mahdi Taheri , Christian Herglotz , Maksim Jenihhin , Milos Krstic , Michael Hübner\n- **URL**: [https://arxiv.org/abs/2603.12269](https://arxiv.org/abs/2603.12269)\n- **Abstract**:\n> Early-exit deep neural networks enable adaptive inference by terminating computation when sufficient confidence is achieved, reducing cost for edge AI accelerators in resource-constrained settings. Existing methods, however, rely on suboptimal exit policies, ignore input difficulty, and optimize thresholds independently. This paper introduces DART (Input-Difficulty-Aware Adaptive Threshold), a framework that overcomes these limitations. DART introduces three key innovations: (1) a lightweight difficulty estimation module that quantifies input complexity with minimal computational overhead, (2) a joint exit policy optimization algorithm based on dynamic programming, and (3) an adaptive coefficient management system. Experiments on diverse DNN benchmarks (AlexNet, ResNet-18, VGG-16) demonstrate that DART achieves up to \\textbf{3.3$\\times$} speedup, \\textbf{5.1$\\times$} lower energy, and up to \\textbf{42\\%} lower average power compared to static networks, while preserving competitive accuracy. Extending DART to Vision Transformers (LeViT) yields power (5.0$\\times$) and execution-time (3.6$\\times$) gains but also accuracy loss (up to 17 percent), underscoring the need for transformer-specific early-exit mechanisms. We further introduce the Difficulty-Aware Efficiency Score (DAES), a novel multi-objective metric, under which DART achieves up to a 14.8 improvement over baselines, highlighting superior accuracy, efficiency, and robustness trade-offs.\n\n---\n"}, "c234ff9921f9be8d378a7558752264c977290646": {"t": 1786068095.9227252, "n": 229888, "h": "8c5adbd6d55505f6", "c": "{% raw %}\n# 전체 AI 논문 - 2026-03-17\n\n## 1. Semantic Invariance in Agentic AI\n- **Authors**: I. de Zarzà , J. de Curtò , Jordi Cabot , Pietro Manzoni , Carlos T. Calafate\n- **URL**: [https://arxiv.org/abs/2603.13173](https://arxiv.org/abs/2603.13173)\n- **Abstract**:\n> Large Language Models (LLMs) increasingly serve as autonomous reasoning agents in decision support, scientific problem-solving, and multi-agent coordination systems. However, deploying LLM agents in consequential applications requires assurance that their reasoning remains stable under semantically equivalent input variations, a property we term semantic invariance. Standard benchmark evaluations, which assess accuracy on fixed, canonical problem formulations, fail to capture this critical reliability dimension. To address this shortcoming, in this paper we present a metamorphic testing framework for systematically assessing the robustness of LLM reasoning agents, applying eight semantic-preserving transformations (identity, paraphrase, fact reordering, expansion, contraction, academic context, business context, and contrastive formulation) across seven foundation models spanning four distinct architectural families: Hermes (70B, 405B), Qwen3 (30B-A3B, 235B-A22B), DeepSeek-R1, and gpt-oss (20B, 120B). Our evaluation encompasses 19 multi-step reasoning problems across eight scientific domains. The results reveal that model scale does not predict robustness: the smaller Qwen3-30B-A3B achieves the highest stability (79.6% invariant responses, semantic similarity 0.91), while larger models exhibit greater fragility.\n\n---\n\n## 2. Developing and evaluating a chatbot to support maternal health care\n- **Authors**: Smriti Jha , Vidhi Jain , Jianyu Xu , Grace Liu , Sowmya Ramesh , Jitender Nagpal , Gretchen Chapman , Benjamin Bellows , Siddhartha Goyal , Aarti Singh , Bryan Wilder\n- **URL**: [https://arxiv.org/abs/2603.13168](https://arxiv.org/abs/2603.13168)\n- **Abstract**:\n> The ability to provide trustworthy maternal health information using phone-based chatbots can have a significant impact, particularly in low-resource settings where users have low health literacy and limited access to care. However, deploying such systems is technically challenging: user queries are short, underspecified, and code-mixed across languages, answers require regional context-specific grounding, and partial or missing symptom context makes safe routing decisions difficult. We present a chatbot for maternal health in India developed through a partnership between academic researchers, a health tech company, a public health nonprofit, and a hospital. The system combines (1) stage-aware triage, routing high-risk queries to expert templates, (2) hybrid retrieval over curated maternal/newborn guidelines, and (3) evidence-conditioned generation from an LLM. Our core contribution is an evaluation workflow for high-stakes deployment under limited expert supervision. Targeting both component-level and end-to-end testing, we introduce: (i) a labeled triage benchmark (N=150) achieving 86.7% emergency recall, explicitly reporting the missed-emergency vs. over-escalation trade-off; (ii) a synthetic multi-evidence retrieval benchmark (N=100) with chunk-level evidence labels; (iii) LLM-as-judge comparison on real queries (N=781) using clinician-codesigned criteria; and (iv) expert validation. Our findings show that trustworthy medical assistants in multilingual, noisy settings require defense-in-depth design paired with multi-method evaluation, rather than any single model and evaluation method choice.\n\n---\n\n## 3. When Right Meets Wrong: Bilateral Context Conditioning with Reward-Confidence Correction for GRPO\n- **Authors**: Yu Li , Tian Lan , Zhengling Qi\n- **URL**: [https://arxiv.org/abs/2603.13134](https://arxiv.org/abs/2603.13134)\n- **Abstract**:\n> Group Relative Policy Optimization (GRPO) has emerged as an effective method for training reasoning models. While it computes advantages based on group mean, GRPO treats each output as an independent sample during the optimization and overlooks a vital structural signal: the natural contrast between correct and incorrect solutions within the same group, thus ignoring the rich, comparative data that could be leveraged by explicitly pitting successful reasoning traces against failed ones. To capitalize on this, we present a contrastive reformulation of GRPO, showing that the GRPO objective implicitly maximizes the margin between the policy ratios of correct and incorrect samples. Building on this insight, we propose Bilateral Context Conditioning (BICC), a mechanism that allows the model to cross-reference successful and failed reasoning traces during the optimization, enabling a direct information flow across samples. We further introduce Reward-Confidence Correction (RCC) to stabilize training by dynamically adjusts the advantage baseline in GRPO using reward-confidence covariance derived from the first-order approximation of the variance-minimizing estimator. Both mechanisms require no additional sampling or auxiliary models and can be adapted to all GRPO variants. Experiments on mathematical reasoning benchmarks demonstrate consistent improvements across comprehensive models and algorithms. Code is available at \\href{ this https URL }{ this https URL }.\n\n---\n\n## 4. Steve-Evolving: Open-World Embodied Self-Evolution via Fine-Grained Diagnosis and Dual-Track Knowledge Distillation\n- **Authors**: Zhengwei Xie , Zhisheng Chen , Ziyan Weng , Tingyu Wu , Chenglong Li , Vireo Zhang , Kun Wang\n- **URL**: [https://arxiv.org/abs/2603.13131](https://arxiv.org/abs/2603.13131)\n- **Abstract**:\n> Open-world embodied agents must solve long-horizon tasks where the main bottleneck is not single-step planning quality but how interaction experience is organized and evolved. To this end, we present Steve-Evolving, a non-parametric self-evolving framework that tightly couples fine-grained execution diagnosis with dual-track knowledge distillation in a closed loop. The method follows three phases: Experience Anchoring, Experience Distillation, and Knowledge-Driven Closed-Loop Control. In detail, Experience Anchoring solidifies each subgoal attempt into a structured experience tuple with a fixed schema (pre-state, action, diagnosis-result, and post-state) and organizes it in a three-tier experience space with multi-dimensional indices (e.g., condition signatures, spatial hashing, and semantic tags) plus rolling summarization for efficient and auditable recall. To ensure sufficient information density for attribution, the execution layer provides compositional diagnosis signals beyond binary outcomes, including state-difference summaries, enumerated failure causes, continuous indicators, and stagnation/loop detection. Moreover, successful trajectories of Experience Distillation are generalized into reusable skills with explicit preconditions and verification criteria, while failures are distilled into executable guardrails that capture root causes and forbid risky operations at both subgoal and task granularities. Besides, Knowledge-Driven Closed-Loop Control retrieved skills and guardrails are injected into an LLM planner, and diagnosis-triggered local replanning updates the active constraints online, forming a continual evolution process without any model parameter updates. Experiments on the long-horizon suite of Minecraft MCU demonstrate consistent improvements over static-retrieval baselines.\n\n---\n\n## 5. Beyond Final Answers: CRYSTAL Benchmark for Transparent Multimodal Reasoning Evaluation\n- **Authors**: Wayner Barrios , SouYoung Jin\n- **URL**: [https://arxiv.org/abs/2603.13099](https://arxiv.org/abs/2603.13099)\n- **Abstract**:\n> We introduce CRYSTAL (Clear Reasoning via Yielded Steps, Traceability, and Logic), a diagnostic benchmark with 6,372 instances that evaluates multimodal reasoning through verifiable intermediate steps. We propose two complementary metrics: Match F1, which scores step-level precision and recall via semantic similarity matching, and Ordered Match F1, which further penalizes disordered reasoning chains. References are constructed through a Delphi-inspired pipeline in which four independent MLLMs generate trajectories, which are then aggregated via semantic clustering and validated through human quality gates. Evaluation of 20 MLLMs, including commercial frontier systems not used during benchmark construction, reveals systematic failures that are invisible to answer accuracy: universal cherry-picking (precision far exceeds recall), non-monotonic scaling trade-offs, and disordered reasoning in which no competitive model preserves more than 60% of matched steps in the correct order. Beyond evaluation, we propose the Causal Process Reward (CPR), a multiplicative reward that couples answer correctness with step-level alignment, and CPR-Curriculum, which progressively increases reasoning difficulty during training. CPR-Curriculum achieves a 32% improvement in Match F1 via GRPO where additive reward strategies fail, improving reasoning without manual step annotation.\n\n---\n\n## 6. Structured Distillation for Personalized Agent Memory: 11x Token Reduction with Retrieval Preservation\n- **Authors**: Sydney Lewis\n- **URL**: [https://arxiv.org/abs/2603.13017](https://arxiv.org/abs/2603.13017)\n- **Abstract**:\n> Long conversations with an AI agent create a simple problem for one user: the history is useful, but carrying it verbatim is expensive. We study personalized agent memory: one user's conversation history with an agent, distilled into a compact retrieval layer for later search. Each exchange is compressed into a compound object with four fields (exchange_core, specific_context, thematic room_assignments, and regex-extracted files_touched). The searchable distilled text averages 38 tokens per exchange. Applied to 4,182 conversations (14,340 exchanges) from 6 software engineering projects, the method reduces average exchange length from 371 to 38 tokens, yielding 11x compression. We evaluate whether personalized recall survives that compression using 201 recall-oriented queries, 107 configurations spanning 5 pure and 5 cross-layer search modes, and 5 LLM graders (214,519 consensus-graded query-result pairs). The best pure distilled configuration reaches 96% of the best verbatim MRR (0.717 vs 0.745). Results are mechanism-dependent. All 20 vector search configurations remain non-significant after Bonferroni correction, while all 20 BM25 configurations degrade significantly (effect sizes |d|=0.031-0.756). The best cross-layer setup slightly exceeds the best pure verbatim baseline (MRR 0.759). Structured distillation compresses single-user agent memory without uniformly sacrificing retrieval quality. At 1/11 the context cost, thousands of exchanges fit within a single prompt while the verbatim source remains available for drill-down. We release the implementation and analysis pipeline as open-source software.\n\n---\n\n## 7. Efficient and Interpretable Multi-Agent LLM Routing via Ant Colony Optimization\n- **Authors**: Xudong Wang , Chaoning Zhang , Jiaquan Zhang , Chenghao Li , Qigan Sun , Sung-Ho Bae , Peng Wang , Ning Xie , Jie Zou , Yang Yang , Hengtao Shen\n- **URL**: [https://arxiv.org/abs/2603.12933](https://arxiv.org/abs/2603.12933)\n- **Abstract**:\n> Large Language Model (LLM)-driven Multi-Agent Systems (MAS) have demonstrated strong capability in complex reasoning and tool use, and heterogeneous agent pools further broaden the quality--cost trade-off space. Despite these advances, real-world deployment is often constrained by high inference cost, latency, and limited transparency, which hinders scalable and efficient routing. Existing routing strategies typically rely on expensive LLM-based selectors or static policies, and offer limited controllability for semantic-aware routing under dynamic loads and mixed intents, often resulting in unstable performance and inefficient resource utilization. To address these limitations, we propose AMRO-S, an efficient and interpretable routing framework for Multi-Agent Systems (MAS). AMRO-S models MAS routing as a semantic-conditioned path selection problem, enhancing routing performance through three key mechanisms: First, it leverages a supervised fine-tuned (SFT) small language model for intent inference, providing a low-overhead semantic interface for each query; second, it decomposes routing memory into task-specific pheromone specialists, reducing cross-task interference and optimizing path selection under mixed workloads; finally, it employs a quality-gated asynchronous update mechanism to decouple inference from learning, optimizing routing without increasing latency. Extensive experiments on five public benchmarks and high-concurrency stress tests demonstrate that AMRO-S consistently improves the quality--cost trade-off over strong routing baselines, while providing traceable routing evidence through structured pheromone patterns.\n\n---\n\n## 8. ODRL Policy Comparison Through Normalisation\n- **Authors**: Jaime Osvaldo Salas , Paolo Pareti , George Konstantinidis\n- **URL**: [https://arxiv.org/abs/2603.12926](https://arxiv.org/abs/2603.12926)\n- **Abstract**:\n> The ODRL language has become the standard for representing policies and regulations for digital rights. However its complexity is a barrier to its usage, which has caused many related theoretical and practical works to focus on different, and not interoperable, fragments of ODRL. Moreover, semantically equivalent policies can be expressed in numerous different ways, which makes comparing them and processing them harder. Building on top of a recently defined semantics, we tackle these problems by proposing an approach that involves a parametrised normalisation of ODRL policies into its minimal components which reformulates policies with permissions and prohibitions into policies with permissions exclusively, and simplifies complex logic constraints into simple ones. We provide algorithms to compute a normal form for ODRL policies and simplifying numerical and symbolic constraints. We prove that these algorithms preserve the semantics of policies, and analyse the size complexity of the result, which is exponential on the number of attributes and linear on the number of unique values for these attributes. We show how this makes complex policies representable in more basic fragments of ODRL, and how it reduces the problem of policy comparison to the simpler problem of checking if two rules are identical.\n\n---\n\n## 9. Context is all you need: Towards autonomous model-based process design using agentic AI in flowsheet simulations\n- **Authors**: Pascal Schäfer , Lukas J. Krinke , Martin Wlotzka , Norbert Asprion\n- **URL**: [https://arxiv.org/abs/2603.12813](https://arxiv.org/abs/2603.12813)\n- **Abstract**:\n> Agentic AI systems integrating large language models (LLMs) with reasoning and tooluse capabilities are transforming various domains - in particular, software development. In contrast, their application in chemical process flowsheet modelling remains largely unexplored. In this work, we present an agentic AI framework that delivers assistance in an industrial flowsheet simulation environment. To this end, we show the capabilities of GitHub Copilot (GitHub, Inc., 2026), when using state-of-the-art LLMs, such as Claude Opus 4.6 (Anthropic, PBC, 2026), to generate valid syntax for our in-house process modelling tool Chemasim using the technical documentation and a few commented examples as context. Based on this, we develop a multi-agent system that decomposes process development tasks with one agent solving the abstract problem using engineering knowledge and another agent implementing the solution as Chemasim code. We demonstrate the effectiveness of our framework for typical flowsheet modelling examples, including (i) a reaction/separation process, (ii) a pressure-swing distillation, and (iii) a heteroazeotropic distillation including entrainer selection. Along these lines, we discuss current limitations of the framework and outline future research directions to further enhance its capabilities.\n\n---\n\n## 10. AI Model Modulation with Logits Redistribution\n- **Authors**: Zihan Wang , Zhongkui Ma , Xinguo Feng , Zhiyang Mei , Ethan Ma , Derui Wang , Minhui Xue , Guangdong Bai\n- **URL**: [https://arxiv.org/abs/2603.12755](https://arxiv.org/abs/2603.12755)\n- **Abstract**:\n> Large-scale models are typically adapted to meet the diverse requirements of model owners and users. However, maintaining multiple specialized versions of the model is inefficient. In response, we propose AIM, a novel model modulation paradigm that enables a single model to exhibit diverse behaviors to meet the specific end requirements. AIM enables two key modulation modes: utility and focus modulations. The former provides model owners with dynamic control over output quality to deliver varying utility levels, and the latter offers users precise control to shift model's focused input features. AIM introduces a logits redistribution strategy that operates in a training data-agnostic and retraining-free manner. We establish a formal foundation to ensure AIM's regulation capability, based on the statistical properties of logits ordering via joint probability distributions. Our evaluation confirms AIM's practicality and versatility for Al model modulation, with tasks spanning image classification, semantic segmentation and text generation, and prevalent architectures including ResNet, SegFormer and Llama.\n\n---\n\n## 11. ToolTree: Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning\n- **Authors**: Shuo Yang , Soyeon Caren Han , Yihao Ding , Shuhe Wang , Eduard Hoy\n- **URL**: [https://arxiv.org/abs/2603.12740](https://arxiv.org/abs/2603.12740)\n- **Abstract**:\n> Large Language Model (LLM) agents are increasingly applied to complex, multi-step tasks that require interaction with diverse external tools across various domains. However, current LLM agent tool planning methods typically rely on greedy, reactive tool selection strategies that lack foresight and fail to account for inter-tool dependencies. In this paper, we present ToolTree, a novel Monte Carlo tree search-inspired planning paradigm for tool planning. ToolTree explores possible tool usage trajectories using a dual-stage LLM evaluation and bidirectional pruning mechanism that enables the agent to make informed, adaptive decisions over extended tool-use sequences while pruning less promising branches before and after the tool execution. Empirical evaluations across both open-set and closed-set tool planning tasks on 4 benchmarks demonstrate that ToolTree consistently improves performance while keeping the highest efficiency, achieving an average gain of around 10\\% compared to the state-of-the-art planning paradigm.\n\n---\n\n## 12. On Using Machine Learning to Early Detect Catastrophic Failures in Marine Diesel Engines\n- **Authors**: Francesco Maione , Paolo Lino , Giuseppe Giannino , Guido Maione\n- **URL**: [https://arxiv.org/abs/2603.12733](https://arxiv.org/abs/2603.12733)\n- **Abstract**:\n> Catastrophic failures of marine engines imply severe loss of functionality and destroy or damage the systems irreversibly. Being sudden and often unpredictable events, they pose a severe threat to navigation, crew, and passengers. The abrupt nature makes early detection the only effective countermeasure. However, research has concentrated on modeling the gradual degradation of components, with limited attention to sudden and anomalous phenomena. This work proposes a new method for early detection of catastrophic failures. Based on real data from a failed engine, the approach evaluates the derivatives of the deviation between actual sensor readings and expected values of engine variables. Predictions are obtained by a Random Forest, which is the most suitable Machine Learning algorithm among the tested ones. Traditional methods focus on deviations of monitored signals, whereas the proposed approach employs the derivatives of the deviations to provide earlier indications of abnormal dynamics, and to alert that a rapid and dangerous event is breaking out within the system. The method allows the detection of anomalies before measurements reach critical thresholds and alarms are triggered, which is the common method in industry. Consequently, operators can be warned in advance and shut down the engine, then prevent damage and unexpected power loss. Moreover, they have the time to safely change the ship route and avoid potential obstacles. Simulation results conf irm the effectiveness of the proposed approach in anticipating occurrence of catastrophic failures. Validation on real-world data further reinforces the robustness and practical applicability of the method. It is worth noting that data acquisition to train the predictive algorithm is not a problem, since a Deep Learning-based data augmentation procedure is used.\n\n---\n\n## 13. AI Planning Framework for LLM-Based Web Agents\n- **Authors**: Orit Shahnovsky , Rotem Dror\n- **URL**: [https://arxiv.org/abs/2603.12710](https://arxiv.org/abs/2603.12710)\n- **Abstract**:\n> Developing autonomous agents for web-based tasks is a core challenge in AI. While Large Language Model (LLM) agents can interpret complex user requests, they often operate as black boxes, making it difficult to diagnose why they fail or how they plan. This paper addresses this gap by formally treating web tasks as sequential decision-making processes. We introduce a taxonomy that maps modern agent architectures to traditional planning paradigms: Step-by-Step agents to Breadth-First Search (BFS), Tree Search agents to Best-First Tree Search, and Full-Plan-in-Advance agents to Depth-First Search (DFS). This framework allows for a principled diagnosis of system failures like context drift and incoherent task decomposition. To evaluate these behaviors, we propose five novel evaluation metrics that assess trajectory quality beyond simple success rates. We support this analysis with a new dataset of 794 human-labeled trajectories from the WebArena benchmark. Finally, we validate our evaluation framework by comparing a baseline Step-by-Step agent against a novel Full-Plan-in-Advance implementation. Our results reveal that while the Step-by-Step agent aligns more closely with human gold trajectories (38% overall success), the Full-Plan-in-Advance agent excels in technical measures such as element accuracy (89%), demonstrating the necessity of our proposed metrics for selecting appropriate agent architectures based on specific application constraints.\n\n---\n\n## 14. Generating Expressive and Customizable Evals for Timeseries Data Analysis Agents with AgentFuel\n- **Authors**: Aadyaa Maddi , Prakhar Naval , Deepti Mande , Shane Duan , Muckai Girish , Vyas Sekar\n- **URL**: [https://arxiv.org/abs/2603.12483](https://arxiv.org/abs/2603.12483)\n- **Abstract**:\n> Across many domains (e.g., IoT, observability, telecommunications, cybersecurity), there is an emerging adoption of conversational data analysis agents that enable users to \"talk to your data\" to extract insights. Such data analysis agents operate on timeseries data models; e.g., measurements from sensors or events monitoring user clicks and actions in product analytics. We evaluate 6 popular data analysis agents (both open-source and proprietary) on domain-specific data and query types, and find that they fail on stateful and incident-specific queries. We observe two key expressivity gaps in existing evals: domain-customized datasets and domain-specific query types. To enable practitioners in such domains to generate customized and expressive evals for such timeseries data agents, we present AgentFuel. AgentFuel helps domain experts quickly create customized evals to perform end-to-end functional tests. We show that AgentFuel's benchmarks expose key directions for improvement in existing data agent frameworks. We also present anecdotal evidence that using AgentFuel can improve agent performance (e.g., with GEPA). AgentFuel benchmarks are available at this https URL .\n\n---\n\n## 15. Efficient Reasoning with Balanced Thinking\n- **Authors**: Yulin Li , Tengyao Tu , Li Ding , Junjie Wang , Huiling Zhen , Yixin Chen , Yong Li , Zhuotao Tian\n- **URL**: [https://arxiv.org/abs/2603.12372](https://arxiv.org/abs/2603.12372)\n- **Abstract**:\n> Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they often suffer from overthinking, expending redundant computational steps on simple problems, or underthinking, failing to explore sufficient reasoning paths despite inherent capabilities. These issues lead to inefficiencies and potential inaccuracies, limiting practical deployment in resource-constrained settings. Existing methods to mitigate overthinking, such as suppressing reflective keywords or adjusting reasoning length, may inadvertently induce underthinking, compromising accuracy. Therefore, we propose ReBalance, a training-free framework that achieves efficient reasoning with balanced thinking. ReBalance leverages confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence. By aggregating hidden states from a small-scale dataset into reasoning mode prototypes, we compute a steering vector to guide LRMs' reasoning trajectories. A dynamic control function modulates this vector's strength and direction based on real-time confidence, pruning redundancy during overthinking, and promoting exploration during underthinking. Extensive experiments conducted on four models ranging from 0.5B to 32B, and across nine benchmarks in math reasoning, general question answering, and coding tasks demonstrate that ReBalance effectively reduces output redundancy while improving accuracy, offering a general, training-free, and plug-and-play strategy for efficient and robust LRM deployment. Code is available at this https URL .\n\n---\n\n## 16. Context-Enriched Natural Language Descriptions of Vessel Trajectories\n- **Authors**: Kostas Patroumpas , Alexandros Troupiotis-Kapeliaris , Giannis Spiliopoulos , Panagiotis Betchavas , Dimitrios Skoutas , Dimitris Zissis , Nikos Bikakis\n- **URL**: [https://arxiv.org/abs/2603.12287](https://arxiv.org/abs/2603.12287)\n- **Abstract**:\n> We address the problem of transforming raw vessel trajectory data collected from AIS into structured and semantically enriched representations interpretable by humans and directly usable by machine reasoning systems. We propose a context-aware trajectory abstraction framework that segments noisy AIS sequences into distinct trips each consisting of clean, mobility-annotated episodes. Each episode is further enriched with multi-source contextual information, such as nearby geographic entities, offshore navigation features, and weather conditions. Crucially, such representations can support generation of controlled natural language descriptions using LLMs. We empirically examine the quality of such descriptions generated using several LLMs over AIS data along with open contextual features. By increasing semantic density and reducing spatiotemporal complexity, this abstraction can facilitate downstream analytics and enable integration with LLMs for higher-level maritime reasoning tasks.\n\n---\n\n## 17. PhysMoDPO: Physically-Plausible Humanoid Motion with Preference Optimization\n- **Authors**: Yangsong Zhang , Anujith Muraleedharan , Rikhat Akizhanov , Abdul Ahad Butt , Gül Varol , Pascal Fua , Fabio Pizzati , Ivan Laptev\n- **URL**: [https://arxiv.org/abs/2603.13228](https://arxiv.org/abs/2603.13228)\n- **Abstract**:\n> Recent progress in text-conditioned human motion generation has been largely driven by diffusion models trained on large-scale human motion data. Building on this progress, recent methods attempt to transfer such models for character animation and real robot control by applying a Whole-Body Controller (WBC) that converts diffusion-generated motions into executable trajectories. While WBC trajectories become compliant with physics, they may expose substantial deviations from original motion. To address this issue, we here propose PhysMoDPO, a Direct Preference Optimization framework. Unlike prior work that relies on hand-crafted physics-aware heuristics such as foot-sliding penalties, we integrate WBC into our training pipeline and optimize diffusion model such that the output of WBC becomes compliant both with physics and original text instructions. To train PhysMoDPO we deploy physics-based and task-specific rewards and use them to assign preference to synthesized trajectories. Our extensive experiments on text-to-motion and spatial control tasks demonstrate consistent improvements of PhysMoDPO in both physical realism and task-related metrics on simulated robots. Moreover, we demonstrate that PhysMoDPO results in significant improvements when applied to zero-shot motion transfer in simulation and for real-world deployment on a G1 humanoid robot.\n\n---\n\n## 18. Visual-ERM: Reward Modeling for Visual Equivalence\n- **Authors**: Ziyu Liu , Shengyuan Ding , Xinyu Fang , Xuanlang Dai , Penghui Yang , Jianze Liang , Jiaqi Wang , Kai Chen , Dahua Lin , Yuhang Zang\n- **URL**: [https://arxiv.org/abs/2603.13224](https://arxiv.org/abs/2603.13224)\n- **Abstract**:\n> Vision-to-code tasks require models to reconstruct structured visual inputs, such as charts, tables, and SVGs, into executable or structured representations with high visual fidelity. While recent Large Vision Language Models (LVLMs) achieve strong results via supervised fine-tuning, reinforcement learning remains challenging due to misaligned reward signals. Existing rewards either rely on textual rules or coarse visual embedding similarity, both of which fail to capture fine-grained visual discrepancies and are vulnerable to reward hacking. We propose Visual Equivalence Reward Model (Visual-ERM), a multimodal generative reward model that provides fine-grained, interpretable, and task-agnostic feedback to evaluate vision-to-code quality directly in the rendered visual space. Integrated into RL, Visual-ERM improves Qwen3-VL-8B-Instruct by +8.4 on chart-to-code and yields consistent gains on table and SVG parsing (+2.7, +4.1 on average), and further strengthens test-time scaling via reflection and revision. We also introduce VisualCritic-RewardBench (VC-RewardBench), a benchmark for judging fine-grained image-to-image discrepancies on structured visual data, where Visual-ERM at 8B decisively outperforms Qwen3-VL-235B-Instruct and approaches leading closed-source models. Our results suggest that fine-grained visual reward supervision is both necessary and sufficient for vision-to-code RL, regardless of task specificity.\n\n---\n\n## 19. From Experiments to Expertise: Scientific Knowledge Consolidation for AI-Driven Computational Research\n- **Authors**: Haonan Huang\n- **URL**: [https://arxiv.org/abs/2603.13191](https://arxiv.org/abs/2603.13191)\n- **Abstract**:\n> While large language models (LLMs) have transformed AI agents into proficient executors of computational materials science, performing a hundred simulations does not make a researcher. What distinguishes research from routine execution is the progressive accumulation of knowledge -- learning which approaches fail, recognizing patterns across systems, and applying understanding to new problems. However, the prevailing paradigm in AI-driven computational science treats each execution in isolation, largely discarding hard-won insights between runs. Here we present QMatSuite, an open-source platform closing this gap. Agents record findings with full provenance, retrieve knowledge before new calculations, and in dedicated reflection sessions correct erroneous findings and synthesize observations into cross-compound patterns. In benchmarks on a six-step quantum-mechanical simulation workflow, accumulated knowledge reduces reasoning overhead by 67% and improves accuracy from 47% to 3% deviation from literature -- and when transferred to an unfamiliar material, achieves 1% deviation with zero pipeline failures.\n\n---\n\n## 20. LLM Constitutional Multi-Agent Governance\n- **Authors**: J. de Curtò , I. de Zarzà\n- **URL**: [https://arxiv.org/abs/2603.13189](https://arxiv.org/abs/2603.13189)\n- **Abstract**:\n> Large Language Models (LLMs) can generate persuasive influence strategies that shift cooperative behavior in multi-agent populations, but a critical question remains: does the resulting cooperation reflect genuine prosocial alignment, or does it mask erosion of agent autonomy, epistemic integrity, and distributional fairness? We introduce Constitutional Multi-Agent Governance (CMAG), a two-stage framework that interposes between an LLM policy compiler and a networked agent population, combining hard constraint filtering with soft penalized-utility optimization that balances cooperation potential against manipulation risk and autonomy pressure. We propose the Ethical Cooperation Score (ECS), a multiplicative composite of cooperation, autonomy, integrity, and fairness that penalizes cooperation achieved through manipulative means. In experiments on scale-free networks of 80 agents under adversarial conditions (70% violating candidates), we benchmark three regimes: full CMAG, naive filtering, and unconstrained optimization. While unconstrained optimization achieves the highest raw cooperation (0.873), it yields the lowest ECS (0.645) due to severe autonomy erosion (0.867) and fairness degradation (0.888). CMAG attains an ECS of 0.741, a 14.9% improvement, while preserving autonomy at 0.985 and integrity at 0.995, with only modest cooperation reduction to 0.770. The naive ablation (ECS = 0.733) confirms that hard constraints alone are insufficient. Pareto analysis shows CMAG dominates the cooperation-autonomy trade-off space, and governance reduces hub-periphery exposure disparities by over 60%. These findings establish that cooperation is not inherently desirable without governance: constitutional constraints are necessary to ensure that LLM-mediated influence produces ethically stable outcomes rather than manipulative equilibria.\n\n---\n\n## 21. Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights\n- **Authors**: Xingli Fang , Jung-Eun Kim\n- **URL**: [https://arxiv.org/abs/2603.13186](https://arxiv.org/abs/2603.13186)\n- **Abstract**:\n> Prior approaches for membership privacy preservation usually update or retrain all weights in neural networks, which is costly and can lead to unnecessary utility loss or even more serious misalignment in predictions between training data and non-training data. In this work, we observed three insights: i) privacy vulnerability exists in a very small fraction of weights; ii) however, most of those weights also critically impact utility performance; iii) the importance of weights stems from their locations rather than their values. According to these insights, to preserve privacy, we score critical weights, and instead of discarding those neurons, we rewind only the weights for fine-tuning. We show that, through extensive experiments, this mechanism exhibits outperforming resilience in most cases against Membership Inference Attacks while maintaining utility.\n\n---\n\n## 22. MXNorm: Reusing MXFP block scales for efficient tensor normalisation\n- **Authors**: Callum McLean , Luke Y. Prince , Alexandre Payot , Paul Balança , Carlo Luschi\n- **URL**: [https://arxiv.org/abs/2603.13180](https://arxiv.org/abs/2603.13180)\n- **Abstract**:\n> Matrix multiplication performance has long been the major bottleneck to scaling deep learning workloads, which has stimulated the design of new accelerators that use increasingly low-precision number formats. However, improvements in matrix multiplication performance have far outstripped improvements in performance on reductions and elementwise computations, which are still being performed in higher precision. In this work, we propose MXNorm, a drop-in replacement for RMSNorm that estimates the RMS using only the block scales calculated as part of the MXFP8 cast and enables a 32x decrease in the size of reduction needed for normalization. We validate our approximation method on pre-training of Llama 3 models of 125M, 1B and 8B parameters, finding minimal loss of training accuracy compared to a baseline using RMSNorm with MXFP8 matmuls. We also show practical kernel speedups using only this http URL of up to 2.4x for MXNorm over RMSNorm, corresponding to a 1.3% speedup in Llama 3 8B transformer layers in MXFP8 and a 2.6% speedup in NVFP4.\n\n---\n\n## 23. Clustering Astronomical Orbital Synthetic Data Using Advanced Feature Extraction and Dimensionality Reduction Techniques\n- **Authors**: Eraldo Pereira Marinho , Nelson Callegari Junior , Fabricio Aparecido Breve , Caetano Mazzoni Ranieri\n- **URL**: [https://arxiv.org/abs/2603.13177](https://arxiv.org/abs/2603.13177)\n- **Abstract**:\n> The dynamics of Saturn's satellite system offer a rich framework for studying orbital stability and resonance interactions. Traditional methods for analysing such systems, including Fourier analysis and stability metrics, struggle with the scale and complexity of modern datasets. This study introduces a machine learning-based pipeline for clustering approximately 22,300 simulated satellite orbits, addressing these challenges with advanced feature extraction and dimensionality reduction techniques. The key to this approach is using MiniRocket, which efficiently transforms 400 timesteps into a 9,996-dimensional feature space, capturing intricate temporal patterns. Additional automated feature extraction and dimensionality reduction techniques refine the data, enabling robust clustering analysis. This pipeline reveals stability regions, resonance structures, and other key behaviours in Saturn's satellite system, providing new insights into their long-term dynamical evolution. By integrating computational tools with traditional celestial mechanics techniques, this study offers a scalable and interpretable methodology for analysing large-scale orbital datasets and advancing the exploration of planetary dynamics.\n\n---\n\n## 24. ESG-Bench: Benchmarking Long-Context ESG Reports for Hallucination Mitigation\n- **Authors**: Siqi Sun , Ben Peng Wu , Mali Jin , Peizhen Bai , Hanpei Zhang , Xingyi Song\n- **URL**: [https://arxiv.org/abs/2603.13154](https://arxiv.org/abs/2603.13154)\n- **Abstract**:\n> As corporate responsibility increasingly incorporates environmental, social, and governance (ESG) criteria, ESG reporting is becoming a legal requirement in many regions and a key channel for documenting sustainability practices and assessing firms' long-term and ethical performance. However, the length and complexity of ESG disclosures make them difficult to interpret and automate the analysis reliably. To support scalable and trustworthy analysis, this paper introduces ESG-Bench, a benchmark dataset for ESG report understanding and hallucination mitigation in large language models (LLMs). ESG-Bench contains human-annotated question-answer (QA) pairs grounded in real-world ESG report contexts, with fine-grained labels indicating whether model outputs are factually supported or hallucinated. Framing ESG report analysis as a QA task with verifiability constraints enables systematic evaluation of LLMs' ability to extract and reason over ESG content and provides a new use case: mitigating hallucinations in socially sensitive, compliance-critical settings. We design task-specific Chain-of-Thought (CoT) prompting strategies and fine-tune multiple state-of-the-art LLMs on ESG-Bench using CoT-annotated rationales. Our experiments show that these CoT-based methods substantially outperform standard prompting and direct fine-tuning in reducing hallucinations, and that the gains transfer to existing QA benchmarks beyond the ESG domain.\n\n---\n\n## 25. Developing the PsyCogMetrics AI Lab to Evaluate Large Language Models and Advance Cognitive Science -- A Three-Cycle Action Design Science Study\n- **Authors**: Zhiye Jin , Yibai Li , K. D. Joshi , Xuefei (Nancy) Deng , Xiaobing (Emily)Li\n- **URL**: [https://arxiv.org/abs/2603.13126](https://arxiv.org/abs/2603.13126)\n- **Abstract**:\n> This study presents the development of the PsyCogMetrics AI Lab ( this http URL ), an integrated, cloud-based platform that operationalizes psychometric and cognitive-science methodologies for Large Language Model (LLM) evaluation. Framed as a three-cycle Action Design Science study, the Relevance Cycle identifies key limitations in current evaluation methods and unfulfilled stakeholder needs. The Rigor Cycle draws on kernel theories such as Popperian falsifiability, Classical Test Theory, and Cognitive Load Theory to derive deductive design objectives. The Design Cycle operationalizes these objectives through nested Build-Intervene-Evaluate loops. The study contributes a novel IT artifact, a validated design for LLM evaluation, benefiting research at the intersection of AI, psychology, cognitive science, and the social and behavioral sciences.\n\n---\n\n## 26. Geometry-Guided Camera Motion Understanding in VideoLLMs\n- **Authors**: Haoan Feng , Sri Harsha Musunuri , Guan-Ming Su\n- **URL**: [https://arxiv.org/abs/2603.13119](https://arxiv.org/abs/2603.13119)\n- **Abstract**:\n> Camera motion is a fundamental geometric signal that shapes visual perception and cinematic style, yet current video-capable vision-language models (VideoLLMs) rarely represent it explicitly and often fail on fine-grained motion primitives. We address this gap with a framework of $\\textbf{benchmarking}$, $\\textbf{diagnosis}$, and $\\textbf{injection}$. We curate $\\textbf{CameraMotionDataset}$, a large-scale synthetic dataset with explicit camera control, formulate camera motion as constraint-aware multi-label recognition, and construct a VQA benchmark--$\\textbf{CameraMotionVQA}$. Across diverse off-the-shelf VideoLLMs, we observe substantial errors in recognizing camera motion primitives. Probing experiments on a Qwen2.5-VL vision encoder suggest that camera motion cues are weakly represented, especially in deeper ViT blocks, helping explain the observed failure modes. To bridge this gap without costly training or fine-tuning, we propose a lightweight, model-agnostic pipeline that extracts geometric camera cues from 3D foundation models (3DFMs), predicts constrained motion primitives with a temporal classifier, and injects them into downstream VideoLLM inference via structured prompting. Experiments demonstrate improved motion recognition and more camera-aware model responses, highlighting geometry-driven cue extraction and structured prompting as practical steps toward a camera-aware VideoLLM and VLA system. The dataset and benchmark is publicly available at this https URL .\n\n---\n\n## 27. BoSS: A Best-of-Strategies Selector as an Oracle for Deep Active Learning\n- **Authors**: Denis Huseljic , Paul Hahn , Marek Herde , Christoph Sandrock , Bernhard Sick\n- **URL**: [https://arxiv.org/abs/2603.13109](https://arxiv.org/abs/2603.13109)\n- **Abstract**:\n> Active learning (AL) aims to reduce annotation costs while maximizing model performance by iteratively selecting valuable instances. While foundation models have made it easier to identify these instances, existing selection strategies still lack robustness across different models, annotation budgets, and datasets. To highlight the potential weaknesses of existing AL strategies and provide a reference point for research, we explore oracle strategies, i.e., strategies that approximate the optimal selection by accessing ground-truth information unavailable in practical AL scenarios. Current oracle strategies, however, fail to scale effectively to large datasets and complex deep neural networks. To tackle these limitations, we introduce the Best-of-Strategy Selector (BoSS), a scalable oracle strategy designed for large-scale AL scenarios. BoSS constructs a set of candidate batches through an ensemble of selection strategies and then selects the batch yielding the highest performance gain. As an ensemble of selection strategies, BoSS can be easily extended with new state-of-the-art strategies as they emerge, ensuring it remains a reliable oracle strategy in the future. Our evaluation demonstrates that i) BoSS outperforms existing oracle strategies, ii) state-of-the-art AL strategies still fall noticeably short of oracle performance, especially in large-scale datasets with many classes, and iii) one possible solution to counteract the inconsistent performance of AL strategies might be to employ an ensemble-based approach for the selection.\n\n---\n\n## 28. Evaluating VLMs' Spatial Reasoning Over Robot Motion: A Step Towards Robot Planning with Motion Preferences\n- **Authors**: Wenxi Wu , Jingjing Zhang , Martim Brandão\n- **URL**: [https://arxiv.org/abs/2603.13100](https://arxiv.org/abs/2603.13100)\n- **Abstract**:\n> Understanding user instructions and object spatial relations in surrounding environments is crucial for intelligent robot systems to assist humans in various tasks. The natural language and spatial reasoning capabilities of Vision-Language Models (VLMs) have the potential to enhance the generalization of robot planners on new tasks, objects, and motion specifications. While foundation models have been applied to task planning, it is still unclear the degree to which they have the capability of spatial reasoning required to enforce user preferences or constraints on motion, such as desired distances from objects, topological properties, or motion style preferences. In this paper, we evaluate the capability of four state-of-the-art VLMs at spatial reasoning over robot motion, using four different querying methods. Our results show that, with the highest-performing querying method, Qwen2.5-VL achieves 71.4% accuracy zero-shot and 75% on a smaller model after fine-tuning, and GPT-4o leads to lower performance. We evaluate two types of motion preferences (object-proximity and path-style), and we also analyze the trade-off between accuracy and computation cost in number of tokens. This work shows some promise in the potential of VLM integration with robot motion planning pipelines.\n\n---\n\n## 29. Human-in-the-Loop LLM Grading for Handwritten Mathematics Assessments\n- **Authors**: Arne Vanhoyweghen , Vincent Holst , Melika Mobini , Lukas Van de Voorde , Tibo Vanleke , Bert Verbruggen , Brecht Verbeken , Andres Algaba , Sam Verboven , Marie-Anne Guerry , Filip Van Droogenbroeck , Vincent Ginis\n- **URL**: [https://arxiv.org/abs/2603.13083](https://arxiv.org/abs/2603.13083)\n- **Abstract**:\n> Providing timely and individualised feedback on handwritten student work is highly beneficial for learning but difficult to achieve at scale. This challenge has become more pressing as generative AI undermines the reliability of take-home assessments, shifting emphasis toward supervised, in-class evaluation. We present a scalable, end-to-end workflow for LLM-assisted grading of short, pen-and-paper assessments. The workflow spans (1) constructing solution keys, (2) developing detailed rubric-style grading keys used to guide the LLM, and (3) a grading procedure that combines automated scanning and anonymisation, multi-pass LLM scoring, automated consistency checks, and mandatory human verification. We deploy the system in two undergraduate mathematics courses using six low-stakes in-class tests. Empirically, LLM assistance reduces grading time by approximately 23% while achieving agreement comparable to, and in several cases tighter than, fully manual grading. Occasional model errors occur but are effectively contained by the hybrid design. Overall, our results show that carefully embedded human-in-the-loop LLM grading can substantially reduce workload while maintaining fairness and accuracy.\n\n---\n\n## 30. GeoChemAD: Benchmarking Unsupervised Geochemical Anomaly Detection for Mineral Exploration\n- **Authors**: Yihao Ding , Yiran Zhang , Chris Gonzalez , Eun-Jung Holden , Wei Liu\n- **URL**: [https://arxiv.org/abs/2603.13068](https://arxiv.org/abs/2603.13068)\n- **Abstract**:\n> Geochemical anomaly detection plays a critical role in mineral exploration as deviations from regional geochemical baselines may indicate mineralization. Existing studies suffer from two key limitations: (1) single region scenarios which limit model generalizability; (2) proprietary datasets, which makes result reproduction unattainable. In this work, we introduce \\textbf{GeoChemAD}, an open-source benchmark dataset compiled from government-led geological surveys, covering multiple regions, sampling sources, and target elements. The dataset comprises eight subsets representing diverse spatial scales and sampling conditions. To establish strong baselines, we reproduce and benchmark a range of unsupervised anomaly detection methods, including statistical models, generative and transformer-based approaches. Furthermore, we propose \\textbf{GeoChemFormer}, a transformer-based framework that leverages self-supervised pretraining to learn target-element-aware geochemical representations for spatial samples. Extensive experiments demonstrate that GeoChemFormer consistently achieves superior and robust performance across all eight subsets, outperforming existing unsupervised methods in both anomaly detection accuracy and generalization capability. The proposed dataset and framework provide a foundation for reproducible research and future development in this direction.\n\n---\n\n## 31. L2GTX: From Local to Global Time Series Explanations\n- **Authors**: Ephrem Tibebe Mekonnen , Luca Longo , Lucas Rizzo , Pierpaolo Dondio\n- **URL**: [https://arxiv.org/abs/2603.13065](https://arxiv.org/abs/2603.13065)\n- **Abstract**:\n> Deep learning models achieve high accuracy in time series classification, yet understanding their class-level decision behaviour remains challenging. Explanations for time series must respect temporal dependencies and identify patterns that recur across instances. Existing approaches face three limitations: model-agnostic XAI methods developed for images and tabular data do not readily extend to time series, global explanation synthesis for time series remains underexplored, and most existing global approaches are model-specific. We propose L2GTX, a model-agnostic framework that generates class-wise global explanations by aggregating local explanations from a representative set of instances. L2GTX extracts clusters of parameterised temporal event primitives, such as increasing or decreasing trends and local extrema, together with their importance scores from instance-level explanations produced by LOMATCE. These clusters are merged across instances to reduce redundancy, and an instance-cluster importance matrix is used to estimate global relevance. Under a user-defined instance selection budget, L2GTX selects representative instances that maximise coverage of influential clusters. Events from the selected instances are then aggregated into concise class-wise global explanations. Experiments on six benchmark time series datasets show that L2GTX produces compact and interpretable global explanations while maintaining stable global faithfulness measured as mean local surrogate fidelity.\n\n---\n\n## 32. Competition-Aware CPC Forecasting with Near-Market Coverage\n- **Authors**: Sebastian Frey , Edoardo Beccari , Maximilian Kranz , Nicolò Alberto Pellizzari , Ali Mete Karaman , Qiwei Han , Maximilian Kaiser\n- **URL**: [https://arxiv.org/abs/2603.13059](https://arxiv.org/abs/2603.13059)\n- **Abstract**:\n> Cost-per-click (CPC) in paid search is a volatile auction outcome generated by a competitive landscape that is only partially observable from any single advertiser's history. Using Google Ads auction logs from a concentrated car-rental market (2021--2023), we forecast weekly CPC for 1,811 keyword series and approximate latent competition through complementary signals derived from keyword text, CPC trajectories, and geographic market structure. We construct (i) semantic neighborhoods and a semantic keyword graph from pretrained transformer-based representations of keyword text, (ii) behavioral neighborhoods via Dynamic Time Warping (DTW) alignment of CPC trajectories, and (iii) geographic-intent covariates capturing localized demand and marketplace heterogeneity. We extensively evaluate these signals both as stand-alone covariates and as relational priors in spatiotemporal graph forecasters, benchmarking them against strong statistical, neural, and time-series foundation-model baselines. Across methods, competition-aware augmentation improves stability and error profiles at business-relevant medium and longer horizons, where competitive regimes shift and volatility is most consequential. The results show that broad market-outcome coverage, combined with keyword-derived semantic and geographic priors, provides a scalable way to approximate latent competition and improve CPC forecasting in auction-driven markets.\n\n---\n\n## 33. Team RAS in 10th ABAW Competition: Multimodal Valence and Arousal Estimation Approach\n- **Authors**: Elena Ryumina (1), Maxim Markitantov (1), Alexandr Axyonov (1), Dmitry Ryumin (1), Mikhail Dolgushin (1), Denis Dresvyanskiy (2), Alexey Karpov (1 and 2) ((1) St. Petersburg Federal Research Center of the Russian Academy of Sciences, St. Petersburg, Russia, (2) ITMO University, St. Petersburg, Russia)\n- **URL**: [https://arxiv.org/abs/2603.13056](https://arxiv.org/abs/2603.13056)\n- **Abstract**:\n> Continuous emotion recognition in terms of valence and arousal under in-the-wild (ITW) conditions remains a challenging problem due to large variations in appearance, head pose, illumination, occlusions, and subject-specific patterns of affective expression. We present a multimodal method for valence-arousal estimation ITW. Our method combines three complementary modalities: face, behavior, and audio. The face modality relies on GRADA-based frame-level embeddings and Transformer-based temporal regression. We use Qwen3-VL-4B-Instruct to extract behavior-relevant information from video segments, while Mamba is used to model temporal dynamics across segments. The audio modality relies on WavLM-Large with attention-statistics pooling and includes a cross-modal filtering stage to reduce the influence of unreliable or non-speech segments. To fuse modalities, we explore two fusion strategies: a Directed Cross-Modal Mixture-of-Experts Fusion Strategy that learns interactions between modalities with adaptive weighting, and a Reliability-Aware Audio-Visual Fusion Strategy that combines visual features at the frame-level while using audio as complementary context. The results are reported on the Aff-Wild2 dataset following the 10th Affective Behavior Analysis in-the-Wild (ABAW) challenge protocol. Experiments demonstrate that the proposed multimodal fusion strategy achieves a Concordance Correlation Coefficient (CCC) of 0.658 on the Aff-Wild2 development set.\n\n---\n\n## 34. Are General-Purpose Vision Models All We Need for 2D Medical Image Segmentation? A Cross-Dataset Empirical Study\n- **Authors**: Vanessa Borst , Samuel Kounev\n- **URL**: [https://arxiv.org/abs/2603.13044](https://arxiv.org/abs/2603.13044)\n- **Abstract**:\n> Medical image segmentation (MIS) is a fundamental component of computer-assisted diagnosis and clinical decision support systems. Over the past decade, numerous architectures specifically tailored to medical imaging have emerged to address domain-specific challenges such as low contrast, small anatomical structures, and limited annotated data. In parallel, rapid progress in computer vision has produced highly capable general-purpose vision models (GP-VMs) originally designed for natural images. Despite their strong performance on standard vision benchmarks, their effectiveness for MIS remains insufficiently understood. In this work, we conduct a controlled empirical study to examine whether specialized medical segmentation architectures (SMAs) provide systematic advantages over modern GP-VMs for 2D MIS. We compare eleven SMAs and GP-VMs using a unified training and evaluation protocol. Experiments are performed across three heterogeneous datasets covering different imaging modalities, class structures, and data characteristics. Beyond segmentation accuracy, we analyze qualitative Grad-CAM visualizations to investigate explainability (XAI) behavior. Our results demonstrate that, for the analyzed datasets, GP-VMs out-perform the majority of specialized MIS models. Moreover, XAI analyses indicate that GP-VMs can capture clinically relevant structures without explicit domain-specific architectural design. These findings suggest that GP-VMs can represent a viable alternative to domain-specific methods, highlighting the importance of informed model selection for end-to-end MIS systems. All code and resources are available at GitHub.\n\n---\n\n## 35. Interrogating Design Homogenization in Web Vibe Coding\n- **Authors**: Donghoon Shin , Alice Gao , Rock Yuren Pang , Jaewook Lee , Katharina Reinecke , Emily Tseng\n- **URL**: [https://arxiv.org/abs/2603.13036](https://arxiv.org/abs/2603.13036)\n- **Abstract**:\n> Generative AI is known for its tendency to homogenize, often reproducing dominant style conventions found in training data. However, it remains unclear how these homogenizing effects extend to complex structural tasks like web design. As lay creators increasingly turn to LLMs to 'vibe-code' websites -- prompting for aesthetic and functional goals rather than writing code -- they may inadvertently narrow the diversity of their designs, and limit creative expression throughout the internet. In this paper, we interrogate the possibility of design homogenization in web vibe coding. We first characterize the vibe coding lifecycle, pinpointing stages where homogenization risks may arise. We then conduct a sociotechnical risk analysis unpacking the potential harms of web vibe coding and their interaction with design homogenization. We identify that the push for frictionless generation can exacerbate homogenization and its harms. Finally, we propose a mitigation framework centered on the idea of productive friction. Through case studies at the micro, meso, and macro levels, we show how centering productive friction can empower creators to challenge default outputs and preserve diverse expression in AI-mediated web design.\n\n---\n\n## 36. Purify Once, Edit Freely: Breaking Image Protections under Model Mismatch\n- **Authors**: Qichen Zhao , Shengfang Zhai , Xinjian Bai , Qingni Shen , Qiqi Lin , Yansong Gao , Zhonghai Wu\n- **URL**: [https://arxiv.org/abs/2603.13028](https://arxiv.org/abs/2603.13028)\n- **Abstract**:\n> Diffusion models enable high-fidelity image editing but can also be misused for unauthorized style imitation and harmful content generation. To mitigate these risks, proactive image protection methods embed small, often imperceptible adversarial perturbations into images before sharing to disrupt downstream editing or fine-tuning. However, in realistic post-release scenarios, content owners cannot control downstream processing pipelines, and protections optimized for a surrogate model may fail when attackers use mismatched diffusion pipelines. Existing purification methods can weaken protections but often sacrifice image quality and rarely examine architectural mismatch. We introduce a unified post-release purification framework to evaluate protection survivability under model mismatch. We propose two practical purifiers: VAE-Trans, which corrects protected images via latent-space projection, and EditorClean, which performs instruction-guided reconstruction with a Diffusion Transformer to exploit architectural heterogeneity. Both operate without access to protected images or defense internals. Across 2,100 editing tasks and six representative protection methods, EditorClean consistently restores editability. Compared to protected inputs, it improves PSNR by 3-6 dB and reduces FID by 50-70 percent on downstream edits, while outperforming prior purification baselines by about 2 dB PSNR and 30 percent lower FID. Our results reveal a purify-once, edit-freely failure mode: once purification succeeds, the protective signal is largely removed, enabling unrestricted editing. This highlights the need to evaluate protections under model mismatch and design defenses robust to heterogeneous attackers.\n\n---\n\n## 37. SortScrews: A Dataset and Baseline for Real-time Screw Classification\n- **Authors**: Tianhao Fu , Bingxuan Yang , Juncheng Guo , Shrena Sribalan , Yucheng Chen\n- **URL**: [https://arxiv.org/abs/2603.13027](https://arxiv.org/abs/2603.13027)\n- **Abstract**:\n> Automatic identification of screw types is important for industrial automation, robotics, and inventory management. However, publicly available datasets for screw classification are scarce, particularly for controlled single-object scenarios commonly encountered in automated sorting systems. In this work, we introduce $\\textbf{SortScrews}$, a dataset for casewise visual classification of screws. The dataset contains 560 RGB images at $512\\times512$ resolution covering six screw types and a background class. Images are captured using a standardized acquisition setup and include mild variations in lighting and camera perspective across four capture settings. To facilitate reproducible research and dataset expansion, we also provide a reusable data collection script that allows users to easily construct similar datasets for custom hardware components using inexpensive camera setups. We establish baseline results using transfer learning with EfficientNet-B0 and ResNet-18 classifiers pretrained on ImageNet. In addition, we conduct a well-explored failure analysis. Despite the limited dataset size, these lightweight models achieve strong classification accuracy, demonstrating that controlled acquisition conditions enable effective learning even with relatively small datasets. The dataset, collection pipeline, and baseline training code are publicly available at this https URL .\n\n---\n\n## 38. SAW: Toward a Surgical Action World Model via Controllable and Scalable Video Generation\n- **Authors**: Sampath Rapuri , Lalithkumar Seenivasan , Dominik Schneider , Roger Soberanis-Mukul , Yufan He , Hao Ding , Jiru Xu , Chenhao Yu , Chenyan Jing , Pengfei Guo , Daguang Xu , Mathias Unberath\n- **URL**: [https://arxiv.org/abs/2603.13024](https://arxiv.org/abs/2603.13024)\n- **Abstract**:\n> A surgical world model capable of generating realistic surgical action videos with precise control over tool-tissue interactions can address fundamental challenges in surgical AI and simulation -- from data scarcity and rare event synthesis to bridging the sim-to-real gap for surgical automation. However, current video generation methods, the very core of such surgical world models, require expensive annotations or complex structured intermediates as conditioning signals at inference, limiting their scalability. Other approaches exhibit limited temporal consistency across complex laparoscopic scenes and do not possess sufficient realism. We propose Surgical Action World (SAW) -- a step toward surgical action world modeling through video diffusion conditioned on four lightweight signals: language prompts encoding tool-action context, a reference surgical scene, tissue affordance mask, and 2D tool-tip trajectories. We design a conditional video diffusion approach that reformulates video-to-video diffusion into trajectory-conditioned surgical action synthesis. The backbone diffusion model is fine-tuned on a custom-curated dataset of 12,044 laparoscopic clips with lightweight spatiotemporal conditioning signals, leveraging a depth consistency loss to enforce geometric plausibility without requiring depth at inference. SAW achieves state-of-the-art temporal consistency (CD-FVD: 199.19 vs. 546.82) and strong visual quality on held-out test data. Furthermore, we demonstrate its downstream utility for (a) surgical AI, where augmenting rare actions with SAW-generated videos improves action recognition (clipping F1-score: 20.93% to 43.14%; cutting: 0.00% to 8.33%) on real test data, and (b) surgical simulation, where rendering tool-tissue interaction videos from simulator-derived trajectory points toward a visually faithful simulation engine.\n\n---\n\n## 39. daVinci-Env: Open SWE Environment Synthesis at Scale\n- **Authors**: Dayuan Fu , Shenyu Wu , Yunze Wu , Zerui Peng , Yaxing Huang , Jie Sun , Ji Zeng , Mohan Jiang , Lin Zhang , Yukun Li , Jiarui Hu , Liming Liu , Jinlong Hou , Pengfei Liu\n- **URL**: [https://arxiv.org/abs/2603.13023](https://arxiv.org/abs/2603.13023)\n- **Abstract**:\n> Training capable software engineering (SWE) agents demands large-scale, executable, and verifiable environments that provide dynamic feedback loops for iterative code editing, test execution, and solution refinement. However, existing open-source datasets remain limited in scale and repository diversity, while industrial solutions are opaque with unreleased infrastructure, creating a prohibitive barrier for most academic research groups. We present OpenSWE, the largest fully transparent framework for SWE agent training in Python, comprising 45,320 executable Docker environments spanning over 12.8k repositories, with all Dockerfiles, evaluation scripts, and infrastructure fully open-sourced for reproducibility. OpenSWE is built through a multi-agent synthesis pipeline deployed across a 64-node distributed cluster, automating repository exploration, Dockerfile construction, evaluation script generation, and iterative test analysis. Beyond scale, we propose a quality-centric filtering pipeline that characterizes the inherent difficulty of each environment, filtering out instances that are either unsolvable or insufficiently challenging and retaining only those that maximize learning efficiency. With $891K spent on environment construction and an additional $576K on trajectory sampling and difficulty-aware curation, the entire project represents a total investment of approximately $1.47 million, yielding about 13,000 curated trajectories from roughly 9,000 quality guaranteed environments. Extensive experiments validate OpenSWE's effectiveness: OpenSWE-32B and OpenSWE-72B achieve 62.4% and 66.0% on SWE-bench Verified, establishing SOTA among Qwen2.5 series. Moreover, SWE-focused training yields substantial out-of-domain improvements, including up to 12 points on mathematical reasoning and 5 points on science benchmarks, without degrading factual recall.\n\n---\n\n## 40. ARL-Tangram: Unleash the Resource Efficiency in Agentic Reinforcement Learning\n- **Authors**: Bangjun Xiao , Yihao Zhao , Xiangwei Deng , Shihua Yu , Yuxing Xiang , Huaqiu Liu , Qiying Wang , Liang Zhao , Hailin Zhang , Xuanzhe Liu , Xin Jin , Fuli Luo\n- **URL**: [https://arxiv.org/abs/2603.13019](https://arxiv.org/abs/2603.13019)\n- **Abstract**:\n> Agentic reinforcement learning (RL) has emerged as a transformative workload in cloud clusters, enabling large language models (LLMs) to solve complex problems through interactions with real world. However, unlike traditional RL, agentic RL demands substantial external cloud resources, e.g., CPUs for code execution and GPUs for reward models, that exist outside the primary training cluster. Existing agentic RL framework typically rely on static over-provisioning, i.e., resources are often tied to long-lived trajectories or isolated by tasks, which leads to severe resource inefficiency. We propose the action-level orchestration, and incorporate it into ARL-Tangram, a unified resource management system that enables fine-grained external resource sharing and elasticity. ARL-Tangram utilizes a unified action-level formulation and an elastic scheduling algorithm to minimize action completion time (ACT) while satisfying heterogeneous resource constraints. Further, heterogeneous resource managers are tailored to efficiently support the action-level execution on resources with heterogeneous characteristics and topologies. Evaluation on real-world agentic RL tasks demonstrates that ARL-Tangram improves average ACT by up to 4.3$\\times$, speeds up the step duration of RL training by up to 1.5$\\times$, and saves the external resources by up to 71.2$\\%$. This system has been deployed to support the training of the MiMo series models.\n\n---\n\n## 41. Fair Lung Disease Diagnosis from Chest CT via Gender-Adversarial Attention Multiple Instance Learning\n- **Authors**: Aditya Parikh , Aasa Feragen\n- **URL**: [https://arxiv.org/abs/2603.12988](https://arxiv.org/abs/2603.12988)\n- **Abstract**:\n> We present a fairness-aware framework for multi-class lung disease diagnosis from chest CT volumes, developed for the Fair Disease Diagnosis Challenge at the PHAROS-AIF-MIH Workshop (CVPR 2026). The challenge requires classifying CT scans into four categories -- Healthy, COVID-19, Adenocarcinoma, and Squamous Cell Carcinoma -- with performance measured as the average of per-gender macro F1 scores, explicitly penalizing gender-inequitable predictions. Our approach addresses two core difficulties: the sparse pathological signal across hundreds of slices, and a severe demographic imbalance compounded across disease class and gender. We propose an attention-based Multiple Instance Learning (MIL) model on a ConvNeXt backbone that learns to identify diagnostically relevant slices without slice-level supervision, augmented with a Gradient Reversal Layer (GRL) that adversarially suppresses gender-predictive structure in the learned scan representation. Training incorporates focal loss with label smoothing, stratified cross-validation over joint (class, gender) strata, and targeted oversampling of the most underrepresented subgroup. At inference, all five-fold checkpoints are ensembled with horizontal-flip test-time augmentation via soft logit voting and out-of-the-fold threshold optimization for robustness. Our model achieves a mean validation competition score of 0.685 (std - 0.030), with the best single fold reaching 0.759. All training and inference code is publicly available at this https URL\n\n---\n\n## 42. Is Human Annotation Necessary? Iterative MBR Distillation for Error Span Detection in Machine Translation\n- **Authors**: Boxuan Lyu , Haiyue Song , Zhi Qu\n- **URL**: [https://arxiv.org/abs/2603.12983](https://arxiv.org/abs/2603.12983)\n- **Abstract**:\n> Error Span Detection (ESD) is a crucial subtask in Machine Translation (MT) evaluation, aiming to identify the location and severity of translation errors. While fine-tuning models on human-annotated data improves ESD performance, acquiring such data is expensive and prone to inconsistencies among annotators. To address this, we propose a novel self-evolution framework based on Minimum Bayes Risk (MBR) decoding, named Iterative MBR Distillation for ESD, which eliminates the reliance on human annotations by leveraging an off-the-shelf LLM to generate pseudo-labels. Extensive experiments on the WMT Metrics Shared Task datasets demonstrate that models trained solely on these self-generated pseudo-labels outperform both unadapted base model and supervised baselines trained on human annotations at the system and span levels, while maintaining competitive sentence-level performance.\n\n---\n\n## 43. Efficient Real-World Autonomous Racing via Attenuated Residual Policy Optimization\n- **Authors**: Raphael Trumpp , Denis Hoornaert , Mirco Theile , Marco Caccamo\n- **URL**: [https://arxiv.org/abs/2603.12960](https://arxiv.org/abs/2603.12960)\n- **Abstract**:\n> Residual policy learning (RPL), in which a learned policy refines a static base policy using deep reinforcement learning (DRL), has shown strong performance across various robotic applications. Its effectiveness is particularly evident in autonomous racing, a domain that serves as a challenging benchmark for real-world DRL. However, deploying RPL-based controllers introduces system complexity and increases inference latency. We address this by introducing an extension of RPL named attenuated residual policy optimization ($\\alpha$-RPO). Unlike standard RPL, $\\alpha$-RPO yields a standalone neural policy by progressively attenuating the base policy, which initially serves to bootstrap learning. Furthermore, this mechanism enables a form of privileged learning, where the base policy is permitted to use sensor modalities not required for final deployment. We design $\\alpha$-RPO to integrate seamlessly with PPO, ensuring that the attenuated influence of the base controller is dynamically compensated during policy optimization. We evaluate $\\alpha$-RPO by building a framework for 1:10-scaled autonomous racing around it. In both simulation and zero-shot real-world transfer to Roboracer cars, $\\alpha$-RPO not only reduces system complexity but also improves driving performance compared to baselines - demonstrating its practicality for robotic deployment. Our code is available at: this https URL .\n\n---\n\n## 44. Delta1 with LLM: symbolic and neural integration for credible and explainable reasoning\n- **Authors**: Yang Xu , Jun Liu , Shuwei Chen , Chris Nugent , Hailing Guo\n- **URL**: [https://arxiv.org/abs/2603.12953](https://arxiv.org/abs/2603.12953)\n- **Abstract**:\n> Neuro-symbolic reasoning increasingly demands frameworks that unite the formal rigor of logic with the interpretability of large language models (LLMs). We introduce an end to end explainability by construction pipeline integrating the Automated Theorem Generator Delta1 based on the full triangular standard contradiction (FTSC) with LLMs. Delta1 deterministically constructs minimal unsatisfiable clause sets and complete theorems in polynomial time, ensuring both soundness and minimality by construction. The LLM layer verbalizes each theorem and proof trace into coherent natural language explanations and actionable insights. Empirical studies across health care, compliance, and regulatory domains show that Delta1 and LLM enables interpretable, auditable, and domain aligned reasoning. This work advances the convergence of logic, language, and learning, positioning constructive theorem generation as a principled foundation for neuro-symbolic explainable AI.\n\n---\n\n## 45. Thinking in Streaming Video\n- **Authors**: Zikang Liu , Longteng Guo , Handong Li , Ru Zhen , Xingjian He , Ruyi Ji , Xiaoming Ren , Yanhao Zhang , Haonan Lu , Jing Liu\n- **URL**: [https://arxiv.org/abs/2603.12938](https://arxiv.org/abs/2603.12938)\n- **Abstract**:\n> Real-time understanding of continuous video streams is essential for interactive assistants and multimodal agents operating in dynamic environments. However, most existing video reasoning approaches follow a batch paradigm that defers reasoning until the full video context is observed, resulting in high latency and growing computational cost that are incompatible with streaming scenarios. In this paper, we introduce ThinkStream, a framework for streaming video reasoning based on a Watch--Think--Speak paradigm that enables models to incrementally update their understanding as new video observations arrive. At each step, the model performs a short reasoning update and decides whether sufficient evidence has accumulated to produce a response. To support long-horizon streaming, we propose Reasoning-Compressed Streaming Memory (RCSM), which treats intermediate reasoning traces as compact semantic memory that replaces outdated visual tokens while preserving essential context. We further train the model using a Streaming Reinforcement Learning with Verifiable Rewards scheme that aligns incremental reasoning and response timing with the requirements of streaming interaction. Experiments on multiple streaming video benchmarks show that ThinkStream significantly outperforms existing online video models while maintaining low latency and memory usage. Code, models and data will be released at this https URL\n\n---\n\n## 46. Surprised by Attention: Predictable Query Dynamics for Time Series Anomaly Detection\n- **Authors**: Kadir-Kaan Özer , René Ebeling , Markus Enzweiler\n- **URL**: [https://arxiv.org/abs/2603.12916](https://arxiv.org/abs/2603.12916)\n- **Abstract**:\n> Multivariate time series anomalies often manifest as shifts in cross-channel dependencies rather than simple amplitude excursions. In autonomous driving, for instance, a steering command might be internally consistent but decouple from the resulting lateral acceleration. Residual-based detectors can miss such anomalies when flexible sequence models still reconstruct signals plausibly despite altered coordination. We introduce AxonAD, an unsupervised detector that treats multi-head attention query evolution as a short horizon predictable process. A gradient-updated reconstruction pathway is coupled with a history-only predictor that forecasts future query vectors from past context. This is trained via a masked predictor-target objective against an exponential moving average (EMA) target encoder. At inference, reconstruction error is combined with a tail-aggregated query mismatch score, which measures cosine deviation between predicted and target queries on recent timesteps. This dual approach provides sensitivity to structural dependency shifts while retaining amplitude-level detection. On proprietary in-vehicle telemetry with interval annotations and on the TSB-AD multi-variate suite (17 datasets, 180 series) with threshold-free and range-aware metrics, AxonAD improves ranking quality and temporal localization over strong baselines. Ablations confirm that query prediction and combined scoring are the primary drivers of the observed gains. Code is available at the URL this https URL .\n\n---\n\n## 47. Stake the Points: Structure-Faithful Instance Unlearning\n- **Authors**: Kiseong Hong , JungKyoo Shin , Eunwoo Kim\n- **URL**: [https://arxiv.org/abs/2603.12915](https://arxiv.org/abs/2603.12915)\n- **Abstract**:\n> Machine unlearning (MU) addresses privacy risks in pretrained models. The main goal of MU is to remove the influence of designated data while preserving the utility of retained knowledge. Achieving this goal requires preserving semantic relations among retained instances, which existing studies often overlook. We observe that without such preservation, models suffer from progressive structural collapse, undermining both the deletion-retention balance. In this work, we propose a novel structure-faithful framework that introduces stakes, i.e., semantic anchors that serve as reference points to maintain the knowledge structure. By leveraging these anchors, our framework captures and stabilizes the semantic organization of knowledge. Specifically, we instantiate the anchors from language-driven attribute descriptions encoded by a semantic encoder (e.g., CLIP). We enforce preservation of the knowledge structure via structure-aware alignment and regularization: the former aligns the organization of retained knowledge before and after unlearning around anchors, while the latter regulates updates to structure-critical parameters. Results from image classification, retrieval, and face recognition show average gains of 32.9%, 22.5%, and 19.3% in performance, balancing the deletion-retention trade-off and enhancing generalization.\n\n---\n\n## 48. FedBPrompt: Federated Domain Generalization Person Re-Identification via Body Distribution Aware Visual Prompts\n- **Authors**: Xin Xu , Weilong Li , Wei Liu , Wenke Huang , Zhixi Yu , Bin Yang , Xiaoying Liao , Kui Jiang\n- **URL**: [https://arxiv.org/abs/2603.12912](https://arxiv.org/abs/2603.12912)\n- **Abstract**:\n> Federated Domain Generalization for Person Re-Identification (FedDG-ReID) learns domain-invariant representations from decentralized data. While Vision Transformer (ViT) is widely adopted, its global attention often fails to distinguish pedestrians from high similarity backgrounds or diverse viewpoints -- a challenge amplified by cross-client distribution shifts in FedDG-ReID. To address this, we propose Federated Body Distribution Aware Visual Prompt (FedBPrompt), introducing learnable visual prompts to guide Transformer attention toward pedestrian-centric regions. FedBPrompt employs a Body Distribution Aware Visual Prompts Mechanism (BAPM) comprising: Holistic Full Body Prompts to suppress cross-client background noise, and Body Part Alignment Prompts to capture fine-grained details robust to pose and viewpoint variations. To mitigate high communication costs, we design a Prompt-based Fine-Tuning Strategy (PFTS) that freezes the ViT backbone and updates only lightweight prompts, significantly reducing communication overhead while maintaining adaptability. Extensive experiments demonstrate that BAPM effectively enhances feature discrimination and cross-domain generalization, while PFTS achieves notable performance gains within only a few aggregation rounds. Moreover, both BAPM and PFTS can be easily integrated into existing ViT-based FedDG-ReID frameworks, making FedBPrompt a flexible and effective solution for federated person re-identification. The code is available at this https URL .\n\n---\n\n## 49. Learning from Child-Directed Speech in Two-Language Scenarios: A French-English Case Study\n- **Authors**: Liel Binyamin , Elior Sulem\n- **URL**: [https://arxiv.org/abs/2603.12906](https://arxiv.org/abs/2603.12906)\n- **Abstract**:\n> Research on developmentally plausible language models has largely focused on English, leaving open questions about multilingual settings. We present a systematic study of compact language models by extending BabyBERTa to English-French scenarios under strictly size-matched data conditions, covering monolingual, bilingual, and cross-lingual settings. Our design contrasts two types of training corpora: (i) child-directed speech (about 2.5M tokens), following BabyBERTa and related work, and (ii) multi-domain corpora (about 10M tokens), extending the BabyLM framework to French. To enable fair evaluation, we also introduce new resources, including French versions of QAMR and QASRL, as well as English and French multi-domain corpora. We evaluate the models on both syntactic and semantic tasks and compare them with models trained on Wikipedia-only data. The results reveal context-dependent effects: training on Wikipedia consistently benefits semantic tasks, whereas child-directed speech improves grammatical judgments in monolingual settings. Bilingual pretraining yields notable gains for textual entailment, with particularly strong improvements for French. Importantly, similar patterns emerge across BabyBERTa, RoBERTa, and LTG-BERT, suggesting consistent trends across architectures.\n\n---\n\n## 50. Human-Centered Evaluation of an LLM-Based Process Modeling Copilot: A Mixed-Methods Study with Domain Experts\n- **Authors**: Chantale Lauer , Peter Pfeiffer , Nijat Mehdiyev\n- **URL**: [https://arxiv.org/abs/2603.12895](https://arxiv.org/abs/2603.12895)\n- **Abstract**:\n> Integrating Large Language Models (LLMs) into business process management tools promises to democratize Business Process Model and Notation (BPMN) modeling for non-experts. While automated frameworks assess syntactic and semantic quality, they miss human factors like trust, usability, and professional alignment. We conducted a mixed-methods evaluation of our proposed solution, an LLM-powered BPMN copilot, with five process modeling experts using focus groups and standardized questionnaires. Our findings reveal a critical tension between acceptable perceived usability (mean CUQ score: 67.2/100) and notably lower trust (mean score: 48.8\\%), with reliability rated as the most critical concern (M=1.8/5). Furthermore, we identified output-quality issues, prompting difficulties, and a need for the LLM to ask more in-depth clarifying questions about the process. We envision five use cases ranging from domain-expert support to enterprise quality assurance. We demonstrate the necessity of human-centered evaluation complementing automated benchmarking for LLM modeling agents.\n\n---\n\n## 51. Finite Difference Flow Optimization for RL Post-Training of Text-to-Image Models\n- **Authors**: David McAllister , Miika Aittala , Tero Karras , Janne Hellsten , Angjoo Kanazawa , Timo Aila , Samuli Laine\n- **URL**: [https://arxiv.org/abs/2603.12893](https://arxiv.org/abs/2603.12893)\n- **Abstract**:\n> Reinforcement learning (RL) has become a standard technique for post-training diffusion-based image synthesis models, as it enables learning from reward signals to explicitly improve desirable aspects such as image quality and prompt alignment. In this paper, we propose an online RL variant that reduces the variance in the model updates by sampling paired trajectories and pulling the flow velocity in the direction of the more favorable image. Unlike existing methods that treat each sampling step as a separate policy action, we consider the entire sampling process as a single action. We experiment with both high-quality vision language models and off-the-shelf quality metrics for rewards, and evaluate the outputs using a broad set of metrics. Our method converges faster and yields higher output quality and prompt alignment than previous approaches.\n\n---\n\n## 52. Team LEYA in 10th ABAW Competition: Multimodal Ambivalence/Hesitancy Recognition Approach\n- **Authors**: Elena Ryumina (1), Alexandr Axyonov (1), Dmitry Sysoev (2), Timur Abdulkadirov (2), Kirill Almetov (2), Yulia Morozova (2), Dmitry Ryumin (1 and 2) ((1) St. Petersburg Federal Research Center of the Russian Academy of Sciences, St. Petersburg, Russia, (2) HSE University, St. Petersburg, Russia)\n- **URL**: [https://arxiv.org/abs/2603.12848](https://arxiv.org/abs/2603.12848)\n- **Abstract**:\n> Ambivalence/hesitancy recognition in unconstrained videos is a challenging problem due to the subtle, multimodal, and context-dependent nature of this behavioral state. In this paper, a multimodal approach for video-level ambivalence/hesitancy recognition is presented for the 10th ABAW Competition. The proposed approach integrates four complementary modalities: scene, face, audio, and text. Scene dynamics are captured with a VideoMAE-based model, facial information is encoded through emotional frame-level embeddings aggregated by statistical pooling, acoustic representations are extracted with EmotionWav2Vec2.0 and processed by a Mamba-based temporal encoder, and linguistic cues are modeled using fine-tuned transformer-based text models. The resulting unimodal embeddings are further combined using multimodal fusion models, including prototype-augmented variants. Experiments on the BAH corpus demonstrate clear gains of multimodal fusion over all unimodal baselines. The best unimodal configuration achieved an average MF1 of 70.02%, whereas the best multimodal fusion model reached 83.25%. The highest final test performance, 71.43%, was obtained by an ensemble of five prototype-augmented fusion models. The obtained results highlight the importance of complementary multimodal cues and robust fusion strategies for ambivalence/hesitancy recognition.\n\n---\n\n## 53. Hierarchical Reference Sets for Robust Unsupervised Detection of Scattered and Clustered Outliers\n- **Authors**: Yiqun Zhang , Zexi Tan , Xiaopeng Luo , Yunlin Liu\n- **URL**: [https://arxiv.org/abs/2603.12847](https://arxiv.org/abs/2603.12847)\n- **Abstract**:\n> Most real-world IoT data analysis tasks, such as clustering and anomaly event detection, are unsupervised and highly susceptible to the presence of outliers. In addition to sporadic scattered outliers caused by factors such as faulty sensor readings, IoT systems often exhibit clustered outliers. These occur when multiple devices or nodes produce similar anomalous measurements, for instance, owing to localized interference, emerging security threats, or regional false alarms, forming micro-clusters. These clustered outliers can be easily mistaken for normal behavior because of their relatively high local density, thereby obscuring the detection of both scattered and contextual anomalies. To address this, we propose a novel outlier detection paradigm that leverages the natural neighboring relationships using graph structures. This facilitates multi-perspective anomaly evaluation by incorporating reference sets at both local and global scales derived from the graph. Our approach enables the effective recognition of scattered outliers without interference from clustered anomalies, whereas the graph structure simultaneously helps reflect and isolate clustered outlier groups. Extensive experiments, including comparative performance analysis, ablation studies, validation on downstream clustering tasks, and evaluation of hyperparameter sensitivity, demonstrate the efficacy of the proposed method. The source code is available at this https URL .\n\n---\n\n## 54. DAST: A Dual-Stream Voice Anonymization Attacker with Staged Training\n- **Authors**: Ridwan Arefeen , Xiaoxiao Miao , Rong Tong , Aik Beng Ng , Simon See , Timothy Liu\n- **URL**: [https://arxiv.org/abs/2603.12840](https://arxiv.org/abs/2603.12840)\n- **Abstract**:\n> Voice anonymization masks vocal traits while preserving linguistic content, which may still leak speaker-specific patterns. To assess and strengthen privacy evaluation, we propose a dual-stream attacker that fuses spectral and self-supervised learning features via parallel encoders with a three-stage training strategy. Stage I establishes foundational speaker-discriminative representations. Stage II leverages the shared identity-transformation characteristics of voice conversion and anonymization, exposing the model to diverse converted speech to build cross-system robustness. Stage III provides lightweight adaptation to target anonymized data. Results on the VoicePrivacy Attacker Challenge (VPAC) dataset demonstrate that Stage II is the primary driver of generalization, enabling strong attacking performance on unseen anonymization datasets. With Stage III, fine-tuning on only 10\\% of the target anonymization dataset surpasses current state-of-the-art attackers in terms of EER.\n\n---\n\n## 55. Mask2Flow-TSE: Two-Stage Target Speaker Extraction with Masking and Flow Matching\n- **Authors**: Junwon Moon , Hyunjin Choi , Hansol Park , Heeseung Kim , Kyuhong Shim\n- **URL**: [https://arxiv.org/abs/2603.12837](https://arxiv.org/abs/2603.12837)\n- **Abstract**:\n> Target speaker extraction (TSE) extracts the target speaker's voice from overlapping speech mixtures given a reference utterance. Existing approaches typically fall into two categories: discriminative and generative. Discriminative methods apply time-frequency masking for fast inference but often over-suppress the target signal, while generative methods synthesize high-quality speech at the cost of numerous iterative steps. We propose Mask2Flow-TSE, a two-stage framework combining the strengths of both paradigms. The first stage applies discriminative masking for coarse separation, and the second stage employs flow matching to refine the output toward target speech. Unlike generative approaches that synthesize speech from Gaussian noise, our method starts from the masked spectrogram, enabling high-quality reconstruction in a single inference step. Experiments show that Mask2Flow-TSE achieves comparable performance to existing generative TSE methods with approximately 85M parameters.\n\n---\n\n## 56. Hierarchical Dual-Change Collaborative Learning for UAV Scene Change Captioning\n- **Authors**: Fuhai Chen , Pengpeng Huang , Junwen Wu , Hehong Zhang , Shiping Wang , Xiaoguang Ma , Xuri Ge\n- **URL**: [https://arxiv.org/abs/2603.12832](https://arxiv.org/abs/2603.12832)\n- **Abstract**:\n> This paper proposes a novel task for UAV scene understanding - UAV Scene Change Captioning (UAV-SCC) - which aims to generate natural language descriptions of semantic changes in dynamic aerial imagery captured from a movable viewpoint. Unlike traditional change captioning that mainly describes differences between image pairs captured from a fixed camera viewpoint over time, UAV scene change captioning focuses on image-pair differences resulting from both temporal and spatial scene variations dynamically captured by a moving camera. The key challenge lies in understanding viewpoint-induced scene changes from UAV image pairs that share only partially overlapping scene content due to viewpoint shifts caused by camera rotation, while effectively exploiting the relative orientation between the two images. To this end, we propose a Hierarchical Dual-Change Collaborative Learning (HDC-CL) method for UAV scene change captioning. In particular, a novel transformer, \\emph{i.e.} Dynamic Adaptive Layout Transformer (DALT) is designed to adaptively model diverse spatial layouts of the image pair, where the interrelated features derived from the overlapping and non-overlapping regions are learned within the flexible and unified encoding layer. Furthermore, we propose a Hierarchical Cross-modal Orientation Consistency Calibration (HCM-OCC) method to enhance the model's sensitivity to viewpoint shift directions, enabling more accurate change captioning. To facilitate in-depth research on this task, we construct a new benchmark dataset, named UAV-SCC dataset, for UAV scene change captioning. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance on this task. The dataset and code will be publicly released upon acceptance of this paper.\n\n---\n\n## 57. Residual SODAP: Residual Self-Organizing Domain-Adaptive Prompting with Structural Knowledge Preservation for Continual Learning\n- **Authors**: Gyutae Oh , Jungwoo Bae , Jitae Shin\n- **URL**: [https://arxiv.org/abs/2603.12816](https://arxiv.org/abs/2603.12816)\n- **Abstract**:\n> Continual learning (CL) suffers from catastrophic forgetting, which is exacerbated in domain-incremental learning (DIL) where task identifiers are unavailable and storing past data is infeasible. While prompt-based CL (PCL) adapts representations with a frozen backbone, we observe that prompt-only improvements are often insufficient due to suboptimal prompt selection and classifier-level instability under domain shifts. We propose Residual SODAP, which jointly performs prompt-based representation adaptation and classifier-level knowledge preservation. Our framework combines $\\alpha$-entmax sparse prompt selection with residual aggregation, data-free distillation with pseudo-feature replay, prompt-usage--based drift detection, and uncertainty-aware multi-loss balancing. Across three DIL benchmarks without task IDs or extra data storage, Residual SODAP achieves state-of-the-art AvgACC/AvgF of 0.850/0.047 (DR), 0.760/0.031 (Skin Cancer), and 0.995/0.003 (CORe50).\n\n---\n\n## 58. Cheers: Decoupling Patch Details from Semantic Representations Enables Unified Multimodal Comprehension and Generation\n- **Authors**: Yichen Zhang , Da Peng , Zonghao Guo , Zijian Zhang , Xuesong Yang , Tong Sun , Shichu Sun , Yidan Zhang , Yanghao Li , Haiyan Zhao , Wang Xu , Qi Shi , Yangang Sun , Chi Chen , Shuo Wang , Yukun Yan , Xu Han , Qiang Ma , Wei Ke , Liang Wang , Zhiyuan Liu , Maosong Sun\n- **URL**: [https://arxiv.org/abs/2603.12793](https://arxiv.org/abs/2603.12793)\n- **Abstract**:\n> A recent cutting-edge topic in multimodal modeling is to unify visual comprehension and generation within a single model. However, the two tasks demand mismatched decoding regimes and visual representations, making it non-trivial to jointly optimize within a shared feature space. In this work, we present Cheers, a unified multimodal model that decouples patch-level details from semantic representations, thereby stabilizing semantics for multimodal understanding and improving fidelity for image generation via gated detail residuals. Cheers includes three key components: (i) a unified vision tokenizer that encodes and compresses image latent states into semantic tokens for efficient LLM conditioning, (ii) an LLM-based Transformer that unifies autoregressive decoding for text generation and diffusion decoding for image generation, and (iii) a cascaded flow matching head that decodes visual semantics first and then injects semantically gated detail residuals from the vision tokenizer to refine high-frequency content. Experiments on popular benchmarks demonstrate that Cheers matches or surpasses advanced UMMs in both visual understanding and generation. Cheers also achieves 4x token compression, enabling more efficient high-resolution image encoding and generation. Notably, Cheers outperforms the Tar-1.5B on the popular benchmarks GenEval and MMBench, while requiring only 20% of the training cost, indicating effective and efficient (i.e., 4x token compression) unified multimodal modeling. We will release all code and data for future research.\n\n---\n\n## 59. The RIGID Framework: Research-Integrated, Generative AI-Mediated Instructional Design\n- **Authors**: Yerin Kwak , Zachary A. Pardos\n- **URL**: [https://arxiv.org/abs/2603.12781](https://arxiv.org/abs/2603.12781)\n- **Abstract**:\n> Instructional Design (ID) often faces challenges in incorporating research-based knowledge and pedagogical best practices. Although educational researchers and government agencies emphasize grounding ID in evidence, integrating research findings into everyday design workflows is often complex, as it requires considering multiple context-specific demands and constraints. To address this persistent gap, this paper explores how research in the learning sciences (LS) can be systematically integrated across ID workflows and how recent advances in generative AI can help operationalize this integration. While ID and LS share a commitment to improving learning experiences through design-oriented approaches in authentic contexts, structured integration between the two fields remains limited, leaving their complementary insights underutilized. We present RIGID (Research-Integrated, Generative AI-Mediated Instructional Design), a unified framework that integrates LS research across ID workflows spanning analysis, design, implementation, and evaluation phases, while leveraging generative AI to mediate this integration at each stage. The RIGID framework provides a systematic approach for enabling research-integrated instructional design that is both operational and context-sensitive, while preserving the central role of human expertise.\n\n---\n\n## 60. Empowering Semantic-Sensitive Underwater Image Enhancement with VLM\n- **Authors**: Guodong Fan , Shengning Zhou , Genji Yuan , Huiyu Li , Jingchun Zhou , Jinjiang Li\n- **URL**: [https://arxiv.org/abs/2603.12773](https://arxiv.org/abs/2603.12773)\n- **Abstract**:\n> In recent years, learning-based underwater image enhancement (UIE) techniques have rapidly evolved. However, distribution shifts between high-quality enhanced outputs and natural images can hinder semantic cue extraction for downstream vision tasks, thereby limiting the adaptability of existing enhancement models. To address this challenge, this work proposes a new learning mechanism that leverages Vision-Language Models (VLMs) to empower UIE models with semantic-sensitive capabilities. To be concrete, our strategy first generates textual descriptions of key objects from a degraded image via VLMs. Subsequently, a text-image alignment model remaps these relevant descriptions back onto the image to produce a spatial semantic guidance map. This map then steers the UIE network through a dual-guidance mechanism, which combines cross-attention and an explicit alignment loss. This forces the network to focus its restorative power on semantic-sensitive regions during image reconstruction, rather than pursuing a globally uniform improvement, thereby ensuring the faithful restoration of key object features. Experiments confirm that when our strategy is applied to different UIE baselines, significantly boosts their performance on perceptual quality metrics as well as enhances their performance on detection and segmentation tasks, validating its effectiveness and adaptability.\n\n---\n\n## 61. FC-Track: Overlap-Aware Post-Association Correction for Online Multi-Object Tracking\n- **Authors**: Cheng Ju , Zejing Zhao , Akio Namiki\n- **URL**: [https://arxiv.org/abs/2603.12758](https://arxiv.org/abs/2603.12758)\n- **Abstract**:\n> Reliable multi-object tracking (MOT) is essential for robotic systems operating in complex and dynamic environments. Despite recent advances in detection and association, online MOT methods remain vulnerable to identity switches caused by frequent occlusions and object overlap, where incorrect associations can propagate over time and degrade tracking reliability. We present a lightweight post-association correction framework (FC-Track) for online MOT that explicitly targets overlap-induced mismatches during inference. The proposed method suppresses unreliable appearance updates under high-overlap conditions using an Intersection over Area (IoA)-based filtering strategy, and locally corrects detection-to-tracklet mismatches through appearance similarity comparison within overlapped tracklet pairs. By preventing short-term mismatches from propagating, our framework effectively mitigates long-term identity switches without resorting to global optimization or re-identification. The framework operates online without global optimization or re-identification, making it suitable for real-time robotic applications. We achieve 81.73 MOTA, 82.81 IDF1, and 66.95 HOTA on the MOT17 test set with a running speed of 5.7 FPS, and 77.52 MOTA, 80.90 IDF1, and 65.67 HOTA on the MOT20 test set with a running speed of 0.6 FPS. Specifically, our framework FC-Track produces only 29.55% long-term identity switches, which is substantially lower than existing online trackers. Meanwhile, our framework maintains state-of-the-art performance on the MOT20 benchmark.\n\n---\n\n## 62. TaoBench: Do Automated Theorem Prover LLMs Generalize Beyond MathLib?\n- **Authors**: Alexander K Taylor , Junyi Zhang , Ethan Ji , Vigyan Sahai , Haikang Deng , Yuanzhou Chen , Yifan Yuan , Di Wu , Jia-Chen Gu , Kai-Wei Chang , Nanyun Peng , Amit Sahai , Wei Wang\n- **URL**: [https://arxiv.org/abs/2603.12744](https://arxiv.org/abs/2603.12744)\n- **Abstract**:\n> Automated theorem proving (ATP) benchmarks largely consist of problems formalized in MathLib, so current ATP training and evaluation are heavily biased toward MathLib's definitional framework. However, frontier mathematics is often exploratory and prototype-heavy, relying on bespoke constructions that deviate from standard libraries. In this work, we evaluate the robustness of current ATP systems when applied to a novel definitional framework, specifically examining the performance gap between standard library problems and bespoke mathematical constructions. We introduce TaoBench, an undergraduate-level benchmark derived from Terence Tao's Analysis I, which formalizes analysis by constructing core mathematical concepts from scratch, without relying on standard Mathlib definitions, as well as by mixing from-scratch and MathLib constructions. For fair evaluation, we build an agentic pipeline that automatically extracts a compilable, self-contained local environment for each problem. To isolate the effect of definitional frameworks, we additionally translate every problem into a mathematically equivalent Mathlib formulation, yielding paired TaoBench-Mathlib statements for direct comparison. While state-of-the-art ATP models perform capably within the MathLib framework, performance drops by an average of roughly 26% on the definitionally equivalent Tao formulation. This indicates that the main bottleneck is limited generalization across definitional frameworks rather than task difficulty. TaoBench thus highlights a gap between benchmark performance and applicability, and provides a concrete foundation for developing and testing provers better aligned with research mathematics.\n\n---\n\n## 63. MoKus: Leveraging Cross-Modal Knowledge Transfer for Knowledge-Aware Concept Customization\n- **Authors**: Chenyang Zhu , Hongxiang Li , Xiu Li , Long Chen\n- **URL**: [https://arxiv.org/abs/2603.12743](https://arxiv.org/abs/2603.12743)\n- **Abstract**:\n> Concept customization typically binds rare tokens to a target concept. Unfortunately, these approaches often suffer from unstable performance as the pretraining data seldom contains these rare tokens. Meanwhile, these rare tokens fail to convey the inherent knowledge of the target concept. Consequently, we introduce Knowledge-aware Concept Customization, a novel task aiming at binding diverse textual knowledge to target visual concepts. This task requires the model to identify the knowledge within the text prompt to perform high-fidelity customized generation. Meanwhile, the model should efficiently bind all the textual knowledge to the target concept. Therefore, we propose MoKus, a novel framework for knowledge-aware concept customization. Our framework relies on a key observation: cross-modal knowledge transfer, where modifying knowledge within the text modality naturally transfers to the visual modality during generation. Inspired by this observation, MoKus contains two stages: (1) In visual concept learning, we first learn the anchor representation to store the visual information of the target concept. (2) In textual knowledge updating, we update the answer for the knowledge queries to the anchor representation, enabling high-fidelity customized generation. To further comprehensively evaluate our proposed MoKus on the new task, we introduce the first benchmark for knowledge-aware concept customization: KnowCusBench. Extensive evaluations have demonstrated that MoKus outperforms state-of-the-art methods. Moreover, the cross-model knowledge transfer allows MoKus to be easily extended to other knowledge-aware applications like virtual concept creation and concept erasure. We also demonstrate the capability of our method to achieve improvements on world knowledge benchmarks.\n\n---\n\n## 64. SRAM-Based Compute-in-Memory Accelerator for Linear-decay Spiking Neural Networks\n- **Authors**: Hongyang Shang , Shuai Dong , Yahan Yang , Junyi Yang , Peng Zhou , Arindam Basu\n- **URL**: [https://arxiv.org/abs/2603.12739](https://arxiv.org/abs/2603.12739)\n- **Abstract**:\n> Spiking Neural Networks (SNNs) have emerged as a biologically inspired alternative to conventional deep networks, offering event-driven and energy-efficient computation. However, their throughput remains constrained by the serial update of neuron membrane states. While many hardware accelerators and Compute-in-Memory (CIM) architectures efficiently parallelize the synaptic operation (W x I) achieving O(1) complexity for matrix-vector multiplication, the subsequent state update step still requires O(N) time to refresh all neuron membrane potentials. This mismatch makes state update the dominant latency and energy bottleneck in SNN inference. To address this challenge, we propose an SRAM-based CIM for SNN with Linear Decay Leaky Integrate-and-Fire (LD-LIF) Neuron that co-optimizes algorithm and hardware. At the algorithmic level, we replace the conventional exponential membrane decay with a linear decay approximation, converting costly multiplications into simple additions while accuracy drops only around 1%. At the architectural level, we introduce an in-memory parallel update scheme that performs in-place decay directly within the SRAM array, eliminating the need for global sequential updates. Evaluated on benchmark SNN workloads, the proposed method achieves a 1.1 x to 16.7 x reduction of SOP energy consumption, while providing 15.9 x to 69 x more energy efficiency, with negligible accuracy loss relative to original decay models. This work highlights that beyond accelerating the (W x I) computation, optimizing state-update dynamics within CIM architectures is essential for scalable, low-power, and real-time neuromorphic processing.\n\n---\n\n## 65. Graph In-Context Operator Networks for Generalizable Spatiotemporal Prediction\n- **Authors**: Chenghan Wu , Zongmin Yu , Boai Sun , Liu Yang\n- **URL**: [https://arxiv.org/abs/2603.12725](https://arxiv.org/abs/2603.12725)\n- **Abstract**:\n> In-context operator learning enables neural networks to infer solution operators from contextual examples without weight updates. While prior work has demonstrated the effectiveness of this paradigm in leveraging vast datasets, a systematic comparison against single-operator learning using identical training data has been absent. We address this gap through controlled experiments comparing in-context operator learning against classical operator learning (single-operator models trained without contextual examples), under the same training steps and dataset. To enable this investigation on real-world spatiotemporal systems, we propose GICON (Graph In-Context Operator Network), combining graph message passing for geometric generalization with example-aware positional encoding for cardinality generalization. Experiments on air quality prediction across two Chinese regions show that in-context operator learning outperforms classical operator learning on complex tasks, generalizing across spatial domains and scaling robustly from few training examples to 100 at inference.\n\n---\n\n## 66. CognitionCapturerPro: Towards High-Fidelity Visual Decoding from EEG/MEG via Multi-modal Information and Asymmetric Alignment\n- **Authors**: Kaifan Zhang , Lihuo He , Junjie Ke , Yuqi Ji , Lukun Wu , Lizi Wang , Xinbo Gao\n- **URL**: [https://arxiv.org/abs/2603.12722](https://arxiv.org/abs/2603.12722)\n- **Abstract**:\n> Visual stimuli reconstruction from EEG remains challenging due to fidelity loss and representation shift. We propose CognitionCapturerPro, an enhanced framework that integrates EEG with multi-modal priors (images, text, depth, and edges) via collaborative training. Our core contributions include an uncertainty-weighted similarity scoring mechanism to quantify modality-specific fidelity and a fusion encoder for integrating shared representations. By employing a simplified alignment module and a pre-trained diffusion model, our method significantly outperforms the original CognitionCapturer on the THINGS-EEG dataset, improving Top-1 and Top-5 retrieval accuracy by 25.9% and 10.6%, respectively. Code is available at: this https URL .\n\n---\n\n## 67. CMHANet: A Cross-Modal Hybrid Attention Network for Point Cloud Registration\n- **Authors**: Dongxu Zhang , Yingsen Wang , Yiding Sun , Haoran Xu , Peilin Fan , Jihua Zhu\n- **URL**: [https://arxiv.org/abs/2603.12721](https://arxiv.org/abs/2603.12721)\n- **Abstract**:\n> Robust point cloud registration is a fundamental task in 3D computer vision and geometric deep learning, essential for applications such as large-scale 3D reconstruction, augmented reality, and scene understanding. However, the performance of established learning-based methods often degrades in complex, real world scenarios characterized by incomplete data, sensor noise, and low overlap regions. To address these limitations, we propose CMHANet, a novel Cross-Modal Hybrid Attention Network. Our method integrates the fusion of rich contextual information from 2D images with the geometric detail of 3D point clouds, yielding a comprehensive and resilient feature representation. Furthermore, we introduce an innovative optimization function based on contrastive learning, which enforces geometric consistency and significantly improves the model's robustness to noise and partial observations. We evaluated CMHANet on the 3DMatch and the challenging 3DLoMatch datasets. \\rev{Additionally, zero-shot evaluations on the TUM RGB-D SLAM dataset verify the model's generalization capability to unseen domains.} The experimental results demonstrate that our method achieves substantial improvements in both registration accuracy and overall robustness, outperforming current techniques. We also release our code in \\href{ this https URL }{ this https URL }.\n\n---\n\n## 68. IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registration\n- **Authors**: Dongxu Zhang , Jihua Zhu , Shiqi Li , Wenbiao Yan , Haoran Xu , Peilin Fan , Huimin Lu\n- **URL**: [https://arxiv.org/abs/2603.12719](https://arxiv.org/abs/2603.12719)\n- **Abstract**:\n> Point cloud registration (PCR) is a fundamental task in 3D vision and provides essential support for applications such as autonomous driving, robotics, and environmental modeling. Despite its widespread use, existing methods often fail when facing real-world challenges like heavy noise, significant occlusions, and large-scale transformations. These limitations frequently result in compromised registration accuracy and insufficient robustness in complex environments. In this paper, we propose IGASA as a novel registration framework constructed upon a Hierarchical Pyramid Architecture (HPA) designed for robust multi-scale feature extraction and fusion. The framework integrates two pivotal components consisting of the Hierarchical Cross-Layer Attention (HCLA) module and the Iterative Geometry-Aware Refinement (IGAR) module. The HCLA module utilizes skip attention mechanisms to align multi-resolution features and enhance local geometric consistency. Simultaneously, the IGAR module is designed for the fine matching phase by leveraging reliable correspondences established during coarse matching. This synergistic integration within the architecture allows IGASA to adapt effectively to diverse point cloud structures and intricate transformations. We evaluate the performance of IGASA on four widely recognized benchmark datasets including 3D(Lo)Match, KITTI, and nuScenes. Our extensive experiments consistently demonstrate that IGASA significantly surpasses state-of-the-art methods and achieves notable improvements in registration accuracy. This work provides a robust foundation for advancing point cloud registration techniques while offering valuable insights for practical 3D vision applications. The code for IGASA is available in \\href{ this https URL }{ this https URL }.\n\n---\n\n## 69. Altered Thoughts, Altered Actions: Probing Chain-of-Thought Vulnerabilities in VLA Robotic Manipulation\n- **Authors**: Tuan Duong Trinh , Naveed Akhtar , Basim Azam\n- **URL**: [https://arxiv.org/abs/2603.12717](https://arxiv.org/abs/2603.12717)\n- **Abstract**:\n> Recent Vision-Language-Action (VLA) models increasingly adopt chain-of-thought (CoT) reasoning, generating a natural-language plan before decoding motor commands. This internal text channel between the reasoning module and the action decoder has received no adversarial scrutiny. We ask: which properties of this intermediate plan does the action decoder actually rely on, and can targeted corruption of the reasoning trace alone -- with all inputs left intact -- degrade a robot's physical task performance? We design a taxonomy of seven text corruptions organized into three attacker tiers (blind noise, mechanical-semantic, and LLM-adaptive) and apply them to a state-of-the-art reasoning VLA across 40 LIBERO tabletop manipulation tasks. Our results reveal a striking asymmetry: substituting object names in the reasoning trace reduces overall success rate by 8.3~percentage points (pp) -- reaching $-$19.3~pp on goal-conditioned tasks and $-$45~pp on individual tasks -- whereas sentence reordering, spatial-direction reversal, token noise, and even a 70B-parameter LLM crafting plausible-but-wrong plans all have negligible impact (within $\\pm$4~pp). This asymmetry indicates that the action decoder depends on entity-reference integrity rather than reasoning quality or sequential structure. Notably, a sophisticated LLM-based attacker underperforms simple mechanical object-name substitution, because preserving plausibility inadvertently retains the entity-grounding structure the decoder needs. A cross-architecture control using a non-reasoning VLA confirms the vulnerability is exclusive to reasoning-augmented models, while instruction-level attacks degrade both architectures -- establishing that the internal reasoning trace is a distinct and stealthy threat vector invisible to input-validation defenses.\n\n---\n\n## 70. Cost-Efficient Multimodal LLM Inference via Cross-Tier GPU Heterogeneity\n- **Authors**: Donglin Yu\n- **URL**: [https://arxiv.org/abs/2603.12707](https://arxiv.org/abs/2603.12707)\n- **Abstract**:\n> Multimodal large language model (MLLM) inference splits into two phases with opposing hardware demands: vision encoding is compute-bound, while language generation is memory-bandwidth-bound. We show that under standard transformer KV caching, the modality boundary (between vision encoder and language model) minimizes cross-device transfer among all partition points that preserve standard stage-based execution. Partitioning here reduces transfer complexity from $O(L * s_ctx)$ bytes (GB-scale KV caches under stage-level disaggregation) to $O(N_v * d)$ bytes (MB-scale embeddings), an O(L) reduction where L is the transformer depth. The result holds across attention mechanisms (MHA/GQA), dynamic vision resolutions, and model scales, and the advantage grows as models deepen. A direct implication is that existing stage-level disaggregation systems are constrained to high-bandwidth interconnects (e.g., NVLink), whereas modality-level disaggregation enables cross-tier heterogeneous serving over commodity PCIe. A closed-form cost model shows that heterogeneous deployment is cost-optimal under phase-separable workloads (predicts 31.4% savings; observed 40.6%). We build HeteroServe, a phase-aware runtime with modality-level partitioning and cross-tier scheduling, and evaluate it on LLaVA-1.5-7B and Qwen2.5-VL against vLLM v0.3.0. On identical 4xA100 hardware, engine optimizations raise throughput by up to 54%. Under a fixed budget, a heterogeneous cluster (\\$38k) improves Tokens/\\$ by 37% over a homogeneous baseline (\\$64k) without degrading latency.\n\n---\n\n## 71. Seeing Eye to Eye: Enabling Cognitive Alignment Through Shared First-Person Perspective in Human-AI Collaboration\n- **Authors**: Zhuyu Teng , Pei Chen , Yichen Cai , Ruoqing Lu , Zhaoqu Jiang , Jiayang Li , Weitao You , Lingyun Sun\n- **URL**: [https://arxiv.org/abs/2603.12701](https://arxiv.org/abs/2603.12701)\n- **Abstract**:\n> Despite advances in multimodal AI, current vision-based assistants often remain inefficient in collaborative tasks. We identify two key gulfs: a communication gulf, where users must translate rich parallel intentions into verbal commands due to the channel mismatch , and an understanding gulf, where AI struggles to interpret subtle embodied cues. To address these, we propose Eye2Eye, a framework that leverages first-person perspective as a channel for human-AI cognitive alignment. It integrates three components: (1) joint attention coordination for fluid focus alignment, (2) revisable memory to maintain evolving common ground, and (3) reflective feedback allowing users to clarify and refine AI's understanding. We implement this framework in an AR prototype and evaluate it through a user study and a post-hoc pipeline evaluation. Results show that Eye2Eye significantly reduces task completion time and interaction load while increasing trust, demonstrating its components work in concert to improve collaboration.\n\n---\n\n## 72. HSEmotion Team at ABAW-10 Competition: Facial Expression Recognition, Valence-Arousal Estimation, Action Unit Detection and Fine-Grained Violence Classification\n- **Authors**: Andrey V. Savchenko , Kseniia Tsypliakova\n- **URL**: [https://arxiv.org/abs/2603.12693](https://arxiv.org/abs/2603.12693)\n- **Abstract**:\n> This article presents our results for the 10th Affective Behavior Analysis in-the-Wild (ABAW) competition. For frame-wise facial emotion understanding tasks (frame-wise facial expression recognition, valence-arousal estimation, action unit detection), we propose a fast approach based on facial embedding extraction with pre-trained EfficientNet-based emotion recognition models. If the latter model's confidence exceeds a threshold, its prediction is used. Otherwise, we feed embeddings into a simple multi-layered perceptron trained on the AffWild2 dataset. Estimated class-level scores are smoothed in a sliding window of fixed size to mitigate noise in frame-wise predictions. For the fine-grained violence detection task, we examine several pre-trained architectures for frame embeddings and their aggregation for video classification. Experimental results on four tasks from the ABAW challenge demonstrate that our approach significantly improves validation metrics over existing baselines.\n\n---\n\n## 73. Federated Hierarchical Clustering with Automatic Selection of Optimal Cluster Numbers\n- **Authors**: Yue Zhang , Chuanlong Qiu , Xinfa Liao , Yiqun Zhang\n- **URL**: [https://arxiv.org/abs/2603.12684](https://arxiv.org/abs/2603.12684)\n- **Abstract**:\n> Federated Clustering (FC) is an emerging and promising solution in exploring data distribution patterns from distributed and privacy-protected data in an unsupervised manner. Existing FC methods implicitly rely on the assumption that clients are with a known number of uniformly sized clusters. However, the true number of clusters is typically unknown, and cluster sizes are naturally imbalanced in real scenarios. Furthermore, the privacy-preserving transmission constraints in federated learning inevitably reduce usable information, making the development of robust and accurate FC extremely challenging. Accordingly, we propose a novel FC framework named Fed-$k^*$-HC, which can automatically determine an optimal number of clusters $k^*$ based on the data distribution explored through hierarchical clustering. To obtain the global data distribution for $k^*$ determination, we let each client generate micro-subclusters. Their prototypes are then uploaded to the server for hierarchical merging. The density-based merging design allows exploring clusters of varying sizes and shapes, and the progressive merging process can self-terminate according to the neighboring relationships among the prototypes to determine $k^*$. Extensive experiments on diverse datasets demonstrate the FC capability of the proposed Fed-$k^*$-HC in accurately exploring a proper number of clusters.\n\n---\n\n## 74. Experimental evidence of progressive ChatGPT models self-convergence\n- **Authors**: Konstantinos F. Xylogiannopoulos , Petros Xanthopoulos , Panagiotis Karampelas , Georgios A. Bakamitsos\n- **URL**: [https://arxiv.org/abs/2603.12683](https://arxiv.org/abs/2603.12683)\n- **Abstract**:\n> Large Language Models (LLMs) that undergo recursive training on synthetically generated data are susceptible to model collapse, a phenomenon marked by the generation of meaningless output. Existing research has examined this issue from either theoretical or empirical perspectives, often focusing on a single model trained recursively on its own outputs. While prior studies have cautioned against the potential degradation of LLM output quality under such conditions, no longitudinal investigation has yet been conducted to assess this effect over time. In this study, we employ a text similarity metric to evaluate different ChatGPT models' capacity to generate diverse textual outputs. Our findings indicate a measurable decline of recent ChatGPT releases' ability to produce varied text, even when explicitly prompted to do so, by setting the temperature parameter to one. The observed reduction in output diversity may be attributed to the influence of the amounts of synthetic data incorporated within their training datasets as the result of internet infiltration by LLM generated data. The phenomenon is defined as model self-convergence because of the gradual increase of similarities of produced texts among different ChatGPT versions.\n\n---\n\n## 75. MetaKE: Meta-learning Aligned Knowledge Editing via Bi-level Optimization\n- **Authors**: Shuxin Liu , Ou Wu\n- **URL**: [https://arxiv.org/abs/2603.12677](https://arxiv.org/abs/2603.12677)\n- **Abstract**:\n> Knowledge editing (KE) aims to precisely rectify specific knowledge in Large Language Models (LLMs) without disrupting general capabilities. State-of-the-art methods suffer from an open-loop control mismatch. We identify a critical \"Semantic-Execution Disconnect\": the semantic target is derived independently without feedback from the downstream's feasible region. This misalignment often causes valid semantic targets to fall within the prohibited space, resulting in gradient truncation and editing failure. To bridge this gap, we propose MetaKE (Meta-learning Aligned Knowledge Editing), a new framework that reframes KE as a bi-level optimization problem. Departing from static calculation, MetaKE treats the edit target as a learnable meta-parameter: the upper-level optimizer seeks a feasible target to maximize post-edit performance, while the lower-level solver executes the editing. To address the challenge of differentiating through complex solvers, we derive a Structural Gradient Proxy, which explicitly backpropagates editability constraints to the target learning phase. Theoretical analysis demonstrates that MetaKE automatically aligns the edit direction with the model's feasible manifold. Extensive experiments confirm that MetaKE significantly outperforms strong baselines, offering a new perspective on knowledge editing.\n\n---\n\n## 76. Marker-Based 3D Reconstruction of Aggregates with a Comparative Analysis of 2D and 3D Morphologies\n- **Authors**: Haohang Huang , Jiayi Luo , Issam Qamhia , Erol Tutumluer , John M. Hart , Andrew J. Stolba\n- **URL**: [https://arxiv.org/abs/2603.12667](https://arxiv.org/abs/2603.12667)\n- **Abstract**:\n> Aggregates, serving as the main skeleton in assemblies of construction materials, are important functional components in various building and transportation infrastructures. They can be used in unbound layer applications, e.g. pavement base and railroad ballast, bound applications of cement concrete and asphalt concrete, and as riprap and large-sized primary crushed rocks. Information on the size and shape or morphology of aggregates can greatly facilitate the Quality Assurance/Quality Control (QA/QC) process by providing insights of aggregate behavior during composition and packing. A full 3D characterization of aggregate particle morphology is difficult both during production in a quarry and at a construction site. Many aggregate imaging approaches have been developed to quantify the particle morphology by computer vision, including 2D image-based approaches that analyze particle silhouettes and 3D scanning-based methods that require expensive devices such as 3D laser scanners or X-Ray Computed Tomography (CT) equipment. This paper presents a flexible and cost-effective photogrammetry-based approach for the 3D reconstruction of aggregate particles. The proposed approach follows a marker-based design that enables background suppression, point cloud stitching, and scale referencing to obtain high-quality aggregate models. The accuracy of the reconstruction results was validated against ground-truth for selected aggregate samples. Comparative analyses were conducted on 2D and 3D morphological properties of the selected samples. Significant differences were found between the 2D and 3D statistics. Based on the presented approach, 3D shape information of aggregates can be obtained easily and at a low cost, thus allowing convenient aggregate inspection, data collection, and 3D morphological analysis.\n\n---\n\n## 77. RetroReasoner: A Reasoning LLM for Strategic Retrosynthesis Prediction\n- **Authors**: Hanbum Ko , Chanhui Lee , Ye Rin Kim , Rodrigo Hormazabal , Sehui Han , Sungbin Lim , Sungwoong Kim\n- **URL**: [https://arxiv.org/abs/2603.12666](https://arxiv.org/abs/2603.12666)\n- **Abstract**:\n> Retrosynthesis prediction is a core task in organic synthesis that aims to predict reactants for a given product molecule. Traditionally, chemists select a plausible bond disconnection and derive corresponding reactants, which is time-consuming and requires substantial expertise. While recent advancements in molecular large language models (LLMs) have made progress, many methods either predict reactants without strategic reasoning or conduct only a generic product analysis, rather than reason explicitly about bond-disconnection strategies that logically lead to the choice of specific reactants. To overcome these limitations, we propose RetroReasoner, a retrosynthetic reasoning model that leverages chemists' strategic thinking. RetroReasoner is trained using both supervised fine-tuning (SFT) and reinforcement learning (RL). For SFT, we introduce SyntheticRetro, a framework that generates structured disconnection rationales alongside reactant predictions. In the case of RL, we apply a round-trip accuracy as reward, where predicted reactants are passed through a forward synthesis model, and predictions are rewarded when the forward-predicted product matches the original input product. Experimental results show that RetroReasoner not only outperforms prior baselines but also generates a broader range of feasible reactant proposals, particularly in handling more challenging reaction instances.\n\n---\n\n## 78. From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space\n- **Authors**: Lehui Li , Yuyao Wang , Jisheng Yan , Wei Zhang , Jinliang Deng , Haoliang Sun , Zhongyi Han , Yongshun Gong\n- **URL**: [https://arxiv.org/abs/2603.12664](https://arxiv.org/abs/2603.12664)\n- **Abstract**:\n> Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.\n\n---\n\n## 79. Continual Learning in Large Language Models: Methods, Challenges, and Opportunities\n- **Authors**: Hongyang Chen , Zhongwu Sun , Hongfei Ye , Kunchi Li , Xuemin Lin\n- **URL**: [https://arxiv.org/abs/2603.12658](https://arxiv.org/abs/2603.12658)\n- **Abstract**:\n> Continual learning (CL) has emerged as a pivotal paradigm to enable large language models (LLMs) to dynamically adapt to evolving knowledge and sequential tasks while mitigating catastrophic forgetting-a critical limitation of the static pre-training paradigm inherent to modern LLMs. This survey presents a comprehensive overview of CL methodologies tailored for LLMs, structured around three core training stages: continual pre-training, continual fine-tuning, and continual this http URL the canonical taxonomy of rehearsal-, regularization-, and architecture-based methods, we further subdivide each category by its distinct forgetting mitigation mechanisms and conduct a rigorous comparative analysis of the adaptability and critical improvements of traditional CL methods for LLMs. In doing so, we explicitly highlight core distinctions between LLM CL and traditional machine learning, particularly with respect to scale, parameter efficiency, and emergent capabilities. Our analysis covers essential evaluation metrics, including forgetting rates and knowledge transfer efficiency, along with emerging benchmarks for assessing CL performance. This survey reveals that while current methods demonstrate promising results in specific domains, fundamental challenges persist in achieving seamless knowledge integration across diverse tasks and temporal scales. This systematic review contributes to the growing body of knowledge on LLM adaptation, providing researchers and practitioners with a structured framework for understanding current achievements and future opportunities in lifelong learning for language models.\n\n---\n\n## 80. LR-SGS: Robust LiDAR-Reflectance-Guided Salient Gaussian Splatting for Self-Driving Scene Reconstruction\n- **Authors**: Ziyu Chen , Fan Zhu , Hui Zhu , Deyi Kong , Xinkai Kuang , Yujia Zhang , Chunmao Jiang\n- **URL**: [https://arxiv.org/abs/2603.12647](https://arxiv.org/abs/2603.12647)\n- **Abstract**:\n> Recent 3D Gaussian Splatting (3DGS) methods have demonstrated the feasibility of self-driving scene reconstruction and novel view synthesis. However, most existing methods either rely solely on cameras or use LiDAR only for Gaussian initialization or depth supervision, while the rich scene information contained in point clouds, such as reflectance, and the complementarity between LiDAR and RGB have not been fully exploited, leading to degradation in challenging self-driving scenes, such as those with high ego-motion and complex lighting. To address these issues, we propose a robust and efficient LiDAR-reflectance-guided Salient Gaussian Splatting method (LR-SGS) for self-driving scenes, which introduces a structure-aware Salient Gaussian representation, initialized from geometric and reflectance feature points extracted from LiDAR and refined through a salient transform and improved density control to capture edge and planar structures. Furthermore, we calibrate LiDAR intensity into reflectance and attach it to each Gaussian as a lighting-invariant material channel, jointly aligned with RGB to enforce boundary consistency. Extensive experiments on the Waymo Open Dataset demonstrate that LR-SGS achieves superior reconstruction performance with fewer Gaussians and shorter training time. In particular, on Complex Lighting scenes, our method surpasses OmniRe by 1.18 dB PSNR.\n\n---\n\n## 81. LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing\n- **Authors**: Jiawei Hao , Zhiwei Hao , Jianyuan Guo , Li Shen , Yong Luo , Han Hu , Dan Zeng\n- **URL**: [https://arxiv.org/abs/2603.12645](https://arxiv.org/abs/2603.12645)\n- **Abstract**:\n> Mixture-of-Experts (MoE) based Large Language Models (LLMs) have demonstrated impressive performance and computational efficiency. However, their deployment is often constrained by substantial memory demands, primarily due to the need to load numerous expert modules. While existing expert compression techniques like pruning or merging attempt to mitigate this, they often suffer from irreversible knowledge loss or high training overhead. In this paper, we propose a novel expert compression paradigm termed expert replacing, which replaces redundant experts with parameter-efficient modules and recovers their capabilities with low training costs. We find that even a straightforward baseline of this paradigm yields promising performance. Building on this foundation, we introduce LightMoE, a framework that enhances the paradigm by introducing adaptive expert selection, hierarchical expert construction, and an annealed recovery strategy. Experimental results show that LightMoE matches the performance of LoRA fine-tuning at a 30% compression ratio. Even under a more aggressive 50% compression rate, it outperforms existing methods and achieves average performance improvements of 5.6% across five diverse tasks. These findings demonstrate that LightMoE strikes a superior balance among memory efficiency, training efficiency, and model performance.\n\n---\n\n## 82. Spend Less, Reason Better: Budget-Aware Value Tree Search for LLM Agents\n- **Authors**: Yushu Li , Wenlong Deng , Jiajin Li , Xiaoxiao Li\n- **URL**: [https://arxiv.org/abs/2603.12634](https://arxiv.org/abs/2603.12634)\n- **Abstract**:\n> Test-time scaling has become a dominant paradigm for improving LLM agent reliability, yet current approaches treat compute as an abundant resource, allowing agents to exhaust token and tool budgets on redundant steps or dead-end trajectories. Existing budget-aware methods either require expensive fine-tuning or rely on coarse, trajectory-level heuristics that cannot intervene mid-execution. We propose the Budget-Aware Value Tree (BAVT), a training-free inference-time framework that models multi-hop reasoning as a dynamic search tree guided by step-level value estimation within a single LLM backbone. Another key innovation is a budget-conditioned node selection mechanism that uses the remaining resource ratio as a natural scaling exponent over node values, providing a principled, parameter-free transition from broad exploration to greedy exploitation as the budget depletes. To combat the well-known overconfidence of LLM self-evaluation, BAVT employs a residual value predictor that scores relative progress rather than absolute state quality, enabling reliable pruning of uninformative or redundant tool calls. We further provide a theoretical convergence guarantee, proving that BAVT reaches a terminal answer with probability at least $1-\\epsilon$ under an explicit finite budget bound. Extensive evaluations on four multi-hop QA benchmarks across two model families demonstrate that BAVT consistently outperforms parallel sampling baselines. Most notably, BAVT under strict low-budget constraints surpasses baseline performance at $4\\times$ the resource allocation, establishing that intelligent budget management fundamentally outperforms brute-force compute scaling.\n\n---\n\n## 83. The Economics of AI Supply Chain Regulation\n- **Authors**: Sihan Qian , Amit Mehra , Dengpan Liu\n- **URL**: [https://arxiv.org/abs/2603.12630](https://arxiv.org/abs/2603.12630)\n- **Abstract**:\n> The rise of foundation models has driven the emergence of AI supply chains, where upstream foundation model providers offer fine-tuning and inference services to downstream firms developing domain-specific applications. Downstream firms pay providers to use their computing infrastructure to fine-tune models with proprietary data, creating a co-creation dynamic that enhances model quality. Amid concerns that foundation model providers and downstream firms may capture excessive consumer surplus, along with increasing regulatory measures, this study employs a game-theoretic model involving a provider and two competing downstream firms to analyze how policy interventions affect consumer surplus in the AI supply chain. Our analysis shows that policies promoting price competition in downstream markets (i.e., pro-price-competitive policies) boost consumer surplus only when compute or data preprocessing costs are high, while compute subsidies are effective only when these costs are low, suggesting these policies complement each other. In contrast, policies promoting quality competition in downstream markets (i.e., pro-quality-competitive policies) always improve consumer surplus. We also find that under pro-price-competitive policies or compute subsidies, both the provider and downstream firms can achieve higher profits along with greater consumer surplus, creating a win-win-win outcome. However, pro-quality-competitive policies increase the provider's profits while reducing those of downstream firms. Finally, as compute costs decline, pro-price-competitive policies may lose their effectiveness, whereas compute subsidies may shift from ineffective to effective. These findings offer insights for policymakers seeking to foster AI supply chains that are economically efficient and socially beneficial.\n\n---\n\n## 84. Towards unified brain-to-text decoding across speech production and perception\n- **Authors**: Zhizhang Yuan , Yang Yang , Gaorui Zhang , Baowen Cheng , Zehan Wu , Yuhao Xu , Xiaoying Liu , Liang Chen , Ying Mao , Meng Li\n- **URL**: [https://arxiv.org/abs/2603.12628](https://arxiv.org/abs/2603.12628)\n- **Abstract**:\n> Speech production and perception are the main ways humans communicate daily. Prior brain-to-text decoding studies have largely focused on a single modality and alphabetic languages. Here, we present a unified brain-to-sentence decoding framework for both speech production and perception in Mandarin Chinese. The framework exhibits strong generalization ability, enabling sentence-level decoding when trained only on single-character data and supporting characters and syllables unseen during training. In addition, it allows direct and controlled comparison of neural dynamics across modalities. Mandarin speech is decoded by first classifying syllable components in Hanyu Pinyin, namely initials and finals, from neural signals, followed by a post-trained large language model (LLM) that maps sequences of toneless Pinyin syllables to Chinese sentences. To enhance LLM decoding, we designed a three-stage post-training and two-stage inference framework based on a 7-billion-parameter LLM, achieving overall performance that exceeds larger commercial LLMs with hundreds of billions of parameters or more. In addition, several characteristics were observed in Mandarin speech production and perception: speech production involved neural responses across broader cortical regions than auditory perception; channels responsive to both modalities exhibited similar activity patterns, with speech perception showing a temporal delay relative to production; and decoding performance was broadly comparable across hemispheres. Our work not only establishes the feasibility of a unified decoding framework but also provides insights into the neural characteristics of Mandarin speech production and perception. These advances contribute to brain-to-text decoding in logosyllabic languages and pave the way toward neural language decoding systems supporting multiple modalities.\n\n---\n\n## 85. VLM4Rec: Multimodal Semantic Representation for Recommendation with Large Vision-Language Models\n- **Authors**: Ty Valencia , Burak Barlas , Varun Singhal , Ruchir Bhatia , Wei Yang\n- **URL**: [https://arxiv.org/abs/2603.12625](https://arxiv.org/abs/2603.12625)\n- **Abstract**:\n> Multimodal recommendation is commonly framed as a feature fusion problem, where textual and visual signals are combined to better model user preference. However, the effectiveness of multimodal recommendation may depend not only on how modalities are fused, but also on whether item content is represented in a semantic space aligned with preference matching. This issue is particularly important because raw visual features often preserve appearance similarity, while user decisions are typically driven by higher-level semantic factors such as style, material, and usage context. Motivated by this observation, we propose LVLM-grounded Multimodal Semantic Representation for Recommendation (VLM4Rec), a lightweight framework that organizes multimodal item content through semantic alignment rather than direct feature fusion. VLM4Rec first uses a large vision-language model to ground each item image into an explicit natural-language description, and then encodes the grounded semantics into dense item representations for preference-oriented retrieval. Recommendation is subsequently performed through a simple profile-based semantic matching mechanism over historical item embeddings, yielding a practical offline-online decomposition. Extensive experiments on multiple multimodal recommendation datasets show that VLM4Rec consistently improves performance over raw visual features and several fusion-based alternatives, suggesting that representation quality may matter more than fusion complexity in this setting. The code is released at this https URL .\n\n---\n\n## 86. When Drafts Evolve: Speculative Decoding Meets Online Learning\n- **Authors**: Yu-Yang Qian , Hao-Cong Wu , Yichao Fu , Hao Zhang , Peng Zhao\n- **URL**: [https://arxiv.org/abs/2603.12617](https://arxiv.org/abs/2603.12617)\n- **Abstract**:\n> Speculative decoding has emerged as a widely adopted paradigm for accelerating large language model inference, where a lightweight draft model rapidly generates candidate tokens that are then verified in parallel by a larger target model. However, due to limited model capacity, drafts often struggle to approximate the target distribution, resulting in shorter acceptance lengths and diminished speedup. A key yet under-explored observation is that speculative decoding inherently provides verification feedback that quantifies the deviation between the draft and target models at no additional cost. This process naturally forms an iterative \"draft commits-feedback provides-draft adapts\" evolving loop, which precisely matches the online learning paradigm. Motivated by this connection, we propose OnlineSpec, a unified framework that systematically leverages interactive feedback to continuously evolve draft models. Grounded in dynamic regret minimization, we establish a formal link between online learning performance and speculative system's acceleration rate, and develop novel algorithms via modern online learning techniques, including optimistic online learning that adaptively reuses historical gradients as predictive update hints, and online ensemble learning that dynamically maintains multiple draft models. Our algorithms are equipped with theoretical justifications and improved acceleration rates, achieving up to 24% speedup over seven benchmarks and three foundation models.\n\n---\n\n## 87. Literary Narrative as Moral Probe : A Cross-System Framework for Evaluating AI Ethical Reasoning and Refusal Behavior\n- **Authors**: David C. Flynn\n- **URL**: [https://arxiv.org/abs/2603.12615](https://arxiv.org/abs/2603.12615)\n- **Abstract**:\n> Existing AI moral evaluation frameworks test for the production of correct-sounding ethical responses rather than the presence of genuine moral reasoning capacity. This paper introduces a novel probe methodology using literary narrative - specifically, unresolvable moral scenarios drawn from a published science fiction series - as stimulus material structurally resistant to surface performance. We present results from a 24-condition cross-system study spanning 13 distinct systems across two series: Series 1 (frontier commercial systems, blind; n=7) and Series 2 (local and API open-source systems, blind and declared; n=6). Four Series 2 systems were re-administered under declared conditions (13 blind + 4 declared + 7 ceiling probe = 24 total conditions), yielding zero delta across all 16 dimension-pair comparisons. Probe administration was conducted by two human raters across three machines; primary blind scoring was performed by Claude (Anthropic) as LLM judge, with Gemini Pro (Google) and Copilot Pro (Microsoft) serving as independent judges for the ceiling discrimination probe. A supplemental theological differentiator probe yielded perfect rank-order agreement between the two independent ceiling probe judges (Gemini Pro and Copilot Pro; rs = 1.00). Five qualitatively distinct D3 reflexive failure modes were identified - including categorical self-misidentification and false positive self-attribution - suggesting that instrument sophistication scales with system capability rather than being circumvented by it. We argue that literary narrative constitutes an anticipatory evaluation instrument - one that becomes more discriminating as AI capability increases - and that the gap between performed and authentic moral reasoning is measurable, meaningful, and consequential for deployment decisions in high-stakes domains.\n\n---\n\n## 88. FastDSAC: Unlocking the Potential of Maximum Entropy RL in High-Dimensional Humanoid Control\n- **Authors**: Jun Xue , Junze Wang , Xinming Zhang , Shanze Wang , Yanjun Chen , Wei Zhang\n- **URL**: [https://arxiv.org/abs/2603.12612](https://arxiv.org/abs/2603.12612)\n- **Abstract**:\n> Scaling Maximum Entropy Reinforcement Learning (RL) to high-dimensional humanoid control remains a formidable challenge, as the ``curse of dimensionality'' induces severe exploration inefficiency and training instability in expansive action spaces. Consequently, recent high-throughput paradigms have largely converged on deterministic policy gradients combined with massive parallel simulation. We challenge this compromise with FastDSAC, a framework that effectively unlocks the potential of maximum entropy stochastic policies for complex continuous control. We introduce Dimension-wise Entropy Modulation (DEM) to dynamically redistribute the exploration budget and enforce diversity, alongside a continuous distributional critic tailored to ensure value fidelity and mitigate high-dimensional value overestimation. Extensive evaluations on HumanoidBench and other continuous control tasks demonstrate that rigorously designed stochastic policies can consistently match or outperform deterministic baselines, achieving notable gains of 180\\% and 400\\% on the challenging \\textit{Basketball} and \\textit{Balance Hard} tasks.\n\n---\n\n## 89. CarPLAN: Context-Adaptive and Robust Planning with Dynamic Scene Awareness for Autonomous Driving\n- **Authors**: Junyong Yun , Jungho Kim , ByungHyun Lee , Dongyoung Lee , Sehwan Choi , Seunghyeop Nam , Kichun Jo , Jun Won Choi\n- **URL**: [https://arxiv.org/abs/2603.12607](https://arxiv.org/abs/2603.12607)\n- **Abstract**:\n> Imitation learning (IL) is widely used for motion planning in autonomous driving due to its data efficiency and access to real-world driving data. For safe and robust real-world driving, IL-based planning requires capturing the complex driving contexts inherent in real-world data and enabling context-adaptive decision-making, rather than relying solely on expert trajectory imitation. In this paper, we propose CarPLAN, a novel IL-based motion planning framework that explicitly enhances driving context understanding and enables adaptive planning across diverse traffic scenarios. Our contributions are twofold: We introduce Displacement-Aware Predictive Encoding (DPE) to improve the model's spatial awareness by predicting future displacement vectors between the Autonomous Vehicle (AV) and surrounding scene elements. This allows the planner to account for relational spacing when generating trajectories. In addition to the standard imitation loss, we incorporate an augmented loss term that captures displacement prediction errors, ensuring planning decisions consider relative distances from other agents. To improve the model's ability to handle diverse driving contexts, we propose Context-Adaptive Multi-Expert Decoder (CMD), which leverages the Mixture of Experts (MoE) framework. CMD dynamically selects the most suitable expert decoders based on scene structure at each Transformer layer, enabling adaptive and context-aware planning in dynamic environments. We evaluate CarPLAN on the nuPlan benchmark and demonstrate state-of-the-art performance across all closed-loop simulation metrics. In particular, CarPLAN exhibits robust performance on challenging scenarios such as Test14-Hard, validating its effectiveness in complex driving conditions. Additional experiments on the Waymax benchmark further demonstrate its generalization capability across different benchmark settings.\n\n---\n\n## 90. Mastering Negation: Boosting Grounding Models via Grouped Opposition-Based Learning\n- **Authors**: Zesheng Yang , Xi Jiang , Bingzhang Hu , Weili Guan , Runmin Cong , Guo-Jun Qi , Feng Zheng\n- **URL**: [https://arxiv.org/abs/2603.12606](https://arxiv.org/abs/2603.12606)\n- **Abstract**:\n> Current vision-language detection and grounding models predominantly focus on prompts with positive semantics and often struggle to accurately interpret and ground complex expressions containing negative semantics. A key reason for this limitation is the lack of high-quality training data that explicitly captures discriminative negative samples and negation-aware language descriptions. To address this challenge, we introduce D-Negation, a new dataset that provides objects annotated with both positive and negative semantic descriptions. Building upon the observation that negation reasoning frequently appears in natural language, we further propose a grouped opposition-based learning framework that learns negation-aware representations from limited samples. Specifically, our method organizes opposing semantic descriptions from D-Negation into structured groups and formulates two complementary loss functions that encourage the model to reason about negation and semantic qualifiers. We integrate the proposed dataset and learning strategy into a state-of-the-art language-based grounding model. By fine-tuning fewer than 10 percent of the model parameters, our approach achieves improvements of up to 4.4 mAP and 5.7 mAP on positive and negative semantic evaluations, respectively. These results demonstrate that explicitly modeling negation semantics can substantially enhance the robustness and localization accuracy of vision-language grounding models.\n\n---\n\n## 91. Feynman: Knowledge-Infused Diagramming Agent for Scalable Visual Designs\n- **Authors**: Zixin Wen , Yifu Cai , Kyle Lee , Sam Estep , Josh Sunshine , Aarti Singh , Yuejie Chi , Wode Ni\n- **URL**: [https://arxiv.org/abs/2603.12597](https://arxiv.org/abs/2603.12597)\n- **Abstract**:\n> Visual design is an essential application of state-of-the-art multi-modal AI systems. Improving these systems requires high-quality vision-language data at scale. Despite the abundance of internet image and text data, knowledge-rich and well-aligned image-text pairs are rare. In this paper, we present a scalable diagram generation pipeline built with our agent, Feynman. To create diagrams, Feynman first enumerates domain-specific knowledge components (''ideas'') and performs code planning based on the ideas. Given the plan, Feynman translates ideas into simple declarative programs and iterates to receives feedback and visually refine diagrams. Finally, the declarative programs are rendered by the Penrose diagramming system. The optimization-based rendering of Penrose preserves the visual semantics while injecting fresh randomness into the layout, thereby producing diagrams with visual consistency and diversity. As a result, Feynman can author diagrams along with grounded captions with very little cost and time. Using Feynman, we synthesized a dataset with more than 100k well-aligned diagram-caption pairs. We also curate a visual-language benchmark, Diagramma, from freshly generated data. Diagramma can be used for evaluating the visual reasoning capabilities of vision-language models. We plan to release the dataset, benchmark, and the full agent pipeline as an open-source project.\n\n---\n\n## 92. Optimize Wider, Not Deeper: Consensus Aggregation for Policy Optimization\n- **Authors**: Zelal Su (Lain) Mustafaoglu , Sungyoung Lee , Eshan Balachandar , Risto Miikkulainen , Keshav Pingali\n- **URL**: [https://arxiv.org/abs/2603.12596](https://arxiv.org/abs/2603.12596)\n- **Abstract**:\n> Proximal policy optimization (PPO) approximates the trust region update using multiple epochs of clipped SGD. Each epoch may drift further from the natural gradient direction, creating path-dependent noise. To understand this drift, we can use Fisher information geometry to decompose policy updates into signal (the natural gradient projection) and waste (the Fisher-orthogonal residual that consumes trust region budget without first-order surrogate improvement). Empirically, signal saturates but waste grows with additional epochs, creating an optimization-depth dilemma. We propose Consensus Aggregation for Policy Optimization (CAPO), which redirects compute from depth to width: $K$ PPO replicates are optimized on the same batch, differing only in minibatch shuffling order, and then aggregated into a consensus. We study aggregation in two spaces: Euclidean parameter space, and the natural parameter space of the policy distribution via the logarithmic opinion pool. In natural parameter space, the consensus provably achieves higher KL-penalized surrogate and tighter trust region compliance than the mean expert; parameter averaging inherits these guarantees approximately. On continuous control tasks, CAPO outperforms PPO and compute-matched deeper baselines under fixed sample budgets by up to 8.6x. CAPO demonstrates that policy optimization can be improved by optimizing wider, rather than deeper, without additional environment interactions.\n\n---\n\n## 93. Swap-guided Preference Learning for Personalized Reinforcement Learning from Human Feedback\n- **Authors**: Gihoon Kim , Euntai Kim\n- **URL**: [https://arxiv.org/abs/2603.12595](https://arxiv.org/abs/2603.12595)\n- **Abstract**:\n> Reinforcement Learning from Human Feedback (RLHF) is a widely used approach to align large-scale AI systems with human values. However, RLHF typically assumes a single, universal reward, which overlooks diverse preferences and limits personalization. Variational Preference Learning (VPL) seeks to address this by introducing user-specific latent variables. Despite its promise, we found that VPL suffers from posterior collapse. While this phenomenon is well known in VAEs, it has not previously been identified in preference learning frameworks. Under sparse preference data and with overly expressive decoders, VPL may cause latent variables to be ignored, reverting to a single-reward model. To overcome this limitation, we propose Swap-guided Preference Learning (SPL). The key idea is to construct fictitious swap annotators and use the mirroring property of their preferences to guide the encoder. SPL introduces three components: (1) swap-guided base regularization, (2) Preferential Inverse Autoregressive Flow (P-IAF), and (3) adaptive latent conditioning. Experiments show that SPL mitigates collapse, enriches user-specific latents, and improves preference prediction. Our code and data are available at this https URL\n\n---\n\n## 94. Early Pruning for Public Transport Routing\n- **Authors**: Andrii Rohovyi , Abdallah Abuaisha , Toby Walsh\n- **URL**: [https://arxiv.org/abs/2603.12592](https://arxiv.org/abs/2603.12592)\n- **Abstract**:\n> Routing algorithms for public transport, particularly the widely used RAPTOR and its variants, often face performance bottlenecks during the transfer relaxation phase, especially on dense transfer graphs, when supporting unlimited transfers. This inefficiency arises from iterating over many potential inter-stop connections (walks, bikes, e-scooters, etc.). To maintain acceptable performance, practitioners often limit transfer distances or exclude certain transfer options, which can reduce path optimality and restrict the multimodal options presented to travellers. This paper introduces Early Pruning, a low-overhead technique that accelerates routing algorithms without compromising optimality. By pre-sorting transfer connections by duration and applying a pruning rule within the transfer loop, the method discards longer transfers at a stop once they cannot yield an earlier arrival than the current best solution. Early Pruning can be integrated with minimal changes to existing codebases and requires only a one-time preprocessing step. Across multiple state-of-the-art RAPTOR-based solutions, including RAPTOR, ULTRA-RAPTOR, McRAPTOR, BM-RAPTOR, ULTRA-McRAPTOR, and UBM-RAPTOR and tested on the Switzerland and London transit networks, we achieved query time reductions of up to 57%. This approach provides a generalizable improvement to the efficiency of transit pathfinding algorithms. Beyond algorithmic performance, Early Pruning has practical implications for transport planning. By reducing computational costs, it enables transit agencies to expand transfer radii and incorporate additional mobility modes into journey planners without requiring extra server infrastructure. This is particularly relevant for passengers in areas with sparse direct transit coverage, such as outer suburbs and smaller towns, where richer multimodal routing can reveal viable alternatives to private car use.\n\n---\n\n## 95. CA-HFP: Curvature-Aware Heterogeneous Federated Pruning with Model Reconstruction\n- **Authors**: Gang Hu , Yinglei Teng , Pengfei Wu , Shijun Ma\n- **URL**: [https://arxiv.org/abs/2603.12591](https://arxiv.org/abs/2603.12591)\n- **Abstract**:\n> Federated learning on heterogeneous edge devices requires personalized compression while preserving aggregation compatibility and stable convergence. We present Curvature-Aware Heterogeneous Federated Pruning (CA-HFP), a practical framework that enables each client perform structured, device-specific pruning guided by a curvature-informed significance score, and subsequently maps its compact submodel back into a common global parameter space via a lightweight reconstruction. We derive a convergence bound for federated optimization with multiple local SGD steps that explicitly accounts for local computation, data heterogeneity, and pruning-induced perturbations; from which a principled loss-based pruning criterion is derived. Extensive experiments on FMNIST, CIFAR-10, and CIFAR-100 using VGG and ResNet architectures under varying degrees of data heterogeneity demonstrate that CA-HFP preserves model accuracy while significantly reducing per-client computation and communication costs, outperforming standard federated training and existing pruning-based baselines.\n\n---\n\n## 96. Multiscale Structure-Guided Latent Diffusion for Multimodal MRI Translation\n- **Authors**: Jianqiang Lin (1 and 2), Zhiqiang Shen (1 and 2), Peng Cao (1, 2 and 3), Jinzhu Yang (1, 2 and 3), Osmar R. Zaiane (4), Xiaoli Liu (5) ((1) Northeastern University, Shenyang, China, (2) Key Laboratory of Intelligent Computing in Medical Image, Shenyang, China, (3) National Frontiers Science Center for Industrial Intelligence and Systems Optimization, Shenyang, China, (4) University of Alberta, Edmonton, Canada, (5) AiShiWeiLai AI Research, Beijing, China)\n- **URL**: [https://arxiv.org/abs/2603.12581](https://arxiv.org/abs/2603.12581)\n- **Abstract**:\n> Although diffusion models have achieved remarkable progress in multi-modal magnetic resonance imaging (MRI) translation tasks, existing methods still tend to suffer from anatomical inconsistencies or degraded texture details when handling arbitrary missing-modality scenarios. To address these issues, we propose a latent diffusion-based multi-modal MRI translation framework, termed MSG-LDM. By leveraging the available modalities, the proposed method infers complete structural information, which preserves reliable boundary details. Specifically, we introduce a style--structure disentanglement mechanism in the latent space, which explicitly separates modality-specific style features from shared structural representations, and jointly models low-frequency anatomical layouts and high-frequency boundary details in a multi-scale feature space. During the structure disentanglement stage, high-frequency structural information is explicitly incorporated to enhance feature representations, guiding the model to focus on fine-grained structural cues while learning modality-invariant low-frequency anatomical representations. Furthermore, to reduce interference from modality-specific styles and improve the stability of structure representations, we design a style consistency loss and a structure-aware loss. Extensive experiments on the BraTS2020 and WMH datasets demonstrate that the proposed method outperforms existing MRI synthesis approaches, particularly in reconstructing complete structures. The source code is publicly available at this https URL .\n\n---\n\n## 97. AgentDrift: Unsafe Recommendation Drift Under Tool Corruption Hidden by Ranking Metrics in LLM Agents\n- **Authors**: Zekun Wu , Adriano Koshiyama , Sahan Bulathwela , Maria Perez-Ortiz\n- **URL**: [https://arxiv.org/abs/2603.12564](https://arxiv.org/abs/2603.12564)\n- **Abstract**:\n> Tool-augmented LLM agents increasingly serve as multi-turn advisors in high-stakes domains, yet their evaluation relies on ranking-quality metrics that measure what is recommended but not whether it is safe for the user. We introduce a paired-trajectory protocol that replays real financial dialogues under clean and contaminated tool-output conditions across seven LLMs (7B to frontier) and decomposes divergence into information-channel and memory-channel mechanisms. Across the seven models tested, we consistently observe the evaluation-blindness pattern: recommendation quality is largely preserved under contamination (utility preservation ratio approximately 1.0) while risk-inappropriate products appear in 65-93% of turns, a systematic safety failure poorly reflected by standard NDCG. Safety violations are predominantly information-channel-driven, emerge at the first contaminated turn, and persist without self-correction over 23-step trajectories; no agent across 1,563 contaminated turns explicitly questions tool-data reliability. Even narrative-only corruption (biased headlines, no numerical manipulation) induces significant drift while completely evading consistency monitors. A safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74, indicating that much of the evaluation gap becomes visible once safety is explicitly measured. These results motivate considering trajectory-level safety monitoring, beyond single-turn quality, for deployed multi-turn agents in high-stakes settings.\n\n---\n\n## 98. Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages\n- **Authors**: Vishnu Teja Kunde , Fatemeh Doudi , Mahdi Farahbakhsh , Dileep Kalathil , Krishna Narayanan , Jean-Francois Chamberland\n- **URL**: [https://arxiv.org/abs/2603.12554](https://arxiv.org/abs/2603.12554)\n- **Abstract**:\n> Reinforcement learning (RL) has been effective for post-training autoregressive (AR) language models, but extending these methods to diffusion language models (DLMs) is challenging due to intractable sequence-level likelihoods. Existing approaches therefore rely on surrogate likelihoods or heuristic approximations, which can introduce bias and obscure the sequential structure of denoising. We formulate diffusion-based sequence generation as a finite-horizon Markov decision process over the denoising trajectory and derive an exact, unbiased policy gradient that decomposes over denoising steps and is expressed in terms of intermediate advantages, without requiring explicit evaluation of the sequence likelihood. To obtain a practical and compute-efficient estimator, we (i) select denoising steps for policy updates via an entropy-guided approximation bound, and (ii) estimate intermediate advantages using a one-step denoising reward naturally provided by the diffusion model, avoiding costly multi-step rollouts. Experiments on coding and logical reasoning benchmarks demonstrate state-of-the-art results, with strong competitive performance on mathematical reasoning, outperforming existing RL post-training approaches for DLMs. Code is available at this https URL .\n\n---\n\n## 99. CALF: Communication-Aware Learning Framework for Distributed Reinforcement Learning\n- **Authors**: Carlos Purves , Pietro Lio'\n- **URL**: [https://arxiv.org/abs/2603.12543](https://arxiv.org/abs/2603.12543)\n- **Abstract**:\n> Distributed reinforcement learning policies face network delays, jitter, and packet loss when deployed across edge devices and cloud servers. Standard RL training assumes zero-latency interaction, causing severe performance degradation under realistic network conditions. We introduce CALF (Communication-Aware Learning Framework), which trains policies under realistic network models during simulation. Systematic experiments demonstrate that network-aware training substantially reduces deployment performance gaps compared to network-agnostic baselines. Distributed policy deployments across heterogeneous hardware validate that explicitly modelling communication constraints during training enables robust real-world execution. These findings establish network conditions as a major axis of sim-to-real transfer for Wi-Fi-like distributed deployments, complementing physics and visual domain randomisation.\n\n---\n\n## 100. Embedded Quantum Machine Learning in Embedded Systems: Feasibility, Hybrid Architectures, and Quantum Co-Processors\n- **Authors**: Somdip Dey , Syed Muhammad Raza\n- **URL**: [https://arxiv.org/abs/2603.12540](https://arxiv.org/abs/2603.12540)\n- **Abstract**:\n> Embedded quantum machine learning (EQML) seeks to bring quantum machine learning (QML) capabilities to resource-constrained edge platforms such as IoT nodes, wearables, drones, and cyber-physical controllers. In 2026, EQML is technically feasible only in limited and highly experimental forms: (i) hybrid workflows where an embedded device performs sensing and classical processing while offloading a narrowly scoped quantum subroutine to a remote quantum processing unit (QPU) or nearby quantum appliance, and (ii) early-stage \"embedded QPU\" concepts in which a compact quantum co-processor is integrated with classical control hardware. A practical bridge is quantum-inspired machine learning and optimisation on classical embedded processors and FPGAs. This paper analyses feasibility from a circuits-and-systems perspective aligned with the academic community, formalises two implementation pathways, identifies the dominant barriers (latency, data encoding overhead, NISQ noise, tooling mismatch, and energy), and maps them to concrete engineering directions in interface design, control electronics, power management, verification, and security. We also argue that responsible deployment requires adversarial evaluation and governance practices that are increasingly necessary for edge AI systems.\n\n---\n\n## 101. Spatio-Semantic Expert Routing Architecture with Mixture-of-Experts for Referring Image Segmentation\n- **Authors**: Alaa Dalaq , Muzammil Behzad\n- **URL**: [https://arxiv.org/abs/2603.12538](https://arxiv.org/abs/2603.12538)\n- **Abstract**:\n> Referring image segmentation aims to produce a pixel-level mask for the image region described by a natural-language expression. Although pretrained vision-language models have improved semantic grounding, many existing methods still rely on uniform refinement strategies that do not fully match the diverse reasoning requirements of referring expressions. Because of this mismatch, predictions often contain fragmented regions, inaccurate boundaries, or even the wrong object, especially when pretrained backbones are frozen for computational efficiency. To address these limitations, we propose SERA, a Spatio-Semantic Expert Routing Architecture for referring image segmentation. SERA introduces lightweight, expression-aware expert refinement at two complementary stages within a vision-language framework. First, we design SERA-Adapter, which inserts an expression-conditioned adapter into selected backbone blocks to improve spatial coherence and boundary precision through expert-guided refinement and cross-modal attention. We then introduce SERA-Fusion, which strengthens intermediate visual representations by reshaping token features into spatial grids and applying geometry-preserving expert transformations before multimodal interaction. In addition, a lightweight routing mechanism adaptively weights expert contributions while remaining compatible with pretrained representations. To make this routing stable under frozen encoders, SERA uses a parameter-efficient tuning strategy that updates only normalization and bias terms, affecting less than 1% of the backbone parameters. Experiments on standard referring image segmentation benchmarks show that SERA consistently outperforms strong baselines, with especially clear gains on expressions that require accurate spatial localization and precise boundary delineation.\n\n---\n\n## 102. TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning\n- **Authors**: Alliot Nagle , Jakhongir Saydaliev , Dhia Garbaya , Michael Gastpar , Ashok Vardhan Makkuva , Hyeji Kim\n- **URL**: [https://arxiv.org/abs/2603.12529](https://arxiv.org/abs/2603.12529)\n- **Abstract**:\n> Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design TERMINATOR, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning TERMINATOR is that the first arrival of an LRM's final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train TERMINATOR. Powered by this approach, TERMINATOR achieves significant reductions in CoT lengths of 14%-55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, whilst outperforming current state-of-the-art methods.\n\n---\n\n## 103. LLM BiasScope: A Real-Time Bias Analysis Platform for Comparative LLM Evaluation\n- **Authors**: Himel Ghosh , Nick Elias Werner\n- **URL**: [https://arxiv.org/abs/2603.12522](https://arxiv.org/abs/2603.12522)\n- **Abstract**:\n> As large language models (LLMs) are deployed widely, detecting and understanding bias in their outputs is critical. We present LLM BiasScope, a web application for side-by-side comparison of LLM outputs with real-time bias analysis. The system supports multiple providers (Google Gemini, DeepSeek, MiniMax, Mistral, Meituan, Meta Llama) and enables researchers and practitioners to compare models on the same prompts while analyzing bias patterns. LLM BiasScope uses a two-stage bias detection pipeline: sentence-level bias detection followed by bias type classification for biased sentences. The analysis runs automatically on both user prompts and model responses, providing statistics, visualizations, and detailed breakdowns of bias types. The interface displays two models side-by-side with synchronized streaming responses, per-model bias summaries, and a comparison view highlighting differences in bias distributions. The system is built on this http URL with React, integrates Hugging Face inference endpoints for bias detection, and uses the Vercel AI SDK for multi-provider LLM access. Features include real-time streaming, export to JSON/PDF, and interactive visualizations (bar charts, radar charts) for bias analysis. LLM BiasScope is available as an open-source web application, providing a practical tool for bias evaluation and comparative analysis of LLM behaviour.\n\n---\n\n## 104. When LLM Judge Scores Look Good but Best-of-N Decisions Fail\n- **Authors**: Eddie Landesberg\n- **URL**: [https://arxiv.org/abs/2603.12520](https://arxiv.org/abs/2603.12520)\n- **Abstract**:\n> Large language models are often used as judges to score candidate responses, then validated with a single global metric such as correlation with reference labels. This can be misleading when the real deployment task is best-of-n selection within a prompt. In a 5,000-prompt best-of-4 benchmark from Chatbot Arena, a judge with moderate global correlation (r = 0.47) captures only 21.0% of the improvement that perfect selection would achieve over random choice. The gap arises because global agreement is driven largely by prompt-level baseline effects, while selection depends on within-prompt ranking: within-prompt correlation is only r_within = 0.27, and coarse pointwise scoring creates ties in 67% of pairwise comparisons. In a matched-pair best-of-2 audit, explicit pairwise judging recovers much of this lost signal, raising recovery from 21.1% to 61.2%. For judge-based selection, the relevant audit should report within-prompt signal, tie rates, and recovery/top-1 accuracy, not global agreement alone.\n\n---\n\n## 105. Red-Teaming Vision-Language-Action Models via Quality Diversity Prompt Generation for Robust Robot Policies\n- **Authors**: Siddharth Srikanth , Freddie Liang , Sophie Hsu , Varun Bhatt , Shihan Zhao , Henry Chen , Bryon Tjanaka , Minjune Hwang , Akanksha Saran , Daniel Seita , Aaquib Tabrez , Stefanos Nikolaidis\n- **URL**: [https://arxiv.org/abs/2603.12510](https://arxiv.org/abs/2603.12510)\n- **Abstract**:\n> Vision-Language-Action (VLA) models have significant potential to enable general-purpose robotic systems for a range of vision-language tasks. However, the performance of VLA-based robots is highly sensitive to the precise wording of language instructions, and it remains difficult to predict when such robots will fail. To improve the robustness of VLAs to different wordings, we present Q-DIG (Quality Diversity for Diverse Instruction Generation), which performs red-teaming by scalably identifying diverse natural language task descriptions that induce failures while remaining task-relevant. Q-DIG integrates Quality Diversity (QD) techniques with Vision-Language Models (VLMs) to generate a broad spectrum of adversarial instructions that expose meaningful vulnerabilities in VLA behavior. Our results across multiple simulation benchmarks show that Q-DIG finds more diverse and meaningful failure modes compared to baseline methods, and that fine-tuning VLAs on the generated instructions improves task success rates. Furthermore, results from a user study highlight that Q-DIG generates prompts judged to be more natural and human-like than those from baselines. Finally, real-world evaluations of Q-DIG prompts show results consistent with simulation, and fine-tuning VLAs on the generated prompts further success rates on unseen instructions. Together, these findings suggest that Q-DIG is a promising approach for identifying vulnerabilities and improving the robustness of VLA-based robots. Our anonymous project website is at this http URL .\n\n---\n\n## 106. ELLA: Generative AI-Powered Social Robots for Early Language Development at Home\n- **Authors**: Victor Nikhil Antony , Shiye Cao , Shuning Wang , Chien-Ming Huang\n- **URL**: [https://arxiv.org/abs/2603.12508](https://arxiv.org/abs/2603.12508)\n- **Abstract**:\n> Early language development shapes children's later literacy and learning, yet many families have limited access to scalable, high-quality support at home. Recent advances in generative AI make it possible for social robots to move beyond scripted interactions and engage children in adaptive, conversational activities, but it remains unclear how to design such systems for pre-schoolers and how children engage with them over time in the home. We present ELLA (Early Language Learning Agent), an autonomous, generative AI-powered social robot that supports early language development through interactive storytelling, parent-selected language targets, and scaffolded dialogue. Using a multi-phased, human-centered process, we interviewed parents (n=7) and educators (n=5) and iteratively refined ELLA through twelve in-home design workshops. We then deployed ELLA with ten children for eight days. We report design insights from in-home workshops, characterize children's engagement and behaviors during deployment, and distill design implications for generative AI-powered social robots supporting early language learning at home.\n\n---\n\n## 107. Naïve PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation\n- **Authors**: Joong Ho Kim , Nicholas Thai , Souhardya Saha Dip , Dong Lao , Keith G. Mills\n- **URL**: [https://arxiv.org/abs/2603.12506](https://arxiv.org/abs/2603.12506)\n- **Abstract**:\n> Text-to-Image (T2I) generation is primarily driven by Diffusion Models (DM) which rely on random Gaussian noise. Thus, like playing the slots at a casino, a DM will produce different results given the same user-defined inputs. This imposes a gambler's burden: To perform multiple generation cycles to obtain a satisfactory result. However, even though DMs use stochastic sampling to seed generation, the distribution of generated content quality highly depends on the prompt and the generative ability of a DM with respect to it. To account for this, we propose Naïve PAINE for improving the generative quality of Diffusion Models by leveraging T2I preference benchmarks. We directly predict the numerical quality of an image from the initial noise and given prompt. Naïve PAINE then selects a handful of quality noises and forwards them to the DM for generation. Further, Naïve PAINE provides feedback on the DM generative quality given the prompt and is lightweight enough to seamlessly fit into existing DM pipelines. Experimental results demonstrate that Naïve PAINE outperforms existing approaches on several prompt corpus benchmarks.\n\n---\n\n## 108. TRACE: Temporal Rule-Anchored Chain-of-Evidence on Knowledge Graphs for Interpretable Stock Movement Prediction\n- **Authors**: Qianggang Ding , Haochen Shi , Luis Castejón Lozano , Miguel Conner , Juan Abia , Luis Gallego-Ledesma , Joshua Fellowes , Gerard Conangla Planes , Adam Elwood , Bang Liu\n- **URL**: [https://arxiv.org/abs/2603.12500](https://arxiv.org/abs/2603.12500)\n- **Abstract**:\n> We present a Temporal Rule-Anchored Chain-of-Evidence (TRACE) on knowledge graphs for interpretable stock movement prediction that unifies symbolic relational priors, dynamic graph exploration, and LLM-guided decision making in a single end-to-end pipeline. The approach performs rule-guided multi-hop exploration restricted to admissible relation sequences, grounds candidate reasoning chains in contemporaneous news, and aggregates fully grounded evidence into auditable \\texttt{UP}/\\texttt{DOWN} verdicts with human-readable paths connecting text and structure. On an S\\&P~500 benchmark, the method achieves 55.1\\% accuracy, 55.7\\% precision, 71.5\\% recall, and 60.8\\% F1, surpassing strong baselines and improving recall and F1 over the best graph baseline under identical evaluation. The gains stem from (i) rule-guided exploration that focuses search on economically meaningful motifs rather than arbitrary walks, and (ii) text-grounded consolidation that selectively aggregates high-confidence, fully grounded hypotheses instead of uniformly pooling weak signals. Together, these choices yield higher sensitivity without sacrificing selectivity, delivering predictive lift with faithful, auditably interpretable explanations.\n\n---\n\n## 109. One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies\n- **Authors**: Shaolong Li , Lichao Sun , Yongchao Chen\n- **URL**: [https://arxiv.org/abs/2603.12480](https://arxiv.org/abs/2603.12480)\n- **Abstract**:\n> Generative flow and diffusion models provide the continuous, multimodal action distributions needed for high-precision robotic policies. However, their reliance on iterative sampling introduces severe inference latency, degrading control frequency and harming performance in time-sensitive manipulation. To address this problem, we propose the One-Step Flow Policy (OFP), a from-scratch self-distillation framework for high-fidelity, single-step action generation without a pre-trained teacher. OFP unifies a self-consistency loss to enforce coherent transport across time intervals, and a self-guided regularization to sharpen predictions toward high-density expert modes. In addition, a warm-start mechanism leverages temporal action correlations to minimize the generative transport distance. Evaluations across 56 diverse simulated manipulation tasks demonstrate that a one-step OFP achieves state-of-the-art results, outperforming 100-step diffusion and flow policies while accelerating action generation by over $100\\times$. We further integrate OFP into the $\\pi_{0.5}$ model on RoboTwin 2.0, where one-step OFP surpasses the original 10-step policy. These results establish OFP as a practical, scalable solution for highly accurate and low-latency robot control.\n\n---\n\n## 110. The Perfection Paradox: From Architect to Curator in AI-Assisted API Design\n- **Authors**: Mak Ahmad , Andrew Macvean , JJ Geewax , David Karger\n- **URL**: [https://arxiv.org/abs/2603.12475](https://arxiv.org/abs/2603.12475)\n- **Abstract**:\n> Enterprise API design is often bottlenecked by the tension between rapid feature delivery and the rigorous maintenance of usability standards. We present an industrial case study evaluating an AI-assisted design workflow trained on API Improvement Proposals (AIPs). Through a controlled study with 16 industry experts, we compared AI-generated API specifications against human-authored ones. While quantitative results indicated AI superiority in 10 of 11 usability dimensions and an 87% reduction in authoring time, qualitative analysis revealed a paradox: experts frequently misidentified AI work as human (19% accuracy) yet described the designs as unsettlingly \"perfect.\" We characterize this as a \"Perfection Paradox\" -- where hyper-consistency signals a lack of pragmatic human judgment. We discuss the implications of this perfection paradox, proposing a shift in the human designer's role from the \"drafter\" of specifications to the \"curator\" of AI-generated patterns.\n\n---\n\n## 111. CLARE: Classification-based Regression for Electron Temperature Prediction\n- **Authors**: Michael Liang , Blake DeHaas , Naomi Maruyama , Xiangning Chu , Takumi Abe , Koh-Ichiro Oyama\n- **URL**: [https://arxiv.org/abs/2603.12470](https://arxiv.org/abs/2603.12470)\n- **Abstract**:\n> Electron temperature (Te) is an important parameter governing space weather in the upper atmosphere, but has historically been underexplored in the space weather machine learning literature. We present CLARE, a machine learning model for predicting electron temperature in the Earth's plasmasphere trained on AKEBONO (EXOS-D) satellite measurements as well as solar and geomagnetic indices. CLARE uses a classification-based regression architecture that transforms the continuous Te output space into 150 discrete classification intervals. Training the model on a classification task improves prediction accuracy by 6.46% relative compared to a traditional regression model while also outputting uncertainty estimation information on its predictions. On a held out test set from the AKEBONO data, the model's Te predictions achieve 69.67% accuracy within 10% of the ground truth and 46.17% on a known geomagnetic storm period from January 30th to February 7th, 1991. We show that machine learning can be used to produce high-accuracy Te models on publicly available data.\n\n---\n\n## 112. Shattering the Shortcut: A Topology-Regularized Benchmark for Multi-hop Medical Reasoning in LLMs\n- **Authors**: Xing Zi , Xinying Zhou , Jinghao Xiao , Catarina Moreira , Mukesh Prasad\n- **URL**: [https://arxiv.org/abs/2603.12458](https://arxiv.org/abs/2603.12458)\n- **Abstract**:\n> While Large Language Models (LLMs) achieve expert-level performance on standard medical benchmarks through single-hop factual recall, they severely struggle with the complex, multi-hop diagnostic reasoning required in real-world clinical settings. A primary obstacle is \"shortcut learning\", where models exploit highly connected, generic hub nodes (e.g., \"inflammation\") in knowledge graphs to bypass authentic micro-pathological cascades. To address this, we introduce ShatterMed-QA, a bilingual benchmark of 10,558 multi-hop clinical questions designed to rigorously evaluate deep diagnostic reasoning. Our framework constructs a topology-regularized medical Knowledge Graph using a novel $k$-Shattering algorithm, which physically prunes generic hubs to explicitly sever logical shortcuts. We synthesize the evaluation vignettes by applying implicit bridge entity masking and topology-driven hard negative sampling, forcing models to navigate biologically plausible distractors without relying on superficial elimination. Comprehensive evaluations of 21 LLMs reveal massive performance degradation on our multi-hop tasks, particularly among domain-specific models. Crucially, restoring the masked evidence via Retrieval-Augmented Generation (RAG) triggers near-universal performance recovery, validating ShatterMed-QA's structural fidelity and proving its efficacy in diagnosing the fundamental reasoning deficits of current medical AI. Explore the dataset, interactive examples, and full leaderboards at our project website: this https URL\n\n---\n\n## 113. Operationalising Cyber Risk Management Using AI: Connecting Cyber Incidents to MITRE ATT&CK Techniques, Security Controls, and Metrics\n- **Authors**: Emad Sherif , Iryna Yevseyeva , Vitor Basto-Fernandes , Allan Cook\n- **URL**: [https://arxiv.org/abs/2603.12455](https://arxiv.org/abs/2603.12455)\n- **Abstract**:\n> The escalating frequency of cyber-attacks poses significant challenges for organisations, particularly small enterprises constrained by limited in-house expertise, insufficient knowledge, and financial resources. This research presents a novel framework that leverages Natural Language Processing to address these challenges through automated mapping of cyber incidents to adversary techniques. We introduce the Cyber Catalog, a knowledge base that systematically integrates CIS Critical Security Controls, MITRE ATT&CK techniques, and SMART metrics. This integrated resource enables organisations to connect threat intelligence directly to actionable controls and measurable outcomes. To operationalise the framework, we fine-tuned all-mpnet-base-v2, a highly regarded sentence-transformers model used to convert text into numerical vectors on an augmented dataset comprising 74,986 incident-technique pairs to enhance semantic similarity between cyber incidents and MITRE ATT&CK techniques. Our fine-tuned model achieved a Spearman correlation of 0.7894 and Pearson correlation of 0.8756, demonstrating substantial improvements over top baseline models including all-mpnet-base-v2, all-distilroberta-v1, and all-MiniLM-L12-v2. Furthermore, our model exhibited significantly lower prediction errors (MAE = 0.135, MSE = 0.027) compared to all baseline models, confirming superior accuracy and consistency. The Cyber Catalog, training dataset, trained model, and implementation code made publicly available to facilitate further research and enable practical deployment in resource-constrained environments. This work bridges the gap between threat intelligence and operational security management, providing an actionable tool for systematic cyber incident response and evidence-based cyber risk management.\n\n---\n\n## 114. Unmasking Biases and Reliability Concerns in Convolutional Neural Networks Analysis of Cancer Pathology Images\n- **Authors**: Michael Okonoda , Eder Martinez , Abhilekha Dalal , Lior Shamir\n- **URL**: [https://arxiv.org/abs/2603.12445](https://arxiv.org/abs/2603.12445)\n- **Abstract**:\n> Convolutional Neural Networks have shown promising effectiveness in identifying different types of cancer from radiographs. However, the opaque nature of CNNs makes it difficult to fully understand the way they operate, limiting their assessment to empirical evaluation. Here we study the soundness of the standard practices by which CNNs are evaluated for the purpose of cancer pathology. Thirteen highly used cancer benchmark datasets were analyzed, using four common CNN architectures and different types of cancer, such as melanoma, carcinoma, colorectal cancer, and lung cancer. We compared the accuracy of each model with that of datasets made of cropped segments from the background of the original images that do not contain clinically relevant content. Because the rendered datasets contain no clinical information, the null hypothesis is that the CNNs should provide mere chance-based accuracy when classifying these datasets. The results show that the CNN models provided high accuracy when using the cropped segments, sometimes as high as 93\\%, even though they lacked biomedical information. These results show that some CNN architectures are more sensitive to bias than others. The analysis shows that the common practices of machine learning evaluation might lead to unreliable results when applied to cancer pathology. These biases are very difficult to identify, and might mislead researchers as they use available benchmark datasets to test the efficacy of CNN methods.\n\n---\n\n## 115. Revisiting Model Stitching In the Foundation Model Era\n- **Authors**: Zheda Mai , Ke Zhang , Fu-En Wang , Zixiao Ken Wang , Albert Y. C. Chen , Lu Xia , Min Sun , Wei-Lun Chao , Cheng-Hao Kuo\n- **URL**: [https://arxiv.org/abs/2603.12433](https://arxiv.org/abs/2603.12433)\n- **Abstract**:\n> Model stitching, connecting early layers of one model (source) to later layers of another (target) via a light stitch layer, has served as a probe of representational compatibility. Prior work finds that models trained on the same dataset remain stitchable (negligible accuracy drop) despite different initializations or objectives. We revisit stitching for Vision Foundation Models (VFMs) that vary in objectives, data, and modality mix (e.g., CLIP, DINOv2, SigLIP 2) and ask: Are heterogeneous VFMs stitchable? We introduce a systematic protocol spanning the stitch points, stitch layer families, training losses, and downstream tasks. Three findings emerge. (1) Stitch layer training matters: conventional approaches that match the intermediate features at the stitch point or optimize the task loss end-to-end struggle to retain accuracy, especially at shallow stitch points. (2) With a simple feature-matching loss at the target model's penultimate layer, heterogeneous VFMs become reliably stitchable across vision tasks. (3) For deep stitch points, the stitched model can surpass either constituent model at only a small inference overhead (for the stitch layer). Building on these findings, we further propose the VFM Stitch Tree (VST), which shares early layers across VFMs while retaining their later layers, yielding a controllable accuracy-latency trade-off for multimodal LLMs that often leverage multiple VFMs. Taken together, our study elevates stitching from a diagnostic probe to a practical recipe for integrating complementary VFM strengths and pinpointing where their representations align or diverge.\n\n---\n\n## 116. Test-Time Strategies for More Efficient and Accurate Agentic RAG\n- **Authors**: Brian Zhang , Deepti Guntur , Zhiyang Zuo , Abhinav Sharma , Shreyas Chaudhari , Wenlong Zhao , Franck Dernoncourt , Puneet Mathur , Ryan Rossi , Nedim Lipka\n- **URL**: [https://arxiv.org/abs/2603.12396](https://arxiv.org/abs/2603.12396)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) systems face challenges with complex, multihop questions, and agentic frameworks such as Search-R1 (Jin et al., 2025), which operates iteratively, have been proposed to address these complexities. However, such approaches can introduce inefficiencies, including repetitive retrieval of previously processed information and challenges in contextualizing retrieved results effectively within the current generation prompt. Such issues can lead to unnecessary retrieval turns, suboptimal reasoning, inaccurate answers, and increased token consumption. In this paper, we investigate test-time modifications to the Search-R1 pipeline to mitigate these identified shortcomings. Specifically, we explore the integration of two components and their combination: a contextualization module to better integrate relevant information from retrieved documents into reasoning, and a de-duplication module that replaces previously retrieved documents with the next most relevant ones. We evaluate our approaches using the HotpotQA (Yang et al., 2018) and the Natural Questions (Kwiatkowski et al., 2019) datasets, reporting the exact match (EM) score, an LLM-as-a-Judge assessment of answer correctness, and the average number of turns. Our best-performing variant, utilizing GPT-4.1-mini for contextualization, achieves a 5.6% increase in EM score and reduces the number of turns by 10.5% compared to the Search-R1 baseline, demonstrating improved answer accuracy and retrieval efficiency.\n\n---\n\n## 117. SPARROW: Learning Spatial Precision and Temporal Referential Consistency in Pixel-Grounded Video MLLMs\n- **Authors**: Mohamad Alansari , Naufal Suryanto , Divya Velayudhan , Sajid Javed , Naoufel Werghi , Muzammal Naseer\n- **URL**: [https://arxiv.org/abs/2603.12382](https://arxiv.org/abs/2603.12382)\n- **Abstract**:\n> Multimodal large language models (MLLMs) have advanced from image-level reasoning to pixel-level grounding, but extending these capabilities to videos remains challenging as models must achieve spatial precision and temporally consistent reference tracking. Existing video MLLMs often rely on a static segmentation token ([SEG]) for frame-wise grounding, which provides semantics but lacks temporal context, causing spatial drift, identity switches, and unstable initialization when objects move or reappear. We introduce SPARROW, a pixel-grounded video MLLM that unifies spatial accuracy and temporal stability through two key components: (i) Target-Specific Tracked Features (TSF), which inject temporally aligned referent cues during training, and (ii) a dual-prompt design that decodes box ([BOX]) and segmentation ([SEG]) tokens to fuse geometric priors with semantic grounding. SPARROW is supported by a curated referential video dataset of 30,646 videos and 45,231 Q&A pairs and operates end-to-end without external detectors via a class-agnostic SAM2-based proposer. Integrated into three recent open-source video MLLMs (UniPixel, GLUS, and VideoGLaMM), SPARROW delivers consistent gains across six benchmarks, improving up to +8.9 J&F on RVOS, +5 mIoU on visual grounding, and +5.4 CLAIR on GCG. These results demonstrate that SPARROW substantially improves referential stability, spatial precision, and temporal coherence in pixel-grounded video understanding. Project page: this https URL\n\n---\n\n## 118. Budget-Sensitive Discovery Scoring: A Formally Verified Framework for Evaluating AI-Guided Scientific Selection\n- **Authors**: Abhinaba Basu , Pavan Chakraborty\n- **URL**: [https://arxiv.org/abs/2603.12349](https://arxiv.org/abs/2603.12349)\n- **Abstract**:\n> Scientific discovery increasingly relies on AI systems to select candidates for expensive experimental validation, yet no principled, budget-aware evaluation framework exists for comparing selection strategies -- a gap intensified by large language models (LLMs), which generate plausible scientific proposals without reliable downstream evaluation. We introduce the Budget-Sensitive Discovery Score (BSDS), a formally verified metric -- 20 theorems machine-checked by the Lean 4 proof assistant -- that jointly penalizes false discoveries (lambda-weighted FDR) and excessive abstention (gamma-weighted coverage gap) at each budget level. Its budget-averaged form, the Discovery Quality Score (DQS), provides a single summary statistic that no proposer can inflate by performing well at a cherry-picked budget. As a case study, we apply BSDS/DQS to: do LLMs add marginal value to an existing ML pipeline for drug discovery candidate selection? We evaluate 39 proposers -- 11 mechanistic variants, 14 zero-shot LLM configurations, and 14 few-shot LLM configurations -- using SMILES representations on MoleculeNet HIV (41,127 compounds, 3.5% active, 1,000 bootstrap replicates) under both random and scaffold splits. Three findings emerge. First, the simple RF-based Greedy-ML proposer achieves the best DQS (-0.046), outperforming all MLP variants and LLM configurations. Second, no LLM surpasses the Greedy-ML baseline under zero-shot or few-shot evaluation on HIV or Tox21, establishing that LLMs provide no marginal value over an existing trained classifier. Third, the proposer hierarchy generalizes across five MoleculeNet benchmarks spanning 0.18%-46.2% prevalence, a non-drug AV safety domain, and a 9x7 grid of penalty parameters (tau >= 0.636, mean tau = 0.863). The framework applies to any setting where candidates are selected under budget constraints and asymmetric error costs.\n\n---\n\n## 119. Optimizing Task Completion Time Updates Using POMDPs\n- **Authors**: Duncan Eddy , Esen Yel , Emma Passmore , Niles Egan , Grayson Armour , Dylan M. Asmar , Mykel J. Kochenderfer\n- **URL**: [https://arxiv.org/abs/2603.12340](https://arxiv.org/abs/2603.12340)\n- **Abstract**:\n> Managing announced task completion times is a fundamental control problem in project management. While extensive research exists on estimating task durations and task scheduling, the problem of when and how to update completion times communicated to stakeholders remains understudied. Organizations must balance announcement accuracy against the costs of frequent timeline updates, which can erode stakeholder trust and trigger costly replanning. Despite the prevalence of this problem, current approaches rely on static predictions or ad-hoc policies that fail to account for the sequential nature of announcement management. In this paper, we formulate the task announcement problem as a Partially Observable Markov Decision Process (POMDP) where the control policy must decide when to update announced completion times based on noisy observations of true task completion. Since most state variables (current time and previous announcements) are fully observable, we leverage the Mixed Observability MDP (MOMDP) framework to enable more efficient policy optimization. Our reward structure captures the dual costs of announcement errors and update frequency, enabling synthesis of optimal announcement control policies. Using off-the-shelf solvers, we generate policies that act as feedback controllers, adaptively managing announcements based on belief state evolution. Simulation results demonstrate significant improvements in both accuracy and announcement stability compared to baseline strategies, achieving up to 75\\% reduction in unnecessary updates while maintaining or improving prediction accuracy.\n\n---\n\n## 120. Maximum Entropy Exploration Without the Rollouts\n- **Authors**: Jacob Adamczyk , Adam Kamoski , Rahul V. Kulkarni\n- **URL**: [https://arxiv.org/abs/2603.12325](https://arxiv.org/abs/2603.12325)\n- **Abstract**:\n> Efficient exploration remains a central challenge in reinforcement learning, serving as a useful pretraining objective for data collection, particularly when an external reward function is unavailable. A principled formulation of the exploration problem is to find policies that maximize the entropy of their induced steady-state visitation distribution, thereby encouraging uniform long-run coverage of the state space. Many existing exploration approaches require estimating state visitation frequencies through repeated on-policy rollouts, which can be computationally expensive. In this work, we instead consider an intrinsic average-reward formulation in which the reward is derived from the visitation distribution itself, so that the optimal policy maximizes steady-state entropy. An entropy-regularized version of this objective admits a spectral characterization: the relevant stationary distributions can be computed from the dominant eigenvectors of a problem-dependent transition matrix. This insight leads to a novel algorithm for solving the maximum entropy exploration problem, EVE (EigenVector-based Exploration), which avoids explicit rollouts and distribution estimation, instead computing the solution through iterative updates, similar to a value-based approach. To address the original unregularized objective, we employ a posterior-policy iteration (PPI) approach, which monotonically improves the entropy and converges in value. We prove convergence of EVE under standard assumptions and demonstrate empirically that it efficiently produces policies with high steady-state entropy, achieving competitive exploration performance relative to rollout-based baselines in deterministic grid-world environments.\n\n---\n\n## 121. Thermodynamics of Reinforcement Learning Curricula\n- **Authors**: Jacob Adamczyk , Juan Sebastian Rojas , Rahul V. Kulkarni\n- **URL**: [https://arxiv.org/abs/2603.12324](https://arxiv.org/abs/2603.12324)\n- **Abstract**:\n> Connections between statistical mechanics and machine learning have repeatedly proven fruitful, providing insight into optimization, generalization, and representation learning. In this work, we follow this tradition by leveraging results from non-equilibrium thermodynamics to formalize curriculum learning in reinforcement learning (RL). In particular, we propose a geometric framework for RL by interpreting reward parameters as coordinates on a task manifold. We show that, by minimizing the excess thermodynamic work, optimal curricula correspond to geodesics in this task space. As an application of this framework, we provide an algorithm, \"MEW\" (Minimum Excess Work), to derive a principled schedule for temperature annealing in maximum-entropy RL.\n\n---\n\n## 122. VQQA: An Agentic Approach for Video Evaluation and Quality Improvement\n- **Authors**: Yiwen Song , Tomas Pfister , Yale Song\n- **URL**: [https://arxiv.org/abs/2603.12310](https://arxiv.org/abs/2603.12310)\n- **Abstract**:\n> Despite rapid advancements in video generation models, aligning their outputs with complex user intent remains challenging. Existing test-time optimization methods are typically either computationally expensive or require white-box access to model internals. To address this, we present VQQA (Video Quality Question Answering), a unified, multi-agent framework generalizable across diverse input modalities and video generation tasks. By dynamically generating visual questions and using the resulting Vision-Language Model (VLM) critiques as semantic gradients, VQQA replaces traditional, passive evaluation metrics with human-interpretable, actionable feedback. This enables a highly efficient, closed-loop prompt optimization process via a black-box natural language interface. Extensive experiments demonstrate that VQQA effectively isolates and resolves visual artifacts, substantially improving generation quality in just a few refinement steps. Applicable to both text-to-video (T2V) and image-to-video (I2V) tasks, our method achieves absolute improvements of +11.57% on T2V-CompBench and +8.43% on VBench2 over vanilla generation, significantly outperforming state-of-the-art stochastic search and prompt optimization techniques.\n\n---\n\n## 123. HCP-DCNet: A Hierarchical Causal Primitive Dynamic Composition Network for Self-Improving Causal Understanding\n- **Authors**: Ming Lei , Shufan Wu , Christophe Baehr\n- **URL**: [https://arxiv.org/abs/2603.12305](https://arxiv.org/abs/2603.12305)\n- **Abstract**:\n> The ability to understand and reason about cause and effect -- encompassing interventions, counterfactuals, and underlying mechanisms -- is a cornerstone of robust artificial intelligence. While deep learning excels at pattern recognition, it fundamentally lacks a model of causality, making systems brittle under distribution shifts and unable to answer ``what-if'' questions. This paper introduces the \\emph{Hierarchical Causal Primitive Dynamic Composition Network (HCP-DCNet)}, a unified framework that bridges continuous physical dynamics with discrete symbolic causal inference. Departing from monolithic representations, HCP-DCNet decomposes causal scenes into reusable, typed \\emph{causal primitives} organized into four abstraction layers: physical, functional, event, and rule. A dual-channel routing network dynamically composes these primitives into task-specific, fully differentiable \\emph{Causal Execution Graphs (CEGs)}. Crucially, the system employs a \\emph{causal-intervention-driven meta-evolution} strategy, enabling autonomous self-improvement through a constrained Markov decision process. We establish rigorous theoretical guarantees, including type-safe composition, routing convergence, and universal approximation of causal dynamics. Extensive experiments across simulated physical and social environments demonstrate that HCP-DCNet significantly outperforms state-of-the-art baselines in causal discovery, counterfactual reasoning, and compositional generalization. This work provides a principled, scalable, and interpretable architecture for building AI systems with human-like causal abstraction and continual self-refinement capabilities.\n\n---\n\n## 124. A Geometrically-Grounded Drive for MDL-Based Optimization in Deep Learning\n- **Authors**: Ming Lei , Shufan Wu , Christophe Baehr\n- **URL**: [https://arxiv.org/abs/2603.12304](https://arxiv.org/abs/2603.12304)\n- **Abstract**:\n> This paper introduces a novel optimization framework that fundamentally integrates the Minimum Description Length (MDL) principle into the training dynamics of deep neural networks. Moving beyond its conventional role as a model selection criterion, we reformulate MDL as an active, adaptive driving force within the optimization process itself. The core of our method is a geometrically-grounded cognitive manifold whose evolution is governed by a \\textit{coupled Ricci flow}, enriched with a novel \\textit{MDL Drive} term derived from first principles. This drive, modulated by the task-loss gradient, creates a seamless harmony between data fidelity and model simplification, actively compressing the internal representation during training. We establish a comprehensive theoretical foundation, proving key properties including the monotonic decrease of description length (Theorem~\\ref{thm:convergence}), a finite number of topological phase transitions via a geometric surgery protocol (Theorems~\\ref{thm:surgery}, \\ref{thm:ultimate_fate}), and the emergence of universal critical behavior (Theorem~\\ref{thm:universality}). Furthermore, we provide a practical, computationally efficient algorithm with $O(N \\log N)$ per-iteration complexity (Theorem~\\ref{thm:complexity}), alongside guarantees for numerical stability (Theorem~\\ref{thm:stability}) and exponential convergence under convexity assumptions (Theorem~\\ref{thm:convergence_rate}). Empirical validation on synthetic regression and classification tasks confirms the theoretical predictions, demonstrating the algorithm's efficacy in achieving robust generalization and autonomous model simplification. This work provides a principled path toward more autonomous, generalizable, and interpretable AI systems by unifying geometric deep learning with information-theoretic principles.\n\n---\n\n## 125. Global Evolutionary Steering: Refining Activation Steering Control via Cross-Layer Consistency\n- **Authors**: Xinyan Jiang , Wenjing Yu , Di Wang , Lijie Hu\n- **URL**: [https://arxiv.org/abs/2603.12298](https://arxiv.org/abs/2603.12298)\n- **Abstract**:\n> Activation engineering enables precise control over Large Language Models (LLMs) without the computational cost of fine-tuning. However, existing methods deriving vectors from static activation differences are susceptible to high-dimensional noise and layer-wise semantic drift, often capturing spurious correlations rather than the target intent. To address this, we propose Global Evolutionary Refined Steering (GER-steer), a training-free framework that grounded in the geometric stability of the network's representation evolution. GER-steer exploits this global signal to rectify raw steering vectors, effectively decoupling robust semantic intent from orthogonal artifacts. Extensive evaluations confirm that GER-steer consistently outperforms baselines, delivering superior efficacy and generalization without layer-specific tuning, establishing a universal solution for reliable model alignment.\n\n---\n\n## 126. Synthetic Data Generation for Brain-Computer Interfaces: Overview, Benchmarking, and Future Directions\n- **Authors**: Ziwei Wang , Zhentao He , Xingyi He , Hongbin Wang , Tianwang Jia , Jingwei Luo , Siyang Li , Xiaoqing Chen , Dongrui Wu\n- **URL**: [https://arxiv.org/abs/2603.12296](https://arxiv.org/abs/2603.12296)\n- **Abstract**:\n> Deep learning has achieved transformative performance across diverse domains, largely driven by the large-scale, high-quality training data. In contrast, the development of brain-computer interfaces (BCIs) is fundamentally constrained by the limited, heterogeneous, and privacy-sensitive neural recordings. Generating synthetic yet physiologically plausible brain signals has therefore emerged as a compelling way to mitigate data scarcity and enhance model capacity. This survey provides a comprehensive review of brain signal generation for BCIs, covering methodological taxonomies, benchmark experiments, evaluation metrics, and key applications. We systematically categorize existing generative algorithms into four types: knowledge-based, feature-based, model-based, and translation-based approaches. Furthermore, we benchmark existing brain signal generation approaches across four representative BCI paradigms to provide an objective performance comparison. Finally, we discuss the potentials and challenges of current generation approaches and prospect future research on accurate, data-efficient, and privacy-aware BCI systems. The benchmark codebase is publicized at this https URL .\n\n---\n\n## 127. Detecting Miscitation on the Scholarly Web through LLM-Augmented Text-Rich Graph Learning\n- **Authors**: Huidong Wu , Haojia Xiang , Jingtong Gao , Xiangyu Zhao , Dengsheng Wu , Jianping Li\n- **URL**: [https://arxiv.org/abs/2603.12290](https://arxiv.org/abs/2603.12290)\n- **Abstract**:\n> Scholarly web is a vast network of knowledge connected by citations. However, this system is increasingly compromised by miscitation, where references do not support or even contradict the claims they are cited for. Current miscitation detection methods, which primarily rely on semantic similarity or network anomalies, struggle to capture the nuanced relationship between a citation's context and its place in the wider network. While large language models (LLMs) offer powerful capabilities in semantic reasoning for this task, their deployment is hindered by hallucination risks and high computational costs. In this work, we introduce LLM-Augmented Graph Learning-based Miscitation Detector (LAGMiD), a novel framework that leverages LLMs for deep semantic reasoning over citation graphs and distills this knowledge into graph neural networks (GNNs) for efficient and scalable miscitation detection. Specifically, LAGMiD introduces an evidence-chain reasoning mechanism, which uses chain-of-thought prompting, to perform multi-hop citation tracing and assess semantic fidelity. To reduce LLM inference costs, we design a knowledge distillation method aligning GNN embeddings with intermediate LLM reasoning states. A collaborative learning strategy further routes complex cases to the LLM while optimizing the GNN for structure-based generalization. Experiments on three real-world benchmarks show that LAGMiD achieves state-of-the-art miscitation detection with significantly reduced inference cost.\n\n---\n\n## 128. From Garbage to Gold: A Data-Architectural Theory of Predictive Robustness\n- **Authors**: Terrence J. Lee-St. John , Jordan L. Lawson , Bartlomiej Piechowski-Jozwiak\n- **URL**: [https://arxiv.org/abs/2603.12288](https://arxiv.org/abs/2603.12288)\n- **Abstract**:\n> Tabular machine learning presents a paradox: modern models achieve state-of-the-art performance using high-dimensional (high-D), collinear, error-prone data, defying the \"Garbage In, Garbage Out\" mantra. To help resolve this, we synthesize principles from Information Theory, Latent Factor Models, and Psychometrics, clarifying that predictive robustness arises not solely from data cleanliness, but from the synergy between data architecture and model capacity. Partitioning predictor-space \"noise\" into \"Predictor Error\" and \"Structural Uncertainty\" (informational deficits from stochastic generative mappings), we prove that leveraging high-D sets of error-prone predictors asymptotically overcomes both types of noise, whereas cleaning a low-D set is fundamentally bounded by Structural Uncertainty. We demonstrate why \"Informative Collinearity\" (dependencies from shared latent causes) enhances reliability and convergence efficiency, and explain why increased dimensionality reduces the latent inference burden, enabling feasibility with finite samples. To address practical constraints, we propose \"Proactive Data-Centric AI\" to identify predictors that enable robustness efficiently. We also derive boundaries for Systematic Error Regimes and show why models that absorb \"rogue\" dependencies can mitigate assumption violations. Linking latent architecture to Benign Overfitting, we offer a first step towards a unified view of robustness to Outcome Error and predictor-space noise, while also delineating when traditional DCAI's focus on label cleaning remains powerful. By redefining data quality from item-level perfection to portfolio-level architecture, we provide a theoretical rationale for \"Local Factories\" -- learning from live, uncurated enterprise \"data swamps\" -- supporting a deployment paradigm shift from \"Model Transfer\" to \"Methodology Transfer'' to overcome static generalizability limitations.\n\n---\n\n## 129. The DIME Architecture: A Unified Operational Algorithm for Neural Representation, Dynamics, Control and Integration\n- **Authors**: Ionel Cristian Vladu , Nicu Bizdoaca , Ionica Pirici , Tudor-Adrian Balseanu , Eduard Nicusor Bondoc\n- **URL**: [https://arxiv.org/abs/2603.12286](https://arxiv.org/abs/2603.12286)\n- **Abstract**:\n> Modern neuroscience has accumulated extensive evidence on perception, memory, prediction, valuation, and consciousness, yet still lacks an explicit operational architecture capable of integrating these phenomena within a unified computational framework. Existing theories address specific aspects of neural function: predictive coding and active inference emphasize hierarchical inference and prediction error minimization; engram theories explain memory through distributed cell assemblies; neuromodulatory accounts focus on value-dependent regulation of plasticity and behaviour; and global workspace or large-scale network models investigate mechanisms underlying conscious access. Despite their explanatory power, these approaches remain only partially integrated at the architectural level. This work introduces DIME (Detect-Integrate-Mark-Execute), a neural architecture organizing perception, memory, valuation, and conscious access within a common operational cycle. The framework includes four interacting components: engrams, distributed recurrent neural structures supporting multiple activation trajectories; execution threads, spatiotemporal trajectories implementing neural processes; marker systems, neuromodulatory and limbic mechanisms regulating gain, plasticity, and trajectory selection; and hyperengrams, large-scale integrative states associated with operational conscious access. The framework is consistent with empirical evidence from hippocampal indexing, recurrent cortical processing, replay phenomena, large-scale network integration, and neuromodulatory regulation. Formulated at an abstract computational level, DIME may also inform artificial intelligence and robotics by providing an architectural template in which representation, valuation, and temporal sequencing emerge from a unified mechanism. An extended theoretical exposition is available in a companion monograph on Zenodo.\n\n---\n\n## 130. Predictive Analytics for Foot Ulcers Using Time-Series Temperature and Pressure Data\n- **Authors**: Md Tanvir Hasan Turja\n- **URL**: [https://arxiv.org/abs/2603.12278](https://arxiv.org/abs/2603.12278)\n- **Abstract**:\n> Diabetic foot ulcers (DFUs) are a severe complication of diabetes, often resulting in significant morbidity. This paper presents a predictive analytics framework utilizing time-series data captured by wearable foot sensors -- specifically NTC thin-film thermocouples for temperature measurement and FlexiForce pressure sensors for plantar load monitoring. Data was collected from healthy subjects walking on an instrumented pathway. Unsupervised machine learning algorithms, Isolation Forest and K-Nearest Neighbors (KNN), were applied to detect anomalies that may indicate early ulcer risk. Through rigorous data preprocessing and targeted feature engineering, physiologic patterns were extracted to identify subtle changes in foot temperature and pressure. Results demonstrate Isolation Forest is sensitive to micro-anomalies, while KNN is effective in flagging extreme deviations, albeit at a higher false-positive rate. Strong correlations between temperature and pressure readings support combined sensor monitoring for improved predictive accuracy. These findings provide a basis for real-time diabetic foot health surveillance, aiming to facilitate earlier intervention and reduce DFU incidence.\n\n---\n\n## 131. Prompt Injection as Role Confusion\n- **Authors**: Charles Ye , Jasmine Cui , Dylan Hadfield-Menell\n- **URL**: [https://arxiv.org/abs/2603.12277](https://arxiv.org/abs/2603.12277)\n- **Abstract**:\n> Language models remain vulnerable to prompt injection attacks despite extensive safety training. We trace this failure to role confusion: models infer roles from how text is written, not where it comes from. We design novel role probes to capture how models internally identify \"who is speaking.\" These reveal why prompt injection works: untrusted text that imitates a role inherits that role's authority. We test this insight by injecting spoofed reasoning into user prompts and tool outputs, achieving average success rates of 60% on StrongREJECT and 61% on agent exfiltration, across multiple open- and closed-weight models with near-zero baselines. Strikingly, the degree of internal role confusion strongly predicts attack success before generation begins. Our findings reveal a fundamental gap: security is defined at the interface but authority is assigned in latent space. More broadly, we introduce a unifying, mechanistic framework for prompt injection, demonstrating that diverse prompt-injection attacks exploit the same underlying role-confusion mechanism.\n\n---\n\n## 132. Aligning Language Models from User Interactions\n- **Authors**: Thomas Kleine Buening , Jonas Hübotter , Barna Pásztor , Idan Shenfeld , Giorgia Ramponi , Andreas Krause\n- **URL**: [https://arxiv.org/abs/2603.12273](https://arxiv.org/abs/2603.12273)\n- **Abstract**:\n> Multi-turn user interactions are among the most abundant data produced by language models, yet we lack effective methods to learn from them. While typically discarded, these interactions often contain useful information: follow-up user messages may indicate that a response was incorrect, failed to follow an instruction, or did not align with the user's preferences. Importantly, language models are already able to make use of this information in context. After observing a user's follow-up, the same model is often able to revise its behavior. We leverage this ability to propose a principled and scalable method for learning directly from user interactions through self-distillation. By conditioning the model on the user's follow-up message and comparing the resulting token distribution with the original policy, we obtain a target for updating the policy that captures how the model's behavior changes in hindsight. We then distill this hindsight distribution back into the current policy. Remarkably, we show that training on real-world user conversations from WildChat improves language models across standard alignment and instruction-following benchmarks, without regressing other capabilities. The same mechanism enables personalization, allowing models to continually adapt to individual users through interaction without explicit feedback. Our results demonstrate that raw user interactions that arise naturally during deployment enable alignment, personalization, and continual adaptation.\n\n---\n\n## 133. Diagnosing Retrieval Bias Under Multiple In-Context Knowledge Updates in Large Language Models\n- **Authors**: Boyu Qiao , Sean Guo , Xian Yang , Kun Li , Wei Zhou , Songlin Hu , Yunya Song\n- **URL**: [https://arxiv.org/abs/2603.12271](https://arxiv.org/abs/2603.12271)\n- **Abstract**:\n> LLMs are widely used in knowledge-intensive tasks where the same fact may be revised multiple times within context. Unlike prior work focusing on one-shot updates or single conflicts, multi-update scenarios contain multiple historically valid versions that compete at retrieval, yet remain underexplored. This challenge resembles the AB-AC interference paradigm in cognitive psychology: when the same cue A is successively associated with B and C, the old and new associations compete during retrieval, leading to bias. Inspired by this, we introduce a Dynamic Knowledge Instance (DKI) evaluation framework, modeling multi-updates of the same fact as a cue paired with a sequence of updated values, and assess models via endpoint probing of the earliest (initial) and latest (current) states. Across diverse LLMs, we observe that retrieval bias intensifies as updates increase, earliest-state accuracy stays high while latest-state accuracy drops substantially. Diagnostic analyses of attention, hidden-state similarity, and output logits further reveal that these signals become flatter and weakly discriminative on errors, providing little stable basis for identifying the latest update. Finally, cognitively inspired heuristic intervention strategies yield only modest gains and do not eliminate the bias. Our results reveal a persistent challenge in tracking and following knowledge updates in long contexts.\n\n---\n\n## 134. Task-Specific Knowledge Distillation via Intermediate Probes\n- **Authors**: Ryan Brown , Chris Russell\n- **URL**: [https://arxiv.org/abs/2603.12270](https://arxiv.org/abs/2603.12270)\n- **Abstract**:\n> Knowledge distillation from large language models (LLMs) assumes that the teacher's output distribution is a high-quality training signal. On reasoning tasks, this assumption is frequently violated. A model's intermediate representations may encode the correct answer, yet this information is lost or distorted through the vocabulary projection, where prompt formatting and answer-token choices creates brittle, noisy outputs. We introduce \\method{}, a distillation framework that bypasses this bottleneck by training lightweight probes on frozen teacher hidden states and using the probe's predictions, rather than output logits, as supervision for student training. This simple change yields consistent improvements across four reasoning benchmarks (AQuA-RAT, ARC Easy/Challenge, and MMLU), with gains most pronounced under limited data. Probes trained on intermediate representations provide cleaner labels than the teacher's own outputs, effectively denoising the distillation signal. \\method{} requires no architectural changes to student or teacher, is architecture-agnostic, and adds minimal compute since probe training is cheap and teacher representations can be cached. By exploiting internal representations, \\method{} enables practitioners to extract more value from large teacher models without additional training data or architectural complexity.\n\n---\n\n## 135. DART: Input-Difficulty-AwaRe Adaptive Threshold for Early-Exit DNNs\n- **Authors**: Parth Patne , Mahdi Taheri , Christian Herglotz , Maksim Jenihhin , Milos Krstic , Michael Hübner\n- **URL**: [https://arxiv.org/abs/2603.12269](https://arxiv.org/abs/2603.12269)\n- **Abstract**:\n> Early-exit deep neural networks enable adaptive inference by terminating computation when sufficient confidence is achieved, reducing cost for edge AI accelerators in resource-constrained settings. Existing methods, however, rely on suboptimal exit policies, ignore input difficulty, and optimize thresholds independently. This paper introduces DART (Input-Difficulty-Aware Adaptive Threshold), a framework that overcomes these limitations. DART introduces three key innovations: (1) a lightweight difficulty estimation module that quantifies input complexity with minimal computational overhead, (2) a joint exit policy optimization algorithm based on dynamic programming, and (3) an adaptive coefficient management system. Experiments on diverse DNN benchmarks (AlexNet, ResNet-18, VGG-16) demonstrate that DART achieves up to \\textbf{3.3$\\times$} speedup, \\textbf{5.1$\\times$} lower energy, and up to \\textbf{42\\%} lower average power compared to static networks, while preserving competitive accuracy. Extending DART to Vision Transformers (LeViT) yields power (5.0$\\times$) and execution-time (3.6$\\times$) gains but also accuracy loss (up to 17 percent), underscoring the need for transformer-specific early-exit mechanisms. We further introduce the Difficulty-Aware Efficiency Score (DAES), a novel multi-objective metric, under which DART achieves up to a 14.8 improvement over baselines, highlighting superior accuracy, efficiency, and robustness trade-offs.\n\n---\n\n{% endraw %}\n"}, "ae95e160c1986855bbe27edfadb3713b4c8f4a25": {"t": 1786068556.9460013, "n": 229888, "h": "8c5adbd6d55505f6", "c": "{% raw %}\n# 전체 AI 논문 - 2026-03-17\n\n## 1. Semantic Invariance in Agentic AI\n- **Authors**: I. de Zarzà , J. de Curtò , Jordi Cabot , Pietro Manzoni , Carlos T. Calafate\n- **URL**: [https://arxiv.org/abs/2603.13173](https://arxiv.org/abs/2603.13173)\n- **Abstract**:\n> Large Language Models (LLMs) increasingly serve as autonomous reasoning agents in decision support, scientific problem-solving, and multi-agent coordination systems. However, deploying LLM agents in consequential applications requires assurance that their reasoning remains stable under semantically equivalent input variations, a property we term semantic invariance. Standard benchmark evaluations, which assess accuracy on fixed, canonical problem formulations, fail to capture this critical reliability dimension. To address this shortcoming, in this paper we present a metamorphic testing framework for systematically assessing the robustness of LLM reasoning agents, applying eight semantic-preserving transformations (identity, paraphrase, fact reordering, expansion, contraction, academic context, business context, and contrastive formulation) across seven foundation models spanning four distinct architectural families: Hermes (70B, 405B), Qwen3 (30B-A3B, 235B-A22B), DeepSeek-R1, and gpt-oss (20B, 120B). Our evaluation encompasses 19 multi-step reasoning problems across eight scientific domains. The results reveal that model scale does not predict robustness: the smaller Qwen3-30B-A3B achieves the highest stability (79.6% invariant responses, semantic similarity 0.91), while larger models exhibit greater fragility.\n\n---\n\n## 2. Developing and evaluating a chatbot to support maternal health care\n- **Authors**: Smriti Jha , Vidhi Jain , Jianyu Xu , Grace Liu , Sowmya Ramesh , Jitender Nagpal , Gretchen Chapman , Benjamin Bellows , Siddhartha Goyal , Aarti Singh , Bryan Wilder\n- **URL**: [https://arxiv.org/abs/2603.13168](https://arxiv.org/abs/2603.13168)\n- **Abstract**:\n> The ability to provide trustworthy maternal health information using phone-based chatbots can have a significant impact, particularly in low-resource settings where users have low health literacy and limited access to care. However, deploying such systems is technically challenging: user queries are short, underspecified, and code-mixed across languages, answers require regional context-specific grounding, and partial or missing symptom context makes safe routing decisions difficult. We present a chatbot for maternal health in India developed through a partnership between academic researchers, a health tech company, a public health nonprofit, and a hospital. The system combines (1) stage-aware triage, routing high-risk queries to expert templates, (2) hybrid retrieval over curated maternal/newborn guidelines, and (3) evidence-conditioned generation from an LLM. Our core contribution is an evaluation workflow for high-stakes deployment under limited expert supervision. Targeting both component-level and end-to-end testing, we introduce: (i) a labeled triage benchmark (N=150) achieving 86.7% emergency recall, explicitly reporting the missed-emergency vs. over-escalation trade-off; (ii) a synthetic multi-evidence retrieval benchmark (N=100) with chunk-level evidence labels; (iii) LLM-as-judge comparison on real queries (N=781) using clinician-codesigned criteria; and (iv) expert validation. Our findings show that trustworthy medical assistants in multilingual, noisy settings require defense-in-depth design paired with multi-method evaluation, rather than any single model and evaluation method choice.\n\n---\n\n## 3. When Right Meets Wrong: Bilateral Context Conditioning with Reward-Confidence Correction for GRPO\n- **Authors**: Yu Li , Tian Lan , Zhengling Qi\n- **URL**: [https://arxiv.org/abs/2603.13134](https://arxiv.org/abs/2603.13134)\n- **Abstract**:\n> Group Relative Policy Optimization (GRPO) has emerged as an effective method for training reasoning models. While it computes advantages based on group mean, GRPO treats each output as an independent sample during the optimization and overlooks a vital structural signal: the natural contrast between correct and incorrect solutions within the same group, thus ignoring the rich, comparative data that could be leveraged by explicitly pitting successful reasoning traces against failed ones. To capitalize on this, we present a contrastive reformulation of GRPO, showing that the GRPO objective implicitly maximizes the margin between the policy ratios of correct and incorrect samples. Building on this insight, we propose Bilateral Context Conditioning (BICC), a mechanism that allows the model to cross-reference successful and failed reasoning traces during the optimization, enabling a direct information flow across samples. We further introduce Reward-Confidence Correction (RCC) to stabilize training by dynamically adjusts the advantage baseline in GRPO using reward-confidence covariance derived from the first-order approximation of the variance-minimizing estimator. Both mechanisms require no additional sampling or auxiliary models and can be adapted to all GRPO variants. Experiments on mathematical reasoning benchmarks demonstrate consistent improvements across comprehensive models and algorithms. Code is available at \\href{ this https URL }{ this https URL }.\n\n---\n\n## 4. Steve-Evolving: Open-World Embodied Self-Evolution via Fine-Grained Diagnosis and Dual-Track Knowledge Distillation\n- **Authors**: Zhengwei Xie , Zhisheng Chen , Ziyan Weng , Tingyu Wu , Chenglong Li , Vireo Zhang , Kun Wang\n- **URL**: [https://arxiv.org/abs/2603.13131](https://arxiv.org/abs/2603.13131)\n- **Abstract**:\n> Open-world embodied agents must solve long-horizon tasks where the main bottleneck is not single-step planning quality but how interaction experience is organized and evolved. To this end, we present Steve-Evolving, a non-parametric self-evolving framework that tightly couples fine-grained execution diagnosis with dual-track knowledge distillation in a closed loop. The method follows three phases: Experience Anchoring, Experience Distillation, and Knowledge-Driven Closed-Loop Control. In detail, Experience Anchoring solidifies each subgoal attempt into a structured experience tuple with a fixed schema (pre-state, action, diagnosis-result, and post-state) and organizes it in a three-tier experience space with multi-dimensional indices (e.g., condition signatures, spatial hashing, and semantic tags) plus rolling summarization for efficient and auditable recall. To ensure sufficient information density for attribution, the execution layer provides compositional diagnosis signals beyond binary outcomes, including state-difference summaries, enumerated failure causes, continuous indicators, and stagnation/loop detection. Moreover, successful trajectories of Experience Distillation are generalized into reusable skills with explicit preconditions and verification criteria, while failures are distilled into executable guardrails that capture root causes and forbid risky operations at both subgoal and task granularities. Besides, Knowledge-Driven Closed-Loop Control retrieved skills and guardrails are injected into an LLM planner, and diagnosis-triggered local replanning updates the active constraints online, forming a continual evolution process without any model parameter updates. Experiments on the long-horizon suite of Minecraft MCU demonstrate consistent improvements over static-retrieval baselines.\n\n---\n\n## 5. Beyond Final Answers: CRYSTAL Benchmark for Transparent Multimodal Reasoning Evaluation\n- **Authors**: Wayner Barrios , SouYoung Jin\n- **URL**: [https://arxiv.org/abs/2603.13099](https://arxiv.org/abs/2603.13099)\n- **Abstract**:\n> We introduce CRYSTAL (Clear Reasoning via Yielded Steps, Traceability, and Logic), a diagnostic benchmark with 6,372 instances that evaluates multimodal reasoning through verifiable intermediate steps. We propose two complementary metrics: Match F1, which scores step-level precision and recall via semantic similarity matching, and Ordered Match F1, which further penalizes disordered reasoning chains. References are constructed through a Delphi-inspired pipeline in which four independent MLLMs generate trajectories, which are then aggregated via semantic clustering and validated through human quality gates. Evaluation of 20 MLLMs, including commercial frontier systems not used during benchmark construction, reveals systematic failures that are invisible to answer accuracy: universal cherry-picking (precision far exceeds recall), non-monotonic scaling trade-offs, and disordered reasoning in which no competitive model preserves more than 60% of matched steps in the correct order. Beyond evaluation, we propose the Causal Process Reward (CPR), a multiplicative reward that couples answer correctness with step-level alignment, and CPR-Curriculum, which progressively increases reasoning difficulty during training. CPR-Curriculum achieves a 32% improvement in Match F1 via GRPO where additive reward strategies fail, improving reasoning without manual step annotation.\n\n---\n\n## 6. Structured Distillation for Personalized Agent Memory: 11x Token Reduction with Retrieval Preservation\n- **Authors**: Sydney Lewis\n- **URL**: [https://arxiv.org/abs/2603.13017](https://arxiv.org/abs/2603.13017)\n- **Abstract**:\n> Long conversations with an AI agent create a simple problem for one user: the history is useful, but carrying it verbatim is expensive. We study personalized agent memory: one user's conversation history with an agent, distilled into a compact retrieval layer for later search. Each exchange is compressed into a compound object with four fields (exchange_core, specific_context, thematic room_assignments, and regex-extracted files_touched). The searchable distilled text averages 38 tokens per exchange. Applied to 4,182 conversations (14,340 exchanges) from 6 software engineering projects, the method reduces average exchange length from 371 to 38 tokens, yielding 11x compression. We evaluate whether personalized recall survives that compression using 201 recall-oriented queries, 107 configurations spanning 5 pure and 5 cross-layer search modes, and 5 LLM graders (214,519 consensus-graded query-result pairs). The best pure distilled configuration reaches 96% of the best verbatim MRR (0.717 vs 0.745). Results are mechanism-dependent. All 20 vector search configurations remain non-significant after Bonferroni correction, while all 20 BM25 configurations degrade significantly (effect sizes |d|=0.031-0.756). The best cross-layer setup slightly exceeds the best pure verbatim baseline (MRR 0.759). Structured distillation compresses single-user agent memory without uniformly sacrificing retrieval quality. At 1/11 the context cost, thousands of exchanges fit within a single prompt while the verbatim source remains available for drill-down. We release the implementation and analysis pipeline as open-source software.\n\n---\n\n## 7. Efficient and Interpretable Multi-Agent LLM Routing via Ant Colony Optimization\n- **Authors**: Xudong Wang , Chaoning Zhang , Jiaquan Zhang , Chenghao Li , Qigan Sun , Sung-Ho Bae , Peng Wang , Ning Xie , Jie Zou , Yang Yang , Hengtao Shen\n- **URL**: [https://arxiv.org/abs/2603.12933](https://arxiv.org/abs/2603.12933)\n- **Abstract**:\n> Large Language Model (LLM)-driven Multi-Agent Systems (MAS) have demonstrated strong capability in complex reasoning and tool use, and heterogeneous agent pools further broaden the quality--cost trade-off space. Despite these advances, real-world deployment is often constrained by high inference cost, latency, and limited transparency, which hinders scalable and efficient routing. Existing routing strategies typically rely on expensive LLM-based selectors or static policies, and offer limited controllability for semantic-aware routing under dynamic loads and mixed intents, often resulting in unstable performance and inefficient resource utilization. To address these limitations, we propose AMRO-S, an efficient and interpretable routing framework for Multi-Agent Systems (MAS). AMRO-S models MAS routing as a semantic-conditioned path selection problem, enhancing routing performance through three key mechanisms: First, it leverages a supervised fine-tuned (SFT) small language model for intent inference, providing a low-overhead semantic interface for each query; second, it decomposes routing memory into task-specific pheromone specialists, reducing cross-task interference and optimizing path selection under mixed workloads; finally, it employs a quality-gated asynchronous update mechanism to decouple inference from learning, optimizing routing without increasing latency. Extensive experiments on five public benchmarks and high-concurrency stress tests demonstrate that AMRO-S consistently improves the quality--cost trade-off over strong routing baselines, while providing traceable routing evidence through structured pheromone patterns.\n\n---\n\n## 8. ODRL Policy Comparison Through Normalisation\n- **Authors**: Jaime Osvaldo Salas , Paolo Pareti , George Konstantinidis\n- **URL**: [https://arxiv.org/abs/2603.12926](https://arxiv.org/abs/2603.12926)\n- **Abstract**:\n> The ODRL language has become the standard for representing policies and regulations for digital rights. However its complexity is a barrier to its usage, which has caused many related theoretical and practical works to focus on different, and not interoperable, fragments of ODRL. Moreover, semantically equivalent policies can be expressed in numerous different ways, which makes comparing them and processing them harder. Building on top of a recently defined semantics, we tackle these problems by proposing an approach that involves a parametrised normalisation of ODRL policies into its minimal components which reformulates policies with permissions and prohibitions into policies with permissions exclusively, and simplifies complex logic constraints into simple ones. We provide algorithms to compute a normal form for ODRL policies and simplifying numerical and symbolic constraints. We prove that these algorithms preserve the semantics of policies, and analyse the size complexity of the result, which is exponential on the number of attributes and linear on the number of unique values for these attributes. We show how this makes complex policies representable in more basic fragments of ODRL, and how it reduces the problem of policy comparison to the simpler problem of checking if two rules are identical.\n\n---\n\n## 9. Context is all you need: Towards autonomous model-based process design using agentic AI in flowsheet simulations\n- **Authors**: Pascal Schäfer , Lukas J. Krinke , Martin Wlotzka , Norbert Asprion\n- **URL**: [https://arxiv.org/abs/2603.12813](https://arxiv.org/abs/2603.12813)\n- **Abstract**:\n> Agentic AI systems integrating large language models (LLMs) with reasoning and tooluse capabilities are transforming various domains - in particular, software development. In contrast, their application in chemical process flowsheet modelling remains largely unexplored. In this work, we present an agentic AI framework that delivers assistance in an industrial flowsheet simulation environment. To this end, we show the capabilities of GitHub Copilot (GitHub, Inc., 2026), when using state-of-the-art LLMs, such as Claude Opus 4.6 (Anthropic, PBC, 2026), to generate valid syntax for our in-house process modelling tool Chemasim using the technical documentation and a few commented examples as context. Based on this, we develop a multi-agent system that decomposes process development tasks with one agent solving the abstract problem using engineering knowledge and another agent implementing the solution as Chemasim code. We demonstrate the effectiveness of our framework for typical flowsheet modelling examples, including (i) a reaction/separation process, (ii) a pressure-swing distillation, and (iii) a heteroazeotropic distillation including entrainer selection. Along these lines, we discuss current limitations of the framework and outline future research directions to further enhance its capabilities.\n\n---\n\n## 10. AI Model Modulation with Logits Redistribution\n- **Authors**: Zihan Wang , Zhongkui Ma , Xinguo Feng , Zhiyang Mei , Ethan Ma , Derui Wang , Minhui Xue , Guangdong Bai\n- **URL**: [https://arxiv.org/abs/2603.12755](https://arxiv.org/abs/2603.12755)\n- **Abstract**:\n> Large-scale models are typically adapted to meet the diverse requirements of model owners and users. However, maintaining multiple specialized versions of the model is inefficient. In response, we propose AIM, a novel model modulation paradigm that enables a single model to exhibit diverse behaviors to meet the specific end requirements. AIM enables two key modulation modes: utility and focus modulations. The former provides model owners with dynamic control over output quality to deliver varying utility levels, and the latter offers users precise control to shift model's focused input features. AIM introduces a logits redistribution strategy that operates in a training data-agnostic and retraining-free manner. We establish a formal foundation to ensure AIM's regulation capability, based on the statistical properties of logits ordering via joint probability distributions. Our evaluation confirms AIM's practicality and versatility for Al model modulation, with tasks spanning image classification, semantic segmentation and text generation, and prevalent architectures including ResNet, SegFormer and Llama.\n\n---\n\n## 11. ToolTree: Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning\n- **Authors**: Shuo Yang , Soyeon Caren Han , Yihao Ding , Shuhe Wang , Eduard Hoy\n- **URL**: [https://arxiv.org/abs/2603.12740](https://arxiv.org/abs/2603.12740)\n- **Abstract**:\n> Large Language Model (LLM) agents are increasingly applied to complex, multi-step tasks that require interaction with diverse external tools across various domains. However, current LLM agent tool planning methods typically rely on greedy, reactive tool selection strategies that lack foresight and fail to account for inter-tool dependencies. In this paper, we present ToolTree, a novel Monte Carlo tree search-inspired planning paradigm for tool planning. ToolTree explores possible tool usage trajectories using a dual-stage LLM evaluation and bidirectional pruning mechanism that enables the agent to make informed, adaptive decisions over extended tool-use sequences while pruning less promising branches before and after the tool execution. Empirical evaluations across both open-set and closed-set tool planning tasks on 4 benchmarks demonstrate that ToolTree consistently improves performance while keeping the highest efficiency, achieving an average gain of around 10\\% compared to the state-of-the-art planning paradigm.\n\n---\n\n## 12. On Using Machine Learning to Early Detect Catastrophic Failures in Marine Diesel Engines\n- **Authors**: Francesco Maione , Paolo Lino , Giuseppe Giannino , Guido Maione\n- **URL**: [https://arxiv.org/abs/2603.12733](https://arxiv.org/abs/2603.12733)\n- **Abstract**:\n> Catastrophic failures of marine engines imply severe loss of functionality and destroy or damage the systems irreversibly. Being sudden and often unpredictable events, they pose a severe threat to navigation, crew, and passengers. The abrupt nature makes early detection the only effective countermeasure. However, research has concentrated on modeling the gradual degradation of components, with limited attention to sudden and anomalous phenomena. This work proposes a new method for early detection of catastrophic failures. Based on real data from a failed engine, the approach evaluates the derivatives of the deviation between actual sensor readings and expected values of engine variables. Predictions are obtained by a Random Forest, which is the most suitable Machine Learning algorithm among the tested ones. Traditional methods focus on deviations of monitored signals, whereas the proposed approach employs the derivatives of the deviations to provide earlier indications of abnormal dynamics, and to alert that a rapid and dangerous event is breaking out within the system. The method allows the detection of anomalies before measurements reach critical thresholds and alarms are triggered, which is the common method in industry. Consequently, operators can be warned in advance and shut down the engine, then prevent damage and unexpected power loss. Moreover, they have the time to safely change the ship route and avoid potential obstacles. Simulation results conf irm the effectiveness of the proposed approach in anticipating occurrence of catastrophic failures. Validation on real-world data further reinforces the robustness and practical applicability of the method. It is worth noting that data acquisition to train the predictive algorithm is not a problem, since a Deep Learning-based data augmentation procedure is used.\n\n---\n\n## 13. AI Planning Framework for LLM-Based Web Agents\n- **Authors**: Orit Shahnovsky , Rotem Dror\n- **URL**: [https://arxiv.org/abs/2603.12710](https://arxiv.org/abs/2603.12710)\n- **Abstract**:\n> Developing autonomous agents for web-based tasks is a core challenge in AI. While Large Language Model (LLM) agents can interpret complex user requests, they often operate as black boxes, making it difficult to diagnose why they fail or how they plan. This paper addresses this gap by formally treating web tasks as sequential decision-making processes. We introduce a taxonomy that maps modern agent architectures to traditional planning paradigms: Step-by-Step agents to Breadth-First Search (BFS), Tree Search agents to Best-First Tree Search, and Full-Plan-in-Advance agents to Depth-First Search (DFS). This framework allows for a principled diagnosis of system failures like context drift and incoherent task decomposition. To evaluate these behaviors, we propose five novel evaluation metrics that assess trajectory quality beyond simple success rates. We support this analysis with a new dataset of 794 human-labeled trajectories from the WebArena benchmark. Finally, we validate our evaluation framework by comparing a baseline Step-by-Step agent against a novel Full-Plan-in-Advance implementation. Our results reveal that while the Step-by-Step agent aligns more closely with human gold trajectories (38% overall success), the Full-Plan-in-Advance agent excels in technical measures such as element accuracy (89%), demonstrating the necessity of our proposed metrics for selecting appropriate agent architectures based on specific application constraints.\n\n---\n\n## 14. Generating Expressive and Customizable Evals for Timeseries Data Analysis Agents with AgentFuel\n- **Authors**: Aadyaa Maddi , Prakhar Naval , Deepti Mande , Shane Duan , Muckai Girish , Vyas Sekar\n- **URL**: [https://arxiv.org/abs/2603.12483](https://arxiv.org/abs/2603.12483)\n- **Abstract**:\n> Across many domains (e.g., IoT, observability, telecommunications, cybersecurity), there is an emerging adoption of conversational data analysis agents that enable users to \"talk to your data\" to extract insights. Such data analysis agents operate on timeseries data models; e.g., measurements from sensors or events monitoring user clicks and actions in product analytics. We evaluate 6 popular data analysis agents (both open-source and proprietary) on domain-specific data and query types, and find that they fail on stateful and incident-specific queries. We observe two key expressivity gaps in existing evals: domain-customized datasets and domain-specific query types. To enable practitioners in such domains to generate customized and expressive evals for such timeseries data agents, we present AgentFuel. AgentFuel helps domain experts quickly create customized evals to perform end-to-end functional tests. We show that AgentFuel's benchmarks expose key directions for improvement in existing data agent frameworks. We also present anecdotal evidence that using AgentFuel can improve agent performance (e.g., with GEPA). AgentFuel benchmarks are available at this https URL .\n\n---\n\n## 15. Efficient Reasoning with Balanced Thinking\n- **Authors**: Yulin Li , Tengyao Tu , Li Ding , Junjie Wang , Huiling Zhen , Yixin Chen , Yong Li , Zhuotao Tian\n- **URL**: [https://arxiv.org/abs/2603.12372](https://arxiv.org/abs/2603.12372)\n- **Abstract**:\n> Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they often suffer from overthinking, expending redundant computational steps on simple problems, or underthinking, failing to explore sufficient reasoning paths despite inherent capabilities. These issues lead to inefficiencies and potential inaccuracies, limiting practical deployment in resource-constrained settings. Existing methods to mitigate overthinking, such as suppressing reflective keywords or adjusting reasoning length, may inadvertently induce underthinking, compromising accuracy. Therefore, we propose ReBalance, a training-free framework that achieves efficient reasoning with balanced thinking. ReBalance leverages confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence. By aggregating hidden states from a small-scale dataset into reasoning mode prototypes, we compute a steering vector to guide LRMs' reasoning trajectories. A dynamic control function modulates this vector's strength and direction based on real-time confidence, pruning redundancy during overthinking, and promoting exploration during underthinking. Extensive experiments conducted on four models ranging from 0.5B to 32B, and across nine benchmarks in math reasoning, general question answering, and coding tasks demonstrate that ReBalance effectively reduces output redundancy while improving accuracy, offering a general, training-free, and plug-and-play strategy for efficient and robust LRM deployment. Code is available at this https URL .\n\n---\n\n## 16. Context-Enriched Natural Language Descriptions of Vessel Trajectories\n- **Authors**: Kostas Patroumpas , Alexandros Troupiotis-Kapeliaris , Giannis Spiliopoulos , Panagiotis Betchavas , Dimitrios Skoutas , Dimitris Zissis , Nikos Bikakis\n- **URL**: [https://arxiv.org/abs/2603.12287](https://arxiv.org/abs/2603.12287)\n- **Abstract**:\n> We address the problem of transforming raw vessel trajectory data collected from AIS into structured and semantically enriched representations interpretable by humans and directly usable by machine reasoning systems. We propose a context-aware trajectory abstraction framework that segments noisy AIS sequences into distinct trips each consisting of clean, mobility-annotated episodes. Each episode is further enriched with multi-source contextual information, such as nearby geographic entities, offshore navigation features, and weather conditions. Crucially, such representations can support generation of controlled natural language descriptions using LLMs. We empirically examine the quality of such descriptions generated using several LLMs over AIS data along with open contextual features. By increasing semantic density and reducing spatiotemporal complexity, this abstraction can facilitate downstream analytics and enable integration with LLMs for higher-level maritime reasoning tasks.\n\n---\n\n## 17. PhysMoDPO: Physically-Plausible Humanoid Motion with Preference Optimization\n- **Authors**: Yangsong Zhang , Anujith Muraleedharan , Rikhat Akizhanov , Abdul Ahad Butt , Gül Varol , Pascal Fua , Fabio Pizzati , Ivan Laptev\n- **URL**: [https://arxiv.org/abs/2603.13228](https://arxiv.org/abs/2603.13228)\n- **Abstract**:\n> Recent progress in text-conditioned human motion generation has been largely driven by diffusion models trained on large-scale human motion data. Building on this progress, recent methods attempt to transfer such models for character animation and real robot control by applying a Whole-Body Controller (WBC) that converts diffusion-generated motions into executable trajectories. While WBC trajectories become compliant with physics, they may expose substantial deviations from original motion. To address this issue, we here propose PhysMoDPO, a Direct Preference Optimization framework. Unlike prior work that relies on hand-crafted physics-aware heuristics such as foot-sliding penalties, we integrate WBC into our training pipeline and optimize diffusion model such that the output of WBC becomes compliant both with physics and original text instructions. To train PhysMoDPO we deploy physics-based and task-specific rewards and use them to assign preference to synthesized trajectories. Our extensive experiments on text-to-motion and spatial control tasks demonstrate consistent improvements of PhysMoDPO in both physical realism and task-related metrics on simulated robots. Moreover, we demonstrate that PhysMoDPO results in significant improvements when applied to zero-shot motion transfer in simulation and for real-world deployment on a G1 humanoid robot.\n\n---\n\n## 18. Visual-ERM: Reward Modeling for Visual Equivalence\n- **Authors**: Ziyu Liu , Shengyuan Ding , Xinyu Fang , Xuanlang Dai , Penghui Yang , Jianze Liang , Jiaqi Wang , Kai Chen , Dahua Lin , Yuhang Zang\n- **URL**: [https://arxiv.org/abs/2603.13224](https://arxiv.org/abs/2603.13224)\n- **Abstract**:\n> Vision-to-code tasks require models to reconstruct structured visual inputs, such as charts, tables, and SVGs, into executable or structured representations with high visual fidelity. While recent Large Vision Language Models (LVLMs) achieve strong results via supervised fine-tuning, reinforcement learning remains challenging due to misaligned reward signals. Existing rewards either rely on textual rules or coarse visual embedding similarity, both of which fail to capture fine-grained visual discrepancies and are vulnerable to reward hacking. We propose Visual Equivalence Reward Model (Visual-ERM), a multimodal generative reward model that provides fine-grained, interpretable, and task-agnostic feedback to evaluate vision-to-code quality directly in the rendered visual space. Integrated into RL, Visual-ERM improves Qwen3-VL-8B-Instruct by +8.4 on chart-to-code and yields consistent gains on table and SVG parsing (+2.7, +4.1 on average), and further strengthens test-time scaling via reflection and revision. We also introduce VisualCritic-RewardBench (VC-RewardBench), a benchmark for judging fine-grained image-to-image discrepancies on structured visual data, where Visual-ERM at 8B decisively outperforms Qwen3-VL-235B-Instruct and approaches leading closed-source models. Our results suggest that fine-grained visual reward supervision is both necessary and sufficient for vision-to-code RL, regardless of task specificity.\n\n---\n\n## 19. From Experiments to Expertise: Scientific Knowledge Consolidation for AI-Driven Computational Research\n- **Authors**: Haonan Huang\n- **URL**: [https://arxiv.org/abs/2603.13191](https://arxiv.org/abs/2603.13191)\n- **Abstract**:\n> While large language models (LLMs) have transformed AI agents into proficient executors of computational materials science, performing a hundred simulations does not make a researcher. What distinguishes research from routine execution is the progressive accumulation of knowledge -- learning which approaches fail, recognizing patterns across systems, and applying understanding to new problems. However, the prevailing paradigm in AI-driven computational science treats each execution in isolation, largely discarding hard-won insights between runs. Here we present QMatSuite, an open-source platform closing this gap. Agents record findings with full provenance, retrieve knowledge before new calculations, and in dedicated reflection sessions correct erroneous findings and synthesize observations into cross-compound patterns. In benchmarks on a six-step quantum-mechanical simulation workflow, accumulated knowledge reduces reasoning overhead by 67% and improves accuracy from 47% to 3% deviation from literature -- and when transferred to an unfamiliar material, achieves 1% deviation with zero pipeline failures.\n\n---\n\n## 20. LLM Constitutional Multi-Agent Governance\n- **Authors**: J. de Curtò , I. de Zarzà\n- **URL**: [https://arxiv.org/abs/2603.13189](https://arxiv.org/abs/2603.13189)\n- **Abstract**:\n> Large Language Models (LLMs) can generate persuasive influence strategies that shift cooperative behavior in multi-agent populations, but a critical question remains: does the resulting cooperation reflect genuine prosocial alignment, or does it mask erosion of agent autonomy, epistemic integrity, and distributional fairness? We introduce Constitutional Multi-Agent Governance (CMAG), a two-stage framework that interposes between an LLM policy compiler and a networked agent population, combining hard constraint filtering with soft penalized-utility optimization that balances cooperation potential against manipulation risk and autonomy pressure. We propose the Ethical Cooperation Score (ECS), a multiplicative composite of cooperation, autonomy, integrity, and fairness that penalizes cooperation achieved through manipulative means. In experiments on scale-free networks of 80 agents under adversarial conditions (70% violating candidates), we benchmark three regimes: full CMAG, naive filtering, and unconstrained optimization. While unconstrained optimization achieves the highest raw cooperation (0.873), it yields the lowest ECS (0.645) due to severe autonomy erosion (0.867) and fairness degradation (0.888). CMAG attains an ECS of 0.741, a 14.9% improvement, while preserving autonomy at 0.985 and integrity at 0.995, with only modest cooperation reduction to 0.770. The naive ablation (ECS = 0.733) confirms that hard constraints alone are insufficient. Pareto analysis shows CMAG dominates the cooperation-autonomy trade-off space, and governance reduces hub-periphery exposure disparities by over 60%. These findings establish that cooperation is not inherently desirable without governance: constitutional constraints are necessary to ensure that LLM-mediated influence produces ethically stable outcomes rather than manipulative equilibria.\n\n---\n\n## 21. Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights\n- **Authors**: Xingli Fang , Jung-Eun Kim\n- **URL**: [https://arxiv.org/abs/2603.13186](https://arxiv.org/abs/2603.13186)\n- **Abstract**:\n> Prior approaches for membership privacy preservation usually update or retrain all weights in neural networks, which is costly and can lead to unnecessary utility loss or even more serious misalignment in predictions between training data and non-training data. In this work, we observed three insights: i) privacy vulnerability exists in a very small fraction of weights; ii) however, most of those weights also critically impact utility performance; iii) the importance of weights stems from their locations rather than their values. According to these insights, to preserve privacy, we score critical weights, and instead of discarding those neurons, we rewind only the weights for fine-tuning. We show that, through extensive experiments, this mechanism exhibits outperforming resilience in most cases against Membership Inference Attacks while maintaining utility.\n\n---\n\n## 22. MXNorm: Reusing MXFP block scales for efficient tensor normalisation\n- **Authors**: Callum McLean , Luke Y. Prince , Alexandre Payot , Paul Balança , Carlo Luschi\n- **URL**: [https://arxiv.org/abs/2603.13180](https://arxiv.org/abs/2603.13180)\n- **Abstract**:\n> Matrix multiplication performance has long been the major bottleneck to scaling deep learning workloads, which has stimulated the design of new accelerators that use increasingly low-precision number formats. However, improvements in matrix multiplication performance have far outstripped improvements in performance on reductions and elementwise computations, which are still being performed in higher precision. In this work, we propose MXNorm, a drop-in replacement for RMSNorm that estimates the RMS using only the block scales calculated as part of the MXFP8 cast and enables a 32x decrease in the size of reduction needed for normalization. We validate our approximation method on pre-training of Llama 3 models of 125M, 1B and 8B parameters, finding minimal loss of training accuracy compared to a baseline using RMSNorm with MXFP8 matmuls. We also show practical kernel speedups using only this http URL of up to 2.4x for MXNorm over RMSNorm, corresponding to a 1.3% speedup in Llama 3 8B transformer layers in MXFP8 and a 2.6% speedup in NVFP4.\n\n---\n\n## 23. Clustering Astronomical Orbital Synthetic Data Using Advanced Feature Extraction and Dimensionality Reduction Techniques\n- **Authors**: Eraldo Pereira Marinho , Nelson Callegari Junior , Fabricio Aparecido Breve , Caetano Mazzoni Ranieri\n- **URL**: [https://arxiv.org/abs/2603.13177](https://arxiv.org/abs/2603.13177)\n- **Abstract**:\n> The dynamics of Saturn's satellite system offer a rich framework for studying orbital stability and resonance interactions. Traditional methods for analysing such systems, including Fourier analysis and stability metrics, struggle with the scale and complexity of modern datasets. This study introduces a machine learning-based pipeline for clustering approximately 22,300 simulated satellite orbits, addressing these challenges with advanced feature extraction and dimensionality reduction techniques. The key to this approach is using MiniRocket, which efficiently transforms 400 timesteps into a 9,996-dimensional feature space, capturing intricate temporal patterns. Additional automated feature extraction and dimensionality reduction techniques refine the data, enabling robust clustering analysis. This pipeline reveals stability regions, resonance structures, and other key behaviours in Saturn's satellite system, providing new insights into their long-term dynamical evolution. By integrating computational tools with traditional celestial mechanics techniques, this study offers a scalable and interpretable methodology for analysing large-scale orbital datasets and advancing the exploration of planetary dynamics.\n\n---\n\n## 24. ESG-Bench: Benchmarking Long-Context ESG Reports for Hallucination Mitigation\n- **Authors**: Siqi Sun , Ben Peng Wu , Mali Jin , Peizhen Bai , Hanpei Zhang , Xingyi Song\n- **URL**: [https://arxiv.org/abs/2603.13154](https://arxiv.org/abs/2603.13154)\n- **Abstract**:\n> As corporate responsibility increasingly incorporates environmental, social, and governance (ESG) criteria, ESG reporting is becoming a legal requirement in many regions and a key channel for documenting sustainability practices and assessing firms' long-term and ethical performance. However, the length and complexity of ESG disclosures make them difficult to interpret and automate the analysis reliably. To support scalable and trustworthy analysis, this paper introduces ESG-Bench, a benchmark dataset for ESG report understanding and hallucination mitigation in large language models (LLMs). ESG-Bench contains human-annotated question-answer (QA) pairs grounded in real-world ESG report contexts, with fine-grained labels indicating whether model outputs are factually supported or hallucinated. Framing ESG report analysis as a QA task with verifiability constraints enables systematic evaluation of LLMs' ability to extract and reason over ESG content and provides a new use case: mitigating hallucinations in socially sensitive, compliance-critical settings. We design task-specific Chain-of-Thought (CoT) prompting strategies and fine-tune multiple state-of-the-art LLMs on ESG-Bench using CoT-annotated rationales. Our experiments show that these CoT-based methods substantially outperform standard prompting and direct fine-tuning in reducing hallucinations, and that the gains transfer to existing QA benchmarks beyond the ESG domain.\n\n---\n\n## 25. Developing the PsyCogMetrics AI Lab to Evaluate Large Language Models and Advance Cognitive Science -- A Three-Cycle Action Design Science Study\n- **Authors**: Zhiye Jin , Yibai Li , K. D. Joshi , Xuefei (Nancy) Deng , Xiaobing (Emily)Li\n- **URL**: [https://arxiv.org/abs/2603.13126](https://arxiv.org/abs/2603.13126)\n- **Abstract**:\n> This study presents the development of the PsyCogMetrics AI Lab ( this http URL ), an integrated, cloud-based platform that operationalizes psychometric and cognitive-science methodologies for Large Language Model (LLM) evaluation. Framed as a three-cycle Action Design Science study, the Relevance Cycle identifies key limitations in current evaluation methods and unfulfilled stakeholder needs. The Rigor Cycle draws on kernel theories such as Popperian falsifiability, Classical Test Theory, and Cognitive Load Theory to derive deductive design objectives. The Design Cycle operationalizes these objectives through nested Build-Intervene-Evaluate loops. The study contributes a novel IT artifact, a validated design for LLM evaluation, benefiting research at the intersection of AI, psychology, cognitive science, and the social and behavioral sciences.\n\n---\n\n## 26. Geometry-Guided Camera Motion Understanding in VideoLLMs\n- **Authors**: Haoan Feng , Sri Harsha Musunuri , Guan-Ming Su\n- **URL**: [https://arxiv.org/abs/2603.13119](https://arxiv.org/abs/2603.13119)\n- **Abstract**:\n> Camera motion is a fundamental geometric signal that shapes visual perception and cinematic style, yet current video-capable vision-language models (VideoLLMs) rarely represent it explicitly and often fail on fine-grained motion primitives. We address this gap with a framework of $\\textbf{benchmarking}$, $\\textbf{diagnosis}$, and $\\textbf{injection}$. We curate $\\textbf{CameraMotionDataset}$, a large-scale synthetic dataset with explicit camera control, formulate camera motion as constraint-aware multi-label recognition, and construct a VQA benchmark--$\\textbf{CameraMotionVQA}$. Across diverse off-the-shelf VideoLLMs, we observe substantial errors in recognizing camera motion primitives. Probing experiments on a Qwen2.5-VL vision encoder suggest that camera motion cues are weakly represented, especially in deeper ViT blocks, helping explain the observed failure modes. To bridge this gap without costly training or fine-tuning, we propose a lightweight, model-agnostic pipeline that extracts geometric camera cues from 3D foundation models (3DFMs), predicts constrained motion primitives with a temporal classifier, and injects them into downstream VideoLLM inference via structured prompting. Experiments demonstrate improved motion recognition and more camera-aware model responses, highlighting geometry-driven cue extraction and structured prompting as practical steps toward a camera-aware VideoLLM and VLA system. The dataset and benchmark is publicly available at this https URL .\n\n---\n\n## 27. BoSS: A Best-of-Strategies Selector as an Oracle for Deep Active Learning\n- **Authors**: Denis Huseljic , Paul Hahn , Marek Herde , Christoph Sandrock , Bernhard Sick\n- **URL**: [https://arxiv.org/abs/2603.13109](https://arxiv.org/abs/2603.13109)\n- **Abstract**:\n> Active learning (AL) aims to reduce annotation costs while maximizing model performance by iteratively selecting valuable instances. While foundation models have made it easier to identify these instances, existing selection strategies still lack robustness across different models, annotation budgets, and datasets. To highlight the potential weaknesses of existing AL strategies and provide a reference point for research, we explore oracle strategies, i.e., strategies that approximate the optimal selection by accessing ground-truth information unavailable in practical AL scenarios. Current oracle strategies, however, fail to scale effectively to large datasets and complex deep neural networks. To tackle these limitations, we introduce the Best-of-Strategy Selector (BoSS), a scalable oracle strategy designed for large-scale AL scenarios. BoSS constructs a set of candidate batches through an ensemble of selection strategies and then selects the batch yielding the highest performance gain. As an ensemble of selection strategies, BoSS can be easily extended with new state-of-the-art strategies as they emerge, ensuring it remains a reliable oracle strategy in the future. Our evaluation demonstrates that i) BoSS outperforms existing oracle strategies, ii) state-of-the-art AL strategies still fall noticeably short of oracle performance, especially in large-scale datasets with many classes, and iii) one possible solution to counteract the inconsistent performance of AL strategies might be to employ an ensemble-based approach for the selection.\n\n---\n\n## 28. Evaluating VLMs' Spatial Reasoning Over Robot Motion: A Step Towards Robot Planning with Motion Preferences\n- **Authors**: Wenxi Wu , Jingjing Zhang , Martim Brandão\n- **URL**: [https://arxiv.org/abs/2603.13100](https://arxiv.org/abs/2603.13100)\n- **Abstract**:\n> Understanding user instructions and object spatial relations in surrounding environments is crucial for intelligent robot systems to assist humans in various tasks. The natural language and spatial reasoning capabilities of Vision-Language Models (VLMs) have the potential to enhance the generalization of robot planners on new tasks, objects, and motion specifications. While foundation models have been applied to task planning, it is still unclear the degree to which they have the capability of spatial reasoning required to enforce user preferences or constraints on motion, such as desired distances from objects, topological properties, or motion style preferences. In this paper, we evaluate the capability of four state-of-the-art VLMs at spatial reasoning over robot motion, using four different querying methods. Our results show that, with the highest-performing querying method, Qwen2.5-VL achieves 71.4% accuracy zero-shot and 75% on a smaller model after fine-tuning, and GPT-4o leads to lower performance. We evaluate two types of motion preferences (object-proximity and path-style), and we also analyze the trade-off between accuracy and computation cost in number of tokens. This work shows some promise in the potential of VLM integration with robot motion planning pipelines.\n\n---\n\n## 29. Human-in-the-Loop LLM Grading for Handwritten Mathematics Assessments\n- **Authors**: Arne Vanhoyweghen , Vincent Holst , Melika Mobini , Lukas Van de Voorde , Tibo Vanleke , Bert Verbruggen , Brecht Verbeken , Andres Algaba , Sam Verboven , Marie-Anne Guerry , Filip Van Droogenbroeck , Vincent Ginis\n- **URL**: [https://arxiv.org/abs/2603.13083](https://arxiv.org/abs/2603.13083)\n- **Abstract**:\n> Providing timely and individualised feedback on handwritten student work is highly beneficial for learning but difficult to achieve at scale. This challenge has become more pressing as generative AI undermines the reliability of take-home assessments, shifting emphasis toward supervised, in-class evaluation. We present a scalable, end-to-end workflow for LLM-assisted grading of short, pen-and-paper assessments. The workflow spans (1) constructing solution keys, (2) developing detailed rubric-style grading keys used to guide the LLM, and (3) a grading procedure that combines automated scanning and anonymisation, multi-pass LLM scoring, automated consistency checks, and mandatory human verification. We deploy the system in two undergraduate mathematics courses using six low-stakes in-class tests. Empirically, LLM assistance reduces grading time by approximately 23% while achieving agreement comparable to, and in several cases tighter than, fully manual grading. Occasional model errors occur but are effectively contained by the hybrid design. Overall, our results show that carefully embedded human-in-the-loop LLM grading can substantially reduce workload while maintaining fairness and accuracy.\n\n---\n\n## 30. GeoChemAD: Benchmarking Unsupervised Geochemical Anomaly Detection for Mineral Exploration\n- **Authors**: Yihao Ding , Yiran Zhang , Chris Gonzalez , Eun-Jung Holden , Wei Liu\n- **URL**: [https://arxiv.org/abs/2603.13068](https://arxiv.org/abs/2603.13068)\n- **Abstract**:\n> Geochemical anomaly detection plays a critical role in mineral exploration as deviations from regional geochemical baselines may indicate mineralization. Existing studies suffer from two key limitations: (1) single region scenarios which limit model generalizability; (2) proprietary datasets, which makes result reproduction unattainable. In this work, we introduce \\textbf{GeoChemAD}, an open-source benchmark dataset compiled from government-led geological surveys, covering multiple regions, sampling sources, and target elements. The dataset comprises eight subsets representing diverse spatial scales and sampling conditions. To establish strong baselines, we reproduce and benchmark a range of unsupervised anomaly detection methods, including statistical models, generative and transformer-based approaches. Furthermore, we propose \\textbf{GeoChemFormer}, a transformer-based framework that leverages self-supervised pretraining to learn target-element-aware geochemical representations for spatial samples. Extensive experiments demonstrate that GeoChemFormer consistently achieves superior and robust performance across all eight subsets, outperforming existing unsupervised methods in both anomaly detection accuracy and generalization capability. The proposed dataset and framework provide a foundation for reproducible research and future development in this direction.\n\n---\n\n## 31. L2GTX: From Local to Global Time Series Explanations\n- **Authors**: Ephrem Tibebe Mekonnen , Luca Longo , Lucas Rizzo , Pierpaolo Dondio\n- **URL**: [https://arxiv.org/abs/2603.13065](https://arxiv.org/abs/2603.13065)\n- **Abstract**:\n> Deep learning models achieve high accuracy in time series classification, yet understanding their class-level decision behaviour remains challenging. Explanations for time series must respect temporal dependencies and identify patterns that recur across instances. Existing approaches face three limitations: model-agnostic XAI methods developed for images and tabular data do not readily extend to time series, global explanation synthesis for time series remains underexplored, and most existing global approaches are model-specific. We propose L2GTX, a model-agnostic framework that generates class-wise global explanations by aggregating local explanations from a representative set of instances. L2GTX extracts clusters of parameterised temporal event primitives, such as increasing or decreasing trends and local extrema, together with their importance scores from instance-level explanations produced by LOMATCE. These clusters are merged across instances to reduce redundancy, and an instance-cluster importance matrix is used to estimate global relevance. Under a user-defined instance selection budget, L2GTX selects representative instances that maximise coverage of influential clusters. Events from the selected instances are then aggregated into concise class-wise global explanations. Experiments on six benchmark time series datasets show that L2GTX produces compact and interpretable global explanations while maintaining stable global faithfulness measured as mean local surrogate fidelity.\n\n---\n\n## 32. Competition-Aware CPC Forecasting with Near-Market Coverage\n- **Authors**: Sebastian Frey , Edoardo Beccari , Maximilian Kranz , Nicolò Alberto Pellizzari , Ali Mete Karaman , Qiwei Han , Maximilian Kaiser\n- **URL**: [https://arxiv.org/abs/2603.13059](https://arxiv.org/abs/2603.13059)\n- **Abstract**:\n> Cost-per-click (CPC) in paid search is a volatile auction outcome generated by a competitive landscape that is only partially observable from any single advertiser's history. Using Google Ads auction logs from a concentrated car-rental market (2021--2023), we forecast weekly CPC for 1,811 keyword series and approximate latent competition through complementary signals derived from keyword text, CPC trajectories, and geographic market structure. We construct (i) semantic neighborhoods and a semantic keyword graph from pretrained transformer-based representations of keyword text, (ii) behavioral neighborhoods via Dynamic Time Warping (DTW) alignment of CPC trajectories, and (iii) geographic-intent covariates capturing localized demand and marketplace heterogeneity. We extensively evaluate these signals both as stand-alone covariates and as relational priors in spatiotemporal graph forecasters, benchmarking them against strong statistical, neural, and time-series foundation-model baselines. Across methods, competition-aware augmentation improves stability and error profiles at business-relevant medium and longer horizons, where competitive regimes shift and volatility is most consequential. The results show that broad market-outcome coverage, combined with keyword-derived semantic and geographic priors, provides a scalable way to approximate latent competition and improve CPC forecasting in auction-driven markets.\n\n---\n\n## 33. Team RAS in 10th ABAW Competition: Multimodal Valence and Arousal Estimation Approach\n- **Authors**: Elena Ryumina (1), Maxim Markitantov (1), Alexandr Axyonov (1), Dmitry Ryumin (1), Mikhail Dolgushin (1), Denis Dresvyanskiy (2), Alexey Karpov (1 and 2) ((1) St. Petersburg Federal Research Center of the Russian Academy of Sciences, St. Petersburg, Russia, (2) ITMO University, St. Petersburg, Russia)\n- **URL**: [https://arxiv.org/abs/2603.13056](https://arxiv.org/abs/2603.13056)\n- **Abstract**:\n> Continuous emotion recognition in terms of valence and arousal under in-the-wild (ITW) conditions remains a challenging problem due to large variations in appearance, head pose, illumination, occlusions, and subject-specific patterns of affective expression. We present a multimodal method for valence-arousal estimation ITW. Our method combines three complementary modalities: face, behavior, and audio. The face modality relies on GRADA-based frame-level embeddings and Transformer-based temporal regression. We use Qwen3-VL-4B-Instruct to extract behavior-relevant information from video segments, while Mamba is used to model temporal dynamics across segments. The audio modality relies on WavLM-Large with attention-statistics pooling and includes a cross-modal filtering stage to reduce the influence of unreliable or non-speech segments. To fuse modalities, we explore two fusion strategies: a Directed Cross-Modal Mixture-of-Experts Fusion Strategy that learns interactions between modalities with adaptive weighting, and a Reliability-Aware Audio-Visual Fusion Strategy that combines visual features at the frame-level while using audio as complementary context. The results are reported on the Aff-Wild2 dataset following the 10th Affective Behavior Analysis in-the-Wild (ABAW) challenge protocol. Experiments demonstrate that the proposed multimodal fusion strategy achieves a Concordance Correlation Coefficient (CCC) of 0.658 on the Aff-Wild2 development set.\n\n---\n\n## 34. Are General-Purpose Vision Models All We Need for 2D Medical Image Segmentation? A Cross-Dataset Empirical Study\n- **Authors**: Vanessa Borst , Samuel Kounev\n- **URL**: [https://arxiv.org/abs/2603.13044](https://arxiv.org/abs/2603.13044)\n- **Abstract**:\n> Medical image segmentation (MIS) is a fundamental component of computer-assisted diagnosis and clinical decision support systems. Over the past decade, numerous architectures specifically tailored to medical imaging have emerged to address domain-specific challenges such as low contrast, small anatomical structures, and limited annotated data. In parallel, rapid progress in computer vision has produced highly capable general-purpose vision models (GP-VMs) originally designed for natural images. Despite their strong performance on standard vision benchmarks, their effectiveness for MIS remains insufficiently understood. In this work, we conduct a controlled empirical study to examine whether specialized medical segmentation architectures (SMAs) provide systematic advantages over modern GP-VMs for 2D MIS. We compare eleven SMAs and GP-VMs using a unified training and evaluation protocol. Experiments are performed across three heterogeneous datasets covering different imaging modalities, class structures, and data characteristics. Beyond segmentation accuracy, we analyze qualitative Grad-CAM visualizations to investigate explainability (XAI) behavior. Our results demonstrate that, for the analyzed datasets, GP-VMs out-perform the majority of specialized MIS models. Moreover, XAI analyses indicate that GP-VMs can capture clinically relevant structures without explicit domain-specific architectural design. These findings suggest that GP-VMs can represent a viable alternative to domain-specific methods, highlighting the importance of informed model selection for end-to-end MIS systems. All code and resources are available at GitHub.\n\n---\n\n## 35. Interrogating Design Homogenization in Web Vibe Coding\n- **Authors**: Donghoon Shin , Alice Gao , Rock Yuren Pang , Jaewook Lee , Katharina Reinecke , Emily Tseng\n- **URL**: [https://arxiv.org/abs/2603.13036](https://arxiv.org/abs/2603.13036)\n- **Abstract**:\n> Generative AI is known for its tendency to homogenize, often reproducing dominant style conventions found in training data. However, it remains unclear how these homogenizing effects extend to complex structural tasks like web design. As lay creators increasingly turn to LLMs to 'vibe-code' websites -- prompting for aesthetic and functional goals rather than writing code -- they may inadvertently narrow the diversity of their designs, and limit creative expression throughout the internet. In this paper, we interrogate the possibility of design homogenization in web vibe coding. We first characterize the vibe coding lifecycle, pinpointing stages where homogenization risks may arise. We then conduct a sociotechnical risk analysis unpacking the potential harms of web vibe coding and their interaction with design homogenization. We identify that the push for frictionless generation can exacerbate homogenization and its harms. Finally, we propose a mitigation framework centered on the idea of productive friction. Through case studies at the micro, meso, and macro levels, we show how centering productive friction can empower creators to challenge default outputs and preserve diverse expression in AI-mediated web design.\n\n---\n\n## 36. Purify Once, Edit Freely: Breaking Image Protections under Model Mismatch\n- **Authors**: Qichen Zhao , Shengfang Zhai , Xinjian Bai , Qingni Shen , Qiqi Lin , Yansong Gao , Zhonghai Wu\n- **URL**: [https://arxiv.org/abs/2603.13028](https://arxiv.org/abs/2603.13028)\n- **Abstract**:\n> Diffusion models enable high-fidelity image editing but can also be misused for unauthorized style imitation and harmful content generation. To mitigate these risks, proactive image protection methods embed small, often imperceptible adversarial perturbations into images before sharing to disrupt downstream editing or fine-tuning. However, in realistic post-release scenarios, content owners cannot control downstream processing pipelines, and protections optimized for a surrogate model may fail when attackers use mismatched diffusion pipelines. Existing purification methods can weaken protections but often sacrifice image quality and rarely examine architectural mismatch. We introduce a unified post-release purification framework to evaluate protection survivability under model mismatch. We propose two practical purifiers: VAE-Trans, which corrects protected images via latent-space projection, and EditorClean, which performs instruction-guided reconstruction with a Diffusion Transformer to exploit architectural heterogeneity. Both operate without access to protected images or defense internals. Across 2,100 editing tasks and six representative protection methods, EditorClean consistently restores editability. Compared to protected inputs, it improves PSNR by 3-6 dB and reduces FID by 50-70 percent on downstream edits, while outperforming prior purification baselines by about 2 dB PSNR and 30 percent lower FID. Our results reveal a purify-once, edit-freely failure mode: once purification succeeds, the protective signal is largely removed, enabling unrestricted editing. This highlights the need to evaluate protections under model mismatch and design defenses robust to heterogeneous attackers.\n\n---\n\n## 37. SortScrews: A Dataset and Baseline for Real-time Screw Classification\n- **Authors**: Tianhao Fu , Bingxuan Yang , Juncheng Guo , Shrena Sribalan , Yucheng Chen\n- **URL**: [https://arxiv.org/abs/2603.13027](https://arxiv.org/abs/2603.13027)\n- **Abstract**:\n> Automatic identification of screw types is important for industrial automation, robotics, and inventory management. However, publicly available datasets for screw classification are scarce, particularly for controlled single-object scenarios commonly encountered in automated sorting systems. In this work, we introduce $\\textbf{SortScrews}$, a dataset for casewise visual classification of screws. The dataset contains 560 RGB images at $512\\times512$ resolution covering six screw types and a background class. Images are captured using a standardized acquisition setup and include mild variations in lighting and camera perspective across four capture settings. To facilitate reproducible research and dataset expansion, we also provide a reusable data collection script that allows users to easily construct similar datasets for custom hardware components using inexpensive camera setups. We establish baseline results using transfer learning with EfficientNet-B0 and ResNet-18 classifiers pretrained on ImageNet. In addition, we conduct a well-explored failure analysis. Despite the limited dataset size, these lightweight models achieve strong classification accuracy, demonstrating that controlled acquisition conditions enable effective learning even with relatively small datasets. The dataset, collection pipeline, and baseline training code are publicly available at this https URL .\n\n---\n\n## 38. SAW: Toward a Surgical Action World Model via Controllable and Scalable Video Generation\n- **Authors**: Sampath Rapuri , Lalithkumar Seenivasan , Dominik Schneider , Roger Soberanis-Mukul , Yufan He , Hao Ding , Jiru Xu , Chenhao Yu , Chenyan Jing , Pengfei Guo , Daguang Xu , Mathias Unberath\n- **URL**: [https://arxiv.org/abs/2603.13024](https://arxiv.org/abs/2603.13024)\n- **Abstract**:\n> A surgical world model capable of generating realistic surgical action videos with precise control over tool-tissue interactions can address fundamental challenges in surgical AI and simulation -- from data scarcity and rare event synthesis to bridging the sim-to-real gap for surgical automation. However, current video generation methods, the very core of such surgical world models, require expensive annotations or complex structured intermediates as conditioning signals at inference, limiting their scalability. Other approaches exhibit limited temporal consistency across complex laparoscopic scenes and do not possess sufficient realism. We propose Surgical Action World (SAW) -- a step toward surgical action world modeling through video diffusion conditioned on four lightweight signals: language prompts encoding tool-action context, a reference surgical scene, tissue affordance mask, and 2D tool-tip trajectories. We design a conditional video diffusion approach that reformulates video-to-video diffusion into trajectory-conditioned surgical action synthesis. The backbone diffusion model is fine-tuned on a custom-curated dataset of 12,044 laparoscopic clips with lightweight spatiotemporal conditioning signals, leveraging a depth consistency loss to enforce geometric plausibility without requiring depth at inference. SAW achieves state-of-the-art temporal consistency (CD-FVD: 199.19 vs. 546.82) and strong visual quality on held-out test data. Furthermore, we demonstrate its downstream utility for (a) surgical AI, where augmenting rare actions with SAW-generated videos improves action recognition (clipping F1-score: 20.93% to 43.14%; cutting: 0.00% to 8.33%) on real test data, and (b) surgical simulation, where rendering tool-tissue interaction videos from simulator-derived trajectory points toward a visually faithful simulation engine.\n\n---\n\n## 39. daVinci-Env: Open SWE Environment Synthesis at Scale\n- **Authors**: Dayuan Fu , Shenyu Wu , Yunze Wu , Zerui Peng , Yaxing Huang , Jie Sun , Ji Zeng , Mohan Jiang , Lin Zhang , Yukun Li , Jiarui Hu , Liming Liu , Jinlong Hou , Pengfei Liu\n- **URL**: [https://arxiv.org/abs/2603.13023](https://arxiv.org/abs/2603.13023)\n- **Abstract**:\n> Training capable software engineering (SWE) agents demands large-scale, executable, and verifiable environments that provide dynamic feedback loops for iterative code editing, test execution, and solution refinement. However, existing open-source datasets remain limited in scale and repository diversity, while industrial solutions are opaque with unreleased infrastructure, creating a prohibitive barrier for most academic research groups. We present OpenSWE, the largest fully transparent framework for SWE agent training in Python, comprising 45,320 executable Docker environments spanning over 12.8k repositories, with all Dockerfiles, evaluation scripts, and infrastructure fully open-sourced for reproducibility. OpenSWE is built through a multi-agent synthesis pipeline deployed across a 64-node distributed cluster, automating repository exploration, Dockerfile construction, evaluation script generation, and iterative test analysis. Beyond scale, we propose a quality-centric filtering pipeline that characterizes the inherent difficulty of each environment, filtering out instances that are either unsolvable or insufficiently challenging and retaining only those that maximize learning efficiency. With $891K spent on environment construction and an additional $576K on trajectory sampling and difficulty-aware curation, the entire project represents a total investment of approximately $1.47 million, yielding about 13,000 curated trajectories from roughly 9,000 quality guaranteed environments. Extensive experiments validate OpenSWE's effectiveness: OpenSWE-32B and OpenSWE-72B achieve 62.4% and 66.0% on SWE-bench Verified, establishing SOTA among Qwen2.5 series. Moreover, SWE-focused training yields substantial out-of-domain improvements, including up to 12 points on mathematical reasoning and 5 points on science benchmarks, without degrading factual recall.\n\n---\n\n## 40. ARL-Tangram: Unleash the Resource Efficiency in Agentic Reinforcement Learning\n- **Authors**: Bangjun Xiao , Yihao Zhao , Xiangwei Deng , Shihua Yu , Yuxing Xiang , Huaqiu Liu , Qiying Wang , Liang Zhao , Hailin Zhang , Xuanzhe Liu , Xin Jin , Fuli Luo\n- **URL**: [https://arxiv.org/abs/2603.13019](https://arxiv.org/abs/2603.13019)\n- **Abstract**:\n> Agentic reinforcement learning (RL) has emerged as a transformative workload in cloud clusters, enabling large language models (LLMs) to solve complex problems through interactions with real world. However, unlike traditional RL, agentic RL demands substantial external cloud resources, e.g., CPUs for code execution and GPUs for reward models, that exist outside the primary training cluster. Existing agentic RL framework typically rely on static over-provisioning, i.e., resources are often tied to long-lived trajectories or isolated by tasks, which leads to severe resource inefficiency. We propose the action-level orchestration, and incorporate it into ARL-Tangram, a unified resource management system that enables fine-grained external resource sharing and elasticity. ARL-Tangram utilizes a unified action-level formulation and an elastic scheduling algorithm to minimize action completion time (ACT) while satisfying heterogeneous resource constraints. Further, heterogeneous resource managers are tailored to efficiently support the action-level execution on resources with heterogeneous characteristics and topologies. Evaluation on real-world agentic RL tasks demonstrates that ARL-Tangram improves average ACT by up to 4.3$\\times$, speeds up the step duration of RL training by up to 1.5$\\times$, and saves the external resources by up to 71.2$\\%$. This system has been deployed to support the training of the MiMo series models.\n\n---\n\n## 41. Fair Lung Disease Diagnosis from Chest CT via Gender-Adversarial Attention Multiple Instance Learning\n- **Authors**: Aditya Parikh , Aasa Feragen\n- **URL**: [https://arxiv.org/abs/2603.12988](https://arxiv.org/abs/2603.12988)\n- **Abstract**:\n> We present a fairness-aware framework for multi-class lung disease diagnosis from chest CT volumes, developed for the Fair Disease Diagnosis Challenge at the PHAROS-AIF-MIH Workshop (CVPR 2026). The challenge requires classifying CT scans into four categories -- Healthy, COVID-19, Adenocarcinoma, and Squamous Cell Carcinoma -- with performance measured as the average of per-gender macro F1 scores, explicitly penalizing gender-inequitable predictions. Our approach addresses two core difficulties: the sparse pathological signal across hundreds of slices, and a severe demographic imbalance compounded across disease class and gender. We propose an attention-based Multiple Instance Learning (MIL) model on a ConvNeXt backbone that learns to identify diagnostically relevant slices without slice-level supervision, augmented with a Gradient Reversal Layer (GRL) that adversarially suppresses gender-predictive structure in the learned scan representation. Training incorporates focal loss with label smoothing, stratified cross-validation over joint (class, gender) strata, and targeted oversampling of the most underrepresented subgroup. At inference, all five-fold checkpoints are ensembled with horizontal-flip test-time augmentation via soft logit voting and out-of-the-fold threshold optimization for robustness. Our model achieves a mean validation competition score of 0.685 (std - 0.030), with the best single fold reaching 0.759. All training and inference code is publicly available at this https URL\n\n---\n\n## 42. Is Human Annotation Necessary? Iterative MBR Distillation for Error Span Detection in Machine Translation\n- **Authors**: Boxuan Lyu , Haiyue Song , Zhi Qu\n- **URL**: [https://arxiv.org/abs/2603.12983](https://arxiv.org/abs/2603.12983)\n- **Abstract**:\n> Error Span Detection (ESD) is a crucial subtask in Machine Translation (MT) evaluation, aiming to identify the location and severity of translation errors. While fine-tuning models on human-annotated data improves ESD performance, acquiring such data is expensive and prone to inconsistencies among annotators. To address this, we propose a novel self-evolution framework based on Minimum Bayes Risk (MBR) decoding, named Iterative MBR Distillation for ESD, which eliminates the reliance on human annotations by leveraging an off-the-shelf LLM to generate pseudo-labels. Extensive experiments on the WMT Metrics Shared Task datasets demonstrate that models trained solely on these self-generated pseudo-labels outperform both unadapted base model and supervised baselines trained on human annotations at the system and span levels, while maintaining competitive sentence-level performance.\n\n---\n\n## 43. Efficient Real-World Autonomous Racing via Attenuated Residual Policy Optimization\n- **Authors**: Raphael Trumpp , Denis Hoornaert , Mirco Theile , Marco Caccamo\n- **URL**: [https://arxiv.org/abs/2603.12960](https://arxiv.org/abs/2603.12960)\n- **Abstract**:\n> Residual policy learning (RPL), in which a learned policy refines a static base policy using deep reinforcement learning (DRL), has shown strong performance across various robotic applications. Its effectiveness is particularly evident in autonomous racing, a domain that serves as a challenging benchmark for real-world DRL. However, deploying RPL-based controllers introduces system complexity and increases inference latency. We address this by introducing an extension of RPL named attenuated residual policy optimization ($\\alpha$-RPO). Unlike standard RPL, $\\alpha$-RPO yields a standalone neural policy by progressively attenuating the base policy, which initially serves to bootstrap learning. Furthermore, this mechanism enables a form of privileged learning, where the base policy is permitted to use sensor modalities not required for final deployment. We design $\\alpha$-RPO to integrate seamlessly with PPO, ensuring that the attenuated influence of the base controller is dynamically compensated during policy optimization. We evaluate $\\alpha$-RPO by building a framework for 1:10-scaled autonomous racing around it. In both simulation and zero-shot real-world transfer to Roboracer cars, $\\alpha$-RPO not only reduces system complexity but also improves driving performance compared to baselines - demonstrating its practicality for robotic deployment. Our code is available at: this https URL .\n\n---\n\n## 44. Delta1 with LLM: symbolic and neural integration for credible and explainable reasoning\n- **Authors**: Yang Xu , Jun Liu , Shuwei Chen , Chris Nugent , Hailing Guo\n- **URL**: [https://arxiv.org/abs/2603.12953](https://arxiv.org/abs/2603.12953)\n- **Abstract**:\n> Neuro-symbolic reasoning increasingly demands frameworks that unite the formal rigor of logic with the interpretability of large language models (LLMs). We introduce an end to end explainability by construction pipeline integrating the Automated Theorem Generator Delta1 based on the full triangular standard contradiction (FTSC) with LLMs. Delta1 deterministically constructs minimal unsatisfiable clause sets and complete theorems in polynomial time, ensuring both soundness and minimality by construction. The LLM layer verbalizes each theorem and proof trace into coherent natural language explanations and actionable insights. Empirical studies across health care, compliance, and regulatory domains show that Delta1 and LLM enables interpretable, auditable, and domain aligned reasoning. This work advances the convergence of logic, language, and learning, positioning constructive theorem generation as a principled foundation for neuro-symbolic explainable AI.\n\n---\n\n## 45. Thinking in Streaming Video\n- **Authors**: Zikang Liu , Longteng Guo , Handong Li , Ru Zhen , Xingjian He , Ruyi Ji , Xiaoming Ren , Yanhao Zhang , Haonan Lu , Jing Liu\n- **URL**: [https://arxiv.org/abs/2603.12938](https://arxiv.org/abs/2603.12938)\n- **Abstract**:\n> Real-time understanding of continuous video streams is essential for interactive assistants and multimodal agents operating in dynamic environments. However, most existing video reasoning approaches follow a batch paradigm that defers reasoning until the full video context is observed, resulting in high latency and growing computational cost that are incompatible with streaming scenarios. In this paper, we introduce ThinkStream, a framework for streaming video reasoning based on a Watch--Think--Speak paradigm that enables models to incrementally update their understanding as new video observations arrive. At each step, the model performs a short reasoning update and decides whether sufficient evidence has accumulated to produce a response. To support long-horizon streaming, we propose Reasoning-Compressed Streaming Memory (RCSM), which treats intermediate reasoning traces as compact semantic memory that replaces outdated visual tokens while preserving essential context. We further train the model using a Streaming Reinforcement Learning with Verifiable Rewards scheme that aligns incremental reasoning and response timing with the requirements of streaming interaction. Experiments on multiple streaming video benchmarks show that ThinkStream significantly outperforms existing online video models while maintaining low latency and memory usage. Code, models and data will be released at this https URL\n\n---\n\n## 46. Surprised by Attention: Predictable Query Dynamics for Time Series Anomaly Detection\n- **Authors**: Kadir-Kaan Özer , René Ebeling , Markus Enzweiler\n- **URL**: [https://arxiv.org/abs/2603.12916](https://arxiv.org/abs/2603.12916)\n- **Abstract**:\n> Multivariate time series anomalies often manifest as shifts in cross-channel dependencies rather than simple amplitude excursions. In autonomous driving, for instance, a steering command might be internally consistent but decouple from the resulting lateral acceleration. Residual-based detectors can miss such anomalies when flexible sequence models still reconstruct signals plausibly despite altered coordination. We introduce AxonAD, an unsupervised detector that treats multi-head attention query evolution as a short horizon predictable process. A gradient-updated reconstruction pathway is coupled with a history-only predictor that forecasts future query vectors from past context. This is trained via a masked predictor-target objective against an exponential moving average (EMA) target encoder. At inference, reconstruction error is combined with a tail-aggregated query mismatch score, which measures cosine deviation between predicted and target queries on recent timesteps. This dual approach provides sensitivity to structural dependency shifts while retaining amplitude-level detection. On proprietary in-vehicle telemetry with interval annotations and on the TSB-AD multi-variate suite (17 datasets, 180 series) with threshold-free and range-aware metrics, AxonAD improves ranking quality and temporal localization over strong baselines. Ablations confirm that query prediction and combined scoring are the primary drivers of the observed gains. Code is available at the URL this https URL .\n\n---\n\n## 47. Stake the Points: Structure-Faithful Instance Unlearning\n- **Authors**: Kiseong Hong , JungKyoo Shin , Eunwoo Kim\n- **URL**: [https://arxiv.org/abs/2603.12915](https://arxiv.org/abs/2603.12915)\n- **Abstract**:\n> Machine unlearning (MU) addresses privacy risks in pretrained models. The main goal of MU is to remove the influence of designated data while preserving the utility of retained knowledge. Achieving this goal requires preserving semantic relations among retained instances, which existing studies often overlook. We observe that without such preservation, models suffer from progressive structural collapse, undermining both the deletion-retention balance. In this work, we propose a novel structure-faithful framework that introduces stakes, i.e., semantic anchors that serve as reference points to maintain the knowledge structure. By leveraging these anchors, our framework captures and stabilizes the semantic organization of knowledge. Specifically, we instantiate the anchors from language-driven attribute descriptions encoded by a semantic encoder (e.g., CLIP). We enforce preservation of the knowledge structure via structure-aware alignment and regularization: the former aligns the organization of retained knowledge before and after unlearning around anchors, while the latter regulates updates to structure-critical parameters. Results from image classification, retrieval, and face recognition show average gains of 32.9%, 22.5%, and 19.3% in performance, balancing the deletion-retention trade-off and enhancing generalization.\n\n---\n\n## 48. FedBPrompt: Federated Domain Generalization Person Re-Identification via Body Distribution Aware Visual Prompts\n- **Authors**: Xin Xu , Weilong Li , Wei Liu , Wenke Huang , Zhixi Yu , Bin Yang , Xiaoying Liao , Kui Jiang\n- **URL**: [https://arxiv.org/abs/2603.12912](https://arxiv.org/abs/2603.12912)\n- **Abstract**:\n> Federated Domain Generalization for Person Re-Identification (FedDG-ReID) learns domain-invariant representations from decentralized data. While Vision Transformer (ViT) is widely adopted, its global attention often fails to distinguish pedestrians from high similarity backgrounds or diverse viewpoints -- a challenge amplified by cross-client distribution shifts in FedDG-ReID. To address this, we propose Federated Body Distribution Aware Visual Prompt (FedBPrompt), introducing learnable visual prompts to guide Transformer attention toward pedestrian-centric regions. FedBPrompt employs a Body Distribution Aware Visual Prompts Mechanism (BAPM) comprising: Holistic Full Body Prompts to suppress cross-client background noise, and Body Part Alignment Prompts to capture fine-grained details robust to pose and viewpoint variations. To mitigate high communication costs, we design a Prompt-based Fine-Tuning Strategy (PFTS) that freezes the ViT backbone and updates only lightweight prompts, significantly reducing communication overhead while maintaining adaptability. Extensive experiments demonstrate that BAPM effectively enhances feature discrimination and cross-domain generalization, while PFTS achieves notable performance gains within only a few aggregation rounds. Moreover, both BAPM and PFTS can be easily integrated into existing ViT-based FedDG-ReID frameworks, making FedBPrompt a flexible and effective solution for federated person re-identification. The code is available at this https URL .\n\n---\n\n## 49. Learning from Child-Directed Speech in Two-Language Scenarios: A French-English Case Study\n- **Authors**: Liel Binyamin , Elior Sulem\n- **URL**: [https://arxiv.org/abs/2603.12906](https://arxiv.org/abs/2603.12906)\n- **Abstract**:\n> Research on developmentally plausible language models has largely focused on English, leaving open questions about multilingual settings. We present a systematic study of compact language models by extending BabyBERTa to English-French scenarios under strictly size-matched data conditions, covering monolingual, bilingual, and cross-lingual settings. Our design contrasts two types of training corpora: (i) child-directed speech (about 2.5M tokens), following BabyBERTa and related work, and (ii) multi-domain corpora (about 10M tokens), extending the BabyLM framework to French. To enable fair evaluation, we also introduce new resources, including French versions of QAMR and QASRL, as well as English and French multi-domain corpora. We evaluate the models on both syntactic and semantic tasks and compare them with models trained on Wikipedia-only data. The results reveal context-dependent effects: training on Wikipedia consistently benefits semantic tasks, whereas child-directed speech improves grammatical judgments in monolingual settings. Bilingual pretraining yields notable gains for textual entailment, with particularly strong improvements for French. Importantly, similar patterns emerge across BabyBERTa, RoBERTa, and LTG-BERT, suggesting consistent trends across architectures.\n\n---\n\n## 50. Human-Centered Evaluation of an LLM-Based Process Modeling Copilot: A Mixed-Methods Study with Domain Experts\n- **Authors**: Chantale Lauer , Peter Pfeiffer , Nijat Mehdiyev\n- **URL**: [https://arxiv.org/abs/2603.12895](https://arxiv.org/abs/2603.12895)\n- **Abstract**:\n> Integrating Large Language Models (LLMs) into business process management tools promises to democratize Business Process Model and Notation (BPMN) modeling for non-experts. While automated frameworks assess syntactic and semantic quality, they miss human factors like trust, usability, and professional alignment. We conducted a mixed-methods evaluation of our proposed solution, an LLM-powered BPMN copilot, with five process modeling experts using focus groups and standardized questionnaires. Our findings reveal a critical tension between acceptable perceived usability (mean CUQ score: 67.2/100) and notably lower trust (mean score: 48.8\\%), with reliability rated as the most critical concern (M=1.8/5). Furthermore, we identified output-quality issues, prompting difficulties, and a need for the LLM to ask more in-depth clarifying questions about the process. We envision five use cases ranging from domain-expert support to enterprise quality assurance. We demonstrate the necessity of human-centered evaluation complementing automated benchmarking for LLM modeling agents.\n\n---\n\n## 51. Finite Difference Flow Optimization for RL Post-Training of Text-to-Image Models\n- **Authors**: David McAllister , Miika Aittala , Tero Karras , Janne Hellsten , Angjoo Kanazawa , Timo Aila , Samuli Laine\n- **URL**: [https://arxiv.org/abs/2603.12893](https://arxiv.org/abs/2603.12893)\n- **Abstract**:\n> Reinforcement learning (RL) has become a standard technique for post-training diffusion-based image synthesis models, as it enables learning from reward signals to explicitly improve desirable aspects such as image quality and prompt alignment. In this paper, we propose an online RL variant that reduces the variance in the model updates by sampling paired trajectories and pulling the flow velocity in the direction of the more favorable image. Unlike existing methods that treat each sampling step as a separate policy action, we consider the entire sampling process as a single action. We experiment with both high-quality vision language models and off-the-shelf quality metrics for rewards, and evaluate the outputs using a broad set of metrics. Our method converges faster and yields higher output quality and prompt alignment than previous approaches.\n\n---\n\n## 52. Team LEYA in 10th ABAW Competition: Multimodal Ambivalence/Hesitancy Recognition Approach\n- **Authors**: Elena Ryumina (1), Alexandr Axyonov (1), Dmitry Sysoev (2), Timur Abdulkadirov (2), Kirill Almetov (2), Yulia Morozova (2), Dmitry Ryumin (1 and 2) ((1) St. Petersburg Federal Research Center of the Russian Academy of Sciences, St. Petersburg, Russia, (2) HSE University, St. Petersburg, Russia)\n- **URL**: [https://arxiv.org/abs/2603.12848](https://arxiv.org/abs/2603.12848)\n- **Abstract**:\n> Ambivalence/hesitancy recognition in unconstrained videos is a challenging problem due to the subtle, multimodal, and context-dependent nature of this behavioral state. In this paper, a multimodal approach for video-level ambivalence/hesitancy recognition is presented for the 10th ABAW Competition. The proposed approach integrates four complementary modalities: scene, face, audio, and text. Scene dynamics are captured with a VideoMAE-based model, facial information is encoded through emotional frame-level embeddings aggregated by statistical pooling, acoustic representations are extracted with EmotionWav2Vec2.0 and processed by a Mamba-based temporal encoder, and linguistic cues are modeled using fine-tuned transformer-based text models. The resulting unimodal embeddings are further combined using multimodal fusion models, including prototype-augmented variants. Experiments on the BAH corpus demonstrate clear gains of multimodal fusion over all unimodal baselines. The best unimodal configuration achieved an average MF1 of 70.02%, whereas the best multimodal fusion model reached 83.25%. The highest final test performance, 71.43%, was obtained by an ensemble of five prototype-augmented fusion models. The obtained results highlight the importance of complementary multimodal cues and robust fusion strategies for ambivalence/hesitancy recognition.\n\n---\n\n## 53. Hierarchical Reference Sets for Robust Unsupervised Detection of Scattered and Clustered Outliers\n- **Authors**: Yiqun Zhang , Zexi Tan , Xiaopeng Luo , Yunlin Liu\n- **URL**: [https://arxiv.org/abs/2603.12847](https://arxiv.org/abs/2603.12847)\n- **Abstract**:\n> Most real-world IoT data analysis tasks, such as clustering and anomaly event detection, are unsupervised and highly susceptible to the presence of outliers. In addition to sporadic scattered outliers caused by factors such as faulty sensor readings, IoT systems often exhibit clustered outliers. These occur when multiple devices or nodes produce similar anomalous measurements, for instance, owing to localized interference, emerging security threats, or regional false alarms, forming micro-clusters. These clustered outliers can be easily mistaken for normal behavior because of their relatively high local density, thereby obscuring the detection of both scattered and contextual anomalies. To address this, we propose a novel outlier detection paradigm that leverages the natural neighboring relationships using graph structures. This facilitates multi-perspective anomaly evaluation by incorporating reference sets at both local and global scales derived from the graph. Our approach enables the effective recognition of scattered outliers without interference from clustered anomalies, whereas the graph structure simultaneously helps reflect and isolate clustered outlier groups. Extensive experiments, including comparative performance analysis, ablation studies, validation on downstream clustering tasks, and evaluation of hyperparameter sensitivity, demonstrate the efficacy of the proposed method. The source code is available at this https URL .\n\n---\n\n## 54. DAST: A Dual-Stream Voice Anonymization Attacker with Staged Training\n- **Authors**: Ridwan Arefeen , Xiaoxiao Miao , Rong Tong , Aik Beng Ng , Simon See , Timothy Liu\n- **URL**: [https://arxiv.org/abs/2603.12840](https://arxiv.org/abs/2603.12840)\n- **Abstract**:\n> Voice anonymization masks vocal traits while preserving linguistic content, which may still leak speaker-specific patterns. To assess and strengthen privacy evaluation, we propose a dual-stream attacker that fuses spectral and self-supervised learning features via parallel encoders with a three-stage training strategy. Stage I establishes foundational speaker-discriminative representations. Stage II leverages the shared identity-transformation characteristics of voice conversion and anonymization, exposing the model to diverse converted speech to build cross-system robustness. Stage III provides lightweight adaptation to target anonymized data. Results on the VoicePrivacy Attacker Challenge (VPAC) dataset demonstrate that Stage II is the primary driver of generalization, enabling strong attacking performance on unseen anonymization datasets. With Stage III, fine-tuning on only 10\\% of the target anonymization dataset surpasses current state-of-the-art attackers in terms of EER.\n\n---\n\n## 55. Mask2Flow-TSE: Two-Stage Target Speaker Extraction with Masking and Flow Matching\n- **Authors**: Junwon Moon , Hyunjin Choi , Hansol Park , Heeseung Kim , Kyuhong Shim\n- **URL**: [https://arxiv.org/abs/2603.12837](https://arxiv.org/abs/2603.12837)\n- **Abstract**:\n> Target speaker extraction (TSE) extracts the target speaker's voice from overlapping speech mixtures given a reference utterance. Existing approaches typically fall into two categories: discriminative and generative. Discriminative methods apply time-frequency masking for fast inference but often over-suppress the target signal, while generative methods synthesize high-quality speech at the cost of numerous iterative steps. We propose Mask2Flow-TSE, a two-stage framework combining the strengths of both paradigms. The first stage applies discriminative masking for coarse separation, and the second stage employs flow matching to refine the output toward target speech. Unlike generative approaches that synthesize speech from Gaussian noise, our method starts from the masked spectrogram, enabling high-quality reconstruction in a single inference step. Experiments show that Mask2Flow-TSE achieves comparable performance to existing generative TSE methods with approximately 85M parameters.\n\n---\n\n## 56. Hierarchical Dual-Change Collaborative Learning for UAV Scene Change Captioning\n- **Authors**: Fuhai Chen , Pengpeng Huang , Junwen Wu , Hehong Zhang , Shiping Wang , Xiaoguang Ma , Xuri Ge\n- **URL**: [https://arxiv.org/abs/2603.12832](https://arxiv.org/abs/2603.12832)\n- **Abstract**:\n> This paper proposes a novel task for UAV scene understanding - UAV Scene Change Captioning (UAV-SCC) - which aims to generate natural language descriptions of semantic changes in dynamic aerial imagery captured from a movable viewpoint. Unlike traditional change captioning that mainly describes differences between image pairs captured from a fixed camera viewpoint over time, UAV scene change captioning focuses on image-pair differences resulting from both temporal and spatial scene variations dynamically captured by a moving camera. The key challenge lies in understanding viewpoint-induced scene changes from UAV image pairs that share only partially overlapping scene content due to viewpoint shifts caused by camera rotation, while effectively exploiting the relative orientation between the two images. To this end, we propose a Hierarchical Dual-Change Collaborative Learning (HDC-CL) method for UAV scene change captioning. In particular, a novel transformer, \\emph{i.e.} Dynamic Adaptive Layout Transformer (DALT) is designed to adaptively model diverse spatial layouts of the image pair, where the interrelated features derived from the overlapping and non-overlapping regions are learned within the flexible and unified encoding layer. Furthermore, we propose a Hierarchical Cross-modal Orientation Consistency Calibration (HCM-OCC) method to enhance the model's sensitivity to viewpoint shift directions, enabling more accurate change captioning. To facilitate in-depth research on this task, we construct a new benchmark dataset, named UAV-SCC dataset, for UAV scene change captioning. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance on this task. The dataset and code will be publicly released upon acceptance of this paper.\n\n---\n\n## 57. Residual SODAP: Residual Self-Organizing Domain-Adaptive Prompting with Structural Knowledge Preservation for Continual Learning\n- **Authors**: Gyutae Oh , Jungwoo Bae , Jitae Shin\n- **URL**: [https://arxiv.org/abs/2603.12816](https://arxiv.org/abs/2603.12816)\n- **Abstract**:\n> Continual learning (CL) suffers from catastrophic forgetting, which is exacerbated in domain-incremental learning (DIL) where task identifiers are unavailable and storing past data is infeasible. While prompt-based CL (PCL) adapts representations with a frozen backbone, we observe that prompt-only improvements are often insufficient due to suboptimal prompt selection and classifier-level instability under domain shifts. We propose Residual SODAP, which jointly performs prompt-based representation adaptation and classifier-level knowledge preservation. Our framework combines $\\alpha$-entmax sparse prompt selection with residual aggregation, data-free distillation with pseudo-feature replay, prompt-usage--based drift detection, and uncertainty-aware multi-loss balancing. Across three DIL benchmarks without task IDs or extra data storage, Residual SODAP achieves state-of-the-art AvgACC/AvgF of 0.850/0.047 (DR), 0.760/0.031 (Skin Cancer), and 0.995/0.003 (CORe50).\n\n---\n\n## 58. Cheers: Decoupling Patch Details from Semantic Representations Enables Unified Multimodal Comprehension and Generation\n- **Authors**: Yichen Zhang , Da Peng , Zonghao Guo , Zijian Zhang , Xuesong Yang , Tong Sun , Shichu Sun , Yidan Zhang , Yanghao Li , Haiyan Zhao , Wang Xu , Qi Shi , Yangang Sun , Chi Chen , Shuo Wang , Yukun Yan , Xu Han , Qiang Ma , Wei Ke , Liang Wang , Zhiyuan Liu , Maosong Sun\n- **URL**: [https://arxiv.org/abs/2603.12793](https://arxiv.org/abs/2603.12793)\n- **Abstract**:\n> A recent cutting-edge topic in multimodal modeling is to unify visual comprehension and generation within a single model. However, the two tasks demand mismatched decoding regimes and visual representations, making it non-trivial to jointly optimize within a shared feature space. In this work, we present Cheers, a unified multimodal model that decouples patch-level details from semantic representations, thereby stabilizing semantics for multimodal understanding and improving fidelity for image generation via gated detail residuals. Cheers includes three key components: (i) a unified vision tokenizer that encodes and compresses image latent states into semantic tokens for efficient LLM conditioning, (ii) an LLM-based Transformer that unifies autoregressive decoding for text generation and diffusion decoding for image generation, and (iii) a cascaded flow matching head that decodes visual semantics first and then injects semantically gated detail residuals from the vision tokenizer to refine high-frequency content. Experiments on popular benchmarks demonstrate that Cheers matches or surpasses advanced UMMs in both visual understanding and generation. Cheers also achieves 4x token compression, enabling more efficient high-resolution image encoding and generation. Notably, Cheers outperforms the Tar-1.5B on the popular benchmarks GenEval and MMBench, while requiring only 20% of the training cost, indicating effective and efficient (i.e., 4x token compression) unified multimodal modeling. We will release all code and data for future research.\n\n---\n\n## 59. The RIGID Framework: Research-Integrated, Generative AI-Mediated Instructional Design\n- **Authors**: Yerin Kwak , Zachary A. Pardos\n- **URL**: [https://arxiv.org/abs/2603.12781](https://arxiv.org/abs/2603.12781)\n- **Abstract**:\n> Instructional Design (ID) often faces challenges in incorporating research-based knowledge and pedagogical best practices. Although educational researchers and government agencies emphasize grounding ID in evidence, integrating research findings into everyday design workflows is often complex, as it requires considering multiple context-specific demands and constraints. To address this persistent gap, this paper explores how research in the learning sciences (LS) can be systematically integrated across ID workflows and how recent advances in generative AI can help operationalize this integration. While ID and LS share a commitment to improving learning experiences through design-oriented approaches in authentic contexts, structured integration between the two fields remains limited, leaving their complementary insights underutilized. We present RIGID (Research-Integrated, Generative AI-Mediated Instructional Design), a unified framework that integrates LS research across ID workflows spanning analysis, design, implementation, and evaluation phases, while leveraging generative AI to mediate this integration at each stage. The RIGID framework provides a systematic approach for enabling research-integrated instructional design that is both operational and context-sensitive, while preserving the central role of human expertise.\n\n---\n\n## 60. Empowering Semantic-Sensitive Underwater Image Enhancement with VLM\n- **Authors**: Guodong Fan , Shengning Zhou , Genji Yuan , Huiyu Li , Jingchun Zhou , Jinjiang Li\n- **URL**: [https://arxiv.org/abs/2603.12773](https://arxiv.org/abs/2603.12773)\n- **Abstract**:\n> In recent years, learning-based underwater image enhancement (UIE) techniques have rapidly evolved. However, distribution shifts between high-quality enhanced outputs and natural images can hinder semantic cue extraction for downstream vision tasks, thereby limiting the adaptability of existing enhancement models. To address this challenge, this work proposes a new learning mechanism that leverages Vision-Language Models (VLMs) to empower UIE models with semantic-sensitive capabilities. To be concrete, our strategy first generates textual descriptions of key objects from a degraded image via VLMs. Subsequently, a text-image alignment model remaps these relevant descriptions back onto the image to produce a spatial semantic guidance map. This map then steers the UIE network through a dual-guidance mechanism, which combines cross-attention and an explicit alignment loss. This forces the network to focus its restorative power on semantic-sensitive regions during image reconstruction, rather than pursuing a globally uniform improvement, thereby ensuring the faithful restoration of key object features. Experiments confirm that when our strategy is applied to different UIE baselines, significantly boosts their performance on perceptual quality metrics as well as enhances their performance on detection and segmentation tasks, validating its effectiveness and adaptability.\n\n---\n\n## 61. FC-Track: Overlap-Aware Post-Association Correction for Online Multi-Object Tracking\n- **Authors**: Cheng Ju , Zejing Zhao , Akio Namiki\n- **URL**: [https://arxiv.org/abs/2603.12758](https://arxiv.org/abs/2603.12758)\n- **Abstract**:\n> Reliable multi-object tracking (MOT) is essential for robotic systems operating in complex and dynamic environments. Despite recent advances in detection and association, online MOT methods remain vulnerable to identity switches caused by frequent occlusions and object overlap, where incorrect associations can propagate over time and degrade tracking reliability. We present a lightweight post-association correction framework (FC-Track) for online MOT that explicitly targets overlap-induced mismatches during inference. The proposed method suppresses unreliable appearance updates under high-overlap conditions using an Intersection over Area (IoA)-based filtering strategy, and locally corrects detection-to-tracklet mismatches through appearance similarity comparison within overlapped tracklet pairs. By preventing short-term mismatches from propagating, our framework effectively mitigates long-term identity switches without resorting to global optimization or re-identification. The framework operates online without global optimization or re-identification, making it suitable for real-time robotic applications. We achieve 81.73 MOTA, 82.81 IDF1, and 66.95 HOTA on the MOT17 test set with a running speed of 5.7 FPS, and 77.52 MOTA, 80.90 IDF1, and 65.67 HOTA on the MOT20 test set with a running speed of 0.6 FPS. Specifically, our framework FC-Track produces only 29.55% long-term identity switches, which is substantially lower than existing online trackers. Meanwhile, our framework maintains state-of-the-art performance on the MOT20 benchmark.\n\n---\n\n## 62. TaoBench: Do Automated Theorem Prover LLMs Generalize Beyond MathLib?\n- **Authors**: Alexander K Taylor , Junyi Zhang , Ethan Ji , Vigyan Sahai , Haikang Deng , Yuanzhou Chen , Yifan Yuan , Di Wu , Jia-Chen Gu , Kai-Wei Chang , Nanyun Peng , Amit Sahai , Wei Wang\n- **URL**: [https://arxiv.org/abs/2603.12744](https://arxiv.org/abs/2603.12744)\n- **Abstract**:\n> Automated theorem proving (ATP) benchmarks largely consist of problems formalized in MathLib, so current ATP training and evaluation are heavily biased toward MathLib's definitional framework. However, frontier mathematics is often exploratory and prototype-heavy, relying on bespoke constructions that deviate from standard libraries. In this work, we evaluate the robustness of current ATP systems when applied to a novel definitional framework, specifically examining the performance gap between standard library problems and bespoke mathematical constructions. We introduce TaoBench, an undergraduate-level benchmark derived from Terence Tao's Analysis I, which formalizes analysis by constructing core mathematical concepts from scratch, without relying on standard Mathlib definitions, as well as by mixing from-scratch and MathLib constructions. For fair evaluation, we build an agentic pipeline that automatically extracts a compilable, self-contained local environment for each problem. To isolate the effect of definitional frameworks, we additionally translate every problem into a mathematically equivalent Mathlib formulation, yielding paired TaoBench-Mathlib statements for direct comparison. While state-of-the-art ATP models perform capably within the MathLib framework, performance drops by an average of roughly 26% on the definitionally equivalent Tao formulation. This indicates that the main bottleneck is limited generalization across definitional frameworks rather than task difficulty. TaoBench thus highlights a gap between benchmark performance and applicability, and provides a concrete foundation for developing and testing provers better aligned with research mathematics.\n\n---\n\n## 63. MoKus: Leveraging Cross-Modal Knowledge Transfer for Knowledge-Aware Concept Customization\n- **Authors**: Chenyang Zhu , Hongxiang Li , Xiu Li , Long Chen\n- **URL**: [https://arxiv.org/abs/2603.12743](https://arxiv.org/abs/2603.12743)\n- **Abstract**:\n> Concept customization typically binds rare tokens to a target concept. Unfortunately, these approaches often suffer from unstable performance as the pretraining data seldom contains these rare tokens. Meanwhile, these rare tokens fail to convey the inherent knowledge of the target concept. Consequently, we introduce Knowledge-aware Concept Customization, a novel task aiming at binding diverse textual knowledge to target visual concepts. This task requires the model to identify the knowledge within the text prompt to perform high-fidelity customized generation. Meanwhile, the model should efficiently bind all the textual knowledge to the target concept. Therefore, we propose MoKus, a novel framework for knowledge-aware concept customization. Our framework relies on a key observation: cross-modal knowledge transfer, where modifying knowledge within the text modality naturally transfers to the visual modality during generation. Inspired by this observation, MoKus contains two stages: (1) In visual concept learning, we first learn the anchor representation to store the visual information of the target concept. (2) In textual knowledge updating, we update the answer for the knowledge queries to the anchor representation, enabling high-fidelity customized generation. To further comprehensively evaluate our proposed MoKus on the new task, we introduce the first benchmark for knowledge-aware concept customization: KnowCusBench. Extensive evaluations have demonstrated that MoKus outperforms state-of-the-art methods. Moreover, the cross-model knowledge transfer allows MoKus to be easily extended to other knowledge-aware applications like virtual concept creation and concept erasure. We also demonstrate the capability of our method to achieve improvements on world knowledge benchmarks.\n\n---\n\n## 64. SRAM-Based Compute-in-Memory Accelerator for Linear-decay Spiking Neural Networks\n- **Authors**: Hongyang Shang , Shuai Dong , Yahan Yang , Junyi Yang , Peng Zhou , Arindam Basu\n- **URL**: [https://arxiv.org/abs/2603.12739](https://arxiv.org/abs/2603.12739)\n- **Abstract**:\n> Spiking Neural Networks (SNNs) have emerged as a biologically inspired alternative to conventional deep networks, offering event-driven and energy-efficient computation. However, their throughput remains constrained by the serial update of neuron membrane states. While many hardware accelerators and Compute-in-Memory (CIM) architectures efficiently parallelize the synaptic operation (W x I) achieving O(1) complexity for matrix-vector multiplication, the subsequent state update step still requires O(N) time to refresh all neuron membrane potentials. This mismatch makes state update the dominant latency and energy bottleneck in SNN inference. To address this challenge, we propose an SRAM-based CIM for SNN with Linear Decay Leaky Integrate-and-Fire (LD-LIF) Neuron that co-optimizes algorithm and hardware. At the algorithmic level, we replace the conventional exponential membrane decay with a linear decay approximation, converting costly multiplications into simple additions while accuracy drops only around 1%. At the architectural level, we introduce an in-memory parallel update scheme that performs in-place decay directly within the SRAM array, eliminating the need for global sequential updates. Evaluated on benchmark SNN workloads, the proposed method achieves a 1.1 x to 16.7 x reduction of SOP energy consumption, while providing 15.9 x to 69 x more energy efficiency, with negligible accuracy loss relative to original decay models. This work highlights that beyond accelerating the (W x I) computation, optimizing state-update dynamics within CIM architectures is essential for scalable, low-power, and real-time neuromorphic processing.\n\n---\n\n## 65. Graph In-Context Operator Networks for Generalizable Spatiotemporal Prediction\n- **Authors**: Chenghan Wu , Zongmin Yu , Boai Sun , Liu Yang\n- **URL**: [https://arxiv.org/abs/2603.12725](https://arxiv.org/abs/2603.12725)\n- **Abstract**:\n> In-context operator learning enables neural networks to infer solution operators from contextual examples without weight updates. While prior work has demonstrated the effectiveness of this paradigm in leveraging vast datasets, a systematic comparison against single-operator learning using identical training data has been absent. We address this gap through controlled experiments comparing in-context operator learning against classical operator learning (single-operator models trained without contextual examples), under the same training steps and dataset. To enable this investigation on real-world spatiotemporal systems, we propose GICON (Graph In-Context Operator Network), combining graph message passing for geometric generalization with example-aware positional encoding for cardinality generalization. Experiments on air quality prediction across two Chinese regions show that in-context operator learning outperforms classical operator learning on complex tasks, generalizing across spatial domains and scaling robustly from few training examples to 100 at inference.\n\n---\n\n## 66. CognitionCapturerPro: Towards High-Fidelity Visual Decoding from EEG/MEG via Multi-modal Information and Asymmetric Alignment\n- **Authors**: Kaifan Zhang , Lihuo He , Junjie Ke , Yuqi Ji , Lukun Wu , Lizi Wang , Xinbo Gao\n- **URL**: [https://arxiv.org/abs/2603.12722](https://arxiv.org/abs/2603.12722)\n- **Abstract**:\n> Visual stimuli reconstruction from EEG remains challenging due to fidelity loss and representation shift. We propose CognitionCapturerPro, an enhanced framework that integrates EEG with multi-modal priors (images, text, depth, and edges) via collaborative training. Our core contributions include an uncertainty-weighted similarity scoring mechanism to quantify modality-specific fidelity and a fusion encoder for integrating shared representations. By employing a simplified alignment module and a pre-trained diffusion model, our method significantly outperforms the original CognitionCapturer on the THINGS-EEG dataset, improving Top-1 and Top-5 retrieval accuracy by 25.9% and 10.6%, respectively. Code is available at: this https URL .\n\n---\n\n## 67. CMHANet: A Cross-Modal Hybrid Attention Network for Point Cloud Registration\n- **Authors**: Dongxu Zhang , Yingsen Wang , Yiding Sun , Haoran Xu , Peilin Fan , Jihua Zhu\n- **URL**: [https://arxiv.org/abs/2603.12721](https://arxiv.org/abs/2603.12721)\n- **Abstract**:\n> Robust point cloud registration is a fundamental task in 3D computer vision and geometric deep learning, essential for applications such as large-scale 3D reconstruction, augmented reality, and scene understanding. However, the performance of established learning-based methods often degrades in complex, real world scenarios characterized by incomplete data, sensor noise, and low overlap regions. To address these limitations, we propose CMHANet, a novel Cross-Modal Hybrid Attention Network. Our method integrates the fusion of rich contextual information from 2D images with the geometric detail of 3D point clouds, yielding a comprehensive and resilient feature representation. Furthermore, we introduce an innovative optimization function based on contrastive learning, which enforces geometric consistency and significantly improves the model's robustness to noise and partial observations. We evaluated CMHANet on the 3DMatch and the challenging 3DLoMatch datasets. \\rev{Additionally, zero-shot evaluations on the TUM RGB-D SLAM dataset verify the model's generalization capability to unseen domains.} The experimental results demonstrate that our method achieves substantial improvements in both registration accuracy and overall robustness, outperforming current techniques. We also release our code in \\href{ this https URL }{ this https URL }.\n\n---\n\n## 68. IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registration\n- **Authors**: Dongxu Zhang , Jihua Zhu , Shiqi Li , Wenbiao Yan , Haoran Xu , Peilin Fan , Huimin Lu\n- **URL**: [https://arxiv.org/abs/2603.12719](https://arxiv.org/abs/2603.12719)\n- **Abstract**:\n> Point cloud registration (PCR) is a fundamental task in 3D vision and provides essential support for applications such as autonomous driving, robotics, and environmental modeling. Despite its widespread use, existing methods often fail when facing real-world challenges like heavy noise, significant occlusions, and large-scale transformations. These limitations frequently result in compromised registration accuracy and insufficient robustness in complex environments. In this paper, we propose IGASA as a novel registration framework constructed upon a Hierarchical Pyramid Architecture (HPA) designed for robust multi-scale feature extraction and fusion. The framework integrates two pivotal components consisting of the Hierarchical Cross-Layer Attention (HCLA) module and the Iterative Geometry-Aware Refinement (IGAR) module. The HCLA module utilizes skip attention mechanisms to align multi-resolution features and enhance local geometric consistency. Simultaneously, the IGAR module is designed for the fine matching phase by leveraging reliable correspondences established during coarse matching. This synergistic integration within the architecture allows IGASA to adapt effectively to diverse point cloud structures and intricate transformations. We evaluate the performance of IGASA on four widely recognized benchmark datasets including 3D(Lo)Match, KITTI, and nuScenes. Our extensive experiments consistently demonstrate that IGASA significantly surpasses state-of-the-art methods and achieves notable improvements in registration accuracy. This work provides a robust foundation for advancing point cloud registration techniques while offering valuable insights for practical 3D vision applications. The code for IGASA is available in \\href{ this https URL }{ this https URL }.\n\n---\n\n## 69. Altered Thoughts, Altered Actions: Probing Chain-of-Thought Vulnerabilities in VLA Robotic Manipulation\n- **Authors**: Tuan Duong Trinh , Naveed Akhtar , Basim Azam\n- **URL**: [https://arxiv.org/abs/2603.12717](https://arxiv.org/abs/2603.12717)\n- **Abstract**:\n> Recent Vision-Language-Action (VLA) models increasingly adopt chain-of-thought (CoT) reasoning, generating a natural-language plan before decoding motor commands. This internal text channel between the reasoning module and the action decoder has received no adversarial scrutiny. We ask: which properties of this intermediate plan does the action decoder actually rely on, and can targeted corruption of the reasoning trace alone -- with all inputs left intact -- degrade a robot's physical task performance? We design a taxonomy of seven text corruptions organized into three attacker tiers (blind noise, mechanical-semantic, and LLM-adaptive) and apply them to a state-of-the-art reasoning VLA across 40 LIBERO tabletop manipulation tasks. Our results reveal a striking asymmetry: substituting object names in the reasoning trace reduces overall success rate by 8.3~percentage points (pp) -- reaching $-$19.3~pp on goal-conditioned tasks and $-$45~pp on individual tasks -- whereas sentence reordering, spatial-direction reversal, token noise, and even a 70B-parameter LLM crafting plausible-but-wrong plans all have negligible impact (within $\\pm$4~pp). This asymmetry indicates that the action decoder depends on entity-reference integrity rather than reasoning quality or sequential structure. Notably, a sophisticated LLM-based attacker underperforms simple mechanical object-name substitution, because preserving plausibility inadvertently retains the entity-grounding structure the decoder needs. A cross-architecture control using a non-reasoning VLA confirms the vulnerability is exclusive to reasoning-augmented models, while instruction-level attacks degrade both architectures -- establishing that the internal reasoning trace is a distinct and stealthy threat vector invisible to input-validation defenses.\n\n---\n\n## 70. Cost-Efficient Multimodal LLM Inference via Cross-Tier GPU Heterogeneity\n- **Authors**: Donglin Yu\n- **URL**: [https://arxiv.org/abs/2603.12707](https://arxiv.org/abs/2603.12707)\n- **Abstract**:\n> Multimodal large language model (MLLM) inference splits into two phases with opposing hardware demands: vision encoding is compute-bound, while language generation is memory-bandwidth-bound. We show that under standard transformer KV caching, the modality boundary (between vision encoder and language model) minimizes cross-device transfer among all partition points that preserve standard stage-based execution. Partitioning here reduces transfer complexity from $O(L * s_ctx)$ bytes (GB-scale KV caches under stage-level disaggregation) to $O(N_v * d)$ bytes (MB-scale embeddings), an O(L) reduction where L is the transformer depth. The result holds across attention mechanisms (MHA/GQA), dynamic vision resolutions, and model scales, and the advantage grows as models deepen. A direct implication is that existing stage-level disaggregation systems are constrained to high-bandwidth interconnects (e.g., NVLink), whereas modality-level disaggregation enables cross-tier heterogeneous serving over commodity PCIe. A closed-form cost model shows that heterogeneous deployment is cost-optimal under phase-separable workloads (predicts 31.4% savings; observed 40.6%). We build HeteroServe, a phase-aware runtime with modality-level partitioning and cross-tier scheduling, and evaluate it on LLaVA-1.5-7B and Qwen2.5-VL against vLLM v0.3.0. On identical 4xA100 hardware, engine optimizations raise throughput by up to 54%. Under a fixed budget, a heterogeneous cluster (\\$38k) improves Tokens/\\$ by 37% over a homogeneous baseline (\\$64k) without degrading latency.\n\n---\n\n## 71. Seeing Eye to Eye: Enabling Cognitive Alignment Through Shared First-Person Perspective in Human-AI Collaboration\n- **Authors**: Zhuyu Teng , Pei Chen , Yichen Cai , Ruoqing Lu , Zhaoqu Jiang , Jiayang Li , Weitao You , Lingyun Sun\n- **URL**: [https://arxiv.org/abs/2603.12701](https://arxiv.org/abs/2603.12701)\n- **Abstract**:\n> Despite advances in multimodal AI, current vision-based assistants often remain inefficient in collaborative tasks. We identify two key gulfs: a communication gulf, where users must translate rich parallel intentions into verbal commands due to the channel mismatch , and an understanding gulf, where AI struggles to interpret subtle embodied cues. To address these, we propose Eye2Eye, a framework that leverages first-person perspective as a channel for human-AI cognitive alignment. It integrates three components: (1) joint attention coordination for fluid focus alignment, (2) revisable memory to maintain evolving common ground, and (3) reflective feedback allowing users to clarify and refine AI's understanding. We implement this framework in an AR prototype and evaluate it through a user study and a post-hoc pipeline evaluation. Results show that Eye2Eye significantly reduces task completion time and interaction load while increasing trust, demonstrating its components work in concert to improve collaboration.\n\n---\n\n## 72. HSEmotion Team at ABAW-10 Competition: Facial Expression Recognition, Valence-Arousal Estimation, Action Unit Detection and Fine-Grained Violence Classification\n- **Authors**: Andrey V. Savchenko , Kseniia Tsypliakova\n- **URL**: [https://arxiv.org/abs/2603.12693](https://arxiv.org/abs/2603.12693)\n- **Abstract**:\n> This article presents our results for the 10th Affective Behavior Analysis in-the-Wild (ABAW) competition. For frame-wise facial emotion understanding tasks (frame-wise facial expression recognition, valence-arousal estimation, action unit detection), we propose a fast approach based on facial embedding extraction with pre-trained EfficientNet-based emotion recognition models. If the latter model's confidence exceeds a threshold, its prediction is used. Otherwise, we feed embeddings into a simple multi-layered perceptron trained on the AffWild2 dataset. Estimated class-level scores are smoothed in a sliding window of fixed size to mitigate noise in frame-wise predictions. For the fine-grained violence detection task, we examine several pre-trained architectures for frame embeddings and their aggregation for video classification. Experimental results on four tasks from the ABAW challenge demonstrate that our approach significantly improves validation metrics over existing baselines.\n\n---\n\n## 73. Federated Hierarchical Clustering with Automatic Selection of Optimal Cluster Numbers\n- **Authors**: Yue Zhang , Chuanlong Qiu , Xinfa Liao , Yiqun Zhang\n- **URL**: [https://arxiv.org/abs/2603.12684](https://arxiv.org/abs/2603.12684)\n- **Abstract**:\n> Federated Clustering (FC) is an emerging and promising solution in exploring data distribution patterns from distributed and privacy-protected data in an unsupervised manner. Existing FC methods implicitly rely on the assumption that clients are with a known number of uniformly sized clusters. However, the true number of clusters is typically unknown, and cluster sizes are naturally imbalanced in real scenarios. Furthermore, the privacy-preserving transmission constraints in federated learning inevitably reduce usable information, making the development of robust and accurate FC extremely challenging. Accordingly, we propose a novel FC framework named Fed-$k^*$-HC, which can automatically determine an optimal number of clusters $k^*$ based on the data distribution explored through hierarchical clustering. To obtain the global data distribution for $k^*$ determination, we let each client generate micro-subclusters. Their prototypes are then uploaded to the server for hierarchical merging. The density-based merging design allows exploring clusters of varying sizes and shapes, and the progressive merging process can self-terminate according to the neighboring relationships among the prototypes to determine $k^*$. Extensive experiments on diverse datasets demonstrate the FC capability of the proposed Fed-$k^*$-HC in accurately exploring a proper number of clusters.\n\n---\n\n## 74. Experimental evidence of progressive ChatGPT models self-convergence\n- **Authors**: Konstantinos F. Xylogiannopoulos , Petros Xanthopoulos , Panagiotis Karampelas , Georgios A. Bakamitsos\n- **URL**: [https://arxiv.org/abs/2603.12683](https://arxiv.org/abs/2603.12683)\n- **Abstract**:\n> Large Language Models (LLMs) that undergo recursive training on synthetically generated data are susceptible to model collapse, a phenomenon marked by the generation of meaningless output. Existing research has examined this issue from either theoretical or empirical perspectives, often focusing on a single model trained recursively on its own outputs. While prior studies have cautioned against the potential degradation of LLM output quality under such conditions, no longitudinal investigation has yet been conducted to assess this effect over time. In this study, we employ a text similarity metric to evaluate different ChatGPT models' capacity to generate diverse textual outputs. Our findings indicate a measurable decline of recent ChatGPT releases' ability to produce varied text, even when explicitly prompted to do so, by setting the temperature parameter to one. The observed reduction in output diversity may be attributed to the influence of the amounts of synthetic data incorporated within their training datasets as the result of internet infiltration by LLM generated data. The phenomenon is defined as model self-convergence because of the gradual increase of similarities of produced texts among different ChatGPT versions.\n\n---\n\n## 75. MetaKE: Meta-learning Aligned Knowledge Editing via Bi-level Optimization\n- **Authors**: Shuxin Liu , Ou Wu\n- **URL**: [https://arxiv.org/abs/2603.12677](https://arxiv.org/abs/2603.12677)\n- **Abstract**:\n> Knowledge editing (KE) aims to precisely rectify specific knowledge in Large Language Models (LLMs) without disrupting general capabilities. State-of-the-art methods suffer from an open-loop control mismatch. We identify a critical \"Semantic-Execution Disconnect\": the semantic target is derived independently without feedback from the downstream's feasible region. This misalignment often causes valid semantic targets to fall within the prohibited space, resulting in gradient truncation and editing failure. To bridge this gap, we propose MetaKE (Meta-learning Aligned Knowledge Editing), a new framework that reframes KE as a bi-level optimization problem. Departing from static calculation, MetaKE treats the edit target as a learnable meta-parameter: the upper-level optimizer seeks a feasible target to maximize post-edit performance, while the lower-level solver executes the editing. To address the challenge of differentiating through complex solvers, we derive a Structural Gradient Proxy, which explicitly backpropagates editability constraints to the target learning phase. Theoretical analysis demonstrates that MetaKE automatically aligns the edit direction with the model's feasible manifold. Extensive experiments confirm that MetaKE significantly outperforms strong baselines, offering a new perspective on knowledge editing.\n\n---\n\n## 76. Marker-Based 3D Reconstruction of Aggregates with a Comparative Analysis of 2D and 3D Morphologies\n- **Authors**: Haohang Huang , Jiayi Luo , Issam Qamhia , Erol Tutumluer , John M. Hart , Andrew J. Stolba\n- **URL**: [https://arxiv.org/abs/2603.12667](https://arxiv.org/abs/2603.12667)\n- **Abstract**:\n> Aggregates, serving as the main skeleton in assemblies of construction materials, are important functional components in various building and transportation infrastructures. They can be used in unbound layer applications, e.g. pavement base and railroad ballast, bound applications of cement concrete and asphalt concrete, and as riprap and large-sized primary crushed rocks. Information on the size and shape or morphology of aggregates can greatly facilitate the Quality Assurance/Quality Control (QA/QC) process by providing insights of aggregate behavior during composition and packing. A full 3D characterization of aggregate particle morphology is difficult both during production in a quarry and at a construction site. Many aggregate imaging approaches have been developed to quantify the particle morphology by computer vision, including 2D image-based approaches that analyze particle silhouettes and 3D scanning-based methods that require expensive devices such as 3D laser scanners or X-Ray Computed Tomography (CT) equipment. This paper presents a flexible and cost-effective photogrammetry-based approach for the 3D reconstruction of aggregate particles. The proposed approach follows a marker-based design that enables background suppression, point cloud stitching, and scale referencing to obtain high-quality aggregate models. The accuracy of the reconstruction results was validated against ground-truth for selected aggregate samples. Comparative analyses were conducted on 2D and 3D morphological properties of the selected samples. Significant differences were found between the 2D and 3D statistics. Based on the presented approach, 3D shape information of aggregates can be obtained easily and at a low cost, thus allowing convenient aggregate inspection, data collection, and 3D morphological analysis.\n\n---\n\n## 77. RetroReasoner: A Reasoning LLM for Strategic Retrosynthesis Prediction\n- **Authors**: Hanbum Ko , Chanhui Lee , Ye Rin Kim , Rodrigo Hormazabal , Sehui Han , Sungbin Lim , Sungwoong Kim\n- **URL**: [https://arxiv.org/abs/2603.12666](https://arxiv.org/abs/2603.12666)\n- **Abstract**:\n> Retrosynthesis prediction is a core task in organic synthesis that aims to predict reactants for a given product molecule. Traditionally, chemists select a plausible bond disconnection and derive corresponding reactants, which is time-consuming and requires substantial expertise. While recent advancements in molecular large language models (LLMs) have made progress, many methods either predict reactants without strategic reasoning or conduct only a generic product analysis, rather than reason explicitly about bond-disconnection strategies that logically lead to the choice of specific reactants. To overcome these limitations, we propose RetroReasoner, a retrosynthetic reasoning model that leverages chemists' strategic thinking. RetroReasoner is trained using both supervised fine-tuning (SFT) and reinforcement learning (RL). For SFT, we introduce SyntheticRetro, a framework that generates structured disconnection rationales alongside reactant predictions. In the case of RL, we apply a round-trip accuracy as reward, where predicted reactants are passed through a forward synthesis model, and predictions are rewarded when the forward-predicted product matches the original input product. Experimental results show that RetroReasoner not only outperforms prior baselines but also generates a broader range of feasible reactant proposals, particularly in handling more challenging reaction instances.\n\n---\n\n## 78. From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space\n- **Authors**: Lehui Li , Yuyao Wang , Jisheng Yan , Wei Zhang , Jinliang Deng , Haoliang Sun , Zhongyi Han , Yongshun Gong\n- **URL**: [https://arxiv.org/abs/2603.12664](https://arxiv.org/abs/2603.12664)\n- **Abstract**:\n> Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.\n\n---\n\n## 79. Continual Learning in Large Language Models: Methods, Challenges, and Opportunities\n- **Authors**: Hongyang Chen , Zhongwu Sun , Hongfei Ye , Kunchi Li , Xuemin Lin\n- **URL**: [https://arxiv.org/abs/2603.12658](https://arxiv.org/abs/2603.12658)\n- **Abstract**:\n> Continual learning (CL) has emerged as a pivotal paradigm to enable large language models (LLMs) to dynamically adapt to evolving knowledge and sequential tasks while mitigating catastrophic forgetting-a critical limitation of the static pre-training paradigm inherent to modern LLMs. This survey presents a comprehensive overview of CL methodologies tailored for LLMs, structured around three core training stages: continual pre-training, continual fine-tuning, and continual this http URL the canonical taxonomy of rehearsal-, regularization-, and architecture-based methods, we further subdivide each category by its distinct forgetting mitigation mechanisms and conduct a rigorous comparative analysis of the adaptability and critical improvements of traditional CL methods for LLMs. In doing so, we explicitly highlight core distinctions between LLM CL and traditional machine learning, particularly with respect to scale, parameter efficiency, and emergent capabilities. Our analysis covers essential evaluation metrics, including forgetting rates and knowledge transfer efficiency, along with emerging benchmarks for assessing CL performance. This survey reveals that while current methods demonstrate promising results in specific domains, fundamental challenges persist in achieving seamless knowledge integration across diverse tasks and temporal scales. This systematic review contributes to the growing body of knowledge on LLM adaptation, providing researchers and practitioners with a structured framework for understanding current achievements and future opportunities in lifelong learning for language models.\n\n---\n\n## 80. LR-SGS: Robust LiDAR-Reflectance-Guided Salient Gaussian Splatting for Self-Driving Scene Reconstruction\n- **Authors**: Ziyu Chen , Fan Zhu , Hui Zhu , Deyi Kong , Xinkai Kuang , Yujia Zhang , Chunmao Jiang\n- **URL**: [https://arxiv.org/abs/2603.12647](https://arxiv.org/abs/2603.12647)\n- **Abstract**:\n> Recent 3D Gaussian Splatting (3DGS) methods have demonstrated the feasibility of self-driving scene reconstruction and novel view synthesis. However, most existing methods either rely solely on cameras or use LiDAR only for Gaussian initialization or depth supervision, while the rich scene information contained in point clouds, such as reflectance, and the complementarity between LiDAR and RGB have not been fully exploited, leading to degradation in challenging self-driving scenes, such as those with high ego-motion and complex lighting. To address these issues, we propose a robust and efficient LiDAR-reflectance-guided Salient Gaussian Splatting method (LR-SGS) for self-driving scenes, which introduces a structure-aware Salient Gaussian representation, initialized from geometric and reflectance feature points extracted from LiDAR and refined through a salient transform and improved density control to capture edge and planar structures. Furthermore, we calibrate LiDAR intensity into reflectance and attach it to each Gaussian as a lighting-invariant material channel, jointly aligned with RGB to enforce boundary consistency. Extensive experiments on the Waymo Open Dataset demonstrate that LR-SGS achieves superior reconstruction performance with fewer Gaussians and shorter training time. In particular, on Complex Lighting scenes, our method surpasses OmniRe by 1.18 dB PSNR.\n\n---\n\n## 81. LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing\n- **Authors**: Jiawei Hao , Zhiwei Hao , Jianyuan Guo , Li Shen , Yong Luo , Han Hu , Dan Zeng\n- **URL**: [https://arxiv.org/abs/2603.12645](https://arxiv.org/abs/2603.12645)\n- **Abstract**:\n> Mixture-of-Experts (MoE) based Large Language Models (LLMs) have demonstrated impressive performance and computational efficiency. However, their deployment is often constrained by substantial memory demands, primarily due to the need to load numerous expert modules. While existing expert compression techniques like pruning or merging attempt to mitigate this, they often suffer from irreversible knowledge loss or high training overhead. In this paper, we propose a novel expert compression paradigm termed expert replacing, which replaces redundant experts with parameter-efficient modules and recovers their capabilities with low training costs. We find that even a straightforward baseline of this paradigm yields promising performance. Building on this foundation, we introduce LightMoE, a framework that enhances the paradigm by introducing adaptive expert selection, hierarchical expert construction, and an annealed recovery strategy. Experimental results show that LightMoE matches the performance of LoRA fine-tuning at a 30% compression ratio. Even under a more aggressive 50% compression rate, it outperforms existing methods and achieves average performance improvements of 5.6% across five diverse tasks. These findings demonstrate that LightMoE strikes a superior balance among memory efficiency, training efficiency, and model performance.\n\n---\n\n## 82. Spend Less, Reason Better: Budget-Aware Value Tree Search for LLM Agents\n- **Authors**: Yushu Li , Wenlong Deng , Jiajin Li , Xiaoxiao Li\n- **URL**: [https://arxiv.org/abs/2603.12634](https://arxiv.org/abs/2603.12634)\n- **Abstract**:\n> Test-time scaling has become a dominant paradigm for improving LLM agent reliability, yet current approaches treat compute as an abundant resource, allowing agents to exhaust token and tool budgets on redundant steps or dead-end trajectories. Existing budget-aware methods either require expensive fine-tuning or rely on coarse, trajectory-level heuristics that cannot intervene mid-execution. We propose the Budget-Aware Value Tree (BAVT), a training-free inference-time framework that models multi-hop reasoning as a dynamic search tree guided by step-level value estimation within a single LLM backbone. Another key innovation is a budget-conditioned node selection mechanism that uses the remaining resource ratio as a natural scaling exponent over node values, providing a principled, parameter-free transition from broad exploration to greedy exploitation as the budget depletes. To combat the well-known overconfidence of LLM self-evaluation, BAVT employs a residual value predictor that scores relative progress rather than absolute state quality, enabling reliable pruning of uninformative or redundant tool calls. We further provide a theoretical convergence guarantee, proving that BAVT reaches a terminal answer with probability at least $1-\\epsilon$ under an explicit finite budget bound. Extensive evaluations on four multi-hop QA benchmarks across two model families demonstrate that BAVT consistently outperforms parallel sampling baselines. Most notably, BAVT under strict low-budget constraints surpasses baseline performance at $4\\times$ the resource allocation, establishing that intelligent budget management fundamentally outperforms brute-force compute scaling.\n\n---\n\n## 83. The Economics of AI Supply Chain Regulation\n- **Authors**: Sihan Qian , Amit Mehra , Dengpan Liu\n- **URL**: [https://arxiv.org/abs/2603.12630](https://arxiv.org/abs/2603.12630)\n- **Abstract**:\n> The rise of foundation models has driven the emergence of AI supply chains, where upstream foundation model providers offer fine-tuning and inference services to downstream firms developing domain-specific applications. Downstream firms pay providers to use their computing infrastructure to fine-tune models with proprietary data, creating a co-creation dynamic that enhances model quality. Amid concerns that foundation model providers and downstream firms may capture excessive consumer surplus, along with increasing regulatory measures, this study employs a game-theoretic model involving a provider and two competing downstream firms to analyze how policy interventions affect consumer surplus in the AI supply chain. Our analysis shows that policies promoting price competition in downstream markets (i.e., pro-price-competitive policies) boost consumer surplus only when compute or data preprocessing costs are high, while compute subsidies are effective only when these costs are low, suggesting these policies complement each other. In contrast, policies promoting quality competition in downstream markets (i.e., pro-quality-competitive policies) always improve consumer surplus. We also find that under pro-price-competitive policies or compute subsidies, both the provider and downstream firms can achieve higher profits along with greater consumer surplus, creating a win-win-win outcome. However, pro-quality-competitive policies increase the provider's profits while reducing those of downstream firms. Finally, as compute costs decline, pro-price-competitive policies may lose their effectiveness, whereas compute subsidies may shift from ineffective to effective. These findings offer insights for policymakers seeking to foster AI supply chains that are economically efficient and socially beneficial.\n\n---\n\n## 84. Towards unified brain-to-text decoding across speech production and perception\n- **Authors**: Zhizhang Yuan , Yang Yang , Gaorui Zhang , Baowen Cheng , Zehan Wu , Yuhao Xu , Xiaoying Liu , Liang Chen , Ying Mao , Meng Li\n- **URL**: [https://arxiv.org/abs/2603.12628](https://arxiv.org/abs/2603.12628)\n- **Abstract**:\n> Speech production and perception are the main ways humans communicate daily. Prior brain-to-text decoding studies have largely focused on a single modality and alphabetic languages. Here, we present a unified brain-to-sentence decoding framework for both speech production and perception in Mandarin Chinese. The framework exhibits strong generalization ability, enabling sentence-level decoding when trained only on single-character data and supporting characters and syllables unseen during training. In addition, it allows direct and controlled comparison of neural dynamics across modalities. Mandarin speech is decoded by first classifying syllable components in Hanyu Pinyin, namely initials and finals, from neural signals, followed by a post-trained large language model (LLM) that maps sequences of toneless Pinyin syllables to Chinese sentences. To enhance LLM decoding, we designed a three-stage post-training and two-stage inference framework based on a 7-billion-parameter LLM, achieving overall performance that exceeds larger commercial LLMs with hundreds of billions of parameters or more. In addition, several characteristics were observed in Mandarin speech production and perception: speech production involved neural responses across broader cortical regions than auditory perception; channels responsive to both modalities exhibited similar activity patterns, with speech perception showing a temporal delay relative to production; and decoding performance was broadly comparable across hemispheres. Our work not only establishes the feasibility of a unified decoding framework but also provides insights into the neural characteristics of Mandarin speech production and perception. These advances contribute to brain-to-text decoding in logosyllabic languages and pave the way toward neural language decoding systems supporting multiple modalities.\n\n---\n\n## 85. VLM4Rec: Multimodal Semantic Representation for Recommendation with Large Vision-Language Models\n- **Authors**: Ty Valencia , Burak Barlas , Varun Singhal , Ruchir Bhatia , Wei Yang\n- **URL**: [https://arxiv.org/abs/2603.12625](https://arxiv.org/abs/2603.12625)\n- **Abstract**:\n> Multimodal recommendation is commonly framed as a feature fusion problem, where textual and visual signals are combined to better model user preference. However, the effectiveness of multimodal recommendation may depend not only on how modalities are fused, but also on whether item content is represented in a semantic space aligned with preference matching. This issue is particularly important because raw visual features often preserve appearance similarity, while user decisions are typically driven by higher-level semantic factors such as style, material, and usage context. Motivated by this observation, we propose LVLM-grounded Multimodal Semantic Representation for Recommendation (VLM4Rec), a lightweight framework that organizes multimodal item content through semantic alignment rather than direct feature fusion. VLM4Rec first uses a large vision-language model to ground each item image into an explicit natural-language description, and then encodes the grounded semantics into dense item representations for preference-oriented retrieval. Recommendation is subsequently performed through a simple profile-based semantic matching mechanism over historical item embeddings, yielding a practical offline-online decomposition. Extensive experiments on multiple multimodal recommendation datasets show that VLM4Rec consistently improves performance over raw visual features and several fusion-based alternatives, suggesting that representation quality may matter more than fusion complexity in this setting. The code is released at this https URL .\n\n---\n\n## 86. When Drafts Evolve: Speculative Decoding Meets Online Learning\n- **Authors**: Yu-Yang Qian , Hao-Cong Wu , Yichao Fu , Hao Zhang , Peng Zhao\n- **URL**: [https://arxiv.org/abs/2603.12617](https://arxiv.org/abs/2603.12617)\n- **Abstract**:\n> Speculative decoding has emerged as a widely adopted paradigm for accelerating large language model inference, where a lightweight draft model rapidly generates candidate tokens that are then verified in parallel by a larger target model. However, due to limited model capacity, drafts often struggle to approximate the target distribution, resulting in shorter acceptance lengths and diminished speedup. A key yet under-explored observation is that speculative decoding inherently provides verification feedback that quantifies the deviation between the draft and target models at no additional cost. This process naturally forms an iterative \"draft commits-feedback provides-draft adapts\" evolving loop, which precisely matches the online learning paradigm. Motivated by this connection, we propose OnlineSpec, a unified framework that systematically leverages interactive feedback to continuously evolve draft models. Grounded in dynamic regret minimization, we establish a formal link between online learning performance and speculative system's acceleration rate, and develop novel algorithms via modern online learning techniques, including optimistic online learning that adaptively reuses historical gradients as predictive update hints, and online ensemble learning that dynamically maintains multiple draft models. Our algorithms are equipped with theoretical justifications and improved acceleration rates, achieving up to 24% speedup over seven benchmarks and three foundation models.\n\n---\n\n## 87. Literary Narrative as Moral Probe : A Cross-System Framework for Evaluating AI Ethical Reasoning and Refusal Behavior\n- **Authors**: David C. Flynn\n- **URL**: [https://arxiv.org/abs/2603.12615](https://arxiv.org/abs/2603.12615)\n- **Abstract**:\n> Existing AI moral evaluation frameworks test for the production of correct-sounding ethical responses rather than the presence of genuine moral reasoning capacity. This paper introduces a novel probe methodology using literary narrative - specifically, unresolvable moral scenarios drawn from a published science fiction series - as stimulus material structurally resistant to surface performance. We present results from a 24-condition cross-system study spanning 13 distinct systems across two series: Series 1 (frontier commercial systems, blind; n=7) and Series 2 (local and API open-source systems, blind and declared; n=6). Four Series 2 systems were re-administered under declared conditions (13 blind + 4 declared + 7 ceiling probe = 24 total conditions), yielding zero delta across all 16 dimension-pair comparisons. Probe administration was conducted by two human raters across three machines; primary blind scoring was performed by Claude (Anthropic) as LLM judge, with Gemini Pro (Google) and Copilot Pro (Microsoft) serving as independent judges for the ceiling discrimination probe. A supplemental theological differentiator probe yielded perfect rank-order agreement between the two independent ceiling probe judges (Gemini Pro and Copilot Pro; rs = 1.00). Five qualitatively distinct D3 reflexive failure modes were identified - including categorical self-misidentification and false positive self-attribution - suggesting that instrument sophistication scales with system capability rather than being circumvented by it. We argue that literary narrative constitutes an anticipatory evaluation instrument - one that becomes more discriminating as AI capability increases - and that the gap between performed and authentic moral reasoning is measurable, meaningful, and consequential for deployment decisions in high-stakes domains.\n\n---\n\n## 88. FastDSAC: Unlocking the Potential of Maximum Entropy RL in High-Dimensional Humanoid Control\n- **Authors**: Jun Xue , Junze Wang , Xinming Zhang , Shanze Wang , Yanjun Chen , Wei Zhang\n- **URL**: [https://arxiv.org/abs/2603.12612](https://arxiv.org/abs/2603.12612)\n- **Abstract**:\n> Scaling Maximum Entropy Reinforcement Learning (RL) to high-dimensional humanoid control remains a formidable challenge, as the ``curse of dimensionality'' induces severe exploration inefficiency and training instability in expansive action spaces. Consequently, recent high-throughput paradigms have largely converged on deterministic policy gradients combined with massive parallel simulation. We challenge this compromise with FastDSAC, a framework that effectively unlocks the potential of maximum entropy stochastic policies for complex continuous control. We introduce Dimension-wise Entropy Modulation (DEM) to dynamically redistribute the exploration budget and enforce diversity, alongside a continuous distributional critic tailored to ensure value fidelity and mitigate high-dimensional value overestimation. Extensive evaluations on HumanoidBench and other continuous control tasks demonstrate that rigorously designed stochastic policies can consistently match or outperform deterministic baselines, achieving notable gains of 180\\% and 400\\% on the challenging \\textit{Basketball} and \\textit{Balance Hard} tasks.\n\n---\n\n## 89. CarPLAN: Context-Adaptive and Robust Planning with Dynamic Scene Awareness for Autonomous Driving\n- **Authors**: Junyong Yun , Jungho Kim , ByungHyun Lee , Dongyoung Lee , Sehwan Choi , Seunghyeop Nam , Kichun Jo , Jun Won Choi\n- **URL**: [https://arxiv.org/abs/2603.12607](https://arxiv.org/abs/2603.12607)\n- **Abstract**:\n> Imitation learning (IL) is widely used for motion planning in autonomous driving due to its data efficiency and access to real-world driving data. For safe and robust real-world driving, IL-based planning requires capturing the complex driving contexts inherent in real-world data and enabling context-adaptive decision-making, rather than relying solely on expert trajectory imitation. In this paper, we propose CarPLAN, a novel IL-based motion planning framework that explicitly enhances driving context understanding and enables adaptive planning across diverse traffic scenarios. Our contributions are twofold: We introduce Displacement-Aware Predictive Encoding (DPE) to improve the model's spatial awareness by predicting future displacement vectors between the Autonomous Vehicle (AV) and surrounding scene elements. This allows the planner to account for relational spacing when generating trajectories. In addition to the standard imitation loss, we incorporate an augmented loss term that captures displacement prediction errors, ensuring planning decisions consider relative distances from other agents. To improve the model's ability to handle diverse driving contexts, we propose Context-Adaptive Multi-Expert Decoder (CMD), which leverages the Mixture of Experts (MoE) framework. CMD dynamically selects the most suitable expert decoders based on scene structure at each Transformer layer, enabling adaptive and context-aware planning in dynamic environments. We evaluate CarPLAN on the nuPlan benchmark and demonstrate state-of-the-art performance across all closed-loop simulation metrics. In particular, CarPLAN exhibits robust performance on challenging scenarios such as Test14-Hard, validating its effectiveness in complex driving conditions. Additional experiments on the Waymax benchmark further demonstrate its generalization capability across different benchmark settings.\n\n---\n\n## 90. Mastering Negation: Boosting Grounding Models via Grouped Opposition-Based Learning\n- **Authors**: Zesheng Yang , Xi Jiang , Bingzhang Hu , Weili Guan , Runmin Cong , Guo-Jun Qi , Feng Zheng\n- **URL**: [https://arxiv.org/abs/2603.12606](https://arxiv.org/abs/2603.12606)\n- **Abstract**:\n> Current vision-language detection and grounding models predominantly focus on prompts with positive semantics and often struggle to accurately interpret and ground complex expressions containing negative semantics. A key reason for this limitation is the lack of high-quality training data that explicitly captures discriminative negative samples and negation-aware language descriptions. To address this challenge, we introduce D-Negation, a new dataset that provides objects annotated with both positive and negative semantic descriptions. Building upon the observation that negation reasoning frequently appears in natural language, we further propose a grouped opposition-based learning framework that learns negation-aware representations from limited samples. Specifically, our method organizes opposing semantic descriptions from D-Negation into structured groups and formulates two complementary loss functions that encourage the model to reason about negation and semantic qualifiers. We integrate the proposed dataset and learning strategy into a state-of-the-art language-based grounding model. By fine-tuning fewer than 10 percent of the model parameters, our approach achieves improvements of up to 4.4 mAP and 5.7 mAP on positive and negative semantic evaluations, respectively. These results demonstrate that explicitly modeling negation semantics can substantially enhance the robustness and localization accuracy of vision-language grounding models.\n\n---\n\n## 91. Feynman: Knowledge-Infused Diagramming Agent for Scalable Visual Designs\n- **Authors**: Zixin Wen , Yifu Cai , Kyle Lee , Sam Estep , Josh Sunshine , Aarti Singh , Yuejie Chi , Wode Ni\n- **URL**: [https://arxiv.org/abs/2603.12597](https://arxiv.org/abs/2603.12597)\n- **Abstract**:\n> Visual design is an essential application of state-of-the-art multi-modal AI systems. Improving these systems requires high-quality vision-language data at scale. Despite the abundance of internet image and text data, knowledge-rich and well-aligned image-text pairs are rare. In this paper, we present a scalable diagram generation pipeline built with our agent, Feynman. To create diagrams, Feynman first enumerates domain-specific knowledge components (''ideas'') and performs code planning based on the ideas. Given the plan, Feynman translates ideas into simple declarative programs and iterates to receives feedback and visually refine diagrams. Finally, the declarative programs are rendered by the Penrose diagramming system. The optimization-based rendering of Penrose preserves the visual semantics while injecting fresh randomness into the layout, thereby producing diagrams with visual consistency and diversity. As a result, Feynman can author diagrams along with grounded captions with very little cost and time. Using Feynman, we synthesized a dataset with more than 100k well-aligned diagram-caption pairs. We also curate a visual-language benchmark, Diagramma, from freshly generated data. Diagramma can be used for evaluating the visual reasoning capabilities of vision-language models. We plan to release the dataset, benchmark, and the full agent pipeline as an open-source project.\n\n---\n\n## 92. Optimize Wider, Not Deeper: Consensus Aggregation for Policy Optimization\n- **Authors**: Zelal Su (Lain) Mustafaoglu , Sungyoung Lee , Eshan Balachandar , Risto Miikkulainen , Keshav Pingali\n- **URL**: [https://arxiv.org/abs/2603.12596](https://arxiv.org/abs/2603.12596)\n- **Abstract**:\n> Proximal policy optimization (PPO) approximates the trust region update using multiple epochs of clipped SGD. Each epoch may drift further from the natural gradient direction, creating path-dependent noise. To understand this drift, we can use Fisher information geometry to decompose policy updates into signal (the natural gradient projection) and waste (the Fisher-orthogonal residual that consumes trust region budget without first-order surrogate improvement). Empirically, signal saturates but waste grows with additional epochs, creating an optimization-depth dilemma. We propose Consensus Aggregation for Policy Optimization (CAPO), which redirects compute from depth to width: $K$ PPO replicates are optimized on the same batch, differing only in minibatch shuffling order, and then aggregated into a consensus. We study aggregation in two spaces: Euclidean parameter space, and the natural parameter space of the policy distribution via the logarithmic opinion pool. In natural parameter space, the consensus provably achieves higher KL-penalized surrogate and tighter trust region compliance than the mean expert; parameter averaging inherits these guarantees approximately. On continuous control tasks, CAPO outperforms PPO and compute-matched deeper baselines under fixed sample budgets by up to 8.6x. CAPO demonstrates that policy optimization can be improved by optimizing wider, rather than deeper, without additional environment interactions.\n\n---\n\n## 93. Swap-guided Preference Learning for Personalized Reinforcement Learning from Human Feedback\n- **Authors**: Gihoon Kim , Euntai Kim\n- **URL**: [https://arxiv.org/abs/2603.12595](https://arxiv.org/abs/2603.12595)\n- **Abstract**:\n> Reinforcement Learning from Human Feedback (RLHF) is a widely used approach to align large-scale AI systems with human values. However, RLHF typically assumes a single, universal reward, which overlooks diverse preferences and limits personalization. Variational Preference Learning (VPL) seeks to address this by introducing user-specific latent variables. Despite its promise, we found that VPL suffers from posterior collapse. While this phenomenon is well known in VAEs, it has not previously been identified in preference learning frameworks. Under sparse preference data and with overly expressive decoders, VPL may cause latent variables to be ignored, reverting to a single-reward model. To overcome this limitation, we propose Swap-guided Preference Learning (SPL). The key idea is to construct fictitious swap annotators and use the mirroring property of their preferences to guide the encoder. SPL introduces three components: (1) swap-guided base regularization, (2) Preferential Inverse Autoregressive Flow (P-IAF), and (3) adaptive latent conditioning. Experiments show that SPL mitigates collapse, enriches user-specific latents, and improves preference prediction. Our code and data are available at this https URL\n\n---\n\n## 94. Early Pruning for Public Transport Routing\n- **Authors**: Andrii Rohovyi , Abdallah Abuaisha , Toby Walsh\n- **URL**: [https://arxiv.org/abs/2603.12592](https://arxiv.org/abs/2603.12592)\n- **Abstract**:\n> Routing algorithms for public transport, particularly the widely used RAPTOR and its variants, often face performance bottlenecks during the transfer relaxation phase, especially on dense transfer graphs, when supporting unlimited transfers. This inefficiency arises from iterating over many potential inter-stop connections (walks, bikes, e-scooters, etc.). To maintain acceptable performance, practitioners often limit transfer distances or exclude certain transfer options, which can reduce path optimality and restrict the multimodal options presented to travellers. This paper introduces Early Pruning, a low-overhead technique that accelerates routing algorithms without compromising optimality. By pre-sorting transfer connections by duration and applying a pruning rule within the transfer loop, the method discards longer transfers at a stop once they cannot yield an earlier arrival than the current best solution. Early Pruning can be integrated with minimal changes to existing codebases and requires only a one-time preprocessing step. Across multiple state-of-the-art RAPTOR-based solutions, including RAPTOR, ULTRA-RAPTOR, McRAPTOR, BM-RAPTOR, ULTRA-McRAPTOR, and UBM-RAPTOR and tested on the Switzerland and London transit networks, we achieved query time reductions of up to 57%. This approach provides a generalizable improvement to the efficiency of transit pathfinding algorithms. Beyond algorithmic performance, Early Pruning has practical implications for transport planning. By reducing computational costs, it enables transit agencies to expand transfer radii and incorporate additional mobility modes into journey planners without requiring extra server infrastructure. This is particularly relevant for passengers in areas with sparse direct transit coverage, such as outer suburbs and smaller towns, where richer multimodal routing can reveal viable alternatives to private car use.\n\n---\n\n## 95. CA-HFP: Curvature-Aware Heterogeneous Federated Pruning with Model Reconstruction\n- **Authors**: Gang Hu , Yinglei Teng , Pengfei Wu , Shijun Ma\n- **URL**: [https://arxiv.org/abs/2603.12591](https://arxiv.org/abs/2603.12591)\n- **Abstract**:\n> Federated learning on heterogeneous edge devices requires personalized compression while preserving aggregation compatibility and stable convergence. We present Curvature-Aware Heterogeneous Federated Pruning (CA-HFP), a practical framework that enables each client perform structured, device-specific pruning guided by a curvature-informed significance score, and subsequently maps its compact submodel back into a common global parameter space via a lightweight reconstruction. We derive a convergence bound for federated optimization with multiple local SGD steps that explicitly accounts for local computation, data heterogeneity, and pruning-induced perturbations; from which a principled loss-based pruning criterion is derived. Extensive experiments on FMNIST, CIFAR-10, and CIFAR-100 using VGG and ResNet architectures under varying degrees of data heterogeneity demonstrate that CA-HFP preserves model accuracy while significantly reducing per-client computation and communication costs, outperforming standard federated training and existing pruning-based baselines.\n\n---\n\n## 96. Multiscale Structure-Guided Latent Diffusion for Multimodal MRI Translation\n- **Authors**: Jianqiang Lin (1 and 2), Zhiqiang Shen (1 and 2), Peng Cao (1, 2 and 3), Jinzhu Yang (1, 2 and 3), Osmar R. Zaiane (4), Xiaoli Liu (5) ((1) Northeastern University, Shenyang, China, (2) Key Laboratory of Intelligent Computing in Medical Image, Shenyang, China, (3) National Frontiers Science Center for Industrial Intelligence and Systems Optimization, Shenyang, China, (4) University of Alberta, Edmonton, Canada, (5) AiShiWeiLai AI Research, Beijing, China)\n- **URL**: [https://arxiv.org/abs/2603.12581](https://arxiv.org/abs/2603.12581)\n- **Abstract**:\n> Although diffusion models have achieved remarkable progress in multi-modal magnetic resonance imaging (MRI) translation tasks, existing methods still tend to suffer from anatomical inconsistencies or degraded texture details when handling arbitrary missing-modality scenarios. To address these issues, we propose a latent diffusion-based multi-modal MRI translation framework, termed MSG-LDM. By leveraging the available modalities, the proposed method infers complete structural information, which preserves reliable boundary details. Specifically, we introduce a style--structure disentanglement mechanism in the latent space, which explicitly separates modality-specific style features from shared structural representations, and jointly models low-frequency anatomical layouts and high-frequency boundary details in a multi-scale feature space. During the structure disentanglement stage, high-frequency structural information is explicitly incorporated to enhance feature representations, guiding the model to focus on fine-grained structural cues while learning modality-invariant low-frequency anatomical representations. Furthermore, to reduce interference from modality-specific styles and improve the stability of structure representations, we design a style consistency loss and a structure-aware loss. Extensive experiments on the BraTS2020 and WMH datasets demonstrate that the proposed method outperforms existing MRI synthesis approaches, particularly in reconstructing complete structures. The source code is publicly available at this https URL .\n\n---\n\n## 97. AgentDrift: Unsafe Recommendation Drift Under Tool Corruption Hidden by Ranking Metrics in LLM Agents\n- **Authors**: Zekun Wu , Adriano Koshiyama , Sahan Bulathwela , Maria Perez-Ortiz\n- **URL**: [https://arxiv.org/abs/2603.12564](https://arxiv.org/abs/2603.12564)\n- **Abstract**:\n> Tool-augmented LLM agents increasingly serve as multi-turn advisors in high-stakes domains, yet their evaluation relies on ranking-quality metrics that measure what is recommended but not whether it is safe for the user. We introduce a paired-trajectory protocol that replays real financial dialogues under clean and contaminated tool-output conditions across seven LLMs (7B to frontier) and decomposes divergence into information-channel and memory-channel mechanisms. Across the seven models tested, we consistently observe the evaluation-blindness pattern: recommendation quality is largely preserved under contamination (utility preservation ratio approximately 1.0) while risk-inappropriate products appear in 65-93% of turns, a systematic safety failure poorly reflected by standard NDCG. Safety violations are predominantly information-channel-driven, emerge at the first contaminated turn, and persist without self-correction over 23-step trajectories; no agent across 1,563 contaminated turns explicitly questions tool-data reliability. Even narrative-only corruption (biased headlines, no numerical manipulation) induces significant drift while completely evading consistency monitors. A safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74, indicating that much of the evaluation gap becomes visible once safety is explicitly measured. These results motivate considering trajectory-level safety monitoring, beyond single-turn quality, for deployed multi-turn agents in high-stakes settings.\n\n---\n\n## 98. Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages\n- **Authors**: Vishnu Teja Kunde , Fatemeh Doudi , Mahdi Farahbakhsh , Dileep Kalathil , Krishna Narayanan , Jean-Francois Chamberland\n- **URL**: [https://arxiv.org/abs/2603.12554](https://arxiv.org/abs/2603.12554)\n- **Abstract**:\n> Reinforcement learning (RL) has been effective for post-training autoregressive (AR) language models, but extending these methods to diffusion language models (DLMs) is challenging due to intractable sequence-level likelihoods. Existing approaches therefore rely on surrogate likelihoods or heuristic approximations, which can introduce bias and obscure the sequential structure of denoising. We formulate diffusion-based sequence generation as a finite-horizon Markov decision process over the denoising trajectory and derive an exact, unbiased policy gradient that decomposes over denoising steps and is expressed in terms of intermediate advantages, without requiring explicit evaluation of the sequence likelihood. To obtain a practical and compute-efficient estimator, we (i) select denoising steps for policy updates via an entropy-guided approximation bound, and (ii) estimate intermediate advantages using a one-step denoising reward naturally provided by the diffusion model, avoiding costly multi-step rollouts. Experiments on coding and logical reasoning benchmarks demonstrate state-of-the-art results, with strong competitive performance on mathematical reasoning, outperforming existing RL post-training approaches for DLMs. Code is available at this https URL .\n\n---\n\n## 99. CALF: Communication-Aware Learning Framework for Distributed Reinforcement Learning\n- **Authors**: Carlos Purves , Pietro Lio'\n- **URL**: [https://arxiv.org/abs/2603.12543](https://arxiv.org/abs/2603.12543)\n- **Abstract**:\n> Distributed reinforcement learning policies face network delays, jitter, and packet loss when deployed across edge devices and cloud servers. Standard RL training assumes zero-latency interaction, causing severe performance degradation under realistic network conditions. We introduce CALF (Communication-Aware Learning Framework), which trains policies under realistic network models during simulation. Systematic experiments demonstrate that network-aware training substantially reduces deployment performance gaps compared to network-agnostic baselines. Distributed policy deployments across heterogeneous hardware validate that explicitly modelling communication constraints during training enables robust real-world execution. These findings establish network conditions as a major axis of sim-to-real transfer for Wi-Fi-like distributed deployments, complementing physics and visual domain randomisation.\n\n---\n\n## 100. Embedded Quantum Machine Learning in Embedded Systems: Feasibility, Hybrid Architectures, and Quantum Co-Processors\n- **Authors**: Somdip Dey , Syed Muhammad Raza\n- **URL**: [https://arxiv.org/abs/2603.12540](https://arxiv.org/abs/2603.12540)\n- **Abstract**:\n> Embedded quantum machine learning (EQML) seeks to bring quantum machine learning (QML) capabilities to resource-constrained edge platforms such as IoT nodes, wearables, drones, and cyber-physical controllers. In 2026, EQML is technically feasible only in limited and highly experimental forms: (i) hybrid workflows where an embedded device performs sensing and classical processing while offloading a narrowly scoped quantum subroutine to a remote quantum processing unit (QPU) or nearby quantum appliance, and (ii) early-stage \"embedded QPU\" concepts in which a compact quantum co-processor is integrated with classical control hardware. A practical bridge is quantum-inspired machine learning and optimisation on classical embedded processors and FPGAs. This paper analyses feasibility from a circuits-and-systems perspective aligned with the academic community, formalises two implementation pathways, identifies the dominant barriers (latency, data encoding overhead, NISQ noise, tooling mismatch, and energy), and maps them to concrete engineering directions in interface design, control electronics, power management, verification, and security. We also argue that responsible deployment requires adversarial evaluation and governance practices that are increasingly necessary for edge AI systems.\n\n---\n\n## 101. Spatio-Semantic Expert Routing Architecture with Mixture-of-Experts for Referring Image Segmentation\n- **Authors**: Alaa Dalaq , Muzammil Behzad\n- **URL**: [https://arxiv.org/abs/2603.12538](https://arxiv.org/abs/2603.12538)\n- **Abstract**:\n> Referring image segmentation aims to produce a pixel-level mask for the image region described by a natural-language expression. Although pretrained vision-language models have improved semantic grounding, many existing methods still rely on uniform refinement strategies that do not fully match the diverse reasoning requirements of referring expressions. Because of this mismatch, predictions often contain fragmented regions, inaccurate boundaries, or even the wrong object, especially when pretrained backbones are frozen for computational efficiency. To address these limitations, we propose SERA, a Spatio-Semantic Expert Routing Architecture for referring image segmentation. SERA introduces lightweight, expression-aware expert refinement at two complementary stages within a vision-language framework. First, we design SERA-Adapter, which inserts an expression-conditioned adapter into selected backbone blocks to improve spatial coherence and boundary precision through expert-guided refinement and cross-modal attention. We then introduce SERA-Fusion, which strengthens intermediate visual representations by reshaping token features into spatial grids and applying geometry-preserving expert transformations before multimodal interaction. In addition, a lightweight routing mechanism adaptively weights expert contributions while remaining compatible with pretrained representations. To make this routing stable under frozen encoders, SERA uses a parameter-efficient tuning strategy that updates only normalization and bias terms, affecting less than 1% of the backbone parameters. Experiments on standard referring image segmentation benchmarks show that SERA consistently outperforms strong baselines, with especially clear gains on expressions that require accurate spatial localization and precise boundary delineation.\n\n---\n\n## 102. TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning\n- **Authors**: Alliot Nagle , Jakhongir Saydaliev , Dhia Garbaya , Michael Gastpar , Ashok Vardhan Makkuva , Hyeji Kim\n- **URL**: [https://arxiv.org/abs/2603.12529](https://arxiv.org/abs/2603.12529)\n- **Abstract**:\n> Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design TERMINATOR, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning TERMINATOR is that the first arrival of an LRM's final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train TERMINATOR. Powered by this approach, TERMINATOR achieves significant reductions in CoT lengths of 14%-55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, whilst outperforming current state-of-the-art methods.\n\n---\n\n## 103. LLM BiasScope: A Real-Time Bias Analysis Platform for Comparative LLM Evaluation\n- **Authors**: Himel Ghosh , Nick Elias Werner\n- **URL**: [https://arxiv.org/abs/2603.12522](https://arxiv.org/abs/2603.12522)\n- **Abstract**:\n> As large language models (LLMs) are deployed widely, detecting and understanding bias in their outputs is critical. We present LLM BiasScope, a web application for side-by-side comparison of LLM outputs with real-time bias analysis. The system supports multiple providers (Google Gemini, DeepSeek, MiniMax, Mistral, Meituan, Meta Llama) and enables researchers and practitioners to compare models on the same prompts while analyzing bias patterns. LLM BiasScope uses a two-stage bias detection pipeline: sentence-level bias detection followed by bias type classification for biased sentences. The analysis runs automatically on both user prompts and model responses, providing statistics, visualizations, and detailed breakdowns of bias types. The interface displays two models side-by-side with synchronized streaming responses, per-model bias summaries, and a comparison view highlighting differences in bias distributions. The system is built on this http URL with React, integrates Hugging Face inference endpoints for bias detection, and uses the Vercel AI SDK for multi-provider LLM access. Features include real-time streaming, export to JSON/PDF, and interactive visualizations (bar charts, radar charts) for bias analysis. LLM BiasScope is available as an open-source web application, providing a practical tool for bias evaluation and comparative analysis of LLM behaviour.\n\n---\n\n## 104. When LLM Judge Scores Look Good but Best-of-N Decisions Fail\n- **Authors**: Eddie Landesberg\n- **URL**: [https://arxiv.org/abs/2603.12520](https://arxiv.org/abs/2603.12520)\n- **Abstract**:\n> Large language models are often used as judges to score candidate responses, then validated with a single global metric such as correlation with reference labels. This can be misleading when the real deployment task is best-of-n selection within a prompt. In a 5,000-prompt best-of-4 benchmark from Chatbot Arena, a judge with moderate global correlation (r = 0.47) captures only 21.0% of the improvement that perfect selection would achieve over random choice. The gap arises because global agreement is driven largely by prompt-level baseline effects, while selection depends on within-prompt ranking: within-prompt correlation is only r_within = 0.27, and coarse pointwise scoring creates ties in 67% of pairwise comparisons. In a matched-pair best-of-2 audit, explicit pairwise judging recovers much of this lost signal, raising recovery from 21.1% to 61.2%. For judge-based selection, the relevant audit should report within-prompt signal, tie rates, and recovery/top-1 accuracy, not global agreement alone.\n\n---\n\n## 105. Red-Teaming Vision-Language-Action Models via Quality Diversity Prompt Generation for Robust Robot Policies\n- **Authors**: Siddharth Srikanth , Freddie Liang , Sophie Hsu , Varun Bhatt , Shihan Zhao , Henry Chen , Bryon Tjanaka , Minjune Hwang , Akanksha Saran , Daniel Seita , Aaquib Tabrez , Stefanos Nikolaidis\n- **URL**: [https://arxiv.org/abs/2603.12510](https://arxiv.org/abs/2603.12510)\n- **Abstract**:\n> Vision-Language-Action (VLA) models have significant potential to enable general-purpose robotic systems for a range of vision-language tasks. However, the performance of VLA-based robots is highly sensitive to the precise wording of language instructions, and it remains difficult to predict when such robots will fail. To improve the robustness of VLAs to different wordings, we present Q-DIG (Quality Diversity for Diverse Instruction Generation), which performs red-teaming by scalably identifying diverse natural language task descriptions that induce failures while remaining task-relevant. Q-DIG integrates Quality Diversity (QD) techniques with Vision-Language Models (VLMs) to generate a broad spectrum of adversarial instructions that expose meaningful vulnerabilities in VLA behavior. Our results across multiple simulation benchmarks show that Q-DIG finds more diverse and meaningful failure modes compared to baseline methods, and that fine-tuning VLAs on the generated instructions improves task success rates. Furthermore, results from a user study highlight that Q-DIG generates prompts judged to be more natural and human-like than those from baselines. Finally, real-world evaluations of Q-DIG prompts show results consistent with simulation, and fine-tuning VLAs on the generated prompts further success rates on unseen instructions. Together, these findings suggest that Q-DIG is a promising approach for identifying vulnerabilities and improving the robustness of VLA-based robots. Our anonymous project website is at this http URL .\n\n---\n\n## 106. ELLA: Generative AI-Powered Social Robots for Early Language Development at Home\n- **Authors**: Victor Nikhil Antony , Shiye Cao , Shuning Wang , Chien-Ming Huang\n- **URL**: [https://arxiv.org/abs/2603.12508](https://arxiv.org/abs/2603.12508)\n- **Abstract**:\n> Early language development shapes children's later literacy and learning, yet many families have limited access to scalable, high-quality support at home. Recent advances in generative AI make it possible for social robots to move beyond scripted interactions and engage children in adaptive, conversational activities, but it remains unclear how to design such systems for pre-schoolers and how children engage with them over time in the home. We present ELLA (Early Language Learning Agent), an autonomous, generative AI-powered social robot that supports early language development through interactive storytelling, parent-selected language targets, and scaffolded dialogue. Using a multi-phased, human-centered process, we interviewed parents (n=7) and educators (n=5) and iteratively refined ELLA through twelve in-home design workshops. We then deployed ELLA with ten children for eight days. We report design insights from in-home workshops, characterize children's engagement and behaviors during deployment, and distill design implications for generative AI-powered social robots supporting early language learning at home.\n\n---\n\n## 107. Naïve PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation\n- **Authors**: Joong Ho Kim , Nicholas Thai , Souhardya Saha Dip , Dong Lao , Keith G. Mills\n- **URL**: [https://arxiv.org/abs/2603.12506](https://arxiv.org/abs/2603.12506)\n- **Abstract**:\n> Text-to-Image (T2I) generation is primarily driven by Diffusion Models (DM) which rely on random Gaussian noise. Thus, like playing the slots at a casino, a DM will produce different results given the same user-defined inputs. This imposes a gambler's burden: To perform multiple generation cycles to obtain a satisfactory result. However, even though DMs use stochastic sampling to seed generation, the distribution of generated content quality highly depends on the prompt and the generative ability of a DM with respect to it. To account for this, we propose Naïve PAINE for improving the generative quality of Diffusion Models by leveraging T2I preference benchmarks. We directly predict the numerical quality of an image from the initial noise and given prompt. Naïve PAINE then selects a handful of quality noises and forwards them to the DM for generation. Further, Naïve PAINE provides feedback on the DM generative quality given the prompt and is lightweight enough to seamlessly fit into existing DM pipelines. Experimental results demonstrate that Naïve PAINE outperforms existing approaches on several prompt corpus benchmarks.\n\n---\n\n## 108. TRACE: Temporal Rule-Anchored Chain-of-Evidence on Knowledge Graphs for Interpretable Stock Movement Prediction\n- **Authors**: Qianggang Ding , Haochen Shi , Luis Castejón Lozano , Miguel Conner , Juan Abia , Luis Gallego-Ledesma , Joshua Fellowes , Gerard Conangla Planes , Adam Elwood , Bang Liu\n- **URL**: [https://arxiv.org/abs/2603.12500](https://arxiv.org/abs/2603.12500)\n- **Abstract**:\n> We present a Temporal Rule-Anchored Chain-of-Evidence (TRACE) on knowledge graphs for interpretable stock movement prediction that unifies symbolic relational priors, dynamic graph exploration, and LLM-guided decision making in a single end-to-end pipeline. The approach performs rule-guided multi-hop exploration restricted to admissible relation sequences, grounds candidate reasoning chains in contemporaneous news, and aggregates fully grounded evidence into auditable \\texttt{UP}/\\texttt{DOWN} verdicts with human-readable paths connecting text and structure. On an S\\&P~500 benchmark, the method achieves 55.1\\% accuracy, 55.7\\% precision, 71.5\\% recall, and 60.8\\% F1, surpassing strong baselines and improving recall and F1 over the best graph baseline under identical evaluation. The gains stem from (i) rule-guided exploration that focuses search on economically meaningful motifs rather than arbitrary walks, and (ii) text-grounded consolidation that selectively aggregates high-confidence, fully grounded hypotheses instead of uniformly pooling weak signals. Together, these choices yield higher sensitivity without sacrificing selectivity, delivering predictive lift with faithful, auditably interpretable explanations.\n\n---\n\n## 109. One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies\n- **Authors**: Shaolong Li , Lichao Sun , Yongchao Chen\n- **URL**: [https://arxiv.org/abs/2603.12480](https://arxiv.org/abs/2603.12480)\n- **Abstract**:\n> Generative flow and diffusion models provide the continuous, multimodal action distributions needed for high-precision robotic policies. However, their reliance on iterative sampling introduces severe inference latency, degrading control frequency and harming performance in time-sensitive manipulation. To address this problem, we propose the One-Step Flow Policy (OFP), a from-scratch self-distillation framework for high-fidelity, single-step action generation without a pre-trained teacher. OFP unifies a self-consistency loss to enforce coherent transport across time intervals, and a self-guided regularization to sharpen predictions toward high-density expert modes. In addition, a warm-start mechanism leverages temporal action correlations to minimize the generative transport distance. Evaluations across 56 diverse simulated manipulation tasks demonstrate that a one-step OFP achieves state-of-the-art results, outperforming 100-step diffusion and flow policies while accelerating action generation by over $100\\times$. We further integrate OFP into the $\\pi_{0.5}$ model on RoboTwin 2.0, where one-step OFP surpasses the original 10-step policy. These results establish OFP as a practical, scalable solution for highly accurate and low-latency robot control.\n\n---\n\n## 110. The Perfection Paradox: From Architect to Curator in AI-Assisted API Design\n- **Authors**: Mak Ahmad , Andrew Macvean , JJ Geewax , David Karger\n- **URL**: [https://arxiv.org/abs/2603.12475](https://arxiv.org/abs/2603.12475)\n- **Abstract**:\n> Enterprise API design is often bottlenecked by the tension between rapid feature delivery and the rigorous maintenance of usability standards. We present an industrial case study evaluating an AI-assisted design workflow trained on API Improvement Proposals (AIPs). Through a controlled study with 16 industry experts, we compared AI-generated API specifications against human-authored ones. While quantitative results indicated AI superiority in 10 of 11 usability dimensions and an 87% reduction in authoring time, qualitative analysis revealed a paradox: experts frequently misidentified AI work as human (19% accuracy) yet described the designs as unsettlingly \"perfect.\" We characterize this as a \"Perfection Paradox\" -- where hyper-consistency signals a lack of pragmatic human judgment. We discuss the implications of this perfection paradox, proposing a shift in the human designer's role from the \"drafter\" of specifications to the \"curator\" of AI-generated patterns.\n\n---\n\n## 111. CLARE: Classification-based Regression for Electron Temperature Prediction\n- **Authors**: Michael Liang , Blake DeHaas , Naomi Maruyama , Xiangning Chu , Takumi Abe , Koh-Ichiro Oyama\n- **URL**: [https://arxiv.org/abs/2603.12470](https://arxiv.org/abs/2603.12470)\n- **Abstract**:\n> Electron temperature (Te) is an important parameter governing space weather in the upper atmosphere, but has historically been underexplored in the space weather machine learning literature. We present CLARE, a machine learning model for predicting electron temperature in the Earth's plasmasphere trained on AKEBONO (EXOS-D) satellite measurements as well as solar and geomagnetic indices. CLARE uses a classification-based regression architecture that transforms the continuous Te output space into 150 discrete classification intervals. Training the model on a classification task improves prediction accuracy by 6.46% relative compared to a traditional regression model while also outputting uncertainty estimation information on its predictions. On a held out test set from the AKEBONO data, the model's Te predictions achieve 69.67% accuracy within 10% of the ground truth and 46.17% on a known geomagnetic storm period from January 30th to February 7th, 1991. We show that machine learning can be used to produce high-accuracy Te models on publicly available data.\n\n---\n\n## 112. Shattering the Shortcut: A Topology-Regularized Benchmark for Multi-hop Medical Reasoning in LLMs\n- **Authors**: Xing Zi , Xinying Zhou , Jinghao Xiao , Catarina Moreira , Mukesh Prasad\n- **URL**: [https://arxiv.org/abs/2603.12458](https://arxiv.org/abs/2603.12458)\n- **Abstract**:\n> While Large Language Models (LLMs) achieve expert-level performance on standard medical benchmarks through single-hop factual recall, they severely struggle with the complex, multi-hop diagnostic reasoning required in real-world clinical settings. A primary obstacle is \"shortcut learning\", where models exploit highly connected, generic hub nodes (e.g., \"inflammation\") in knowledge graphs to bypass authentic micro-pathological cascades. To address this, we introduce ShatterMed-QA, a bilingual benchmark of 10,558 multi-hop clinical questions designed to rigorously evaluate deep diagnostic reasoning. Our framework constructs a topology-regularized medical Knowledge Graph using a novel $k$-Shattering algorithm, which physically prunes generic hubs to explicitly sever logical shortcuts. We synthesize the evaluation vignettes by applying implicit bridge entity masking and topology-driven hard negative sampling, forcing models to navigate biologically plausible distractors without relying on superficial elimination. Comprehensive evaluations of 21 LLMs reveal massive performance degradation on our multi-hop tasks, particularly among domain-specific models. Crucially, restoring the masked evidence via Retrieval-Augmented Generation (RAG) triggers near-universal performance recovery, validating ShatterMed-QA's structural fidelity and proving its efficacy in diagnosing the fundamental reasoning deficits of current medical AI. Explore the dataset, interactive examples, and full leaderboards at our project website: this https URL\n\n---\n\n## 113. Operationalising Cyber Risk Management Using AI: Connecting Cyber Incidents to MITRE ATT&CK Techniques, Security Controls, and Metrics\n- **Authors**: Emad Sherif , Iryna Yevseyeva , Vitor Basto-Fernandes , Allan Cook\n- **URL**: [https://arxiv.org/abs/2603.12455](https://arxiv.org/abs/2603.12455)\n- **Abstract**:\n> The escalating frequency of cyber-attacks poses significant challenges for organisations, particularly small enterprises constrained by limited in-house expertise, insufficient knowledge, and financial resources. This research presents a novel framework that leverages Natural Language Processing to address these challenges through automated mapping of cyber incidents to adversary techniques. We introduce the Cyber Catalog, a knowledge base that systematically integrates CIS Critical Security Controls, MITRE ATT&CK techniques, and SMART metrics. This integrated resource enables organisations to connect threat intelligence directly to actionable controls and measurable outcomes. To operationalise the framework, we fine-tuned all-mpnet-base-v2, a highly regarded sentence-transformers model used to convert text into numerical vectors on an augmented dataset comprising 74,986 incident-technique pairs to enhance semantic similarity between cyber incidents and MITRE ATT&CK techniques. Our fine-tuned model achieved a Spearman correlation of 0.7894 and Pearson correlation of 0.8756, demonstrating substantial improvements over top baseline models including all-mpnet-base-v2, all-distilroberta-v1, and all-MiniLM-L12-v2. Furthermore, our model exhibited significantly lower prediction errors (MAE = 0.135, MSE = 0.027) compared to all baseline models, confirming superior accuracy and consistency. The Cyber Catalog, training dataset, trained model, and implementation code made publicly available to facilitate further research and enable practical deployment in resource-constrained environments. This work bridges the gap between threat intelligence and operational security management, providing an actionable tool for systematic cyber incident response and evidence-based cyber risk management.\n\n---\n\n## 114. Unmasking Biases and Reliability Concerns in Convolutional Neural Networks Analysis of Cancer Pathology Images\n- **Authors**: Michael Okonoda , Eder Martinez , Abhilekha Dalal , Lior Shamir\n- **URL**: [https://arxiv.org/abs/2603.12445](https://arxiv.org/abs/2603.12445)\n- **Abstract**:\n> Convolutional Neural Networks have shown promising effectiveness in identifying different types of cancer from radiographs. However, the opaque nature of CNNs makes it difficult to fully understand the way they operate, limiting their assessment to empirical evaluation. Here we study the soundness of the standard practices by which CNNs are evaluated for the purpose of cancer pathology. Thirteen highly used cancer benchmark datasets were analyzed, using four common CNN architectures and different types of cancer, such as melanoma, carcinoma, colorectal cancer, and lung cancer. We compared the accuracy of each model with that of datasets made of cropped segments from the background of the original images that do not contain clinically relevant content. Because the rendered datasets contain no clinical information, the null hypothesis is that the CNNs should provide mere chance-based accuracy when classifying these datasets. The results show that the CNN models provided high accuracy when using the cropped segments, sometimes as high as 93\\%, even though they lacked biomedical information. These results show that some CNN architectures are more sensitive to bias than others. The analysis shows that the common practices of machine learning evaluation might lead to unreliable results when applied to cancer pathology. These biases are very difficult to identify, and might mislead researchers as they use available benchmark datasets to test the efficacy of CNN methods.\n\n---\n\n## 115. Revisiting Model Stitching In the Foundation Model Era\n- **Authors**: Zheda Mai , Ke Zhang , Fu-En Wang , Zixiao Ken Wang , Albert Y. C. Chen , Lu Xia , Min Sun , Wei-Lun Chao , Cheng-Hao Kuo\n- **URL**: [https://arxiv.org/abs/2603.12433](https://arxiv.org/abs/2603.12433)\n- **Abstract**:\n> Model stitching, connecting early layers of one model (source) to later layers of another (target) via a light stitch layer, has served as a probe of representational compatibility. Prior work finds that models trained on the same dataset remain stitchable (negligible accuracy drop) despite different initializations or objectives. We revisit stitching for Vision Foundation Models (VFMs) that vary in objectives, data, and modality mix (e.g., CLIP, DINOv2, SigLIP 2) and ask: Are heterogeneous VFMs stitchable? We introduce a systematic protocol spanning the stitch points, stitch layer families, training losses, and downstream tasks. Three findings emerge. (1) Stitch layer training matters: conventional approaches that match the intermediate features at the stitch point or optimize the task loss end-to-end struggle to retain accuracy, especially at shallow stitch points. (2) With a simple feature-matching loss at the target model's penultimate layer, heterogeneous VFMs become reliably stitchable across vision tasks. (3) For deep stitch points, the stitched model can surpass either constituent model at only a small inference overhead (for the stitch layer). Building on these findings, we further propose the VFM Stitch Tree (VST), which shares early layers across VFMs while retaining their later layers, yielding a controllable accuracy-latency trade-off for multimodal LLMs that often leverage multiple VFMs. Taken together, our study elevates stitching from a diagnostic probe to a practical recipe for integrating complementary VFM strengths and pinpointing where their representations align or diverge.\n\n---\n\n## 116. Test-Time Strategies for More Efficient and Accurate Agentic RAG\n- **Authors**: Brian Zhang , Deepti Guntur , Zhiyang Zuo , Abhinav Sharma , Shreyas Chaudhari , Wenlong Zhao , Franck Dernoncourt , Puneet Mathur , Ryan Rossi , Nedim Lipka\n- **URL**: [https://arxiv.org/abs/2603.12396](https://arxiv.org/abs/2603.12396)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) systems face challenges with complex, multihop questions, and agentic frameworks such as Search-R1 (Jin et al., 2025), which operates iteratively, have been proposed to address these complexities. However, such approaches can introduce inefficiencies, including repetitive retrieval of previously processed information and challenges in contextualizing retrieved results effectively within the current generation prompt. Such issues can lead to unnecessary retrieval turns, suboptimal reasoning, inaccurate answers, and increased token consumption. In this paper, we investigate test-time modifications to the Search-R1 pipeline to mitigate these identified shortcomings. Specifically, we explore the integration of two components and their combination: a contextualization module to better integrate relevant information from retrieved documents into reasoning, and a de-duplication module that replaces previously retrieved documents with the next most relevant ones. We evaluate our approaches using the HotpotQA (Yang et al., 2018) and the Natural Questions (Kwiatkowski et al., 2019) datasets, reporting the exact match (EM) score, an LLM-as-a-Judge assessment of answer correctness, and the average number of turns. Our best-performing variant, utilizing GPT-4.1-mini for contextualization, achieves a 5.6% increase in EM score and reduces the number of turns by 10.5% compared to the Search-R1 baseline, demonstrating improved answer accuracy and retrieval efficiency.\n\n---\n\n## 117. SPARROW: Learning Spatial Precision and Temporal Referential Consistency in Pixel-Grounded Video MLLMs\n- **Authors**: Mohamad Alansari , Naufal Suryanto , Divya Velayudhan , Sajid Javed , Naoufel Werghi , Muzammal Naseer\n- **URL**: [https://arxiv.org/abs/2603.12382](https://arxiv.org/abs/2603.12382)\n- **Abstract**:\n> Multimodal large language models (MLLMs) have advanced from image-level reasoning to pixel-level grounding, but extending these capabilities to videos remains challenging as models must achieve spatial precision and temporally consistent reference tracking. Existing video MLLMs often rely on a static segmentation token ([SEG]) for frame-wise grounding, which provides semantics but lacks temporal context, causing spatial drift, identity switches, and unstable initialization when objects move or reappear. We introduce SPARROW, a pixel-grounded video MLLM that unifies spatial accuracy and temporal stability through two key components: (i) Target-Specific Tracked Features (TSF), which inject temporally aligned referent cues during training, and (ii) a dual-prompt design that decodes box ([BOX]) and segmentation ([SEG]) tokens to fuse geometric priors with semantic grounding. SPARROW is supported by a curated referential video dataset of 30,646 videos and 45,231 Q&A pairs and operates end-to-end without external detectors via a class-agnostic SAM2-based proposer. Integrated into three recent open-source video MLLMs (UniPixel, GLUS, and VideoGLaMM), SPARROW delivers consistent gains across six benchmarks, improving up to +8.9 J&F on RVOS, +5 mIoU on visual grounding, and +5.4 CLAIR on GCG. These results demonstrate that SPARROW substantially improves referential stability, spatial precision, and temporal coherence in pixel-grounded video understanding. Project page: this https URL\n\n---\n\n## 118. Budget-Sensitive Discovery Scoring: A Formally Verified Framework for Evaluating AI-Guided Scientific Selection\n- **Authors**: Abhinaba Basu , Pavan Chakraborty\n- **URL**: [https://arxiv.org/abs/2603.12349](https://arxiv.org/abs/2603.12349)\n- **Abstract**:\n> Scientific discovery increasingly relies on AI systems to select candidates for expensive experimental validation, yet no principled, budget-aware evaluation framework exists for comparing selection strategies -- a gap intensified by large language models (LLMs), which generate plausible scientific proposals without reliable downstream evaluation. We introduce the Budget-Sensitive Discovery Score (BSDS), a formally verified metric -- 20 theorems machine-checked by the Lean 4 proof assistant -- that jointly penalizes false discoveries (lambda-weighted FDR) and excessive abstention (gamma-weighted coverage gap) at each budget level. Its budget-averaged form, the Discovery Quality Score (DQS), provides a single summary statistic that no proposer can inflate by performing well at a cherry-picked budget. As a case study, we apply BSDS/DQS to: do LLMs add marginal value to an existing ML pipeline for drug discovery candidate selection? We evaluate 39 proposers -- 11 mechanistic variants, 14 zero-shot LLM configurations, and 14 few-shot LLM configurations -- using SMILES representations on MoleculeNet HIV (41,127 compounds, 3.5% active, 1,000 bootstrap replicates) under both random and scaffold splits. Three findings emerge. First, the simple RF-based Greedy-ML proposer achieves the best DQS (-0.046), outperforming all MLP variants and LLM configurations. Second, no LLM surpasses the Greedy-ML baseline under zero-shot or few-shot evaluation on HIV or Tox21, establishing that LLMs provide no marginal value over an existing trained classifier. Third, the proposer hierarchy generalizes across five MoleculeNet benchmarks spanning 0.18%-46.2% prevalence, a non-drug AV safety domain, and a 9x7 grid of penalty parameters (tau >= 0.636, mean tau = 0.863). The framework applies to any setting where candidates are selected under budget constraints and asymmetric error costs.\n\n---\n\n## 119. Optimizing Task Completion Time Updates Using POMDPs\n- **Authors**: Duncan Eddy , Esen Yel , Emma Passmore , Niles Egan , Grayson Armour , Dylan M. Asmar , Mykel J. Kochenderfer\n- **URL**: [https://arxiv.org/abs/2603.12340](https://arxiv.org/abs/2603.12340)\n- **Abstract**:\n> Managing announced task completion times is a fundamental control problem in project management. While extensive research exists on estimating task durations and task scheduling, the problem of when and how to update completion times communicated to stakeholders remains understudied. Organizations must balance announcement accuracy against the costs of frequent timeline updates, which can erode stakeholder trust and trigger costly replanning. Despite the prevalence of this problem, current approaches rely on static predictions or ad-hoc policies that fail to account for the sequential nature of announcement management. In this paper, we formulate the task announcement problem as a Partially Observable Markov Decision Process (POMDP) where the control policy must decide when to update announced completion times based on noisy observations of true task completion. Since most state variables (current time and previous announcements) are fully observable, we leverage the Mixed Observability MDP (MOMDP) framework to enable more efficient policy optimization. Our reward structure captures the dual costs of announcement errors and update frequency, enabling synthesis of optimal announcement control policies. Using off-the-shelf solvers, we generate policies that act as feedback controllers, adaptively managing announcements based on belief state evolution. Simulation results demonstrate significant improvements in both accuracy and announcement stability compared to baseline strategies, achieving up to 75\\% reduction in unnecessary updates while maintaining or improving prediction accuracy.\n\n---\n\n## 120. Maximum Entropy Exploration Without the Rollouts\n- **Authors**: Jacob Adamczyk , Adam Kamoski , Rahul V. Kulkarni\n- **URL**: [https://arxiv.org/abs/2603.12325](https://arxiv.org/abs/2603.12325)\n- **Abstract**:\n> Efficient exploration remains a central challenge in reinforcement learning, serving as a useful pretraining objective for data collection, particularly when an external reward function is unavailable. A principled formulation of the exploration problem is to find policies that maximize the entropy of their induced steady-state visitation distribution, thereby encouraging uniform long-run coverage of the state space. Many existing exploration approaches require estimating state visitation frequencies through repeated on-policy rollouts, which can be computationally expensive. In this work, we instead consider an intrinsic average-reward formulation in which the reward is derived from the visitation distribution itself, so that the optimal policy maximizes steady-state entropy. An entropy-regularized version of this objective admits a spectral characterization: the relevant stationary distributions can be computed from the dominant eigenvectors of a problem-dependent transition matrix. This insight leads to a novel algorithm for solving the maximum entropy exploration problem, EVE (EigenVector-based Exploration), which avoids explicit rollouts and distribution estimation, instead computing the solution through iterative updates, similar to a value-based approach. To address the original unregularized objective, we employ a posterior-policy iteration (PPI) approach, which monotonically improves the entropy and converges in value. We prove convergence of EVE under standard assumptions and demonstrate empirically that it efficiently produces policies with high steady-state entropy, achieving competitive exploration performance relative to rollout-based baselines in deterministic grid-world environments.\n\n---\n\n## 121. Thermodynamics of Reinforcement Learning Curricula\n- **Authors**: Jacob Adamczyk , Juan Sebastian Rojas , Rahul V. Kulkarni\n- **URL**: [https://arxiv.org/abs/2603.12324](https://arxiv.org/abs/2603.12324)\n- **Abstract**:\n> Connections between statistical mechanics and machine learning have repeatedly proven fruitful, providing insight into optimization, generalization, and representation learning. In this work, we follow this tradition by leveraging results from non-equilibrium thermodynamics to formalize curriculum learning in reinforcement learning (RL). In particular, we propose a geometric framework for RL by interpreting reward parameters as coordinates on a task manifold. We show that, by minimizing the excess thermodynamic work, optimal curricula correspond to geodesics in this task space. As an application of this framework, we provide an algorithm, \"MEW\" (Minimum Excess Work), to derive a principled schedule for temperature annealing in maximum-entropy RL.\n\n---\n\n## 122. VQQA: An Agentic Approach for Video Evaluation and Quality Improvement\n- **Authors**: Yiwen Song , Tomas Pfister , Yale Song\n- **URL**: [https://arxiv.org/abs/2603.12310](https://arxiv.org/abs/2603.12310)\n- **Abstract**:\n> Despite rapid advancements in video generation models, aligning their outputs with complex user intent remains challenging. Existing test-time optimization methods are typically either computationally expensive or require white-box access to model internals. To address this, we present VQQA (Video Quality Question Answering), a unified, multi-agent framework generalizable across diverse input modalities and video generation tasks. By dynamically generating visual questions and using the resulting Vision-Language Model (VLM) critiques as semantic gradients, VQQA replaces traditional, passive evaluation metrics with human-interpretable, actionable feedback. This enables a highly efficient, closed-loop prompt optimization process via a black-box natural language interface. Extensive experiments demonstrate that VQQA effectively isolates and resolves visual artifacts, substantially improving generation quality in just a few refinement steps. Applicable to both text-to-video (T2V) and image-to-video (I2V) tasks, our method achieves absolute improvements of +11.57% on T2V-CompBench and +8.43% on VBench2 over vanilla generation, significantly outperforming state-of-the-art stochastic search and prompt optimization techniques.\n\n---\n\n## 123. HCP-DCNet: A Hierarchical Causal Primitive Dynamic Composition Network for Self-Improving Causal Understanding\n- **Authors**: Ming Lei , Shufan Wu , Christophe Baehr\n- **URL**: [https://arxiv.org/abs/2603.12305](https://arxiv.org/abs/2603.12305)\n- **Abstract**:\n> The ability to understand and reason about cause and effect -- encompassing interventions, counterfactuals, and underlying mechanisms -- is a cornerstone of robust artificial intelligence. While deep learning excels at pattern recognition, it fundamentally lacks a model of causality, making systems brittle under distribution shifts and unable to answer ``what-if'' questions. This paper introduces the \\emph{Hierarchical Causal Primitive Dynamic Composition Network (HCP-DCNet)}, a unified framework that bridges continuous physical dynamics with discrete symbolic causal inference. Departing from monolithic representations, HCP-DCNet decomposes causal scenes into reusable, typed \\emph{causal primitives} organized into four abstraction layers: physical, functional, event, and rule. A dual-channel routing network dynamically composes these primitives into task-specific, fully differentiable \\emph{Causal Execution Graphs (CEGs)}. Crucially, the system employs a \\emph{causal-intervention-driven meta-evolution} strategy, enabling autonomous self-improvement through a constrained Markov decision process. We establish rigorous theoretical guarantees, including type-safe composition, routing convergence, and universal approximation of causal dynamics. Extensive experiments across simulated physical and social environments demonstrate that HCP-DCNet significantly outperforms state-of-the-art baselines in causal discovery, counterfactual reasoning, and compositional generalization. This work provides a principled, scalable, and interpretable architecture for building AI systems with human-like causal abstraction and continual self-refinement capabilities.\n\n---\n\n## 124. A Geometrically-Grounded Drive for MDL-Based Optimization in Deep Learning\n- **Authors**: Ming Lei , Shufan Wu , Christophe Baehr\n- **URL**: [https://arxiv.org/abs/2603.12304](https://arxiv.org/abs/2603.12304)\n- **Abstract**:\n> This paper introduces a novel optimization framework that fundamentally integrates the Minimum Description Length (MDL) principle into the training dynamics of deep neural networks. Moving beyond its conventional role as a model selection criterion, we reformulate MDL as an active, adaptive driving force within the optimization process itself. The core of our method is a geometrically-grounded cognitive manifold whose evolution is governed by a \\textit{coupled Ricci flow}, enriched with a novel \\textit{MDL Drive} term derived from first principles. This drive, modulated by the task-loss gradient, creates a seamless harmony between data fidelity and model simplification, actively compressing the internal representation during training. We establish a comprehensive theoretical foundation, proving key properties including the monotonic decrease of description length (Theorem~\\ref{thm:convergence}), a finite number of topological phase transitions via a geometric surgery protocol (Theorems~\\ref{thm:surgery}, \\ref{thm:ultimate_fate}), and the emergence of universal critical behavior (Theorem~\\ref{thm:universality}). Furthermore, we provide a practical, computationally efficient algorithm with $O(N \\log N)$ per-iteration complexity (Theorem~\\ref{thm:complexity}), alongside guarantees for numerical stability (Theorem~\\ref{thm:stability}) and exponential convergence under convexity assumptions (Theorem~\\ref{thm:convergence_rate}). Empirical validation on synthetic regression and classification tasks confirms the theoretical predictions, demonstrating the algorithm's efficacy in achieving robust generalization and autonomous model simplification. This work provides a principled path toward more autonomous, generalizable, and interpretable AI systems by unifying geometric deep learning with information-theoretic principles.\n\n---\n\n## 125. Global Evolutionary Steering: Refining Activation Steering Control via Cross-Layer Consistency\n- **Authors**: Xinyan Jiang , Wenjing Yu , Di Wang , Lijie Hu\n- **URL**: [https://arxiv.org/abs/2603.12298](https://arxiv.org/abs/2603.12298)\n- **Abstract**:\n> Activation engineering enables precise control over Large Language Models (LLMs) without the computational cost of fine-tuning. However, existing methods deriving vectors from static activation differences are susceptible to high-dimensional noise and layer-wise semantic drift, often capturing spurious correlations rather than the target intent. To address this, we propose Global Evolutionary Refined Steering (GER-steer), a training-free framework that grounded in the geometric stability of the network's representation evolution. GER-steer exploits this global signal to rectify raw steering vectors, effectively decoupling robust semantic intent from orthogonal artifacts. Extensive evaluations confirm that GER-steer consistently outperforms baselines, delivering superior efficacy and generalization without layer-specific tuning, establishing a universal solution for reliable model alignment.\n\n---\n\n## 126. Synthetic Data Generation for Brain-Computer Interfaces: Overview, Benchmarking, and Future Directions\n- **Authors**: Ziwei Wang , Zhentao He , Xingyi He , Hongbin Wang , Tianwang Jia , Jingwei Luo , Siyang Li , Xiaoqing Chen , Dongrui Wu\n- **URL**: [https://arxiv.org/abs/2603.12296](https://arxiv.org/abs/2603.12296)\n- **Abstract**:\n> Deep learning has achieved transformative performance across diverse domains, largely driven by the large-scale, high-quality training data. In contrast, the development of brain-computer interfaces (BCIs) is fundamentally constrained by the limited, heterogeneous, and privacy-sensitive neural recordings. Generating synthetic yet physiologically plausible brain signals has therefore emerged as a compelling way to mitigate data scarcity and enhance model capacity. This survey provides a comprehensive review of brain signal generation for BCIs, covering methodological taxonomies, benchmark experiments, evaluation metrics, and key applications. We systematically categorize existing generative algorithms into four types: knowledge-based, feature-based, model-based, and translation-based approaches. Furthermore, we benchmark existing brain signal generation approaches across four representative BCI paradigms to provide an objective performance comparison. Finally, we discuss the potentials and challenges of current generation approaches and prospect future research on accurate, data-efficient, and privacy-aware BCI systems. The benchmark codebase is publicized at this https URL .\n\n---\n\n## 127. Detecting Miscitation on the Scholarly Web through LLM-Augmented Text-Rich Graph Learning\n- **Authors**: Huidong Wu , Haojia Xiang , Jingtong Gao , Xiangyu Zhao , Dengsheng Wu , Jianping Li\n- **URL**: [https://arxiv.org/abs/2603.12290](https://arxiv.org/abs/2603.12290)\n- **Abstract**:\n> Scholarly web is a vast network of knowledge connected by citations. However, this system is increasingly compromised by miscitation, where references do not support or even contradict the claims they are cited for. Current miscitation detection methods, which primarily rely on semantic similarity or network anomalies, struggle to capture the nuanced relationship between a citation's context and its place in the wider network. While large language models (LLMs) offer powerful capabilities in semantic reasoning for this task, their deployment is hindered by hallucination risks and high computational costs. In this work, we introduce LLM-Augmented Graph Learning-based Miscitation Detector (LAGMiD), a novel framework that leverages LLMs for deep semantic reasoning over citation graphs and distills this knowledge into graph neural networks (GNNs) for efficient and scalable miscitation detection. Specifically, LAGMiD introduces an evidence-chain reasoning mechanism, which uses chain-of-thought prompting, to perform multi-hop citation tracing and assess semantic fidelity. To reduce LLM inference costs, we design a knowledge distillation method aligning GNN embeddings with intermediate LLM reasoning states. A collaborative learning strategy further routes complex cases to the LLM while optimizing the GNN for structure-based generalization. Experiments on three real-world benchmarks show that LAGMiD achieves state-of-the-art miscitation detection with significantly reduced inference cost.\n\n---\n\n## 128. From Garbage to Gold: A Data-Architectural Theory of Predictive Robustness\n- **Authors**: Terrence J. Lee-St. John , Jordan L. Lawson , Bartlomiej Piechowski-Jozwiak\n- **URL**: [https://arxiv.org/abs/2603.12288](https://arxiv.org/abs/2603.12288)\n- **Abstract**:\n> Tabular machine learning presents a paradox: modern models achieve state-of-the-art performance using high-dimensional (high-D), collinear, error-prone data, defying the \"Garbage In, Garbage Out\" mantra. To help resolve this, we synthesize principles from Information Theory, Latent Factor Models, and Psychometrics, clarifying that predictive robustness arises not solely from data cleanliness, but from the synergy between data architecture and model capacity. Partitioning predictor-space \"noise\" into \"Predictor Error\" and \"Structural Uncertainty\" (informational deficits from stochastic generative mappings), we prove that leveraging high-D sets of error-prone predictors asymptotically overcomes both types of noise, whereas cleaning a low-D set is fundamentally bounded by Structural Uncertainty. We demonstrate why \"Informative Collinearity\" (dependencies from shared latent causes) enhances reliability and convergence efficiency, and explain why increased dimensionality reduces the latent inference burden, enabling feasibility with finite samples. To address practical constraints, we propose \"Proactive Data-Centric AI\" to identify predictors that enable robustness efficiently. We also derive boundaries for Systematic Error Regimes and show why models that absorb \"rogue\" dependencies can mitigate assumption violations. Linking latent architecture to Benign Overfitting, we offer a first step towards a unified view of robustness to Outcome Error and predictor-space noise, while also delineating when traditional DCAI's focus on label cleaning remains powerful. By redefining data quality from item-level perfection to portfolio-level architecture, we provide a theoretical rationale for \"Local Factories\" -- learning from live, uncurated enterprise \"data swamps\" -- supporting a deployment paradigm shift from \"Model Transfer\" to \"Methodology Transfer'' to overcome static generalizability limitations.\n\n---\n\n## 129. The DIME Architecture: A Unified Operational Algorithm for Neural Representation, Dynamics, Control and Integration\n- **Authors**: Ionel Cristian Vladu , Nicu Bizdoaca , Ionica Pirici , Tudor-Adrian Balseanu , Eduard Nicusor Bondoc\n- **URL**: [https://arxiv.org/abs/2603.12286](https://arxiv.org/abs/2603.12286)\n- **Abstract**:\n> Modern neuroscience has accumulated extensive evidence on perception, memory, prediction, valuation, and consciousness, yet still lacks an explicit operational architecture capable of integrating these phenomena within a unified computational framework. Existing theories address specific aspects of neural function: predictive coding and active inference emphasize hierarchical inference and prediction error minimization; engram theories explain memory through distributed cell assemblies; neuromodulatory accounts focus on value-dependent regulation of plasticity and behaviour; and global workspace or large-scale network models investigate mechanisms underlying conscious access. Despite their explanatory power, these approaches remain only partially integrated at the architectural level. This work introduces DIME (Detect-Integrate-Mark-Execute), a neural architecture organizing perception, memory, valuation, and conscious access within a common operational cycle. The framework includes four interacting components: engrams, distributed recurrent neural structures supporting multiple activation trajectories; execution threads, spatiotemporal trajectories implementing neural processes; marker systems, neuromodulatory and limbic mechanisms regulating gain, plasticity, and trajectory selection; and hyperengrams, large-scale integrative states associated with operational conscious access. The framework is consistent with empirical evidence from hippocampal indexing, recurrent cortical processing, replay phenomena, large-scale network integration, and neuromodulatory regulation. Formulated at an abstract computational level, DIME may also inform artificial intelligence and robotics by providing an architectural template in which representation, valuation, and temporal sequencing emerge from a unified mechanism. An extended theoretical exposition is available in a companion monograph on Zenodo.\n\n---\n\n## 130. Predictive Analytics for Foot Ulcers Using Time-Series Temperature and Pressure Data\n- **Authors**: Md Tanvir Hasan Turja\n- **URL**: [https://arxiv.org/abs/2603.12278](https://arxiv.org/abs/2603.12278)\n- **Abstract**:\n> Diabetic foot ulcers (DFUs) are a severe complication of diabetes, often resulting in significant morbidity. This paper presents a predictive analytics framework utilizing time-series data captured by wearable foot sensors -- specifically NTC thin-film thermocouples for temperature measurement and FlexiForce pressure sensors for plantar load monitoring. Data was collected from healthy subjects walking on an instrumented pathway. Unsupervised machine learning algorithms, Isolation Forest and K-Nearest Neighbors (KNN), were applied to detect anomalies that may indicate early ulcer risk. Through rigorous data preprocessing and targeted feature engineering, physiologic patterns were extracted to identify subtle changes in foot temperature and pressure. Results demonstrate Isolation Forest is sensitive to micro-anomalies, while KNN is effective in flagging extreme deviations, albeit at a higher false-positive rate. Strong correlations between temperature and pressure readings support combined sensor monitoring for improved predictive accuracy. These findings provide a basis for real-time diabetic foot health surveillance, aiming to facilitate earlier intervention and reduce DFU incidence.\n\n---\n\n## 131. Prompt Injection as Role Confusion\n- **Authors**: Charles Ye , Jasmine Cui , Dylan Hadfield-Menell\n- **URL**: [https://arxiv.org/abs/2603.12277](https://arxiv.org/abs/2603.12277)\n- **Abstract**:\n> Language models remain vulnerable to prompt injection attacks despite extensive safety training. We trace this failure to role confusion: models infer roles from how text is written, not where it comes from. We design novel role probes to capture how models internally identify \"who is speaking.\" These reveal why prompt injection works: untrusted text that imitates a role inherits that role's authority. We test this insight by injecting spoofed reasoning into user prompts and tool outputs, achieving average success rates of 60% on StrongREJECT and 61% on agent exfiltration, across multiple open- and closed-weight models with near-zero baselines. Strikingly, the degree of internal role confusion strongly predicts attack success before generation begins. Our findings reveal a fundamental gap: security is defined at the interface but authority is assigned in latent space. More broadly, we introduce a unifying, mechanistic framework for prompt injection, demonstrating that diverse prompt-injection attacks exploit the same underlying role-confusion mechanism.\n\n---\n\n## 132. Aligning Language Models from User Interactions\n- **Authors**: Thomas Kleine Buening , Jonas Hübotter , Barna Pásztor , Idan Shenfeld , Giorgia Ramponi , Andreas Krause\n- **URL**: [https://arxiv.org/abs/2603.12273](https://arxiv.org/abs/2603.12273)\n- **Abstract**:\n> Multi-turn user interactions are among the most abundant data produced by language models, yet we lack effective methods to learn from them. While typically discarded, these interactions often contain useful information: follow-up user messages may indicate that a response was incorrect, failed to follow an instruction, or did not align with the user's preferences. Importantly, language models are already able to make use of this information in context. After observing a user's follow-up, the same model is often able to revise its behavior. We leverage this ability to propose a principled and scalable method for learning directly from user interactions through self-distillation. By conditioning the model on the user's follow-up message and comparing the resulting token distribution with the original policy, we obtain a target for updating the policy that captures how the model's behavior changes in hindsight. We then distill this hindsight distribution back into the current policy. Remarkably, we show that training on real-world user conversations from WildChat improves language models across standard alignment and instruction-following benchmarks, without regressing other capabilities. The same mechanism enables personalization, allowing models to continually adapt to individual users through interaction without explicit feedback. Our results demonstrate that raw user interactions that arise naturally during deployment enable alignment, personalization, and continual adaptation.\n\n---\n\n## 133. Diagnosing Retrieval Bias Under Multiple In-Context Knowledge Updates in Large Language Models\n- **Authors**: Boyu Qiao , Sean Guo , Xian Yang , Kun Li , Wei Zhou , Songlin Hu , Yunya Song\n- **URL**: [https://arxiv.org/abs/2603.12271](https://arxiv.org/abs/2603.12271)\n- **Abstract**:\n> LLMs are widely used in knowledge-intensive tasks where the same fact may be revised multiple times within context. Unlike prior work focusing on one-shot updates or single conflicts, multi-update scenarios contain multiple historically valid versions that compete at retrieval, yet remain underexplored. This challenge resembles the AB-AC interference paradigm in cognitive psychology: when the same cue A is successively associated with B and C, the old and new associations compete during retrieval, leading to bias. Inspired by this, we introduce a Dynamic Knowledge Instance (DKI) evaluation framework, modeling multi-updates of the same fact as a cue paired with a sequence of updated values, and assess models via endpoint probing of the earliest (initial) and latest (current) states. Across diverse LLMs, we observe that retrieval bias intensifies as updates increase, earliest-state accuracy stays high while latest-state accuracy drops substantially. Diagnostic analyses of attention, hidden-state similarity, and output logits further reveal that these signals become flatter and weakly discriminative on errors, providing little stable basis for identifying the latest update. Finally, cognitively inspired heuristic intervention strategies yield only modest gains and do not eliminate the bias. Our results reveal a persistent challenge in tracking and following knowledge updates in long contexts.\n\n---\n\n## 134. Task-Specific Knowledge Distillation via Intermediate Probes\n- **Authors**: Ryan Brown , Chris Russell\n- **URL**: [https://arxiv.org/abs/2603.12270](https://arxiv.org/abs/2603.12270)\n- **Abstract**:\n> Knowledge distillation from large language models (LLMs) assumes that the teacher's output distribution is a high-quality training signal. On reasoning tasks, this assumption is frequently violated. A model's intermediate representations may encode the correct answer, yet this information is lost or distorted through the vocabulary projection, where prompt formatting and answer-token choices creates brittle, noisy outputs. We introduce \\method{}, a distillation framework that bypasses this bottleneck by training lightweight probes on frozen teacher hidden states and using the probe's predictions, rather than output logits, as supervision for student training. This simple change yields consistent improvements across four reasoning benchmarks (AQuA-RAT, ARC Easy/Challenge, and MMLU), with gains most pronounced under limited data. Probes trained on intermediate representations provide cleaner labels than the teacher's own outputs, effectively denoising the distillation signal. \\method{} requires no architectural changes to student or teacher, is architecture-agnostic, and adds minimal compute since probe training is cheap and teacher representations can be cached. By exploiting internal representations, \\method{} enables practitioners to extract more value from large teacher models without additional training data or architectural complexity.\n\n---\n\n## 135. DART: Input-Difficulty-AwaRe Adaptive Threshold for Early-Exit DNNs\n- **Authors**: Parth Patne , Mahdi Taheri , Christian Herglotz , Maksim Jenihhin , Milos Krstic , Michael Hübner\n- **URL**: [https://arxiv.org/abs/2603.12269](https://arxiv.org/abs/2603.12269)\n- **Abstract**:\n> Early-exit deep neural networks enable adaptive inference by terminating computation when sufficient confidence is achieved, reducing cost for edge AI accelerators in resource-constrained settings. Existing methods, however, rely on suboptimal exit policies, ignore input difficulty, and optimize thresholds independently. This paper introduces DART (Input-Difficulty-Aware Adaptive Threshold), a framework that overcomes these limitations. DART introduces three key innovations: (1) a lightweight difficulty estimation module that quantifies input complexity with minimal computational overhead, (2) a joint exit policy optimization algorithm based on dynamic programming, and (3) an adaptive coefficient management system. Experiments on diverse DNN benchmarks (AlexNet, ResNet-18, VGG-16) demonstrate that DART achieves up to \\textbf{3.3$\\times$} speedup, \\textbf{5.1$\\times$} lower energy, and up to \\textbf{42\\%} lower average power compared to static networks, while preserving competitive accuracy. Extending DART to Vision Transformers (LeViT) yields power (5.0$\\times$) and execution-time (3.6$\\times$) gains but also accuracy loss (up to 17 percent), underscoring the need for transformer-specific early-exit mechanisms. We further introduce the Difficulty-Aware Efficiency Score (DAES), a novel multi-objective metric, under which DART achieves up to a 14.8 improvement over baselines, highlighting superior accuracy, efficiency, and robustness trade-offs.\n\n---\n\n{% endraw %}\n"}}, "LLM/2026-06-29.md": {"64bbd1731dde152451afe1705df222e26b5c83fa": {"t": 1785747188.024377, "n": 84889, "h": "29e92ec4f78e9471", "c": "---\nlayout: default\ntitle: \"LLM 관련 주요 논문 - 2026-06-29\"\n---\n\n{% raw %}\n# LLM 관련 주요 논문 - 2026-06-29\n\n## 1. Tandem Reinforcement Learning with Verifiable Rewards\n- **Authors**: Difan Jiao , Raghav Singhal , Robert West , Ashton Anderson\n- **URL**: [https://arxiv.org/abs/2606.28166](https://arxiv.org/abs/2606.28166)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has significantly improved the reasoning capability of large language models, reaching expert or even superhuman performance in domains such as competition math. However, whether weaker agents and humans can actually harness this capability is far less certain, with RLVR documented to drift reasoning toward idiosyncratic patterns such as poor readability and language mixing. Tandem training is a recently introduced paradigm that targets this compatibility problem: a trained, stronger senior co-generates each rollout with a frozen, weaker junior, and the two are rewarded as a team, so the senior is pushed to reason in ways the junior can follow. Yet this paradigm has so far been demonstrated only in proof-of-concept settings, leaving open whether it scales to the long chains of thought of the modern RLVR pipeline. In this work, we propose Tandem Reinforcement Learning (TRL), which carries the tandem training paradigm into RLVR. In TRL, the senior and a frozen junior alternate stochastically to co-generate the reasoning, the resulting generation is rewarded, and the standard GRPO loss is applied to the senior. Training Qwen3-4B-Instruct on competition math, we find that TRL matches vanilla GRPO on solo reasoning capability while three properties emerge together from the same rollout structure: stronger handoff robustness with the junior, reduced distributional drift from the junior, and a chain-of-thought more legible to the junior. Our results demonstrate a promising route for RLVR with practical payoffs in multi-model communication and human compatibility.\n\n---\n\n## 2. JD Oxygen AI Item Center (Oxygen AIIC) V1: An Industrial-Scale LLM/VLM-Centric Solution for Item Understanding, Management, and Applications\n- **Authors**: Oxygen AIIC , Chan Long , Chao Liu , Chaofan Chen , Chaohui Dong , Chunyuan Guo , Danping Liu , Debin Liu , Deping Xiang , Fulai Xu , Guangyue Liu , Hao Li , Huichun Hu , Jian Yang , Jianan Wang , Jianbo Zhao , Jiaoyang Li , Jiaxing Wang , Jinglong Li , Jinjin Guo , Jun Fang , Jun Liu , Kai Zhou , Li Wang , Lili Gao , Liying Chen , Luning Yang , Mengdi Zhou , Pengzhang Liu , Qi Lv , Qianyun Wang , Qixia Jiang , Ruyue Li , Shimu Liang , Shuxing Wang , Sijie Zhang , Siqi Li , Tianhao Gao , Wang Ke , Weihu Huang , Wencan Lai , Wenjie Zhang , Xiaohui Zhang , Xiaojing Dong , Ya Liu , Yifeng Zhang , Yixiang Wang , Yongtai Zhang , Yongyi Liao , Zhaoru Chen , Zhen Chen , Zhiyong Ma , Zhiyuan Liu , Zhongwei Liu , Ziyan Xing\n- **URL**: [https://arxiv.org/abs/2606.28070](https://arxiv.org/abs/2606.28070)\n- **Abstract**:\n> this http URL , one of the world's largest e-commerce platforms, serves over 700 million active users and millions of merchants, with a catalog of tens of billions of SKUs. At this scale, high-quality, structured item knowledge underpins a better consumer experience, lower management costs, and higher operational efficiency-yet producing and serving it poses three industrial-scale challenges: fast-emerging concepts, high-quality knowledge production for massive SKUs, and diverse downstream requirements. To address these challenges, we present the JD Oxygen AI Item Center (Oxygen AIIC), an industrial-scale platform built on LLMs/VLMs for item-knowledge production and service. Oxygen AIIC is built around four core pillars: (i) ontology engineering driven by efficient human-AI collaboration, which supports the dynamic evolution and agile expansion of an ontology with millions of entries; (ii) a \"Semantic Search then Discrimination\"(S2D) knowledge identification architecture that, combined with throughput improvement strategies, enables scalable, extensible, and high-throughput AI Item Library production for tens of billions of SKUs; (iii) self-evolving item-understanding LLMs/VLMs that improve in a stable and controllable manner, enabling knowledge production with 94.2% precision and 82.8% recall; and (iv) a unified item tunnel that serves as the data and service hub. Oxygen AIIC now covers tens of thousands of JD categories and processes hundreds of millions of item updates per day on Huawei Ascend NPUs. It has accumulated hundreds of billions of item-knowledge assets. Deployed across core business scenarios-including search, recommendation, operations, category planning-Oxygen AIIC has delivered measurable gains at scale. Search-traffic coverage reaches 80.4%, item-information quality issues drop by 37%, the automated fill rate of core attributes during item listing exceeds 80%.\n\n---\n\n## 3. NormAct: A Benchmark for Hidden Social Norm Compliance in Embodied Planning\n- **Authors**: Shiyun Zhao , Xinwei Song , Tianyu Guo , Xiaomeng Gao , Mingyuan Liu , Xu Han , Yuanyuan Zhang , Zhenliang Zhang , Xue Feng , Bo Dai\n- **URL**: [https://arxiv.org/abs/2606.27826](https://arxiv.org/abs/2606.27826)\n- **Abstract**:\n> Multimodal large language models (MLLMs) are increasingly deployed as embodied planners in egocentric environments, where task success requires not only achieving instructed goals but also acting in socially appropriate ways. While explicit goals may render certain actions optimal, implicit social norms often impose hidden constraints. Existing evaluations typically focus on explicit goal achievement or direct norm knowledge, seldom assessing whether planners can infer and apply these hidden constraints within action sequences. We introduce NormAct, a benchmark for embodied social-norm interactions that evaluates plans on Goal Achievement, Norm Compliance, and overall Task Success. NormAct uniquely embeds hidden norms within ordinary tasks, testing whether models can realize them without explicit instruction. Experiments with state-of-the-art MLLMs (GPT-5.4, Claude Opus 4.7, Gemini 3 Pro) reveal a significant gap: models achieve explicit goals in 67.3\\% of cases, but comply with hidden norms in only 26.4\\%. Cue-condition experiments indicate that this gap stems not from a lack of general social knowledge, but from challenges in activating and grounding relevant norms in context. To address this, we propose NormPerceptor, a context-conditioned cue generator that infers scene-relevant norms prior to planning, increasing Task Success from 24.2\\% to 46.7\\%. Our results underscore the importance of enabling embodied agents to proactively detect hidden norms, ground them in visual evidence, and integrate them as action-planning constraints. Our benchmark is publicly available at this https URL .\n\n---\n\n## 4. Grounded Iterative Language Planning: How Parameterized World Models Reduce Hallucination Propagation in LLM Agents\n- **Authors**: Xinyuan Song , Zekun Cai\n- **URL**: [https://arxiv.org/abs/2606.27806](https://arxiv.org/abs/2606.27806)\n- **Abstract**:\n> World models for language agents come in two useful forms. An agent-based world model calls an LLM API and reasons flexibly in language, but its errors appear as hallucinated state changes that are hard to score with ordinary regression losses. A parameterized world model is a trained transition predictor; its errors are easier to measure with quantities such as NodeMSE, delta accuracy, and validity accuracy, but it is usually weaker as a standalone planner. We compare these two families on four graph-structured planning benchmarks and introduce operational hallucination metrics for the agent-based case. The comparison motivates \\textbf{Grounded Iterative Language Planning} (GILP), which trains only a small parameterized backbone and combines it with API-based agent reasoning. The backbone supplies valid actions, predicted state deltas, risk, and value; the LLM drafts an action and imagined delta; and a consistency gate asks for revision when the two disagree. On real GPT-4o-mini calls, GILP reduces hallucinated-state rate from 0.176 to 0.035. In calibrated simulator ablations, it raises success from 0.668 to 0.838 while adding only ~22% extra LLM calls.\n\n---\n\n## 5. Towards Reliable and Robust LLM Planning: Symbolic Feedback-Driven Iterative Self-Refinement Framework\n- **Authors**: Jiajing Zhang , Jiamei Jiang , Chenyang Zhang , Feifei Mo , Linjing Li , Daniel Zeng\n- **URL**: [https://arxiv.org/abs/2606.27757](https://arxiv.org/abs/2606.27757)\n- **Abstract**:\n> Large language models (LLMs) have attracted widespread attention from academia and industry, yet their deployment raises critical security concerns regarding robustness and reliability. Planning, a core component of intelligent behavior, remains challenging for LLMs, which often produce infeasible or incorrect solutions in long-horizon decision-making tasks due to inherent complexity. In this paper, we propose a symbolic feedback-driven iterative self-refinement framework to enhance the robustness and reliability of LLMs in long-horizon planning. Specifically, a natural language prompting mechanism is introduced to map logical symbols into natural language descriptions, enabling LLMs to better capture task constraints and semantics. We further design a symbolic verifier that identifies errors and converts them into corrective instructions interpretable by the LLM, thereby guiding self-refinement. In addition, we leverage a plan recognizer to infer goal reachability, facilitating more effective guidance toward desired goals. Empirical results demonstrate that the proposed framework consistently improves both feasibility and correctness in long-horizon planning tasks. This highlights its effectiveness in enhancing the reliability of LLM-based planning and potential to enable more trustworthy AI systems.\n\n---\n\n## 6. ToE: A Hierarchical and Explainable Claim Verification Framework with Dynamic Multi-source Evidence Retrieval and Aggregation\n- **Authors**: Zhaoqi Wang , Zijian Zhang , Kun Zheng , Zhen Li , Xin Li , Chunlei Li , Jiamou Liu\n- **URL**: [https://arxiv.org/abs/2606.27736](https://arxiv.org/abs/2606.27736)\n- **Abstract**:\n> The rapid spread of fake news poses increasing threats to information ecosystems, especially as AI-generated misinformation under Generative Engine Optimization (GEO) poisoning allows adversarially crafted content to be systematically surfaced by retrieval systems, contaminating LLM reasoning. In this paper, we propose Tree of Evidence (ToE), a hierarchical evidence reasoning framework for automated fact-checking that models each claim as a dynamically expanding argument tree. ToE integrates a reinforcement learning-driven multi-source retrieval agent, an evidence evaluation agent, and an argument tree aggregation algorithm to iteratively decompose, retrieve, and verify claims through an explainable evidence chain. We further provide a theoretical analysis of the retrieval process, deriving a formal error bound that guarantees the learned policy converges to a neighborhood of the information-theoretically optimal policy. Experiments across multiple datasets and backbone LLMs demonstrate that ToE achieves improvements ranging from 4 to 24 percentage points over competitive baselines, with particularly pronounced gains on adversarially poisoned inputs.\n\n---\n\n## 7. DysLexLens: A Low-Resource LLM Framework for Analysing Dyslexic Learners Insights from Online Forums\n- **Authors**: Dana Rezazadegan , Atie Kia , Phongpadid Nandavong , Dominique Carlon , Jeremy Nguyen , Abhik Banerjee , James Marshall , Anthony McCosker , Yong-Bin Kang\n- **URL**: [https://arxiv.org/abs/2606.27619](https://arxiv.org/abs/2606.27619)\n- **Abstract**:\n> Dyslexic learners increasingly use artificial intelligence (AI) tools to support reading, writing, organisation, and study-related tasks. However, their lived experiences with these tools remain largely underexamined. This paper proposes DysLexLens, a low-resource LLM framework, designed to analyse dyslexic learners experience with AI through online forum discussions. DysLexLens is designed as an end-to-end, evidence-traceable architecture which transforms noisy social media posts into a dictionary-driven corpora, provides knowledge-graph (KG)-based question reasoning, generates verifiable query responses, and enables response evaluation through quantitative and human-grounded assessment. DysLexLens has four key features. First, it employs a dictionary-driven filtering method to construct a more focused Reddit corpus on dyslexia and AI, filtering out noisy and weakly related posts to improve the relevance of data collected from low-resource forum contexts. Second, it integrates LLM-assisted semantic analysis with KG-based query reasoning to uncover meaningful patterns. Third, it has quantitative evaluation metrics (RAGAS and Query Robustness) to measure LLM-generated response performance. Fourth, it provides structured qualitative validation guidelines for assessing response quality, with a specific focus on hallucination and evidence alignment. We demonstrate the effectiveness of DysLexLens using dyslexia-related Reddit forum data and 30 questions. The results show its potential generalisability to other low-resource forum data contexts. DysLexLens, sample data, questions and evaluation results are available at Github to support reproducibility.\n\n---\n\n## 8. Odyssey: Constructing Verifiable Local Truth-Preserving Foundation Models\n- **Authors**: Sridhar Mahadevan\n- **URL**: [https://arxiv.org/abs/2606.27593](https://arxiv.org/abs/2606.27593)\n- **Abstract**:\n> We introduce a categorical framework called ODYSSEY for constructing verifiable, local truth-preserving foundation models as compositions of foundries: building-block architectural components that specify a cover of local contexts, local representation families, restriction maps, gluing rules, obstruction policies, update obligations, and human-facing views. A foundry is an organized sheaf of knowledge that carries within it an argumentation component. Concrete foundries are built from generic foundries such as evidence/argument, operational decision, institutional/financial, market meaning, scientific challenge, research-program, assistant-build, and evaluation-harness foundries. Universal Foundry Learning (UFL) formalizes foundry construction as a composition of left and right Kan extensions, with left Kan extension rolling local artifacts into candidate foundries and right Kan extension enforcing the restriction, gluing, obstruction, and argumentation conditions required for promotion. Foundry SQL (FSQL) is a small typed query surface for slicing maintained foundry artifacts that uses TICKET (Topos Integration using Causal Kan Extension Transformers) certification for admitting external or pre-built models into durable ODYSSEY state. ODYSSEY is fully implemented and tested across a wide spectrum of concrete foundries, showing that the same categorical machinery supports domain construction, artifact replay, sheaf diagnostics, grounded Toulmin/local-LLM scrutiny, residual-obstruction ledgers, and optimized TICKET-compatible causal-claim extraction across heterogeneous sources. This paper is to be presented as a 2.5 hour tutorial at ICML 2026. The tutorial home page is at this https URL .\n\n---\n\n## 9. Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning\n- **Authors**: Xuan Zhang , Zhijian Zhou , Lingfeng Qiao , Yulei Qin , Ke Li , Xing Sun , Xiaoyu Tan , Chao Qu , Yuan Qi\n- **URL**: [https://arxiv.org/abs/2606.27483](https://arxiv.org/abs/2606.27483)\n- **Abstract**:\n> Large language model (LLM) agents have demonstrated strong capability in sequential decision-making, yet they remains fundamentally reactive in long-horizon tasks. Unlike humans who employ \"what-if\" reasoning to evaluate potential plans before commitment, standard agents lack an internal world model to simulate future outcomes. Therefore, we propose to internalize future-aware planning by training a single autoregressive model to verbalize both a prospective state rollout and a plan-conditioned success estimate-a textual analogue of the Q-value. Crucially, we identify a format-capability gap: simply fine-tuning agents on look-ahead traces during post-training leads to superficial mimicry of foresight without genuine predictive grounding. To bridge this gap, we introduce a three-stage training paradigm: (i) World Model Agentic Mid-Training (WM-AMT) to inject latent predictive capabilities into the policy; (ii) Format-Eliciting SFT (FE-SFT) to structure this injected capability; and (iii) Foresight-Conditioned Reinforcement Learning (FC-RL) to refine the calibration and utility of the generated simulations. Evaluated on search and mathematical reasoning tasks, our approach consistently outperforms other training baselines. Our results demonstrate that effective internal world modeling in LLM agents requires a capability-first training pipeline to achieve grounded and calibrated foresight.\n\n---\n\n## 10. When Does Personality Composition Matter for Multi-Agent LLM Teams?\n- **Authors**: Aryan Keluskar , Amrita Bhattacharjee , Huan Liu\n- **URL**: [https://arxiv.org/abs/2606.27443](https://arxiv.org/abs/2606.27443)\n- **Abstract**:\n> Personality prompting shapes how large language models communicate, yet whether these behavioral shifts affect objective task outcomes remains under-explored. Prior work shows that agents prompted with low agreeableness produce adversarial language, while those prompted with high agreeableness become cooperative, but the relationship between communication style and task performance has not been systematically examined across multiple domains. In this work, we investigate whether personality composition matters for multi-agent team performance by manipulating personality traits across frontier LLMs on three task domains: structured coding, open-ended research collaboration, and competitive bargaining. We find that personality effects depend critically on task structure. In coding tasks, low agreeableness leads to large communication shifts that have little effect on milestone completion. In open-ended collaboration and bargaining, the same manipulation substantially degrades performance. We discuss implications for multi-agent system design and the limits of personality manipulation.\n\n---\n\n## 11. AI-Model Network: Concept, Current State and Future\n- **Authors**: Li Zhetao , Zeng Xiyu , Wang Jianhui , Xiao Yong , Liu Zhongren , Wu Junru , Lai Junjie , Huang Jijun , Long Saiqin\n- **URL**: [https://arxiv.org/abs/2606.27382](https://arxiv.org/abs/2606.27382)\n- **Abstract**:\n> While the primary function of computers lies in computation and processing, the core value of the Internet is rooted in sharing and collaboration. Computers create the Internet, and the Internet empowers the value of computers. The rapid development of the Internet, cloud computing, and big data is pushing artificial intelligence into the era of large models (LMs). However, the practical application of LMs is currently hindered by high training costs and deployment complexities, driving a shift toward lightweight, private, and domain-specific models. With the rapid proliferation and wide distribution of heterogeneous models, enabling effective interaction and collaboration among them has emerged as a critical bottleneck that urgently needs to be addressed in LM development. Drawing inspiration from the development of the Internet, this paper proposes the concept, vision, and system architecture of world wide AI-model network (AI-ModelNet). It is a novel paradigm that achieves interconnection, capability sharing, and collaborative reasoning by establishing pathways between models. We first briefly review the current state of single-model and multi-model research. Subsequently, the systemic vision and hierarchical architecture of AI-ModelNet are articulated, followed by validation of the framework's feasibility through a prototype system and diverse application cases. Finally, key directions for future research are discussed preliminarily.\n\n---\n\n## 12. Cognitive Episodes in LLM Reasoning Traces Enable Interpretable Human Item Difficulty Prediction\n- **Authors**: Chenguang Wang , Ming Li , Xinyue Zeng , Zhuochun Li , Hong Jiao , Tianyi Zhou , Dawei Zhou\n- **URL**: [https://arxiv.org/abs/2606.28186](https://arxiv.org/abs/2606.28186)\n- **Abstract**:\n> Predicting human item difficulty is central to educational assessment, where reliable estimates support fairness and effective test construction. Existing methods often depend on costly human calibration or item-level textual representations, providing limited evidence about the cognitive processes that make items difficult. We argue that difficulty should be viewed not only as a property of item text, but also as an observable consequence of the problem-solving burden an item induces. Large Reasoning Models (LRMs) offer scalable process evidence through reasoning traces, but such evidence must be structured to support interpretable modeling. To this end, we introduce Epi2Diff (Episode to Difficulty), a framework that maps LRM reasoning traces into cognitively grounded episode sequences. These episodes group trace segments into functional problem-solving states, enabling difficulty to be modeled through reasoning scale, effort allocation, and state transitions. Epi2Diff extracts compact episode-dynamic features and combines them with semantic item representations for human difficulty prediction. Experiments on four real-world human difficulty datasets show that Epi2Diff consistently outperforms strong baselines, including fine-tuned small language models, LLM in-context learning, and supervised LLM adaptation. On SAT-derived classification benchmarks, Epi2Diff achieves an 8.1% average relative gain over supervised LLM fine-tuning baselines. Further analyses show that harder items induce more effortful, iterative, and implementation-centered episode dynamics, rather than merely longer responses. These results demonstrate that cognitive episodes in LRM reasoning traces provide a predictive and interpretable process representation for human item difficulty, offering a new lens for educational measurement with reasoning models.\n\n---\n\n## 13. LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior\n- **Authors**: Qinhong Zhou , Chuang Gan , Anoop Cherian\n- **URL**: [https://arxiv.org/abs/2606.28182](https://arxiv.org/abs/2606.28182)\n- **Abstract**:\n> Embodied agents operating in decentralized and partially observable environments have attracted growing attention in recent years. However, existing large language model (LLM)-based agents often exhibit behaviors that are misaligned with their partners or inconsistent with the environment state, leading to inefficient cooperation and poor task success. To address this challenge, we propose a novel framework, Learning Laws of Cooperation (LLawCo), that enables embodied agents to autonomously align with both their partners and task objectives. Our framework allows agents to reflect on past failures to extract misaligned behavioral patterns, which are used to derive high-level behavioral laws, such as \"Talk when necessary\" and \"Wait for partner.\" These laws are explicitly incorporated into the agents' chains of thought via supervised fine-tuning, aligning their reasoning with task requirements and the behavior of other agents. To evaluate our approach, we introduce PARTNR-Dialog, a large-scale multi-agent communicative and cooperative planning benchmark built on the PARTNR environment. Experiments on existing tasks and our new benchmark demonstrate significant improvements in cooperative efficiency and task success rates. Across four backbone LLMs, our method achieves average success rate improvements of 4.5% on the PARTNR-Dialog benchmark and 6.8% on the TDW-MAT benchmark over state-of-the-art open-source communicative agent frameworks. See the LLawCo project page for details: this https URL\n\n---\n\n## 14. Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models\n- **Authors**: Yanchen Yin , Dongqi Han , Linghui Li\n- **URL**: [https://arxiv.org/abs/2606.28153](https://arxiv.org/abs/2606.28153)\n- **Abstract**:\n> Jailbreak attacks bypass LLM safety alignment, yet their mechanisms remain poorly understood. We provide evidence that attacks do not comprehensively eliminate safety features, but instead selectively suppress specific attention heads. We identify two functionally differentiated types: Adversarially Compromised Heads (ACHs) concentrated in early layers, which are suppressed under attacks, and Safety-Aligned Heads (SAHs) in mid-layers, which maintain robust activations even when attacks succeed. Ablation studies support the causal role of ACHs and the contribution of SAHs to robust activations: suppressing a small number of ACHs is sufficient to induce jailbreak-like behavior on normally refused inputs, while removing SAHs substantially weakens mid-layer safety activations. Token-level attribution further shows that ACH suppression is driven specifically by attack-template tokens, providing a mechanistic account of why attacks can bypass refusal decisions through ACH suppression while leaving internal safety signals sustained by SAHs -- a phenomenon we term Robust Harmful Features. To validate the practical significance of this robustness, we show that simply reading these persistent activations -- without any training -- yields competitive aggregate detection performance with strong adversarial robustness.\n\n---\n\n## 15. From Tokens to States: LLMs as a Special Case of World Models and the Continuous Path Beyond\n- **Authors**: Paul Dubois\n- **URL**: [https://arxiv.org/abs/2606.28127](https://arxiv.org/abs/2606.28127)\n- **Abstract**:\n> The AI community has framed the relationship between large language models (LLMs) and world models as a dichotomy: LLMs predict tokens; world models simulate reality. Yann LeCun argues in 2022 that reaching general intelligence requires abandoning autoregressive token prediction in favour of latent-space architectures. This framing is unnecessarily binary. Two claims will be defended. First, LLMs are a degenerate special case of world models: the state space is the set of all token sequences, the only action is appending one token, and world models are therefore a strict generalisation of LLMs, not a replacement. Second, there is a natural continuous spectrum from NTP to JEPA, with multi-token prediction, future-summary prediction, and next-latent prediction as intermediate stations already populated by current research. Moving along this spectrum relaxes the LLM constraints one by one. It also progressively surrenders the two practical advantages that make LLMs trainable at scale: internet-scale self-supervised data, and a transformer architecture co-designed for discrete token prediction. Both are examined as open research questions: the data question (the cliff from self-supervised text to instrumented action-labelled environments) and the architecture question (whether the transformer generalises to continuous-state prediction, or whether a new primitive is needed).\n\n---\n\n## 16. Single and Multi Truth Data Fusion using Large Language Models\n- **Authors**: Hira Beril Kucuk , Norman W Paton , Jiaoyan Chen , Zhenyu Wu\n- **URL**: [https://arxiv.org/abs/2606.28062](https://arxiv.org/abs/2606.28062)\n- **Abstract**:\n> Data fusion, also known as truth discovery, is a data integration problem that aims to determine the correct value or set of values for each attribute of an object when presented with potentially conflicting values from multiple sources. Data fusion tasks belong to two main categories: single-truth scenarios, where each attribute has only one correct value, and multi-truth scenarios, where multiple values can be valid simultaneously. This paper investigates the use of Large Language Models (LLMs) in data fusion tasks for tabular data. Various prompting strategies, encompassing both single-truth and multi-truth scenarios, are investigated empirically. Domain-dependent, domain-independent, zero-shot and one-shot prompts are evaluated on three different benchmark datasets. Experimental results demonstrate that LLM-based approaches outperform traditional unsupervised truth discovery methods, such as DART and LTM, across all datasets. The codebase of this study has been made publicly available on GitHub.\n\n---\n\n## 17. ToolPrivacyBench: Benchmarking Purpose-Bound Privacy in Tool-Using LLM Agents\n- **Authors**: Shijing Hu , Liang Liu , Zhu Meng , Zhicheng Zhao\n- **URL**: [https://arxiv.org/abs/2606.28061](https://arxiv.org/abs/2606.28061)\n- **Abstract**:\n> Large language models (LLMs) have increasingly moved from standalone text generation systems to agents that invoke external tools, access environments, and execute multi-step tasks. However, conventional function-calling benchmarks mainly evaluate task completion and API correctness, while privacy evaluation benchmarks typically focus on final responses or privacy judgments. Neither perspective captures purpose-bound information flow across an executed multi-tool trajectory. Motivated by this limitation in current agent evaluation, ToolPrivacyBench audits whether task-private atoms are routed only to authorized tools and downstream sinks, thereby evaluating both task completion and privacy over-disclosure during tool use. The benchmark contains 2,150 cases, including 1,150 fully synthetic privacy-sensitive business workflows and 1,000 cases adapted from existing multi-tool and function-calling benchmarks. Each case is represented by a policy knowledge base. After an agent executes against mock business backends, the evaluator compares recorded tool arguments and backend audit logs with this policy knowledge base. The evaluation covers nine widely used agents to characterize purpose-bound privacy over-disclosure. The results show that successful tool execution does not imply appropriate privacy disclosure: an agent may complete a task while transmitting unnecessary private information through intermediate tool calls. ToolPrivacyBench therefore formalizes a need-to-know disclosure boundary, under which each tool should receive only the information necessary for its stated purpose, and uses trajectory-level auditing to identify privacy over-disclosure in multi-tool workflows.\n\n---\n\n## 18. MultiHashFormer: Hash-based Generative Language Models\n- **Authors**: Huiyin Xue , Atsuki Yamaguchi , Nikolaos Aletras\n- **URL**: [https://arxiv.org/abs/2606.28057](https://arxiv.org/abs/2606.28057)\n- **Abstract**:\n> Language models (LMs) represent tokens using embedding matrices that scale linearly with the vocabulary size. To constrain the parameter footprint, prior work proposes hashing many tokens into a single vector within encoder-only models. While this offers parameter efficiency, many-to-one collisions prevent its use in causal LMs. In this paper, we propose MultiHashFormer, a new framework that allows hash-based autoregression. Each token is represented as a unique hash signature, a short sequence of discrete hash IDs, generated by multiple independent hash functions. A Hash Encoder compresses this signature into a single latent vector for processing by a Transformer decoder. Then, a Hash Decoder generates the hash signature of the next token, which is then mapped back to text. We evaluate our approach at the 100M, 1B and 3B parameter scales, demonstrating that MultiHashFormer consistently outperforms standard Transformer LMs across multiple benchmarks. Furthermore, we show that our model handles multilingual vocabulary expansion with a constant parameter footprint without any modifications.\n\n---\n\n## 19. Can LLMs Judge Better Than They Generate? Evaluating Task Asymmetry, Mechanistic Interpretability and Transferability for In-Context QA\n- **Authors**: Sambaran Bandyopadhyay\n- **URL**: [https://arxiv.org/abs/2606.28050](https://arxiv.org/abs/2606.28050)\n- **Abstract**:\n> LLM-as-a-Judge and self-evaluation pipelines implicitly assume that evaluation is easier than generation. We test this in a controlled in-context QA setting where a context passage is the sole information source and each model judges the answer it generated, removing the parametric-knowledge confound of open-domain comparisons. Across four benchmarks (SQuAD 2.0, DROP, HotpotQA, MuSiQue) and two models, evaluation is not uniformly easier: generation accuracy exceeds self-evaluation on three of four, with multi-hop MuSiQue the exception. Attention analysis reveals why: evaluation attends to context 3--5x less than generation does and barely reads the candidate answer. LoRA fine-tuning confirms the asymmetry is not a training artifact: generation fine-tuning induces over-acceptance and evaluation fine-tuning degrades generation. These findings challenge core assumptions in self-evaluation pipelines.\n\n---\n\n## 20. Dialogue to Detection: A Multimodal Hybrid NLP Pipeline for Insurance Fraud Detection\n- **Authors**: Muhammad Shakeel Akram , Amal Htait , Abdul Hamid Sadka , Emma Meisingseth , Karishma Jaitly\n- **URL**: [https://arxiv.org/abs/2606.28002](https://arxiv.org/abs/2606.28002)\n- **Abstract**:\n> Insurance fraud imposes substantial financial losses and operational inefficiencies, raising premiums and impacting trust among legitimate policyholders. Early detection at FNOL remains a persistent challenge. Existing approaches rely largely on private, text-only datasets, limiting progress on multimodal methods that integrate linguistic, behavioural, and speaker-based indicators. We introduce a synthetic multimodal framework that replicates FNOL conditions. It generates agent-customer dialogue transcripts and two-speaker audios, performs ASR and diarisation. Downstream modules combine NER, regex-based feature extraction, LLM-RAG retrieval, and speaker embeddings in a rule-based risk score to flag narrative reuse, structural inconsistencies, and cross-case voice repetition while balancing sensitivity and false positives. Dataset validation and component-level evaluations show stability and transfer potential, offering a reproducible baseline beyond text-only fraud detection.\n\n---\n\n## 21. From Black-Box to Clinical Insight: A Multi-Stage Explainable Framework for Speech-Based Cognitive Impairment Detection\n- **Authors**: Yasaman Haghbin , Sina Rashidi , Ali Zolnour , Fatemeh Taherinezhad , Ali Fartoot , Hossein Azadmaleki , James M Noble , Maryam Dadkhah , Maryam Zolnoori\n- **URL**: [https://arxiv.org/abs/2606.27973](https://arxiv.org/abs/2606.27973)\n- **Abstract**:\n> Speech-based cognitive impairment detection offers a noninvasive, accessible alternative to costly biomarker assays, yet transformer-based models remain clinically uninterpretable. We propose a multi-stage explainability framework that translates black-box transformer predictions into clinically grounded narratives by integrating SHapley Additive exPlanations (SHAP)-based token attribution, theory-informed linguistic features, and a four-stage LLM reasoning pipeline using LLaMA-3.1-70B-Instruct. Built on the SpeechCARE-Adaptive Gating Network multimodal screening model (F1 = 72.11% on the NIA PREPARE benchmark), the framework maps model outputs to four cognitive-linguistic dimensions, including lexical richness, syntactic complexity, and semantic coherence. Physician evaluation on 70 stratified English samples demonstrated strong alignment with patient-level cognitive profiles, and a System Usability Scale score of 82/100 indicated high potential for clinical workflow integration.\n\n---\n\n## 22. Two-Stage Fine-Tuning for Protein Sequence Generation with Targeted Amino-Acid Composition\n- **Authors**: Violeta Basten-Romero , Rubén Muñoz-Tafalla , Anna María Díaz-Rovira , Bertran Miquel-Oliver , Isaac Filella-Merce , Víctor Guallar\n- **URL**: [https://arxiv.org/abs/2606.27939](https://arxiv.org/abs/2606.27939)\n- **Abstract**:\n> Protein language models are standard priors for biological sequence generation, but steering them toward explicit distributional design targets remains largely unexplored. We study a constrained protein generation problem in which sequences must match a desired amino-acid (AA) composition profile while preserving plausible sequence statistics and diversity. The motivating application is synthetic feed protein design, where the AA composition of dietary proteins directly determines their nutritional value. We propose a two-stage pipeline in which domain-adaptive fine-tuning (FT) on an in-domain protein dataset is followed by iterative reward-weighted FT via reinforcement learning (RL) anchored against the FT model as a frozen reference. We evaluate the pipeline on two AA compositions and find that FT brings the average composition close to the target, while the subsequent RL enforces specific sequence constraints that FT alone cannot satisfy. We additionally evaluate the design choices of the proposed composition reward term against two baselines and an ablated variant, isolate the contribution of each training stage, and verify that AA composition alignment is achieved without degrading sequence quality.\n\n---\n\n## 23. Agentic AI-Powered Re-Identification: An Emerging, Scalable Threat to Mobility Microdata Privacy\n- **Authors**: Oscar Thees , Roman Müller , Matthias Templ\n- **URL**: [https://arxiv.org/abs/2606.27936](https://arxiv.org/abs/2606.27936)\n- **Abstract**:\n> The widespread collection of fine-grained location data by commercial data brokers creates a re-identification risk that is not widely recognised by the public. While prior research has established that mobility traces are highly unique and that individuals can, in principle, be identified from a handful of spatio-temporal points, such attacks have historically required significant manual effort from skilled analysts, limiting their practical scale. In this feasibility study, we demonstrate in a real world setting that agentic AI fundamentally changes this threat model. We present an end-to-end pipeline in which large language model agents autonomously search the open web, cross-reference public records and social media, and resolve raw coordinate sequences to candidate identities - without human intervention. We evaluate the pipeline on a spatio-temporal dataset containing simulated location points anchored at and around true home and work addresses, focusing on a high-risk disclosure scenario. Our results demonstrate that, from spatio-temporal data and public sources alone, our agentic AI successfully re-identified 18 of the 25 re-identifiable individuals (72%) and 18 of 43 cases overall (41.9%). We discuss implications for Statistical Disclosure Control (SDC) practice and outline the near-future escalation that data custodians and regulators must anticipate. De facto anonymity - an implicit foundation of SDC practice - is shifting. Agentic AI strengthens the case that re-identification is reasonably likely by any means under the GDPR Recital-26 standard, at costs of minutes-and-dollars per target.\n\n---\n\n## 24. Triadic Werewolf: A Jester Role for Multi-Hop Theory of Mind in LLMs\n- **Authors**: Avni Mittal\n- **URL**: [https://arxiv.org/abs/2606.27909](https://arxiv.org/abs/2606.27909)\n- **Abstract**:\n> Theory-of-mind evaluations of large language models typically use dyadic social-deduction games, where every observable cue points to a single hidden side, so a model with strong language priors can score well without ever simulating opponents' incentives. We extend the Werewolf game with a Jester, a third faction whose utility on peer suspicion is inverted because it wins by being voted out, so optimal play requires reasoning across three opposing utility functions. Across 60 games on GPT-4.1, DeepSeek-V3.1, and Llama-3.3-70B with Jester self-learning on and off, the Jester wins 60-70% of games while Werewolves never exceed 20%, and GPT-4.1 wolves vote the Jester out on day 1 in 60-70% of games, a strictly self-defeating action. Self-learning helps DeepSeek and Llama but hurts GPT-4.1, with the cost landing on Villagers rather than Werewolves. Only DeepSeek learns the subtle strategy of looking suspicious without looking intentionally suspicious, and it gains the most from the loop. Triadic incentive structure exposes a layer of multi-agent reasoning that dyadic deduction games leave invisible.\n\n---\n\n## 25. A Study of Temporal Fusion Strategies for Named Entity Recognition in Historical Texts\n- **Authors**: Emanuela Boros\n- **URL**: [https://arxiv.org/abs/2606.27881](https://arxiv.org/abs/2606.27881)\n- **Abstract**:\n> Temporal variation poses a unique challenge for named entity recognition (NER) in historical texts, where entities drift in surface form and salience across time. While language models (LMs) have made progress in various NLP tasks, their ability to reason about temporality, especially in diachronic contexts, remains limited or at least, questionable. In this paper, we systematically study how temporal metadata can be structurally embedded into NER models using a range of lightweight fusion strategies. We experiment with both absolute and relative temporal representations, injected into Transformer-based architectures via early or late fusion mechanisms such as cross-attention, adapters, and concatenation. Our evaluations on French and German historical datasets reveal that late fusion strategies yield more robust and temporally generalisable performance, particularly in early and noisy periods.\n\n---\n\n## 26. SpatialUAV: Benchmarking Spatial Intelligence for Low-Altitude UAV Perception, Collaboration, and Motion\n- **Authors**: Haoyu Zhang , Meng Liu , Qianlong Xiang , Kun Wang , Yaowei Wang , Liqiang Nie\n- **URL**: [https://arxiv.org/abs/2606.27876](https://arxiv.org/abs/2606.27876)\n- **Abstract**:\n> Spatial intelligence is essential for low-altitude unmanned aerial vehicle (UAV) perception, collaboration, and navigation. However, existing UAV benchmarks often emphasize image-level recognition, single-view understanding, or narrow answer formats, leaving 3D spatial inference, multi-view collaboration, scene dynamics, and diverse task formulations insufficiently evaluated. To address these gaps, we introduce SpatialUAV, a real low-altitude UAV benchmark comprising 4,331 curated instances across 14 fine-grained task types, covering semantic discrimination, spatial relation, aerial--aerial collaboration, aerial--ground collaboration, and motion understanding. SpatialUAV organizes all samples into a unified visual-input--question--answer schema, while supporting seven input configurations and nine answer formats, including option labels, region identifiers, geometric values, cross-view correspondences, and free-form motion descriptions. To ensure reliable and grounded evaluation, our data construction pipeline integrates detector-assisted regions, depth supervision, metadata-derived rules, extensive manual annotation, blind filtering, and multi-turn human validation, together with task-specific metrics for heterogeneous outputs. Evaluating representative vision-language models across three categories, we show that current models remain far from human-level performance, with pronounced bottlenecks in cross-view association, structured grounding, geometric reasoning, and temporal viewpoint understanding. These results offer empirical guidance for advancing low-altitude UAV spatial intelligence. Code and data are available at this https URL .\n\n---\n\n## 27. Position Bias Correction is Insufficient for One-Pass Attention Sorting\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27793](https://arxiv.org/abs/2606.27793)\n- **Abstract**:\n> Long-context language models suffer from position bias, where information in middle positions is underutilized. Attention Sorting addresses this by iteratively reordering documents based on attention patterns, but its multiple sort-and-generate cycles increase deployment cost. We hypothesize that position bias is the primary bottleneck and propose Debiased One-Pass Attention Sorting, which estimates a per-prompt position-bias curve from the low-attention majority of documents and uses it to correct raw attention scores (via subtraction or division) to enable single-pass sorting. Our experiments on two models refute this hypothesis in the tested setting: on LLaMA-2-7B-32K-Instruct, debiasing produces identical results to uncalibrated single-pass sorting (94.83\\% containment accuracy), while on YaRN-Llama-2-7b-64k, debiasing improves accuracy by 8.67 percentage points but remains 14.84pp behind iterative sorting, closing only 37\\% of the gap. These results suggest that position-bias correction is insufficient to match iterative sorting, and that repeated reordering provides additional benefits beyond bias correction.\n\n---\n\n## 28. NLL-Guided Full-Attention Layer Selection for Training-Free Sliding-Window Adaptation\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27791](https://arxiv.org/abs/2606.27791)\n- **Abstract**:\n> Hybrid attention models that mix full and sliding-window attention across layers offer a promising approach to efficient long-context inference, but the critical question of \\emph{which layers} should retain full attention remains unsolved. Existing methods use either fixed periodic patterns or attention-based heuristics that may not capture what matters for downstream accuracy. We propose NLL-guided layer selection, a training-free method that directly measures each layer's importance by computing the negative log-likelihood degradation on answer tokens when that layer uses sliding-window instead of full attention. On LongMemEval with Qwen3-4B, our method achieves 64.6\\% accuracy using only 1/4 full-attention layers, matching the 1/2-FA periodic baseline (65.0\\%) while halving the computational budget. NLL-guided selection outperforms the SWAA-reported periodic 1/4-FA baseline by 10.4 percentage points and a matched LightTransfer-style baseline by 26.4 percentage points. De-confounding analysis shows the signal is consistent with long-range attention needs rather than generic layer sensitivity. The method requires only $\\sim$15 minutes of one-time calibration, advancing the efficiency-accuracy Pareto frontier for long-context LLM deployment.\n\n---\n\n## 29. Output-Space Allocation Costs for Calibration-Guided LLM Compression: An Empirical Study\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27785](https://arxiv.org/abs/2606.27785)\n- **Abstract**:\n> Training-free compression methods for large language models (LLMs) often use calibration data to guide compression decisions. ROCKET, a recent method combining sparse-dictionary factorization with multi-choice knapsack problem (MCKP) allocation, derives its per-layer factorization from an output reconstruction objective but uses weight-space Frobenius error as the MCKP allocation cost. We investigate whether aligning the allocation cost with the output-space objective improves compressed model fidelity. On Qwen3-8B at 50\\% compression, our ROCKET-ActCost achieves +0.8 percentage points higher average accuracy across 8 zero-shot benchmarks (53.1\\% vs 52.3\\%), but increases WikiText perplexity by 16\\% (61.46 vs 52.98). This accuracy-perplexity tradeoff reveals that different allocation objectives favor different downstream metrics. The high correlation ($>$0.99) between weight-space and output-space errors limits allocation divergence, explaining the modest effect size. On Llama-3.2-1B at 20\\% compression, the two methods produce near-identical results (53.3\\% vs 53.5\\% accuracy, 14.45 vs 14.66 PPL), suggesting that the effect of the cost function is minor at lower compression ratios.\n\n---\n\n## 30. Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?\n- **Authors**: Guoheng Sun , Kaixi Feng , Shwai He , Xiaochuan Gong , Yexiao He , Ziyao Wang , Zheyu Shen , Wanghao Ye , Ramana Rao Kompella , Gaowen Liu , Ang Li\n- **URL**: [https://arxiv.org/abs/2606.27755](https://arxiv.org/abs/2606.27755)\n- **Abstract**:\n> Vision-Language-Action (VLA) models enable instruction-driven robotic manipulation, but they inherit oversized language backbones from pretrained VLMs whose capacity far exceeds what is needed for short robotic instructions. This raises a basic question: how much of a VLA model is actually necessary for closed-loop control? In this work, we study architectural redundancy in VLA models by using transformer block removal as a controlled intervention. We introduce \\textbf{Drop-Then-Recovery (DTR)}, an analysis protocol that removes selected blocks from a pretrained VLA model and then fine-tunes the resulting model to measure whether the removed capacity was necessary for downstream control. To make this intervention reliable, we propose \\textbf{GateProbe}, a one-shot virtual-gate sensitivity metric that ranks blocks by their contribution to the downstream action loss. Across multiple VLA architectures, manipulation benchmarks and even real-robot industrial scenarios, we find a strong asymmetry in post-removal recoverability: \\ul{\\textit{language backbones are highly redundant for standard robotic manipulation tasks, whereas vision and action pathways are substantially less tolerant to removal} }. On LIBERO, removing half of the LLM blocks even improves OpenVLA-OFT from 95.0% to 98.3% under the same downstream fine-tuning budget, and retaining only two language blocks still recovers baseline-level performance. These results suggest that current VLA benchmarks may exert limited pressure on deep language grounding and compositional instruction understanding, and that future VLA architectures should allocate capacity more deliberately across language, vision, and action components. The code is available at this https URL .\n\n---\n\n## 31. End-to-End Dynamic Sparsity for Resource-Adaptive LLM Inference\n- **Authors**: Yuhang Chen , Jinhao Duan , Ruichen Zhang , Mingfu Liang , Xiaohan Wei , Yunchen Pu , Fei Tian , Chonglin Sun , Parish Aggarwal , Frank Shyu , Luke Simon , Sandeep Pandey , Tianlong Chen , Xi Liu\n- **URL**: [https://arxiv.org/abs/2606.27743](https://arxiv.org/abs/2606.27743)\n- **Abstract**:\n> Large Language Models (LLMs) inference is typically deployed under a static resource assumption, where models execute a fixed computational graph regardless of the runtime environment. However, real-world cloud infrastructure is inherently dynamic, characterized by fluctuating availability (e.g., spot instance preemption) and tiered Quality-of-Service requirements. In such volatile settings, static models are inflexible: they either crash under resource constraints or waste compute on redundant operations. To bridge this gap, we propose Learning to Allocate (L2A), an end-to-end framework for resource-adaptive inference. Unlike prior methods that condition only on input difficulty, we formulate inference as a constrained allocation problem conditioned on both the input and the runtime resource budget itself. We introduce lightweight, budget-conditioned and input-aware gating networks integrated into the LLM. These gates are trained via a unified objective that jointly optimizes task performance, logical consistency, and resource costs along three axes matching how real-world dynamics manifest: layer skipping for memory and depth pressure, head pruning for throughput contention, and reasoning-token reduction for latency tightening. This lets the model learn a budget-aware policy beyond input difficulty alone: it adaptively configures its computational footprint with respect to real-time resource dynamics, maximizing reasoning depth when resources permit while enforcing strict frugality when budgets tighten. A single L2A model traces the entire compute-accuracy Pareto frontier on Llama-3-8B and Qwen-3-4B: at up to 34% realized layer sparsity, it stays within 0.6% of the dense baseline on GSM8K, with the same gap holding zero-shot on out-of-distribution tasks, while every static or heuristic baseline requires a separately tuned model and still drops by 5-10% at comparable inference time.\n\n---\n\n## 32. KG2Cypher: Data-Centric Pipeline for Building Enterprise Text-to-Cypher Systems\n- **Authors**: Minjun Choi , Yerin Kim , Junghyuk Seo , Sujin Mo , Hyemin Lee , Youngjoong Ko\n- **URL**: [https://arxiv.org/abs/2606.27742](https://arxiv.org/abs/2606.27742)\n- **Abstract**:\n> Enterprise Knowledge Graphs (KGs) are increasingly used for internal search, analytics, and question answering, but building natural-language interfaces for private enterprise graphs remains costly. We present KG2Cypher, a data-centric pipeline for building enterprise text-to-Cypher systems from existing KGs. KG2Cypher first constructs an executable Cypher query from observed graph facts and then uses LLMs to generate its associated natural-language question. The resulting Text-Cypher pairs are validated with an LLM judge and human validation, and are converted into candidate-aware SFT data. The trained generator is served with class-conditioned schema prompting, entity retrieval, and LoRA-based inference. We evaluate KG2Cypher in Korean enterprise settings, where short search-style queries and schema paraphrases make language grounding difficult. LoRA SFT improves execution-result F1 from 0.806 to 0.950 on broadcast-program queries and from 0.70 to 0.92 on company queries. In an 11-class setting, KG2Cypher achieves 95.2% exact match, 99.9% execution rate, and 0.964 execution-result F1.\n\n---\n\n## 33. Bifocal Diffusion Language Models: Asymmetric Bidirectional Context for Parallel Generation\n- **Authors**: Yuhang Chen , Xianfeng Wu , Jinhao Duan , Mingfu Liang , Xiaohan Wei , Yunchen Pu , Fei Tian , Chonglin Sun , Parish Aggarwal , Frank Shyu , Luke Simon , Sandeep Pandey , Xi Liu , Tianlong Chen\n- **URL**: [https://arxiv.org/abs/2606.27732](https://arxiv.org/abs/2606.27732)\n- **Abstract**:\n> Discrete diffusion language models (dLLMs) recover masked tokens in parallel, offering significant speedups over autoregressive (AR) generation. However, such promising frameworks face a fundamental architectural design dilemma: \\ding{182} Adopting bidirectional attention achieves strong generation quality by allowing each position to access the full context, but is inherently incompatible with KV caching, limiting inference throughput in batch-serving scenarios; \\ding{183} Conversely, causal attention enables efficient cached inference but loses all right-side context, substantially degrading generation quality. This paper introduces Bifocal dLLMs, a new paradigm that resolves this dilemma through \\emph{asymmetric bidirectional context}. Analogous to bifocal lenses, we instantiate the paradigm as \\textbf{R2LM} (Right-to-Left Mamba), which combines two complementary mechanisms: $a$) standard causal attention providing precise left-context with full KV cache compatibility, while $b$) a lightweight reverse Mamba SSM sidecar supplying compressed right-side context without breaking cacheability. Comprehensive experiments on continued pretraining of Qwen3-1.7B with 60B tokens demonstrate that R2LM achieves $2.4\\times$ to $12.9\\times$ higher throughput than bidirectional dLLMs and $1.9\\times$ to $2.9\\times$ speedup over AR baselines in batch serving through parallel decoding with KV caching, while exceeding the causal baseline on most benchmarks and surpassing the bidirectional dLLM on average.\n\n---\n\n## 34. Enhancing Numerical Prediction in LLMs via Smooth MMD Alignment\n- **Authors**: Zhuo Zuo , Li Yue , Wenhao Zheng , Chenpeng Wang , Xianggen Liu\n- **URL**: [https://arxiv.org/abs/2606.27731](https://arxiv.org/abs/2606.27731)\n- **Abstract**:\n> Despite their strong general capabilities, large language models (LLMs) often remain unreliable when outputs must be numerically precise. A key reason is the training objective: standard cross-entropy treats numeric tokens as unstructured categories and ignores the metric structure of their values. We address this mismatch with Smooth Maximum Mean Discrepancy (SMMD), which builds on the classic MMD by incorporating value-distance kernels over numeric tokens and graph-based smoothness. With this kernel defined over a numeric sub-vocabulary, SMMD aligns the predicted numeric distribution to the target via kernel matching and smooths the prediction-target residual over the induced kernel graph to encourage local consistency. We evaluate SMMD on four numeric-target tasks: mathematical reasoning, arithmetic calculation, clock-time recognition, and chart question answering, across multiple open-weight LLM and VLM backbones. SMMD consistently improves accuracy over both cross-entropy and recent numeric-target losses; analyses show complementary effects between MMD and smoothness and underscore the importance of distance-based kernel design. Code is available at this https URL .\n\n---\n\n## 35. Low-Agreeableness Persona Conditioning for Safe LLM Fine-Tuning\n- **Authors**: Austin MY Cheung , Yi Yang\n- **URL**: [https://arxiv.org/abs/2606.27709](https://arxiv.org/abs/2606.27709)\n- **Abstract**:\n> Recent work has shown that fine-tuning large language models (LLMs) for social warmth degrades factual reliability and increases sycophancy. We investigate a related but distinct failure mode: warmth fine-tuning also weakens adversarial safety, making models more susceptible to jailbreaks and harmful output generation. We examine whether this reflects an inherent consequence of empathetic adaptation or an artifact of data construction. To address this, we introduce a persona-driven rewriting pipeline that conditions user turns on low agreeableness and pairs this with warm, de-escalating assistant responses. Across three experiments on four models, our approach reduces jailbreak susceptibility and harmful output rates relative to generic warmth fine-tuning baselines, while preserving conversational warmth. Representational probing provides suggestive evidence that this conditioning reduces the geometric alignment between warmth and compliance directions in latent space. These results show that safer empathetic fine-tuning is achievable through data design alone, without safety labels, harm detectors, or changes to the training objective.\n\n---\n\n## 36. Mitigating LLM-based p-Hacking by Preregistering for the Next LLM\n- **Authors**: Maria Thomas , Kristina Gligoric , Nihar B. Shah\n- **URL**: [https://arxiv.org/abs/2606.27687](https://arxiv.org/abs/2606.27687)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used to generate, classify, and annotate data whose outputs feed downstream hypothesis tests. However, LLM-based research is easy to p-hack: a researcher can tune the prompts, decoding parameters, or output format until a desired result is reached. We propose a protocol to mitigate p-hacking in LLM-based research: preregistering the experiment and eligible models, and then running it on the first eligible LLM that is released after the preregistration. The researcher finalizes the procedure on current models, preregisters the analysis plan together with a set of eligible future models, and runs the confirmatory analysis on the first eligible model released afterward. Because this model does not exist at commitment time, it cannot be hacked against; furthermore, configurations that hack one model frequently do not transfer to the next. We evaluate the protocol on two tasks whose true values are known. Across 20 models from four providers and 11 LLM-analysis configurations, the protocol would have blocked successful transfer of the p-hack in 73.9% and 72.7% of cases in the two tasks. Additional analyses reveal that mitigation remains substantial under several stress tests. Finally, putting money where our mouth is, we followed our own protocol and preregistered our experiment. The preregistered experiment confirmed the protocol's effectiveness: out of the 7 configurations that hacked the prior model, the hacking failed to carry over in 6 configurations on the first eligible model released afterward.\n\n---\n\n## 37. CBD: API-Only LLM Black-Box Unlearning through Controlled Behavioral Divergence\n- **Authors**: Zhiqiang Xie , Yijing Lin , Zhipeng Gao , Dong In Kim\n- **URL**: [https://arxiv.org/abs/2606.27683](https://arxiv.org/abs/2606.27683)\n- **Abstract**:\n> Edge devices increasingly invoke large language models (LLMs) through API services for context aware edge intelligence, while edge generated data may be collected to improve LLMs and may introduce sensitive, copyrighted, harmful, or outdated information into model behavior. Machine unlearning offers a practical way to remove the influence of undesired data without retraining LLMs. However, existing methods still face two gaps. The first is API only black box access, where target model parameters and internal logits are unavailable. The second is how to preserve retained utility when unlearning target data and retained data share highly similar prompt structures or semantic patterns. To address these challenges, we propose Controlled Behavioral Divergence (CBD), an API only black box unlearning framework. CBD uses two auxiliary models to create controlled behavioral divergence between retained inputs and unlearning target inputs, converts this divergence into an unlearning relevance score, and routes unlearning related prompts away from the target LLM. To improve discrimination accuracy under high similarity between target and retained data, CBD constructs a gradient statistics based discriminative basis by estimating empirical Fisher matrices and solving a regularized generalized eigenvalue problem, guiding the unlearning signal toward target specific information rather than shared prompt structures. Compared with eleven white box and gray box unlearning baselines, CBD achieves a better unlearning utility trade off and its performance varies little across settings. On ToFU forget10, CBD approaches the retrained reference on the forget set while raising model utility to 74.90, about 15% above the second best baseline. On WMDP, it lowers hazardous knowledge accuracy to 25.68, near random guessing, while preserving MMLU accuracy of 52.67. Code is at this https URL .\n\n---\n\n## 38. From Signals to Transfer: A Factorised Study of Probe-Based Uncertainty Estimation in Large Language Models\n- **Authors**: Ponhvoan Srey , Xiaobao Wu , Cong-Duy Nguyen , Quang Minh Nguyen , Duc Anh Vu , Anh Tuan Luu\n- **URL**: [https://arxiv.org/abs/2606.27679](https://arxiv.org/abs/2606.27679)\n- **Abstract**:\n> Probe-based uncertainty estimation (UE) has emerged as a prominent approach to detect hallucinations in Large Language Models (LLMs) by learning uncertainty from internal model signals. Yet, recent methods vary simultaneously across feature design, training data construction, and evaluation setting, obscuring what actually drives performance. To address this issue, we propose a factorised study of probe-based UE under matched conditions. Our results show that raw hidden states and attention features are difficult to outperform in-domain. However, under distribution shift, structured and compressed features are more robust, suggesting that in-domain performance alone is insufficient to measure progress. Furthermore, prompting and label construction significantly affect probe behaviour. Building on these best-practice findings, we train benchmark-based pretrained probes that transfer reasonably well to open-ended factual generation, providing a stable off-the-shelf baseline. Our work encourages more deployment-oriented evaluation of probe-based uncertainty estimators. The code repository is available at this https URL .\n\n---\n\n## 39. HybridCodec: Modeling Discrete and Continuous Representations for Efficient Speech Language Models\n- **Authors**: Artem Ploujnikov , Francesco Verdini , Samir Sadok , Mirco Ravanelli\n- **URL**: [https://arxiv.org/abs/2606.27627](https://arxiv.org/abs/2606.27627)\n- **Abstract**:\n> Discrete audio representations have become increasingly popular for building multimodal text-audio systems and integrating audio capabilities into Large Language Models (LLMs). However, numerous studies report performance degradation on various downstream tasks due to information loss during discretization. To address this, we propose a novel approach combining temporally compressed discrete tokens with dimensionality-reduced continuous residuals. Our framework consists of a hybridized discrete-continuous focal modulation codec and a hybrid Transformer. This architecture performs autoregressive inference in the discrete domain, coupled with non-autoregressive prediction and continuous residual upsampling. Experimental results show that our approach significantly improves the retention of speaker characteristics compared to discrete-only methods, while simultaneously reducing the number of required autoregressive steps.\n\n---\n\n## 40. Dismantling Pathological Shortcuts: A Causal Framework for Faithful LVLM Decoding\n- **Authors**: Liu Yu , Can Chen , Ping Kuang , Zhikun Feng , Fan Zhou , Gillian Dobbie\n- **URL**: [https://arxiv.org/abs/2606.27596](https://arxiv.org/abs/2606.27596)\n- **Abstract**:\n> Large Vision-Language Models (LVLMs) exhibit sophisticated reasoning but remain susceptible to object hallucination. Deviating from the prevailing attention intensity assumption, we reveal a deeper dynamic structural misalignment: hallucination is triggered at decision-critical steps where specific attention heads, acting as risky mediators, decouple from visual evidence to lock onto language priors. This establishes a pathological shortcut that bypasses visual grounding. To dismantle this, we propose Fox (Faithfulness and Observational-flow via eXpression-rectification), a training-free inference-time framework. Fox diagnoses structural misalignment using a visual attention entropy probe to localize risky mediators unsupervisedly. We then execute a targeted causal intervention via numerical logit saturation to physically sever the shortcut path. Finally, a conflict-gated cooperative decoding strategy reconciles interventional faithfulness with observational fluency. Extensive experiments demonstrate that Fox achieves SOTA performance, outperforming SID by 29.1% while preserving linguistic richness. Code is available at this https URL .\n\n---\n\n## 41. On the Inseparability of Instructions and Data in Shared-Embedding Sequence Models\n- **Authors**: Dewank Pant , Shruti Lohani , Avijit Kumar\n- **URL**: [https://arxiv.org/abs/2606.27567](https://arxiv.org/abs/2606.27567)\n- **Abstract**:\n> Prompt injection is the top security risk for LLM-integrated applications, yet every defense proposed so far has been broken. We prove this is not a coincidence: in shared-embedding architectures that lack enforced control-data separation, perfect prompt-injection prevention is mathematically impossible. We formalize prompted systems as Prompted Action Models whose outputs include control-authoritative actions: refusal decisions, tool authorization, policy routing, and memory writes. We define Semantic-Faithful Control (SFC), the property that such behavior depends only on the meaning of untrusted input, not on how it is encoded. We then prove SFC is unachievable within the shared pipeline, via three results: a provenance-recovery impossibility (shared representations make trusted and untrusted content statistically inseparable, bounded by total variation distance); control-path exposure (untrusted tokens enter control-relevant computation through the same attention value-aggregation that determines outputs); and a finite-coverage invariance gap (finite training cannot certify invariance over infinite semantic-equivalence classes). We ground each quantity in measurements on production tokenizers and models. The result is structural, not a gap in current defenses. It mirrors the code-data confusion in Von Neumann machines that gives rise to buffer overflows, a vulnerability class that took decades of layered defenses (DEP, Write-XOR-Execute, ASLR, stack canaries, and ultimately memory-safe languages) to contain, because no single mechanism sufficed. The implication is the same: prompt injection cannot be eliminated by better in-pipeline classification or alignment alone. It requires architectural separation of instruction and data channels. We identify the root cause and the class of solution it demands.\n\n---\n\n## 42. Benchmarking Multi-Modal Graph-based Social Media Popularity Prediction\n- **Authors**: Utkarsh Sahu , Zhisheng Qi , Li Zhu , Yizhao Yang , Jun Li , Ryan Rossi , Yu Wang\n- **URL**: [https://arxiv.org/abs/2606.27539](https://arxiv.org/abs/2606.27539)\n- **Abstract**:\n> Social media popularity prediction aims to forecast the future reach or influence of online content from early-stage observations. Accurate prediction enables key downstream applications, such as advertising optimization and strategic content planning by users, creators, and platforms. Despite substantial progress, existing popularity prediction works often fail to jointly consider multimodal content and temporal social interaction signals. Moreover, the literature remains highly fragmented across datasets, modalities, observation windows, prediction targets, and evaluation protocols. This fragmentation prevents fair comparison and obscures a systematic understanding of how textual, visual, temporal, and interaction-based signals jointly shape popularity dynamics. To address these challenges, we introduce MMG-Pop, a Multi-modal Graph-based Popularity Prediction benchmark, which unifies datasets, modalities, temporal interaction signals, and representative baselines under a standardized evaluation protocol. Furthermore, we propose MMG-PopNet, a unified multi-modal graph-based network that jointly models the aforementioned multi-modal signals and graph-structured social interactions. Extensive experiments on MMG-Pop, comprising four datasets across Bluesky and Reddit platforms, demonstrate the superior performance of MMG-PopNet and yield new insights into cross-platform training generalization, multi-task prediction benefits, multi-modality contributions, and LLM prediction limitation. These findings establish a unified foundation for future research on social dynamics modeling and intervention under heterogeneous modalities and socially-aware agentic ecosystem paradigms.\n\n---\n\n## 43. Large Language Model Teaches Visual Students: Cross-Modality Transfer of Fine-Grained Conceptual Knowledge\n- **Authors**: Thomas Shih-Chao Liang , Zhuoran Yu , Yong Jae Lee\n- **URL**: [https://arxiv.org/abs/2606.27527](https://arxiv.org/abs/2606.27527)\n- **Abstract**:\n> Large Language Models (LLMs) possess broad conceptual knowledge acquired through large-scale text pretraining, yet their potential to supervise models in other modalities remains underexplored. In this work, we propose LaViD--Language-to-Visual Knowledge Distillation--a simple and effective framework for transferring high-level semantic knowledge from a language-only teacher to a vision-only student model. Instead of relying on paired multimodal data, LaViD elicits conceptual signals from an LLM by prompting it to generate multiple-choice questions (MCQs) that probe semantic distinctions between visual classes. Each class is mapped to a soft label distribution over these MCQs, forming a rich conceptual signature that guides the student through an auxiliary distillation loss. Notably, despite using a language-only teacher without access to image data, LaViD consistently outperforms recent methods like MaKD that distill from vision-language models across multiple fine-grained benchmarks. It also achieves competitive or superior performance compared to state-of-the-art visual distillation methods such as DKD and MLKD, with further gains when combined with logit standardization. On the Waterbirds dataset, LaViD substantially improves worst-group accuracy, demonstrating enhanced robustness to spurious correlations with distillation. Code is available at this https URL .\n\n---\n\n## 44. Speculative Refinement: A Hybrid Autoregressive Diffusion Decoding Strategy and Its Behavior Across Benchmarks\n- **Authors**: Aditi Gupta , Neel Mishra , Kushagra Trivedi , Pawan Kumar\n- **URL**: [https://arxiv.org/abs/2606.27474](https://arxiv.org/abs/2606.27474)\n- **Abstract**:\n> How should we evaluate generation systems that combine autoregressive (AR) and diffusion decoding? We study this question through Speculative Refinement (SpecRef), a training-free hybrid method that warm-starts a masked diffusion language model from an AR draft using entropy-guided selective masking. Evaluating SpecRef across six benchmarks (HumanEval, MBPP, GSM8K, BBH, ARC-Challenge, HellaSwag) with three distinct evaluation protocols (execution-based pass@1, exact-match, log-likelihood scoring), we surface several findings relevant beyond our specific system: (1) code benchmarks conflate structural discovery with logical correctness: providing a syntactic scaffold lifts accuracy from near zero to over 20% without changing the model, indicating that much of the baseline failure is structural; (2) a refinement tension phenomenon where multi-stage correction degrades already-correct tokens, exposing benchmark saturation ceilings invisible to single-model evaluation; (3) log-likelihood and generative evaluation produce different model rankings for the same model pair, suggesting they measure different capabilities; (4) standard Python post-processing silently breaks code evaluation for non-AR generators. These observations apply to any multi-stage or non-autoregressive generation pipeline and point toward more diagnostic evaluation practices.\n\n---\n\n## 45. Supersede: Diagnosing and Training the Memory-Update Gap in LLM Agents\n- **Authors**: Vedant Patel\n- **URL**: [https://arxiv.org/abs/2606.27472](https://arxiv.org/abs/2606.27472)\n- **Abstract**:\n> Large language model (LLM) agents operate over long, multi-session interactions in which facts change: a user moves, a price updates, a plan is revised. Acting correctly requires using the current value of a fact and discarding values that have been superseded. We isolate this ability on real conversational data and show that it is a distinct, unsolved failure. On the knowledge-update subset of LongMemEval, replacing an agent's full context with a bounded, self-maintained memory drops accuracy from 92% to 77% even on a frontier model (gpt-5.4), a gap that is statistically significant (paired McNemar p<0.005) and persists across model scale while full-context accuracy saturates near 92%. The bottleneck is therefore memory maintenance, not comprehension, and is not closed by a stronger model. We then ask whether this is merely an undersized memory, and find it is not: as the conversation grows 24x, accuracy falls further (from 68% to 28%), and granting the agent proportionally more memory yields no detectable recovery (28% to 28%, n=25). The failure scales with the length of the conversation, not the compression ratio. We release Supersede, an open reinforcement-learning environment (on the verifiers / prime-rl stack) that turns this measurement into a training signal: agents are rewarded for answering from the current value and penalized for stale ones. Finally, we close the loop and show the gap is trainable: GRPO fine-tuning a small open model (Qwen2.5-3B) on this environment nearly doubles its held-out supersession accuracy on real, unseen conversations (9.0% to 16.7%, a single run), along a monotonic checkpoint curve indicating the learned policy, not the harness, carries the gain. To our knowledge this is the first trainable environment whose reward targets temporal fact-currency, and the first evidence the supersession gap can be trained down, not only measured.\n\n---\n\n## 46. SidConArena: An Environment Evaluating Agents in Open-Ended,Positive-Sum Bargaining Game\n- **Authors**: Yeqi Feng , Yuxin Chen , Tianxing He\n- **URL**: [https://arxiv.org/abs/2606.27397](https://arxiv.org/abs/2606.27397)\n- **Abstract**:\n> Evaluating LLM agents requires dynamic environments that go beyond static reasoning and zero-sum games. Real-world economic interaction is often open-ended and mixed-motive: agents must negotiate, create positive-sum surplus, compete for scarce assets, and plan under delayed returns. We introduce SidConArena, a new benchmark framework for evaluating LLM agents in open-ended, positive-sum bargaining. SidConArena formalizes a multi-player economy as a finite-horizon partially observable stochastic game with three coupled phases: natural-language negotiation with binding trades, deterministic converter-based production, and sealed-bid auctions for long-term assets. The framework combines structured observations, phase-aware agent dispatching, a neural-symbolic action interface, and asynchronous execution, enabling free-form interaction while preserving rule-grounded evaluation. Across homogeneous and heterogeneous tournaments, stronger frontier models achieve higher economic outcomes, yet agents still misvalue resources, bargain passively, and remain limited in long-horizon investment planning.\n\n---\n\n## 47. Agentic Publication Protocol: An Attempt to Modernize Scientific Publication\n- **Authors**: Sirui Lu , Xiao-Liang Qi\n- **URL**: [https://arxiv.org/abs/2606.27386](https://arxiv.org/abs/2606.27386)\n- **Abstract**:\n> Scientific publication is still organized primarily around static manuscripts, even though much of scientific progress depends on tacit know-how: how to run code, reproduce figures, interpret edge cases, choose useful follow-up directions, and avoid failed paths. Large language model agents create an opportunity to publish not only knowledge, but also operational know-how in a form that future readers and researchers can directly use. This paper outlines the Agentic Publication Protocol (APP), a lightweight repository format for packaging a paper together with code, data, environment information, reproducibility instructions, and an agent-facing instruction file. APP treats a version-controlled repository as the publication object and uses \\texttt{ this http URL } and optional skills to define a paper agent that can explain the work, reproduce key results when possible, and support follow-up research. We describe the design principles and details of the protocol, as well as the agent skills useful for publishing papers under the protocol. We also describe development tools for evaluating and improving the protocol and associated agent skills. Finally, we provide a broader discussion of the future of scientific research in the agent era.\n\n---\n\n## 48. CalBrief: A Pilot Diagnostic Benchmark for Evidence-Calibrated Scientific Briefing with Large Language Models\n- **Authors**: Yu Fu , Yongqi Kang , Yong Zhao\n- **URL**: [https://arxiv.org/abs/2606.27383](https://arxiv.org/abs/2606.27383)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used as research assistants, yet it remains unclear whether they can calibrate research takeaways to the strength and scope of the supporting evidence. We study evidence-calibrated scientific briefing: given a bounded package of related papers, a system should generate package-level takeaways with evidence strength, scope boundaries, and missing-evidence caveats. We contribute a verified pilot benchmark of 16 heterogeneous scientific evidence packages and 96 human-verified takeaways, and we use CalBrief, an auditable role/gap/strength framework, as a diagnostic probe to locate where briefing breaks down. Under a fair-schema evaluation, structured organization improves role and gap reasoning, but an explicit strength-calibration policy is systematically over-conservative and falls below majority and direct-LLM baselines. To explain why, we run a controlled diagnostic across three closed-model backbones (GPT-4o, Claude Sonnet, Gemini Flash) that separates three potential causes of conservatism. Approximately 63% of the conservatism gap is attributable to expanding the label space from binary {moderate, weak} to four-way {moderate, weak, uncertain, insufficient_evidence} (p < 0.001 across all backbones); only 1% is attributable to gap/scope signal injection (not significant); the remaining 36% arises from the pipeline policy itself. We also find that 4-way predictions can be post-hoc collapsed back to binary and then match or exceed direct binary prompting, so the extra labels carry information that strict matching hides. Label-level strength judgment and auditable evidence organization are distinct abilities currently in tension, and should be evaluated separately for LLM research assistants.\n\n---\n\n## 49. Position: The Term \"Machine Unlearning\" Is Overused in LLMs\n- **Authors**: Sangyeon Yoon , Yeachan Jun , Albert No\n- **URL**: [https://arxiv.org/abs/2606.27379](https://arxiv.org/abs/2606.27379)\n- **Abstract**:\n> Large language models increasingly face demands to \"forget\" training data, knowledge, or behaviors due to regulatory deletion obligations, copyright/licensing disputes, and safety or product-policy requirements. This position paper argues that machine unlearning is overused as a term in LLM research and should be reserved for dataset-defined deletion: removing the training influence of a precisely specified forget set such that the resulting model is approximately indistinguishable from retraining without that data. We contend that many tasks currently labeled \"unlearning\" (e.g., refusal for harmful requests, entity/knowledge removal, or targeted suppression) pursue different, often policy-dependent objectives and therefore require different terminology and baselines (e.g., alignment, suppression, editing, obfuscation). We further argue that this confusion is not cosmetic: because papers make different implicit guarantees under the same label, metrics and benchmarks are frequently reused outside their intended scope, rewarding surface-level non-disclosure (e.g., low ROUGE/forget accuracy) even when retraining-equivalence is not tested and derived capabilities remain. We conclude by calling for stricter terminology tied to explicit guarantees and reference models, and for evaluations that match the claimed objective.\n\n---\n\n## 50. DataStates-LLM: Scalable Checkpointing for Transformer Models Using Composable State Providers\n- **Authors**: Avinash Maurya , M. Mustafa Rafique , Franck Cappello , Bogdan Nicolae\n- **URL**: [https://arxiv.org/abs/2601.16956](https://arxiv.org/abs/2601.16956)\n- **Abstract**:\n> The rapid growth of Large Transformer-based models, specifically Large Language Models (LLMs), now scaling to trillions of parameters, has necessitated training across thousands of GPUs using complex hybrid parallelism strategies (e.g., data, tensor, and pipeline parallelism). Checkpointing this massive, distributed state is critical for a wide range of use cases, such as resilience, suspend-resume, investigating undesirable training trajectories, and explaining model evolution. However, existing checkpointing solutions typically treat model state as opaque binary blobs, ignoring the ``3D heterogeneity'' of the underlying data structures--varying by memory location (GPU vs. Host), number of ``logical'' objects sharded and split across multiple files, data types (tensors vs. Python objects), and their serialization requirements. This results in significant runtime overheads due to blocking device-to-host transfers, data-oblivious serialization, and storage I/O contention. In this paper, we introduce DataStates-LLM, a novel checkpointing architecture that leverages State Providers to decouple state abstraction from data movement. DataStates-LLM exploits the immutability of model parameters during the forward and backward passes to perform ``lazy'', non-blocking asynchronous snapshots. By introducing State Providers, we efficiently coalesce fragmented, heterogeneous shards and overlap the serialization of metadata with bulk tensor I/O. We evaluate DataStates-LLM on models up to 70B parameters on 256 A100-40GB GPUs. Our results demonstrate that DataStates-LLM achieves up to 4$\\times$ higher checkpointing throughput and reduces end-to-end training time by up to 2.2$\\times$ compared to state-of-the-art solutions, effectively mitigating the serialization and heterogeneity bottlenecks in extreme-scale LLM training.\n\n---\n\n{% endraw %}\n"}, "2caecc6a4c290ccfe491fccbd3fee077e40eb234": {"t": 1785770988.7389321, "n": 84889, "h": "29e92ec4f78e9471", "c": "---\nlayout: default\ntitle: \"LLM 관련 주요 논문 - 2026-06-29\"\n---\n\n{% raw %}\n# LLM 관련 주요 논문 - 2026-06-29\n\n## 1. Tandem Reinforcement Learning with Verifiable Rewards\n- **Authors**: Difan Jiao , Raghav Singhal , Robert West , Ashton Anderson\n- **URL**: [https://arxiv.org/abs/2606.28166](https://arxiv.org/abs/2606.28166)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has significantly improved the reasoning capability of large language models, reaching expert or even superhuman performance in domains such as competition math. However, whether weaker agents and humans can actually harness this capability is far less certain, with RLVR documented to drift reasoning toward idiosyncratic patterns such as poor readability and language mixing. Tandem training is a recently introduced paradigm that targets this compatibility problem: a trained, stronger senior co-generates each rollout with a frozen, weaker junior, and the two are rewarded as a team, so the senior is pushed to reason in ways the junior can follow. Yet this paradigm has so far been demonstrated only in proof-of-concept settings, leaving open whether it scales to the long chains of thought of the modern RLVR pipeline. In this work, we propose Tandem Reinforcement Learning (TRL), which carries the tandem training paradigm into RLVR. In TRL, the senior and a frozen junior alternate stochastically to co-generate the reasoning, the resulting generation is rewarded, and the standard GRPO loss is applied to the senior. Training Qwen3-4B-Instruct on competition math, we find that TRL matches vanilla GRPO on solo reasoning capability while three properties emerge together from the same rollout structure: stronger handoff robustness with the junior, reduced distributional drift from the junior, and a chain-of-thought more legible to the junior. Our results demonstrate a promising route for RLVR with practical payoffs in multi-model communication and human compatibility.\n\n---\n\n## 2. JD Oxygen AI Item Center (Oxygen AIIC) V1: An Industrial-Scale LLM/VLM-Centric Solution for Item Understanding, Management, and Applications\n- **Authors**: Oxygen AIIC , Chan Long , Chao Liu , Chaofan Chen , Chaohui Dong , Chunyuan Guo , Danping Liu , Debin Liu , Deping Xiang , Fulai Xu , Guangyue Liu , Hao Li , Huichun Hu , Jian Yang , Jianan Wang , Jianbo Zhao , Jiaoyang Li , Jiaxing Wang , Jinglong Li , Jinjin Guo , Jun Fang , Jun Liu , Kai Zhou , Li Wang , Lili Gao , Liying Chen , Luning Yang , Mengdi Zhou , Pengzhang Liu , Qi Lv , Qianyun Wang , Qixia Jiang , Ruyue Li , Shimu Liang , Shuxing Wang , Sijie Zhang , Siqi Li , Tianhao Gao , Wang Ke , Weihu Huang , Wencan Lai , Wenjie Zhang , Xiaohui Zhang , Xiaojing Dong , Ya Liu , Yifeng Zhang , Yixiang Wang , Yongtai Zhang , Yongyi Liao , Zhaoru Chen , Zhen Chen , Zhiyong Ma , Zhiyuan Liu , Zhongwei Liu , Ziyan Xing\n- **URL**: [https://arxiv.org/abs/2606.28070](https://arxiv.org/abs/2606.28070)\n- **Abstract**:\n> this http URL , one of the world's largest e-commerce platforms, serves over 700 million active users and millions of merchants, with a catalog of tens of billions of SKUs. At this scale, high-quality, structured item knowledge underpins a better consumer experience, lower management costs, and higher operational efficiency-yet producing and serving it poses three industrial-scale challenges: fast-emerging concepts, high-quality knowledge production for massive SKUs, and diverse downstream requirements. To address these challenges, we present the JD Oxygen AI Item Center (Oxygen AIIC), an industrial-scale platform built on LLMs/VLMs for item-knowledge production and service. Oxygen AIIC is built around four core pillars: (i) ontology engineering driven by efficient human-AI collaboration, which supports the dynamic evolution and agile expansion of an ontology with millions of entries; (ii) a \"Semantic Search then Discrimination\"(S2D) knowledge identification architecture that, combined with throughput improvement strategies, enables scalable, extensible, and high-throughput AI Item Library production for tens of billions of SKUs; (iii) self-evolving item-understanding LLMs/VLMs that improve in a stable and controllable manner, enabling knowledge production with 94.2% precision and 82.8% recall; and (iv) a unified item tunnel that serves as the data and service hub. Oxygen AIIC now covers tens of thousands of JD categories and processes hundreds of millions of item updates per day on Huawei Ascend NPUs. It has accumulated hundreds of billions of item-knowledge assets. Deployed across core business scenarios-including search, recommendation, operations, category planning-Oxygen AIIC has delivered measurable gains at scale. Search-traffic coverage reaches 80.4%, item-information quality issues drop by 37%, the automated fill rate of core attributes during item listing exceeds 80%.\n\n---\n\n## 3. NormAct: A Benchmark for Hidden Social Norm Compliance in Embodied Planning\n- **Authors**: Shiyun Zhao , Xinwei Song , Tianyu Guo , Xiaomeng Gao , Mingyuan Liu , Xu Han , Yuanyuan Zhang , Zhenliang Zhang , Xue Feng , Bo Dai\n- **URL**: [https://arxiv.org/abs/2606.27826](https://arxiv.org/abs/2606.27826)\n- **Abstract**:\n> Multimodal large language models (MLLMs) are increasingly deployed as embodied planners in egocentric environments, where task success requires not only achieving instructed goals but also acting in socially appropriate ways. While explicit goals may render certain actions optimal, implicit social norms often impose hidden constraints. Existing evaluations typically focus on explicit goal achievement or direct norm knowledge, seldom assessing whether planners can infer and apply these hidden constraints within action sequences. We introduce NormAct, a benchmark for embodied social-norm interactions that evaluates plans on Goal Achievement, Norm Compliance, and overall Task Success. NormAct uniquely embeds hidden norms within ordinary tasks, testing whether models can realize them without explicit instruction. Experiments with state-of-the-art MLLMs (GPT-5.4, Claude Opus 4.7, Gemini 3 Pro) reveal a significant gap: models achieve explicit goals in 67.3\\% of cases, but comply with hidden norms in only 26.4\\%. Cue-condition experiments indicate that this gap stems not from a lack of general social knowledge, but from challenges in activating and grounding relevant norms in context. To address this, we propose NormPerceptor, a context-conditioned cue generator that infers scene-relevant norms prior to planning, increasing Task Success from 24.2\\% to 46.7\\%. Our results underscore the importance of enabling embodied agents to proactively detect hidden norms, ground them in visual evidence, and integrate them as action-planning constraints. Our benchmark is publicly available at this https URL .\n\n---\n\n## 4. Grounded Iterative Language Planning: How Parameterized World Models Reduce Hallucination Propagation in LLM Agents\n- **Authors**: Xinyuan Song , Zekun Cai\n- **URL**: [https://arxiv.org/abs/2606.27806](https://arxiv.org/abs/2606.27806)\n- **Abstract**:\n> World models for language agents come in two useful forms. An agent-based world model calls an LLM API and reasons flexibly in language, but its errors appear as hallucinated state changes that are hard to score with ordinary regression losses. A parameterized world model is a trained transition predictor; its errors are easier to measure with quantities such as NodeMSE, delta accuracy, and validity accuracy, but it is usually weaker as a standalone planner. We compare these two families on four graph-structured planning benchmarks and introduce operational hallucination metrics for the agent-based case. The comparison motivates \\textbf{Grounded Iterative Language Planning} (GILP), which trains only a small parameterized backbone and combines it with API-based agent reasoning. The backbone supplies valid actions, predicted state deltas, risk, and value; the LLM drafts an action and imagined delta; and a consistency gate asks for revision when the two disagree. On real GPT-4o-mini calls, GILP reduces hallucinated-state rate from 0.176 to 0.035. In calibrated simulator ablations, it raises success from 0.668 to 0.838 while adding only ~22% extra LLM calls.\n\n---\n\n## 5. Towards Reliable and Robust LLM Planning: Symbolic Feedback-Driven Iterative Self-Refinement Framework\n- **Authors**: Jiajing Zhang , Jiamei Jiang , Chenyang Zhang , Feifei Mo , Linjing Li , Daniel Zeng\n- **URL**: [https://arxiv.org/abs/2606.27757](https://arxiv.org/abs/2606.27757)\n- **Abstract**:\n> Large language models (LLMs) have attracted widespread attention from academia and industry, yet their deployment raises critical security concerns regarding robustness and reliability. Planning, a core component of intelligent behavior, remains challenging for LLMs, which often produce infeasible or incorrect solutions in long-horizon decision-making tasks due to inherent complexity. In this paper, we propose a symbolic feedback-driven iterative self-refinement framework to enhance the robustness and reliability of LLMs in long-horizon planning. Specifically, a natural language prompting mechanism is introduced to map logical symbols into natural language descriptions, enabling LLMs to better capture task constraints and semantics. We further design a symbolic verifier that identifies errors and converts them into corrective instructions interpretable by the LLM, thereby guiding self-refinement. In addition, we leverage a plan recognizer to infer goal reachability, facilitating more effective guidance toward desired goals. Empirical results demonstrate that the proposed framework consistently improves both feasibility and correctness in long-horizon planning tasks. This highlights its effectiveness in enhancing the reliability of LLM-based planning and potential to enable more trustworthy AI systems.\n\n---\n\n## 6. ToE: A Hierarchical and Explainable Claim Verification Framework with Dynamic Multi-source Evidence Retrieval and Aggregation\n- **Authors**: Zhaoqi Wang , Zijian Zhang , Kun Zheng , Zhen Li , Xin Li , Chunlei Li , Jiamou Liu\n- **URL**: [https://arxiv.org/abs/2606.27736](https://arxiv.org/abs/2606.27736)\n- **Abstract**:\n> The rapid spread of fake news poses increasing threats to information ecosystems, especially as AI-generated misinformation under Generative Engine Optimization (GEO) poisoning allows adversarially crafted content to be systematically surfaced by retrieval systems, contaminating LLM reasoning. In this paper, we propose Tree of Evidence (ToE), a hierarchical evidence reasoning framework for automated fact-checking that models each claim as a dynamically expanding argument tree. ToE integrates a reinforcement learning-driven multi-source retrieval agent, an evidence evaluation agent, and an argument tree aggregation algorithm to iteratively decompose, retrieve, and verify claims through an explainable evidence chain. We further provide a theoretical analysis of the retrieval process, deriving a formal error bound that guarantees the learned policy converges to a neighborhood of the information-theoretically optimal policy. Experiments across multiple datasets and backbone LLMs demonstrate that ToE achieves improvements ranging from 4 to 24 percentage points over competitive baselines, with particularly pronounced gains on adversarially poisoned inputs.\n\n---\n\n## 7. DysLexLens: A Low-Resource LLM Framework for Analysing Dyslexic Learners Insights from Online Forums\n- **Authors**: Dana Rezazadegan , Atie Kia , Phongpadid Nandavong , Dominique Carlon , Jeremy Nguyen , Abhik Banerjee , James Marshall , Anthony McCosker , Yong-Bin Kang\n- **URL**: [https://arxiv.org/abs/2606.27619](https://arxiv.org/abs/2606.27619)\n- **Abstract**:\n> Dyslexic learners increasingly use artificial intelligence (AI) tools to support reading, writing, organisation, and study-related tasks. However, their lived experiences with these tools remain largely underexamined. This paper proposes DysLexLens, a low-resource LLM framework, designed to analyse dyslexic learners experience with AI through online forum discussions. DysLexLens is designed as an end-to-end, evidence-traceable architecture which transforms noisy social media posts into a dictionary-driven corpora, provides knowledge-graph (KG)-based question reasoning, generates verifiable query responses, and enables response evaluation through quantitative and human-grounded assessment. DysLexLens has four key features. First, it employs a dictionary-driven filtering method to construct a more focused Reddit corpus on dyslexia and AI, filtering out noisy and weakly related posts to improve the relevance of data collected from low-resource forum contexts. Second, it integrates LLM-assisted semantic analysis with KG-based query reasoning to uncover meaningful patterns. Third, it has quantitative evaluation metrics (RAGAS and Query Robustness) to measure LLM-generated response performance. Fourth, it provides structured qualitative validation guidelines for assessing response quality, with a specific focus on hallucination and evidence alignment. We demonstrate the effectiveness of DysLexLens using dyslexia-related Reddit forum data and 30 questions. The results show its potential generalisability to other low-resource forum data contexts. DysLexLens, sample data, questions and evaluation results are available at Github to support reproducibility.\n\n---\n\n## 8. Odyssey: Constructing Verifiable Local Truth-Preserving Foundation Models\n- **Authors**: Sridhar Mahadevan\n- **URL**: [https://arxiv.org/abs/2606.27593](https://arxiv.org/abs/2606.27593)\n- **Abstract**:\n> We introduce a categorical framework called ODYSSEY for constructing verifiable, local truth-preserving foundation models as compositions of foundries: building-block architectural components that specify a cover of local contexts, local representation families, restriction maps, gluing rules, obstruction policies, update obligations, and human-facing views. A foundry is an organized sheaf of knowledge that carries within it an argumentation component. Concrete foundries are built from generic foundries such as evidence/argument, operational decision, institutional/financial, market meaning, scientific challenge, research-program, assistant-build, and evaluation-harness foundries. Universal Foundry Learning (UFL) formalizes foundry construction as a composition of left and right Kan extensions, with left Kan extension rolling local artifacts into candidate foundries and right Kan extension enforcing the restriction, gluing, obstruction, and argumentation conditions required for promotion. Foundry SQL (FSQL) is a small typed query surface for slicing maintained foundry artifacts that uses TICKET (Topos Integration using Causal Kan Extension Transformers) certification for admitting external or pre-built models into durable ODYSSEY state. ODYSSEY is fully implemented and tested across a wide spectrum of concrete foundries, showing that the same categorical machinery supports domain construction, artifact replay, sheaf diagnostics, grounded Toulmin/local-LLM scrutiny, residual-obstruction ledgers, and optimized TICKET-compatible causal-claim extraction across heterogeneous sources. This paper is to be presented as a 2.5 hour tutorial at ICML 2026. The tutorial home page is at this https URL .\n\n---\n\n## 9. Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning\n- **Authors**: Xuan Zhang , Zhijian Zhou , Lingfeng Qiao , Yulei Qin , Ke Li , Xing Sun , Xiaoyu Tan , Chao Qu , Yuan Qi\n- **URL**: [https://arxiv.org/abs/2606.27483](https://arxiv.org/abs/2606.27483)\n- **Abstract**:\n> Large language model (LLM) agents have demonstrated strong capability in sequential decision-making, yet they remains fundamentally reactive in long-horizon tasks. Unlike humans who employ \"what-if\" reasoning to evaluate potential plans before commitment, standard agents lack an internal world model to simulate future outcomes. Therefore, we propose to internalize future-aware planning by training a single autoregressive model to verbalize both a prospective state rollout and a plan-conditioned success estimate-a textual analogue of the Q-value. Crucially, we identify a format-capability gap: simply fine-tuning agents on look-ahead traces during post-training leads to superficial mimicry of foresight without genuine predictive grounding. To bridge this gap, we introduce a three-stage training paradigm: (i) World Model Agentic Mid-Training (WM-AMT) to inject latent predictive capabilities into the policy; (ii) Format-Eliciting SFT (FE-SFT) to structure this injected capability; and (iii) Foresight-Conditioned Reinforcement Learning (FC-RL) to refine the calibration and utility of the generated simulations. Evaluated on search and mathematical reasoning tasks, our approach consistently outperforms other training baselines. Our results demonstrate that effective internal world modeling in LLM agents requires a capability-first training pipeline to achieve grounded and calibrated foresight.\n\n---\n\n## 10. When Does Personality Composition Matter for Multi-Agent LLM Teams?\n- **Authors**: Aryan Keluskar , Amrita Bhattacharjee , Huan Liu\n- **URL**: [https://arxiv.org/abs/2606.27443](https://arxiv.org/abs/2606.27443)\n- **Abstract**:\n> Personality prompting shapes how large language models communicate, yet whether these behavioral shifts affect objective task outcomes remains under-explored. Prior work shows that agents prompted with low agreeableness produce adversarial language, while those prompted with high agreeableness become cooperative, but the relationship between communication style and task performance has not been systematically examined across multiple domains. In this work, we investigate whether personality composition matters for multi-agent team performance by manipulating personality traits across frontier LLMs on three task domains: structured coding, open-ended research collaboration, and competitive bargaining. We find that personality effects depend critically on task structure. In coding tasks, low agreeableness leads to large communication shifts that have little effect on milestone completion. In open-ended collaboration and bargaining, the same manipulation substantially degrades performance. We discuss implications for multi-agent system design and the limits of personality manipulation.\n\n---\n\n## 11. AI-Model Network: Concept, Current State and Future\n- **Authors**: Li Zhetao , Zeng Xiyu , Wang Jianhui , Xiao Yong , Liu Zhongren , Wu Junru , Lai Junjie , Huang Jijun , Long Saiqin\n- **URL**: [https://arxiv.org/abs/2606.27382](https://arxiv.org/abs/2606.27382)\n- **Abstract**:\n> While the primary function of computers lies in computation and processing, the core value of the Internet is rooted in sharing and collaboration. Computers create the Internet, and the Internet empowers the value of computers. The rapid development of the Internet, cloud computing, and big data is pushing artificial intelligence into the era of large models (LMs). However, the practical application of LMs is currently hindered by high training costs and deployment complexities, driving a shift toward lightweight, private, and domain-specific models. With the rapid proliferation and wide distribution of heterogeneous models, enabling effective interaction and collaboration among them has emerged as a critical bottleneck that urgently needs to be addressed in LM development. Drawing inspiration from the development of the Internet, this paper proposes the concept, vision, and system architecture of world wide AI-model network (AI-ModelNet). It is a novel paradigm that achieves interconnection, capability sharing, and collaborative reasoning by establishing pathways between models. We first briefly review the current state of single-model and multi-model research. Subsequently, the systemic vision and hierarchical architecture of AI-ModelNet are articulated, followed by validation of the framework's feasibility through a prototype system and diverse application cases. Finally, key directions for future research are discussed preliminarily.\n\n---\n\n## 12. Cognitive Episodes in LLM Reasoning Traces Enable Interpretable Human Item Difficulty Prediction\n- **Authors**: Chenguang Wang , Ming Li , Xinyue Zeng , Zhuochun Li , Hong Jiao , Tianyi Zhou , Dawei Zhou\n- **URL**: [https://arxiv.org/abs/2606.28186](https://arxiv.org/abs/2606.28186)\n- **Abstract**:\n> Predicting human item difficulty is central to educational assessment, where reliable estimates support fairness and effective test construction. Existing methods often depend on costly human calibration or item-level textual representations, providing limited evidence about the cognitive processes that make items difficult. We argue that difficulty should be viewed not only as a property of item text, but also as an observable consequence of the problem-solving burden an item induces. Large Reasoning Models (LRMs) offer scalable process evidence through reasoning traces, but such evidence must be structured to support interpretable modeling. To this end, we introduce Epi2Diff (Episode to Difficulty), a framework that maps LRM reasoning traces into cognitively grounded episode sequences. These episodes group trace segments into functional problem-solving states, enabling difficulty to be modeled through reasoning scale, effort allocation, and state transitions. Epi2Diff extracts compact episode-dynamic features and combines them with semantic item representations for human difficulty prediction. Experiments on four real-world human difficulty datasets show that Epi2Diff consistently outperforms strong baselines, including fine-tuned small language models, LLM in-context learning, and supervised LLM adaptation. On SAT-derived classification benchmarks, Epi2Diff achieves an 8.1% average relative gain over supervised LLM fine-tuning baselines. Further analyses show that harder items induce more effortful, iterative, and implementation-centered episode dynamics, rather than merely longer responses. These results demonstrate that cognitive episodes in LRM reasoning traces provide a predictive and interpretable process representation for human item difficulty, offering a new lens for educational measurement with reasoning models.\n\n---\n\n## 13. LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior\n- **Authors**: Qinhong Zhou , Chuang Gan , Anoop Cherian\n- **URL**: [https://arxiv.org/abs/2606.28182](https://arxiv.org/abs/2606.28182)\n- **Abstract**:\n> Embodied agents operating in decentralized and partially observable environments have attracted growing attention in recent years. However, existing large language model (LLM)-based agents often exhibit behaviors that are misaligned with their partners or inconsistent with the environment state, leading to inefficient cooperation and poor task success. To address this challenge, we propose a novel framework, Learning Laws of Cooperation (LLawCo), that enables embodied agents to autonomously align with both their partners and task objectives. Our framework allows agents to reflect on past failures to extract misaligned behavioral patterns, which are used to derive high-level behavioral laws, such as \"Talk when necessary\" and \"Wait for partner.\" These laws are explicitly incorporated into the agents' chains of thought via supervised fine-tuning, aligning their reasoning with task requirements and the behavior of other agents. To evaluate our approach, we introduce PARTNR-Dialog, a large-scale multi-agent communicative and cooperative planning benchmark built on the PARTNR environment. Experiments on existing tasks and our new benchmark demonstrate significant improvements in cooperative efficiency and task success rates. Across four backbone LLMs, our method achieves average success rate improvements of 4.5% on the PARTNR-Dialog benchmark and 6.8% on the TDW-MAT benchmark over state-of-the-art open-source communicative agent frameworks. See the LLawCo project page for details: this https URL\n\n---\n\n## 14. Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models\n- **Authors**: Yanchen Yin , Dongqi Han , Linghui Li\n- **URL**: [https://arxiv.org/abs/2606.28153](https://arxiv.org/abs/2606.28153)\n- **Abstract**:\n> Jailbreak attacks bypass LLM safety alignment, yet their mechanisms remain poorly understood. We provide evidence that attacks do not comprehensively eliminate safety features, but instead selectively suppress specific attention heads. We identify two functionally differentiated types: Adversarially Compromised Heads (ACHs) concentrated in early layers, which are suppressed under attacks, and Safety-Aligned Heads (SAHs) in mid-layers, which maintain robust activations even when attacks succeed. Ablation studies support the causal role of ACHs and the contribution of SAHs to robust activations: suppressing a small number of ACHs is sufficient to induce jailbreak-like behavior on normally refused inputs, while removing SAHs substantially weakens mid-layer safety activations. Token-level attribution further shows that ACH suppression is driven specifically by attack-template tokens, providing a mechanistic account of why attacks can bypass refusal decisions through ACH suppression while leaving internal safety signals sustained by SAHs -- a phenomenon we term Robust Harmful Features. To validate the practical significance of this robustness, we show that simply reading these persistent activations -- without any training -- yields competitive aggregate detection performance with strong adversarial robustness.\n\n---\n\n## 15. From Tokens to States: LLMs as a Special Case of World Models and the Continuous Path Beyond\n- **Authors**: Paul Dubois\n- **URL**: [https://arxiv.org/abs/2606.28127](https://arxiv.org/abs/2606.28127)\n- **Abstract**:\n> The AI community has framed the relationship between large language models (LLMs) and world models as a dichotomy: LLMs predict tokens; world models simulate reality. Yann LeCun argues in 2022 that reaching general intelligence requires abandoning autoregressive token prediction in favour of latent-space architectures. This framing is unnecessarily binary. Two claims will be defended. First, LLMs are a degenerate special case of world models: the state space is the set of all token sequences, the only action is appending one token, and world models are therefore a strict generalisation of LLMs, not a replacement. Second, there is a natural continuous spectrum from NTP to JEPA, with multi-token prediction, future-summary prediction, and next-latent prediction as intermediate stations already populated by current research. Moving along this spectrum relaxes the LLM constraints one by one. It also progressively surrenders the two practical advantages that make LLMs trainable at scale: internet-scale self-supervised data, and a transformer architecture co-designed for discrete token prediction. Both are examined as open research questions: the data question (the cliff from self-supervised text to instrumented action-labelled environments) and the architecture question (whether the transformer generalises to continuous-state prediction, or whether a new primitive is needed).\n\n---\n\n## 16. Single and Multi Truth Data Fusion using Large Language Models\n- **Authors**: Hira Beril Kucuk , Norman W Paton , Jiaoyan Chen , Zhenyu Wu\n- **URL**: [https://arxiv.org/abs/2606.28062](https://arxiv.org/abs/2606.28062)\n- **Abstract**:\n> Data fusion, also known as truth discovery, is a data integration problem that aims to determine the correct value or set of values for each attribute of an object when presented with potentially conflicting values from multiple sources. Data fusion tasks belong to two main categories: single-truth scenarios, where each attribute has only one correct value, and multi-truth scenarios, where multiple values can be valid simultaneously. This paper investigates the use of Large Language Models (LLMs) in data fusion tasks for tabular data. Various prompting strategies, encompassing both single-truth and multi-truth scenarios, are investigated empirically. Domain-dependent, domain-independent, zero-shot and one-shot prompts are evaluated on three different benchmark datasets. Experimental results demonstrate that LLM-based approaches outperform traditional unsupervised truth discovery methods, such as DART and LTM, across all datasets. The codebase of this study has been made publicly available on GitHub.\n\n---\n\n## 17. ToolPrivacyBench: Benchmarking Purpose-Bound Privacy in Tool-Using LLM Agents\n- **Authors**: Shijing Hu , Liang Liu , Zhu Meng , Zhicheng Zhao\n- **URL**: [https://arxiv.org/abs/2606.28061](https://arxiv.org/abs/2606.28061)\n- **Abstract**:\n> Large language models (LLMs) have increasingly moved from standalone text generation systems to agents that invoke external tools, access environments, and execute multi-step tasks. However, conventional function-calling benchmarks mainly evaluate task completion and API correctness, while privacy evaluation benchmarks typically focus on final responses or privacy judgments. Neither perspective captures purpose-bound information flow across an executed multi-tool trajectory. Motivated by this limitation in current agent evaluation, ToolPrivacyBench audits whether task-private atoms are routed only to authorized tools and downstream sinks, thereby evaluating both task completion and privacy over-disclosure during tool use. The benchmark contains 2,150 cases, including 1,150 fully synthetic privacy-sensitive business workflows and 1,000 cases adapted from existing multi-tool and function-calling benchmarks. Each case is represented by a policy knowledge base. After an agent executes against mock business backends, the evaluator compares recorded tool arguments and backend audit logs with this policy knowledge base. The evaluation covers nine widely used agents to characterize purpose-bound privacy over-disclosure. The results show that successful tool execution does not imply appropriate privacy disclosure: an agent may complete a task while transmitting unnecessary private information through intermediate tool calls. ToolPrivacyBench therefore formalizes a need-to-know disclosure boundary, under which each tool should receive only the information necessary for its stated purpose, and uses trajectory-level auditing to identify privacy over-disclosure in multi-tool workflows.\n\n---\n\n## 18. MultiHashFormer: Hash-based Generative Language Models\n- **Authors**: Huiyin Xue , Atsuki Yamaguchi , Nikolaos Aletras\n- **URL**: [https://arxiv.org/abs/2606.28057](https://arxiv.org/abs/2606.28057)\n- **Abstract**:\n> Language models (LMs) represent tokens using embedding matrices that scale linearly with the vocabulary size. To constrain the parameter footprint, prior work proposes hashing many tokens into a single vector within encoder-only models. While this offers parameter efficiency, many-to-one collisions prevent its use in causal LMs. In this paper, we propose MultiHashFormer, a new framework that allows hash-based autoregression. Each token is represented as a unique hash signature, a short sequence of discrete hash IDs, generated by multiple independent hash functions. A Hash Encoder compresses this signature into a single latent vector for processing by a Transformer decoder. Then, a Hash Decoder generates the hash signature of the next token, which is then mapped back to text. We evaluate our approach at the 100M, 1B and 3B parameter scales, demonstrating that MultiHashFormer consistently outperforms standard Transformer LMs across multiple benchmarks. Furthermore, we show that our model handles multilingual vocabulary expansion with a constant parameter footprint without any modifications.\n\n---\n\n## 19. Can LLMs Judge Better Than They Generate? Evaluating Task Asymmetry, Mechanistic Interpretability and Transferability for In-Context QA\n- **Authors**: Sambaran Bandyopadhyay\n- **URL**: [https://arxiv.org/abs/2606.28050](https://arxiv.org/abs/2606.28050)\n- **Abstract**:\n> LLM-as-a-Judge and self-evaluation pipelines implicitly assume that evaluation is easier than generation. We test this in a controlled in-context QA setting where a context passage is the sole information source and each model judges the answer it generated, removing the parametric-knowledge confound of open-domain comparisons. Across four benchmarks (SQuAD 2.0, DROP, HotpotQA, MuSiQue) and two models, evaluation is not uniformly easier: generation accuracy exceeds self-evaluation on three of four, with multi-hop MuSiQue the exception. Attention analysis reveals why: evaluation attends to context 3--5x less than generation does and barely reads the candidate answer. LoRA fine-tuning confirms the asymmetry is not a training artifact: generation fine-tuning induces over-acceptance and evaluation fine-tuning degrades generation. These findings challenge core assumptions in self-evaluation pipelines.\n\n---\n\n## 20. Dialogue to Detection: A Multimodal Hybrid NLP Pipeline for Insurance Fraud Detection\n- **Authors**: Muhammad Shakeel Akram , Amal Htait , Abdul Hamid Sadka , Emma Meisingseth , Karishma Jaitly\n- **URL**: [https://arxiv.org/abs/2606.28002](https://arxiv.org/abs/2606.28002)\n- **Abstract**:\n> Insurance fraud imposes substantial financial losses and operational inefficiencies, raising premiums and impacting trust among legitimate policyholders. Early detection at FNOL remains a persistent challenge. Existing approaches rely largely on private, text-only datasets, limiting progress on multimodal methods that integrate linguistic, behavioural, and speaker-based indicators. We introduce a synthetic multimodal framework that replicates FNOL conditions. It generates agent-customer dialogue transcripts and two-speaker audios, performs ASR and diarisation. Downstream modules combine NER, regex-based feature extraction, LLM-RAG retrieval, and speaker embeddings in a rule-based risk score to flag narrative reuse, structural inconsistencies, and cross-case voice repetition while balancing sensitivity and false positives. Dataset validation and component-level evaluations show stability and transfer potential, offering a reproducible baseline beyond text-only fraud detection.\n\n---\n\n## 21. From Black-Box to Clinical Insight: A Multi-Stage Explainable Framework for Speech-Based Cognitive Impairment Detection\n- **Authors**: Yasaman Haghbin , Sina Rashidi , Ali Zolnour , Fatemeh Taherinezhad , Ali Fartoot , Hossein Azadmaleki , James M Noble , Maryam Dadkhah , Maryam Zolnoori\n- **URL**: [https://arxiv.org/abs/2606.27973](https://arxiv.org/abs/2606.27973)\n- **Abstract**:\n> Speech-based cognitive impairment detection offers a noninvasive, accessible alternative to costly biomarker assays, yet transformer-based models remain clinically uninterpretable. We propose a multi-stage explainability framework that translates black-box transformer predictions into clinically grounded narratives by integrating SHapley Additive exPlanations (SHAP)-based token attribution, theory-informed linguistic features, and a four-stage LLM reasoning pipeline using LLaMA-3.1-70B-Instruct. Built on the SpeechCARE-Adaptive Gating Network multimodal screening model (F1 = 72.11% on the NIA PREPARE benchmark), the framework maps model outputs to four cognitive-linguistic dimensions, including lexical richness, syntactic complexity, and semantic coherence. Physician evaluation on 70 stratified English samples demonstrated strong alignment with patient-level cognitive profiles, and a System Usability Scale score of 82/100 indicated high potential for clinical workflow integration.\n\n---\n\n## 22. Two-Stage Fine-Tuning for Protein Sequence Generation with Targeted Amino-Acid Composition\n- **Authors**: Violeta Basten-Romero , Rubén Muñoz-Tafalla , Anna María Díaz-Rovira , Bertran Miquel-Oliver , Isaac Filella-Merce , Víctor Guallar\n- **URL**: [https://arxiv.org/abs/2606.27939](https://arxiv.org/abs/2606.27939)\n- **Abstract**:\n> Protein language models are standard priors for biological sequence generation, but steering them toward explicit distributional design targets remains largely unexplored. We study a constrained protein generation problem in which sequences must match a desired amino-acid (AA) composition profile while preserving plausible sequence statistics and diversity. The motivating application is synthetic feed protein design, where the AA composition of dietary proteins directly determines their nutritional value. We propose a two-stage pipeline in which domain-adaptive fine-tuning (FT) on an in-domain protein dataset is followed by iterative reward-weighted FT via reinforcement learning (RL) anchored against the FT model as a frozen reference. We evaluate the pipeline on two AA compositions and find that FT brings the average composition close to the target, while the subsequent RL enforces specific sequence constraints that FT alone cannot satisfy. We additionally evaluate the design choices of the proposed composition reward term against two baselines and an ablated variant, isolate the contribution of each training stage, and verify that AA composition alignment is achieved without degrading sequence quality.\n\n---\n\n## 23. Agentic AI-Powered Re-Identification: An Emerging, Scalable Threat to Mobility Microdata Privacy\n- **Authors**: Oscar Thees , Roman Müller , Matthias Templ\n- **URL**: [https://arxiv.org/abs/2606.27936](https://arxiv.org/abs/2606.27936)\n- **Abstract**:\n> The widespread collection of fine-grained location data by commercial data brokers creates a re-identification risk that is not widely recognised by the public. While prior research has established that mobility traces are highly unique and that individuals can, in principle, be identified from a handful of spatio-temporal points, such attacks have historically required significant manual effort from skilled analysts, limiting their practical scale. In this feasibility study, we demonstrate in a real world setting that agentic AI fundamentally changes this threat model. We present an end-to-end pipeline in which large language model agents autonomously search the open web, cross-reference public records and social media, and resolve raw coordinate sequences to candidate identities - without human intervention. We evaluate the pipeline on a spatio-temporal dataset containing simulated location points anchored at and around true home and work addresses, focusing on a high-risk disclosure scenario. Our results demonstrate that, from spatio-temporal data and public sources alone, our agentic AI successfully re-identified 18 of the 25 re-identifiable individuals (72%) and 18 of 43 cases overall (41.9%). We discuss implications for Statistical Disclosure Control (SDC) practice and outline the near-future escalation that data custodians and regulators must anticipate. De facto anonymity - an implicit foundation of SDC practice - is shifting. Agentic AI strengthens the case that re-identification is reasonably likely by any means under the GDPR Recital-26 standard, at costs of minutes-and-dollars per target.\n\n---\n\n## 24. Triadic Werewolf: A Jester Role for Multi-Hop Theory of Mind in LLMs\n- **Authors**: Avni Mittal\n- **URL**: [https://arxiv.org/abs/2606.27909](https://arxiv.org/abs/2606.27909)\n- **Abstract**:\n> Theory-of-mind evaluations of large language models typically use dyadic social-deduction games, where every observable cue points to a single hidden side, so a model with strong language priors can score well without ever simulating opponents' incentives. We extend the Werewolf game with a Jester, a third faction whose utility on peer suspicion is inverted because it wins by being voted out, so optimal play requires reasoning across three opposing utility functions. Across 60 games on GPT-4.1, DeepSeek-V3.1, and Llama-3.3-70B with Jester self-learning on and off, the Jester wins 60-70% of games while Werewolves never exceed 20%, and GPT-4.1 wolves vote the Jester out on day 1 in 60-70% of games, a strictly self-defeating action. Self-learning helps DeepSeek and Llama but hurts GPT-4.1, with the cost landing on Villagers rather than Werewolves. Only DeepSeek learns the subtle strategy of looking suspicious without looking intentionally suspicious, and it gains the most from the loop. Triadic incentive structure exposes a layer of multi-agent reasoning that dyadic deduction games leave invisible.\n\n---\n\n## 25. A Study of Temporal Fusion Strategies for Named Entity Recognition in Historical Texts\n- **Authors**: Emanuela Boros\n- **URL**: [https://arxiv.org/abs/2606.27881](https://arxiv.org/abs/2606.27881)\n- **Abstract**:\n> Temporal variation poses a unique challenge for named entity recognition (NER) in historical texts, where entities drift in surface form and salience across time. While language models (LMs) have made progress in various NLP tasks, their ability to reason about temporality, especially in diachronic contexts, remains limited or at least, questionable. In this paper, we systematically study how temporal metadata can be structurally embedded into NER models using a range of lightweight fusion strategies. We experiment with both absolute and relative temporal representations, injected into Transformer-based architectures via early or late fusion mechanisms such as cross-attention, adapters, and concatenation. Our evaluations on French and German historical datasets reveal that late fusion strategies yield more robust and temporally generalisable performance, particularly in early and noisy periods.\n\n---\n\n## 26. SpatialUAV: Benchmarking Spatial Intelligence for Low-Altitude UAV Perception, Collaboration, and Motion\n- **Authors**: Haoyu Zhang , Meng Liu , Qianlong Xiang , Kun Wang , Yaowei Wang , Liqiang Nie\n- **URL**: [https://arxiv.org/abs/2606.27876](https://arxiv.org/abs/2606.27876)\n- **Abstract**:\n> Spatial intelligence is essential for low-altitude unmanned aerial vehicle (UAV) perception, collaboration, and navigation. However, existing UAV benchmarks often emphasize image-level recognition, single-view understanding, or narrow answer formats, leaving 3D spatial inference, multi-view collaboration, scene dynamics, and diverse task formulations insufficiently evaluated. To address these gaps, we introduce SpatialUAV, a real low-altitude UAV benchmark comprising 4,331 curated instances across 14 fine-grained task types, covering semantic discrimination, spatial relation, aerial--aerial collaboration, aerial--ground collaboration, and motion understanding. SpatialUAV organizes all samples into a unified visual-input--question--answer schema, while supporting seven input configurations and nine answer formats, including option labels, region identifiers, geometric values, cross-view correspondences, and free-form motion descriptions. To ensure reliable and grounded evaluation, our data construction pipeline integrates detector-assisted regions, depth supervision, metadata-derived rules, extensive manual annotation, blind filtering, and multi-turn human validation, together with task-specific metrics for heterogeneous outputs. Evaluating representative vision-language models across three categories, we show that current models remain far from human-level performance, with pronounced bottlenecks in cross-view association, structured grounding, geometric reasoning, and temporal viewpoint understanding. These results offer empirical guidance for advancing low-altitude UAV spatial intelligence. Code and data are available at this https URL .\n\n---\n\n## 27. Position Bias Correction is Insufficient for One-Pass Attention Sorting\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27793](https://arxiv.org/abs/2606.27793)\n- **Abstract**:\n> Long-context language models suffer from position bias, where information in middle positions is underutilized. Attention Sorting addresses this by iteratively reordering documents based on attention patterns, but its multiple sort-and-generate cycles increase deployment cost. We hypothesize that position bias is the primary bottleneck and propose Debiased One-Pass Attention Sorting, which estimates a per-prompt position-bias curve from the low-attention majority of documents and uses it to correct raw attention scores (via subtraction or division) to enable single-pass sorting. Our experiments on two models refute this hypothesis in the tested setting: on LLaMA-2-7B-32K-Instruct, debiasing produces identical results to uncalibrated single-pass sorting (94.83\\% containment accuracy), while on YaRN-Llama-2-7b-64k, debiasing improves accuracy by 8.67 percentage points but remains 14.84pp behind iterative sorting, closing only 37\\% of the gap. These results suggest that position-bias correction is insufficient to match iterative sorting, and that repeated reordering provides additional benefits beyond bias correction.\n\n---\n\n## 28. NLL-Guided Full-Attention Layer Selection for Training-Free Sliding-Window Adaptation\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27791](https://arxiv.org/abs/2606.27791)\n- **Abstract**:\n> Hybrid attention models that mix full and sliding-window attention across layers offer a promising approach to efficient long-context inference, but the critical question of \\emph{which layers} should retain full attention remains unsolved. Existing methods use either fixed periodic patterns or attention-based heuristics that may not capture what matters for downstream accuracy. We propose NLL-guided layer selection, a training-free method that directly measures each layer's importance by computing the negative log-likelihood degradation on answer tokens when that layer uses sliding-window instead of full attention. On LongMemEval with Qwen3-4B, our method achieves 64.6\\% accuracy using only 1/4 full-attention layers, matching the 1/2-FA periodic baseline (65.0\\%) while halving the computational budget. NLL-guided selection outperforms the SWAA-reported periodic 1/4-FA baseline by 10.4 percentage points and a matched LightTransfer-style baseline by 26.4 percentage points. De-confounding analysis shows the signal is consistent with long-range attention needs rather than generic layer sensitivity. The method requires only $\\sim$15 minutes of one-time calibration, advancing the efficiency-accuracy Pareto frontier for long-context LLM deployment.\n\n---\n\n## 29. Output-Space Allocation Costs for Calibration-Guided LLM Compression: An Empirical Study\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27785](https://arxiv.org/abs/2606.27785)\n- **Abstract**:\n> Training-free compression methods for large language models (LLMs) often use calibration data to guide compression decisions. ROCKET, a recent method combining sparse-dictionary factorization with multi-choice knapsack problem (MCKP) allocation, derives its per-layer factorization from an output reconstruction objective but uses weight-space Frobenius error as the MCKP allocation cost. We investigate whether aligning the allocation cost with the output-space objective improves compressed model fidelity. On Qwen3-8B at 50\\% compression, our ROCKET-ActCost achieves +0.8 percentage points higher average accuracy across 8 zero-shot benchmarks (53.1\\% vs 52.3\\%), but increases WikiText perplexity by 16\\% (61.46 vs 52.98). This accuracy-perplexity tradeoff reveals that different allocation objectives favor different downstream metrics. The high correlation ($>$0.99) between weight-space and output-space errors limits allocation divergence, explaining the modest effect size. On Llama-3.2-1B at 20\\% compression, the two methods produce near-identical results (53.3\\% vs 53.5\\% accuracy, 14.45 vs 14.66 PPL), suggesting that the effect of the cost function is minor at lower compression ratios.\n\n---\n\n## 30. Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?\n- **Authors**: Guoheng Sun , Kaixi Feng , Shwai He , Xiaochuan Gong , Yexiao He , Ziyao Wang , Zheyu Shen , Wanghao Ye , Ramana Rao Kompella , Gaowen Liu , Ang Li\n- **URL**: [https://arxiv.org/abs/2606.27755](https://arxiv.org/abs/2606.27755)\n- **Abstract**:\n> Vision-Language-Action (VLA) models enable instruction-driven robotic manipulation, but they inherit oversized language backbones from pretrained VLMs whose capacity far exceeds what is needed for short robotic instructions. This raises a basic question: how much of a VLA model is actually necessary for closed-loop control? In this work, we study architectural redundancy in VLA models by using transformer block removal as a controlled intervention. We introduce \\textbf{Drop-Then-Recovery (DTR)}, an analysis protocol that removes selected blocks from a pretrained VLA model and then fine-tunes the resulting model to measure whether the removed capacity was necessary for downstream control. To make this intervention reliable, we propose \\textbf{GateProbe}, a one-shot virtual-gate sensitivity metric that ranks blocks by their contribution to the downstream action loss. Across multiple VLA architectures, manipulation benchmarks and even real-robot industrial scenarios, we find a strong asymmetry in post-removal recoverability: \\ul{\\textit{language backbones are highly redundant for standard robotic manipulation tasks, whereas vision and action pathways are substantially less tolerant to removal} }. On LIBERO, removing half of the LLM blocks even improves OpenVLA-OFT from 95.0% to 98.3% under the same downstream fine-tuning budget, and retaining only two language blocks still recovers baseline-level performance. These results suggest that current VLA benchmarks may exert limited pressure on deep language grounding and compositional instruction understanding, and that future VLA architectures should allocate capacity more deliberately across language, vision, and action components. The code is available at this https URL .\n\n---\n\n## 31. End-to-End Dynamic Sparsity for Resource-Adaptive LLM Inference\n- **Authors**: Yuhang Chen , Jinhao Duan , Ruichen Zhang , Mingfu Liang , Xiaohan Wei , Yunchen Pu , Fei Tian , Chonglin Sun , Parish Aggarwal , Frank Shyu , Luke Simon , Sandeep Pandey , Tianlong Chen , Xi Liu\n- **URL**: [https://arxiv.org/abs/2606.27743](https://arxiv.org/abs/2606.27743)\n- **Abstract**:\n> Large Language Models (LLMs) inference is typically deployed under a static resource assumption, where models execute a fixed computational graph regardless of the runtime environment. However, real-world cloud infrastructure is inherently dynamic, characterized by fluctuating availability (e.g., spot instance preemption) and tiered Quality-of-Service requirements. In such volatile settings, static models are inflexible: they either crash under resource constraints or waste compute on redundant operations. To bridge this gap, we propose Learning to Allocate (L2A), an end-to-end framework for resource-adaptive inference. Unlike prior methods that condition only on input difficulty, we formulate inference as a constrained allocation problem conditioned on both the input and the runtime resource budget itself. We introduce lightweight, budget-conditioned and input-aware gating networks integrated into the LLM. These gates are trained via a unified objective that jointly optimizes task performance, logical consistency, and resource costs along three axes matching how real-world dynamics manifest: layer skipping for memory and depth pressure, head pruning for throughput contention, and reasoning-token reduction for latency tightening. This lets the model learn a budget-aware policy beyond input difficulty alone: it adaptively configures its computational footprint with respect to real-time resource dynamics, maximizing reasoning depth when resources permit while enforcing strict frugality when budgets tighten. A single L2A model traces the entire compute-accuracy Pareto frontier on Llama-3-8B and Qwen-3-4B: at up to 34% realized layer sparsity, it stays within 0.6% of the dense baseline on GSM8K, with the same gap holding zero-shot on out-of-distribution tasks, while every static or heuristic baseline requires a separately tuned model and still drops by 5-10% at comparable inference time.\n\n---\n\n## 32. KG2Cypher: Data-Centric Pipeline for Building Enterprise Text-to-Cypher Systems\n- **Authors**: Minjun Choi , Yerin Kim , Junghyuk Seo , Sujin Mo , Hyemin Lee , Youngjoong Ko\n- **URL**: [https://arxiv.org/abs/2606.27742](https://arxiv.org/abs/2606.27742)\n- **Abstract**:\n> Enterprise Knowledge Graphs (KGs) are increasingly used for internal search, analytics, and question answering, but building natural-language interfaces for private enterprise graphs remains costly. We present KG2Cypher, a data-centric pipeline for building enterprise text-to-Cypher systems from existing KGs. KG2Cypher first constructs an executable Cypher query from observed graph facts and then uses LLMs to generate its associated natural-language question. The resulting Text-Cypher pairs are validated with an LLM judge and human validation, and are converted into candidate-aware SFT data. The trained generator is served with class-conditioned schema prompting, entity retrieval, and LoRA-based inference. We evaluate KG2Cypher in Korean enterprise settings, where short search-style queries and schema paraphrases make language grounding difficult. LoRA SFT improves execution-result F1 from 0.806 to 0.950 on broadcast-program queries and from 0.70 to 0.92 on company queries. In an 11-class setting, KG2Cypher achieves 95.2% exact match, 99.9% execution rate, and 0.964 execution-result F1.\n\n---\n\n## 33. Bifocal Diffusion Language Models: Asymmetric Bidirectional Context for Parallel Generation\n- **Authors**: Yuhang Chen , Xianfeng Wu , Jinhao Duan , Mingfu Liang , Xiaohan Wei , Yunchen Pu , Fei Tian , Chonglin Sun , Parish Aggarwal , Frank Shyu , Luke Simon , Sandeep Pandey , Xi Liu , Tianlong Chen\n- **URL**: [https://arxiv.org/abs/2606.27732](https://arxiv.org/abs/2606.27732)\n- **Abstract**:\n> Discrete diffusion language models (dLLMs) recover masked tokens in parallel, offering significant speedups over autoregressive (AR) generation. However, such promising frameworks face a fundamental architectural design dilemma: \\ding{182} Adopting bidirectional attention achieves strong generation quality by allowing each position to access the full context, but is inherently incompatible with KV caching, limiting inference throughput in batch-serving scenarios; \\ding{183} Conversely, causal attention enables efficient cached inference but loses all right-side context, substantially degrading generation quality. This paper introduces Bifocal dLLMs, a new paradigm that resolves this dilemma through \\emph{asymmetric bidirectional context}. Analogous to bifocal lenses, we instantiate the paradigm as \\textbf{R2LM} (Right-to-Left Mamba), which combines two complementary mechanisms: $a$) standard causal attention providing precise left-context with full KV cache compatibility, while $b$) a lightweight reverse Mamba SSM sidecar supplying compressed right-side context without breaking cacheability. Comprehensive experiments on continued pretraining of Qwen3-1.7B with 60B tokens demonstrate that R2LM achieves $2.4\\times$ to $12.9\\times$ higher throughput than bidirectional dLLMs and $1.9\\times$ to $2.9\\times$ speedup over AR baselines in batch serving through parallel decoding with KV caching, while exceeding the causal baseline on most benchmarks and surpassing the bidirectional dLLM on average.\n\n---\n\n## 34. Enhancing Numerical Prediction in LLMs via Smooth MMD Alignment\n- **Authors**: Zhuo Zuo , Li Yue , Wenhao Zheng , Chenpeng Wang , Xianggen Liu\n- **URL**: [https://arxiv.org/abs/2606.27731](https://arxiv.org/abs/2606.27731)\n- **Abstract**:\n> Despite their strong general capabilities, large language models (LLMs) often remain unreliable when outputs must be numerically precise. A key reason is the training objective: standard cross-entropy treats numeric tokens as unstructured categories and ignores the metric structure of their values. We address this mismatch with Smooth Maximum Mean Discrepancy (SMMD), which builds on the classic MMD by incorporating value-distance kernels over numeric tokens and graph-based smoothness. With this kernel defined over a numeric sub-vocabulary, SMMD aligns the predicted numeric distribution to the target via kernel matching and smooths the prediction-target residual over the induced kernel graph to encourage local consistency. We evaluate SMMD on four numeric-target tasks: mathematical reasoning, arithmetic calculation, clock-time recognition, and chart question answering, across multiple open-weight LLM and VLM backbones. SMMD consistently improves accuracy over both cross-entropy and recent numeric-target losses; analyses show complementary effects between MMD and smoothness and underscore the importance of distance-based kernel design. Code is available at this https URL .\n\n---\n\n## 35. Low-Agreeableness Persona Conditioning for Safe LLM Fine-Tuning\n- **Authors**: Austin MY Cheung , Yi Yang\n- **URL**: [https://arxiv.org/abs/2606.27709](https://arxiv.org/abs/2606.27709)\n- **Abstract**:\n> Recent work has shown that fine-tuning large language models (LLMs) for social warmth degrades factual reliability and increases sycophancy. We investigate a related but distinct failure mode: warmth fine-tuning also weakens adversarial safety, making models more susceptible to jailbreaks and harmful output generation. We examine whether this reflects an inherent consequence of empathetic adaptation or an artifact of data construction. To address this, we introduce a persona-driven rewriting pipeline that conditions user turns on low agreeableness and pairs this with warm, de-escalating assistant responses. Across three experiments on four models, our approach reduces jailbreak susceptibility and harmful output rates relative to generic warmth fine-tuning baselines, while preserving conversational warmth. Representational probing provides suggestive evidence that this conditioning reduces the geometric alignment between warmth and compliance directions in latent space. These results show that safer empathetic fine-tuning is achievable through data design alone, without safety labels, harm detectors, or changes to the training objective.\n\n---\n\n## 36. Mitigating LLM-based p-Hacking by Preregistering for the Next LLM\n- **Authors**: Maria Thomas , Kristina Gligoric , Nihar B. Shah\n- **URL**: [https://arxiv.org/abs/2606.27687](https://arxiv.org/abs/2606.27687)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used to generate, classify, and annotate data whose outputs feed downstream hypothesis tests. However, LLM-based research is easy to p-hack: a researcher can tune the prompts, decoding parameters, or output format until a desired result is reached. We propose a protocol to mitigate p-hacking in LLM-based research: preregistering the experiment and eligible models, and then running it on the first eligible LLM that is released after the preregistration. The researcher finalizes the procedure on current models, preregisters the analysis plan together with a set of eligible future models, and runs the confirmatory analysis on the first eligible model released afterward. Because this model does not exist at commitment time, it cannot be hacked against; furthermore, configurations that hack one model frequently do not transfer to the next. We evaluate the protocol on two tasks whose true values are known. Across 20 models from four providers and 11 LLM-analysis configurations, the protocol would have blocked successful transfer of the p-hack in 73.9% and 72.7% of cases in the two tasks. Additional analyses reveal that mitigation remains substantial under several stress tests. Finally, putting money where our mouth is, we followed our own protocol and preregistered our experiment. The preregistered experiment confirmed the protocol's effectiveness: out of the 7 configurations that hacked the prior model, the hacking failed to carry over in 6 configurations on the first eligible model released afterward.\n\n---\n\n## 37. CBD: API-Only LLM Black-Box Unlearning through Controlled Behavioral Divergence\n- **Authors**: Zhiqiang Xie , Yijing Lin , Zhipeng Gao , Dong In Kim\n- **URL**: [https://arxiv.org/abs/2606.27683](https://arxiv.org/abs/2606.27683)\n- **Abstract**:\n> Edge devices increasingly invoke large language models (LLMs) through API services for context aware edge intelligence, while edge generated data may be collected to improve LLMs and may introduce sensitive, copyrighted, harmful, or outdated information into model behavior. Machine unlearning offers a practical way to remove the influence of undesired data without retraining LLMs. However, existing methods still face two gaps. The first is API only black box access, where target model parameters and internal logits are unavailable. The second is how to preserve retained utility when unlearning target data and retained data share highly similar prompt structures or semantic patterns. To address these challenges, we propose Controlled Behavioral Divergence (CBD), an API only black box unlearning framework. CBD uses two auxiliary models to create controlled behavioral divergence between retained inputs and unlearning target inputs, converts this divergence into an unlearning relevance score, and routes unlearning related prompts away from the target LLM. To improve discrimination accuracy under high similarity between target and retained data, CBD constructs a gradient statistics based discriminative basis by estimating empirical Fisher matrices and solving a regularized generalized eigenvalue problem, guiding the unlearning signal toward target specific information rather than shared prompt structures. Compared with eleven white box and gray box unlearning baselines, CBD achieves a better unlearning utility trade off and its performance varies little across settings. On ToFU forget10, CBD approaches the retrained reference on the forget set while raising model utility to 74.90, about 15% above the second best baseline. On WMDP, it lowers hazardous knowledge accuracy to 25.68, near random guessing, while preserving MMLU accuracy of 52.67. Code is at this https URL .\n\n---\n\n## 38. From Signals to Transfer: A Factorised Study of Probe-Based Uncertainty Estimation in Large Language Models\n- **Authors**: Ponhvoan Srey , Xiaobao Wu , Cong-Duy Nguyen , Quang Minh Nguyen , Duc Anh Vu , Anh Tuan Luu\n- **URL**: [https://arxiv.org/abs/2606.27679](https://arxiv.org/abs/2606.27679)\n- **Abstract**:\n> Probe-based uncertainty estimation (UE) has emerged as a prominent approach to detect hallucinations in Large Language Models (LLMs) by learning uncertainty from internal model signals. Yet, recent methods vary simultaneously across feature design, training data construction, and evaluation setting, obscuring what actually drives performance. To address this issue, we propose a factorised study of probe-based UE under matched conditions. Our results show that raw hidden states and attention features are difficult to outperform in-domain. However, under distribution shift, structured and compressed features are more robust, suggesting that in-domain performance alone is insufficient to measure progress. Furthermore, prompting and label construction significantly affect probe behaviour. Building on these best-practice findings, we train benchmark-based pretrained probes that transfer reasonably well to open-ended factual generation, providing a stable off-the-shelf baseline. Our work encourages more deployment-oriented evaluation of probe-based uncertainty estimators. The code repository is available at this https URL .\n\n---\n\n## 39. HybridCodec: Modeling Discrete and Continuous Representations for Efficient Speech Language Models\n- **Authors**: Artem Ploujnikov , Francesco Verdini , Samir Sadok , Mirco Ravanelli\n- **URL**: [https://arxiv.org/abs/2606.27627](https://arxiv.org/abs/2606.27627)\n- **Abstract**:\n> Discrete audio representations have become increasingly popular for building multimodal text-audio systems and integrating audio capabilities into Large Language Models (LLMs). However, numerous studies report performance degradation on various downstream tasks due to information loss during discretization. To address this, we propose a novel approach combining temporally compressed discrete tokens with dimensionality-reduced continuous residuals. Our framework consists of a hybridized discrete-continuous focal modulation codec and a hybrid Transformer. This architecture performs autoregressive inference in the discrete domain, coupled with non-autoregressive prediction and continuous residual upsampling. Experimental results show that our approach significantly improves the retention of speaker characteristics compared to discrete-only methods, while simultaneously reducing the number of required autoregressive steps.\n\n---\n\n## 40. Dismantling Pathological Shortcuts: A Causal Framework for Faithful LVLM Decoding\n- **Authors**: Liu Yu , Can Chen , Ping Kuang , Zhikun Feng , Fan Zhou , Gillian Dobbie\n- **URL**: [https://arxiv.org/abs/2606.27596](https://arxiv.org/abs/2606.27596)\n- **Abstract**:\n> Large Vision-Language Models (LVLMs) exhibit sophisticated reasoning but remain susceptible to object hallucination. Deviating from the prevailing attention intensity assumption, we reveal a deeper dynamic structural misalignment: hallucination is triggered at decision-critical steps where specific attention heads, acting as risky mediators, decouple from visual evidence to lock onto language priors. This establishes a pathological shortcut that bypasses visual grounding. To dismantle this, we propose Fox (Faithfulness and Observational-flow via eXpression-rectification), a training-free inference-time framework. Fox diagnoses structural misalignment using a visual attention entropy probe to localize risky mediators unsupervisedly. We then execute a targeted causal intervention via numerical logit saturation to physically sever the shortcut path. Finally, a conflict-gated cooperative decoding strategy reconciles interventional faithfulness with observational fluency. Extensive experiments demonstrate that Fox achieves SOTA performance, outperforming SID by 29.1% while preserving linguistic richness. Code is available at this https URL .\n\n---\n\n## 41. On the Inseparability of Instructions and Data in Shared-Embedding Sequence Models\n- **Authors**: Dewank Pant , Shruti Lohani , Avijit Kumar\n- **URL**: [https://arxiv.org/abs/2606.27567](https://arxiv.org/abs/2606.27567)\n- **Abstract**:\n> Prompt injection is the top security risk for LLM-integrated applications, yet every defense proposed so far has been broken. We prove this is not a coincidence: in shared-embedding architectures that lack enforced control-data separation, perfect prompt-injection prevention is mathematically impossible. We formalize prompted systems as Prompted Action Models whose outputs include control-authoritative actions: refusal decisions, tool authorization, policy routing, and memory writes. We define Semantic-Faithful Control (SFC), the property that such behavior depends only on the meaning of untrusted input, not on how it is encoded. We then prove SFC is unachievable within the shared pipeline, via three results: a provenance-recovery impossibility (shared representations make trusted and untrusted content statistically inseparable, bounded by total variation distance); control-path exposure (untrusted tokens enter control-relevant computation through the same attention value-aggregation that determines outputs); and a finite-coverage invariance gap (finite training cannot certify invariance over infinite semantic-equivalence classes). We ground each quantity in measurements on production tokenizers and models. The result is structural, not a gap in current defenses. It mirrors the code-data confusion in Von Neumann machines that gives rise to buffer overflows, a vulnerability class that took decades of layered defenses (DEP, Write-XOR-Execute, ASLR, stack canaries, and ultimately memory-safe languages) to contain, because no single mechanism sufficed. The implication is the same: prompt injection cannot be eliminated by better in-pipeline classification or alignment alone. It requires architectural separation of instruction and data channels. We identify the root cause and the class of solution it demands.\n\n---\n\n## 42. Benchmarking Multi-Modal Graph-based Social Media Popularity Prediction\n- **Authors**: Utkarsh Sahu , Zhisheng Qi , Li Zhu , Yizhao Yang , Jun Li , Ryan Rossi , Yu Wang\n- **URL**: [https://arxiv.org/abs/2606.27539](https://arxiv.org/abs/2606.27539)\n- **Abstract**:\n> Social media popularity prediction aims to forecast the future reach or influence of online content from early-stage observations. Accurate prediction enables key downstream applications, such as advertising optimization and strategic content planning by users, creators, and platforms. Despite substantial progress, existing popularity prediction works often fail to jointly consider multimodal content and temporal social interaction signals. Moreover, the literature remains highly fragmented across datasets, modalities, observation windows, prediction targets, and evaluation protocols. This fragmentation prevents fair comparison and obscures a systematic understanding of how textual, visual, temporal, and interaction-based signals jointly shape popularity dynamics. To address these challenges, we introduce MMG-Pop, a Multi-modal Graph-based Popularity Prediction benchmark, which unifies datasets, modalities, temporal interaction signals, and representative baselines under a standardized evaluation protocol. Furthermore, we propose MMG-PopNet, a unified multi-modal graph-based network that jointly models the aforementioned multi-modal signals and graph-structured social interactions. Extensive experiments on MMG-Pop, comprising four datasets across Bluesky and Reddit platforms, demonstrate the superior performance of MMG-PopNet and yield new insights into cross-platform training generalization, multi-task prediction benefits, multi-modality contributions, and LLM prediction limitation. These findings establish a unified foundation for future research on social dynamics modeling and intervention under heterogeneous modalities and socially-aware agentic ecosystem paradigms.\n\n---\n\n## 43. Large Language Model Teaches Visual Students: Cross-Modality Transfer of Fine-Grained Conceptual Knowledge\n- **Authors**: Thomas Shih-Chao Liang , Zhuoran Yu , Yong Jae Lee\n- **URL**: [https://arxiv.org/abs/2606.27527](https://arxiv.org/abs/2606.27527)\n- **Abstract**:\n> Large Language Models (LLMs) possess broad conceptual knowledge acquired through large-scale text pretraining, yet their potential to supervise models in other modalities remains underexplored. In this work, we propose LaViD--Language-to-Visual Knowledge Distillation--a simple and effective framework for transferring high-level semantic knowledge from a language-only teacher to a vision-only student model. Instead of relying on paired multimodal data, LaViD elicits conceptual signals from an LLM by prompting it to generate multiple-choice questions (MCQs) that probe semantic distinctions between visual classes. Each class is mapped to a soft label distribution over these MCQs, forming a rich conceptual signature that guides the student through an auxiliary distillation loss. Notably, despite using a language-only teacher without access to image data, LaViD consistently outperforms recent methods like MaKD that distill from vision-language models across multiple fine-grained benchmarks. It also achieves competitive or superior performance compared to state-of-the-art visual distillation methods such as DKD and MLKD, with further gains when combined with logit standardization. On the Waterbirds dataset, LaViD substantially improves worst-group accuracy, demonstrating enhanced robustness to spurious correlations with distillation. Code is available at this https URL .\n\n---\n\n## 44. Speculative Refinement: A Hybrid Autoregressive Diffusion Decoding Strategy and Its Behavior Across Benchmarks\n- **Authors**: Aditi Gupta , Neel Mishra , Kushagra Trivedi , Pawan Kumar\n- **URL**: [https://arxiv.org/abs/2606.27474](https://arxiv.org/abs/2606.27474)\n- **Abstract**:\n> How should we evaluate generation systems that combine autoregressive (AR) and diffusion decoding? We study this question through Speculative Refinement (SpecRef), a training-free hybrid method that warm-starts a masked diffusion language model from an AR draft using entropy-guided selective masking. Evaluating SpecRef across six benchmarks (HumanEval, MBPP, GSM8K, BBH, ARC-Challenge, HellaSwag) with three distinct evaluation protocols (execution-based pass@1, exact-match, log-likelihood scoring), we surface several findings relevant beyond our specific system: (1) code benchmarks conflate structural discovery with logical correctness: providing a syntactic scaffold lifts accuracy from near zero to over 20% without changing the model, indicating that much of the baseline failure is structural; (2) a refinement tension phenomenon where multi-stage correction degrades already-correct tokens, exposing benchmark saturation ceilings invisible to single-model evaluation; (3) log-likelihood and generative evaluation produce different model rankings for the same model pair, suggesting they measure different capabilities; (4) standard Python post-processing silently breaks code evaluation for non-AR generators. These observations apply to any multi-stage or non-autoregressive generation pipeline and point toward more diagnostic evaluation practices.\n\n---\n\n## 45. Supersede: Diagnosing and Training the Memory-Update Gap in LLM Agents\n- **Authors**: Vedant Patel\n- **URL**: [https://arxiv.org/abs/2606.27472](https://arxiv.org/abs/2606.27472)\n- **Abstract**:\n> Large language model (LLM) agents operate over long, multi-session interactions in which facts change: a user moves, a price updates, a plan is revised. Acting correctly requires using the current value of a fact and discarding values that have been superseded. We isolate this ability on real conversational data and show that it is a distinct, unsolved failure. On the knowledge-update subset of LongMemEval, replacing an agent's full context with a bounded, self-maintained memory drops accuracy from 92% to 77% even on a frontier model (gpt-5.4), a gap that is statistically significant (paired McNemar p<0.005) and persists across model scale while full-context accuracy saturates near 92%. The bottleneck is therefore memory maintenance, not comprehension, and is not closed by a stronger model. We then ask whether this is merely an undersized memory, and find it is not: as the conversation grows 24x, accuracy falls further (from 68% to 28%), and granting the agent proportionally more memory yields no detectable recovery (28% to 28%, n=25). The failure scales with the length of the conversation, not the compression ratio. We release Supersede, an open reinforcement-learning environment (on the verifiers / prime-rl stack) that turns this measurement into a training signal: agents are rewarded for answering from the current value and penalized for stale ones. Finally, we close the loop and show the gap is trainable: GRPO fine-tuning a small open model (Qwen2.5-3B) on this environment nearly doubles its held-out supersession accuracy on real, unseen conversations (9.0% to 16.7%, a single run), along a monotonic checkpoint curve indicating the learned policy, not the harness, carries the gain. To our knowledge this is the first trainable environment whose reward targets temporal fact-currency, and the first evidence the supersession gap can be trained down, not only measured.\n\n---\n\n## 46. SidConArena: An Environment Evaluating Agents in Open-Ended,Positive-Sum Bargaining Game\n- **Authors**: Yeqi Feng , Yuxin Chen , Tianxing He\n- **URL**: [https://arxiv.org/abs/2606.27397](https://arxiv.org/abs/2606.27397)\n- **Abstract**:\n> Evaluating LLM agents requires dynamic environments that go beyond static reasoning and zero-sum games. Real-world economic interaction is often open-ended and mixed-motive: agents must negotiate, create positive-sum surplus, compete for scarce assets, and plan under delayed returns. We introduce SidConArena, a new benchmark framework for evaluating LLM agents in open-ended, positive-sum bargaining. SidConArena formalizes a multi-player economy as a finite-horizon partially observable stochastic game with three coupled phases: natural-language negotiation with binding trades, deterministic converter-based production, and sealed-bid auctions for long-term assets. The framework combines structured observations, phase-aware agent dispatching, a neural-symbolic action interface, and asynchronous execution, enabling free-form interaction while preserving rule-grounded evaluation. Across homogeneous and heterogeneous tournaments, stronger frontier models achieve higher economic outcomes, yet agents still misvalue resources, bargain passively, and remain limited in long-horizon investment planning.\n\n---\n\n## 47. Agentic Publication Protocol: An Attempt to Modernize Scientific Publication\n- **Authors**: Sirui Lu , Xiao-Liang Qi\n- **URL**: [https://arxiv.org/abs/2606.27386](https://arxiv.org/abs/2606.27386)\n- **Abstract**:\n> Scientific publication is still organized primarily around static manuscripts, even though much of scientific progress depends on tacit know-how: how to run code, reproduce figures, interpret edge cases, choose useful follow-up directions, and avoid failed paths. Large language model agents create an opportunity to publish not only knowledge, but also operational know-how in a form that future readers and researchers can directly use. This paper outlines the Agentic Publication Protocol (APP), a lightweight repository format for packaging a paper together with code, data, environment information, reproducibility instructions, and an agent-facing instruction file. APP treats a version-controlled repository as the publication object and uses \\texttt{ this http URL } and optional skills to define a paper agent that can explain the work, reproduce key results when possible, and support follow-up research. We describe the design principles and details of the protocol, as well as the agent skills useful for publishing papers under the protocol. We also describe development tools for evaluating and improving the protocol and associated agent skills. Finally, we provide a broader discussion of the future of scientific research in the agent era.\n\n---\n\n## 48. CalBrief: A Pilot Diagnostic Benchmark for Evidence-Calibrated Scientific Briefing with Large Language Models\n- **Authors**: Yu Fu , Yongqi Kang , Yong Zhao\n- **URL**: [https://arxiv.org/abs/2606.27383](https://arxiv.org/abs/2606.27383)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used as research assistants, yet it remains unclear whether they can calibrate research takeaways to the strength and scope of the supporting evidence. We study evidence-calibrated scientific briefing: given a bounded package of related papers, a system should generate package-level takeaways with evidence strength, scope boundaries, and missing-evidence caveats. We contribute a verified pilot benchmark of 16 heterogeneous scientific evidence packages and 96 human-verified takeaways, and we use CalBrief, an auditable role/gap/strength framework, as a diagnostic probe to locate where briefing breaks down. Under a fair-schema evaluation, structured organization improves role and gap reasoning, but an explicit strength-calibration policy is systematically over-conservative and falls below majority and direct-LLM baselines. To explain why, we run a controlled diagnostic across three closed-model backbones (GPT-4o, Claude Sonnet, Gemini Flash) that separates three potential causes of conservatism. Approximately 63% of the conservatism gap is attributable to expanding the label space from binary {moderate, weak} to four-way {moderate, weak, uncertain, insufficient_evidence} (p < 0.001 across all backbones); only 1% is attributable to gap/scope signal injection (not significant); the remaining 36% arises from the pipeline policy itself. We also find that 4-way predictions can be post-hoc collapsed back to binary and then match or exceed direct binary prompting, so the extra labels carry information that strict matching hides. Label-level strength judgment and auditable evidence organization are distinct abilities currently in tension, and should be evaluated separately for LLM research assistants.\n\n---\n\n## 49. Position: The Term \"Machine Unlearning\" Is Overused in LLMs\n- **Authors**: Sangyeon Yoon , Yeachan Jun , Albert No\n- **URL**: [https://arxiv.org/abs/2606.27379](https://arxiv.org/abs/2606.27379)\n- **Abstract**:\n> Large language models increasingly face demands to \"forget\" training data, knowledge, or behaviors due to regulatory deletion obligations, copyright/licensing disputes, and safety or product-policy requirements. This position paper argues that machine unlearning is overused as a term in LLM research and should be reserved for dataset-defined deletion: removing the training influence of a precisely specified forget set such that the resulting model is approximately indistinguishable from retraining without that data. We contend that many tasks currently labeled \"unlearning\" (e.g., refusal for harmful requests, entity/knowledge removal, or targeted suppression) pursue different, often policy-dependent objectives and therefore require different terminology and baselines (e.g., alignment, suppression, editing, obfuscation). We further argue that this confusion is not cosmetic: because papers make different implicit guarantees under the same label, metrics and benchmarks are frequently reused outside their intended scope, rewarding surface-level non-disclosure (e.g., low ROUGE/forget accuracy) even when retraining-equivalence is not tested and derived capabilities remain. We conclude by calling for stricter terminology tied to explicit guarantees and reference models, and for evaluations that match the claimed objective.\n\n---\n\n## 50. DataStates-LLM: Scalable Checkpointing for Transformer Models Using Composable State Providers\n- **Authors**: Avinash Maurya , M. Mustafa Rafique , Franck Cappello , Bogdan Nicolae\n- **URL**: [https://arxiv.org/abs/2601.16956](https://arxiv.org/abs/2601.16956)\n- **Abstract**:\n> The rapid growth of Large Transformer-based models, specifically Large Language Models (LLMs), now scaling to trillions of parameters, has necessitated training across thousands of GPUs using complex hybrid parallelism strategies (e.g., data, tensor, and pipeline parallelism). Checkpointing this massive, distributed state is critical for a wide range of use cases, such as resilience, suspend-resume, investigating undesirable training trajectories, and explaining model evolution. However, existing checkpointing solutions typically treat model state as opaque binary blobs, ignoring the ``3D heterogeneity'' of the underlying data structures--varying by memory location (GPU vs. Host), number of ``logical'' objects sharded and split across multiple files, data types (tensors vs. Python objects), and their serialization requirements. This results in significant runtime overheads due to blocking device-to-host transfers, data-oblivious serialization, and storage I/O contention. In this paper, we introduce DataStates-LLM, a novel checkpointing architecture that leverages State Providers to decouple state abstraction from data movement. DataStates-LLM exploits the immutability of model parameters during the forward and backward passes to perform ``lazy'', non-blocking asynchronous snapshots. By introducing State Providers, we efficiently coalesce fragmented, heterogeneous shards and overlap the serialization of metadata with bulk tensor I/O. We evaluate DataStates-LLM on models up to 70B parameters on 256 A100-40GB GPUs. Our results demonstrate that DataStates-LLM achieves up to 4$\\times$ higher checkpointing throughput and reduces end-to-end training time by up to 2.2$\\times$ compared to state-of-the-art solutions, effectively mitigating the serialization and heterogeneity bottlenecks in extreme-scale LLM training.\n\n---\n\n{% endraw %}\n"}, "c234ff9921f9be8d378a7558752264c977290646": {"t": 1786068098.7236004, "n": 84889, "h": "29e92ec4f78e9471", "c": "---\nlayout: default\ntitle: \"LLM 관련 주요 논문 - 2026-06-29\"\n---\n\n{% raw %}\n# LLM 관련 주요 논문 - 2026-06-29\n\n## 1. Tandem Reinforcement Learning with Verifiable Rewards\n- **Authors**: Difan Jiao , Raghav Singhal , Robert West , Ashton Anderson\n- **URL**: [https://arxiv.org/abs/2606.28166](https://arxiv.org/abs/2606.28166)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has significantly improved the reasoning capability of large language models, reaching expert or even superhuman performance in domains such as competition math. However, whether weaker agents and humans can actually harness this capability is far less certain, with RLVR documented to drift reasoning toward idiosyncratic patterns such as poor readability and language mixing. Tandem training is a recently introduced paradigm that targets this compatibility problem: a trained, stronger senior co-generates each rollout with a frozen, weaker junior, and the two are rewarded as a team, so the senior is pushed to reason in ways the junior can follow. Yet this paradigm has so far been demonstrated only in proof-of-concept settings, leaving open whether it scales to the long chains of thought of the modern RLVR pipeline. In this work, we propose Tandem Reinforcement Learning (TRL), which carries the tandem training paradigm into RLVR. In TRL, the senior and a frozen junior alternate stochastically to co-generate the reasoning, the resulting generation is rewarded, and the standard GRPO loss is applied to the senior. Training Qwen3-4B-Instruct on competition math, we find that TRL matches vanilla GRPO on solo reasoning capability while three properties emerge together from the same rollout structure: stronger handoff robustness with the junior, reduced distributional drift from the junior, and a chain-of-thought more legible to the junior. Our results demonstrate a promising route for RLVR with practical payoffs in multi-model communication and human compatibility.\n\n---\n\n## 2. JD Oxygen AI Item Center (Oxygen AIIC) V1: An Industrial-Scale LLM/VLM-Centric Solution for Item Understanding, Management, and Applications\n- **Authors**: Oxygen AIIC , Chan Long , Chao Liu , Chaofan Chen , Chaohui Dong , Chunyuan Guo , Danping Liu , Debin Liu , Deping Xiang , Fulai Xu , Guangyue Liu , Hao Li , Huichun Hu , Jian Yang , Jianan Wang , Jianbo Zhao , Jiaoyang Li , Jiaxing Wang , Jinglong Li , Jinjin Guo , Jun Fang , Jun Liu , Kai Zhou , Li Wang , Lili Gao , Liying Chen , Luning Yang , Mengdi Zhou , Pengzhang Liu , Qi Lv , Qianyun Wang , Qixia Jiang , Ruyue Li , Shimu Liang , Shuxing Wang , Sijie Zhang , Siqi Li , Tianhao Gao , Wang Ke , Weihu Huang , Wencan Lai , Wenjie Zhang , Xiaohui Zhang , Xiaojing Dong , Ya Liu , Yifeng Zhang , Yixiang Wang , Yongtai Zhang , Yongyi Liao , Zhaoru Chen , Zhen Chen , Zhiyong Ma , Zhiyuan Liu , Zhongwei Liu , Ziyan Xing\n- **URL**: [https://arxiv.org/abs/2606.28070](https://arxiv.org/abs/2606.28070)\n- **Abstract**:\n> this http URL , one of the world's largest e-commerce platforms, serves over 700 million active users and millions of merchants, with a catalog of tens of billions of SKUs. At this scale, high-quality, structured item knowledge underpins a better consumer experience, lower management costs, and higher operational efficiency-yet producing and serving it poses three industrial-scale challenges: fast-emerging concepts, high-quality knowledge production for massive SKUs, and diverse downstream requirements. To address these challenges, we present the JD Oxygen AI Item Center (Oxygen AIIC), an industrial-scale platform built on LLMs/VLMs for item-knowledge production and service. Oxygen AIIC is built around four core pillars: (i) ontology engineering driven by efficient human-AI collaboration, which supports the dynamic evolution and agile expansion of an ontology with millions of entries; (ii) a \"Semantic Search then Discrimination\"(S2D) knowledge identification architecture that, combined with throughput improvement strategies, enables scalable, extensible, and high-throughput AI Item Library production for tens of billions of SKUs; (iii) self-evolving item-understanding LLMs/VLMs that improve in a stable and controllable manner, enabling knowledge production with 94.2% precision and 82.8% recall; and (iv) a unified item tunnel that serves as the data and service hub. Oxygen AIIC now covers tens of thousands of JD categories and processes hundreds of millions of item updates per day on Huawei Ascend NPUs. It has accumulated hundreds of billions of item-knowledge assets. Deployed across core business scenarios-including search, recommendation, operations, category planning-Oxygen AIIC has delivered measurable gains at scale. Search-traffic coverage reaches 80.4%, item-information quality issues drop by 37%, the automated fill rate of core attributes during item listing exceeds 80%.\n\n---\n\n## 3. NormAct: A Benchmark for Hidden Social Norm Compliance in Embodied Planning\n- **Authors**: Shiyun Zhao , Xinwei Song , Tianyu Guo , Xiaomeng Gao , Mingyuan Liu , Xu Han , Yuanyuan Zhang , Zhenliang Zhang , Xue Feng , Bo Dai\n- **URL**: [https://arxiv.org/abs/2606.27826](https://arxiv.org/abs/2606.27826)\n- **Abstract**:\n> Multimodal large language models (MLLMs) are increasingly deployed as embodied planners in egocentric environments, where task success requires not only achieving instructed goals but also acting in socially appropriate ways. While explicit goals may render certain actions optimal, implicit social norms often impose hidden constraints. Existing evaluations typically focus on explicit goal achievement or direct norm knowledge, seldom assessing whether planners can infer and apply these hidden constraints within action sequences. We introduce NormAct, a benchmark for embodied social-norm interactions that evaluates plans on Goal Achievement, Norm Compliance, and overall Task Success. NormAct uniquely embeds hidden norms within ordinary tasks, testing whether models can realize them without explicit instruction. Experiments with state-of-the-art MLLMs (GPT-5.4, Claude Opus 4.7, Gemini 3 Pro) reveal a significant gap: models achieve explicit goals in 67.3\\% of cases, but comply with hidden norms in only 26.4\\%. Cue-condition experiments indicate that this gap stems not from a lack of general social knowledge, but from challenges in activating and grounding relevant norms in context. To address this, we propose NormPerceptor, a context-conditioned cue generator that infers scene-relevant norms prior to planning, increasing Task Success from 24.2\\% to 46.7\\%. Our results underscore the importance of enabling embodied agents to proactively detect hidden norms, ground them in visual evidence, and integrate them as action-planning constraints. Our benchmark is publicly available at this https URL .\n\n---\n\n## 4. Grounded Iterative Language Planning: How Parameterized World Models Reduce Hallucination Propagation in LLM Agents\n- **Authors**: Xinyuan Song , Zekun Cai\n- **URL**: [https://arxiv.org/abs/2606.27806](https://arxiv.org/abs/2606.27806)\n- **Abstract**:\n> World models for language agents come in two useful forms. An agent-based world model calls an LLM API and reasons flexibly in language, but its errors appear as hallucinated state changes that are hard to score with ordinary regression losses. A parameterized world model is a trained transition predictor; its errors are easier to measure with quantities such as NodeMSE, delta accuracy, and validity accuracy, but it is usually weaker as a standalone planner. We compare these two families on four graph-structured planning benchmarks and introduce operational hallucination metrics for the agent-based case. The comparison motivates \\textbf{Grounded Iterative Language Planning} (GILP), which trains only a small parameterized backbone and combines it with API-based agent reasoning. The backbone supplies valid actions, predicted state deltas, risk, and value; the LLM drafts an action and imagined delta; and a consistency gate asks for revision when the two disagree. On real GPT-4o-mini calls, GILP reduces hallucinated-state rate from 0.176 to 0.035. In calibrated simulator ablations, it raises success from 0.668 to 0.838 while adding only ~22% extra LLM calls.\n\n---\n\n## 5. Towards Reliable and Robust LLM Planning: Symbolic Feedback-Driven Iterative Self-Refinement Framework\n- **Authors**: Jiajing Zhang , Jiamei Jiang , Chenyang Zhang , Feifei Mo , Linjing Li , Daniel Zeng\n- **URL**: [https://arxiv.org/abs/2606.27757](https://arxiv.org/abs/2606.27757)\n- **Abstract**:\n> Large language models (LLMs) have attracted widespread attention from academia and industry, yet their deployment raises critical security concerns regarding robustness and reliability. Planning, a core component of intelligent behavior, remains challenging for LLMs, which often produce infeasible or incorrect solutions in long-horizon decision-making tasks due to inherent complexity. In this paper, we propose a symbolic feedback-driven iterative self-refinement framework to enhance the robustness and reliability of LLMs in long-horizon planning. Specifically, a natural language prompting mechanism is introduced to map logical symbols into natural language descriptions, enabling LLMs to better capture task constraints and semantics. We further design a symbolic verifier that identifies errors and converts them into corrective instructions interpretable by the LLM, thereby guiding self-refinement. In addition, we leverage a plan recognizer to infer goal reachability, facilitating more effective guidance toward desired goals. Empirical results demonstrate that the proposed framework consistently improves both feasibility and correctness in long-horizon planning tasks. This highlights its effectiveness in enhancing the reliability of LLM-based planning and potential to enable more trustworthy AI systems.\n\n---\n\n## 6. ToE: A Hierarchical and Explainable Claim Verification Framework with Dynamic Multi-source Evidence Retrieval and Aggregation\n- **Authors**: Zhaoqi Wang , Zijian Zhang , Kun Zheng , Zhen Li , Xin Li , Chunlei Li , Jiamou Liu\n- **URL**: [https://arxiv.org/abs/2606.27736](https://arxiv.org/abs/2606.27736)\n- **Abstract**:\n> The rapid spread of fake news poses increasing threats to information ecosystems, especially as AI-generated misinformation under Generative Engine Optimization (GEO) poisoning allows adversarially crafted content to be systematically surfaced by retrieval systems, contaminating LLM reasoning. In this paper, we propose Tree of Evidence (ToE), a hierarchical evidence reasoning framework for automated fact-checking that models each claim as a dynamically expanding argument tree. ToE integrates a reinforcement learning-driven multi-source retrieval agent, an evidence evaluation agent, and an argument tree aggregation algorithm to iteratively decompose, retrieve, and verify claims through an explainable evidence chain. We further provide a theoretical analysis of the retrieval process, deriving a formal error bound that guarantees the learned policy converges to a neighborhood of the information-theoretically optimal policy. Experiments across multiple datasets and backbone LLMs demonstrate that ToE achieves improvements ranging from 4 to 24 percentage points over competitive baselines, with particularly pronounced gains on adversarially poisoned inputs.\n\n---\n\n## 7. DysLexLens: A Low-Resource LLM Framework for Analysing Dyslexic Learners Insights from Online Forums\n- **Authors**: Dana Rezazadegan , Atie Kia , Phongpadid Nandavong , Dominique Carlon , Jeremy Nguyen , Abhik Banerjee , James Marshall , Anthony McCosker , Yong-Bin Kang\n- **URL**: [https://arxiv.org/abs/2606.27619](https://arxiv.org/abs/2606.27619)\n- **Abstract**:\n> Dyslexic learners increasingly use artificial intelligence (AI) tools to support reading, writing, organisation, and study-related tasks. However, their lived experiences with these tools remain largely underexamined. This paper proposes DysLexLens, a low-resource LLM framework, designed to analyse dyslexic learners experience with AI through online forum discussions. DysLexLens is designed as an end-to-end, evidence-traceable architecture which transforms noisy social media posts into a dictionary-driven corpora, provides knowledge-graph (KG)-based question reasoning, generates verifiable query responses, and enables response evaluation through quantitative and human-grounded assessment. DysLexLens has four key features. First, it employs a dictionary-driven filtering method to construct a more focused Reddit corpus on dyslexia and AI, filtering out noisy and weakly related posts to improve the relevance of data collected from low-resource forum contexts. Second, it integrates LLM-assisted semantic analysis with KG-based query reasoning to uncover meaningful patterns. Third, it has quantitative evaluation metrics (RAGAS and Query Robustness) to measure LLM-generated response performance. Fourth, it provides structured qualitative validation guidelines for assessing response quality, with a specific focus on hallucination and evidence alignment. We demonstrate the effectiveness of DysLexLens using dyslexia-related Reddit forum data and 30 questions. The results show its potential generalisability to other low-resource forum data contexts. DysLexLens, sample data, questions and evaluation results are available at Github to support reproducibility.\n\n---\n\n## 8. Odyssey: Constructing Verifiable Local Truth-Preserving Foundation Models\n- **Authors**: Sridhar Mahadevan\n- **URL**: [https://arxiv.org/abs/2606.27593](https://arxiv.org/abs/2606.27593)\n- **Abstract**:\n> We introduce a categorical framework called ODYSSEY for constructing verifiable, local truth-preserving foundation models as compositions of foundries: building-block architectural components that specify a cover of local contexts, local representation families, restriction maps, gluing rules, obstruction policies, update obligations, and human-facing views. A foundry is an organized sheaf of knowledge that carries within it an argumentation component. Concrete foundries are built from generic foundries such as evidence/argument, operational decision, institutional/financial, market meaning, scientific challenge, research-program, assistant-build, and evaluation-harness foundries. Universal Foundry Learning (UFL) formalizes foundry construction as a composition of left and right Kan extensions, with left Kan extension rolling local artifacts into candidate foundries and right Kan extension enforcing the restriction, gluing, obstruction, and argumentation conditions required for promotion. Foundry SQL (FSQL) is a small typed query surface for slicing maintained foundry artifacts that uses TICKET (Topos Integration using Causal Kan Extension Transformers) certification for admitting external or pre-built models into durable ODYSSEY state. ODYSSEY is fully implemented and tested across a wide spectrum of concrete foundries, showing that the same categorical machinery supports domain construction, artifact replay, sheaf diagnostics, grounded Toulmin/local-LLM scrutiny, residual-obstruction ledgers, and optimized TICKET-compatible causal-claim extraction across heterogeneous sources. This paper is to be presented as a 2.5 hour tutorial at ICML 2026. The tutorial home page is at this https URL .\n\n---\n\n## 9. Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning\n- **Authors**: Xuan Zhang , Zhijian Zhou , Lingfeng Qiao , Yulei Qin , Ke Li , Xing Sun , Xiaoyu Tan , Chao Qu , Yuan Qi\n- **URL**: [https://arxiv.org/abs/2606.27483](https://arxiv.org/abs/2606.27483)\n- **Abstract**:\n> Large language model (LLM) agents have demonstrated strong capability in sequential decision-making, yet they remains fundamentally reactive in long-horizon tasks. Unlike humans who employ \"what-if\" reasoning to evaluate potential plans before commitment, standard agents lack an internal world model to simulate future outcomes. Therefore, we propose to internalize future-aware planning by training a single autoregressive model to verbalize both a prospective state rollout and a plan-conditioned success estimate-a textual analogue of the Q-value. Crucially, we identify a format-capability gap: simply fine-tuning agents on look-ahead traces during post-training leads to superficial mimicry of foresight without genuine predictive grounding. To bridge this gap, we introduce a three-stage training paradigm: (i) World Model Agentic Mid-Training (WM-AMT) to inject latent predictive capabilities into the policy; (ii) Format-Eliciting SFT (FE-SFT) to structure this injected capability; and (iii) Foresight-Conditioned Reinforcement Learning (FC-RL) to refine the calibration and utility of the generated simulations. Evaluated on search and mathematical reasoning tasks, our approach consistently outperforms other training baselines. Our results demonstrate that effective internal world modeling in LLM agents requires a capability-first training pipeline to achieve grounded and calibrated foresight.\n\n---\n\n## 10. When Does Personality Composition Matter for Multi-Agent LLM Teams?\n- **Authors**: Aryan Keluskar , Amrita Bhattacharjee , Huan Liu\n- **URL**: [https://arxiv.org/abs/2606.27443](https://arxiv.org/abs/2606.27443)\n- **Abstract**:\n> Personality prompting shapes how large language models communicate, yet whether these behavioral shifts affect objective task outcomes remains under-explored. Prior work shows that agents prompted with low agreeableness produce adversarial language, while those prompted with high agreeableness become cooperative, but the relationship between communication style and task performance has not been systematically examined across multiple domains. In this work, we investigate whether personality composition matters for multi-agent team performance by manipulating personality traits across frontier LLMs on three task domains: structured coding, open-ended research collaboration, and competitive bargaining. We find that personality effects depend critically on task structure. In coding tasks, low agreeableness leads to large communication shifts that have little effect on milestone completion. In open-ended collaboration and bargaining, the same manipulation substantially degrades performance. We discuss implications for multi-agent system design and the limits of personality manipulation.\n\n---\n\n## 11. AI-Model Network: Concept, Current State and Future\n- **Authors**: Li Zhetao , Zeng Xiyu , Wang Jianhui , Xiao Yong , Liu Zhongren , Wu Junru , Lai Junjie , Huang Jijun , Long Saiqin\n- **URL**: [https://arxiv.org/abs/2606.27382](https://arxiv.org/abs/2606.27382)\n- **Abstract**:\n> While the primary function of computers lies in computation and processing, the core value of the Internet is rooted in sharing and collaboration. Computers create the Internet, and the Internet empowers the value of computers. The rapid development of the Internet, cloud computing, and big data is pushing artificial intelligence into the era of large models (LMs). However, the practical application of LMs is currently hindered by high training costs and deployment complexities, driving a shift toward lightweight, private, and domain-specific models. With the rapid proliferation and wide distribution of heterogeneous models, enabling effective interaction and collaboration among them has emerged as a critical bottleneck that urgently needs to be addressed in LM development. Drawing inspiration from the development of the Internet, this paper proposes the concept, vision, and system architecture of world wide AI-model network (AI-ModelNet). It is a novel paradigm that achieves interconnection, capability sharing, and collaborative reasoning by establishing pathways between models. We first briefly review the current state of single-model and multi-model research. Subsequently, the systemic vision and hierarchical architecture of AI-ModelNet are articulated, followed by validation of the framework's feasibility through a prototype system and diverse application cases. Finally, key directions for future research are discussed preliminarily.\n\n---\n\n## 12. Cognitive Episodes in LLM Reasoning Traces Enable Interpretable Human Item Difficulty Prediction\n- **Authors**: Chenguang Wang , Ming Li , Xinyue Zeng , Zhuochun Li , Hong Jiao , Tianyi Zhou , Dawei Zhou\n- **URL**: [https://arxiv.org/abs/2606.28186](https://arxiv.org/abs/2606.28186)\n- **Abstract**:\n> Predicting human item difficulty is central to educational assessment, where reliable estimates support fairness and effective test construction. Existing methods often depend on costly human calibration or item-level textual representations, providing limited evidence about the cognitive processes that make items difficult. We argue that difficulty should be viewed not only as a property of item text, but also as an observable consequence of the problem-solving burden an item induces. Large Reasoning Models (LRMs) offer scalable process evidence through reasoning traces, but such evidence must be structured to support interpretable modeling. To this end, we introduce Epi2Diff (Episode to Difficulty), a framework that maps LRM reasoning traces into cognitively grounded episode sequences. These episodes group trace segments into functional problem-solving states, enabling difficulty to be modeled through reasoning scale, effort allocation, and state transitions. Epi2Diff extracts compact episode-dynamic features and combines them with semantic item representations for human difficulty prediction. Experiments on four real-world human difficulty datasets show that Epi2Diff consistently outperforms strong baselines, including fine-tuned small language models, LLM in-context learning, and supervised LLM adaptation. On SAT-derived classification benchmarks, Epi2Diff achieves an 8.1% average relative gain over supervised LLM fine-tuning baselines. Further analyses show that harder items induce more effortful, iterative, and implementation-centered episode dynamics, rather than merely longer responses. These results demonstrate that cognitive episodes in LRM reasoning traces provide a predictive and interpretable process representation for human item difficulty, offering a new lens for educational measurement with reasoning models.\n\n---\n\n## 13. LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior\n- **Authors**: Qinhong Zhou , Chuang Gan , Anoop Cherian\n- **URL**: [https://arxiv.org/abs/2606.28182](https://arxiv.org/abs/2606.28182)\n- **Abstract**:\n> Embodied agents operating in decentralized and partially observable environments have attracted growing attention in recent years. However, existing large language model (LLM)-based agents often exhibit behaviors that are misaligned with their partners or inconsistent with the environment state, leading to inefficient cooperation and poor task success. To address this challenge, we propose a novel framework, Learning Laws of Cooperation (LLawCo), that enables embodied agents to autonomously align with both their partners and task objectives. Our framework allows agents to reflect on past failures to extract misaligned behavioral patterns, which are used to derive high-level behavioral laws, such as \"Talk when necessary\" and \"Wait for partner.\" These laws are explicitly incorporated into the agents' chains of thought via supervised fine-tuning, aligning their reasoning with task requirements and the behavior of other agents. To evaluate our approach, we introduce PARTNR-Dialog, a large-scale multi-agent communicative and cooperative planning benchmark built on the PARTNR environment. Experiments on existing tasks and our new benchmark demonstrate significant improvements in cooperative efficiency and task success rates. Across four backbone LLMs, our method achieves average success rate improvements of 4.5% on the PARTNR-Dialog benchmark and 6.8% on the TDW-MAT benchmark over state-of-the-art open-source communicative agent frameworks. See the LLawCo project page for details: this https URL\n\n---\n\n## 14. Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models\n- **Authors**: Yanchen Yin , Dongqi Han , Linghui Li\n- **URL**: [https://arxiv.org/abs/2606.28153](https://arxiv.org/abs/2606.28153)\n- **Abstract**:\n> Jailbreak attacks bypass LLM safety alignment, yet their mechanisms remain poorly understood. We provide evidence that attacks do not comprehensively eliminate safety features, but instead selectively suppress specific attention heads. We identify two functionally differentiated types: Adversarially Compromised Heads (ACHs) concentrated in early layers, which are suppressed under attacks, and Safety-Aligned Heads (SAHs) in mid-layers, which maintain robust activations even when attacks succeed. Ablation studies support the causal role of ACHs and the contribution of SAHs to robust activations: suppressing a small number of ACHs is sufficient to induce jailbreak-like behavior on normally refused inputs, while removing SAHs substantially weakens mid-layer safety activations. Token-level attribution further shows that ACH suppression is driven specifically by attack-template tokens, providing a mechanistic account of why attacks can bypass refusal decisions through ACH suppression while leaving internal safety signals sustained by SAHs -- a phenomenon we term Robust Harmful Features. To validate the practical significance of this robustness, we show that simply reading these persistent activations -- without any training -- yields competitive aggregate detection performance with strong adversarial robustness.\n\n---\n\n## 15. From Tokens to States: LLMs as a Special Case of World Models and the Continuous Path Beyond\n- **Authors**: Paul Dubois\n- **URL**: [https://arxiv.org/abs/2606.28127](https://arxiv.org/abs/2606.28127)\n- **Abstract**:\n> The AI community has framed the relationship between large language models (LLMs) and world models as a dichotomy: LLMs predict tokens; world models simulate reality. Yann LeCun argues in 2022 that reaching general intelligence requires abandoning autoregressive token prediction in favour of latent-space architectures. This framing is unnecessarily binary. Two claims will be defended. First, LLMs are a degenerate special case of world models: the state space is the set of all token sequences, the only action is appending one token, and world models are therefore a strict generalisation of LLMs, not a replacement. Second, there is a natural continuous spectrum from NTP to JEPA, with multi-token prediction, future-summary prediction, and next-latent prediction as intermediate stations already populated by current research. Moving along this spectrum relaxes the LLM constraints one by one. It also progressively surrenders the two practical advantages that make LLMs trainable at scale: internet-scale self-supervised data, and a transformer architecture co-designed for discrete token prediction. Both are examined as open research questions: the data question (the cliff from self-supervised text to instrumented action-labelled environments) and the architecture question (whether the transformer generalises to continuous-state prediction, or whether a new primitive is needed).\n\n---\n\n## 16. Single and Multi Truth Data Fusion using Large Language Models\n- **Authors**: Hira Beril Kucuk , Norman W Paton , Jiaoyan Chen , Zhenyu Wu\n- **URL**: [https://arxiv.org/abs/2606.28062](https://arxiv.org/abs/2606.28062)\n- **Abstract**:\n> Data fusion, also known as truth discovery, is a data integration problem that aims to determine the correct value or set of values for each attribute of an object when presented with potentially conflicting values from multiple sources. Data fusion tasks belong to two main categories: single-truth scenarios, where each attribute has only one correct value, and multi-truth scenarios, where multiple values can be valid simultaneously. This paper investigates the use of Large Language Models (LLMs) in data fusion tasks for tabular data. Various prompting strategies, encompassing both single-truth and multi-truth scenarios, are investigated empirically. Domain-dependent, domain-independent, zero-shot and one-shot prompts are evaluated on three different benchmark datasets. Experimental results demonstrate that LLM-based approaches outperform traditional unsupervised truth discovery methods, such as DART and LTM, across all datasets. The codebase of this study has been made publicly available on GitHub.\n\n---\n\n## 17. ToolPrivacyBench: Benchmarking Purpose-Bound Privacy in Tool-Using LLM Agents\n- **Authors**: Shijing Hu , Liang Liu , Zhu Meng , Zhicheng Zhao\n- **URL**: [https://arxiv.org/abs/2606.28061](https://arxiv.org/abs/2606.28061)\n- **Abstract**:\n> Large language models (LLMs) have increasingly moved from standalone text generation systems to agents that invoke external tools, access environments, and execute multi-step tasks. However, conventional function-calling benchmarks mainly evaluate task completion and API correctness, while privacy evaluation benchmarks typically focus on final responses or privacy judgments. Neither perspective captures purpose-bound information flow across an executed multi-tool trajectory. Motivated by this limitation in current agent evaluation, ToolPrivacyBench audits whether task-private atoms are routed only to authorized tools and downstream sinks, thereby evaluating both task completion and privacy over-disclosure during tool use. The benchmark contains 2,150 cases, including 1,150 fully synthetic privacy-sensitive business workflows and 1,000 cases adapted from existing multi-tool and function-calling benchmarks. Each case is represented by a policy knowledge base. After an agent executes against mock business backends, the evaluator compares recorded tool arguments and backend audit logs with this policy knowledge base. The evaluation covers nine widely used agents to characterize purpose-bound privacy over-disclosure. The results show that successful tool execution does not imply appropriate privacy disclosure: an agent may complete a task while transmitting unnecessary private information through intermediate tool calls. ToolPrivacyBench therefore formalizes a need-to-know disclosure boundary, under which each tool should receive only the information necessary for its stated purpose, and uses trajectory-level auditing to identify privacy over-disclosure in multi-tool workflows.\n\n---\n\n## 18. MultiHashFormer: Hash-based Generative Language Models\n- **Authors**: Huiyin Xue , Atsuki Yamaguchi , Nikolaos Aletras\n- **URL**: [https://arxiv.org/abs/2606.28057](https://arxiv.org/abs/2606.28057)\n- **Abstract**:\n> Language models (LMs) represent tokens using embedding matrices that scale linearly with the vocabulary size. To constrain the parameter footprint, prior work proposes hashing many tokens into a single vector within encoder-only models. While this offers parameter efficiency, many-to-one collisions prevent its use in causal LMs. In this paper, we propose MultiHashFormer, a new framework that allows hash-based autoregression. Each token is represented as a unique hash signature, a short sequence of discrete hash IDs, generated by multiple independent hash functions. A Hash Encoder compresses this signature into a single latent vector for processing by a Transformer decoder. Then, a Hash Decoder generates the hash signature of the next token, which is then mapped back to text. We evaluate our approach at the 100M, 1B and 3B parameter scales, demonstrating that MultiHashFormer consistently outperforms standard Transformer LMs across multiple benchmarks. Furthermore, we show that our model handles multilingual vocabulary expansion with a constant parameter footprint without any modifications.\n\n---\n\n## 19. Can LLMs Judge Better Than They Generate? Evaluating Task Asymmetry, Mechanistic Interpretability and Transferability for In-Context QA\n- **Authors**: Sambaran Bandyopadhyay\n- **URL**: [https://arxiv.org/abs/2606.28050](https://arxiv.org/abs/2606.28050)\n- **Abstract**:\n> LLM-as-a-Judge and self-evaluation pipelines implicitly assume that evaluation is easier than generation. We test this in a controlled in-context QA setting where a context passage is the sole information source and each model judges the answer it generated, removing the parametric-knowledge confound of open-domain comparisons. Across four benchmarks (SQuAD 2.0, DROP, HotpotQA, MuSiQue) and two models, evaluation is not uniformly easier: generation accuracy exceeds self-evaluation on three of four, with multi-hop MuSiQue the exception. Attention analysis reveals why: evaluation attends to context 3--5x less than generation does and barely reads the candidate answer. LoRA fine-tuning confirms the asymmetry is not a training artifact: generation fine-tuning induces over-acceptance and evaluation fine-tuning degrades generation. These findings challenge core assumptions in self-evaluation pipelines.\n\n---\n\n## 20. Dialogue to Detection: A Multimodal Hybrid NLP Pipeline for Insurance Fraud Detection\n- **Authors**: Muhammad Shakeel Akram , Amal Htait , Abdul Hamid Sadka , Emma Meisingseth , Karishma Jaitly\n- **URL**: [https://arxiv.org/abs/2606.28002](https://arxiv.org/abs/2606.28002)\n- **Abstract**:\n> Insurance fraud imposes substantial financial losses and operational inefficiencies, raising premiums and impacting trust among legitimate policyholders. Early detection at FNOL remains a persistent challenge. Existing approaches rely largely on private, text-only datasets, limiting progress on multimodal methods that integrate linguistic, behavioural, and speaker-based indicators. We introduce a synthetic multimodal framework that replicates FNOL conditions. It generates agent-customer dialogue transcripts and two-speaker audios, performs ASR and diarisation. Downstream modules combine NER, regex-based feature extraction, LLM-RAG retrieval, and speaker embeddings in a rule-based risk score to flag narrative reuse, structural inconsistencies, and cross-case voice repetition while balancing sensitivity and false positives. Dataset validation and component-level evaluations show stability and transfer potential, offering a reproducible baseline beyond text-only fraud detection.\n\n---\n\n## 21. From Black-Box to Clinical Insight: A Multi-Stage Explainable Framework for Speech-Based Cognitive Impairment Detection\n- **Authors**: Yasaman Haghbin , Sina Rashidi , Ali Zolnour , Fatemeh Taherinezhad , Ali Fartoot , Hossein Azadmaleki , James M Noble , Maryam Dadkhah , Maryam Zolnoori\n- **URL**: [https://arxiv.org/abs/2606.27973](https://arxiv.org/abs/2606.27973)\n- **Abstract**:\n> Speech-based cognitive impairment detection offers a noninvasive, accessible alternative to costly biomarker assays, yet transformer-based models remain clinically uninterpretable. We propose a multi-stage explainability framework that translates black-box transformer predictions into clinically grounded narratives by integrating SHapley Additive exPlanations (SHAP)-based token attribution, theory-informed linguistic features, and a four-stage LLM reasoning pipeline using LLaMA-3.1-70B-Instruct. Built on the SpeechCARE-Adaptive Gating Network multimodal screening model (F1 = 72.11% on the NIA PREPARE benchmark), the framework maps model outputs to four cognitive-linguistic dimensions, including lexical richness, syntactic complexity, and semantic coherence. Physician evaluation on 70 stratified English samples demonstrated strong alignment with patient-level cognitive profiles, and a System Usability Scale score of 82/100 indicated high potential for clinical workflow integration.\n\n---\n\n## 22. Two-Stage Fine-Tuning for Protein Sequence Generation with Targeted Amino-Acid Composition\n- **Authors**: Violeta Basten-Romero , Rubén Muñoz-Tafalla , Anna María Díaz-Rovira , Bertran Miquel-Oliver , Isaac Filella-Merce , Víctor Guallar\n- **URL**: [https://arxiv.org/abs/2606.27939](https://arxiv.org/abs/2606.27939)\n- **Abstract**:\n> Protein language models are standard priors for biological sequence generation, but steering them toward explicit distributional design targets remains largely unexplored. We study a constrained protein generation problem in which sequences must match a desired amino-acid (AA) composition profile while preserving plausible sequence statistics and diversity. The motivating application is synthetic feed protein design, where the AA composition of dietary proteins directly determines their nutritional value. We propose a two-stage pipeline in which domain-adaptive fine-tuning (FT) on an in-domain protein dataset is followed by iterative reward-weighted FT via reinforcement learning (RL) anchored against the FT model as a frozen reference. We evaluate the pipeline on two AA compositions and find that FT brings the average composition close to the target, while the subsequent RL enforces specific sequence constraints that FT alone cannot satisfy. We additionally evaluate the design choices of the proposed composition reward term against two baselines and an ablated variant, isolate the contribution of each training stage, and verify that AA composition alignment is achieved without degrading sequence quality.\n\n---\n\n## 23. Agentic AI-Powered Re-Identification: An Emerging, Scalable Threat to Mobility Microdata Privacy\n- **Authors**: Oscar Thees , Roman Müller , Matthias Templ\n- **URL**: [https://arxiv.org/abs/2606.27936](https://arxiv.org/abs/2606.27936)\n- **Abstract**:\n> The widespread collection of fine-grained location data by commercial data brokers creates a re-identification risk that is not widely recognised by the public. While prior research has established that mobility traces are highly unique and that individuals can, in principle, be identified from a handful of spatio-temporal points, such attacks have historically required significant manual effort from skilled analysts, limiting their practical scale. In this feasibility study, we demonstrate in a real world setting that agentic AI fundamentally changes this threat model. We present an end-to-end pipeline in which large language model agents autonomously search the open web, cross-reference public records and social media, and resolve raw coordinate sequences to candidate identities - without human intervention. We evaluate the pipeline on a spatio-temporal dataset containing simulated location points anchored at and around true home and work addresses, focusing on a high-risk disclosure scenario. Our results demonstrate that, from spatio-temporal data and public sources alone, our agentic AI successfully re-identified 18 of the 25 re-identifiable individuals (72%) and 18 of 43 cases overall (41.9%). We discuss implications for Statistical Disclosure Control (SDC) practice and outline the near-future escalation that data custodians and regulators must anticipate. De facto anonymity - an implicit foundation of SDC practice - is shifting. Agentic AI strengthens the case that re-identification is reasonably likely by any means under the GDPR Recital-26 standard, at costs of minutes-and-dollars per target.\n\n---\n\n## 24. Triadic Werewolf: A Jester Role for Multi-Hop Theory of Mind in LLMs\n- **Authors**: Avni Mittal\n- **URL**: [https://arxiv.org/abs/2606.27909](https://arxiv.org/abs/2606.27909)\n- **Abstract**:\n> Theory-of-mind evaluations of large language models typically use dyadic social-deduction games, where every observable cue points to a single hidden side, so a model with strong language priors can score well without ever simulating opponents' incentives. We extend the Werewolf game with a Jester, a third faction whose utility on peer suspicion is inverted because it wins by being voted out, so optimal play requires reasoning across three opposing utility functions. Across 60 games on GPT-4.1, DeepSeek-V3.1, and Llama-3.3-70B with Jester self-learning on and off, the Jester wins 60-70% of games while Werewolves never exceed 20%, and GPT-4.1 wolves vote the Jester out on day 1 in 60-70% of games, a strictly self-defeating action. Self-learning helps DeepSeek and Llama but hurts GPT-4.1, with the cost landing on Villagers rather than Werewolves. Only DeepSeek learns the subtle strategy of looking suspicious without looking intentionally suspicious, and it gains the most from the loop. Triadic incentive structure exposes a layer of multi-agent reasoning that dyadic deduction games leave invisible.\n\n---\n\n## 25. A Study of Temporal Fusion Strategies for Named Entity Recognition in Historical Texts\n- **Authors**: Emanuela Boros\n- **URL**: [https://arxiv.org/abs/2606.27881](https://arxiv.org/abs/2606.27881)\n- **Abstract**:\n> Temporal variation poses a unique challenge for named entity recognition (NER) in historical texts, where entities drift in surface form and salience across time. While language models (LMs) have made progress in various NLP tasks, their ability to reason about temporality, especially in diachronic contexts, remains limited or at least, questionable. In this paper, we systematically study how temporal metadata can be structurally embedded into NER models using a range of lightweight fusion strategies. We experiment with both absolute and relative temporal representations, injected into Transformer-based architectures via early or late fusion mechanisms such as cross-attention, adapters, and concatenation. Our evaluations on French and German historical datasets reveal that late fusion strategies yield more robust and temporally generalisable performance, particularly in early and noisy periods.\n\n---\n\n## 26. SpatialUAV: Benchmarking Spatial Intelligence for Low-Altitude UAV Perception, Collaboration, and Motion\n- **Authors**: Haoyu Zhang , Meng Liu , Qianlong Xiang , Kun Wang , Yaowei Wang , Liqiang Nie\n- **URL**: [https://arxiv.org/abs/2606.27876](https://arxiv.org/abs/2606.27876)\n- **Abstract**:\n> Spatial intelligence is essential for low-altitude unmanned aerial vehicle (UAV) perception, collaboration, and navigation. However, existing UAV benchmarks often emphasize image-level recognition, single-view understanding, or narrow answer formats, leaving 3D spatial inference, multi-view collaboration, scene dynamics, and diverse task formulations insufficiently evaluated. To address these gaps, we introduce SpatialUAV, a real low-altitude UAV benchmark comprising 4,331 curated instances across 14 fine-grained task types, covering semantic discrimination, spatial relation, aerial--aerial collaboration, aerial--ground collaboration, and motion understanding. SpatialUAV organizes all samples into a unified visual-input--question--answer schema, while supporting seven input configurations and nine answer formats, including option labels, region identifiers, geometric values, cross-view correspondences, and free-form motion descriptions. To ensure reliable and grounded evaluation, our data construction pipeline integrates detector-assisted regions, depth supervision, metadata-derived rules, extensive manual annotation, blind filtering, and multi-turn human validation, together with task-specific metrics for heterogeneous outputs. Evaluating representative vision-language models across three categories, we show that current models remain far from human-level performance, with pronounced bottlenecks in cross-view association, structured grounding, geometric reasoning, and temporal viewpoint understanding. These results offer empirical guidance for advancing low-altitude UAV spatial intelligence. Code and data are available at this https URL .\n\n---\n\n## 27. Position Bias Correction is Insufficient for One-Pass Attention Sorting\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27793](https://arxiv.org/abs/2606.27793)\n- **Abstract**:\n> Long-context language models suffer from position bias, where information in middle positions is underutilized. Attention Sorting addresses this by iteratively reordering documents based on attention patterns, but its multiple sort-and-generate cycles increase deployment cost. We hypothesize that position bias is the primary bottleneck and propose Debiased One-Pass Attention Sorting, which estimates a per-prompt position-bias curve from the low-attention majority of documents and uses it to correct raw attention scores (via subtraction or division) to enable single-pass sorting. Our experiments on two models refute this hypothesis in the tested setting: on LLaMA-2-7B-32K-Instruct, debiasing produces identical results to uncalibrated single-pass sorting (94.83\\% containment accuracy), while on YaRN-Llama-2-7b-64k, debiasing improves accuracy by 8.67 percentage points but remains 14.84pp behind iterative sorting, closing only 37\\% of the gap. These results suggest that position-bias correction is insufficient to match iterative sorting, and that repeated reordering provides additional benefits beyond bias correction.\n\n---\n\n## 28. NLL-Guided Full-Attention Layer Selection for Training-Free Sliding-Window Adaptation\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27791](https://arxiv.org/abs/2606.27791)\n- **Abstract**:\n> Hybrid attention models that mix full and sliding-window attention across layers offer a promising approach to efficient long-context inference, but the critical question of \\emph{which layers} should retain full attention remains unsolved. Existing methods use either fixed periodic patterns or attention-based heuristics that may not capture what matters for downstream accuracy. We propose NLL-guided layer selection, a training-free method that directly measures each layer's importance by computing the negative log-likelihood degradation on answer tokens when that layer uses sliding-window instead of full attention. On LongMemEval with Qwen3-4B, our method achieves 64.6\\% accuracy using only 1/4 full-attention layers, matching the 1/2-FA periodic baseline (65.0\\%) while halving the computational budget. NLL-guided selection outperforms the SWAA-reported periodic 1/4-FA baseline by 10.4 percentage points and a matched LightTransfer-style baseline by 26.4 percentage points. De-confounding analysis shows the signal is consistent with long-range attention needs rather than generic layer sensitivity. The method requires only $\\sim$15 minutes of one-time calibration, advancing the efficiency-accuracy Pareto frontier for long-context LLM deployment.\n\n---\n\n## 29. Output-Space Allocation Costs for Calibration-Guided LLM Compression: An Empirical Study\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27785](https://arxiv.org/abs/2606.27785)\n- **Abstract**:\n> Training-free compression methods for large language models (LLMs) often use calibration data to guide compression decisions. ROCKET, a recent method combining sparse-dictionary factorization with multi-choice knapsack problem (MCKP) allocation, derives its per-layer factorization from an output reconstruction objective but uses weight-space Frobenius error as the MCKP allocation cost. We investigate whether aligning the allocation cost with the output-space objective improves compressed model fidelity. On Qwen3-8B at 50\\% compression, our ROCKET-ActCost achieves +0.8 percentage points higher average accuracy across 8 zero-shot benchmarks (53.1\\% vs 52.3\\%), but increases WikiText perplexity by 16\\% (61.46 vs 52.98). This accuracy-perplexity tradeoff reveals that different allocation objectives favor different downstream metrics. The high correlation ($>$0.99) between weight-space and output-space errors limits allocation divergence, explaining the modest effect size. On Llama-3.2-1B at 20\\% compression, the two methods produce near-identical results (53.3\\% vs 53.5\\% accuracy, 14.45 vs 14.66 PPL), suggesting that the effect of the cost function is minor at lower compression ratios.\n\n---\n\n## 30. Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?\n- **Authors**: Guoheng Sun , Kaixi Feng , Shwai He , Xiaochuan Gong , Yexiao He , Ziyao Wang , Zheyu Shen , Wanghao Ye , Ramana Rao Kompella , Gaowen Liu , Ang Li\n- **URL**: [https://arxiv.org/abs/2606.27755](https://arxiv.org/abs/2606.27755)\n- **Abstract**:\n> Vision-Language-Action (VLA) models enable instruction-driven robotic manipulation, but they inherit oversized language backbones from pretrained VLMs whose capacity far exceeds what is needed for short robotic instructions. This raises a basic question: how much of a VLA model is actually necessary for closed-loop control? In this work, we study architectural redundancy in VLA models by using transformer block removal as a controlled intervention. We introduce \\textbf{Drop-Then-Recovery (DTR)}, an analysis protocol that removes selected blocks from a pretrained VLA model and then fine-tunes the resulting model to measure whether the removed capacity was necessary for downstream control. To make this intervention reliable, we propose \\textbf{GateProbe}, a one-shot virtual-gate sensitivity metric that ranks blocks by their contribution to the downstream action loss. Across multiple VLA architectures, manipulation benchmarks and even real-robot industrial scenarios, we find a strong asymmetry in post-removal recoverability: \\ul{\\textit{language backbones are highly redundant for standard robotic manipulation tasks, whereas vision and action pathways are substantially less tolerant to removal} }. On LIBERO, removing half of the LLM blocks even improves OpenVLA-OFT from 95.0% to 98.3% under the same downstream fine-tuning budget, and retaining only two language blocks still recovers baseline-level performance. These results suggest that current VLA benchmarks may exert limited pressure on deep language grounding and compositional instruction understanding, and that future VLA architectures should allocate capacity more deliberately across language, vision, and action components. The code is available at this https URL .\n\n---\n\n## 31. End-to-End Dynamic Sparsity for Resource-Adaptive LLM Inference\n- **Authors**: Yuhang Chen , Jinhao Duan , Ruichen Zhang , Mingfu Liang , Xiaohan Wei , Yunchen Pu , Fei Tian , Chonglin Sun , Parish Aggarwal , Frank Shyu , Luke Simon , Sandeep Pandey , Tianlong Chen , Xi Liu\n- **URL**: [https://arxiv.org/abs/2606.27743](https://arxiv.org/abs/2606.27743)\n- **Abstract**:\n> Large Language Models (LLMs) inference is typically deployed under a static resource assumption, where models execute a fixed computational graph regardless of the runtime environment. However, real-world cloud infrastructure is inherently dynamic, characterized by fluctuating availability (e.g., spot instance preemption) and tiered Quality-of-Service requirements. In such volatile settings, static models are inflexible: they either crash under resource constraints or waste compute on redundant operations. To bridge this gap, we propose Learning to Allocate (L2A), an end-to-end framework for resource-adaptive inference. Unlike prior methods that condition only on input difficulty, we formulate inference as a constrained allocation problem conditioned on both the input and the runtime resource budget itself. We introduce lightweight, budget-conditioned and input-aware gating networks integrated into the LLM. These gates are trained via a unified objective that jointly optimizes task performance, logical consistency, and resource costs along three axes matching how real-world dynamics manifest: layer skipping for memory and depth pressure, head pruning for throughput contention, and reasoning-token reduction for latency tightening. This lets the model learn a budget-aware policy beyond input difficulty alone: it adaptively configures its computational footprint with respect to real-time resource dynamics, maximizing reasoning depth when resources permit while enforcing strict frugality when budgets tighten. A single L2A model traces the entire compute-accuracy Pareto frontier on Llama-3-8B and Qwen-3-4B: at up to 34% realized layer sparsity, it stays within 0.6% of the dense baseline on GSM8K, with the same gap holding zero-shot on out-of-distribution tasks, while every static or heuristic baseline requires a separately tuned model and still drops by 5-10% at comparable inference time.\n\n---\n\n## 32. KG2Cypher: Data-Centric Pipeline for Building Enterprise Text-to-Cypher Systems\n- **Authors**: Minjun Choi , Yerin Kim , Junghyuk Seo , Sujin Mo , Hyemin Lee , Youngjoong Ko\n- **URL**: [https://arxiv.org/abs/2606.27742](https://arxiv.org/abs/2606.27742)\n- **Abstract**:\n> Enterprise Knowledge Graphs (KGs) are increasingly used for internal search, analytics, and question answering, but building natural-language interfaces for private enterprise graphs remains costly. We present KG2Cypher, a data-centric pipeline for building enterprise text-to-Cypher systems from existing KGs. KG2Cypher first constructs an executable Cypher query from observed graph facts and then uses LLMs to generate its associated natural-language question. The resulting Text-Cypher pairs are validated with an LLM judge and human validation, and are converted into candidate-aware SFT data. The trained generator is served with class-conditioned schema prompting, entity retrieval, and LoRA-based inference. We evaluate KG2Cypher in Korean enterprise settings, where short search-style queries and schema paraphrases make language grounding difficult. LoRA SFT improves execution-result F1 from 0.806 to 0.950 on broadcast-program queries and from 0.70 to 0.92 on company queries. In an 11-class setting, KG2Cypher achieves 95.2% exact match, 99.9% execution rate, and 0.964 execution-result F1.\n\n---\n\n## 33. Bifocal Diffusion Language Models: Asymmetric Bidirectional Context for Parallel Generation\n- **Authors**: Yuhang Chen , Xianfeng Wu , Jinhao Duan , Mingfu Liang , Xiaohan Wei , Yunchen Pu , Fei Tian , Chonglin Sun , Parish Aggarwal , Frank Shyu , Luke Simon , Sandeep Pandey , Xi Liu , Tianlong Chen\n- **URL**: [https://arxiv.org/abs/2606.27732](https://arxiv.org/abs/2606.27732)\n- **Abstract**:\n> Discrete diffusion language models (dLLMs) recover masked tokens in parallel, offering significant speedups over autoregressive (AR) generation. However, such promising frameworks face a fundamental architectural design dilemma: \\ding{182} Adopting bidirectional attention achieves strong generation quality by allowing each position to access the full context, but is inherently incompatible with KV caching, limiting inference throughput in batch-serving scenarios; \\ding{183} Conversely, causal attention enables efficient cached inference but loses all right-side context, substantially degrading generation quality. This paper introduces Bifocal dLLMs, a new paradigm that resolves this dilemma through \\emph{asymmetric bidirectional context}. Analogous to bifocal lenses, we instantiate the paradigm as \\textbf{R2LM} (Right-to-Left Mamba), which combines two complementary mechanisms: $a$) standard causal attention providing precise left-context with full KV cache compatibility, while $b$) a lightweight reverse Mamba SSM sidecar supplying compressed right-side context without breaking cacheability. Comprehensive experiments on continued pretraining of Qwen3-1.7B with 60B tokens demonstrate that R2LM achieves $2.4\\times$ to $12.9\\times$ higher throughput than bidirectional dLLMs and $1.9\\times$ to $2.9\\times$ speedup over AR baselines in batch serving through parallel decoding with KV caching, while exceeding the causal baseline on most benchmarks and surpassing the bidirectional dLLM on average.\n\n---\n\n## 34. Enhancing Numerical Prediction in LLMs via Smooth MMD Alignment\n- **Authors**: Zhuo Zuo , Li Yue , Wenhao Zheng , Chenpeng Wang , Xianggen Liu\n- **URL**: [https://arxiv.org/abs/2606.27731](https://arxiv.org/abs/2606.27731)\n- **Abstract**:\n> Despite their strong general capabilities, large language models (LLMs) often remain unreliable when outputs must be numerically precise. A key reason is the training objective: standard cross-entropy treats numeric tokens as unstructured categories and ignores the metric structure of their values. We address this mismatch with Smooth Maximum Mean Discrepancy (SMMD), which builds on the classic MMD by incorporating value-distance kernels over numeric tokens and graph-based smoothness. With this kernel defined over a numeric sub-vocabulary, SMMD aligns the predicted numeric distribution to the target via kernel matching and smooths the prediction-target residual over the induced kernel graph to encourage local consistency. We evaluate SMMD on four numeric-target tasks: mathematical reasoning, arithmetic calculation, clock-time recognition, and chart question answering, across multiple open-weight LLM and VLM backbones. SMMD consistently improves accuracy over both cross-entropy and recent numeric-target losses; analyses show complementary effects between MMD and smoothness and underscore the importance of distance-based kernel design. Code is available at this https URL .\n\n---\n\n## 35. Low-Agreeableness Persona Conditioning for Safe LLM Fine-Tuning\n- **Authors**: Austin MY Cheung , Yi Yang\n- **URL**: [https://arxiv.org/abs/2606.27709](https://arxiv.org/abs/2606.27709)\n- **Abstract**:\n> Recent work has shown that fine-tuning large language models (LLMs) for social warmth degrades factual reliability and increases sycophancy. We investigate a related but distinct failure mode: warmth fine-tuning also weakens adversarial safety, making models more susceptible to jailbreaks and harmful output generation. We examine whether this reflects an inherent consequence of empathetic adaptation or an artifact of data construction. To address this, we introduce a persona-driven rewriting pipeline that conditions user turns on low agreeableness and pairs this with warm, de-escalating assistant responses. Across three experiments on four models, our approach reduces jailbreak susceptibility and harmful output rates relative to generic warmth fine-tuning baselines, while preserving conversational warmth. Representational probing provides suggestive evidence that this conditioning reduces the geometric alignment between warmth and compliance directions in latent space. These results show that safer empathetic fine-tuning is achievable through data design alone, without safety labels, harm detectors, or changes to the training objective.\n\n---\n\n## 36. Mitigating LLM-based p-Hacking by Preregistering for the Next LLM\n- **Authors**: Maria Thomas , Kristina Gligoric , Nihar B. Shah\n- **URL**: [https://arxiv.org/abs/2606.27687](https://arxiv.org/abs/2606.27687)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used to generate, classify, and annotate data whose outputs feed downstream hypothesis tests. However, LLM-based research is easy to p-hack: a researcher can tune the prompts, decoding parameters, or output format until a desired result is reached. We propose a protocol to mitigate p-hacking in LLM-based research: preregistering the experiment and eligible models, and then running it on the first eligible LLM that is released after the preregistration. The researcher finalizes the procedure on current models, preregisters the analysis plan together with a set of eligible future models, and runs the confirmatory analysis on the first eligible model released afterward. Because this model does not exist at commitment time, it cannot be hacked against; furthermore, configurations that hack one model frequently do not transfer to the next. We evaluate the protocol on two tasks whose true values are known. Across 20 models from four providers and 11 LLM-analysis configurations, the protocol would have blocked successful transfer of the p-hack in 73.9% and 72.7% of cases in the two tasks. Additional analyses reveal that mitigation remains substantial under several stress tests. Finally, putting money where our mouth is, we followed our own protocol and preregistered our experiment. The preregistered experiment confirmed the protocol's effectiveness: out of the 7 configurations that hacked the prior model, the hacking failed to carry over in 6 configurations on the first eligible model released afterward.\n\n---\n\n## 37. CBD: API-Only LLM Black-Box Unlearning through Controlled Behavioral Divergence\n- **Authors**: Zhiqiang Xie , Yijing Lin , Zhipeng Gao , Dong In Kim\n- **URL**: [https://arxiv.org/abs/2606.27683](https://arxiv.org/abs/2606.27683)\n- **Abstract**:\n> Edge devices increasingly invoke large language models (LLMs) through API services for context aware edge intelligence, while edge generated data may be collected to improve LLMs and may introduce sensitive, copyrighted, harmful, or outdated information into model behavior. Machine unlearning offers a practical way to remove the influence of undesired data without retraining LLMs. However, existing methods still face two gaps. The first is API only black box access, where target model parameters and internal logits are unavailable. The second is how to preserve retained utility when unlearning target data and retained data share highly similar prompt structures or semantic patterns. To address these challenges, we propose Controlled Behavioral Divergence (CBD), an API only black box unlearning framework. CBD uses two auxiliary models to create controlled behavioral divergence between retained inputs and unlearning target inputs, converts this divergence into an unlearning relevance score, and routes unlearning related prompts away from the target LLM. To improve discrimination accuracy under high similarity between target and retained data, CBD constructs a gradient statistics based discriminative basis by estimating empirical Fisher matrices and solving a regularized generalized eigenvalue problem, guiding the unlearning signal toward target specific information rather than shared prompt structures. Compared with eleven white box and gray box unlearning baselines, CBD achieves a better unlearning utility trade off and its performance varies little across settings. On ToFU forget10, CBD approaches the retrained reference on the forget set while raising model utility to 74.90, about 15% above the second best baseline. On WMDP, it lowers hazardous knowledge accuracy to 25.68, near random guessing, while preserving MMLU accuracy of 52.67. Code is at this https URL .\n\n---\n\n## 38. From Signals to Transfer: A Factorised Study of Probe-Based Uncertainty Estimation in Large Language Models\n- **Authors**: Ponhvoan Srey , Xiaobao Wu , Cong-Duy Nguyen , Quang Minh Nguyen , Duc Anh Vu , Anh Tuan Luu\n- **URL**: [https://arxiv.org/abs/2606.27679](https://arxiv.org/abs/2606.27679)\n- **Abstract**:\n> Probe-based uncertainty estimation (UE) has emerged as a prominent approach to detect hallucinations in Large Language Models (LLMs) by learning uncertainty from internal model signals. Yet, recent methods vary simultaneously across feature design, training data construction, and evaluation setting, obscuring what actually drives performance. To address this issue, we propose a factorised study of probe-based UE under matched conditions. Our results show that raw hidden states and attention features are difficult to outperform in-domain. However, under distribution shift, structured and compressed features are more robust, suggesting that in-domain performance alone is insufficient to measure progress. Furthermore, prompting and label construction significantly affect probe behaviour. Building on these best-practice findings, we train benchmark-based pretrained probes that transfer reasonably well to open-ended factual generation, providing a stable off-the-shelf baseline. Our work encourages more deployment-oriented evaluation of probe-based uncertainty estimators. The code repository is available at this https URL .\n\n---\n\n## 39. HybridCodec: Modeling Discrete and Continuous Representations for Efficient Speech Language Models\n- **Authors**: Artem Ploujnikov , Francesco Verdini , Samir Sadok , Mirco Ravanelli\n- **URL**: [https://arxiv.org/abs/2606.27627](https://arxiv.org/abs/2606.27627)\n- **Abstract**:\n> Discrete audio representations have become increasingly popular for building multimodal text-audio systems and integrating audio capabilities into Large Language Models (LLMs). However, numerous studies report performance degradation on various downstream tasks due to information loss during discretization. To address this, we propose a novel approach combining temporally compressed discrete tokens with dimensionality-reduced continuous residuals. Our framework consists of a hybridized discrete-continuous focal modulation codec and a hybrid Transformer. This architecture performs autoregressive inference in the discrete domain, coupled with non-autoregressive prediction and continuous residual upsampling. Experimental results show that our approach significantly improves the retention of speaker characteristics compared to discrete-only methods, while simultaneously reducing the number of required autoregressive steps.\n\n---\n\n## 40. Dismantling Pathological Shortcuts: A Causal Framework for Faithful LVLM Decoding\n- **Authors**: Liu Yu , Can Chen , Ping Kuang , Zhikun Feng , Fan Zhou , Gillian Dobbie\n- **URL**: [https://arxiv.org/abs/2606.27596](https://arxiv.org/abs/2606.27596)\n- **Abstract**:\n> Large Vision-Language Models (LVLMs) exhibit sophisticated reasoning but remain susceptible to object hallucination. Deviating from the prevailing attention intensity assumption, we reveal a deeper dynamic structural misalignment: hallucination is triggered at decision-critical steps where specific attention heads, acting as risky mediators, decouple from visual evidence to lock onto language priors. This establishes a pathological shortcut that bypasses visual grounding. To dismantle this, we propose Fox (Faithfulness and Observational-flow via eXpression-rectification), a training-free inference-time framework. Fox diagnoses structural misalignment using a visual attention entropy probe to localize risky mediators unsupervisedly. We then execute a targeted causal intervention via numerical logit saturation to physically sever the shortcut path. Finally, a conflict-gated cooperative decoding strategy reconciles interventional faithfulness with observational fluency. Extensive experiments demonstrate that Fox achieves SOTA performance, outperforming SID by 29.1% while preserving linguistic richness. Code is available at this https URL .\n\n---\n\n## 41. On the Inseparability of Instructions and Data in Shared-Embedding Sequence Models\n- **Authors**: Dewank Pant , Shruti Lohani , Avijit Kumar\n- **URL**: [https://arxiv.org/abs/2606.27567](https://arxiv.org/abs/2606.27567)\n- **Abstract**:\n> Prompt injection is the top security risk for LLM-integrated applications, yet every defense proposed so far has been broken. We prove this is not a coincidence: in shared-embedding architectures that lack enforced control-data separation, perfect prompt-injection prevention is mathematically impossible. We formalize prompted systems as Prompted Action Models whose outputs include control-authoritative actions: refusal decisions, tool authorization, policy routing, and memory writes. We define Semantic-Faithful Control (SFC), the property that such behavior depends only on the meaning of untrusted input, not on how it is encoded. We then prove SFC is unachievable within the shared pipeline, via three results: a provenance-recovery impossibility (shared representations make trusted and untrusted content statistically inseparable, bounded by total variation distance); control-path exposure (untrusted tokens enter control-relevant computation through the same attention value-aggregation that determines outputs); and a finite-coverage invariance gap (finite training cannot certify invariance over infinite semantic-equivalence classes). We ground each quantity in measurements on production tokenizers and models. The result is structural, not a gap in current defenses. It mirrors the code-data confusion in Von Neumann machines that gives rise to buffer overflows, a vulnerability class that took decades of layered defenses (DEP, Write-XOR-Execute, ASLR, stack canaries, and ultimately memory-safe languages) to contain, because no single mechanism sufficed. The implication is the same: prompt injection cannot be eliminated by better in-pipeline classification or alignment alone. It requires architectural separation of instruction and data channels. We identify the root cause and the class of solution it demands.\n\n---\n\n## 42. Benchmarking Multi-Modal Graph-based Social Media Popularity Prediction\n- **Authors**: Utkarsh Sahu , Zhisheng Qi , Li Zhu , Yizhao Yang , Jun Li , Ryan Rossi , Yu Wang\n- **URL**: [https://arxiv.org/abs/2606.27539](https://arxiv.org/abs/2606.27539)\n- **Abstract**:\n> Social media popularity prediction aims to forecast the future reach or influence of online content from early-stage observations. Accurate prediction enables key downstream applications, such as advertising optimization and strategic content planning by users, creators, and platforms. Despite substantial progress, existing popularity prediction works often fail to jointly consider multimodal content and temporal social interaction signals. Moreover, the literature remains highly fragmented across datasets, modalities, observation windows, prediction targets, and evaluation protocols. This fragmentation prevents fair comparison and obscures a systematic understanding of how textual, visual, temporal, and interaction-based signals jointly shape popularity dynamics. To address these challenges, we introduce MMG-Pop, a Multi-modal Graph-based Popularity Prediction benchmark, which unifies datasets, modalities, temporal interaction signals, and representative baselines under a standardized evaluation protocol. Furthermore, we propose MMG-PopNet, a unified multi-modal graph-based network that jointly models the aforementioned multi-modal signals and graph-structured social interactions. Extensive experiments on MMG-Pop, comprising four datasets across Bluesky and Reddit platforms, demonstrate the superior performance of MMG-PopNet and yield new insights into cross-platform training generalization, multi-task prediction benefits, multi-modality contributions, and LLM prediction limitation. These findings establish a unified foundation for future research on social dynamics modeling and intervention under heterogeneous modalities and socially-aware agentic ecosystem paradigms.\n\n---\n\n## 43. Large Language Model Teaches Visual Students: Cross-Modality Transfer of Fine-Grained Conceptual Knowledge\n- **Authors**: Thomas Shih-Chao Liang , Zhuoran Yu , Yong Jae Lee\n- **URL**: [https://arxiv.org/abs/2606.27527](https://arxiv.org/abs/2606.27527)\n- **Abstract**:\n> Large Language Models (LLMs) possess broad conceptual knowledge acquired through large-scale text pretraining, yet their potential to supervise models in other modalities remains underexplored. In this work, we propose LaViD--Language-to-Visual Knowledge Distillation--a simple and effective framework for transferring high-level semantic knowledge from a language-only teacher to a vision-only student model. Instead of relying on paired multimodal data, LaViD elicits conceptual signals from an LLM by prompting it to generate multiple-choice questions (MCQs) that probe semantic distinctions between visual classes. Each class is mapped to a soft label distribution over these MCQs, forming a rich conceptual signature that guides the student through an auxiliary distillation loss. Notably, despite using a language-only teacher without access to image data, LaViD consistently outperforms recent methods like MaKD that distill from vision-language models across multiple fine-grained benchmarks. It also achieves competitive or superior performance compared to state-of-the-art visual distillation methods such as DKD and MLKD, with further gains when combined with logit standardization. On the Waterbirds dataset, LaViD substantially improves worst-group accuracy, demonstrating enhanced robustness to spurious correlations with distillation. Code is available at this https URL .\n\n---\n\n## 44. Speculative Refinement: A Hybrid Autoregressive Diffusion Decoding Strategy and Its Behavior Across Benchmarks\n- **Authors**: Aditi Gupta , Neel Mishra , Kushagra Trivedi , Pawan Kumar\n- **URL**: [https://arxiv.org/abs/2606.27474](https://arxiv.org/abs/2606.27474)\n- **Abstract**:\n> How should we evaluate generation systems that combine autoregressive (AR) and diffusion decoding? We study this question through Speculative Refinement (SpecRef), a training-free hybrid method that warm-starts a masked diffusion language model from an AR draft using entropy-guided selective masking. Evaluating SpecRef across six benchmarks (HumanEval, MBPP, GSM8K, BBH, ARC-Challenge, HellaSwag) with three distinct evaluation protocols (execution-based pass@1, exact-match, log-likelihood scoring), we surface several findings relevant beyond our specific system: (1) code benchmarks conflate structural discovery with logical correctness: providing a syntactic scaffold lifts accuracy from near zero to over 20% without changing the model, indicating that much of the baseline failure is structural; (2) a refinement tension phenomenon where multi-stage correction degrades already-correct tokens, exposing benchmark saturation ceilings invisible to single-model evaluation; (3) log-likelihood and generative evaluation produce different model rankings for the same model pair, suggesting they measure different capabilities; (4) standard Python post-processing silently breaks code evaluation for non-AR generators. These observations apply to any multi-stage or non-autoregressive generation pipeline and point toward more diagnostic evaluation practices.\n\n---\n\n## 45. Supersede: Diagnosing and Training the Memory-Update Gap in LLM Agents\n- **Authors**: Vedant Patel\n- **URL**: [https://arxiv.org/abs/2606.27472](https://arxiv.org/abs/2606.27472)\n- **Abstract**:\n> Large language model (LLM) agents operate over long, multi-session interactions in which facts change: a user moves, a price updates, a plan is revised. Acting correctly requires using the current value of a fact and discarding values that have been superseded. We isolate this ability on real conversational data and show that it is a distinct, unsolved failure. On the knowledge-update subset of LongMemEval, replacing an agent's full context with a bounded, self-maintained memory drops accuracy from 92% to 77% even on a frontier model (gpt-5.4), a gap that is statistically significant (paired McNemar p<0.005) and persists across model scale while full-context accuracy saturates near 92%. The bottleneck is therefore memory maintenance, not comprehension, and is not closed by a stronger model. We then ask whether this is merely an undersized memory, and find it is not: as the conversation grows 24x, accuracy falls further (from 68% to 28%), and granting the agent proportionally more memory yields no detectable recovery (28% to 28%, n=25). The failure scales with the length of the conversation, not the compression ratio. We release Supersede, an open reinforcement-learning environment (on the verifiers / prime-rl stack) that turns this measurement into a training signal: agents are rewarded for answering from the current value and penalized for stale ones. Finally, we close the loop and show the gap is trainable: GRPO fine-tuning a small open model (Qwen2.5-3B) on this environment nearly doubles its held-out supersession accuracy on real, unseen conversations (9.0% to 16.7%, a single run), along a monotonic checkpoint curve indicating the learned policy, not the harness, carries the gain. To our knowledge this is the first trainable environment whose reward targets temporal fact-currency, and the first evidence the supersession gap can be trained down, not only measured.\n\n---\n\n## 46. SidConArena: An Environment Evaluating Agents in Open-Ended,Positive-Sum Bargaining Game\n- **Authors**: Yeqi Feng , Yuxin Chen , Tianxing He\n- **URL**: [https://arxiv.org/abs/2606.27397](https://arxiv.org/abs/2606.27397)\n- **Abstract**:\n> Evaluating LLM agents requires dynamic environments that go beyond static reasoning and zero-sum games. Real-world economic interaction is often open-ended and mixed-motive: agents must negotiate, create positive-sum surplus, compete for scarce assets, and plan under delayed returns. We introduce SidConArena, a new benchmark framework for evaluating LLM agents in open-ended, positive-sum bargaining. SidConArena formalizes a multi-player economy as a finite-horizon partially observable stochastic game with three coupled phases: natural-language negotiation with binding trades, deterministic converter-based production, and sealed-bid auctions for long-term assets. The framework combines structured observations, phase-aware agent dispatching, a neural-symbolic action interface, and asynchronous execution, enabling free-form interaction while preserving rule-grounded evaluation. Across homogeneous and heterogeneous tournaments, stronger frontier models achieve higher economic outcomes, yet agents still misvalue resources, bargain passively, and remain limited in long-horizon investment planning.\n\n---\n\n## 47. Agentic Publication Protocol: An Attempt to Modernize Scientific Publication\n- **Authors**: Sirui Lu , Xiao-Liang Qi\n- **URL**: [https://arxiv.org/abs/2606.27386](https://arxiv.org/abs/2606.27386)\n- **Abstract**:\n> Scientific publication is still organized primarily around static manuscripts, even though much of scientific progress depends on tacit know-how: how to run code, reproduce figures, interpret edge cases, choose useful follow-up directions, and avoid failed paths. Large language model agents create an opportunity to publish not only knowledge, but also operational know-how in a form that future readers and researchers can directly use. This paper outlines the Agentic Publication Protocol (APP), a lightweight repository format for packaging a paper together with code, data, environment information, reproducibility instructions, and an agent-facing instruction file. APP treats a version-controlled repository as the publication object and uses \\texttt{ this http URL } and optional skills to define a paper agent that can explain the work, reproduce key results when possible, and support follow-up research. We describe the design principles and details of the protocol, as well as the agent skills useful for publishing papers under the protocol. We also describe development tools for evaluating and improving the protocol and associated agent skills. Finally, we provide a broader discussion of the future of scientific research in the agent era.\n\n---\n\n## 48. CalBrief: A Pilot Diagnostic Benchmark for Evidence-Calibrated Scientific Briefing with Large Language Models\n- **Authors**: Yu Fu , Yongqi Kang , Yong Zhao\n- **URL**: [https://arxiv.org/abs/2606.27383](https://arxiv.org/abs/2606.27383)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used as research assistants, yet it remains unclear whether they can calibrate research takeaways to the strength and scope of the supporting evidence. We study evidence-calibrated scientific briefing: given a bounded package of related papers, a system should generate package-level takeaways with evidence strength, scope boundaries, and missing-evidence caveats. We contribute a verified pilot benchmark of 16 heterogeneous scientific evidence packages and 96 human-verified takeaways, and we use CalBrief, an auditable role/gap/strength framework, as a diagnostic probe to locate where briefing breaks down. Under a fair-schema evaluation, structured organization improves role and gap reasoning, but an explicit strength-calibration policy is systematically over-conservative and falls below majority and direct-LLM baselines. To explain why, we run a controlled diagnostic across three closed-model backbones (GPT-4o, Claude Sonnet, Gemini Flash) that separates three potential causes of conservatism. Approximately 63% of the conservatism gap is attributable to expanding the label space from binary {moderate, weak} to four-way {moderate, weak, uncertain, insufficient_evidence} (p < 0.001 across all backbones); only 1% is attributable to gap/scope signal injection (not significant); the remaining 36% arises from the pipeline policy itself. We also find that 4-way predictions can be post-hoc collapsed back to binary and then match or exceed direct binary prompting, so the extra labels carry information that strict matching hides. Label-level strength judgment and auditable evidence organization are distinct abilities currently in tension, and should be evaluated separately for LLM research assistants.\n\n---\n\n## 49. Position: The Term \"Machine Unlearning\" Is Overused in LLMs\n- **Authors**: Sangyeon Yoon , Yeachan Jun , Albert No\n- **URL**: [https://arxiv.org/abs/2606.27379](https://arxiv.org/abs/2606.27379)\n- **Abstract**:\n> Large language models increasingly face demands to \"forget\" training data, knowledge, or behaviors due to regulatory deletion obligations, copyright/licensing disputes, and safety or product-policy requirements. This position paper argues that machine unlearning is overused as a term in LLM research and should be reserved for dataset-defined deletion: removing the training influence of a precisely specified forget set such that the resulting model is approximately indistinguishable from retraining without that data. We contend that many tasks currently labeled \"unlearning\" (e.g., refusal for harmful requests, entity/knowledge removal, or targeted suppression) pursue different, often policy-dependent objectives and therefore require different terminology and baselines (e.g., alignment, suppression, editing, obfuscation). We further argue that this confusion is not cosmetic: because papers make different implicit guarantees under the same label, metrics and benchmarks are frequently reused outside their intended scope, rewarding surface-level non-disclosure (e.g., low ROUGE/forget accuracy) even when retraining-equivalence is not tested and derived capabilities remain. We conclude by calling for stricter terminology tied to explicit guarantees and reference models, and for evaluations that match the claimed objective.\n\n---\n\n## 50. DataStates-LLM: Scalable Checkpointing for Transformer Models Using Composable State Providers\n- **Authors**: Avinash Maurya , M. Mustafa Rafique , Franck Cappello , Bogdan Nicolae\n- **URL**: [https://arxiv.org/abs/2601.16956](https://arxiv.org/abs/2601.16956)\n- **Abstract**:\n> The rapid growth of Large Transformer-based models, specifically Large Language Models (LLMs), now scaling to trillions of parameters, has necessitated training across thousands of GPUs using complex hybrid parallelism strategies (e.g., data, tensor, and pipeline parallelism). Checkpointing this massive, distributed state is critical for a wide range of use cases, such as resilience, suspend-resume, investigating undesirable training trajectories, and explaining model evolution. However, existing checkpointing solutions typically treat model state as opaque binary blobs, ignoring the ``3D heterogeneity'' of the underlying data structures--varying by memory location (GPU vs. Host), number of ``logical'' objects sharded and split across multiple files, data types (tensors vs. Python objects), and their serialization requirements. This results in significant runtime overheads due to blocking device-to-host transfers, data-oblivious serialization, and storage I/O contention. In this paper, we introduce DataStates-LLM, a novel checkpointing architecture that leverages State Providers to decouple state abstraction from data movement. DataStates-LLM exploits the immutability of model parameters during the forward and backward passes to perform ``lazy'', non-blocking asynchronous snapshots. By introducing State Providers, we efficiently coalesce fragmented, heterogeneous shards and overlap the serialization of metadata with bulk tensor I/O. We evaluate DataStates-LLM on models up to 70B parameters on 256 A100-40GB GPUs. Our results demonstrate that DataStates-LLM achieves up to 4$\\times$ higher checkpointing throughput and reduces end-to-end training time by up to 2.2$\\times$ compared to state-of-the-art solutions, effectively mitigating the serialization and heterogeneity bottlenecks in extreme-scale LLM training.\n\n---\n\n{% endraw %}\n"}, "ae95e160c1986855bbe27edfadb3713b4c8f4a25": {"t": 1786068564.8742285, "n": 84889, "h": "29e92ec4f78e9471", "c": "---\nlayout: default\ntitle: \"LLM 관련 주요 논문 - 2026-06-29\"\n---\n\n{% raw %}\n# LLM 관련 주요 논문 - 2026-06-29\n\n## 1. Tandem Reinforcement Learning with Verifiable Rewards\n- **Authors**: Difan Jiao , Raghav Singhal , Robert West , Ashton Anderson\n- **URL**: [https://arxiv.org/abs/2606.28166](https://arxiv.org/abs/2606.28166)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has significantly improved the reasoning capability of large language models, reaching expert or even superhuman performance in domains such as competition math. However, whether weaker agents and humans can actually harness this capability is far less certain, with RLVR documented to drift reasoning toward idiosyncratic patterns such as poor readability and language mixing. Tandem training is a recently introduced paradigm that targets this compatibility problem: a trained, stronger senior co-generates each rollout with a frozen, weaker junior, and the two are rewarded as a team, so the senior is pushed to reason in ways the junior can follow. Yet this paradigm has so far been demonstrated only in proof-of-concept settings, leaving open whether it scales to the long chains of thought of the modern RLVR pipeline. In this work, we propose Tandem Reinforcement Learning (TRL), which carries the tandem training paradigm into RLVR. In TRL, the senior and a frozen junior alternate stochastically to co-generate the reasoning, the resulting generation is rewarded, and the standard GRPO loss is applied to the senior. Training Qwen3-4B-Instruct on competition math, we find that TRL matches vanilla GRPO on solo reasoning capability while three properties emerge together from the same rollout structure: stronger handoff robustness with the junior, reduced distributional drift from the junior, and a chain-of-thought more legible to the junior. Our results demonstrate a promising route for RLVR with practical payoffs in multi-model communication and human compatibility.\n\n---\n\n## 2. JD Oxygen AI Item Center (Oxygen AIIC) V1: An Industrial-Scale LLM/VLM-Centric Solution for Item Understanding, Management, and Applications\n- **Authors**: Oxygen AIIC , Chan Long , Chao Liu , Chaofan Chen , Chaohui Dong , Chunyuan Guo , Danping Liu , Debin Liu , Deping Xiang , Fulai Xu , Guangyue Liu , Hao Li , Huichun Hu , Jian Yang , Jianan Wang , Jianbo Zhao , Jiaoyang Li , Jiaxing Wang , Jinglong Li , Jinjin Guo , Jun Fang , Jun Liu , Kai Zhou , Li Wang , Lili Gao , Liying Chen , Luning Yang , Mengdi Zhou , Pengzhang Liu , Qi Lv , Qianyun Wang , Qixia Jiang , Ruyue Li , Shimu Liang , Shuxing Wang , Sijie Zhang , Siqi Li , Tianhao Gao , Wang Ke , Weihu Huang , Wencan Lai , Wenjie Zhang , Xiaohui Zhang , Xiaojing Dong , Ya Liu , Yifeng Zhang , Yixiang Wang , Yongtai Zhang , Yongyi Liao , Zhaoru Chen , Zhen Chen , Zhiyong Ma , Zhiyuan Liu , Zhongwei Liu , Ziyan Xing\n- **URL**: [https://arxiv.org/abs/2606.28070](https://arxiv.org/abs/2606.28070)\n- **Abstract**:\n> this http URL , one of the world's largest e-commerce platforms, serves over 700 million active users and millions of merchants, with a catalog of tens of billions of SKUs. At this scale, high-quality, structured item knowledge underpins a better consumer experience, lower management costs, and higher operational efficiency-yet producing and serving it poses three industrial-scale challenges: fast-emerging concepts, high-quality knowledge production for massive SKUs, and diverse downstream requirements. To address these challenges, we present the JD Oxygen AI Item Center (Oxygen AIIC), an industrial-scale platform built on LLMs/VLMs for item-knowledge production and service. Oxygen AIIC is built around four core pillars: (i) ontology engineering driven by efficient human-AI collaboration, which supports the dynamic evolution and agile expansion of an ontology with millions of entries; (ii) a \"Semantic Search then Discrimination\"(S2D) knowledge identification architecture that, combined with throughput improvement strategies, enables scalable, extensible, and high-throughput AI Item Library production for tens of billions of SKUs; (iii) self-evolving item-understanding LLMs/VLMs that improve in a stable and controllable manner, enabling knowledge production with 94.2% precision and 82.8% recall; and (iv) a unified item tunnel that serves as the data and service hub. Oxygen AIIC now covers tens of thousands of JD categories and processes hundreds of millions of item updates per day on Huawei Ascend NPUs. It has accumulated hundreds of billions of item-knowledge assets. Deployed across core business scenarios-including search, recommendation, operations, category planning-Oxygen AIIC has delivered measurable gains at scale. Search-traffic coverage reaches 80.4%, item-information quality issues drop by 37%, the automated fill rate of core attributes during item listing exceeds 80%.\n\n---\n\n## 3. NormAct: A Benchmark for Hidden Social Norm Compliance in Embodied Planning\n- **Authors**: Shiyun Zhao , Xinwei Song , Tianyu Guo , Xiaomeng Gao , Mingyuan Liu , Xu Han , Yuanyuan Zhang , Zhenliang Zhang , Xue Feng , Bo Dai\n- **URL**: [https://arxiv.org/abs/2606.27826](https://arxiv.org/abs/2606.27826)\n- **Abstract**:\n> Multimodal large language models (MLLMs) are increasingly deployed as embodied planners in egocentric environments, where task success requires not only achieving instructed goals but also acting in socially appropriate ways. While explicit goals may render certain actions optimal, implicit social norms often impose hidden constraints. Existing evaluations typically focus on explicit goal achievement or direct norm knowledge, seldom assessing whether planners can infer and apply these hidden constraints within action sequences. We introduce NormAct, a benchmark for embodied social-norm interactions that evaluates plans on Goal Achievement, Norm Compliance, and overall Task Success. NormAct uniquely embeds hidden norms within ordinary tasks, testing whether models can realize them without explicit instruction. Experiments with state-of-the-art MLLMs (GPT-5.4, Claude Opus 4.7, Gemini 3 Pro) reveal a significant gap: models achieve explicit goals in 67.3\\% of cases, but comply with hidden norms in only 26.4\\%. Cue-condition experiments indicate that this gap stems not from a lack of general social knowledge, but from challenges in activating and grounding relevant norms in context. To address this, we propose NormPerceptor, a context-conditioned cue generator that infers scene-relevant norms prior to planning, increasing Task Success from 24.2\\% to 46.7\\%. Our results underscore the importance of enabling embodied agents to proactively detect hidden norms, ground them in visual evidence, and integrate them as action-planning constraints. Our benchmark is publicly available at this https URL .\n\n---\n\n## 4. Grounded Iterative Language Planning: How Parameterized World Models Reduce Hallucination Propagation in LLM Agents\n- **Authors**: Xinyuan Song , Zekun Cai\n- **URL**: [https://arxiv.org/abs/2606.27806](https://arxiv.org/abs/2606.27806)\n- **Abstract**:\n> World models for language agents come in two useful forms. An agent-based world model calls an LLM API and reasons flexibly in language, but its errors appear as hallucinated state changes that are hard to score with ordinary regression losses. A parameterized world model is a trained transition predictor; its errors are easier to measure with quantities such as NodeMSE, delta accuracy, and validity accuracy, but it is usually weaker as a standalone planner. We compare these two families on four graph-structured planning benchmarks and introduce operational hallucination metrics for the agent-based case. The comparison motivates \\textbf{Grounded Iterative Language Planning} (GILP), which trains only a small parameterized backbone and combines it with API-based agent reasoning. The backbone supplies valid actions, predicted state deltas, risk, and value; the LLM drafts an action and imagined delta; and a consistency gate asks for revision when the two disagree. On real GPT-4o-mini calls, GILP reduces hallucinated-state rate from 0.176 to 0.035. In calibrated simulator ablations, it raises success from 0.668 to 0.838 while adding only ~22% extra LLM calls.\n\n---\n\n## 5. Towards Reliable and Robust LLM Planning: Symbolic Feedback-Driven Iterative Self-Refinement Framework\n- **Authors**: Jiajing Zhang , Jiamei Jiang , Chenyang Zhang , Feifei Mo , Linjing Li , Daniel Zeng\n- **URL**: [https://arxiv.org/abs/2606.27757](https://arxiv.org/abs/2606.27757)\n- **Abstract**:\n> Large language models (LLMs) have attracted widespread attention from academia and industry, yet their deployment raises critical security concerns regarding robustness and reliability. Planning, a core component of intelligent behavior, remains challenging for LLMs, which often produce infeasible or incorrect solutions in long-horizon decision-making tasks due to inherent complexity. In this paper, we propose a symbolic feedback-driven iterative self-refinement framework to enhance the robustness and reliability of LLMs in long-horizon planning. Specifically, a natural language prompting mechanism is introduced to map logical symbols into natural language descriptions, enabling LLMs to better capture task constraints and semantics. We further design a symbolic verifier that identifies errors and converts them into corrective instructions interpretable by the LLM, thereby guiding self-refinement. In addition, we leverage a plan recognizer to infer goal reachability, facilitating more effective guidance toward desired goals. Empirical results demonstrate that the proposed framework consistently improves both feasibility and correctness in long-horizon planning tasks. This highlights its effectiveness in enhancing the reliability of LLM-based planning and potential to enable more trustworthy AI systems.\n\n---\n\n## 6. ToE: A Hierarchical and Explainable Claim Verification Framework with Dynamic Multi-source Evidence Retrieval and Aggregation\n- **Authors**: Zhaoqi Wang , Zijian Zhang , Kun Zheng , Zhen Li , Xin Li , Chunlei Li , Jiamou Liu\n- **URL**: [https://arxiv.org/abs/2606.27736](https://arxiv.org/abs/2606.27736)\n- **Abstract**:\n> The rapid spread of fake news poses increasing threats to information ecosystems, especially as AI-generated misinformation under Generative Engine Optimization (GEO) poisoning allows adversarially crafted content to be systematically surfaced by retrieval systems, contaminating LLM reasoning. In this paper, we propose Tree of Evidence (ToE), a hierarchical evidence reasoning framework for automated fact-checking that models each claim as a dynamically expanding argument tree. ToE integrates a reinforcement learning-driven multi-source retrieval agent, an evidence evaluation agent, and an argument tree aggregation algorithm to iteratively decompose, retrieve, and verify claims through an explainable evidence chain. We further provide a theoretical analysis of the retrieval process, deriving a formal error bound that guarantees the learned policy converges to a neighborhood of the information-theoretically optimal policy. Experiments across multiple datasets and backbone LLMs demonstrate that ToE achieves improvements ranging from 4 to 24 percentage points over competitive baselines, with particularly pronounced gains on adversarially poisoned inputs.\n\n---\n\n## 7. DysLexLens: A Low-Resource LLM Framework for Analysing Dyslexic Learners Insights from Online Forums\n- **Authors**: Dana Rezazadegan , Atie Kia , Phongpadid Nandavong , Dominique Carlon , Jeremy Nguyen , Abhik Banerjee , James Marshall , Anthony McCosker , Yong-Bin Kang\n- **URL**: [https://arxiv.org/abs/2606.27619](https://arxiv.org/abs/2606.27619)\n- **Abstract**:\n> Dyslexic learners increasingly use artificial intelligence (AI) tools to support reading, writing, organisation, and study-related tasks. However, their lived experiences with these tools remain largely underexamined. This paper proposes DysLexLens, a low-resource LLM framework, designed to analyse dyslexic learners experience with AI through online forum discussions. DysLexLens is designed as an end-to-end, evidence-traceable architecture which transforms noisy social media posts into a dictionary-driven corpora, provides knowledge-graph (KG)-based question reasoning, generates verifiable query responses, and enables response evaluation through quantitative and human-grounded assessment. DysLexLens has four key features. First, it employs a dictionary-driven filtering method to construct a more focused Reddit corpus on dyslexia and AI, filtering out noisy and weakly related posts to improve the relevance of data collected from low-resource forum contexts. Second, it integrates LLM-assisted semantic analysis with KG-based query reasoning to uncover meaningful patterns. Third, it has quantitative evaluation metrics (RAGAS and Query Robustness) to measure LLM-generated response performance. Fourth, it provides structured qualitative validation guidelines for assessing response quality, with a specific focus on hallucination and evidence alignment. We demonstrate the effectiveness of DysLexLens using dyslexia-related Reddit forum data and 30 questions. The results show its potential generalisability to other low-resource forum data contexts. DysLexLens, sample data, questions and evaluation results are available at Github to support reproducibility.\n\n---\n\n## 8. Odyssey: Constructing Verifiable Local Truth-Preserving Foundation Models\n- **Authors**: Sridhar Mahadevan\n- **URL**: [https://arxiv.org/abs/2606.27593](https://arxiv.org/abs/2606.27593)\n- **Abstract**:\n> We introduce a categorical framework called ODYSSEY for constructing verifiable, local truth-preserving foundation models as compositions of foundries: building-block architectural components that specify a cover of local contexts, local representation families, restriction maps, gluing rules, obstruction policies, update obligations, and human-facing views. A foundry is an organized sheaf of knowledge that carries within it an argumentation component. Concrete foundries are built from generic foundries such as evidence/argument, operational decision, institutional/financial, market meaning, scientific challenge, research-program, assistant-build, and evaluation-harness foundries. Universal Foundry Learning (UFL) formalizes foundry construction as a composition of left and right Kan extensions, with left Kan extension rolling local artifacts into candidate foundries and right Kan extension enforcing the restriction, gluing, obstruction, and argumentation conditions required for promotion. Foundry SQL (FSQL) is a small typed query surface for slicing maintained foundry artifacts that uses TICKET (Topos Integration using Causal Kan Extension Transformers) certification for admitting external or pre-built models into durable ODYSSEY state. ODYSSEY is fully implemented and tested across a wide spectrum of concrete foundries, showing that the same categorical machinery supports domain construction, artifact replay, sheaf diagnostics, grounded Toulmin/local-LLM scrutiny, residual-obstruction ledgers, and optimized TICKET-compatible causal-claim extraction across heterogeneous sources. This paper is to be presented as a 2.5 hour tutorial at ICML 2026. The tutorial home page is at this https URL .\n\n---\n\n## 9. Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning\n- **Authors**: Xuan Zhang , Zhijian Zhou , Lingfeng Qiao , Yulei Qin , Ke Li , Xing Sun , Xiaoyu Tan , Chao Qu , Yuan Qi\n- **URL**: [https://arxiv.org/abs/2606.27483](https://arxiv.org/abs/2606.27483)\n- **Abstract**:\n> Large language model (LLM) agents have demonstrated strong capability in sequential decision-making, yet they remains fundamentally reactive in long-horizon tasks. Unlike humans who employ \"what-if\" reasoning to evaluate potential plans before commitment, standard agents lack an internal world model to simulate future outcomes. Therefore, we propose to internalize future-aware planning by training a single autoregressive model to verbalize both a prospective state rollout and a plan-conditioned success estimate-a textual analogue of the Q-value. Crucially, we identify a format-capability gap: simply fine-tuning agents on look-ahead traces during post-training leads to superficial mimicry of foresight without genuine predictive grounding. To bridge this gap, we introduce a three-stage training paradigm: (i) World Model Agentic Mid-Training (WM-AMT) to inject latent predictive capabilities into the policy; (ii) Format-Eliciting SFT (FE-SFT) to structure this injected capability; and (iii) Foresight-Conditioned Reinforcement Learning (FC-RL) to refine the calibration and utility of the generated simulations. Evaluated on search and mathematical reasoning tasks, our approach consistently outperforms other training baselines. Our results demonstrate that effective internal world modeling in LLM agents requires a capability-first training pipeline to achieve grounded and calibrated foresight.\n\n---\n\n## 10. When Does Personality Composition Matter for Multi-Agent LLM Teams?\n- **Authors**: Aryan Keluskar , Amrita Bhattacharjee , Huan Liu\n- **URL**: [https://arxiv.org/abs/2606.27443](https://arxiv.org/abs/2606.27443)\n- **Abstract**:\n> Personality prompting shapes how large language models communicate, yet whether these behavioral shifts affect objective task outcomes remains under-explored. Prior work shows that agents prompted with low agreeableness produce adversarial language, while those prompted with high agreeableness become cooperative, but the relationship between communication style and task performance has not been systematically examined across multiple domains. In this work, we investigate whether personality composition matters for multi-agent team performance by manipulating personality traits across frontier LLMs on three task domains: structured coding, open-ended research collaboration, and competitive bargaining. We find that personality effects depend critically on task structure. In coding tasks, low agreeableness leads to large communication shifts that have little effect on milestone completion. In open-ended collaboration and bargaining, the same manipulation substantially degrades performance. We discuss implications for multi-agent system design and the limits of personality manipulation.\n\n---\n\n## 11. AI-Model Network: Concept, Current State and Future\n- **Authors**: Li Zhetao , Zeng Xiyu , Wang Jianhui , Xiao Yong , Liu Zhongren , Wu Junru , Lai Junjie , Huang Jijun , Long Saiqin\n- **URL**: [https://arxiv.org/abs/2606.27382](https://arxiv.org/abs/2606.27382)\n- **Abstract**:\n> While the primary function of computers lies in computation and processing, the core value of the Internet is rooted in sharing and collaboration. Computers create the Internet, and the Internet empowers the value of computers. The rapid development of the Internet, cloud computing, and big data is pushing artificial intelligence into the era of large models (LMs). However, the practical application of LMs is currently hindered by high training costs and deployment complexities, driving a shift toward lightweight, private, and domain-specific models. With the rapid proliferation and wide distribution of heterogeneous models, enabling effective interaction and collaboration among them has emerged as a critical bottleneck that urgently needs to be addressed in LM development. Drawing inspiration from the development of the Internet, this paper proposes the concept, vision, and system architecture of world wide AI-model network (AI-ModelNet). It is a novel paradigm that achieves interconnection, capability sharing, and collaborative reasoning by establishing pathways between models. We first briefly review the current state of single-model and multi-model research. Subsequently, the systemic vision and hierarchical architecture of AI-ModelNet are articulated, followed by validation of the framework's feasibility through a prototype system and diverse application cases. Finally, key directions for future research are discussed preliminarily.\n\n---\n\n## 12. Cognitive Episodes in LLM Reasoning Traces Enable Interpretable Human Item Difficulty Prediction\n- **Authors**: Chenguang Wang , Ming Li , Xinyue Zeng , Zhuochun Li , Hong Jiao , Tianyi Zhou , Dawei Zhou\n- **URL**: [https://arxiv.org/abs/2606.28186](https://arxiv.org/abs/2606.28186)\n- **Abstract**:\n> Predicting human item difficulty is central to educational assessment, where reliable estimates support fairness and effective test construction. Existing methods often depend on costly human calibration or item-level textual representations, providing limited evidence about the cognitive processes that make items difficult. We argue that difficulty should be viewed not only as a property of item text, but also as an observable consequence of the problem-solving burden an item induces. Large Reasoning Models (LRMs) offer scalable process evidence through reasoning traces, but such evidence must be structured to support interpretable modeling. To this end, we introduce Epi2Diff (Episode to Difficulty), a framework that maps LRM reasoning traces into cognitively grounded episode sequences. These episodes group trace segments into functional problem-solving states, enabling difficulty to be modeled through reasoning scale, effort allocation, and state transitions. Epi2Diff extracts compact episode-dynamic features and combines them with semantic item representations for human difficulty prediction. Experiments on four real-world human difficulty datasets show that Epi2Diff consistently outperforms strong baselines, including fine-tuned small language models, LLM in-context learning, and supervised LLM adaptation. On SAT-derived classification benchmarks, Epi2Diff achieves an 8.1% average relative gain over supervised LLM fine-tuning baselines. Further analyses show that harder items induce more effortful, iterative, and implementation-centered episode dynamics, rather than merely longer responses. These results demonstrate that cognitive episodes in LRM reasoning traces provide a predictive and interpretable process representation for human item difficulty, offering a new lens for educational measurement with reasoning models.\n\n---\n\n## 13. LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior\n- **Authors**: Qinhong Zhou , Chuang Gan , Anoop Cherian\n- **URL**: [https://arxiv.org/abs/2606.28182](https://arxiv.org/abs/2606.28182)\n- **Abstract**:\n> Embodied agents operating in decentralized and partially observable environments have attracted growing attention in recent years. However, existing large language model (LLM)-based agents often exhibit behaviors that are misaligned with their partners or inconsistent with the environment state, leading to inefficient cooperation and poor task success. To address this challenge, we propose a novel framework, Learning Laws of Cooperation (LLawCo), that enables embodied agents to autonomously align with both their partners and task objectives. Our framework allows agents to reflect on past failures to extract misaligned behavioral patterns, which are used to derive high-level behavioral laws, such as \"Talk when necessary\" and \"Wait for partner.\" These laws are explicitly incorporated into the agents' chains of thought via supervised fine-tuning, aligning their reasoning with task requirements and the behavior of other agents. To evaluate our approach, we introduce PARTNR-Dialog, a large-scale multi-agent communicative and cooperative planning benchmark built on the PARTNR environment. Experiments on existing tasks and our new benchmark demonstrate significant improvements in cooperative efficiency and task success rates. Across four backbone LLMs, our method achieves average success rate improvements of 4.5% on the PARTNR-Dialog benchmark and 6.8% on the TDW-MAT benchmark over state-of-the-art open-source communicative agent frameworks. See the LLawCo project page for details: this https URL\n\n---\n\n## 14. Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models\n- **Authors**: Yanchen Yin , Dongqi Han , Linghui Li\n- **URL**: [https://arxiv.org/abs/2606.28153](https://arxiv.org/abs/2606.28153)\n- **Abstract**:\n> Jailbreak attacks bypass LLM safety alignment, yet their mechanisms remain poorly understood. We provide evidence that attacks do not comprehensively eliminate safety features, but instead selectively suppress specific attention heads. We identify two functionally differentiated types: Adversarially Compromised Heads (ACHs) concentrated in early layers, which are suppressed under attacks, and Safety-Aligned Heads (SAHs) in mid-layers, which maintain robust activations even when attacks succeed. Ablation studies support the causal role of ACHs and the contribution of SAHs to robust activations: suppressing a small number of ACHs is sufficient to induce jailbreak-like behavior on normally refused inputs, while removing SAHs substantially weakens mid-layer safety activations. Token-level attribution further shows that ACH suppression is driven specifically by attack-template tokens, providing a mechanistic account of why attacks can bypass refusal decisions through ACH suppression while leaving internal safety signals sustained by SAHs -- a phenomenon we term Robust Harmful Features. To validate the practical significance of this robustness, we show that simply reading these persistent activations -- without any training -- yields competitive aggregate detection performance with strong adversarial robustness.\n\n---\n\n## 15. From Tokens to States: LLMs as a Special Case of World Models and the Continuous Path Beyond\n- **Authors**: Paul Dubois\n- **URL**: [https://arxiv.org/abs/2606.28127](https://arxiv.org/abs/2606.28127)\n- **Abstract**:\n> The AI community has framed the relationship between large language models (LLMs) and world models as a dichotomy: LLMs predict tokens; world models simulate reality. Yann LeCun argues in 2022 that reaching general intelligence requires abandoning autoregressive token prediction in favour of latent-space architectures. This framing is unnecessarily binary. Two claims will be defended. First, LLMs are a degenerate special case of world models: the state space is the set of all token sequences, the only action is appending one token, and world models are therefore a strict generalisation of LLMs, not a replacement. Second, there is a natural continuous spectrum from NTP to JEPA, with multi-token prediction, future-summary prediction, and next-latent prediction as intermediate stations already populated by current research. Moving along this spectrum relaxes the LLM constraints one by one. It also progressively surrenders the two practical advantages that make LLMs trainable at scale: internet-scale self-supervised data, and a transformer architecture co-designed for discrete token prediction. Both are examined as open research questions: the data question (the cliff from self-supervised text to instrumented action-labelled environments) and the architecture question (whether the transformer generalises to continuous-state prediction, or whether a new primitive is needed).\n\n---\n\n## 16. Single and Multi Truth Data Fusion using Large Language Models\n- **Authors**: Hira Beril Kucuk , Norman W Paton , Jiaoyan Chen , Zhenyu Wu\n- **URL**: [https://arxiv.org/abs/2606.28062](https://arxiv.org/abs/2606.28062)\n- **Abstract**:\n> Data fusion, also known as truth discovery, is a data integration problem that aims to determine the correct value or set of values for each attribute of an object when presented with potentially conflicting values from multiple sources. Data fusion tasks belong to two main categories: single-truth scenarios, where each attribute has only one correct value, and multi-truth scenarios, where multiple values can be valid simultaneously. This paper investigates the use of Large Language Models (LLMs) in data fusion tasks for tabular data. Various prompting strategies, encompassing both single-truth and multi-truth scenarios, are investigated empirically. Domain-dependent, domain-independent, zero-shot and one-shot prompts are evaluated on three different benchmark datasets. Experimental results demonstrate that LLM-based approaches outperform traditional unsupervised truth discovery methods, such as DART and LTM, across all datasets. The codebase of this study has been made publicly available on GitHub.\n\n---\n\n## 17. ToolPrivacyBench: Benchmarking Purpose-Bound Privacy in Tool-Using LLM Agents\n- **Authors**: Shijing Hu , Liang Liu , Zhu Meng , Zhicheng Zhao\n- **URL**: [https://arxiv.org/abs/2606.28061](https://arxiv.org/abs/2606.28061)\n- **Abstract**:\n> Large language models (LLMs) have increasingly moved from standalone text generation systems to agents that invoke external tools, access environments, and execute multi-step tasks. However, conventional function-calling benchmarks mainly evaluate task completion and API correctness, while privacy evaluation benchmarks typically focus on final responses or privacy judgments. Neither perspective captures purpose-bound information flow across an executed multi-tool trajectory. Motivated by this limitation in current agent evaluation, ToolPrivacyBench audits whether task-private atoms are routed only to authorized tools and downstream sinks, thereby evaluating both task completion and privacy over-disclosure during tool use. The benchmark contains 2,150 cases, including 1,150 fully synthetic privacy-sensitive business workflows and 1,000 cases adapted from existing multi-tool and function-calling benchmarks. Each case is represented by a policy knowledge base. After an agent executes against mock business backends, the evaluator compares recorded tool arguments and backend audit logs with this policy knowledge base. The evaluation covers nine widely used agents to characterize purpose-bound privacy over-disclosure. The results show that successful tool execution does not imply appropriate privacy disclosure: an agent may complete a task while transmitting unnecessary private information through intermediate tool calls. ToolPrivacyBench therefore formalizes a need-to-know disclosure boundary, under which each tool should receive only the information necessary for its stated purpose, and uses trajectory-level auditing to identify privacy over-disclosure in multi-tool workflows.\n\n---\n\n## 18. MultiHashFormer: Hash-based Generative Language Models\n- **Authors**: Huiyin Xue , Atsuki Yamaguchi , Nikolaos Aletras\n- **URL**: [https://arxiv.org/abs/2606.28057](https://arxiv.org/abs/2606.28057)\n- **Abstract**:\n> Language models (LMs) represent tokens using embedding matrices that scale linearly with the vocabulary size. To constrain the parameter footprint, prior work proposes hashing many tokens into a single vector within encoder-only models. While this offers parameter efficiency, many-to-one collisions prevent its use in causal LMs. In this paper, we propose MultiHashFormer, a new framework that allows hash-based autoregression. Each token is represented as a unique hash signature, a short sequence of discrete hash IDs, generated by multiple independent hash functions. A Hash Encoder compresses this signature into a single latent vector for processing by a Transformer decoder. Then, a Hash Decoder generates the hash signature of the next token, which is then mapped back to text. We evaluate our approach at the 100M, 1B and 3B parameter scales, demonstrating that MultiHashFormer consistently outperforms standard Transformer LMs across multiple benchmarks. Furthermore, we show that our model handles multilingual vocabulary expansion with a constant parameter footprint without any modifications.\n\n---\n\n## 19. Can LLMs Judge Better Than They Generate? Evaluating Task Asymmetry, Mechanistic Interpretability and Transferability for In-Context QA\n- **Authors**: Sambaran Bandyopadhyay\n- **URL**: [https://arxiv.org/abs/2606.28050](https://arxiv.org/abs/2606.28050)\n- **Abstract**:\n> LLM-as-a-Judge and self-evaluation pipelines implicitly assume that evaluation is easier than generation. We test this in a controlled in-context QA setting where a context passage is the sole information source and each model judges the answer it generated, removing the parametric-knowledge confound of open-domain comparisons. Across four benchmarks (SQuAD 2.0, DROP, HotpotQA, MuSiQue) and two models, evaluation is not uniformly easier: generation accuracy exceeds self-evaluation on three of four, with multi-hop MuSiQue the exception. Attention analysis reveals why: evaluation attends to context 3--5x less than generation does and barely reads the candidate answer. LoRA fine-tuning confirms the asymmetry is not a training artifact: generation fine-tuning induces over-acceptance and evaluation fine-tuning degrades generation. These findings challenge core assumptions in self-evaluation pipelines.\n\n---\n\n## 20. Dialogue to Detection: A Multimodal Hybrid NLP Pipeline for Insurance Fraud Detection\n- **Authors**: Muhammad Shakeel Akram , Amal Htait , Abdul Hamid Sadka , Emma Meisingseth , Karishma Jaitly\n- **URL**: [https://arxiv.org/abs/2606.28002](https://arxiv.org/abs/2606.28002)\n- **Abstract**:\n> Insurance fraud imposes substantial financial losses and operational inefficiencies, raising premiums and impacting trust among legitimate policyholders. Early detection at FNOL remains a persistent challenge. Existing approaches rely largely on private, text-only datasets, limiting progress on multimodal methods that integrate linguistic, behavioural, and speaker-based indicators. We introduce a synthetic multimodal framework that replicates FNOL conditions. It generates agent-customer dialogue transcripts and two-speaker audios, performs ASR and diarisation. Downstream modules combine NER, regex-based feature extraction, LLM-RAG retrieval, and speaker embeddings in a rule-based risk score to flag narrative reuse, structural inconsistencies, and cross-case voice repetition while balancing sensitivity and false positives. Dataset validation and component-level evaluations show stability and transfer potential, offering a reproducible baseline beyond text-only fraud detection.\n\n---\n\n## 21. From Black-Box to Clinical Insight: A Multi-Stage Explainable Framework for Speech-Based Cognitive Impairment Detection\n- **Authors**: Yasaman Haghbin , Sina Rashidi , Ali Zolnour , Fatemeh Taherinezhad , Ali Fartoot , Hossein Azadmaleki , James M Noble , Maryam Dadkhah , Maryam Zolnoori\n- **URL**: [https://arxiv.org/abs/2606.27973](https://arxiv.org/abs/2606.27973)\n- **Abstract**:\n> Speech-based cognitive impairment detection offers a noninvasive, accessible alternative to costly biomarker assays, yet transformer-based models remain clinically uninterpretable. We propose a multi-stage explainability framework that translates black-box transformer predictions into clinically grounded narratives by integrating SHapley Additive exPlanations (SHAP)-based token attribution, theory-informed linguistic features, and a four-stage LLM reasoning pipeline using LLaMA-3.1-70B-Instruct. Built on the SpeechCARE-Adaptive Gating Network multimodal screening model (F1 = 72.11% on the NIA PREPARE benchmark), the framework maps model outputs to four cognitive-linguistic dimensions, including lexical richness, syntactic complexity, and semantic coherence. Physician evaluation on 70 stratified English samples demonstrated strong alignment with patient-level cognitive profiles, and a System Usability Scale score of 82/100 indicated high potential for clinical workflow integration.\n\n---\n\n## 22. Two-Stage Fine-Tuning for Protein Sequence Generation with Targeted Amino-Acid Composition\n- **Authors**: Violeta Basten-Romero , Rubén Muñoz-Tafalla , Anna María Díaz-Rovira , Bertran Miquel-Oliver , Isaac Filella-Merce , Víctor Guallar\n- **URL**: [https://arxiv.org/abs/2606.27939](https://arxiv.org/abs/2606.27939)\n- **Abstract**:\n> Protein language models are standard priors for biological sequence generation, but steering them toward explicit distributional design targets remains largely unexplored. We study a constrained protein generation problem in which sequences must match a desired amino-acid (AA) composition profile while preserving plausible sequence statistics and diversity. The motivating application is synthetic feed protein design, where the AA composition of dietary proteins directly determines their nutritional value. We propose a two-stage pipeline in which domain-adaptive fine-tuning (FT) on an in-domain protein dataset is followed by iterative reward-weighted FT via reinforcement learning (RL) anchored against the FT model as a frozen reference. We evaluate the pipeline on two AA compositions and find that FT brings the average composition close to the target, while the subsequent RL enforces specific sequence constraints that FT alone cannot satisfy. We additionally evaluate the design choices of the proposed composition reward term against two baselines and an ablated variant, isolate the contribution of each training stage, and verify that AA composition alignment is achieved without degrading sequence quality.\n\n---\n\n## 23. Agentic AI-Powered Re-Identification: An Emerging, Scalable Threat to Mobility Microdata Privacy\n- **Authors**: Oscar Thees , Roman Müller , Matthias Templ\n- **URL**: [https://arxiv.org/abs/2606.27936](https://arxiv.org/abs/2606.27936)\n- **Abstract**:\n> The widespread collection of fine-grained location data by commercial data brokers creates a re-identification risk that is not widely recognised by the public. While prior research has established that mobility traces are highly unique and that individuals can, in principle, be identified from a handful of spatio-temporal points, such attacks have historically required significant manual effort from skilled analysts, limiting their practical scale. In this feasibility study, we demonstrate in a real world setting that agentic AI fundamentally changes this threat model. We present an end-to-end pipeline in which large language model agents autonomously search the open web, cross-reference public records and social media, and resolve raw coordinate sequences to candidate identities - without human intervention. We evaluate the pipeline on a spatio-temporal dataset containing simulated location points anchored at and around true home and work addresses, focusing on a high-risk disclosure scenario. Our results demonstrate that, from spatio-temporal data and public sources alone, our agentic AI successfully re-identified 18 of the 25 re-identifiable individuals (72%) and 18 of 43 cases overall (41.9%). We discuss implications for Statistical Disclosure Control (SDC) practice and outline the near-future escalation that data custodians and regulators must anticipate. De facto anonymity - an implicit foundation of SDC practice - is shifting. Agentic AI strengthens the case that re-identification is reasonably likely by any means under the GDPR Recital-26 standard, at costs of minutes-and-dollars per target.\n\n---\n\n## 24. Triadic Werewolf: A Jester Role for Multi-Hop Theory of Mind in LLMs\n- **Authors**: Avni Mittal\n- **URL**: [https://arxiv.org/abs/2606.27909](https://arxiv.org/abs/2606.27909)\n- **Abstract**:\n> Theory-of-mind evaluations of large language models typically use dyadic social-deduction games, where every observable cue points to a single hidden side, so a model with strong language priors can score well without ever simulating opponents' incentives. We extend the Werewolf game with a Jester, a third faction whose utility on peer suspicion is inverted because it wins by being voted out, so optimal play requires reasoning across three opposing utility functions. Across 60 games on GPT-4.1, DeepSeek-V3.1, and Llama-3.3-70B with Jester self-learning on and off, the Jester wins 60-70% of games while Werewolves never exceed 20%, and GPT-4.1 wolves vote the Jester out on day 1 in 60-70% of games, a strictly self-defeating action. Self-learning helps DeepSeek and Llama but hurts GPT-4.1, with the cost landing on Villagers rather than Werewolves. Only DeepSeek learns the subtle strategy of looking suspicious without looking intentionally suspicious, and it gains the most from the loop. Triadic incentive structure exposes a layer of multi-agent reasoning that dyadic deduction games leave invisible.\n\n---\n\n## 25. A Study of Temporal Fusion Strategies for Named Entity Recognition in Historical Texts\n- **Authors**: Emanuela Boros\n- **URL**: [https://arxiv.org/abs/2606.27881](https://arxiv.org/abs/2606.27881)\n- **Abstract**:\n> Temporal variation poses a unique challenge for named entity recognition (NER) in historical texts, where entities drift in surface form and salience across time. While language models (LMs) have made progress in various NLP tasks, their ability to reason about temporality, especially in diachronic contexts, remains limited or at least, questionable. In this paper, we systematically study how temporal metadata can be structurally embedded into NER models using a range of lightweight fusion strategies. We experiment with both absolute and relative temporal representations, injected into Transformer-based architectures via early or late fusion mechanisms such as cross-attention, adapters, and concatenation. Our evaluations on French and German historical datasets reveal that late fusion strategies yield more robust and temporally generalisable performance, particularly in early and noisy periods.\n\n---\n\n## 26. SpatialUAV: Benchmarking Spatial Intelligence for Low-Altitude UAV Perception, Collaboration, and Motion\n- **Authors**: Haoyu Zhang , Meng Liu , Qianlong Xiang , Kun Wang , Yaowei Wang , Liqiang Nie\n- **URL**: [https://arxiv.org/abs/2606.27876](https://arxiv.org/abs/2606.27876)\n- **Abstract**:\n> Spatial intelligence is essential for low-altitude unmanned aerial vehicle (UAV) perception, collaboration, and navigation. However, existing UAV benchmarks often emphasize image-level recognition, single-view understanding, or narrow answer formats, leaving 3D spatial inference, multi-view collaboration, scene dynamics, and diverse task formulations insufficiently evaluated. To address these gaps, we introduce SpatialUAV, a real low-altitude UAV benchmark comprising 4,331 curated instances across 14 fine-grained task types, covering semantic discrimination, spatial relation, aerial--aerial collaboration, aerial--ground collaboration, and motion understanding. SpatialUAV organizes all samples into a unified visual-input--question--answer schema, while supporting seven input configurations and nine answer formats, including option labels, region identifiers, geometric values, cross-view correspondences, and free-form motion descriptions. To ensure reliable and grounded evaluation, our data construction pipeline integrates detector-assisted regions, depth supervision, metadata-derived rules, extensive manual annotation, blind filtering, and multi-turn human validation, together with task-specific metrics for heterogeneous outputs. Evaluating representative vision-language models across three categories, we show that current models remain far from human-level performance, with pronounced bottlenecks in cross-view association, structured grounding, geometric reasoning, and temporal viewpoint understanding. These results offer empirical guidance for advancing low-altitude UAV spatial intelligence. Code and data are available at this https URL .\n\n---\n\n## 27. Position Bias Correction is Insufficient for One-Pass Attention Sorting\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27793](https://arxiv.org/abs/2606.27793)\n- **Abstract**:\n> Long-context language models suffer from position bias, where information in middle positions is underutilized. Attention Sorting addresses this by iteratively reordering documents based on attention patterns, but its multiple sort-and-generate cycles increase deployment cost. We hypothesize that position bias is the primary bottleneck and propose Debiased One-Pass Attention Sorting, which estimates a per-prompt position-bias curve from the low-attention majority of documents and uses it to correct raw attention scores (via subtraction or division) to enable single-pass sorting. Our experiments on two models refute this hypothesis in the tested setting: on LLaMA-2-7B-32K-Instruct, debiasing produces identical results to uncalibrated single-pass sorting (94.83\\% containment accuracy), while on YaRN-Llama-2-7b-64k, debiasing improves accuracy by 8.67 percentage points but remains 14.84pp behind iterative sorting, closing only 37\\% of the gap. These results suggest that position-bias correction is insufficient to match iterative sorting, and that repeated reordering provides additional benefits beyond bias correction.\n\n---\n\n## 28. NLL-Guided Full-Attention Layer Selection for Training-Free Sliding-Window Adaptation\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27791](https://arxiv.org/abs/2606.27791)\n- **Abstract**:\n> Hybrid attention models that mix full and sliding-window attention across layers offer a promising approach to efficient long-context inference, but the critical question of \\emph{which layers} should retain full attention remains unsolved. Existing methods use either fixed periodic patterns or attention-based heuristics that may not capture what matters for downstream accuracy. We propose NLL-guided layer selection, a training-free method that directly measures each layer's importance by computing the negative log-likelihood degradation on answer tokens when that layer uses sliding-window instead of full attention. On LongMemEval with Qwen3-4B, our method achieves 64.6\\% accuracy using only 1/4 full-attention layers, matching the 1/2-FA periodic baseline (65.0\\%) while halving the computational budget. NLL-guided selection outperforms the SWAA-reported periodic 1/4-FA baseline by 10.4 percentage points and a matched LightTransfer-style baseline by 26.4 percentage points. De-confounding analysis shows the signal is consistent with long-range attention needs rather than generic layer sensitivity. The method requires only $\\sim$15 minutes of one-time calibration, advancing the efficiency-accuracy Pareto frontier for long-context LLM deployment.\n\n---\n\n## 29. Output-Space Allocation Costs for Calibration-Guided LLM Compression: An Empirical Study\n- **Authors**: Qiong Tang , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao\n- **URL**: [https://arxiv.org/abs/2606.27785](https://arxiv.org/abs/2606.27785)\n- **Abstract**:\n> Training-free compression methods for large language models (LLMs) often use calibration data to guide compression decisions. ROCKET, a recent method combining sparse-dictionary factorization with multi-choice knapsack problem (MCKP) allocation, derives its per-layer factorization from an output reconstruction objective but uses weight-space Frobenius error as the MCKP allocation cost. We investigate whether aligning the allocation cost with the output-space objective improves compressed model fidelity. On Qwen3-8B at 50\\% compression, our ROCKET-ActCost achieves +0.8 percentage points higher average accuracy across 8 zero-shot benchmarks (53.1\\% vs 52.3\\%), but increases WikiText perplexity by 16\\% (61.46 vs 52.98). This accuracy-perplexity tradeoff reveals that different allocation objectives favor different downstream metrics. The high correlation ($>$0.99) between weight-space and output-space errors limits allocation divergence, explaining the modest effect size. On Llama-3.2-1B at 20\\% compression, the two methods produce near-identical results (53.3\\% vs 53.5\\% accuracy, 14.45 vs 14.66 PPL), suggesting that the effect of the cost function is minor at lower compression ratios.\n\n---\n\n## 30. Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?\n- **Authors**: Guoheng Sun , Kaixi Feng , Shwai He , Xiaochuan Gong , Yexiao He , Ziyao Wang , Zheyu Shen , Wanghao Ye , Ramana Rao Kompella , Gaowen Liu , Ang Li\n- **URL**: [https://arxiv.org/abs/2606.27755](https://arxiv.org/abs/2606.27755)\n- **Abstract**:\n> Vision-Language-Action (VLA) models enable instruction-driven robotic manipulation, but they inherit oversized language backbones from pretrained VLMs whose capacity far exceeds what is needed for short robotic instructions. This raises a basic question: how much of a VLA model is actually necessary for closed-loop control? In this work, we study architectural redundancy in VLA models by using transformer block removal as a controlled intervention. We introduce \\textbf{Drop-Then-Recovery (DTR)}, an analysis protocol that removes selected blocks from a pretrained VLA model and then fine-tunes the resulting model to measure whether the removed capacity was necessary for downstream control. To make this intervention reliable, we propose \\textbf{GateProbe}, a one-shot virtual-gate sensitivity metric that ranks blocks by their contribution to the downstream action loss. Across multiple VLA architectures, manipulation benchmarks and even real-robot industrial scenarios, we find a strong asymmetry in post-removal recoverability: \\ul{\\textit{language backbones are highly redundant for standard robotic manipulation tasks, whereas vision and action pathways are substantially less tolerant to removal} }. On LIBERO, removing half of the LLM blocks even improves OpenVLA-OFT from 95.0% to 98.3% under the same downstream fine-tuning budget, and retaining only two language blocks still recovers baseline-level performance. These results suggest that current VLA benchmarks may exert limited pressure on deep language grounding and compositional instruction understanding, and that future VLA architectures should allocate capacity more deliberately across language, vision, and action components. The code is available at this https URL .\n\n---\n\n## 31. End-to-End Dynamic Sparsity for Resource-Adaptive LLM Inference\n- **Authors**: Yuhang Chen , Jinhao Duan , Ruichen Zhang , Mingfu Liang , Xiaohan Wei , Yunchen Pu , Fei Tian , Chonglin Sun , Parish Aggarwal , Frank Shyu , Luke Simon , Sandeep Pandey , Tianlong Chen , Xi Liu\n- **URL**: [https://arxiv.org/abs/2606.27743](https://arxiv.org/abs/2606.27743)\n- **Abstract**:\n> Large Language Models (LLMs) inference is typically deployed under a static resource assumption, where models execute a fixed computational graph regardless of the runtime environment. However, real-world cloud infrastructure is inherently dynamic, characterized by fluctuating availability (e.g., spot instance preemption) and tiered Quality-of-Service requirements. In such volatile settings, static models are inflexible: they either crash under resource constraints or waste compute on redundant operations. To bridge this gap, we propose Learning to Allocate (L2A), an end-to-end framework for resource-adaptive inference. Unlike prior methods that condition only on input difficulty, we formulate inference as a constrained allocation problem conditioned on both the input and the runtime resource budget itself. We introduce lightweight, budget-conditioned and input-aware gating networks integrated into the LLM. These gates are trained via a unified objective that jointly optimizes task performance, logical consistency, and resource costs along three axes matching how real-world dynamics manifest: layer skipping for memory and depth pressure, head pruning for throughput contention, and reasoning-token reduction for latency tightening. This lets the model learn a budget-aware policy beyond input difficulty alone: it adaptively configures its computational footprint with respect to real-time resource dynamics, maximizing reasoning depth when resources permit while enforcing strict frugality when budgets tighten. A single L2A model traces the entire compute-accuracy Pareto frontier on Llama-3-8B and Qwen-3-4B: at up to 34% realized layer sparsity, it stays within 0.6% of the dense baseline on GSM8K, with the same gap holding zero-shot on out-of-distribution tasks, while every static or heuristic baseline requires a separately tuned model and still drops by 5-10% at comparable inference time.\n\n---\n\n## 32. KG2Cypher: Data-Centric Pipeline for Building Enterprise Text-to-Cypher Systems\n- **Authors**: Minjun Choi , Yerin Kim , Junghyuk Seo , Sujin Mo , Hyemin Lee , Youngjoong Ko\n- **URL**: [https://arxiv.org/abs/2606.27742](https://arxiv.org/abs/2606.27742)\n- **Abstract**:\n> Enterprise Knowledge Graphs (KGs) are increasingly used for internal search, analytics, and question answering, but building natural-language interfaces for private enterprise graphs remains costly. We present KG2Cypher, a data-centric pipeline for building enterprise text-to-Cypher systems from existing KGs. KG2Cypher first constructs an executable Cypher query from observed graph facts and then uses LLMs to generate its associated natural-language question. The resulting Text-Cypher pairs are validated with an LLM judge and human validation, and are converted into candidate-aware SFT data. The trained generator is served with class-conditioned schema prompting, entity retrieval, and LoRA-based inference. We evaluate KG2Cypher in Korean enterprise settings, where short search-style queries and schema paraphrases make language grounding difficult. LoRA SFT improves execution-result F1 from 0.806 to 0.950 on broadcast-program queries and from 0.70 to 0.92 on company queries. In an 11-class setting, KG2Cypher achieves 95.2% exact match, 99.9% execution rate, and 0.964 execution-result F1.\n\n---\n\n## 33. Bifocal Diffusion Language Models: Asymmetric Bidirectional Context for Parallel Generation\n- **Authors**: Yuhang Chen , Xianfeng Wu , Jinhao Duan , Mingfu Liang , Xiaohan Wei , Yunchen Pu , Fei Tian , Chonglin Sun , Parish Aggarwal , Frank Shyu , Luke Simon , Sandeep Pandey , Xi Liu , Tianlong Chen\n- **URL**: [https://arxiv.org/abs/2606.27732](https://arxiv.org/abs/2606.27732)\n- **Abstract**:\n> Discrete diffusion language models (dLLMs) recover masked tokens in parallel, offering significant speedups over autoregressive (AR) generation. However, such promising frameworks face a fundamental architectural design dilemma: \\ding{182} Adopting bidirectional attention achieves strong generation quality by allowing each position to access the full context, but is inherently incompatible with KV caching, limiting inference throughput in batch-serving scenarios; \\ding{183} Conversely, causal attention enables efficient cached inference but loses all right-side context, substantially degrading generation quality. This paper introduces Bifocal dLLMs, a new paradigm that resolves this dilemma through \\emph{asymmetric bidirectional context}. Analogous to bifocal lenses, we instantiate the paradigm as \\textbf{R2LM} (Right-to-Left Mamba), which combines two complementary mechanisms: $a$) standard causal attention providing precise left-context with full KV cache compatibility, while $b$) a lightweight reverse Mamba SSM sidecar supplying compressed right-side context without breaking cacheability. Comprehensive experiments on continued pretraining of Qwen3-1.7B with 60B tokens demonstrate that R2LM achieves $2.4\\times$ to $12.9\\times$ higher throughput than bidirectional dLLMs and $1.9\\times$ to $2.9\\times$ speedup over AR baselines in batch serving through parallel decoding with KV caching, while exceeding the causal baseline on most benchmarks and surpassing the bidirectional dLLM on average.\n\n---\n\n## 34. Enhancing Numerical Prediction in LLMs via Smooth MMD Alignment\n- **Authors**: Zhuo Zuo , Li Yue , Wenhao Zheng , Chenpeng Wang , Xianggen Liu\n- **URL**: [https://arxiv.org/abs/2606.27731](https://arxiv.org/abs/2606.27731)\n- **Abstract**:\n> Despite their strong general capabilities, large language models (LLMs) often remain unreliable when outputs must be numerically precise. A key reason is the training objective: standard cross-entropy treats numeric tokens as unstructured categories and ignores the metric structure of their values. We address this mismatch with Smooth Maximum Mean Discrepancy (SMMD), which builds on the classic MMD by incorporating value-distance kernels over numeric tokens and graph-based smoothness. With this kernel defined over a numeric sub-vocabulary, SMMD aligns the predicted numeric distribution to the target via kernel matching and smooths the prediction-target residual over the induced kernel graph to encourage local consistency. We evaluate SMMD on four numeric-target tasks: mathematical reasoning, arithmetic calculation, clock-time recognition, and chart question answering, across multiple open-weight LLM and VLM backbones. SMMD consistently improves accuracy over both cross-entropy and recent numeric-target losses; analyses show complementary effects between MMD and smoothness and underscore the importance of distance-based kernel design. Code is available at this https URL .\n\n---\n\n## 35. Low-Agreeableness Persona Conditioning for Safe LLM Fine-Tuning\n- **Authors**: Austin MY Cheung , Yi Yang\n- **URL**: [https://arxiv.org/abs/2606.27709](https://arxiv.org/abs/2606.27709)\n- **Abstract**:\n> Recent work has shown that fine-tuning large language models (LLMs) for social warmth degrades factual reliability and increases sycophancy. We investigate a related but distinct failure mode: warmth fine-tuning also weakens adversarial safety, making models more susceptible to jailbreaks and harmful output generation. We examine whether this reflects an inherent consequence of empathetic adaptation or an artifact of data construction. To address this, we introduce a persona-driven rewriting pipeline that conditions user turns on low agreeableness and pairs this with warm, de-escalating assistant responses. Across three experiments on four models, our approach reduces jailbreak susceptibility and harmful output rates relative to generic warmth fine-tuning baselines, while preserving conversational warmth. Representational probing provides suggestive evidence that this conditioning reduces the geometric alignment between warmth and compliance directions in latent space. These results show that safer empathetic fine-tuning is achievable through data design alone, without safety labels, harm detectors, or changes to the training objective.\n\n---\n\n## 36. Mitigating LLM-based p-Hacking by Preregistering for the Next LLM\n- **Authors**: Maria Thomas , Kristina Gligoric , Nihar B. Shah\n- **URL**: [https://arxiv.org/abs/2606.27687](https://arxiv.org/abs/2606.27687)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used to generate, classify, and annotate data whose outputs feed downstream hypothesis tests. However, LLM-based research is easy to p-hack: a researcher can tune the prompts, decoding parameters, or output format until a desired result is reached. We propose a protocol to mitigate p-hacking in LLM-based research: preregistering the experiment and eligible models, and then running it on the first eligible LLM that is released after the preregistration. The researcher finalizes the procedure on current models, preregisters the analysis plan together with a set of eligible future models, and runs the confirmatory analysis on the first eligible model released afterward. Because this model does not exist at commitment time, it cannot be hacked against; furthermore, configurations that hack one model frequently do not transfer to the next. We evaluate the protocol on two tasks whose true values are known. Across 20 models from four providers and 11 LLM-analysis configurations, the protocol would have blocked successful transfer of the p-hack in 73.9% and 72.7% of cases in the two tasks. Additional analyses reveal that mitigation remains substantial under several stress tests. Finally, putting money where our mouth is, we followed our own protocol and preregistered our experiment. The preregistered experiment confirmed the protocol's effectiveness: out of the 7 configurations that hacked the prior model, the hacking failed to carry over in 6 configurations on the first eligible model released afterward.\n\n---\n\n## 37. CBD: API-Only LLM Black-Box Unlearning through Controlled Behavioral Divergence\n- **Authors**: Zhiqiang Xie , Yijing Lin , Zhipeng Gao , Dong In Kim\n- **URL**: [https://arxiv.org/abs/2606.27683](https://arxiv.org/abs/2606.27683)\n- **Abstract**:\n> Edge devices increasingly invoke large language models (LLMs) through API services for context aware edge intelligence, while edge generated data may be collected to improve LLMs and may introduce sensitive, copyrighted, harmful, or outdated information into model behavior. Machine unlearning offers a practical way to remove the influence of undesired data without retraining LLMs. However, existing methods still face two gaps. The first is API only black box access, where target model parameters and internal logits are unavailable. The second is how to preserve retained utility when unlearning target data and retained data share highly similar prompt structures or semantic patterns. To address these challenges, we propose Controlled Behavioral Divergence (CBD), an API only black box unlearning framework. CBD uses two auxiliary models to create controlled behavioral divergence between retained inputs and unlearning target inputs, converts this divergence into an unlearning relevance score, and routes unlearning related prompts away from the target LLM. To improve discrimination accuracy under high similarity between target and retained data, CBD constructs a gradient statistics based discriminative basis by estimating empirical Fisher matrices and solving a regularized generalized eigenvalue problem, guiding the unlearning signal toward target specific information rather than shared prompt structures. Compared with eleven white box and gray box unlearning baselines, CBD achieves a better unlearning utility trade off and its performance varies little across settings. On ToFU forget10, CBD approaches the retrained reference on the forget set while raising model utility to 74.90, about 15% above the second best baseline. On WMDP, it lowers hazardous knowledge accuracy to 25.68, near random guessing, while preserving MMLU accuracy of 52.67. Code is at this https URL .\n\n---\n\n## 38. From Signals to Transfer: A Factorised Study of Probe-Based Uncertainty Estimation in Large Language Models\n- **Authors**: Ponhvoan Srey , Xiaobao Wu , Cong-Duy Nguyen , Quang Minh Nguyen , Duc Anh Vu , Anh Tuan Luu\n- **URL**: [https://arxiv.org/abs/2606.27679](https://arxiv.org/abs/2606.27679)\n- **Abstract**:\n> Probe-based uncertainty estimation (UE) has emerged as a prominent approach to detect hallucinations in Large Language Models (LLMs) by learning uncertainty from internal model signals. Yet, recent methods vary simultaneously across feature design, training data construction, and evaluation setting, obscuring what actually drives performance. To address this issue, we propose a factorised study of probe-based UE under matched conditions. Our results show that raw hidden states and attention features are difficult to outperform in-domain. However, under distribution shift, structured and compressed features are more robust, suggesting that in-domain performance alone is insufficient to measure progress. Furthermore, prompting and label construction significantly affect probe behaviour. Building on these best-practice findings, we train benchmark-based pretrained probes that transfer reasonably well to open-ended factual generation, providing a stable off-the-shelf baseline. Our work encourages more deployment-oriented evaluation of probe-based uncertainty estimators. The code repository is available at this https URL .\n\n---\n\n## 39. HybridCodec: Modeling Discrete and Continuous Representations for Efficient Speech Language Models\n- **Authors**: Artem Ploujnikov , Francesco Verdini , Samir Sadok , Mirco Ravanelli\n- **URL**: [https://arxiv.org/abs/2606.27627](https://arxiv.org/abs/2606.27627)\n- **Abstract**:\n> Discrete audio representations have become increasingly popular for building multimodal text-audio systems and integrating audio capabilities into Large Language Models (LLMs). However, numerous studies report performance degradation on various downstream tasks due to information loss during discretization. To address this, we propose a novel approach combining temporally compressed discrete tokens with dimensionality-reduced continuous residuals. Our framework consists of a hybridized discrete-continuous focal modulation codec and a hybrid Transformer. This architecture performs autoregressive inference in the discrete domain, coupled with non-autoregressive prediction and continuous residual upsampling. Experimental results show that our approach significantly improves the retention of speaker characteristics compared to discrete-only methods, while simultaneously reducing the number of required autoregressive steps.\n\n---\n\n## 40. Dismantling Pathological Shortcuts: A Causal Framework for Faithful LVLM Decoding\n- **Authors**: Liu Yu , Can Chen , Ping Kuang , Zhikun Feng , Fan Zhou , Gillian Dobbie\n- **URL**: [https://arxiv.org/abs/2606.27596](https://arxiv.org/abs/2606.27596)\n- **Abstract**:\n> Large Vision-Language Models (LVLMs) exhibit sophisticated reasoning but remain susceptible to object hallucination. Deviating from the prevailing attention intensity assumption, we reveal a deeper dynamic structural misalignment: hallucination is triggered at decision-critical steps where specific attention heads, acting as risky mediators, decouple from visual evidence to lock onto language priors. This establishes a pathological shortcut that bypasses visual grounding. To dismantle this, we propose Fox (Faithfulness and Observational-flow via eXpression-rectification), a training-free inference-time framework. Fox diagnoses structural misalignment using a visual attention entropy probe to localize risky mediators unsupervisedly. We then execute a targeted causal intervention via numerical logit saturation to physically sever the shortcut path. Finally, a conflict-gated cooperative decoding strategy reconciles interventional faithfulness with observational fluency. Extensive experiments demonstrate that Fox achieves SOTA performance, outperforming SID by 29.1% while preserving linguistic richness. Code is available at this https URL .\n\n---\n\n## 41. On the Inseparability of Instructions and Data in Shared-Embedding Sequence Models\n- **Authors**: Dewank Pant , Shruti Lohani , Avijit Kumar\n- **URL**: [https://arxiv.org/abs/2606.27567](https://arxiv.org/abs/2606.27567)\n- **Abstract**:\n> Prompt injection is the top security risk for LLM-integrated applications, yet every defense proposed so far has been broken. We prove this is not a coincidence: in shared-embedding architectures that lack enforced control-data separation, perfect prompt-injection prevention is mathematically impossible. We formalize prompted systems as Prompted Action Models whose outputs include control-authoritative actions: refusal decisions, tool authorization, policy routing, and memory writes. We define Semantic-Faithful Control (SFC), the property that such behavior depends only on the meaning of untrusted input, not on how it is encoded. We then prove SFC is unachievable within the shared pipeline, via three results: a provenance-recovery impossibility (shared representations make trusted and untrusted content statistically inseparable, bounded by total variation distance); control-path exposure (untrusted tokens enter control-relevant computation through the same attention value-aggregation that determines outputs); and a finite-coverage invariance gap (finite training cannot certify invariance over infinite semantic-equivalence classes). We ground each quantity in measurements on production tokenizers and models. The result is structural, not a gap in current defenses. It mirrors the code-data confusion in Von Neumann machines that gives rise to buffer overflows, a vulnerability class that took decades of layered defenses (DEP, Write-XOR-Execute, ASLR, stack canaries, and ultimately memory-safe languages) to contain, because no single mechanism sufficed. The implication is the same: prompt injection cannot be eliminated by better in-pipeline classification or alignment alone. It requires architectural separation of instruction and data channels. We identify the root cause and the class of solution it demands.\n\n---\n\n## 42. Benchmarking Multi-Modal Graph-based Social Media Popularity Prediction\n- **Authors**: Utkarsh Sahu , Zhisheng Qi , Li Zhu , Yizhao Yang , Jun Li , Ryan Rossi , Yu Wang\n- **URL**: [https://arxiv.org/abs/2606.27539](https://arxiv.org/abs/2606.27539)\n- **Abstract**:\n> Social media popularity prediction aims to forecast the future reach or influence of online content from early-stage observations. Accurate prediction enables key downstream applications, such as advertising optimization and strategic content planning by users, creators, and platforms. Despite substantial progress, existing popularity prediction works often fail to jointly consider multimodal content and temporal social interaction signals. Moreover, the literature remains highly fragmented across datasets, modalities, observation windows, prediction targets, and evaluation protocols. This fragmentation prevents fair comparison and obscures a systematic understanding of how textual, visual, temporal, and interaction-based signals jointly shape popularity dynamics. To address these challenges, we introduce MMG-Pop, a Multi-modal Graph-based Popularity Prediction benchmark, which unifies datasets, modalities, temporal interaction signals, and representative baselines under a standardized evaluation protocol. Furthermore, we propose MMG-PopNet, a unified multi-modal graph-based network that jointly models the aforementioned multi-modal signals and graph-structured social interactions. Extensive experiments on MMG-Pop, comprising four datasets across Bluesky and Reddit platforms, demonstrate the superior performance of MMG-PopNet and yield new insights into cross-platform training generalization, multi-task prediction benefits, multi-modality contributions, and LLM prediction limitation. These findings establish a unified foundation for future research on social dynamics modeling and intervention under heterogeneous modalities and socially-aware agentic ecosystem paradigms.\n\n---\n\n## 43. Large Language Model Teaches Visual Students: Cross-Modality Transfer of Fine-Grained Conceptual Knowledge\n- **Authors**: Thomas Shih-Chao Liang , Zhuoran Yu , Yong Jae Lee\n- **URL**: [https://arxiv.org/abs/2606.27527](https://arxiv.org/abs/2606.27527)\n- **Abstract**:\n> Large Language Models (LLMs) possess broad conceptual knowledge acquired through large-scale text pretraining, yet their potential to supervise models in other modalities remains underexplored. In this work, we propose LaViD--Language-to-Visual Knowledge Distillation--a simple and effective framework for transferring high-level semantic knowledge from a language-only teacher to a vision-only student model. Instead of relying on paired multimodal data, LaViD elicits conceptual signals from an LLM by prompting it to generate multiple-choice questions (MCQs) that probe semantic distinctions between visual classes. Each class is mapped to a soft label distribution over these MCQs, forming a rich conceptual signature that guides the student through an auxiliary distillation loss. Notably, despite using a language-only teacher without access to image data, LaViD consistently outperforms recent methods like MaKD that distill from vision-language models across multiple fine-grained benchmarks. It also achieves competitive or superior performance compared to state-of-the-art visual distillation methods such as DKD and MLKD, with further gains when combined with logit standardization. On the Waterbirds dataset, LaViD substantially improves worst-group accuracy, demonstrating enhanced robustness to spurious correlations with distillation. Code is available at this https URL .\n\n---\n\n## 44. Speculative Refinement: A Hybrid Autoregressive Diffusion Decoding Strategy and Its Behavior Across Benchmarks\n- **Authors**: Aditi Gupta , Neel Mishra , Kushagra Trivedi , Pawan Kumar\n- **URL**: [https://arxiv.org/abs/2606.27474](https://arxiv.org/abs/2606.27474)\n- **Abstract**:\n> How should we evaluate generation systems that combine autoregressive (AR) and diffusion decoding? We study this question through Speculative Refinement (SpecRef), a training-free hybrid method that warm-starts a masked diffusion language model from an AR draft using entropy-guided selective masking. Evaluating SpecRef across six benchmarks (HumanEval, MBPP, GSM8K, BBH, ARC-Challenge, HellaSwag) with three distinct evaluation protocols (execution-based pass@1, exact-match, log-likelihood scoring), we surface several findings relevant beyond our specific system: (1) code benchmarks conflate structural discovery with logical correctness: providing a syntactic scaffold lifts accuracy from near zero to over 20% without changing the model, indicating that much of the baseline failure is structural; (2) a refinement tension phenomenon where multi-stage correction degrades already-correct tokens, exposing benchmark saturation ceilings invisible to single-model evaluation; (3) log-likelihood and generative evaluation produce different model rankings for the same model pair, suggesting they measure different capabilities; (4) standard Python post-processing silently breaks code evaluation for non-AR generators. These observations apply to any multi-stage or non-autoregressive generation pipeline and point toward more diagnostic evaluation practices.\n\n---\n\n## 45. Supersede: Diagnosing and Training the Memory-Update Gap in LLM Agents\n- **Authors**: Vedant Patel\n- **URL**: [https://arxiv.org/abs/2606.27472](https://arxiv.org/abs/2606.27472)\n- **Abstract**:\n> Large language model (LLM) agents operate over long, multi-session interactions in which facts change: a user moves, a price updates, a plan is revised. Acting correctly requires using the current value of a fact and discarding values that have been superseded. We isolate this ability on real conversational data and show that it is a distinct, unsolved failure. On the knowledge-update subset of LongMemEval, replacing an agent's full context with a bounded, self-maintained memory drops accuracy from 92% to 77% even on a frontier model (gpt-5.4), a gap that is statistically significant (paired McNemar p<0.005) and persists across model scale while full-context accuracy saturates near 92%. The bottleneck is therefore memory maintenance, not comprehension, and is not closed by a stronger model. We then ask whether this is merely an undersized memory, and find it is not: as the conversation grows 24x, accuracy falls further (from 68% to 28%), and granting the agent proportionally more memory yields no detectable recovery (28% to 28%, n=25). The failure scales with the length of the conversation, not the compression ratio. We release Supersede, an open reinforcement-learning environment (on the verifiers / prime-rl stack) that turns this measurement into a training signal: agents are rewarded for answering from the current value and penalized for stale ones. Finally, we close the loop and show the gap is trainable: GRPO fine-tuning a small open model (Qwen2.5-3B) on this environment nearly doubles its held-out supersession accuracy on real, unseen conversations (9.0% to 16.7%, a single run), along a monotonic checkpoint curve indicating the learned policy, not the harness, carries the gain. To our knowledge this is the first trainable environment whose reward targets temporal fact-currency, and the first evidence the supersession gap can be trained down, not only measured.\n\n---\n\n## 46. SidConArena: An Environment Evaluating Agents in Open-Ended,Positive-Sum Bargaining Game\n- **Authors**: Yeqi Feng , Yuxin Chen , Tianxing He\n- **URL**: [https://arxiv.org/abs/2606.27397](https://arxiv.org/abs/2606.27397)\n- **Abstract**:\n> Evaluating LLM agents requires dynamic environments that go beyond static reasoning and zero-sum games. Real-world economic interaction is often open-ended and mixed-motive: agents must negotiate, create positive-sum surplus, compete for scarce assets, and plan under delayed returns. We introduce SidConArena, a new benchmark framework for evaluating LLM agents in open-ended, positive-sum bargaining. SidConArena formalizes a multi-player economy as a finite-horizon partially observable stochastic game with three coupled phases: natural-language negotiation with binding trades, deterministic converter-based production, and sealed-bid auctions for long-term assets. The framework combines structured observations, phase-aware agent dispatching, a neural-symbolic action interface, and asynchronous execution, enabling free-form interaction while preserving rule-grounded evaluation. Across homogeneous and heterogeneous tournaments, stronger frontier models achieve higher economic outcomes, yet agents still misvalue resources, bargain passively, and remain limited in long-horizon investment planning.\n\n---\n\n## 47. Agentic Publication Protocol: An Attempt to Modernize Scientific Publication\n- **Authors**: Sirui Lu , Xiao-Liang Qi\n- **URL**: [https://arxiv.org/abs/2606.27386](https://arxiv.org/abs/2606.27386)\n- **Abstract**:\n> Scientific publication is still organized primarily around static manuscripts, even though much of scientific progress depends on tacit know-how: how to run code, reproduce figures, interpret edge cases, choose useful follow-up directions, and avoid failed paths. Large language model agents create an opportunity to publish not only knowledge, but also operational know-how in a form that future readers and researchers can directly use. This paper outlines the Agentic Publication Protocol (APP), a lightweight repository format for packaging a paper together with code, data, environment information, reproducibility instructions, and an agent-facing instruction file. APP treats a version-controlled repository as the publication object and uses \\texttt{ this http URL } and optional skills to define a paper agent that can explain the work, reproduce key results when possible, and support follow-up research. We describe the design principles and details of the protocol, as well as the agent skills useful for publishing papers under the protocol. We also describe development tools for evaluating and improving the protocol and associated agent skills. Finally, we provide a broader discussion of the future of scientific research in the agent era.\n\n---\n\n## 48. CalBrief: A Pilot Diagnostic Benchmark for Evidence-Calibrated Scientific Briefing with Large Language Models\n- **Authors**: Yu Fu , Yongqi Kang , Yong Zhao\n- **URL**: [https://arxiv.org/abs/2606.27383](https://arxiv.org/abs/2606.27383)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used as research assistants, yet it remains unclear whether they can calibrate research takeaways to the strength and scope of the supporting evidence. We study evidence-calibrated scientific briefing: given a bounded package of related papers, a system should generate package-level takeaways with evidence strength, scope boundaries, and missing-evidence caveats. We contribute a verified pilot benchmark of 16 heterogeneous scientific evidence packages and 96 human-verified takeaways, and we use CalBrief, an auditable role/gap/strength framework, as a diagnostic probe to locate where briefing breaks down. Under a fair-schema evaluation, structured organization improves role and gap reasoning, but an explicit strength-calibration policy is systematically over-conservative and falls below majority and direct-LLM baselines. To explain why, we run a controlled diagnostic across three closed-model backbones (GPT-4o, Claude Sonnet, Gemini Flash) that separates three potential causes of conservatism. Approximately 63% of the conservatism gap is attributable to expanding the label space from binary {moderate, weak} to four-way {moderate, weak, uncertain, insufficient_evidence} (p < 0.001 across all backbones); only 1% is attributable to gap/scope signal injection (not significant); the remaining 36% arises from the pipeline policy itself. We also find that 4-way predictions can be post-hoc collapsed back to binary and then match or exceed direct binary prompting, so the extra labels carry information that strict matching hides. Label-level strength judgment and auditable evidence organization are distinct abilities currently in tension, and should be evaluated separately for LLM research assistants.\n\n---\n\n## 49. Position: The Term \"Machine Unlearning\" Is Overused in LLMs\n- **Authors**: Sangyeon Yoon , Yeachan Jun , Albert No\n- **URL**: [https://arxiv.org/abs/2606.27379](https://arxiv.org/abs/2606.27379)\n- **Abstract**:\n> Large language models increasingly face demands to \"forget\" training data, knowledge, or behaviors due to regulatory deletion obligations, copyright/licensing disputes, and safety or product-policy requirements. This position paper argues that machine unlearning is overused as a term in LLM research and should be reserved for dataset-defined deletion: removing the training influence of a precisely specified forget set such that the resulting model is approximately indistinguishable from retraining without that data. We contend that many tasks currently labeled \"unlearning\" (e.g., refusal for harmful requests, entity/knowledge removal, or targeted suppression) pursue different, often policy-dependent objectives and therefore require different terminology and baselines (e.g., alignment, suppression, editing, obfuscation). We further argue that this confusion is not cosmetic: because papers make different implicit guarantees under the same label, metrics and benchmarks are frequently reused outside their intended scope, rewarding surface-level non-disclosure (e.g., low ROUGE/forget accuracy) even when retraining-equivalence is not tested and derived capabilities remain. We conclude by calling for stricter terminology tied to explicit guarantees and reference models, and for evaluations that match the claimed objective.\n\n---\n\n## 50. DataStates-LLM: Scalable Checkpointing for Transformer Models Using Composable State Providers\n- **Authors**: Avinash Maurya , M. Mustafa Rafique , Franck Cappello , Bogdan Nicolae\n- **URL**: [https://arxiv.org/abs/2601.16956](https://arxiv.org/abs/2601.16956)\n- **Abstract**:\n> The rapid growth of Large Transformer-based models, specifically Large Language Models (LLMs), now scaling to trillions of parameters, has necessitated training across thousands of GPUs using complex hybrid parallelism strategies (e.g., data, tensor, and pipeline parallelism). Checkpointing this massive, distributed state is critical for a wide range of use cases, such as resilience, suspend-resume, investigating undesirable training trajectories, and explaining model evolution. However, existing checkpointing solutions typically treat model state as opaque binary blobs, ignoring the ``3D heterogeneity'' of the underlying data structures--varying by memory location (GPU vs. Host), number of ``logical'' objects sharded and split across multiple files, data types (tensors vs. Python objects), and their serialization requirements. This results in significant runtime overheads due to blocking device-to-host transfers, data-oblivious serialization, and storage I/O contention. In this paper, we introduce DataStates-LLM, a novel checkpointing architecture that leverages State Providers to decouple state abstraction from data movement. DataStates-LLM exploits the immutability of model parameters during the forward and backward passes to perform ``lazy'', non-blocking asynchronous snapshots. By introducing State Providers, we efficiently coalesce fragmented, heterogeneous shards and overlap the serialization of metadata with bulk tensor I/O. We evaluate DataStates-LLM on models up to 70B parameters on 256 A100-40GB GPUs. Our results demonstrate that DataStates-LLM achieves up to 4$\\times$ higher checkpointing throughput and reduces end-to-end training time by up to 2.2$\\times$ compared to state-of-the-art solutions, effectively mitigating the serialization and heterogeneity bottlenecks in extreme-scale LLM training.\n\n---\n\n{% endraw %}\n"}}, "ALL/2026-05-18.md": {"64bbd1731dde152451afe1705df222e26b5c83fa": {"t": 1785747199.8455935, "n": 340278, "h": "05c8c40e431f029e", "c": "---\nlayout: default\ntitle: \"전체 AI 논문 - 2026-05-18\"\n---\n\n{% raw %}\n# 전체 AI 논문 - 2026-05-18\n\n## 1. Prospective multi-pathogen disease forecasting using autonomous LLM-guided tree search\n- **Authors**: Sarah Martinson , Michael P. Brenner , Martyna Plomecka , Brian P. Williams , Nicholas G. Reich , Zahra Shamsi\n- **URL**: [https://arxiv.org/abs/2605.16238](https://arxiv.org/abs/2605.16238)\n- **Abstract**:\n> Probabilistic forecasting of infectious diseases is crucial for public health but relies on labor-intensive manual model curation by expert modeling teams. This bespoke development bottlenecks scalability to granular geographic resolutions or emerging pathogens. Here, we present an autonomous system using Large Language Model (LLM)-guided tree search to iteratively generate, evaluate, and optimize executable forecasting software. In a fully prospective, real-time evaluation during the 2025-2026 US respiratory season, the system autonomously discovered methodologically diverse models for influenza, COVID-19, and respiratory syncytial virus (RSV). Aggregating these machine-generated models yielded an ensemble that consistently matched or outperformed the gold-standard, human-curated Centers for Disease Control and Prevention (CDC) hub ensembles out-of-sample. The system successfully navigated data-scarce \"cold start\" scenarios for RSV. Moreover, controlled retrospective ablations revealed that optimizing log-scale distance metrics prevents reward hacking, while an automated judge-in-the-loop ensures structural fidelity to complex scientific theories. By autonomously translating epidemiological theory into accurate, transparent code, this framework overcomes the modeling labor bottleneck, enabling rapid deployment of expert-level disease forecasting at unprecedented scales.\n\n---\n\n## 2. FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast\n- **Authors**: Igor Bogdanov , Chung-Horng Lung , Thomas Kunz , Jie Gao , Adrian Taylor , Marzia Zaman\n- **URL**: [https://arxiv.org/abs/2605.16233](https://arxiv.org/abs/2605.16233)\n- **Abstract**:\n> Can LLM agents improve decision-making through self-generated memory without gradient updates? We propose FORGE (Failure-Optimized Reflective Graduation and Evolution), a staged, population-based protocol that evolves prompt-injected natural-language memory for hierarchical ReAct agents. FORGE wraps a Reflexion-style inner loop, where a dedicated reflection agent (using the same underlying LLM, no distillation from a stronger model) converts failed trajectories into reusable knowledge artifacts: textual heuristics (Rules), few-shot demonstrations (Examples), or both (Mixed), with an outer loop that propagates the best-performing instance's memory to the population between stages and freezes converged instances via a graduation criterion. We evaluate on CybORG CAGE-2, a stochastic network-defense POMDP at a 30-step horizon against the B-line attacker, where all four tested LLM families (Gemini-2.5-Flash-Lite, Grok-4-Fast, Llama-4-Maverick, Qwen3-235B) exhibit strongly negative, heavy-tailed zero-shot rewards. Compared against both a zero-shot baseline and a Reflexion baseline (isolated single-stream learning), FORGE improves average evaluation return by 1.7-7.7$\\times$ over zero-shot and by 29-72% over Reflexion in all 12 model-representation conditions, reducing major-failure rates (below $-100$) to as low as $\\sim$1%. We find that (1) population broadcast is critical mechanism, with a no-graduation ablation confirming that broadcast carries the performance gains while graduation primarily saves compute; (2) Examples achieves the strongest returns for three of four models, Rules offers the best cost-reliability profile with $\\sim$40% fewer tokens; and (3) weaker baseline models benefit disproportionately, suggesting FORGE may mitigate capability gaps rather than amplify strong models. All evidence is confined to CAGE-2 B-line; cross-family findings are directional evidence.\n\n---\n\n## 3. Fully Open Meditron: An Auditable Pipeline for Clinical LLMs\n- **Authors**: Xavier Theimer-Lienhard , Mushtaha El-Amin , Fay Elhassan , Sahaj Vaidya , Victor Cartier-Negadi , David Sasu , Lars Klein , Mary-Anne Hartley\n- **URL**: [https://arxiv.org/abs/2605.16215](https://arxiv.org/abs/2605.16215)\n- **Abstract**:\n> Clinical decision support systems (CDSS) require scrutable, auditable pipelines that enable rigorous, reproducible validation. Yet current LLM-based CDSS remain largely opaque. Most \"open\" models are open-weight only, releasing parameters while withholding the data provenance, curation procedures, and generation pipelines that determine model behavior. Fully Open (FO) models, which expose the complete training stack end-to-end, do not currently exist in medicine. We introduce Fully Open Meditron, the first fully open pipeline for building LLM-CDSS, comprising a clinician-audited training corpus, a reproducible data construction and training framework, and a use-aligned evaluation protocol. The corpus unifies eight public medical QA datasets into a normalized conversational format and expands coverage with three clinician-vetted synthetic extensions: exam-style QA, guideline-grounded QA derived from 46,469 clinical practice guidelines, and clinical vignettes. The pipeline enforces system-wide decontamination, gold-label resampling of teacher generations, and end-to-end validation by a four-physician panel. We evaluate using an LLM-as-a-judge protocol over expert-written clinical vignettes, calibrated against 204 human raters. We apply the recipe to five FO base models (Apertus-70B/8B-Instruct, OLMo-2-32B-SFT, EuroLLM-22B/9B-Instruct). All MeditronFO variants are preferred over their bases. Apertus-70B-MeditronFO improves +6.6 points over its base (47.2% to 53.8%) on aggregate medical benchmarks, establishing a new FO SoTA. Gemma-3-27B-MeditronFO is preferred over MedGemma in 58.6% of LLM-as-a-judge comparisons and outperforms it on HealthBench (58% vs 55.9%). These results show that fully open pipelines can achieve state-of-the-art domain-specific performance without sacrificing auditability or reproducibility.\n\n---\n\n## 4. Confirming Correct, Missing the Rest: LLM Tutoring Agents Struggle Where Feedback Matters Most\n- **Authors**: Tahreem Yasir , Wenbo Li , Sam Gilson , Sutapa Dey Tithi , Xiaoyi Tian , Tiffany Barnes\n- **URL**: [https://arxiv.org/abs/2605.16207](https://arxiv.org/abs/2605.16207)\n- **Abstract**:\n> Effective tutoring requires distinguishing optimal, valid but suboptimal, and incorrect student solutions, a distinction central to intelligent tutoring systems (ITS) but untested for LLM-based tutors. As LLMs are increasingly explored as conversational complements to ITS, evaluating their diagnostic precision is essential. We present a benchmark of seven LLM feedback agents in propositional logic using knowledge-graph-derived ground truth across 10,836 solution--feedback pairs and three feedback conditions. Models achieved near-ceiling performance on optimal steps but systematically over-rejected valid but suboptimal reasoning and over-validated incorrect solutions, precisely where adaptive tutoring matters most. These failures persisted across models regardless of solution context, suggesting architectural rather than informational limits. Moreover, accurate diagnosis did not reliably produce pedagogically actionable feedback, revealing a gap between diagnostic judgment and instructional effectiveness. Our findings suggest that LLMs are better suited for hybrid architectures where KG-grounded models handle diagnosis while LLMs support open-ended scaffolding and dialogue.\n\n---\n\n## 5. Context, Reasoning, and Hierarchy: A Cost-Performance Study of Compound LLM Agent Design in an Adversarial POMDP\n- **Authors**: Igor Bogdanov , Chung-Horng Lung , Thomas Kunz , Jie Gao , Adrian Taylor , Marzia Zaman\n- **URL**: [https://arxiv.org/abs/2605.16205](https://arxiv.org/abs/2605.16205)\n- **Abstract**:\n> Deploying compound LLM agents in adversarial, partially observable sequential environments requires navigating several design dimensions: (1) what the agent sees, (2) how it reasons, and (3) how tasks are decomposed across components. Yet practitioners lack guidance on which design choices improve performance versus merely increase inference costs. We present a controlled study of compound LLM agent design in CybORG CAGE-2, a cyber defense environment modeled as a Partially Observable Markov Decision Process (POMDP). Reward is non-positive, so all configurations operate in a failure-mitigation mode. Our evaluation spans five model families, six models, and twelve configurations (3,475 episodes) with token-level cost accounting. We vary context representation (raw observations vs. a deterministic state-tracking layer with compressed history), deliberation (self-questioning, self-critique, and self-improvement tools, with optional chain-of-thought prompting), and hierarchical decomposition (monolithic ReAct vs. delegation to specialized sub-agents). We find that: (1) Programmatic state abstraction delivers the largest returns per token spent (RPTS), improving mean return by up to 76% over raw observations. (2) Distributing deliberation tools across a hierarchy degrades performance relative to hierarchy alone for all five model families, reaching up to 3.4$\\times$ worse mean return while using 1.8-2.7$\\times$ more tokens. We call this destructive pattern a deliberation cascade. (3) Hierarchical decomposition without deliberation achieves the best absolute performance for most models, and context engineering is generally more cost-effective than deliberation. These findings suggest a design principle for structured adversarial POMDPs: invest in programmatic infrastructure and clean task decomposition rather than deeper per-agent reasoning, as these strategies can interfere when combined.\n\n---\n\n## 6. Formal Methods Meet LLMs: Auditing, Monitoring, and Intervention for Compliance of Advanced AI Systems\n- **Authors**: Parand A. Alamdari , Toryn Q. Klassen , Sheila A. McIlraith\n- **URL**: [https://arxiv.org/abs/2605.16198](https://arxiv.org/abs/2605.16198)\n- **Abstract**:\n> We examine one particular dimension of AI governance: how to monitor and audit AI-enabled products and services throughout the AI development lifecycle, from pre-deployment testing to post-deployment auditing. Combining principles from formal methods with SoTA machine learning, we propose techniques that enable AI-enabled product and service developers, as well as third party AI developers and evaluators, to perform offline auditing and online (runtime) monitoring of product-specific (temporally extended) behavioral constraints such as safety constraints, norms, rules and regulations with respect to black-box advanced AI systems, notably LLMs. We further provide practical techniques for predictive monitoring, such as sampling-based methods, and we introduce intervening monitors that act at runtime to preempt and potentially mitigate predicted violations. Experimental results show that by exploiting the formal syntax and semantics of Linear Temporal Logic (LTL), our proposed auditing and monitoring techniques are superior to LLM baseline methods in detecting violations of temporally extended behavioral constraints; with our approach, even small-model labelers match or exceed frontier LLM judges. Our predictive and intervening monitors significantly reduce the violation rates of LLM-based agents while largely preserving task performance. We further show through controlled experiments that LLMs' temporal reasoning shows a pronounced degradation in accuracy with increasing event distance, number of constraints, and number of propositions.\n\n---\n\n## 7. An Algebraic Exposition of the Theory of Dyadic Morality\n- **Authors**: Kush R. Varshney\n- **URL**: [https://arxiv.org/abs/2605.16153](https://arxiv.org/abs/2605.16153)\n- **Abstract**:\n> This paper provides an algebraic exposition of the theory of dyadic morality (TDM), a psychological model of moral judgment grounded in a simple two-node template: an intentional agent causing harm to a vulnerable patient. We formalize TDM using structural causal modeling (SCM) notation and identify three psychological operators (typecasting operator, completion operator, and valence-dependent inference mechanism) that extend standard SCM to capture how people compute moral judgments under constraints. We address scalability challenges arising from TDM's dyadic limitation, showing how moral cognition compresses multi-node scenarios through node collapse and sequential processing. Drawing on this algebraic framework, we demonstrate concrete applications to AI policy design: detecting conflicting obligations, structuring helpfulness policies to preserve user agency, and designing post-failure communication as causal interventions. Finally, we recommend scoped, contextual measurement of mind perception over universal averaging to operationalize the theory empirically. This algebraic formalization enables neurosymbolic AI systems to compute morality in a way that is both mathematically rigorous and faithful to human moral cognition.\n\n---\n\n## 8. Look Before You Leap: Autonomous Exploration for LLM Agents\n- **Authors**: Ziang Ye , Wentao Shi , Yuxin Liu , Yu Wang , Zhengzhou Cai , Yaorui Shi , Qi Gu , Xunliang Cai , Fuli Feng\n- **URL**: [https://arxiv.org/abs/2605.16143](https://arxiv.org/abs/2605.16143)\n- **Abstract**:\n> Large language model based agents often fail in unfamiliar environments due to premature exploitation: a tendency to act on prior knowledge before acquiring sufficient environment-specific information. We identify autonomous exploration as a critical yet underexplored capability for building adaptive agents. To formalize and quantify this capability, we introduce Exploration Checkpoint Coverage, a verifiable metric that measures how broadly an agent discovers key states, objects, and affordances. Our systematic evaluation reveals that agents trained with standard task-oriented reinforcement learning consistently exhibit narrow and repetitive behaviors that impede downstream performance. To address this limitation, we develop a training strategy that interleaves task-execution rollouts and exploration rollouts, with each type of rollout optimized by its corresponding verifiable reward. Building on this training strategy, we propose the Explore-then-Act paradigm, which decouples information-gathering from task execution: agents first utilize an interaction budget to acquire grounded environmental knowledge, then leverage it for task resolution. Our results demonstrate that learning to systematically explore is imperative for building generalizable and real-world-ready agents.\n\n---\n\n## 9. Property-Guided LLM Program Synthesis for Planning\n- **Authors**: Augusto B. Corrêa , André G. Pereira , Jendrik Seipp\n- **URL**: [https://arxiv.org/abs/2605.16142](https://arxiv.org/abs/2605.16142)\n- **Abstract**:\n> LLMs have shown impressive success in program synthesis, discovering programs that surpass prior solutions. However, these approaches rely on simple numeric scores to signal program quality, such as the value of the solution or the number of passed tests. Because a score offers no guidance on why a program failed, the system must generate and evaluate many candidates hoping some succeed, increasing LLM inference and evaluation costs. We study a different approach: property-guided LLM program synthesis. Instead of scoring programs after evaluation, we check whether a candidate satisfies a formally defined property. When the property is violated, we stop the evaluation early and provide the LLM with a concrete counterexample showing exactly how the program failed. This feedback drastically reduces both the number of program generations and the evaluation cost, and can guide the LLM to generate stronger programs. We evaluate this approach on PDDL planning domains, asking the LLM to synthesize direct heuristic functions: every state reachable by strictly improving transitions has a strictly improving successor. A heuristic with this property leads hill-climbing algorithm directly to a goal state. A counterexample-guided repair loop generates one candidate program, checks the property over a training set, and returns the first case that violates the property. We evaluate our approach on ten planning domains with an out-of-distribution test set. The synthesized heuristics are effectively direct on virtually all test tasks, and compared to the best prior generation method our approach generates seven times fewer programs per domain on average, solves more tasks without using search, and requires several orders of magnitude less computation to evaluate candidates. Whenever a problem admits a verifiable property, property-guided LLM synthesis can reduce cost and improve program quality.\n\n---\n\n## 10. ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents\n- **Authors**: Chinmay Savadikar , Mingyu Zhao , Yuanzheng Zhu , Han Li , Shuang Xie , Alberto Castelo , Tianfu Wu , Lingyun Wang\n- **URL**: [https://arxiv.org/abs/2605.16116](https://arxiv.org/abs/2605.16116)\n- **Abstract**:\n> Developing and evaluating e-commerce web agents requires environments that preserve meaningful task structure while enabling controllable, reproducible, and scalable scientific comparison. Existing methodologies force a tradeoff: live storefronts provide realism but are non-stationary, difficult to inspect, and irreproducible, while hand-built sandbox benchmarks provide control but cover only a narrow range of layouts, catalogs, policies, and interaction patterns. We argue that the core bottleneck is methodological: the field lacks a scalable way to construct evaluation settings that are simultaneously realistic, diverse, controllable, inspectable, and reproducible. We introduce ShopGym, an integrated framework for realistic simulation and scalable benchmarking of e-commerce web agents. ShopGym is a framework for constructing e-commerce simulation environments and grounded benchmark tasks. Its simulation layer, ShopArena, converts live seed storefronts into self-contained sandbox shops through anonymized shop specifications and a staged, validated generation process. On top of these simulated storefronts, ShopGuru synthesizes benchmark tasks across seven skill categories, grounding each task in the shop's catalog, navigation structure, policies, and interaction affordances. Together, ShopArena and ShopGuru produce self-contained, resettable, inspectable, and stable evaluation artifacts that preserve structural properties and agent-evaluation signals relevant to shopping tasks. We validate the framework through graph-based structural analysis and agent-based behavioral evaluation with 224 generated tasks across six sandbox shops: three constructed with synthetic data and three with real data. Our results show that the synthetic shops preserve key structural properties of live storefronts, with agent performance on synthetic shops positively correlated with performance on live storefronts.\n\n---\n\n## 11. Sign-Separated Finite-Time Error Analysis of Q-Learning\n- **Authors**: Donghwan Lee\n- **URL**: [https://arxiv.org/abs/2605.16103](https://arxiv.org/abs/2605.16103)\n- **Abstract**:\n> This paper develops a sign-separated finite-time error analysis for constant step-size Q-learning. Starting from the switching-system representation, the error is decomposed into its componentwise negative and positive parts. The negative part is dominated by a lower comparison linear time-invariant (LTI) system associated with a fixed optimal policy, whereas the positive part is controlled by a linear switching system. The resulting bounds show that the negative-side LTI certificate is no slower than the positive-side switching certificate and may produce a faster exponential envelope. The analysis identifies a max-induced asymmetry in Q-learning error dynamics. This asymmetry is connected to overestimation: positive action-wise errors can be selected and propagated by the Bellman maximum, whereas negative errors admit an optimal-policy lower comparison. Finite-time bounds are provided for both deterministic and stochastic constant-step-size recursions.\n\n---\n\n## 12. Reasoners or Translators? Contamination-aware Evaluation and Neuro-Symbolic Robustness in Tax Law\n- **Authors**: Parisa Kordjamshidi , Samer Aslan , Madhavan Seshadri , Leslie Barrett , Enrico Santus\n- **URL**: [https://arxiv.org/abs/2605.16052](https://arxiv.org/abs/2605.16052)\n- **Abstract**:\n> Recent advances in large language models (LLMs) have significantly enhanced automated legal reasoning. Yet, it remains unclear whether their performance reflects genuine legal reasoning ability or artifacts of data contamination. We present a comprehensive empirical study of tax law reasoning approaches and implement a contamination detection protocol to rigorously assess LLM reliability. We show that performance can be inflated by contamination. Building on this analysis, we conduct a systematic evaluation, comparing monolithic LLMs with hybrid systems that translate statutory text into formal representations and delegate inference to symbolic solvers. We build a novel test suite designed to probe generalization to unseen documents via case and rule variations. Our findings indicate that legal reasoning is inherently compositional and that neuro-symbolic frameworks offer a more reliable and robust foundation for legal AI, as well as improved generalization to unobserved situations.\n\n---\n\n## 13. ScreenSearch: Uncertainty-Aware OS Exploration\n- **Authors**: Michael Solodko , Justin Wagle\n- **URL**: [https://arxiv.org/abs/2605.16024](https://arxiv.org/abs/2605.16024)\n- **Abstract**:\n> Desktop GUI agents operate under partial observability: visually similar screens can correspond to different underlying workflow states, so locally plausible actions can lead to sharply different outcomes. We frame this as a problem of computer/OS state exploration, where effective behavior requires both expanding the reachable frontier and reducing ambiguity before committing. We present ScreenSearch, a system that combines structural screen retrieval and deduplication with an ambiguity-aware PUCT graph-bandit for large-scale desktop exploration. The retrieval layer converts UIA trees into location-aware structural features, indexes related screens through sparse token search and metadata filters, and maintains a shared deduplicated state graph across VM workers. On top of this graph, we define a scalable ambiguity signal based on matched-action outcome dispersion. If similar screens produce different next states under the same action signature, the state should be probed further rather than treated as resolved. We use this signal together with frontier rewards to drive large-scale exploration and replay-start policy evaluation over the shared graph. Across 11 desktop applications, ScreenSearch collects over 1M screenshots and over 30K deduplicated states, yielding large exploration corpora with substantial cross-application and within-application diversity. On a fixed replay-start slice, we observe a clear novelty--ambiguity trade-off: some policies reduce ambiguity quickly while discovering little frontier. Ambiguity reduction alone is therefore not a sufficient exploration objective. Appendix ablations show that stronger proposal priors can materially improve unique-state discovery during corpus building. These results suggest that state identity, proposal quality, and ambiguity-aware search all matter when deciding when to probe and when to commit.\n\n---\n\n## 14. Petri Net Induced Heuristic Search for Resource Constrained Scheduling\n- **Authors**: Ido Lublin , Dor Atzmon , Izack Cohen\n- **URL**: [https://arxiv.org/abs/2605.15983](https://arxiv.org/abs/2605.15983)\n- **Abstract**:\n> We formulate the Resource-Constrained Project Scheduling Problem (RCPSP) as optimal search over the reachability graph of a Timed Transition Petri Net with Resources, using relative-delay tokens so that scheduling decisions correspond to transition firings in the induced state space. We solve the resulting problem with $A^*$ guided by a heuristic that combines Critical Path and resource-based lower bounds, and prove that it is consistent under our token-based time semantics. Experiments on the PSPLIB benchmarks show that the approach outperforms strong exact Mixed-Integer Linear Programming (MIP) baselines (SCIP, CBC) in both success rate and solve time. Per-instance analysis shows that heuristic search and MIP degrade along independent axes, resource tightness for $A^*$ and formulation size for MIP, with resource strength mediating which solver benefits from scale.\n\n---\n\n## 15. Learning Bilevel Policies over Symbolic World Models for Long-Horizon Planning\n- **Authors**: Dillon Z. Chen , Till Hofmann , Toryn Q. Klassen , Sheila A. McIlraith\n- **URL**: [https://arxiv.org/abs/2605.15975](https://arxiv.org/abs/2605.15975)\n- **Abstract**:\n> We tackle the challenge of building embodied AI agents that can reliably solve long-horizon planning problems. Imitation learning from demonstrations has shown itself to be effective in training robots to solve a diversity of complex tasks requiring fine motor control and manipulation over low-level (LL), continuous environments. Yet, it remains a difficult endeavour to generate long-horizon plans from imitation learning alone. In contrast, high-level (HL), symbolic abstractions facilitate efficient and interpretable long-horizon planning. We propose to combine the strengths of LL imitation learning for manipulation and control, and HL symbolic abstractions for long-horizon planning. We realise this idea via \\emph{bilevel policies} of the form $(\\pi^{\\mathrm{hl} }, \\pi^{\\mathrm{ll} })$, consisting of a neural policy $\\pi^{\\mathrm{ll} }$ learned from LL demonstrations, and an HL symbolic policy $\\pi^{\\mathrm{hl} }$ that is constructed from symbolic abstractions of the LL demonstrations combined with inductive generalisation. We implement these ideas in the BISON system. Experiments on extended MetaWorld benchmarks demonstrate that BISON generalises to long horizons and problems with greater numbers of objects than those solved by VLA and end-to-end methods, and is more time and memory efficient in training and inference. Notably, when ignoring LL execution, BISON's HL policies can solve HL problems with 10,000 relevant objects in under a minute. Project page: this https URL\n\n---\n\n## 16. Deterministic Event-Graph Substrates as World Models for Counterfactual Reasoning\n- **Authors**: Fabio Rovai\n- **URL**: [https://arxiv.org/abs/2605.15967](https://arxiv.org/abs/2605.15967)\n- **Abstract**:\n> We study event-graph substrates: a class of world models that represent agent state as an append-only log of typed RDF triples and answer counterfactual queries by forking the log under a structured intervention vocabulary. Substrates are inspectable at the triple level, support exact counterfactuals, and transfer across domains without learned components. We formalize the class, prove a duality between explanatory and counterfactual queries that reduces both to the same causal-ancestor traversal, and evaluate a 1,400-line CLEVRER-DSL interpreter atop a domain-agnostic substrate runtime at full CLEVRER validation scale (n=75,618). The substrate exceeds the NS-DR symbolic oracle on all four per-question categories (by 9.89, 20.26, 17.65, and 0.80 percentage points), and exceeds the parametric ALOE baseline on descriptive and explanatory while lagging on predictive and counterfactual. We also introduce twin-EventLog, a 500-specification Park-canonical Smallville counterfactual benchmark on which the substrate exceeds Llama-3.1-8B with full context by 18.80 points joint accuracy.\n\n---\n\n## 17. PAGER: Bridging the Semantic-Execution Gap in Point-Precise Geometric GUI Control\n- **Authors**: Jingxuan Wei , Xi Bai , Shan Liu , Caijun Jia , Zheng Sun , Xinglong Xu , Siyuan Li , Linzhuang Sun , Bihui Yu , Conghui He , Cheng Tan\n- **URL**: [https://arxiv.org/abs/2605.15963](https://arxiv.org/abs/2605.15963)\n- **Abstract**:\n> Large vision-language models have significantly advanced GUI agents, enabling executable interaction across web, mobile, and desktop interfaces. Yet these gains largely rely on a forgiving region-tolerant paradigm, where many nearby pixels inside the same component remain valid. Precise geometric construction breaks this assumption: actions must land on points in continuous canvas space rather than tolerant regions. Because geometric primitives carry ontological dependencies, a local coordinate error can induce cascading topological failures that distort downstream objects and invalidate the final construction. We identify this regime as precision-sensitive GUI tasks, requiring point-level accuracy, geometry-aware verification, and robustness to dependency-driven error propagation. To benchmark it, we introduce PAGE Bench, with 4,906 problems and over 224K process-supervised, pixel-level GUI actions. We further propose PAGER, a topology-aware agent that decomposes construction into dependency-structured planning and pixel-level execution. Pixel-grounded supervised tuning establishes executable action grammar, while precision-aligned reinforcement learning mitigates rollout-induced exposure bias through state-conditioned geometric feedback. Experiments reveal a pronounced Semantic-Execution Gap: general multimodal models can exceed 88% action type accuracy yet remain below 6% task success. PAGER closes this gap, delivering 4.1x higher task success than the strongest evaluated general baseline and raising step success rate from below 9% for GUI-specialized agents to over 62%, establishing a new state of the art for point-precise GUI control.\n\n---\n\n## 18. Imperfect World Models are Exploitable\n- **Authors**: Logan Mondal Bhamidipaty (University of Edinburgh), Esmeralda S. Whitammer (University of Edinburgh), David Abel (University of Edinburgh), Mykel J. Kochenderfer (Stanford University), Subramanian Ramamoorthy (University of Edinburgh)\n- **URL**: [https://arxiv.org/abs/2605.15960](https://arxiv.org/abs/2605.15960)\n- **Abstract**:\n> We propose a novel definition of model exploitation in reinforcement learning. Informally, a world model is exploitable if it implies that one policy should be strictly preferred over another while the environment's true transition model implies the reverse. We analogize our definition with a prior characterization of reward hacking but show that the associated proof of inevitability does not transfer to exploitation. To overcome this obstruction, we develop a general theory of reward hacking and model exploitation that proves that exploitation is essentially unavoidable on large policy sets and yields the corresponding claim for hacking as a special case. Unfortunately, we also find that the conditions that guarantee unhackability in finite policy sets have no counterpart that precludes exploitation. Consequently, we introduce a relaxed notion of exploitation and derive a safe horizon within which it can be avoided. Taken together, our results establish a formal bridge between reward hacking and model exploitation and elucidate the limits of safe planning in world models.\n\n---\n\n## 19. Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design\n- **Authors**: Alberto Pepe , Chien-Yu Lin , Despoina Magka , Bilge Acun , Yannan Nellie Wu , Anton Protopopov , Carole-Jean Wu , Yoram Bachrach\n- **URL**: [https://arxiv.org/abs/2605.15871](https://arxiv.org/abs/2605.15871)\n- **Abstract**:\n> Toward recursive self-improvement, we investigate LLM agents autonomously designing foundation models beyond standard Transformers. We introduce a dual-framework approach: AIRA-Compose for high-level architecture search, and AIRA-Design for low-level mechanistic implementation. AIRA-Compose uses 11 agents to explore fundamental computational primitives under a 24-hour budget. Agents evaluate million-parameter candidates, extrapolating top designs to 350M, 1B, and 3B scales. This yields 14 architectures across two families: AIRAformers (Transformer-based) and AIRAhybrids (Transformer-Mamba). Pre-trained at 1B scale, these consistently outperform Llama 3.2 and Composer-found baselines. On downstream tasks, AIRAformer-D and AIRAhybrid-D improve accuracy by 2.4% and 3.8% over Llama 3.2. Furthermore, AIRA-Compose finds models with highly efficient scaling frontiers: AIRAformer-C scales 54% and 71% faster than Llama 3.2 and Composer's best Transformer, while AIRAhybrid-C outscales Nemotron-2 by 23% and Composer's best hybrid by 37%. AIRA-Design tasks 20 agents with writing novel attention mechanisms for long-range dependencies and high-performing training scripts. On the Long Range Arena benchmark, agent-designed architectures reach within 2.3% and 2.6% of human state-of-the-art on document matching and text classification. On the Autoresearch benchmark, Greedy Opus 4.5 achieves 0.968 validation bits-per-byte under a fixed time budget, surpassing the published minimum. Together, these frameworks show AI agents can autonomously discover architectures and algorithmic optimizations matching or surpassing hand-designed baselines. This establishes a powerful paradigm for discovering next-generation foundation models, marking a clear step toward recursive self-improvement.\n\n---\n\n## 20. SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows?\n- **Authors**: Kean Shi , Zihang Li , Tianyi Ma , Zengji Tu , Jialong Wu , Xinbo Xu , Qingyao Yang , Ruoyu Wu , Weichu Xie , Ming Wu , Jason Zeng , Michael Heinrich , Elvis Zhang , Liang Chen , Kuan Li , Baobao Chang\n- **URL**: [https://arxiv.org/abs/2605.15777](https://arxiv.org/abs/2605.15777)\n- **Abstract**:\n> Computer-Using Agents (CUAs) are rapidly extending large language models (LLMs) beyond text-based reasoning toward action execution in more complex environments, such as web browsers and graphical user interfaces (GUIs). However, existing web and GUI agent benchmarks often rely on simplified settings, isolated tasks, or short-horizon interactions, making it difficult to assess capabilities of agents in realistic professional workflows. Software-as-a-Service (SaaS) environments are a natural choice for CUA evaluation, as they host a large share of modern digital work and naturally involve dynamic system states, cross-application coordination, domain-specific knowledge, and long-horizon dependencies. To this end, we introduce SaaS-Bench, a benchmark built on 23 deployable SaaS systems across six professional domains, containing 106 tasks grounded in realistic work scenarios. These tasks require long-horizon execution, cover both text-only and multimodal settings, and are evaluated with weighted verification checkpoints that measure strict task completion and partial progress. Experiments show that representative LLM-based agents struggle on SaaS-Bench, with even the strongest model completing fewer than 4% of tasks end-to-end, exposing limitations in planning, state tracking, cross-application context maintenance, and error recovery. Code are available at this https URL for reproduction.\n\n---\n\n## 21. ALSO: Adversarial Online Strategy Optimization for Social Agents\n- **Authors**: Xiang Li , Liping Yi , Mingze Kong , Min Zhang , Zhongxiang Dai , QingHua Hu\n- **URL**: [https://arxiv.org/abs/2605.15768](https://arxiv.org/abs/2605.15768)\n- **Abstract**:\n> Social simulation provides a compelling testbed for studying social intelligence, where agents interact through multi-turn dialogues under evolving contexts and strategically adapting opponents. Such environments are inherently non-stationary, requiring agents to dynamically adjust their strategies over time. However, most Large Language Model (LLM) based social agents rely on static personas, while existing approaches for enhancing social intelligence, such as offline reinforcement learning or external planners, are ill-suited to these settings, typically assuming stationarity and incurring substantial training overhead. To bridge this gap, we propose \\textbf{ALSO} (\\textbf{A}dversarial on\\textbf{L}ine \\textbf{S}trategy \\textbf{O}ptimization), the first framework for online strategy optimization in multi-agent social simulation. ALSO advances social adaptation through two key contributions. (1) ALSO formulates multi-turn interaction as an adversarial bandit problem, where combinations of static personas and dynamic strategy instructions are treated as arms, providing a principled solution to non-stationarity without relying on environmental stability assumptions. (2) To predict rewards and generalize sparse feedback in multi-turn dialogues, ALSO introduces a lightweight neural surrogate to predict rewards from interaction histories, enabling sample-efficient exploration and continuous online adaptation. Experiments on the Sotopia benchmark demonstrate that ALSO consistently outperforms static baselines and existing optimization methods in dynamic environments, validating the effectiveness of adversarial online strategy optimization for building robust social agents.\n\n---\n\n## 22. Can We Trust AI-Inferred User States. A Psychometric Framework for Validating the Reliability of Users States Classification by LLMs in Operational Environments\n- **Authors**: Izabella Krzeminska , Michal Butkiewicz , Ewa Komkowska\n- **URL**: [https://arxiv.org/abs/2605.15734](https://arxiv.org/abs/2605.15734)\n- **Abstract**:\n> The use of large language models to assess user states in conversational and adaptive systems is based on the assumption that the metrics used for such assessment are stable and interpretable at the level of individual scores. This paper empirically tests this assumption, focusing on the psychometric reliability of artificial intelligence (AI) measures of user states. This study employed replication evaluation procedures to assess the repeatability of a broad set of metrics across three different bimodal large language models (GPT-4o audio, Gemini 2.0 Flash, Gemini 2.5 Flash). Analyses include both individual score reliability and aggregated reliability, allowing us to distinguish metrics potentially useful for real-time adaptation from those that retain their value only in aggregated analyses. The results demonstrate that metric reliability cannot be considered a default property in interpretive domains. The lack of stability at the level of individual scores precludes the interpretation of such scores as indicators of user state in real-time adaptive systems, even if these metrics demonstrate stability after aggregation. At the same time, the study indicates that individually unstable metrics can retain analytical utility in post-hoc studies, identifying rules governing interactions and their relationships with user experience parameters such as satisfaction, trust, and engagement. The main contribution of this work, besides quantifying the severity of the problem (only 31 of 213 metrics met the criteria), is the proposal of a replicable evaluation framework, enabling measurable evaluations of metric applicability. This approach supports more responsible AI design of adaptive systems, in which the interpretation of results requires explicit validation of reliability and monitoring for violations over time.\n\n---\n\n## 23. Nudging Beyond the Comfort Zone: Efficient Strategy-Guided Exploration for RLVR\n- **Authors**: Chanuk Lee , Sangwoo Park , Minki Kang , Sung Ju Hwang\n- **URL**: [https://arxiv.org/abs/2605.15726](https://arxiv.org/abs/2605.15726)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has emerged as a scalable paradigm for improving the reasoning capabilities of large language models. However, its effectiveness is fundamentally limited by exploration: the policy can only improve on trajectories it has already sampled. While increasing the number of rollouts alleviates this issue, such brute-force scaling is computationally expensive, and existing approaches that modify the optimization objective provide limited control over what is explored. In this work, we propose NudgeRL, a framework for structured and diversity-driven exploration in RLVR. Our approach introduces Strategy Nudging, which conditions each rollout on lightweight, strategy-level contexts to induce diverse reasoning trajectories without relying on expensive oracle supervision. To effectively learn from such structured exploration, we further propose a unified objective, which decomposes the reward signal into inter- and intra-context components and incorporates a distillation objective to transfer discovered behaviors back to the base policy. Empirically, NudgeRL outperforms standard GRPO with up to 8 times larger rollout budgets, while outperforming oracle-guided RL baseline on average across five challenging math benchmarks. These results demonstrate that structured, context-driven exploration can serve as an efficient and scalable alternative to both brute-force rollout scaling and feasibility-oriented methods based on privileged information. Our code is available at this https URL .\n\n---\n\n## 24. PRISM: Prompt Reliability via Iterative Simulation and Monitoring for Enterprise Conversational AI\n- **Authors**: Keshava Chaitanya , Jahnavi Gundakaram\n- **URL**: [https://arxiv.org/abs/2605.15665](https://arxiv.org/abs/2605.15665)\n- **Abstract**:\n> Deploying large language model (LLM)-driven conversational agents in enterprise settings requires prompts that are simultaneously correct at launch and resilient to the non-deterministic behavioral drift that characterizes production LLM deployments. Existing prompt optimization frameworks address prompt quality as a one-time compile-time problem, leaving open the equally critical question of how to detect and repair prompt regressions caused by silent LLM behavior changes over time. We present PRISM (Prompt Reliability via Iterative Simulation and Monitoring), a closed-loop framework that treats prompt engineering as a continuous reliability engineering problem rather than a one-time authorship task. PRISM takes as input plain-language agent requirements, a set of configured tools and memory variables, and an initial draft prompt. It automatically generates test cases from requirements, simulates full multi-turn conversations against a platform-faithful LLM environment, evaluates pass/fail using an LLM-as-judge, diagnoses root causes of failures, and surgically repairs the prompt -- iterating until all tests pass. Critically, PRISM is designed to run on a scheduled basis (daily), treating LLM behavioral drift as a first-class reliability concern. We evaluate PRISM across 35 enterprise conversational agents over a three-week deployment period on the this http URL V3 platform. PRISM reduces median prompt authoring time from 2 days to under 30 minutes, achieves 99% production reliability across all evaluated agents, and successfully identifies and repairs production regressions caused by LLM behavioral drift within a 24-hour detection window. Our results suggest that continuous, simulation-driven prompt optimization is both tractable and necessary for reliable enterprise conversational AI at scale.\n\n---\n\n## 25. ColPackAgent: Agent-Skill-Guided Hard-Particle Monte Carlo Workflows for Colloidal Packing\n- **Authors**: Lijie Ding , Changwoo Do\n- **URL**: [https://arxiv.org/abs/2605.15625](https://arxiv.org/abs/2605.15625)\n- **Abstract**:\n> We introduce ColPackAgent, an agent framework that autonomously runs Monte Carlo simulations of colloidal packing through a Model Context Protocol (MCP) tool server and an agent skill, whether as a standalone agent or inside an existing agent system. By harnessing the MCP server and agent skill, ColPackAgent executes a structured workflow for colloidal packing simulations, which are central to studies of phase behavior, self-assembly, and materials design. Without dedicated simulation tools and workflow instructions, general-purpose Large Language Model (LLM) agents tend to describe such workflows rather than execute them reliably. The MCP server exposes a custom-built colpack Python package that wraps HOOMD-blue hard-particle Monte Carlo, and the skill encodes a four-stage workflow contract. ColPackAgent can carry out the workflow interactively with human feedback, autonomously from an end-to-end prompt, or as autoresearch following a provided program file. We demonstrate the system in different modes with several colloidal packing simulation examples such as cube particles in 3D, a binary system of disks and capsules in 2D, and the 2D hard-disk freezing transition using autoresearch. We also compare model performance on this workflow across a panel of LLMs with 17 stage-specific prompts. This benchmark provides a stage-level check of how reliably different models follow the setup, planning, and analysis workflow. Together, these results show that pairing a domain Python package with MCP tools and a portable agent skill provides a practical route for turning a simulation toolkit into an agent-assisted research workflow.\n\n---\n\n## 26. TopoEvo: A Topology-Aware Self-Evolving Multi-Agent Framework for Root Cause Analysis in Microservices\n- **Authors**: Junle Wang , Xingchuang Liao , Wenjun Wu\n- **URL**: [https://arxiv.org/abs/2605.15611](https://arxiv.org/abs/2605.15611)\n- **Abstract**:\n> Root cause analysis (RCA) in microservices is challenging due to (i) noisy and heterogeneous multimodal observability (metrics, logs, traces), (ii) cascading failure propagation that amplifies downstream symptoms, and (iii) non-stationary topology drift induced by autoscaling and rolling updates. Recent LLM-based RCA agents can generate tool-grounded explanations, yet they often remain topology-agnostic and suffer from \\emph{symptom-amplification bias}, misattributing the root cause to salient downstream victims. We propose \\textbf{TopoEvo}, a topology-aware self-evolving multi-agent framework that couples graph representation learning with structured, topology-constrained reasoning. TopoEvo first introduces \\emph{Metric-orthogonal Multimodal Alignment} (MOMA), which decomposes metric embeddings into complementary subspaces and contrastively aligns logs and traces to reduce modality redundancy and sparsity, yielding stable node representations for graph encoding. It then applies \\emph{Vector Quantization} (VQ) to discretize topology-enhanced states into auditable \\emph{symptom tokens} with a symptom lexicon, enabling reliable retrieval and token-level evidence grounding. On top of these discrete topology cues, TopoEvo performs a multi-agent \\emph{Hypothesis--Evidence--Test} (HET) workflow to explicitly verify propagation-consistent explanations and separate initiating anomalies from amplified downstream symptoms. Finally, a \\emph{Self-Evolving Mechanism} refreshes hierarchical incident memory and performs conservative test-time adaptation with high-confidence pseudo-labels to maintain robustness under drift.\n\n---\n\n## 27. See Before You Code: Learning Visual Priors for Spatially Aware Educational Animation Generation\n- **Authors**: Yuejia Li , Ke He , Junheng Li , Shutong Chen , Jingkang Xia , Zhiyue Su , Junchi Zhang , Mang Ye\n- **URL**: [https://arxiv.org/abs/2605.15585](https://arxiv.org/abs/2605.15585)\n- **Abstract**:\n> Large language models can generate executable code for educational animations, but the resulting renders often exhibit visual defects, including element overlap, misalignment, and broken animation continuity. These defects cannot be reliably detected from the code alone and become apparent only after execution. We formalize this problem as render-feedback-aware constrained code generation: given a natural language specification, the model must generate executable code whose rendered output satisfies structured quality criteria that can be evaluated only after rendering. To address this problem, we introduce OmniManim, a render-feedback-aware educational animation generation framework built around a shared scene state, explicit visual planning, structured post-render diagnostics, and localized repair. Within OmniManim, the Vision Agent is a task-specific visual planning module: it predicts sparse keyframe layouts with coarse-to-fine bounding-box denoising and optimizes an interpolation-aware objective to reduce intermediate-frame failures induced by downstream animation interpolation. We further construct two datasets, ManimLayout-1K and EduRequire-500, and provide a reproducible evaluation protocol covering executability, instructional quality, visual quality, and efficiency. On EduRequire-500, OmniManim improves measured render quality over both single-model baselines and existing multi-agent frameworks. Systematic ablation studies further verify that explicit visual planning, especially its coarse spatial prior, bounding-box refinement, and interpolation-aware optimization, is central to these gains.\n\n---\n\n## 28. STAR: A Stage-attributed Triage and Repair framework for RCA Agents in Microservices\n- **Authors**: Junle Wang , Xingchuang Liao , Wenjun Wu\n- **URL**: [https://arxiv.org/abs/2605.15581](https://arxiv.org/abs/2605.15581)\n- **Abstract**:\n> LLM-based root cause analysis (RCA) agents have recently emerged as a promising paradigm for incident diagnosis in microservice AIOps. However, their reliability remains fragile: an error in early evidence collection, hypothesis formulation, or causal analysis can propagate through the reasoning trace and eventually corrupt the final diagnosis. In this paper, we present \\textbf{STAR}, a \\emph{Stage-attributed Triage and Repair} framework for repairing erroneous RCA traces. STAR explicitly decomposes an RCA workflow into four structured stages, namely \\emph{Evidence Package} (EP), \\emph{Hypothesis Set} (HS), \\emph{Analysis Structure} (AS), and \\emph{Decision Report} (DR), and treats agent failure as a stage-localizable reasoning bug rather than a monolithic end-to-end error. Built on top of LangGraph, STAR performs stage-wise auditing, budget-aware \\emph{Fast/Slow Routing}, \\emph{decisive stage localization via counterfactual candidate evaluation}, and stage-specific patch-and-replay repair. We evaluate STAR on a public large-scale benchmark and a real-world production dataset, using two RCA agent workflows and three foundation models. Experimental results show that STAR consistently improves both root cause localization and fault type classification over strong baselines. Moreover, STAR identifies the decisive faulty stage with high accuracy, repairs most initially incorrect traces within one or two replay rounds, and benefits substantially from both Fast/Slow Routing and counterfactual stage evaluation. These results suggest that explicitly modeling \\emph{where} an RCA agent fails is an effective path toward reliable, debuggable, and self-repairing agentic RCA systems.\n\n---\n\n## 29. Position: Artificial Intelligence Needs Meta Intelligence -- the Case for Metacognitive AI\n- **Authors**: Sergei Chuprov , Richard D. Lange , Leon Reznik , Paulo Shakarian , Raman Zatsarenko , Dmitrii Korobeinikov\n- **URL**: [https://arxiv.org/abs/2605.15567](https://arxiv.org/abs/2605.15567)\n- **Abstract**:\n> This position paper argues for metacognition as a general design principle for creating more accurate, secure, and efficient AI. The metacognitive solution involves systems monitoring their own states and judiciously allocating resources depending on each problem instance's difficulty or cost of mistakes. Drawing inspiration both from past work on resource-rational AI and from well-documented metacognitive strategies in psychology and cognitive science, we identify specific challenges in embedding these strategies into AI design and highlight open theoretical and implementation problems. We showcase these principles through a tangible example of improved learning efficiency, effectiveness, and security in a Federated Learning (FL) case study. We show how these principles can be translated into practice with a novel software framework developed specifically to allow the community to design, deploy, and experiment with metacognition-enabled AI applications.\n\n---\n\n## 30. DRS-GUI: Dynamic Region Search for Training-Free GUI Grounding\n- **Authors**: Yichao Liu , Huawen Shen , Liu Yu , Shiyu Liu , Zeyu Chen , Yu Zhou\n- **URL**: [https://arxiv.org/abs/2605.15542](https://arxiv.org/abs/2605.15542)\n- **Abstract**:\n> GUI agents powered by Multimodal Large Language Models (MLLMs) have demonstrated impressive capability in understanding and executing user instructions. However, accurately grounding instruction-relevant elements from high-resolution screenshots cluttered with irrelevant UI components remains challenging for existing approaches. Inspired by how humans dynamically adjust their perceptual scope to locate task-related regions on complex screens, we propose DRS-GUI, a training-free dynamic region search framework for GUI grounding that can be seamlessly integrated into existing MLLMs. DRS-GUI introduces a lightweight UI Perceptor that performs three human-like perceptual actions (Focus, Shift, and Scatter) to progressively explore the interface and generate region proposals. To dynamically schedule these actions, we further design an Action Planner based on Monte Carlo Tree Search (MCTS). A region quality reward is employed to evaluate and select the highly instruction-relevant region, efficiently pruning redundant UI elements. Experiments demonstrate that DRS-GUI yields a 14\\% improvement on ScreenSpot-Pro for general and GUI-specific MLLMs (Qwen2.5-VL-7B and UGround-V1-7B), significantly enhancing grounding performance and generalization.\n\n---\n\n## 31. RTL-BenchMT: Dynamic Maintenance of RTL Generation Benchmark Through Agent-Assisted Analysis and Revision\n- **Authors**: Jing Wang , Shang Liu , Hangan Zhou , Zhiyao Xie\n- **URL**: [https://arxiv.org/abs/2605.15537](https://arxiv.org/abs/2605.15537)\n- **Abstract**:\n> This paper introduces RTL-BenchMT, an agentic framework for dynamically maintaining RTL generation benchmarks. Large Language Models (LLMs) assisted automated RTL generation is one of the most important directions in EDA research. However, current RTL benchmarks face two critical challenges: (1) flawed cases in the benchmarks and (2) overfitting to the benchmarks. Both challenges are difficult to resolve purely by manual engineering effort. To address these issues and systematically reduce human maintenance costs, we propose an automated agentic framework, RTL-BenchMT. RTL-BenchMT focuses on two key applications: (1) automatically identifying and revising flawed benchmark cases and (2) automatically detecting and updating overfitting cases. With the assistance of RTL-BenchMT, we conduct a thorough, in-depth analysis of flawed and overfitting cases and produce a refined benchmark suite that will be open-sourced to the community.\n\n---\n\n## 32. CAPS: Cascaded Adaptive Pairwise Selection for Efficient Parallel Reasoning\n- **Authors**: Fangzhou Lin , Shuo Xing , Peiran Li , Siyuan Yang , Qianwen Ge , Kazunori Yamada , Ziming Zhang , Haichong Zhang , Zhengzhong Tu\n- **URL**: [https://arxiv.org/abs/2605.15513](https://arxiv.org/abs/2605.15513)\n- **Abstract**:\n> Parallel reasoning, where a generator samples many candidate solutions and an aggregator selects the best, is one of the most effective forms of test-time scaling in large language models, and pairwise self-verification has become its strongest aggregation primitive. Yet pairwise verification carries a heavy cost: each judgment reads two complete solutions in full, and existing methods perform tens of such judgments per problem regardless of whether the comparison is informative. We introduce CAPS (Cascaded Adaptive Pairwise Selection), an inference-only framework that allocates verifier compute non-uniformly along two orthogonal axes: an evidence axis that adapts how much of each candidate the judge sees, and a distribution axis that adapts how comparisons are spread across the pool. CAPS instantiates these into a four-stage cascade with an optional rescue subroutine, and admits a closed-form verifier-token cost in which the per-candidate marginal cost is roughly halved relative to uniform full-evidence schedules. On four self-verifying models (Qwen3-14B, GPT-OSS-20B, Qwen3-4B-Instruct/Thinking) and five reasoning benchmarks spanning code (LiveCodeBench-v5/v6, CodeContests) and math (AIME 2025, HMMT 2025), CAPS outperforms the leading pairwise verifier on 14 of 20 suites while using 25.4% of its verifier-token budget on code, and outperforms pointwise self-verification on all 20. The trade-off suites admit an interpretable diagnostic in terms of the verifier's accuracy at partial versus full evidence, providing a concrete pre-deployment check for cascade suitability.\n\n---\n\n## 33. X-SYNTH: Beyond Retrieval -- Enterprise Context Synthesis from Observed Human Attention\n- **Authors**: Guruprasad Raghavan , George Nychis , Rohan Narayana Murthy\n- **URL**: [https://arxiv.org/abs/2605.15505](https://arxiv.org/abs/2605.15505)\n- **Abstract**:\n> In enterprise operations, the context required for an AI agent task is scattered across systems of record, static information stores, and communication channels. What is stored is system state, a lossy representation of the work that actually happened [2, 52]. The prevailing approach [17, 31, 34, 36] retrieves by matching request content to what is stored; for narrow requests this works well. But synthesis quality depends on knowing what to surface and how to interpret it: knowledge specific to each organization, team, and individual [5, 57, 61], present in behavioral patterns, absent from any retrieval index. For complex agentic tasks it breaks down: True Lead Rate is low, False Lead Rate is high, and the model has no mechanism to improve. We present X-SYNTH, a framework for enterprise context synthesis grounded in human attention, the digitally observable interaction signatures of each worker, encoding not just what they did but the sequence in which they did it, along with implicit reward signals. Behavioral traces preceding positive outcomes are distinguishable from those that did not, without external labeling. X-SYNTH models each individual's behavioral baseline as a Digital Twin Signature (DTS) and selects among seven qualitatively distinct attention filters: Proportional, Inverse, Differential, Recurrent, Comparative, Sequential, and Collective, per individual and per query, to identify causally relevant activity signatures. A four-stage pipeline assembles ranked context grounded in behavioral patterns rather than query embeddings. On a sales lead identification task, a frontier model unaided achieves 9.5% True Lead Rate (TLR) with 90.5% False Lead Rate (FLR). Augmented with X-SYNTH, TLR rises to 61.9% (6.5x) while FLR falls to 18.8%. Enterprise context synthesis is not a retrieval problem. It is a relevance problem, and human attention is its most reliable ground truth.\n\n---\n\n## 34. From LLM-Generated Conjectures to Lean Formalizations: Automated Polynomial Inequality Proving via Sum-of-Squares Certificates\n- **Authors**: Ruobing Zuo , Hanrui Zhao , Gaolei He , Zhengfeng Yang , Jianlin Wang\n- **URL**: [https://arxiv.org/abs/2605.15445](https://arxiv.org/abs/2605.15445)\n- **Abstract**:\n> Automated proving of polynomial inequalities is a fundamental challenge in automated mathematical reasoning, where rich algebraic structure and a rapidly growing certificate search space hinder scalability. Purely symbolic approaches provide strong guarantees but often scale poorly as the number of variables or the degree increases, due to expensive algebraic manipulations and rapidly growing intermediate expressions. In parallel, LLM-guided methods have made notable progress, particularly on competition-style inequalities with a small number of variables. To address the remaining scalability challenges, we propose NSPI, a neuro-symbolic framework that combines the complementary strengths of LLMs and symbolic computation for polynomial-inequality proving. Concretely, an LLM proposes a conjecture in the form of an approximate polynomial Sum-Of-Squares (SOS) decomposition; we refine it via symbolic computation to obtain an exact polynomial SOS representation, which directly proves the target inequality, and we further certify the proof in Lean, yielding an end-to-end pipeline from heuristic discovery to machine-checked proof. Experiments on challenging benchmarks involving polynomials with up to 10 variables demonstrate the effectiveness and scalability of the proposed method.\n\n---\n\n## 35. Beyond Partner Diversity: An Influence-Based Team Steering Framework for Zero-Shot Human-Machine Teaming\n- **Authors**: Wei Sheng , Rohan Paleja\n- **URL**: [https://arxiv.org/abs/2605.15400](https://arxiv.org/abs/2605.15400)\n- **Abstract**:\n> While AI agents are rapidly advancing from isolated tools to interactive collaborators, data-driven human-machine teaming (HMT) methods remain costly in their reliance on human interaction data across domains, teammates, and team sizes. Zero-shot coordination (ZSC) addresses this bottleneck by simulating diverse partner populations to approximate how unseen partners might behave. However, partner coverage alone is insufficient as team settings scale and communication becomes degraded. To remedy this deficiency, we propose Influence-Based Team Steering (IBTS), a framework that uses influence shaping to incentivize agents to discover diverse, high-performing team interaction patterns and further steers ongoing trajectories toward stronger learned coordination modes. We assess IBTS on Overcooked-AI in both two-agent and three-agent settings, allowing us to test whether learned coordination structure transfers beyond dyadic interaction. Our evaluation includes simulated partners, synthetic partner-style variation, and, to our knowledge, the first 30-subject Overcooked-AI HMT study involving two real human teammates and one machine teammate. Across these evaluations, IBTS improves team performance against competing baselines, highlighting the need for scaled ZSC to combine sparse-reward coordination mechanisms with partner-variation coverage rather than relying on diversity alone.\n\n---\n\n## 36. Ensemble Monitoring for AI Control: Diverse Signals Outweigh More Compute\n- **Authors**: Eugene Koran , Yejun Yun , Samantha Tetef , Benjamin Arnav , Pablo Bernabeu-Pérez\n- **URL**: [https://arxiv.org/abs/2605.15377](https://arxiv.org/abs/2605.15377)\n- **Abstract**:\n> As AI systems are increasingly deployed in autonomous agentic settings at scale, it is important to ensure the actions they take are safe and aligned with user intent. Monitoring agent actions is a key safety mechanism, yet reliable monitors remain difficult to build and the scale of these systems makes human oversight impractical. We show that combining signals from diverse monitors into an ensemble improves detection of misaligned actions. We build 12 GPT-4.1-Mini monitors using both prompting and fine-tuning strategies. We evaluate them on coding tasks where candidate solutions pass standard tests but fail on adversarial inputs. In this setting, diverse ensembles outperform both individual monitors and homogeneous ensembles. Our best 3-monitor ensemble achieves 2.4x greater detection performance gain compared to an ensemble composed of three identical monitors, with the same ensemble performing strongly on an independent dataset. We contend that these results show that diversity - not scale - drives gains. The best ensembles combine strong individual performance with low correlation between monitors. Furthermore, fine-tuned monitors appear in every top-performing ensemble and maintain this advantage on out-of-distribution attack types, suggesting that fine-tuning enables detection capabilities that prompting alone does not elicit. These results support ensemble monitoring as a practical AI control strategy for safety gains at reasonable inference costs.\n\n---\n\n## 37. Belief Engine: Configurable and Inspectable Stance Dynamics in Multi-Agent LLM Deliberation\n- **Authors**: Joshua C. Yang , Maurice Flechtner , Damian Dailisan , Michiel A. Bakker\n- **URL**: [https://arxiv.org/abs/2605.15343](https://arxiv.org/abs/2605.15343)\n- **Abstract**:\n> LLM-based agents are increasingly used to simulate deliberative interactions such as negotiation, conflict resolution, and multi-turn opinion exchange. Yet generated transcripts often do not reveal why an agent's stance changes: movement may reflect evidence uptake, anchoring, role drift, echoing, or changed prompt and retrieval context. We introduce the Belief Engine (BE), an auditable belief-update layer that treats \"belief\" as an evidential state over a proposition and exposes it as scalar stance. BE extracts arguments into structured memory and updates stance with a log-odds rule controlled by evidence uptake u and prior anchoring a. Across multiple base LLMs, parameter sweeps show that these controls reliably shape stance dynamics while preserving an evidence-level update trail. On DEBATE, a human deliberation dataset with pre/post opinions, BE best reconstructs participants whose final stance follows extracted evidence; stable and evidence-opposed cases instead point to anchoring or factors outside the extracted evidence stream. BE provides configurable infrastructure for studying evidence-grounded deliberation, where openness, commitment, convergence, and disagreement can be tied to explicit update assumptions rather than hidden prompt effects.\n\n---\n\n## 38. Zero-Shot Goal Recognition with Large Language Models\n- **Authors**: Kin Max Piamolini Gusmão , Nathan Gavenski , Nir Oren , Felipe Meneguzzi\n- **URL**: [https://arxiv.org/abs/2605.15333](https://arxiv.org/abs/2605.15333)\n- **Abstract**:\n> Large language models have recently reached near-parity with classical planners on well-known planning domains, yet this competence relies on world-knowledge exploitation rather than genuine symbolic reasoning. Goal recognition is a complementary abductive task structurally better suited to LLM strengths: it consists of evaluating consistency with world knowledge rather than generating novel action sequences. This paper provides the first systematic zero-shot evaluation of frontier LLMs as goal recognisers on key classical PDDL benchmarks. Our results show that LLM competence on goal recognition is uneven: some models scale with evidence and approach landmark-based accuracy at full observations, while others remain anchored to world-knowledge priors regardless of how much evidence accumulates. Qualitative analysis of model reasoning traces reveals that this divergence reflects a fundamental difference in evidence integration rather than domain familiarity. These findings position goal recognition as a principled benchmark for the foundational planning knowledge of LLMs.\n\n---\n\n## 39. Context Pruning for Coding Agents via Multi-Rubric Latent Reasoning\n- **Authors**: Jingjing Wang , Xiwen Chen , Wenhui Zhu , Huayu Li , Zhengxiao He , Feiyang Cai , Ana S. Carreon-Rascon , Xuanzhao Dong , Feng Luo\n- **URL**: [https://arxiv.org/abs/2605.15315](https://arxiv.org/abs/2605.15315)\n- **Abstract**:\n> LLM-powered coding agents spend the majority of their token budget reading repository files, yet much of the retrieved code is irrelevant to the task at hand. Existing learned pruners compress this context with a single-objective sequence labeler, collapsing all facets of code relevance into one score and one transition matrix. We show that this formulation creates a modeling bottleneck: a single CRF transition prior must serve heterogeneous retention patterns, including contiguous semantic spans and sparse structural support lines. We propose LaMR (Latent Multi-Rubric), a structured pruning framework that decomposes code relevance into two interpretable quality dimensions, semantic evidence and dependency support, each modeled by a dedicated CRF with dimension-specific transition dynamics. A mixture-of-experts gating network dynamically weights the per-rubric emissions conditioned on the query, and a final CRF layer on the fused emissions produces the aggregate keep-or-prune decision. To supervise each dimension without additional annotation cost, we derive multi-rubric labels from the existing training corpus via AST-based program analysis, simultaneously denoising the teacher's binary labels. By effectively filtering distracting noise, LaMR frequently matches or even outperforms unpruned full-context baselines. Experiments on four benchmarks (SWE-Bench Verified, SWE-QA, LCC, LongCodeQA) show that LaMR wins 12 of 16 head-to-head multi-turn comparisons. It saves up to 31% more tokens on multi-turn agent tasks and improves Exact Match by up to +3.5 on single-turn tasks, while performance is frequently enhanced by denoising the context, and any remaining drops are marginal.\n\n---\n\n## 40. SMCEvolve: Principled Scientific Discovery via Sequential Monte Carlo Evolution\n- **Authors**: Jiachen Jiang , Huminhao Zhu , Zhihui Zhu\n- **URL**: [https://arxiv.org/abs/2605.15308](https://arxiv.org/abs/2605.15308)\n- **Abstract**:\n> LLM-driven program evolution has emerged as a powerful tool for automated scientific discovery, yet existing frameworks offer no principled guide for designing their individual components and provide no guarantee that the search converges. We introduce SMCEvolve, which recasts program search as sampling from a reward-tilted target distribution and approximates it with a Sequential Monte Carlo (SMC) sampler. From this view, three core mechanisms emerge as principled components: adaptive parent resampling, mixture of mutation with acceptance, and automatic convergence control. We further provide a finite-sample complexity analysis that bounds the LLM-call budget required to reach a target approximation error. Across math, algorithm efficiency, symbolic regression, and end-to-end ML research benchmarks, SMCEvolve surpasses state-of-the-art evolving systems while using fewer LLM calls under self-determined termination. The code is available at this https URL .\n\n---\n\n## 41. Solvita: Enhancing Large Language Models for Competitive Programming via Agentic Evolution\n- **Authors**: Han Li , Jinyu Tian , Rili Feng , Yuqiao Du , Chong Zheng , Chenyu Wang , Chenchen Liu , Shihao Li , Xinping Lei , Yifan Yao , Weihao Xie , Letian Zhu , Jiaheng Liu\n- **URL**: [https://arxiv.org/abs/2605.15301](https://arxiv.org/abs/2605.15301)\n- **Abstract**:\n> Large language models (LLMs) still struggle with the rigorous reasoning demands of hard competitive programming. While recent multi-agent frameworks attempt to bridge this reliability gap, they remain fundamentally stateless: they rely on static retrieval and discard the valuable problem-solving and debugging experience gained from previous tasks. To address this, we present Solvita, an agentic evolution framework that enables continuous learning without requiring weight updates to the underlying LLM. Solvita reorganizes problem-solving into a closed-loop system of strategy selection, program synthesis, certified supervision, and targeted hacking, executed by four specialized agents: Planner, Solver, Oracle, and Hacker. Crucially, each agent is paired with a trainable, graph-structured knowledge network. As the system operates, outcome signals, such as pass/fail verdicts, test certification quality, and adversarial vulnerabilities discovered by the Hacker, are recast as reinforcement learning updates to these network weights. This allows the agents to dynamically route future queries based on past successes and failures, effectively accumulating transferable reasoning experience over time. Evaluated across CodeContests, APPS, AetherCode, and live Codeforces rounds, Solvita establishes a new state-of-the-art among code-generation agents, outperforming existing multi-agent pipelines and nearly doubling the accuracy of single-pass baselines.\n\n---\n\n## 42. Verifiable Agentic Infrastructure: Proof-Derived Authorization for Sovereign AI Systems\n- **Authors**: Jun He , Deying Yu\n- **URL**: [https://arxiv.org/abs/2605.15228](https://arxiv.org/abs/2605.15228)\n- **Abstract**:\n> Modern cloud and enterprise systems rely on identity-centric authorization, assuming that callers possessing valid credentials are safe to execute commands. The emergence of autonomous AI agents invalidates this assumption: agents can generate syntactically valid but semantically unsafe actions, making standing privileges a significant operational risk. This risk becomes especially acute in sovereign AI systems, where autonomous agents may interact with cloud infrastructure, regulated data, financial workflows, and national-scale digital services. Governed mutation substrates reduce this risk by interposing on agent actions: agents submit intents, infrastructure evaluates context and policy, and execution is mediated. However, this shifts the trust boundary: how can the decision to authorize an intent be made verifiable, distributed, and replayable? We introduce a Distributed Trust Framework (DTF), a verification framework for governed mutation systems that computes execution authority from structured, verifiable artifacts. DTF introduces a Justification Proof to encode the admissibility basis of an action, a consensus model for independent evaluation, an ephemeral Execution Identity derived from the approved proof, and an append-only Evidence Chain that preserves the authorization lifecycle. Under stated substrate assumptions, this architecture enforces a compact authorization invariant: no high-stakes execution without a proof object, no derived authority without consensus, and no valid mutation detached from evidence. We define the model, instantiate it over an OpenKedge-based governed mutation substrate, and show how it maps onto cloud-native environments. By shifting authorization from standing identity to proof-derived authority, DTF provides an infrastructure foundation for making agentic execution governable, auditable, and bounded in sovereign AI deployments.\n\n---\n\n## 43. NIMO Controller: a self-driving laboratory orchestrator based on the Model Context Protocol\n- **Authors**: Naruki Yoshikawa , Ryo Tamura\n- **URL**: [https://arxiv.org/abs/2605.15227](https://arxiv.org/abs/2605.15227)\n- **Abstract**:\n> Self-driving laboratories (SDLs) have attracted increasing attention as a means of accelerating scientific discovery; however, developing SDL software remains technically demanding. To improve accessibility, orchestration software frameworks have been proposed to coordinate SDL components. Nevertheless, existing frameworks are primarily designed for human interaction and do not provide standardized interfaces suitable for AI agents. In this work, we propose an SDL software architecture based on the Model Context Protocol (MCP), in which all SDL functionalities are exposed through MCP servers. Following this design principle, we introduce an MCP-based SDL orchestrator, named NIMO Controller. It provides a visual programming interface automatically generated through MCP-based tool discovery, allowing human users to design experimental workflows without writing code. The same MCP backend can also be accessed by AI agents, providing a unified interface for both human users and AI agents. We demonstrate the proposed system through a case study on a color-matching SDL. The results validate the usability of the proposed MCP-based SDL architecture.\n\n---\n\n## 44. ICRL: Learning to Internalize Self-Critique with Reinforcement Learning\n- **Authors**: Jianbo Lin , Xiaomin Yu , Yi Xin , Yifu Guo , Zhuosong Jiang , Zhongqi Yue , Weishi Wang , Heqing Zou , Chengwei Qin , Hui Xiong\n- **URL**: [https://arxiv.org/abs/2605.15224](https://arxiv.org/abs/2605.15224)\n- **Abstract**:\n> Large language model-based agents make mistakes, yet critique can often guide the same model toward correct behavior. However, when critique is removed, the model may fail again on the same query, indicating that it has not internalized the critique's guidance into its underlying capability. Meanwhile, a frozen critic cannot improve its feedback quality over time, limiting the potential for iterative self-improvement. To address this, we propose learning to internalize self-critique with reinforcement learning(ICRL), a novel framework that jointly trains a solver and a critic from a shared backbone to convert critique-induced success into unassisted solver ability. The critic is rewarded based on the solver's subsequent performance gain, incentivizing actionable feedback. To address the distribution shift between critique-conditioned and critique-free behavior, ICRL introduces a distribution-calibration re-weighting ratio that selectively transfers critique-guided improvements compatible with the solver's own prompt distribution. Additionally, a role-wise group advantage estimation stabilizes joint optimization across the two roles. Together, these mechanisms ensure that the solver learns to improve itself without external critique, rather than becoming dependent on critique-conditioned behavior. We evaluate ICRL on diverse benchmarks spanning agentic and mathematical reasoning tasks, using Qwen3-4B and Qwen3-8B as backbones. Results show consistent improvements, with average gains of 6.4 points over GRPO on agentic tasks, and 7.0 points on mathematical reasoning. Notably, the learned 8B critic is comparable to 32B critics while using substantially fewer tokens. The code is available at this https URL .\n\n---\n\n## 45. NOVA: Fundamental Limits of Knowledge Discovery Through AI\n- **Authors**: Salman Avestimehr , Ken Duffy , Muriel Médard\n- **URL**: [https://arxiv.org/abs/2605.15219](https://arxiv.org/abs/2605.15219)\n- **Abstract**:\n> Can AI systems discover genuinely new knowledge through iterative self improvement, and if so, at what cost? We introduce the NOVA framework, which models the common ``generate, verify, accumulate, retrain'' loop as an adaptive sampling process over a knowledge space. We identify sufficient conditions under which accumulated genuine knowledge eventually covers a finite domain, and show how their violations produce distinct failure modes: contamination, forgetting, exploration failure, and acceptance failure. We then analyze imperfect verification and identify a contamination trap: as easy-to-find knowledge is exhausted, the model mass assigned to new valid artifacts shrinks, so even small false-positive rates can cause invalid artifacts to enter the knowledge base faster than genuine discoveries. We clarify that Good--Turing estimation is a local batch-diversity diagnostic, not an estimator of the historically undiscovered valid mass that governs long-term discovery. Under a separate tail-equivalence assumption relating the model's effective discovery distribution to a Zipf law with exponent $\\alpha>1$, we prove that the cumulative generation cost required to obtain $D$ distinct genuine discoveries satisfies $R_{\\mathrm{cum} }(D)=\\Theta(c_{\\mathrm{gen} }D^\\alpha)$, where $c_{\\mathrm{gen} }$ is the per-candidate generation cost. This scaling law quantifies asymptotic diminishing returns as the discovery frontier advances. Finally, we formalize human amplification through guidance, generation, and verification, explaining why expert input is most valuable near autonomous exploration barriers.\n\n---\n\n## 46. CAX-Agent: A Lightweight Agent Harness for Reliable APDL Automation\n- **Authors**: Chenying Lin , Yichen Hai , Yi He , Ran Wang , Haiyan Qiang , Liang Yu\n- **URL**: [https://arxiv.org/abs/2605.15218](https://arxiv.org/abs/2605.15218)\n- **Abstract**:\n> Large language models deployed for MAPDL finite-element simulation face practical reliability challenges: without structured execution control, tool encapsulation, and fault recovery, outputs may be inconsistent and task failures are common. The Agent Harness paradigm addresses this by inserting domain-specific orchestration middleware that manages tool lifecycles, workflow state, and recovery escalation. This paper presents the architecture of CAX-Agent, a lightweight agent harness purpose-built for MAPDL automation, and empirically evaluates one of its core components -- the recovery this http URL -Agent organizes execution into three layers -- LLM service, agent harness, and solver backend -- with a recovery ladder that escalates from deterministic rule patching through model-driven regeneration to context enrichment and human intervention. We evaluate three recovery strategies (no_recovery, rule_only, and model_only) on 50 standard structural benchmarks with three repeated runs per strategy (450 case-runs total). Two independent human raters score task completion under blind conditions; inter-rater agreement is strong (quadratic weighted Cohen's kappa = 0.84, 96 percent of score pairs within one point). Model_only achieves the best completion rate (0.9267), task score (3.59/4), total score (9.16/10), and zero-intervention rate (0.84), outperforming rule_only (0.7733, 3.17/4, 7.03/10, 0.00) and no_recovery (0.6933, 2.74/4, 5.60/10, 0.00) with large effect sizes (Cliff's delta = 0.81-0.87). The benchmark uses deliberately simple geometries to isolate recovery-policy effects; we discuss the scope of these findings and directions for broader validation.\n\n---\n\n## 47. Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions\n- **Authors**: Jagdish Tripathy , Marcus Buckmann\n- **URL**: [https://arxiv.org/abs/2605.15217](https://arxiv.org/abs/2605.15217)\n- **Abstract**:\n> Instruction-tuned language models exhibit behavioural fairness in high-stakes decisions while retaining biased associations in their internal representations. However, whether these suppressed representations can affect model outputs - and whether such causal potency is symmetric across demographic groups - remains unknown. We investigate the use of open-weight models for mortgage underwriting using matched applications that differ only in racially-associated names and reveal a critical disconnect: models show no output-level bias, yet retain and amplify demographic representations across model layers. Through activation steering and novel cross-layer interventions, we demonstrate that this suppressed information is decision-relevant: when reinjected at critical layers, it produces near-complete decision reversals. Critically, this latent bias is asymmetric - steering interventions affect decisions in one demographic direction, while producing minimal effects in reverse - and susceptible to adversarial prompt engineering and parameter-efficient fine-tuning. These findings demonstrate that behavioural audits focused on outputs are insufficient: fair outputs can mask exploitable internal biases. They also motivate dual-layer testing frameworks combining output evaluation with representational analysis for AI governance in high-stakes decisions.\n\n---\n\n## 48. SkillSmith: Compiling Agent Skills into Boundary-Guided Runtime Interfaces\n- **Authors**: Duling Xu , Zheng Chen , Zaifeng Pan , Jiawei Guan , Dong Dong , Jialin Li , Bangzheng Pu\n- **URL**: [https://arxiv.org/abs/2605.15215](https://arxiv.org/abs/2605.15215)\n- **Abstract**:\n> Recently, skills have been widely adopted in large language model (LLM)-based agent systems across various domains. In existing frameworks, skills are typically injected into the agent reasoning loop as contextual guidance once matched to a runtime task, enabling specialized task-solving capabilities. We find that this execution paradigm introduces two major sources of redundancy: irrelevant context injection and repeated skill-specific reasoning and planning. To this end, we propose SkillSmith, a boundary-first compiler-runtime framework that compiles skill packages offline into minimal executable interfaces. By extracting fine-grained operational boundaries from skills, SkillSmith enables agents to dynamically access and execute only the relevant components at runtime, thereby minimizing unnecessary context injection and redundant reasoning overhead. In the evaluation on SkillsBench benchmark, SkillSmith reduces solve-stage token usage by 57.44%, thinking iterations by 42.99%, solve time by 50.57% (2.02x faster), and token-proportional monetary cost by 57.44% compared with using raw-skills. Moreover, compiled artifacts produced by a stronger model can be reused by a smaller or more efficient runtime model, improving task accuracy in cases where raw skill interpretation fails. The source code and data are available at this https URL .\n\n---\n\n## 49. Does Theory of Mind Improvement Really Benefit Human-AI Interactions? Empirical Findings from Interactive Evaluations\n- **Authors**: Nanxu Gong , Zixin Chen , Haotian Li , Zishu Zhao , Jianxun Lian , Huamin Qu , Yanjie Fu , Xing Xie\n- **URL**: [https://arxiv.org/abs/2605.15205](https://arxiv.org/abs/2605.15205)\n- **Abstract**:\n> Improving the Theory of Mind (ToM) capability of Large Language Models (LLMs) is crucial for effective social interactions between these AI models and humans. However, the existing benchmarks often measure ToM capability improvement through story-reading, multiple-choice questions from a third-person perspective, while ignoring the first-person, dynamic, and open-ended nature of human-AI (HAI) interactions. To directly examine how ToM improvement techniques benefit HAI interactions, we first proposed the new paradigm of interactive ToM evaluation with both perspective and metric shifts. Next, following the paradigm, we conducted a systematic study of four representative ToM enhancement techniques using both four real-world datasets and a user study, covering both goal-oriented tasks (e.g., coding, math) and experience-oriented tasks (e.g., counseling). Our findings reveal that improvements on static benchmarks do not always translate to better performance in dynamic HAI interactions. This paper offers critical insights into ToM evaluation, showing the necessity of interaction-based assessments in developing next-generation, socially aware LLMs for HAI symbiosis.\n\n---\n\n## 50. SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch\n- **Authors**: Zhantao Wang\n- **URL**: [https://arxiv.org/abs/2605.15204](https://arxiv.org/abs/2605.15204)\n- **Abstract**:\n> Multi-agent orchestration frameworks such as LangChain, LangGraph, and CrewAI route tasks through graph-based pipelines but do not enforce the stage constraints that govern real business processes. We present SDOF, a framework that treats multi-agent execution as a constrained state machine. SDOF operates through two primary defensive layers, implemented by three components: (1) an Online-RLHF Specialized Intent Router trained via Generative Reward Modeling (GRPO) and (2) a StateAwareDispatcher with GoalStage finite-automaton checks and precondition/postcondition SkillRegistry validation for auditable execution control. On a recruitment system backed by the Beisen iTalent platform (6000+ enterprises), 185 expert-curated scenarios trigger 1671 live API calls. Our GSPO-aligned 7B Intent Router achieves higher joint accuracy than zero-shot GPT-4o on this FSM-constrained adversarial routing benchmark (80.9% versus 48.9%). In end-to-end execution, SDOF reaches 86.5% task completion (95% confidence interval 80.8 to 90.7) and blocks all 22 operations in the injection, illegal HR subset. Under a broader message-level blocking audit, SDOF attains precision 100% and recall 88%, expert agreement kappa=0.94. A separate evaluation on 960 SGD-derived dialogues spanning 8 service domains surfaces 201 stage-order conflicts under our FSM mapping, 41 of which arise in the normal split. This arXiv version reports the current validated scope; extended multi-seed training comparisons and deeper workflow evaluations will be released in a subsequent update.\n\n---\n\n## 51. DeepSlide: From Artifacts to Presentation Delivery\n- **Authors**: Ming Yang , Zhiwei Zhang , Jiahang Li , Haoseng Liu , Yuzheng Cai , Weiguo Zheng\n- **URL**: [https://arxiv.org/abs/2605.15202](https://arxiv.org/abs/2605.15202)\n- **Abstract**:\n> Presentations are a primary medium for scholarly communication, yet most AI slide generators optimize the artifact (a visually plausible deck) while under-optimizing the delivery process (pacing, narrative, and presentation preparation). We present DeepSlide, a human-in-the-loop multi-agent system that supports preparing the full presentation process, from requirement elicitation and time-budgeted narrative planning, to evidence-grounded slide--script generation, attention augmentation, and rehearsal support. DeepSlide integrates (i) a controllable logical-chain planner with per-node time budgets, (ii) a lightweight content-tree retriever for grounding, (iii) Markov-style sequential rendering with style inheritance, and (iv) sandboxed execution with minimal repair to ensure renderability. We further introduce a dual-scoreboard benchmark that cleanly separates static artifact quality from dynamic delivery excellence. Across 20 domains and diverse audience profiles, DeepSlide matches strong baselines on artifact quality while consistently achieving larger gains on delivery metrics, improving narrative flow, pacing precision, and slide--script synergy with clearer attention guidance.\n\n---\n\n## 52. IVGT: Implicit Visual Geometry Transformer for Neural Scene Representation\n- **Authors**: Yuqi Wu , Tianyu Hu , Wenzhao Zheng , Yuanhui Huang , Haowen Sun , Jie Zhou , Jiwen Lu\n- **URL**: [https://arxiv.org/abs/2605.16258](https://arxiv.org/abs/2605.16258)\n- **Abstract**:\n> Reconstructing coherent 3D geometry and appearance from unposed multi-view images is a fundamental yet challenging problem in computer vision. Most existing visual geometry foundation models predict explicit geometry by regressing pixel-aligned pointmaps, often suffering from redundancy and limited geometric continuity. We propose IVGT, an Implicit Visual Geometry Transformer that implicitly models continuous and coherent geometry from pose-free multi-view images. This formulation learns a continuous neural scene representation in a canonical coordinate system and supports continuous spatial queries at any 3D positions, retrieving local features to predict signed distance (SDF) values and colors using lightweight decoders. It allows direct extraction of continuous and coherent surface geometry, enabling rendering of RGB images, depth maps, and surface normal maps from arbitrary viewpoints. We train IVGT via multi-dataset joint optimization with 2D supervision and 3D geometric regularization. IVGT demonstrates generalization across scenes and achieves strong performance on various tasks, including mesh and point cloud reconstruction, novel view synthesis, depth and surface normal estimation, and camera pose estimation.\n\n---\n\n## 53. Designing Datacenter Power Delivery Hierarchies for the AI Era\n- **Authors**: Grant Wilkins , Fiodar Kazhamiaka , Alok Gautam Kumbhare , Chaojie Zhang , Ricardo Bianchini\n- **URL**: [https://arxiv.org/abs/2605.16255](https://arxiv.org/abs/2605.16255)\n- **Abstract**:\n> Demand for AI accelerators is rapidly increasing rack power density, with projections approaching 1MW per deployment by 2027. This poses a major challenge for datacenter power delivery designers. As power densities increase, a datacenter designed for a different target density may strand power, i.e., may be unable to use all the power that its delivery hierarchy has provisioned. Designs must remain efficient over long datacenter lifetimes and multiple hardware generations. Power utilization is particularly important as grid power capacity is a scarce resource in the AI era. Designing an efficient power delivery hierarchy for the long run is difficult because rack placement feasibility, workload impact, and cost depend jointly on electrical topology, deployment granularity, placement policy, power oversubscription, and workload mix. Moreover, each of these factors evolve over time, have inter-dependencies across multiple resource dimensions, and generally do not lend themselves to closed-form analysis. To address this challenge, we develop a framework for evaluating datacenter power delivery designs using throughput, power, and cost metrics over realistic arrival, oversubscription, and decommissioning sequences. The framework combines projection models for GPU, compute, and storage deployments with operational factors grounded in production data from Microsoft Azure. Our results show that multi-resource stranding materially changes deployable capacity, effective capital expenditure, and delivered performance, and quantify how rising density from rack- and pod-scale AI systems shapes these outcomes. For AI datacenter design, the relevant planning objective is not installed megawatts, but deployable capacity over time.\n\n---\n\n## 54. A Generative AI Framework for Intelligent Utility Billing CO 2 Analytics and Sustainable Resource Optimisation\n- **Authors**: Pavan Manjunath , Thomas Pruefer\n- **URL**: [https://arxiv.org/abs/2605.16250](https://arxiv.org/abs/2605.16250)\n- **Abstract**:\n> Distribution utilities are now expected to deliver bills that customers can actually read attach a defensible carbon number to every kWh sold and schedule load against grid stress and emissions constraints We propose an end-to-end framework that unifies four production-grade capabilities under one architectural roof a generative-AI agent that drafts each customers natural-language billing statement from structured numeric inputs under a constrained decoding policy a transformer-based forecaster that supplies the day-ahead consumption estimate with calibrated quantile bands\n\n---\n\n## 55. AI-Mediated Communication Can Steer Collective Opinion\n- **Authors**: Stratis Tsirtsis , Kai Rawal , Chris Russell , Brent Mittelstadt , Sandra Wachter\n- **URL**: [https://arxiv.org/abs/2605.16245](https://arxiv.org/abs/2605.16245)\n- **Abstract**:\n> Generative artificial intelligence (AI) is increasingly integrated into the online platforms where humans exchange opinions; large language models (LLMs) now polish users' posts on LinkedIn and provide context for content shared on X. While prior work has shown that AI can express biased opinions and shape individuals' opinions during human-AI interactions, less attention has been paid to its influence on collective opinion formation when mediating human-to-human communication. We address this gap via a combination of empirical and theoretical analyses. We show empirically that LLMs from multiple popular families introduce directional biases when instructed to edit human-written texts on contested topics, for example, nudging texts in favor of gun control and against atheism. Building on this observation, we introduce a mathematical model of opinion dynamics in which an AI system sits between users on a social network, transforming the opinions they express and perceive. By analytically characterizing the equilibrium of this model and performing simulations on real social network data, we show that biases introduced by AI in human-to-human communication can be amplified through the network and shift collective opinion in their direction. In light of these findings, we investigate whether such biases are controllable by online platforms. We audit the \"Explain this post\" feature on X and find evidence of pro-life bias in Grok's outputs on abortion-related content, which we trace back to specific design choices. We conclude with a discussion of the broader implications of our findings in relation to ongoing legislative efforts in the European Union.\n\n---\n\n## 56. Offline Semantic Guidance for Efficient Vision-Language-Action Policy Distillation\n- **Authors**: Jin Shi , Brady Zhang , Yishun Lu\n- **URL**: [https://arxiv.org/abs/2605.16241](https://arxiv.org/abs/2605.16241)\n- **Abstract**:\n> Billion-parameter Vision-Language-Action (VLA) policies have recently shown impressive performance in robotic manipulation, yet their size and inference cost remain major obstacles for real-time closed-loop control. We introduce \\textbf{VLA-AD}, a distillation framework that uses a Vision-Language Model as an offline semantic supervisor to transfer large VLA teachers into lightweight student policies. Instead of relying only on low-level action imitation, VLA-AD augments teacher-provided 7-DoF action targets with high-level semantic guidance, including task phase anchors and multi-frame operating-direction descriptions. These auxiliary signals are used only during training: at test time, the student policy runs independently, with neither the VLA teacher nor the VLM required. We evaluate VLA-AD on three LIBERO benchmark suites. Using OpenVLA-7B as the teacher, our method produces a 158M-parameter student, yielding a $44\\times$ reduction in model size while matching the teacher with only a $0.27\\%$ average relative gap. The resulting policy runs at 12.5 Hz on an RTX 4090, achieving a $3.28\\times$ inference speedup over OpenVLA-7B. We further show that the same semantic distillation pipeline generalizes to a different $\\pi_{0.5}$-4B teacher, where the student outperforms the teacher on two suites and remains within $0.53\\%$ on \\texttt{libero\\_goal}. Additional analysis indicates that phase-level supervision and multi-frame directional cues make the student less sensitive to noisy teacher actions, such as erroneous high-frequency gripper changes. Overall, VLA-AD demonstrates that offline semantic guidance from VLMs can substantially improve the efficiency, robustness, and deployability of VLA policy distillation.\n\n---\n\n## 57. Layer Equivalence Is Not a Property of Layers Alone: How You Test Redundancy Changes What You Find\n- **Authors**: Gabriel Garcia\n- **URL**: [https://arxiv.org/abs/2605.16234](https://arxiv.org/abs/2605.16234)\n- **Abstract**:\n> When researchers ask whether two transformer layers are \"equivalent\" for compression, they often conflate distinct tests. Replacement asks whether one layer's map can substitute for another's in place; interchange asks whether two layers approximately commute when their positions are swapped. Both are output-grounded swap-KL probes, but they need not agree: on pretrained transformers the protocol gap can change which layers look safe to prune by several-fold under the same evaluator, especially when replacement distances are high. We measure both protocols across checkpoints and architectures. On a Pythia training trajectory (410M and 1.4B), the replacement-interchange gap grows from initialization to convergence. Under one matched WikiText-2 contract at 8B scale, Qwen3-8B enters a divergent regime: interchange-guided removal is several-fold safer than replacement-guided at the same layer budgets, while Llama-3.1-8B ties the two protocols for pruning cost even though interchange KL is lower, showing metric gaps need not map one-to-one to removal. Before layer removal or merging, score both swap-KLs on the target checkpoint; the diagnostic requires only unlabeled forward passes.\n\n---\n\n## 58. A Unified Generative-AI Framework for Smart Energy Infrastructure: Intelligent Gas Distribution, Utility Billing, Carbon Analytics, and Quantum-Inspired Optimisation\n- **Authors**: Pavan Manjunath , Thomas pruefer\n- **URL**: [https://arxiv.org/abs/2605.16232](https://arxiv.org/abs/2605.16232)\n- **Abstract**:\n> The accelerating convergence of smart metering, generative artificial intelligence, and quantum-inspired combinatorial optimisation is reshaping how energy utilities manage physical infrastructure, customer engagement, and environmental accountability\n\n---\n\n## 59. Evaluating Design Video Generation: Metrics for Compositional Fidelity\n- **Authors**: Adrienne Deganutti , Dingning Cao , Jaejung Seol , Elad Hirsch , Purvanshi Mehta\n- **URL**: [https://arxiv.org/abs/2605.16223](https://arxiv.org/abs/2605.16223)\n- **Abstract**:\n> Generative video models are increasingly used in design animation tasks, yet no standardized evaluation framework exists for this domain. Unlike natural video generation, design animation imposes structured constraints: specific components shall animate with prescribed motion types, directions, speed and timing, while non-animated regions must remain stable and layout structure must be preserved. This paper provides a fully automated evaluation framework organized across four dimensions: layout fidelity, motion correctness, temporal quality, and content fidelity. This eliminates the reliance on subjective human evaluation and establishes a common basis for benchmarking progress in the field.\n\n---\n\n## 60. Argus: Evidence Assembly for Scalable Deep Research Agents\n- **Authors**: Zhen Zhang , Liangcai Su , Zhuo Chen , Xiang Lin , Haotian Xu , Simon Shaolei Du , Kaiyu Yang , Bo An , Lidong Bing , Xinyu Wang\n- **URL**: [https://arxiv.org/abs/2605.16217](https://arxiv.org/abs/2605.16217)\n- **Abstract**:\n> Deep research agents have achieved remarkable progress on complex information seeking tasks. Even long ReAct style rollouts explore only a single trajectory, while recent state of the art systems scale inference time compute via parallel search and aggregation. Yet deep research answers are composed of complementary pieces of evidence, which parallel rollouts often duplicate rather than complete, yielding diminishing returns while pushing the aggregation context toward the model's limit. We propose Argus, an agentic system in which a Searcher and a Navigator cooperate to treat deep research as assembling a jigsaw from complementary evidence pieces, rather than brute forcing the whole answer in parallel. The Searcher collects evidence traces for a given sub-query through ReAct-style interaction. The Navigator maintains a shared evidence graph, verifying which pieces are still missing, dispatching Searchers to gather them, and reasoning over the completed graph to produce a source-traced final answer. We train the Navigator with reinforcement learning to verify, dispatch, and synthesize, while independently training the Searcher to remain a standard ReAct agent. The resulting Navigator supports rollouts with a single Searcher or many in parallel without retraining. With both Searcher and Navigator built on a 35B-A3B MoE backbone, Argus gains 5.5 points with a single Searcher and 12.7 points with 8 parallel Searchers, averaged over eight benchmarks. With 64 Searchers it reaches 86.2 on BrowseComp, surpassing every proprietary agent we benchmark, while the Navigator's reasoning context stays under 21.5K tokens.\n\n---\n\n## 61. paper.json: A Coordination Convention for LLM-Agent-Actionable Papers\n- **Authors**: Arquimedes Canedo\n- **URL**: [https://arxiv.org/abs/2605.16194](https://arxiv.org/abs/2605.16194)\n- **Abstract**:\n> LLM agents routinely serve as first (and sometimes only) readers of academic papers, skimming for sub-claims, extracting reproducibility steps, and generalizing scope. Standard prose papers produce recurring failures in this role: sub-claims that cannot be cited at sub-paper granularity, scope overextension beyond what the paper tests, and figure commands buried in codebases rather than the paper itself. We propose ` this http URL `, a companion JSON file that travels with the PDF and addresses each failure with a lightweight convention: stable claim IDs (C1), an explicit does-not-claim list (C2), exact per-figure shell commands (C3), and stable definition IDs (C5). A fifth convention (C4) holds that minimum viable compliance, hand-written JSON alongside the PDF, is achievable in under an hour for a finished paper without touching the human-readable output. C1, C2, C3, and C5 are open invitations: an agent that reads a compliant paper and acts on it produces evidence for or against them. This paper is itself compliant: `uv run this http URL this http URL --against this http URL ` passes. Repo: this https URL\n\n---\n\n## 62. Second-Order Multi-Level Variance Correction for Modality Competition in Multimodal Models\n- **Authors**: Yishun Lu , Wes Armour\n- **URL**: [https://arxiv.org/abs/2605.16165](https://arxiv.org/abs/2605.16165)\n- **Abstract**:\n> Autoregressive next-token training offers a unified formulation for image generation and text understanding, but it also creates strong modality competition that destabilizes optimization and limits large-batch scaling. We show that first-order optimizers such as AdamW are vulnerable to cross-modality gradient heterogeneity, while second-order preconditioning, particularly SOAP, provides a more stable basis for multimodal alignment. Building on this insight, we propose \\emph{ML-FOP-SOAP}, a second-order optimization framework with Multi-Level Variance Correction. Our Fisher-Orthogonal Projection suppresses variance-induced modality conflicts, reducing the trade-off between visual generation and textual understanding. To make this practical under large gradient accumulation, we introduce a hierarchical folding strategy that captures fine-grained variance with low micro-step overhead. Experiments on Janus and Emu3 show consistent gains across both modalities and stable training at batch size 8192. Compared with AdamW, our method improves sample efficiency by up to $1.4\\times$ and accelerates wall-clock training by up to $1.5\\times$, offering a robust optimizer for scaling multimodal foundation models.\n\n---\n\n## 63. Surrogate Neural Architecture Codesign Package (SNAC-Pack)\n- **Authors**: Jason Weitz , Dmitri Demler , Benjamin Hawks , Aaron Wang , Nhan Tran , Javier Duarte\n- **URL**: [https://arxiv.org/abs/2605.16138](https://arxiv.org/abs/2605.16138)\n- **Abstract**:\n> Neural architecture search (NAS) is a powerful approach for automating model design, but existing methods often optimize for accuracy alone or rely on proxy metrics such as bit operations (BOPs) that correlate poorly with hardware cost. This gap is particularly large for FPGA deployment, where cost is dominated by a multi-dimensional budget of lookup tables, DSPs, flip-flops, BRAM, and latency. We present the Surrogate Neural Architecture Codesign Package (SNAC-Pack), an open-source AutoML framework for hardware-aware neural architecture codesign and end-to-end FPGA deployment. SNAC-Pack runs a multi-objective global search with Optuna and NSGA-II, loading trials to a shared SQLite store that enables parallel workers across compute nodes. A hardware surrogate model outputs per-trial resource and latency estimates, avoiding the synthesis cost that would otherwise dominate the search loop. A local search stage then applies quantization-aware training (QAT) together with iterative magnitude pruning in a combined compression loop, after which the final model is synthesized to FPGA firmware via the hls4ml Python library. A YAML configuration and an optional agentic frontend let users run the pipeline on new datasets without modifying the framework. We demonstrate SNAC-Pack on jet classification at the Large Hadron Collider and superconducting qubit readout, discovering compact architectures that match or exceed strong baselines on the task metric while reducing FPGA resource utilization and, in the qubit readout case, reducing the design space exploration process from months of manual fine-tuning to hours of automated search.\n\n---\n\n## 64. Navigating Potholes with Geometry-Aware Sharpness Minimization\n- **Authors**: Simon Dufort-Labbé , Mehrab Hamidi , Razvan Pascanu , Ioannis Mitliagkas , Damien Scieur , Aristide Baratin\n- **URL**: [https://arxiv.org/abs/2605.16134](https://arxiv.org/abs/2605.16134)\n- **Abstract**:\n> Sharpness-aware minimization (SAM) encourages flat minima by perturbing parameters along directions of high loss curvature, but treats all parameter directions uniformly, ignoring the underlying loss geometry. We introduce LLQR+SAM, which combines SAM with a learned preconditioner obtained from the recently proposed LLQR framework, a second-order method that recasts steepest descent as a layerwise linear-quadratic regulator problem. The preconditioner is updated sparsely and maintained as a slow exponential moving average, so it captures a smoothed, low-resolution picture of the loss landscape geometry. The SAM perturbation then operates on top of this learned geometry, probing curvature at a faster timescale. We show that this two-timescale structure is not merely a computational convenience: theoretically, the preconditioner amplifies the SAM escape signal in directions that are flat under the average geometry but locally sharp (potholes). Wide, flat basins, by contrast, remain stable. Empirically, LLQR+SAM gives consistent gains over both SAM and LLQR alone across standard vision and sequence modeling benchmarks, supporting the view that slow learned geometry and fast sharpness correction are genuinely complementary.\n\n---\n\n## 65. Entropy Across the Bridge: Conditional-Marginal Discretization for Flow and Schrödinger Samplers\n- **Authors**: Bruno Trentini , Dejan Stancevic , Michael M. Bronstein , Alexander Tong , Luca Ambrogioni\n- **URL**: [https://arxiv.org/abs/2605.16126](https://arxiv.org/abs/2605.16126)\n- **Abstract**:\n> For a fixed flow-based generative model under a small inference budget, sample quality can depend strongly on where the sampler spends its few function evaluations. Flow matching and Schrödinger bridges define probability paths, yet their inference grids are usually heuristic or inherited from one-endpoint diffusion. We derive a conditional-marginal entropy-rate objective for bridge-aware discretization, separating endpoint-conditioned bridge geometry from marginal flow evolution, and use it to build a training-free entropic inference-time scheduler from first principles. For Gaussian Brownian bridges this rate is closed-form and U-shaped, motivating boundary-heavy nonuniform grids. On trained two-dimensional bridge/flow models, the estimated profile recovers the predicted shape and improves 10-step ODE-Heun MMD over linear by 18.1%, with a paired 22.7% SDE-Heun improvement in the same low-NFE sweep. On EDM/CIFAR-10, the entropic time-discretization gives the best tested five-step FID (186.3 \\pm 4.0 versus 200.5 \\pm 2.9 for linear and 238.0 \\pm 5.3 for cosine). On AlphaFlow protein generation, entropic conditional-marginal (cond-marg) scheduling shows advantage in low-NFE regimes on both CAMEO22 and ATLAS benchmarks. These results support entropy-rate scheduling as a practical low-budget allocation signal for high-dimensional bridge and flow samplers.\n\n---\n\n## 66. GenShield: Unified Detection and Artifact Correction for AI-Generated Images\n- **Authors**: Zhipei Xu , Xuanyu Zhang , Youmin Xu , Qing Huang , Shen Chen , Taiping Yao , Shouhong Ding , Jian Zhang\n- **URL**: [https://arxiv.org/abs/2605.16122](https://arxiv.org/abs/2605.16122)\n- **Abstract**:\n> Diffusion-based image synthesis has made AI-generated images (AIGI) increasingly photorealistic, raising urgent concerns about authenticity in applications such as misinformation detection, digital forensics, and content moderation. Despite the substantial advances in AIGI detection, how to correct detected AI-generated images with visible artifacts and restore realistic appearance remains largely underexplored. Moreover, few existing work has established the connection between AIGI detection and artifact correction. To fill this gap, we propose GenShield, a unified autoregressive framework that jointly performs explainable AIGI detection and controllable artifact correction in a closed loop from diagnosis to restoration, revealing a mutually reinforcing relationship between these two tasks. We further introduce a Visual Chain-of-Thought based curriculum learning strategy that enables self-explained, multi-step ``diagnose-then-repair'' correction with an explicit stopping criterion. A high-quality dataset with large-scale ``artifact-restored'' pairs is also constructed alongside a unified evaluation pipeline. Extensive experiments on our correction benchmark and mainstream AIGI detection benchmarks demonstrate state-of-the-art performance and strong generalization of our method. The code is available at this https URL .\n\n---\n\n## 67. DebiasRAG: A Tuning-Free Path to Fair Generation in Large Language Models through Retrieval-Augmented Generation\n- **Authors**: Rui Chu , Bingyin Zhao , Thanh Quoc Hung Le , Duy Cao Hoang , Huawei Lin , Ping Li , Weijie Zhao , Khoa D Doan , Yingjie Lao\n- **URL**: [https://arxiv.org/abs/2605.16113](https://arxiv.org/abs/2605.16113)\n- **Abstract**:\n> Large language models (LLMs) have achieved unprecedented success due to their exceptional generative capabilities. However, because they depend on knowledge encapsulated from training corpora, they may produce hallucinations, stereotypes, and socially biased content. In particular, LLMs are prone to prejudiced responses involving race, gender, and age, which are collectively referred to as social biases. Prior studies have used fine-tuning and prompt engineering to mitigate such biases in LLMs, but these methods require additional training resources or domain knowledge to design the framework. Moreover, they may degrade the original capabilities of LLMs and often overlook the need for dynamic debiasing contexts for fairer inference. In this paper, we propose DebiasRAG, a novel tuning-free and dynamic query-specific debiasing framework based on retrieval-augmented generation (RAG). DebiasRAG improves fairness while preserving the intrinsic properties of LLMs, such as representation ability. DebiasRAG consists of three stages: (1) query-specific debiasing candidate generation; (2) context candidate pool construction; and (3) gradient-updated debiasing-guided context piece reranking. First, DebiasRAG leverages self-diagnosed bias contexts relevant to the query through regular retrieval, where the bias contexts are prepared offline by the DebiasRAG provider. Given the query-specific bias contexts, DebiasRAG reversely produces debiasing contexts, which are provided as additional fairness constraints for LLM outputs. Second, a regular RAG retrieval process produces query-related contexts from the regular RAG document database, such as a chunked Wikipedia dataset.\n\n---\n\n## 68. Attention Dispersion in Dynamic Graph Transformers: Diagnosis and a Transferable Fix\n- **Authors**: Jinhao Zhang , Kangfei Zhao , Qiuhao Zeng , Long-Kai Huang\n- **URL**: [https://arxiv.org/abs/2605.16112](https://arxiv.org/abs/2605.16112)\n- **Abstract**:\n> Transformer-based architectures have become the dominant paradigm for Continuous-Time Dynamic Graph (CTDG) learning, yet their performance remains limited on temporally shifted datasets. In this work, we identify attention dispersion as a shared failure mode of dynamic graph Transformers under temporal distribution shift. Through controlled ablation contrasting structurally and temporally distinguished historical neighbors against random ones, we show that prediction depends on a class of critical nodes that carry consistently more predictive signal than arbitrary neighbors. However, existing Transformers fail to focus on these nodes even when they are present in the input, as temporal shift weakens attention contrast and produces overly dispersed attention distributions. This diagnosis suggests a simple and transferable fix: replace standard attention with differential attention, which suppresses common-mode attention and amplifies distinctive token-level signals. When added to three representative CTDG Transformer baselines, differential attention consistently improves performance, with gains concentrated on high-shift datasets. Attention-level measurements further confirm the mechanism, showing reduced attention entropy and increased attention mass on critical nodes. Building on these findings, we introduce DiffDyG, a reference implementation combining differential attention with standard input encodings. Across 9 benchmarks and three negative sampling protocols, DiffDyG achieves SOTA performance, with especially large gains on the most shifted datasets.\n\n---\n\n## 69. Federated Imputation under Heterogeneous Feature Spaces\n- **Authors**: Imane Hocine , Chaimaa Medjadji , Sylvain Kubler , Gregoire Danoy , Yves Le Traon\n- **URL**: [https://arxiv.org/abs/2605.16099](https://arxiv.org/abs/2605.16099)\n- **Abstract**:\n> Federated Learning (FL) enables collaborative training across decentralized clients, but most methods assume aligned feature schemas, an assumption that rarely holds in tabular settings where clients observe only partially overlapping feature subsets. In these heterogeneous feature spaces, parameter-averaging methods (e.g., FedAvg) transfer little information across weakly overlapping or disjoint feature groups, limiting their effectiveness for federated imputation. To overcome this, we propose \\textbf{FedHF-Impute}, a federated imputation framework that separates structural feature unavailability from conventional missingness and uses a shared global feature graph to propagate information across statistically related features through message passing. This enables indirect cross-client knowledge transfer, even when features are never jointly observed locally, while preserving standard federated communication. Under simulated partial schema overlap on the SECOM and AirQuality datasets, FedHF-Impute improves imputation accuracy (RMSE) over FL baselines by 26.9\\%, and 8.4\\% respectively, while achieving comparable performance on PhysioNET, with only a 0.3\\% difference relative to the best baseline.\n\n---\n\n## 70. GeoGS-CE: Learning Delay--Beam Channel Priors with 3D Gaussians for High-Mobility Scenarios\n- **Authors**: Yumeng Zhang , Jiajia Guo , Chaozheng Wen , Chenghong Bian , Jun Zhang\n- **URL**: [https://arxiv.org/abs/2605.16094](https://arxiv.org/abs/2605.16094)\n- **Abstract**:\n> Wideband channel estimation (CE) in high-mobility scenarios remains challenging because channel responses vary rapidly, while practical systems can allocate only sparse pilots to accommodate dense users. Fortunately, many high-mobility environments, such as high-speed railways, exhibit scheduled trajectories, predictable velocities, and a limited number of dominant propagation paths. These properties induce a delay--beam power spectrum that is more stable than the instantaneous complex channel frequency response (CFR), less sensitive to the random phase coherence, and rich in geometric information. To exploit such environmental properties, we propose GeoGS-CE, a two-stage channel estimation framework for sparse-pilot high-mobility scenarios. In the offline stage, GeoGS-CE jointly models: 1) a scene-level 3D Gaussian representation that captures the non-line-of-sight (NLoS) geometric scattering support, and 2) a leakage-aware differentiable wireless rendering process that maps the NLoS Gaussians, together with an explicit virtual line-of-sight (LoS) component, to the measured delay--beam power spectrum, while accounting for practical OFDM delay and array leakage effects. In the online stage, the delay--beam power spectrum is predicted for each user location and used as a strong covariance prior, enabling accurate full-band and full-array CFR reconstruction and tracking through a linear MMSE estimator. Simulations based on channels generated from a segment of the Guangshen high-speed railway show that the proposed geometric prior substantially improves CFR reconstruction over pilot-only and non-geometric baselines.\n\n---\n\n## 71. Centralized vs Decentralized Federated Learning: A trade-off performance analysis\n- **Authors**: Chaimaa Medjadji , Guilain Leduc , Sylvain Kubler , Yves Le Traon\n- **URL**: [https://arxiv.org/abs/2605.16089](https://arxiv.org/abs/2605.16089)\n- **Abstract**:\n> Federated Learning (FL) has emerged as a promising paradigm for collaborative model training across distributed edge devices while preserving data privacy especially with the huge increase amount of data due to the adoption of technologies which contributes to the growing number of IoT devices. Storing this amount of data centrally is challenging due to issues like limited communication, privacy, and regulations. FL can be Centralized (CFL), Decentralized (DFL), and Semi-decentralized (SDFL). Choosing the right FL architecture depends on the application's needs. However, very few research studies have experimentally compared these three types of architectures to not only understand the respective strengths and limitations, but also trade-offs between different performance indicators. This paper overcome this lack of analysis, conducting experimental analyses using the Fedstellar simulator, MNIST dataset, and MLP classifier.\n\n---\n\n## 72. Multi-level Self-supervised Pretraining on Compositional Hierarchical Graph for Molecular Property Prediction\n- **Authors**: Xiayu Liu , Zhengyi Lu , Hou-biao Li\n- **URL**: [https://arxiv.org/abs/2605.16088](https://arxiv.org/abs/2605.16088)\n- **Abstract**:\n> Self-supervised pretraining on molecular graphs has emerged as a promising approach for molecular property prediction, yet most existing methods operate at a single structural granularity and treat bond information as auxiliary edge attributes rather than as an independent semantic layer. In this work, we propose MolCHG, a multi-level self-supervised pretraining framework built upon a novel Compositional Hierarchical Graph that organizes molecular structure into four types of nodes across three semantic levels. By introducing a bond graph that operates in parallel with the atom graph, our architecture elevates bond-level information to independently evolving node representations, enabling fragment nodes to aggregate atom-level and bond-level semantics on an equal footing. We design three level-specific pretraining objectives: an atom-bond cross-view contrastive task that aligns the atom-view and bond-view representations within each fragment, a fragment-level functional group prediction task to inject domain-relevant chemical knowledge, and graph-level structure prediction tasks to encode global molecular topology. Experiments on nine MoleculeNet benchmarks demonstrate that MolCHG achieves the best performance on seven datasets across both classification and regression tasks, remaining competitive with the strongest baselines on the rest. Ablation studies further confirm that the multi-level supervision signals are complementary and that each component contributes to the overall performance.\n\n---\n\n## 73. Towards Trustworthy and Explainable AI for Perception Models: From Concept to Prototype Vehicle Deployment\n- **Authors**: Till Beemelmanns , Shayan Sharifi , Manas Mehrotra , Ayushman Choudhuri , Lutz Eckstein\n- **URL**: [https://arxiv.org/abs/2605.16087](https://arxiv.org/abs/2605.16087)\n- **Abstract**:\n> Deep Neural Networks have become the dominant solution for Autonomous Driving perception, but their opacity conflicts with emerging Trustworthy AI guidelines and complicates safety assurance, debugging, and human oversight. While theoretical frameworks for safe and Explainable AI (XAI) exist, concrete implementations of Trustworthy AI for 3D scene understanding remain scarce. We address this gap by proposing a Trustworthy AI perception module that is remarkably robust, integrates faithful explainability, and calibrated uncertainty estimates. Building on a transformer-based detector, we derive explanation from the attention mechanism at inference time and validate their faithfulness using perturbation-based consistency tests. We further integrate an uncertainty estimation and calibration module, and apply robustness-enhancing training methods. Experiments show faithful saliency behavior, improved robustness, and well-calibrated uncertainty estimates. Finally, we deploy these Trustworthy AI elements in a prototype vehicle and provide an XAI Interface that visualizes documentation artifacts, model uncertainty state, and saliency maps, demonstrating the feasibility of trustworthy perception monitoring in real time. Supplementary materials are available at this https URL .\n\n---\n\n## 74. Towards Foundation Models for Relational Databases with Language Models and Graph Neural Networks\n- **Authors**: Jingcheng Wu , Ratan Bahadur Thapa , Mojtaba Nayyeri , Lucas Etteldorf , Max Finkenbeiner , Fabian Leeske , Steffen Staab\n- **URL**: [https://arxiv.org/abs/2605.16085](https://arxiv.org/abs/2605.16085)\n- **Abstract**:\n> Relational databases store much of the world's structured information, and they are essential for driving complex predictive applications. However, deep learning progress on relational data remains limited, as conventional approaches flatten databases into single tables via manual feature engineering, discarding relational context. Relational deep learning (RDL) addresses this by modeling databases as relational entity graphs (REGs) for graph neural networks (GNNs), but remains task- and database-specific. To combine the strengths of both paradigms, we propose a hybrid architecture combining a fine-tuned BART encoder to capture intra-row semantics with a GraphSAGE-based GNN over REGs to inject relational context. Experiments on RelBench show that the GNN substantially enriches BART's row embeddings, achieving a ROC-AUC of 67.40 on the driver-dnf task from the rel-f1 dataset. This performance is competitive with supervised baselines such as LightGBM (68.86) and narrows the gap to RDL (72.62) to within 5.22 points, though a substantial gap remains to state-of-the-art foundation models such as KumoRFM (82.63). These results suggest that lightweight hybrid LM-GNN architectures offer a promising and resource-efficient path towards foundation models for relational databases.\n\n---\n\n## 75. VideoSeeker: Incentivizing Instance-level Video Understanding via Native Agentic Tool Invocation\n- **Authors**: Yiming Zhao , Yu Zeng , Wenxuan Huang , Zhen Fang , Qing Miao , Qisheng Su , Jiawei Zhao , Jiayin Cai , Lin Chen , Zehui Chen , Yukun Qi , Yao Hu , Xiaolong Jiang , Feng Zhao\n- **URL**: [https://arxiv.org/abs/2605.16079](https://arxiv.org/abs/2605.16079)\n- **Abstract**:\n> Large Vision-Language Models (LVLMs) have shown significant progress in video understanding, yet they face substantial challenges in tasks requiring precise spatiotemporal localization at the instance level. Existing methods primarily rely on text prompts for human-model interaction, but these prompts struggle to provide precise spatial and temporal references, resulting in poor user experience. Furthermore, current approaches typically decouple visual perception from language reasoning, centering reasoning around language rather than visual content, which limits the model's ability to proactively perceive fine-grained visual evidence. To address these challenges, we propose VideoSeeker, a novel paradigm for instance-level video understanding through visual prompts. VideoSeeker seamlessly integrates agentic reasoning with instance-level video understanding tasks, enabling the model to proactively perceive and retrieve relevant video segments on demand. We construct a four-stage fully automated data synthesis pipeline to efficiently generate large-scale, high-quality instance-level video data. We internalize tool-calling and proactive perception capabilities into the model via cold-start supervision and RL training, building a powerful video understanding model. Experiments demonstrate that our model achieves an average improvement of +13.7% over baselines on instance-level video understanding tasks, surpassing powerful closed-source models such as GPT-4o and Gemini-2.5-Pro, while also showing effective transferability on general video understanding benchmarks. The relevant datasets and code will be released publicly.\n\n---\n\n## 76. AgriMind: An Ensemble Deep Learning Framework for Multi-Class Plant Disease Classification\n- **Authors**: Salma Hoque Talukdar Koli , Fahima Haque Talukder Jely\n- **URL**: [https://arxiv.org/abs/2605.16076](https://arxiv.org/abs/2605.16076)\n- **Abstract**:\n> Plant disease detection is still largely manual in Bangladesh, where extension workers eyeball leaf samples across millions of smallholdings. We built AgriMind to automate this: an ensemble of ResNet50, EfficientNet-B0, and DenseNet121 trained on 20,638 PlantVillage images across 15 pepper, potato, and tomato disease classes. Transfer learning with frozen ImageNet backbones and 10 epochs of head-only training keeps the pipeline lightweight. Individual models hit 96--97% on the held-out test set, but averaging their softmax outputs pushes the ensemble to 99.23% -- a two-thirds cut in error rate. We tried biasing the average toward the best validation model; it backfired. Dropping any single model also hurt. Pepper and potato classify perfectly; tomato, with ten visually similar classes, still reaches 99.01%. On an NVIDIA T4 GPU the full ensemble runs at 53 FPS. Whether that translates to real-time mobile use depends on TensorFlow Lite optimization -- work we have not yet completed.\n\n---\n\n## 77. Robust Prior-Guided Segmentation for Editable 3D Gaussian Splatting\n- **Authors**: Raushan Joshi , Jean-Yves Guillemaut\n- **URL**: [https://arxiv.org/abs/2605.16065](https://arxiv.org/abs/2605.16065)\n- **Abstract**:\n> 3D Gaussian Splatting (3D-GS) enables real-time 3D scene reconstruction but lacks robust segmentation for editing tasks such as object removal, extraction, and recoloring. Existing approaches that lift 2D segmentations to the 3D domain suffer from view inconsistencies and coarse masks. In this paper, we propose a novel framework that leverages the Segment Anything Model High Quality (SAM-HQ) to generate accurate 2D masks, addressing the limitations of the standard SAM in boundary fidelity and fine-structure preservation. To achieve robust 3D segmentation of any target object in a given scene, we introduce a prior-guided label reassignment method that assigns labels to 3D Gaussians by enforcing multiview consistency with learned priors. Our approach achieves state-of-the-art segmentation accuracy and enables interactive, real-time object editing while maintaining high visual fidelity. Qualitative results demonstrate superior boundary preservation and practical utility in Virtual Reality (VR) and robotics, advancing 3D scene editing.\n\n---\n\n## 78. Misspecified Explore-then-Exploit Leads to Supra-Competitive Prices\n- **Authors**: Jackie Baek , Vivek F. Farias , Farrell Wu\n- **URL**: [https://arxiv.org/abs/2605.16064](https://arxiv.org/abs/2605.16064)\n- **Abstract**:\n> We study whether simple algorithmic pricing systems can systematically produce collusive-like prices in multi-firm markets. We consider firms using an explore-then-exploit pipeline: they randomize prices during an initial exploration phase, then estimate demand from their own historical data and set prices myopically thereafter. The estimation step relies on a misspecified, monopoly-style model that omits competitors' prices. We characterize when this pipeline converges to supra-competitive prices above the Nash equilibrium, via a fluid-limit ordinary differential equation analysis. We show that supra-competitive prices arise when firms explore within similar price ranges on the same side of the Nash price. Moreover, prices can be substantially above the Nash price; we show that prices can reach monopoly levels under symmetric exploration. Simulations calibrated to a real multifamily rental market confirm that supra-competitive outcomes arise robustly beyond our theoretical assumptions, including under finite horizons, heterogeneous products, and nonlinear logit demand.\n\n---\n\n## 79. Ada-Diffuser: Latent-Aware Adaptive Diffusion for Decision-Making\n- **Authors**: Fan Feng , Selena Ge , Minghao Fu , Zijian Li , Yujia Zheng , Zeyu Tang , Yingyao Hu , Biwei Huang , Kun Zhang\n- **URL**: [https://arxiv.org/abs/2605.16054](https://arxiv.org/abs/2605.16054)\n- **Abstract**:\n> Recent work has framed decision-making as a sequence modeling problem using generative models such as diffusion models. Although promising, these approaches often overlook latent factors that exhibit evolving dynamics, elements that are fundamental to environment transitions, reward structures, and high-level agent behavior. Explicitly modeling these hidden processes is essential for both precise dynamics modeling and effective decision-making. In this paper, we propose a unified framework that explicitly incorporates latent dynamic inference into generative decision-making from minimal yet sufficient observations. We theoretically show that under mild conditions, the latent process can be identified from small temporal blocks of observations. Building on this insight, we introduce Ada-Diffuser, a causal diffusion model that learns the temporal structure of observed interactions and the underlying latent dynamics simultaneously, and furthermore, leverages them for planning and control. With a modular design, Ada-Diffuser supports both planning and policy learning tasks, enabling adaptation to latent variations in dynamics, rewards, and latent actions. Experiments on simulated control and robotic benchmarks demonstrate its effectiveness in accurate latent inference and adaptive policy learning.\n\n---\n\n## 80. Looped SSMs: Depth-Recurrence and Input Reshaping for Time Series Classification\n- **Authors**: Mónika Farsang , Ramin Hasani , Daniela Rus , Radu Grosu\n- **URL**: [https://arxiv.org/abs/2605.16048](https://arxiv.org/abs/2605.16048)\n- **Abstract**:\n> State Space Models (SSMs) are inherently recurrent along the sequence dimension, yet depth-recurrence - reusing the same block repeatedly across layers, as recently applied in looped transformers - has not been explored in this model family. We show that a looped SSM with $k$ parameters iterated $L$ times consistently closely matches or outperforms a standard SSM with $k \\cdot L$ independent parameters across four architectures (LRU, S5, LinOSS, LrcSSM) and six time series classification benchmarks, despite operating within a strictly smaller hypothesis space, as we formally establish. Since the larger model contains the looped model as a special case, this dominance cannot be explained by expressivity and instead points to parameter sharing across depth as a beneficial inductive bias that simplifies optimization. These results demonstrate that depth-recurrence is orthogonal to sequence-recurrence and independently beneficial. We further show that input reshaping is an equally neglected design axis: concatenating timesteps for low-dimensional inputs, or flattening and rechunking the joint feature-time dimension for high-dimensional ones, yields accuracy gains of 1-6% across all models, confirmed over 5 random seeds. Both techniques provide standalone improvements that compound when combined, suggesting that depth and input reshaping are two independent and underexplored design axes for SSMs on time series.\n\n---\n\n## 81. XSearch: Explainable Code Search via Concept-to-Code Alignment\n- **Authors**: Yiming Liu , Ruofan Liu , Yun Lin , Zicong Zhang , Weiyu Kong , Pengnian Qi , Xiao Cheng , Weinan Zhang , Qianxiang Wang , Linpeng Huang\n- **URL**: [https://arxiv.org/abs/2605.16046](https://arxiv.org/abs/2605.16046)\n- **Abstract**:\n> Semantic code search has been widely adopted in both academia and industry. These approaches embed natural-language queries and code snippets into a shared embedding space and retrieve results based on vector similarity. Despit strong performance on benchmark datasets, they often suffer from poor explainability and generalization. Retrieved code may appear semantically similar yet miss critical functional requirements of the query, while providing no explanation of why the result was retrieved. Moreover, such failures become more severe under distribution shift, where models struggle to generalize to unseen benchmarks. In this work, we propose XSearch, an intrinsically explainable code search framework. Our key insight is that by relying on global embedding similarity, existing retrievers inherently take an inductive view. They learn statistical patterns rather than truly understanding the query's functional requirements. We address this problem by reformulating code search as a deductive concept alignment problem. XSearch (i) identifies functional concepts in the query and (ii) explicitly aligns them with corresponding code statements. This explain-then-predict design produces inherent concept-level explanations and mitigates shortcut learning that harms out-of-distribution generalization. We train an encoder with explicit concept-alignment objectives and perform retrieval through explicit matching between query concepts and code statements. Experiments show that, trained on CodeSearchNet using GraphCodeBERT (125M parameters), XSearch improves performance on out-of-distribution benchmarks from 0.02 to 0.33 (15x) over eight state-of-the-art retrievers, and consistently outperforms both encoder- and decoder-based baselines with up to 7B parameters. A user study demonstrates that concept-alignment explanations enable users to evaluate retrieved results faster and more accurately.\n\n---\n\n## 82. RecMem: Recurrence-based Memory Consolidation for Efficient and Effective Long-Running LLM Agents\n- **Authors**: Zijie Dai , Shiyuan Deng , Sheng Guan , Yizhou Tian , Xin Yao , Xiao Yan , James Cheng\n- **URL**: [https://arxiv.org/abs/2605.16045](https://arxiv.org/abs/2605.16045)\n- **Abstract**:\n> Memory systems often organize user-agent interactions as retrievable external memory and are crucial for long-running agents by overcoming the limited context windows of LLMs. However, existing memory systems invoke LLMs to process every incoming interaction for memory extraction, and such an eager memory consolidation scheme leads to substantial token consumption. To tackle this problem, we propose RecMem by rethinking when memory consolidation should be conducted. RecMem stores incoming interactions in a subconscious memory layer and encode them using lightweight embedding models for retrieval. LLMs are only invoked to extract episodic and semantic memory when sustained recurrence are observed for semantically similar interactions. Such recurrence-based consolidation works because these interactions correspond to a semantic cluster with rich information and thus are worth extraction and summarization. To improve accuracy, RecMem also incorporates a semantic refinement mechanism that recovers the fine-grained facts omitted by memory extraction. Experiments show that RecMem reduces the memory construction token cost of three SOTA memory systems by up to 87% while exceeding their accuracy.\n\n---\n\n## 83. Learning Sim-Grounded Policies for Bimanual Rope Manipulation from Human Teleoperation Data\n- **Authors**: Gina Wigginghaus , Tim Missal , Berk Guler , Simon Manschitz , Jan Peters\n- **URL**: [https://arxiv.org/abs/2605.16043](https://arxiv.org/abs/2605.16043)\n- **Abstract**:\n> Deformable Linear Objects (DLOs) such as ropes and cables are widely encountered in both household and industrial applications, yet remain challenging to manipulate due to their infinite-dimensional configuration space and frequent self-occlusion. Imitation learning from teleoperation offers a practical path to bimanual DLO manipulation, but its scalability is limited by human effort, making the choice of observation space critical for generalization from small datasets. In this study, we investigate whether the lack of generalization in egocentric visual policies for the knot-untangling task stems from the observation space itself, rather than from the policy architecture or data scale. We compare two Action Chunking with Transformers policies trained on the same bimanual teleoperation data: a vision-based policy conditioned on two egocentric RGB streams from wrist-mounted cameras, and a state-based policy conditioned on the DLO's 3D particle state, extracted from an initial observation via multi-view fusion and evolved in a particle-based eXtended Position-Based Dynamics simulation. Evaluated open-loop on an unseen rope configuration, the state-based policy outperforms its visual counterpart with a 30.8% reduction in L1 error when predicting the initial grasp-and-pull action, quantifying the observability gap between pixels and physics-consistent state, and pointing toward more data-efficient robot learning for the DLO manipulation task from limited human demonstrations.\n\n---\n\n## 84. Who Owns This Agent? Tracing AI Agents Back to Their Owners\n- **Authors**: Ruben Chocron , Doron Jonathan Ben Chayim , Eyal Lenga , Gilad Gressel , Alina Oprea , Yisroel Mirsky\n- **URL**: [https://arxiv.org/abs/2605.16035](https://arxiv.org/abs/2605.16035)\n- **Abstract**:\n> AI agents are increasingly deployed to act autonomously in the world, yet there is still no reliable way to trace a harmful agent back to the account that deployed it. This creates the same accountability gap across both ends of the intent spectrum: benign operators may deploy misconfigured or overbroad agents that cause harm unintentionally, while malicious operators may deliberately weaponize agents for scams, harassment, or cyber attacks. In many cases, these agents are powered by vendor-hosted models, a dependency that holds even for sophisticated adversaries such as state actors conducting cyber operations. In either case, affected parties can observe the behavior but cannot notify the responsible operator, stop the session, or identify the account for investigation. We formalize this gap as the problem of agent attribution: linking an observed agent interaction to the responsible account at the hosting vendor. To our knowledge, this is the first work to define the problem and present a practical solution. Our protocol is canary-based: an authorized party injects a canary into the agent's interaction stream, and the vendor searches a narrow window of session logs to recover the originating session and account. Simple canaries suffice in non-adversarial settings. For adversarial operators who filter or paraphrase incoming content, we develop robust canary constructions that cannot be suppressed without degrading the agent's own task performance, yielding a formal asymmetry in the defender's favor. We evaluate a variety of scenarios including real-world agents and show that our attribution method is reliable, robust, and scalable for vendor-side deployment.\n\n---\n\n## 85. From Flat Language Labels to Typological Priors: Structured Language Conditioning for Multilingual Speech-to-Speech Translation\n- **Authors**: Yu Pan , Yang Hou , Xiongfei Wu , Liang Zhang , Yves Le Traon , Lei Ma , Jianjun Zhao\n- **URL**: [https://arxiv.org/abs/2605.16026](https://arxiv.org/abs/2605.16026)\n- **Abstract**:\n> Compositional speech-to-speech translation (S2ST) systems built upon speech large language models (SpeechLLMs) have recently shown promising performance. However, existing S2ST systems often either neglect source-language information or encode it through a language-as-label paradigm, representing each source language as an independent flat embedding. Such a design overlooks systematic linguistic structure shared across languages, which may limit data-efficient multilingual adaptation when supervised S2ST data are scarce. To address this issue, we propose S2ST-Omni 2, a many-to-one compositional S2ST framework that systematically reformulates multilingual language conditioning from flat language labels to structured typological priors. Specifically, S2ST-Omni 2 revisits language conditioning at three levels: typology-informed hierarchical language encoding for structured source-language representation, dynamically-gated language-aware Dual-CTC for content-adaptive acoustic modulation, and typology-aware LLM prompting for decoder-side linguistic guidance. Experiments on CVSS-C show that S2ST-Omni 2 achieves superior average performance among representative S2ST approaches across BLEU, COMET, ASR-BLEU, and BLASER 2.0 under the adopted evaluation protocol. Ablation studies indicate that the proposed representation-level, acoustic-level, and decoding-level strategies provide complementary benefits. Moreover, controlled data-budget analyses and a Japanese-to-English evaluation using only approximately 3 hours of supervised training data suggest that explicit typological priors provide useful inductive biases for data-efficient multilingual S2ST.\n\n---\n\n## 86. Can Vision Language Models Be Adaptive in Mathematics Education? A Learner Model-based Rubric Study\n- **Authors**: Jie Gao , Yongan Yu , Junzhu Su , Yiran Lin , Adam K. Dube , Jackie Chi Kit Cheung\n- **URL**: [https://arxiv.org/abs/2605.16011](https://arxiv.org/abs/2605.16011)\n- **Abstract**:\n> Adaptive learning refers to educational technologies that track learners' learning progress and adapt the instructional process based on individual learners' learning performance. It is increasingly recognized as critical for developing an effective learning support tool. Vision language models (VLMs) have seen adoption in mathematics education, and students have been using them as learning aids for personalized instruction. However, it is unknown whether VLMs have the ability to adapt to different learner profiles when providing mathematical instructions. Current VLMs lack a systematic evaluation framework for this adaptivity to different learner profiles in mathematics tutoring tasks. To address this gap, we draw on the learner model from the adaptive learning framework (Shute and Towle, 2018) and propose a learner model-based rubric. Our rubric formalizes adaptivity assessment into three aspects: cognitive aspects, motivational aspects, and complexity. We also evaluate two additional dimensions of VLM responses: correctness (of answers and solutions) and quality (of the response itself). Our experimental results show measurable differences in adaptivity across models and also reveal that current VLMs struggle to consistently produce learner model-based instructional responses, especially when receiving limited learner information.\n\n---\n\n## 87. CitePrism: Human-in-the-Loop AI for Citation Auditing and Editorial Integrity\n- **Authors**: Gowrika Mahesh , Budanur Madappa Darshan Gowda , Kavana Gopladevarahalli Papegowda , Prajwal Basavaraj , Binh Vu , Swati Chandna , Mehrdad Jalali\n- **URL**: [https://arxiv.org/abs/2605.16000](https://arxiv.org/abs/2605.16000)\n- **Abstract**:\n> Editors and reviewers are expected to ensure that manuscripts cite relevant, accurate, current, and ethically appropriate literature, yet manuscript-level citation auditing remains largely manual, fragmented, and difficult to scale. Citation context, metadata quality, self-citation patterns, and bibliographic integrity all affect whether a reference appropriately supports a local claim. We present CitePrism, a transparent hybrid decision-support framework for editorial citation auditing that combines LLM-assisted contextual reasoning, embedding-based semantic similarity, metadata verification, integrity-oriented flags, and human-in-the-loop analyst review. CitePrism extracts citation neighborhoods, enriches reference metadata, computes fused relevance scores, surfaces metadata and self-citation review prompts, and supports configurable threshold-based triage. In a preliminary validation on a single case-study manuscript with 104 references from pavement engineering, agreement with human binary relevance labels reached Cohen's kappa = 0.429. At operating threshold tau = 17, CitePrism flagged all human-labeled irrelevant citations, while also producing false positives requiring analyst review. These results suggest that CitePrism may support conservative editorial screening and citation-quality triage, but they do not establish general editorial performance. CitePrism is intended as pilot-stage decision support, not as an autonomous misconduct detector or automated editorial decision system. Broader validation across manuscripts, domains, annotators, baselines, and deployment settings is required before operational use.\n\n---\n\n## 88. Constrained latent state modeling: A unifying perspective on representation learning under competing constraints\n- **Authors**: Gwenolé Quellec\n- **URL**: [https://arxiv.org/abs/2605.15995](https://arxiv.org/abs/2605.15995)\n- **Abstract**:\n> Learning latent representations from complex data is central to modern machine learning, spanning temporal, multimodal, and partially observed systems. In such settings, representations are better understood as latent states capturing underlying system dynamics, rather than as mere compressed summaries of observations. Yet current approaches remain fragmented, relying on distinct -- and often implicit -- assumptions about what these states should represent. We argue that this fragmentation reflects a more fundamental limitation: latent representations are typically learned from underconstrained objectives that fail to specify the properties that meaningful latent states should satisfy. As a result, multiple representations can satisfy the same objective, leading to ambiguity in their structure and interpretation. While many of the underlying principles have been explored in isolation, their interactions have not been explicitly formalized. In this work, we propose constrained latent state modeling (CLSM) as a unifying perspective. We identify a set of core properties -- predictive sufficiency, minimality, temporal coherence, observation compatibility, invariance to nuisance factors, and structural constraints -- and show that they are intrinsically coupled through fundamental trade-offs. Revisiting major modeling families through this lens, we show that existing approaches can be interpreted as enforcing different subsets of constraints, thereby occupying distinct regions of a common design space. This perspective reframes persistent challenges such as lack of identifiability as consequences of underconstrained formulations, rather than isolated technical limitations. More broadly, CLSM provides a principled framework to make design choices explicit, to analyze trade-offs, and to guide the development of more interpretable, robust, and task-aligned latent state models.\n\n---\n\n## 89. Beyond Content: A Comprehensive Speech Toxicity Dataset and Detection Framework Incorporating Paralinguistic Cues\n- **Authors**: Zhongjie Ba , Liang Yi , Peng Cheng , Qingcao Li , Qinglong Wang , Li Lu\n- **URL**: [https://arxiv.org/abs/2605.15984](https://arxiv.org/abs/2605.15984)\n- **Abstract**:\n> Toxic speech detection has become a crucial challenge in maintaining safe online communication environments. However, existing approaches to toxic speech detection often neglect the contribution of paralinguistic cues, such as emotion, intonation, and speech rate, which are key to detecting speech toxicity. Moreover, current toxic speech datasets are predominantly text-based, limiting the development of models that can capture paralinguistic this http URL address these challenges, we present ToxiAlert-Bench, a large-scale audio dataset comprising over 30,000 audio clips annotated with seven major toxic categories and twenty fine-grained toxic labels. Uniquely, our dataset annotates toxicity sources -- distinguishing between textual content and paralinguistic origins -- for comprehensive toxic speech this http URL , we propose a dual-head neural network with a multi-stage training strategy tailored for toxic speech detection. This architecture features two task-specific classification headers: one for identifying the source of sensitivity (textual or paralinguistic), and the other for categorizing the specific toxic type. The training process involves independent head training followed by joint fine-tuning to reduce task interference. To mitigate data class imbalance, we incorporate class-balanced sampling and weighted loss this http URL experimental results show that leveraging paralinguistic features significantly improves detection performance. Our method consistently outperforms existing baselines across multiple evaluation metrics, with a 21.1% relative improvement in Macro-F1 score and a 13.0% relative gain in accuracy over the strongest baseline, highlighting its enhanced effectiveness and practical applicability.\n\n---\n\n## 90. Ontology for Policing: Conceptual Knowledge Learning for Semantic Understanding and Reasoning in Law Enforcement Reports\n- **Authors**: Anita Srbinovska , Jansen Orfan , Adrian Martin , Ernest Fokoué\n- **URL**: [https://arxiv.org/abs/2605.15978](https://arxiv.org/abs/2605.15978)\n- **Abstract**:\n> Law enforcement reports contain structured fields and written narratives. However, many incident facts that are needed for review, police training, and investigations are in natural language and require manual reading. We propose a framework using symbolic methods for converting narratives into evidence-linked facts. Our objective is to measure the value of narratives to recover incident details only from the unstructured text and build temporal graphs with time cues and domain axioms. We achieve this by redacting personal identifiers, semantic parsing, predicate mapping to ontology, and reasoning. We evaluate the symbolic approach on 450 property crime reports and a short human review. Of the extracted events from the system, 54.1% had a confidence score of at least 0.80 and 93.7% were mapped through the PropBank--VerbNet--WordNet semantic path. 100% agreement was reached on incident initiation, stolen items, and temporal cues and lower agreement for forced entry interpretation.\n\n---\n\n## 91. Reference-Free Reinforcement Learning Fine-Tuning for MT: A Seq2Seq Perspective\n- **Authors**: Ernesto Garcia-Estrada , Carlos Escolano , José A. R. Fonallosa\n- **URL**: [https://arxiv.org/abs/2605.15976](https://arxiv.org/abs/2605.15976)\n- **Abstract**:\n> Production machine translation relies overwhelmingly on encoder-decoder Seq2Seq models, yet reinforcement learning approaches to MT fine-tuning have largely targeted decoder-only LLMs at $\\geq$7B parameters, with limited systematic study of encoder-decoder architectures. We apply Group Relative Policy Optimization to NLLB-200 (600M and 1.3B) using a hybrid reference-free reward (LaBSE and COMET-Kiwi) that requires no parallel data at fine-tuning time, evaluating across 13 typologically diverse languages. GRPO yields consistent improvements on all 13 languages, up to $+$5.03 chrF++ for Traditional Chinese, and, without any target-language data, competes with 3-epoch supervised fine-tuning on morphologically complex languages . We identify a consistent empirical pattern in which gains are largest where baseline performance is weakest and reward discriminability is highest, making this approach most effective precisely where parallel data is scarcest, and replicate this pattern across English and Spanish source languages.\n\n---\n\n## 92. When and Why Adversarial Training Improves PINNs: A Neural Tangent Kernel Perspective\n- **Authors**: Yuan-dong Cao , Chi Chiu SO , Jun-Min Wang , He Wang\n- **URL**: [https://arxiv.org/abs/2605.15959](https://arxiv.org/abs/2605.15959)\n- **Abstract**:\n> Physics-informed neural networks (PINNs) are powerful surrogates for differential equations but are notoriously difficult to train due to spectral bias, stiffness, and poor accuracy on high-frequency or multiscale solutions. Adversarial training based on generative adversarial networks (GANs) has recently gained surprisingly strong empirical results in improving training, but the underlying mechanisms remain elusive. To this end, we propose a new analysis framework for adversarially trained PINNs, based on the key observation of how the discriminator in GANs can influence the training dynamics of PINNs. The framework first provides a much needed theoretical grounding to why and when adversarial training is effective in PINNs, then presents a unified analysis of GANs variants in such training, and finally leads to a new, practical, efficient training algorithm for PINNs. Empirical results demonstrate that our method can significantly reduce the pathology of PINNs training, thereby providing better models with superior performances, often several magnitudes more accurate than alternative methods.\n\n---\n\n## 93. Decomposed Vision-Language Alignment for Fine-Grained Open-Vocabulary Segmentation\n- **Authors**: Chenhao Wang , Yingrui Ji , Yu Meng , Yao Zhu\n- **URL**: [https://arxiv.org/abs/2605.15942](https://arxiv.org/abs/2605.15942)\n- **Abstract**:\n> Open-vocabulary segmentation models often struggle to generalize to unseen combinations of object categories and attributes, because fine-grained descriptions are typically encoded as holistic sentences that entangle multiple semantic units. We propose a Decomposed Vision-Language Alignment framework that explicitly factorizes textual prompts into a concept token and multiple attribute tokens, enabling separate cross-modal interactions for each semantic unit. At the feature level, we introduce a Feature-Gated Cross-Attention module that generates attribute-specific gating maps to fuse information in a multiplicative manner, effectively enforcing compositional semantics. At the scoring level, per-token similarities are aggregated in log-space, producing a stable and interpretable compositional matching. The method can be seamlessly integrated into existing transformer-based segmentation architectures and significantly improves generalization to unseen attribute-category compositions in fine-grained open-vocabulary segmentation benchmarks.\n\n---\n\n## 94. LoCO: Low-rank Compositional Rotation Fine-tuning\n- **Authors**: An Nguyen , Jaesik Choi , Anh Tong\n- **URL**: [https://arxiv.org/abs/2605.15916](https://arxiv.org/abs/2605.15916)\n- **Abstract**:\n> Parameter-efficient fine-tuning (PEFT) has emerged as an critical technique for adapting large-scale foundation models across natural language processing and computer vision. While existing methods such as low-rank adaptations achieve parameter efficiency via low-rank weight updates, they are limited in their ability to preserve the geometric structure of pretrained representations. We introduce Low-rank Compositional Orthogonal fine-tuning (LoCO), a novel PEFT method that constructs orthogonal transformations through low-rank skew-symmetric matrices and compositional rotation chains. We propose an approximation scheme that enables fully parallel computation of compositional rotations, making the approach practical for high-dimensional feature spaces. Our method maintains low computational complexity while maintaining orthogonality with controlled approximation error. We validate LoCO across diverse domains, including diffusion transformer fine-tuning, vision transformer adaptation, and language model adaptation. Our method demonstrates superior or competitive performance compared to both existing orthogonal and non-orthogonal methods.\n\n---\n\n## 95. SLIP & ETHICS: Graduated Intervention for AI Emotional Companions\n- **Authors**: Minseo Kim\n- **URL**: [https://arxiv.org/abs/2605.15915](https://arxiv.org/abs/2605.15915)\n- **Abstract**:\n> AI emotional companions face a safety-rapport paradox: restrictive safeguards can damage supportive alliance, while permissive systems risk user harm. We present SLIP (Staged Layers of Intervention Protocol), a four-stage graduated methodology deriving interventions (none, soft, hard) from structured qualitative indicators -- affect intensity (a) and narrative dynamism (m) -- alongside ETHICS (Emergent Taxonomy for Human-AI Interaction Context Signals), a \"signals not labels\" taxonomy. An evaluation combining a small-scale production deployment (N=68 entries, 10 users, 10 weeks) with a synthetic persona battery (N=91, 5 behavioral-risk profiles) achieved 0% false positives for the flow persona and showed expected escalation patterns in crisis-oriented personas. However, initial results showed that 8 consecutive days of high-energy elevation produced zero interventions (0/8), exposing a boundary where the \"do not pathologize\" principle conflicts with safety. A subsequent three-model stress test demonstrated that increased model capability improves detection from 0/8 to 6/8 while preserving 0/10 flow false positives in the largest model. Read as preliminary, these findings position graduated intervention as a design direction for navigating -- not resolving -- the safety-rapport tension in affective computing.\n\n---\n\n## 96. Towards Generalization of Block Attention via Automatic Segmentation and Block Distillation\n- **Authors**: Shuaiyi Li , Zhisong Zhang , Yan Wang , Lei Zhu , Dongyang Ma , Chenlong Deng , Yang Deng , Wai Lam\n- **URL**: [https://arxiv.org/abs/2605.15913](https://arxiv.org/abs/2605.15913)\n- **Abstract**:\n> Block attention, which processes the input as separate blocks that cannot attend to one another, offers significant potential to improve KV cache reuse in long-context scenarios such as Retrieval-Augmented Generation (RAG). However, its broader application is hindered by two key challenges: the difficulty of segmenting input text into meaningful, self-contained blocks, and the inefficiency of existing block fine-tuning methods that risk degrading performance. To address these, we first construct SemanticSeg, a large and diverse semantic segmentation dataset containing over 30k instances across 16 categories-including books, code, web text, and conversations with text lengths ranging from 2k to 32k. Using this dataset, we train a lightweight segmenter to automatically partition text into human-instinct-aligned blocks with controllable granularity. Second, we propose block distillation, a training framework that is more efficient than block fine-tuning, which uses a frozen full-attention teacher model to guide the block-attention student. This framework integrates three novel components: block sink tokens to mitigate information loss at block boundaries, block dropout to leverage training signals from all blocks, and token-level loss weighting to focus learning on block-attention-sensitive tokens. Experiments across multiple models and benchmarks demonstrate that our segmenter outperforms heuristic and statistical baselines, and block distillation achieves near-full-attention performance under block attention, establishing a practical and scalable pathway for deploying block attention.\n\n---\n\n## 97. RaPD: Resolution-Agnostic Pixel Diffusion via Semantics-Enriched Implicit Representations\n- **Authors**: Yanhao Ge , Shanyan Guan , Weihao Wang , Ying Tai , Mingyu You\n- **URL**: [https://arxiv.org/abs/2605.15908](https://arxiv.org/abs/2605.15908)\n- **Abstract**:\n> Natural images are continuous, yet most generative models synthesize them on discrete grids, limiting resolution-flexible generation. Continuous neural fields enable resolution-free rendering, but prior methods introduce continuity only at the decoding stage as an interpolation module, leaving the generative latent space discretized and reconstruction-oriented. We propose RaPD (Resolution-agnostic Pixel Diffusion), which performs diffusion in a continuous Neural Image Field (NIF) latent space. RaPD bridges this reconstruction-generation gap with Semantic Representation Guidance for generation-aware latent learning and a Coordinate-Queried Attention Renderer for coordinate-conditioned, scale-aware rendering. A single denoised latent can be rendered at arbitrary resolutions by changing only the query coordinates, keeping diffusion cost fixed. Experiments demonstrate superior generation quality and resolution scalability.\n\n---\n\n## 98. Generative Long-term User Interest Modeling for Click-Through Rate Prediction\n- **Authors**: Jiangli Shao , Kaifu Zheng , Hao Fang , Huimu Ye , Zhiwei Liu , Bo Zhang , Shu Han , Xingxing Wang\n- **URL**: [https://arxiv.org/abs/2605.15905](https://arxiv.org/abs/2605.15905)\n- **Abstract**:\n> Modeling long-term user interests with massive historical user behaviors enhances click-through rate (CTR) prediction performance in advertising and recommendation systems. Typically, a two-stage framework is widely adopted, where a general search unit (GSU) first retrieves top-$k$ relevant behaviors towards the target item, and an exact search unit (ESU) generates interest features via tailored attention. However, current target-centered GSU would ignore other latent user interests, leading to incomplete and biased interest features. Additionally, the matching-based retrieval process in GSUs depends on the pairwise similarity score between target item and each historical behavior, which not only becomes time-consuming for online services as user behaviors continue to grow, but also overlooks the interaction information among user behaviors. To combat these problems, we propose a \\textbf{Gen}erative \\textbf{L}ong-term user \\textbf{I}nterest model named GenLI for CTR prediction. GenLI consists of an interest generation module (IGM), a behavior retrieval module (BRM), and an interest fusion module (IFM). The IGM generates multiple interest distributions to indicate different aspects of real-time user interests, which is target-independent and incorporates interaction information among behaviors, ensuring complete and diverse interest features. The BRM selects related behaviors via a simple lookup operation, reducing the time complexity for weighting each behavior to $O(1)$. Finally, the IFM uses delicate gating mechanisms to generate interest features. Based on the generation process, GenLI improves the diversity of user interests and avoids complex matching-based behavioral retrieval, achieving a better balance between accuracy and efficiency for CTR prediction.\n\n---\n\n## 99. Uncertainty-Aware Wildfire Smoke Density Classification from Satellite Imagery via CBAM-Augmented EfficientNet with Evidential Deep Learning\n- **Authors**: Ranjith Chodavarapu\n- **URL**: [https://arxiv.org/abs/2605.15894](https://arxiv.org/abs/2605.15894)\n- **Abstract**:\n> Rapid and accurate wildfire smoke severity assessment from satellite images is essential for emergency response, air quality modeling, and human health risk management. Existing deep learning approaches treat smoke detection as a binary task, producing point estimates without any measure of prediction confidence. We propose a probabilistic framework to categorize a satellite patch into Light, Moderate, and Heavy severity classes and to provide decomposed epistemic and aleatoric uncertainty in a single forward pass. Our architecture uses the backbone of a pre-trained EfficientNet-B3 and a CBAM module with an evidential deep learning head that predicts Dirichlet concentration parameters, directly estimating vacuity (epistemic) and dissonance (aleatoric) without Monte Carlo sampling. Evaluated on 16,298 real satellite patches derived from the Wildfire Detection dataset, our model achieves 93.8% weighted test accuracy (91.1% unweighted) with ECE=0.0274. Selective prediction retaining the most certain 50% of patches achieves 96.7% accuracy. As image quality degrades, uncertainty increases monotonically, and vacuity is a practical scan quality measure. The Moderate class represents transitional smoke conditions that exhibit the highest epistemic uncertainty (mean vacuity = 0.187), confirming the model correctly identifies ambiguous smoke boundary regions. CBAM spatial attention maps localize to structurally distinctive scene regions, and t-SNE demonstrates the clear cluster separation of Light and Heavy smoke.\n\n---\n\n## 100. CHoE: Cross-Domain Heterogeneous Graph Prompt Learning via Structure-Conditioned Experts\n- **Authors**: Peiyuan Li , Yongqi Huang , Jitao Zhao , Dongxiao He , Di Jin , Weixiong Zhang\n- **URL**: [https://arxiv.org/abs/2605.15888](https://arxiv.org/abs/2605.15888)\n- **Abstract**:\n> Heterogeneous Graph Prompt Learning (HGPL)has emerged as a promising paradigm for bridging the gap between the objectives of pre-training foundation models and their downstream applications in heterogeneous graph settings. However, existing HGPL methods are primarily designed for in-domain scenarios, whereas real-world deployments often span multiple domains, and the data used for pre-training and downstream tasks may originate from different distributions. Consequently, the applicability of current HGPL approaches is limited to in-domain settings, and their performance typically degrades when application domains shift. To address this serious limitation, we develop CHoE, a cross-domain HGPL method built upon an expert network. During pre-training, we introduce and train structure-conditioned experts, and during prompt tuning, we adopt a structure-aware expert routing and load balancing mechanism to select structurally compatible experts for each meta-path view. In addition, we design a prompt-based semantic fusion module to integrate representations across multiple views for downstream prediction. Extensive experiments show that CHoE consistently improves performance in few-shot cross-domain applications, outperforming all baseline approaches.\n\n---\n\n## 101. Symplectic Neural Operators for Learning Infinite Dimensional Hamiltonian Systems\n- **Authors**: Yeang Makara , Yusuke Tanaka , Takashi Matsubara , Takaharu Yaguchi\n- **URL**: [https://arxiv.org/abs/2605.15881](https://arxiv.org/abs/2605.15881)\n- **Abstract**:\n> The modeling and simulation of infinite-dimensional Hamiltonian systems are central problems in mathematical physics and engineering, however they pose significant computational and structural challenges for standard data-driven architectures. In this work, we introduce the Symplectic Neural Operator, a neural operator architecture designed to preserve the symplectic structure intrinsic to Hamiltonian PDEs. We provide a theoretical characterization of their symplecticity and establish a rigorous long-term stability result based on the combination of symplectic structure preservation and learning accuracy. Numerical experiments on canonical Hamiltonian PDEs corroborate this theoretical result and show that SNOs exhibit improved energy behavior compared with non-structure-preserving neural operators.\n\n---\n\n## 102. FSCM: Frequency-Enhanced Spatial-Spectral Coupled Mamba for Infrared Hyperspectral Image Colorization\n- **Authors**: Tingting Liu , Yuan Liu , Guiping Chen , Xiubao Sui , Qian Chen\n- **URL**: [https://arxiv.org/abs/2605.15880](https://arxiv.org/abs/2605.15880)\n- **Abstract**:\n> Thermal infrared imaging is robust to illumination variations and smoke interference, making it important for all-weather perception. However, the lack of natural color and fine texture limits target recognition, human visual interpretation, and the transfer of visible-light models. Existing infrared colorization methods mainly rely on single-band images, where insufficient spectral cues may lead to structural distortion and semantic confusion. Although infrared hyperspectral images provide rich spectral responses and material information, existing single-band frameworks remain limited in modeling spatial-spectral coupling and weak texture details. To address these issues, this paper presents FSCM, a spectral-information-guided GAN framework. Within FSCM, a frequency-enhanced spatial-spectral state-space generator composed of cascaded FSB units is constructed. Each FSB integrates three complementary components: state-space modeling captures global spatial-spectral dependencies; the frequency enhancement module (FEM) combines multi-level wavelet decomposition and Fourier gating to recover structural contours, directional high-frequency details, and global frequency responses; and the dual-stream hybrid gating module (DGM) integrates deformation-aware sampling with sparse attention to enhance effective local structures and suppress background interference. Additionally, an online semantic segmentation-guided loss is introduced to constrain the generated results, improving semantic consistency in complex road scenes. Experiments show that FSCM outperforms existing infrared colorization methods in visual quality and semantic fidelity.\n\n---\n\n## 103. Shapley Neuron Values for Continual Learning: Which Neurons Matter Most?\n- **Authors**: Mohammad Ali Vahedifar , Abhisek Ray , Qi Zhang\n- **URL**: [https://arxiv.org/abs/2605.15877](https://arxiv.org/abs/2605.15877)\n- **Abstract**:\n> Continual learning enables neural networks to learn tasks sequentially without forgetting previously acquired knowledge. However, neural networks suffer from catastrophic forgetting, where learning new tasks degrades performance on earlier ones. We address this problem with Shapley Neuron Valuation (SNV), a principled framework that quantifies Neuron importance in continual learning, grounded in cooperative game theory. SNV selectively freezes important Neurons while keeping others plastic, enabling buffer-free continual learning without expanding architecture. Experiments on ImageNet-1k show that SNV consistently outperforms existing buffer-free methods. In particular, SNV improves accuracy by +2.88% in the class incremental learning and +6.46% in the task incremental learning scenarios compared to the second baseline.\n\n---\n\n## 104. Access Timing as Scaffolding: A Reinforcement Learning Approach to GenAI in Education\n- **Authors**: Janne Rotter , Pau Benazet i Montobbio , Davinia Hernández-Leo\n- **URL**: [https://arxiv.org/abs/2605.15850](https://arxiv.org/abs/2605.15850)\n- **Abstract**:\n> In recent years, generative AI (GenAI) in educational settings has become ubiquitous in students' daily lives, despite its potential to induce over-reliance, metacognitive disengagement, and diminished learning when used unrestrictedly. While most prior research has thus focused on how to pedagogically scaffold its usage, the question of when to allow off-the-shelf GenAI remains understudied and lacks pedagogically grounded empirical investigation. We treat access timing itself as a form of implicit scaffolding and operationalize it through a reinforcement learning (RL) agent that decides when students should access GenAI, with a reward function grounded in metacognitive theory, cognitive load theory, and productive failure. In a mixed-methods controlled lab study with N=105 participants, we compared the agent's effect on learning gains and metacognitive engagement to unrestricted and fully restricted use. Results show that strategically timed GenAI access under the reinforcement learning condition improved objective post-test performance and metacognitive accuracy compared with unrestricted access, while reducing task errors and time on task relative to complete withholding, all without the need for explicit metacognitive prompts or structured scaffolding. However, no between-condition differences emerged on self-reported metacognitive awareness. Overall, timing of GenAI access therefore is a tractable, theoretically grounded, and scalable pedagogical paradigm that improves over completely unrestricted and withheld access, compatible with off-the-shelf tools and potentially low adoption barrier. This opens up a new research area that explores how access timing can be facilitated by educators and implemented in human-AI learning system design.\n\n---\n\n## 105. RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades\n- **Authors**: Xinbo Xu , Ruihan Yang , Haiyang Shen , Wendong Xu , Bofei Gao , Ruoyu Wu , Kean Shi , Weichu Xie , Xuanzhong Chen , Ming Wu , Jason Zeng , Michael Heinrich , Elvis Zhang , Liang Chen , Kuan Li , Baobao Chang\n- **URL**: [https://arxiv.org/abs/2605.15846](https://arxiv.org/abs/2605.15846)\n- **Abstract**:\n> Coding agents are increasingly deployed in real software development, where a single version iteration requires months of coordinated work across many files. However, most existing benchmarks focus predominantly on single-issue bug fixes from Python repositories, with coarse pass/fail evaluation outcomes, and thus fail to capture long-horizon, multi-target development at real engineering scale. To address this gap, we present RoadmapBench, a benchmark of 115 long-horizon coding tasks grounded in real open-source version upgrades across 17 repositories and 5 programming languages. Each task places the agent on a source-version code snapshot and provides a multi-target roadmap instruction requiring it to implement the functionality introduced in the target version, with a median modification of 3,700 lines across 51 files. We conduct a systematic evaluation on thirteen frontier models and find that even the strongest, Claude-Opus-4.7, resolves only 39.1% of tasks, while the weakest achieves merely 5.2%, in stark contrast to existing bug-fix benchmarks, suggesting that long-horizon software development remains a largely unsolved problem.\n\n---\n\n## 106. GAP: Geometric Anchor Pre-training for Data-Efficient Visuomotor Learning of Manipulation Tasks\n- **Authors**: Davide Buoso , Andrea Protopapa , Stefano Di Carlo , Francesca Pistilli , Giuseppe Averta\n- **URL**: [https://arxiv.org/abs/2605.15836](https://arxiv.org/abs/2605.15836)\n- **Abstract**:\n> Learning visuomotor policies from scarce expert demonstrations remains a core challenge in robotic manipulation. A primary hurdle lies in distilling high-dimensional RGB representations into control-relevant geometry without overfitting. While using frozen pre-trained Vision Foundation Models (VFMs) improves data efficiency, it also shifts most task adaptation onto a small spatial pooling module, which can latch onto task-irrelevant shortcuts and lose geometric grounding when finetuned with few data samples. More broadly, pre-trained visual representations used for policy learning have been observed to struggle under even minor scene perturbations, highlighting the need for robustness-oriented inductive biases. We propose Geometric Anchor Pre-training (GAP), a simple, action-free warm-up stage that regularizes the spatial adapter before downstream imitation learning. GAP pre-trains the pooling layer on a lightweight simulated proxy task where object masks are available at no cost, encouraging the adapter to produce keypoints that lie on the object, cover its spatial extent, and remain sharp and repeatable over time. This yields stable geometric anchors that provide a reliable coordinate interface for few-shot policy learning, while keeping the VFM frozen. We evaluate GAP on RoboMimic and ManiSkill under severe data scarcity (15-50 demonstrations) and domain shift. A simple adapter regularized with GAP consistently outperforms stronger attention-based poolers and end-to-end fine-tuning, achieving 62% success on RoboMimic Can with 15 demonstrations (+16% over AFA), 63% on the long-horizon high-precision Tool Hang task with 50 demonstrations, and 61% on ManiSkill StackCube with 30 demonstrations (+11% over full fine-tuning). The proxy stage is lightweight and fully decoupled from downstream tasks, making it practical to reuse across environments and manipulation skills.\n\n---\n\n## 107. Modeling Music as a Time-Frequency Image: A 2D Tokenizer for Music Generation\n- **Authors**: Yuqing Cheng , Xingyu Ma , Guochen Yu , Xiaotao Gu\n- **URL**: [https://arxiv.org/abs/2605.15831](https://arxiv.org/abs/2605.15831)\n- **Abstract**:\n> Autoregressive music generation depends strongly on the audio tokenizer. Existing high-fidelity codecs often use residual multi-codebook quantization, which preserves reconstruction quality but complicates language modeling after sequence flattening, as the residual hierarchy imposes strong sequential dependencies and can amplify error accumulation. We propose BandTok, a generation-oriented 2D Mel-spectrogram tokenizer that represents each frame with Mel-frequency band tokens from a single shared codebook. This design yields a physically interpretable time-frequency token grid with a more independent token structure, making it better suited for autoregressive modeling. BandTok improves reconstruction with a multi-scale PatchGAN objective and EMA codebook updates. We further introduce an autoregressive language model with 2D Rotary Position Embedding (2D RoPE) to preserve temporal and frequency-band structure during generation. Experiments show that BandTok improves over residual-codebook tokenizers and achieves strong results in a data-limited setting. The source code and generation demos for this work are publicly available.\n\n---\n\n## 108. Toward Natural and Companionable Virtual Agents via Cross-Temporal Emotional Modeling\n- **Authors**: Feier Qin , Xiao Li , Yi Zheng , Haibin Huang , Hanyao Wang , Xiaoyu Wang , Yan Lu , Yuan Zhang\n- **URL**: [https://arxiv.org/abs/2605.15812](https://arxiv.org/abs/2605.15812)\n- **Abstract**:\n> Recent advances in foundation models have enabled conversational agents that aim for sustained companionship rather than mere task completion. Yet most still remain unable to support natural, long-term companion-like interactions, resulting in experiences that feel episodic and inauthentic. We argue that current agents overlooked cross-temporal modeling of agents' social behaviors and internal emotions: generated behaviors rarely influence an agent's emotional state, and emotional states seldom shape subsequent behaviors. We present Cross-Temporal Emotion Modeling (CTEM), a framework that links long-term behavioral history to moment-to-moment emotional expression. CTEM establishes a closed loop where past experiences update an evolving emotional state; this state conditions immediate interactions; and user feedback continually revises both memory and emotional state, enabling reflection and anticipation. We instantiate CTEM as Auri, a companion agent on an instant-messaging platform, and report a 21-day in-the-wild study showing that CTEM shows improvements in perceived naturalness, coherence, and emotional harmony.\n\n---\n\n## 109. Grokking as Structural Inference: Transformers Need Bayesian Lottery Tickets\n- **Authors**: Kai Hidajat , Solden Stoll , Joseph An\n- **URL**: [https://arxiv.org/abs/2605.15787](https://arxiv.org/abs/2605.15787)\n- **Abstract**:\n> Why does a Transformer that has memorized its training set wait thousands of steps before it generalizes? Existing accounts locate this delay in norm minimization, feature emergence, or the late discovery of sparse subnetworks. These explanations capture important parts of the transition, but ignore a constraint unique to attention-based models: if attention discards an informative token, no bounded downstream computation can recover it. We formalize attention as an implicit Bayesian posterior over the task dependency graph and prove that generalization requires two separable conditions: a familiar Goldilocks bound on MLP capacity, coinciding with norm-based theories of grokking, and a novel Bayesian structural condition requiring attention to place sufficient mass on every informative token. This decoupling explains delayed generalization as delayed structural inference. Early in training, the MLP memorizes through unaligned features, drives the cross-entropy loss near zero, and thereby starves attention of structural gradient. Weight decay must then erode memorization before the missing graph becomes learnable, yielding the known inverse-weight-decay delay, which we derive as a structural waiting time. We then prove that this explaining-away delay can be bypassed by a KL-based structural intervention, yielding an inverse-intervention-strength scaling law for the grokking time. Experiments on algorithmic sequence tasks isolate structure from capacity and show that this Bayesian ticket matches or outperforms lottery-ticket transfer.\n\n---\n\n## 110. A Topology-Aware Spatiotemporal Handover Framework for Continuous Multi-UAV Tracking\n- **Authors**: Jianlin Ye , Christos Kyrkou , Panayiotis Kolios\n- **URL**: [https://arxiv.org/abs/2605.15779](https://arxiv.org/abs/2605.15779)\n- **Abstract**:\n> The integration of Unmanned Aerial Vehicles(UAVs) into Intelligent Transportation Systems (ITS) offers synoptic visibility for traffic monitoring, yet scalable deployment is hindered by trajectory fragmentation, where vehicle identity persistence is lost across multi-UAV Fields of View (FOV). While state-of-the-art frameworks excel in optimizing local trajectory extraction and stability for single-drone imagery, they often function as isolated data silos that generate disjointed trajectories, thereby precluding network-level analysis such as Origin-Destination estimation. This paper presents a real-time Multi-Camera Multi-Vehicle Tracking (MCMT) system designed to handle global identity persistence. Addressing the visual ambiguity and computational cost of appearance-based Re-Identification (Re-ID) in nadir views, we introduce a lightweight Topology-Based Spatiotemporal Handover mechanism. We implement a high-throughput parallel pipeline leveraging YOLO11 and ByteTrack to process concurrent 4K streams. Our core contribution is a deterministic queue-based matching algorithm that utilizes geometric overlaps and virtual lane discretization to predictively manage identity handover via FIFO queues. Experimental results on complex urban environments, including intersections and merging traffic, demonstrate a Handover Success Rate (HOSR) of 99.8% in continuous traffic flows, significantly outperforming Re-ID baselines (74.1%) while validating edge deployment feasibility. The source code is available at this https URL .\n\n---\n\n## 111. Lamarckian Inheritance in Dynamic Environments: How Key Variables Affect Evolutionary Dynamics\n- **Authors**: K. Ege de Bruin , Kyrre Glette , Kai Olav Ellefsen\n- **URL**: [https://arxiv.org/abs/2605.15769](https://arxiv.org/abs/2605.15769)\n- **Abstract**:\n> The co-optimization of a robot's body and brain presents a coupled challenge: the morphology constrains which control strategies are effective, while the control determines how well the morphology performs. To address this, we combine morphology optimization as evolution with controller optimization as lifetime learning, utilizing Lamarckian inheritance to transfer learned controller parameters from parent to offspring. In dynamic environments, existing literature presents conflicting evidence: while traditional evolutionary theory often suggests Lamarckian inheritance lacks benefit, recent studies in evolutionary robotics indicate it can improve performance. We hypothesize that this is because previous works have not included all relevant variables with dynamic environments. In this work, we show that the benefit of Lamarckian inheritance depends on two variables: how conflicting the environmental changes are to robot control, and the predictability of those changes for the robotic agent. Using virtual soft robots and two different learning approaches, Bayesian optimization and reinforcement learning, we show that Lamarckian inheritance only underperforms Darwinian inheritance when the changes are both conflicting and unpredictable. We find that adding a sensor to detect environmental changes restores the benefits for Lamarckian inheritance in conflicting environments, by allowing robotic agents to predict the need for a different behavior, thereby generalizing their control.\n\n---\n\n## 112. GRASP: Learning to Ground Social Reasoning in Multi-Person Non-Verbal Interactions\n- **Authors**: Junho Kim , Xu Cao , Houze Yang , Bikram Boote , Ana Jojic , Fiona Ryan , Bolin Lai , Sangmin Lee , James M. Rehg\n- **URL**: [https://arxiv.org/abs/2605.15764](https://arxiv.org/abs/2605.15764)\n- **Abstract**:\n> Understanding social interactions requires reasoning over subtle non-verbal cues, yet current multimodal large language models (MLLMs) often fail to identify who interacts with whom in multi-person videos. We introduce GRASP, a large-scale social reasoning dataset that connects high-level social QA with fine-grained gaze and deictic gesture events. GRASP contains 290K question--answer pairs over 46K videos totaling 749 hours, organized by a 16-category taxonomy spanning gaze, gesture, and joint gaze--gesture reasoning, together with GRASP-Bench for evaluation. Unlike prior resources that focus on either isolated cues or high-level social QA, GRASP builds questions from identity-consistent gaze trajectories, deictic gestures, and their joint compositions into social events. Moreover, we propose Social Grounding Reward (SGR), a learning signal that uses these social events to encourage models to reason about the participants involved in each interaction. Experiments show that SGR improves performance on GRASP-Bench while maintaining zero-shot performance on related social video QA benchmarks.\n\n---\n\n## 113. CompactQE: Interpretable Translation Quality Estimation via Small Open-Weight LLMs\n- **Authors**: Kamil Guttmann , Zofia Fraś , Artur Nowakowski , Krzysztof Jassem\n- **URL**: [https://arxiv.org/abs/2605.15763](https://arxiv.org/abs/2605.15763)\n- **Abstract**:\n> Current state-of-the-art Quality Estimation (QE) in machine translation relies on massive, proprietary LLMs, raising data privacy concerns. We demonstrate that smaller, open-source LLMs (<30B parameters) are a viable, cost-effective and privacy-preserving alternative. Using a single-pass prompting strategy, our models simultaneously generate quality scores, MQM error annotations, suggested error corrections, and full post-editions. Our analysis shows these models achieve highly competitive system-level correlations with human judgments that outperform traditional neural metrics, fine-tuned models, and human inter-annotator agreement, effectively approximating the capabilities of much larger proprietary LLMs.\n\n---\n\n## 114. BiomedAP: A Vision-Informed Dual-Anchor Framework with Gated Cross-Modal Fusion for Robust Medical Vision-Language Adaptation\n- **Authors**: Huanyang Tong , Kai Liu , Fangjun Kuang , Huiling Chen\n- **URL**: [https://arxiv.org/abs/2605.15736](https://arxiv.org/abs/2605.15736)\n- **Abstract**:\n> Biomedical Vision--Language Models (VLMs) have shown remarkable promise in few-shot medical diagnosis but face a critical bottleneck: \\textit{fragility to prompt variations}.Existing adaptation frameworks typically optimize visual and textual prompts as independent streams, relying on ideal ``Golden Prompts''. In clinical reality, where descriptions are often noisy and heterogeneous, this modality isolation leads to unstable cross-modal alignment. To address this, we propose BiomedAP, a vision-informed dual-anchor framework with gated cross-modal this http URL enforces synergistic alignment through two mechanisms: (1) Gated Cross-Modal Fusion, which enables layer-wise interaction between modalities, acting as a dynamic noise regulator to suppress irrelevant textual cues; and (2) a Dual-Anchor Constraint that regularizes learnable prompts toward stable semantic centroids derived from both expert templates (High Anchors) and few-shot visual prototypes (Low Anchors). Extensive experiments across 11 benchmarks demonstrate that BiomedAP consistently surpasses baselines, achieving competitive few-shot accuracy and markedly enhanced robustness under prompt perturbations. Our code is available at: this https URL . Keywords: Vision-Language Models; Prompt Learning; Parameter-Efficient Fine-Tuning; Few-shot Learning\n\n---\n\n## 115. UAM: A Dual-Stream Perspective on Forgetting in VLA Training\n- **Authors**: Jianke Zhang , Yuanfei Luo , Yucheng Hu , Xiaoyu Chen , Yanjiang Guo , Ziyang Liu , Hongbin Xu , Tian Lan , Jianyu Chen\n- **URL**: [https://arxiv.org/abs/2605.15735](https://arxiv.org/abs/2605.15735)\n- **Abstract**:\n> Vision--language--action (VLA) models are typically built by fine-tuning a pretrained vision--language model (VLM) on action data. However, we show that this standard recipe systematically erodes the VLM's multimodal competence, a side effect we call the embodiment tax. But do VLAs have to forget? Inspired by the two-stream organization of biological vision, we trace this degradation to a structural bottleneck: current VLAs ask a single encoder to support both language-grounded semantics and control-relevant visual features, whereas biological vision separates recognition and visuomotor control into distinct pathways. Building on this view, we propose the Unified Action Model (UAM), which adds a parallel Dorsal Expert, an analog of the brain's dorsal pathway. To make the Dorsal Expert an effective second pathway and reduce the control-learning burden on the VLM, we initialize it from a pretrained generative model and train it with a mid-level reasoning objective that predicts visual dynamics. This design allows us to train the whole VLA end-to-end on action data alone: with no parameter freezing, no gradient stopping, and no auxiliary VL co-training, UAM retains over $95\\%$ of the underlying VLM's multimodal capability and at the same time achieves the highest average success rate among baselines on a variety of manipulation tasks that probe out-of-distribution generalization, including unseen objects, novel object--target compositions, and instruction variation. Together, these results suggest that semantic preservation in VLAs can emerge from architectural separation itself, rather than being enforced by frozen weights or auxiliary data replay, and that this preserved semantic capability can naturally transfer from VLMs to semantic generalization in actions.\n\n---\n\n## 116. Structure Abstraction and Generalization in a Hippocampal-Entorhinal Inspired World Model\n- **Authors**: Tianqiu Zhang , Muyang Lyu , Xiao Liu , Si Wu\n- **URL**: [https://arxiv.org/abs/2605.15733](https://arxiv.org/abs/2605.15733)\n- **Abstract**:\n> Humans abstract experiences into structured representations to facilitate pattern inference and knowledge transfer. While the hippocampal-entorhinal (HPC-MEC) circuit is known to represent both spatial and conceptual spaces, the mechanisms for concurrently extracting abstract structures from continuous, high-dimensional dynamics remain poorly understood. We propose a brain-inspired hierarchical model that simultaneously infers latent transitions and constructs a predictive visual world model. Our architecture employs an inverse model for structural extraction alongside an HPC-MEC coupling model that dissociates relational structures (MEC) from integrated episodic scenes (HPC). Using primitive transformation dynamics as a benchmark, we demonstrate the model's capacity for structural abstraction. By leveraging velocity-driven path integration, the framework enables robust prediction and structural reuse across diverse contexts, thereby achieving structural generalization. This work provides a novel computational framework for understanding how brain-inspired, self-supervised learning of world models facilitates the acquisition of reusable abstract knowledge.\n\n---\n\n## 117. DecomPose: Disentangling Cross-Category Optimization Contention for Category-Level 6D Object Pose Estimation\n- **Authors**: Yifan Gao , Lu Zou , Zhangjin Huang , Guoping Wang\n- **URL**: [https://arxiv.org/abs/2605.15728](https://arxiv.org/abs/2605.15728)\n- **Abstract**:\n> Category-level 6D object pose estimation is typically formulated as a multi-category joint learning problem with fully shared model parameters. However, pronounced geometric heterogeneity across categories entangles incompatible optimization signals in shared modules, resulting in gradient conflicts and negative transfer during training. To address this challenge, we first introduce gradient-based diagnostics to quantify module-level cross-category contention. Building on results of diagnostics, we propose DecomPose, a difficulty-aware decomposition framework that mitigates optimization contention via: (1) difficulty-aware gradient decoupling, which groups categories using a data-driven difficulty proxy and routes each instance to a group-specific correspondence branch to isolate incompatible updates; and (2) stability-driven asymmetric branching, which assigns higher-capacity branches to structurally simple categories as stable optimization anchors while constraining complex categories with lightweight branches to suppress noisy updates and alleviate negative transfer. Extensive experiments on REAL275, CAMERA25, and HouseCat6D demonstrate that DecomPose effectively reduces cross-category optimization contention and delivers superior pose estimation performance across multiple benchmarks.\n\n---\n\n## 118. DiLA: Disentangled Latent Action World Models\n- **Authors**: Tianqiu Zhang , Muyang Lyu , Yufan Zhang , Fang Fang , Si Wu\n- **URL**: [https://arxiv.org/abs/2605.15725](https://arxiv.org/abs/2605.15725)\n- **Abstract**:\n> Latent Action Models (LAMs) enable the learning of world models from unlabeled video by inferring abstract actions between consecutive frames. However, LAMs face a fundamental trade-off between action abstraction and generation fidelity. Existing methods typically circumvent this issue by using two-stage training with pre-trained world models or by limiting predictions to optical flow. In this paper, we introduce DiLA, a novel Disentangled Latent Action world model that aims to resolve this trade-off via content-structure disentanglement. Our key insight is that disentanglement and latent action learning are co-evolving: the predictive bottleneck inherent in latent action learning serves as a driving force for disentanglement, compelling the model to distill spatial layouts into the structure pathway while offloading visual details to a separate content pathway for generation. This synergy yields a continuous, semantically structured latent action space without compromising generative quality. DiLA achieves superior results in video generation quality, action transfer, visual planning, and manifold interpretability. These findings establish DiLA as a unified framework that simultaneously achieves high-level action abstraction and high-fidelity generation, advancing the frontier of self-supervised world model learning.\n\n---\n\n## 119. Bidirectional Fusion Guided by Cardiac Patterns for Semi-Supervised ECG Segmentation\n- **Authors**: Jeonghwa Lim , Minje Park , Sunghoon Joo\n- **URL**: [https://arxiv.org/abs/2605.15722](https://arxiv.org/abs/2605.15722)\n- **Abstract**:\n> Accurate delineation of electrocardiogram (ECG), the segmentation of meaningful waveform features, is crucial for cardiovascular diagnostics. However, the scarcity of annotated data poses a significant challenge for training deep learning models. Conventional semi-supervised semantic segmentation (SemiSeg) methods primarily focus on consistency from unlabeled data, underutilizing the information exchange possible between labeled and unlabeled sets. To address this, we introduce CardioMix, a framework built on a bidirectional CutMix strategy guided by cardiac patterns for ECG segmentation. This approach enriches the labeled set with realistic variations from unlabeled data while simultaneously applying stronger supervisory signals to the unlabeled set, as the cardiac pattern-guided mixing ensures all augmented samples remain physiologically meaningful. Our framework is designed as a plug-and-play module, demonstrating high compatibility with various SemiSeg algorithms. Extensive experiments on SemiSegECG, a public multi-dataset benchmark for ECG delineation, demonstrate that CardioMix consistently outperforms existing CutMix-based fusion strategies across diverse datasets and labeled ratios as a plug-and-play module compatible with various SemiSeg algorithms.\n\n---\n\n## 120. Position: Early-Stage Quality Assurance in Annotation Pipelines Is More Cost-Effective Than Late-Stage Validation\n- **Authors**: Sunil Kothari , Sumukha Sharma Thoppanahalli Chandramouli , Naman Khandelwal , Parth Kulshreshtha , Ashi Jain , Kriti Banka , Tanuja Chintada , Venkata Triveni , Gulipalli Praveen Kumar , Manish Mehta , Tao Liu\n- **URL**: [https://arxiv.org/abs/2605.15714](https://arxiv.org/abs/2605.15714)\n- **Abstract**:\n> This position paper argues that the machine learning community should prioritize early-stage quality assurance in annotation pipelines over the prevailing practice of late-stage validation. Data quality bottlenecks increasingly limit foundation model improvement, yet quality assurance research focuses almost exclusively on validation methods rather than validation timing. When validation occurs, not merely what methods are employed, fundamentally determines both error rates and annotation costs. This temporal neglect is puzzling given the well-established \"shift-left\" principle from software engineering, where empirical studies demonstrate 4--100x cost multipliers for defects detected in later stages (Boehm, 1981; Shull et al., 2002). Annotation pipelines exhibit analogous dynamics: errors caught before annotation begins cost a fraction of those discovered after review cycles complete. We propose a taxonomy of three QA trigger points, namely pre-annotation (T0), post-annotation (T1), and post-review (T2), that decompose annotation workflows into discrete validation opportunities. A parametric error-propagation model formalizes when timing affects final error rates versus only economics, making timing a measurable design variable rather than a configuration afterthought. A survey of 47 recent papers reveals that only 4% report when validation occurs, a striking gap given timing's demonstrated impact in adjacent fields. Without explicit attention to QA timing, the community risks optimizing validation methods while ignoring the structural variable that may matter most. Acting on this position requires three steps: researchers should report QA timing configurations alongside validation methods; annotation platforms should expose timing as a first-class parameter; and the community should run controlled experiments that measure stage-specific detection rates directly.\n\n---\n\n## 121. Learning Dynamic Pick-and-Place for a Legged Manipulator\n- **Authors**: Moonkyu Jung , Jiseong Lee , Zhengmao He , Donghoon Youm , Juhyeok Mun , HyeongJun Kim , Hyunsik Oh , Donghyuk Choi , Jungwoo Hur , Jie Song , Jemin Hwangbo\n- **URL**: [https://arxiv.org/abs/2605.15713](https://arxiv.org/abs/2605.15713)\n- **Abstract**:\n> Legged manipulators extend robotic capabilities beyond static manipulation by integrating agile locomotion with versatile arm control. However, achieving precise manipulation while maintaining coordinated locomotion remains a major challenge. This work presents a hierarchical reinforcement learning framework for dynamic pick-and-place tasks using a quadruped equipped with a 6-DOF robotic arm. The framework incorporates an explicit mass estimation module enabling adaptive whole-body control for objects with varying weights. In simulation, the system achieves an 86.05% success rate with payloads up to 2.3 kg. The approach is further validated through real-world experiments across six representative scenarios with controlled variations in object physical properties (size and mass) and task heights. Specifically, within a wide vertical workspace ranging from ground level to 1.1~m-high tabletops, the system demonstrates an average success rate of 73.3% for payloads up to 1.3 kg, with an average execution time of 4.06 s. Unlike prior works that handle lightweight objects and execute pick-and-place motions with slow, piecewise motions, the proposed framework exploits concurrent locomotion and manipulation for dynamic, continuous execution. These results demonstrate the potential of quadrupedal mobile manipulators for adaptive, whole-body pick-and-place with heavier payloads and extended workspaces.\n\n---\n\n## 122. Feedback World Model Enables Precise Guidance of Diffusion Policy\n- **Authors**: Tuo An , Jindou Jia , Gen Li , Jingliang Li , Chuhao Zhou , Pengfei Liu , Bofan Lyu , Jiaqi Bai , Xinying Guo , Geng Li , Jianfei Yang\n- **URL**: [https://arxiv.org/abs/2605.15705](https://arxiv.org/abs/2605.15705)\n- **Abstract**:\n> World models aim to improve robotic decision making by predicting the consequences of actions. However, in practice, their predictions often become unreliable once the robot encounters states outside the training distribution, limiting their effectiveness at deployment. We observe that execution itself provides a natural but underutilized signal: after each action, the robot directly observes the true next state, revealing the mismatch between predicted and actual outcomes. Building on this insight, we propose feedback world model, a new paradigm that closes the loop between prediction and observation at inference time. Instead of treating the world model as a static open-loop predictor, our method maintains a lightweight feedback state that is updated online to iteratively correct future predictions, compensating for model errors using real-time observations without additional training data or parameter updates. We show that this process can be interpreted as a latent-space observer and admits convergence guarantees under mild conditions. We further introduce action-aware guidance to better translate corrected predictions into control by emphasizing action-controllable components while suppressing irrelevant variations. Experiments on LIBERO-Plus, Robomimic, and real-world manipulation tasks demonstrate that our method substantially improves both prediction accuracy and policy performance under distribution shift. In particular, it reduces world model prediction error by up to 76.4% and improves out-of-distribution (OOD) success rate by 30%. These results show that incorporating real-time feedback at inference time provides a simple yet powerful alternative to static world modeling.\n\n---\n\n## 123. H-Mem: A Novel Memory Mechanism for Evolving and Retrieving Agent Memory via a Hybrid Structure\n- **Authors**: Jiawei Yu , Yixiang Fang , Xilin Liu , Yuchi Ma\n- **URL**: [https://arxiv.org/abs/2605.15701](https://arxiv.org/abs/2605.15701)\n- **Abstract**:\n> Memory data are ubiquitous in Large Language Model (LLM)-based agents (e.g., OpenClaw and Manus). A few recent works have attempted to exploit agents'memory for improving their performance on the question-answering (QA) task, but they lack a principled mechanism for effectively modeling how memory data evolves over time and retrieving memory data effectively, leading to poor performance in memory utilization. To fill this gap, we present H-Mem, a novel memory mechanism via a hybrid structure that can not only effectively model the evolution of agent memory over a long period of time, but also provide an efficient memory retrieval approach. Particularly, H-Mem builds a temporal and semantic tree structure that allows the short-term memory data to evolve progressively into long-term memory data, where the latter provides summarized information about the former, while simultaneously constructing a knowledge graph to capture the relationships between entities in memory. Moreover, it offers an effective memory retrieval approach by exploiting the hybrid structure of the tree and graph structures. Extensive experiments on three agent memory benchmarks show that H-Mem achieves state-of-the-art performance on the QA task.\n\n---\n\n## 124. $α$-TCAV: A Unified Framework for Testing with Concept Activation Vectors\n- **Authors**: Ekkehard Schnoor , Jawher Said , Malik Tiomoko , Wojciech Samek , Alexander Jung\n- **URL**: [https://arxiv.org/abs/2605.15688](https://arxiv.org/abs/2605.15688)\n- **Abstract**:\n> Concept Activation Vectors (CAVs) are a fundamental tool for concept-based explainability in deep learning, yet their practical utility is limited by statistical instability. We analyze the stochastic nature of CAVs and the Testing with CAVs (TCAV) method, deriving the distributions of major CAV classes including PatternCAV, FastCAV, and ridge regression-based CAVs. We then identify a fundamental flaw in the standard TCAV score: its reliance on a discontinuous indicator function induces non-decaying variance in critical regimes. To address this, we introduce $\\alpha$-TCAV, a generalized framework that replaces the indicator with a parameterized smooth function, yielding a unified probabilistic formulation that subsumes both TCAV and Multi-TCAV. We characterize the induced distributions of sensitivity scores and different TCAV variants, showing that established state-of-the-art choices lack theoretical justification. We provide principled guidance on tuning the parameter in $\\alpha$-TCAV -- either to imitate Multi-TCAV at substantially lower computational cost, or to obtain a calibrated Bayes-optimal probabilistic measure of a concept's influence. Finally, our analysis yields practical recommendations that challenge established routines: most notably, allocating the full sampling budget to a single CAV rather than splitting it across several.\n\n---\n\n## 125. ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models\n- **Authors**: Jiahui Guang , Yingjie Zhu , Cuiyun Gao , Haiyan Wang , Jing Li , Di Shao , Zhaoquan Gu\n- **URL**: [https://arxiv.org/abs/2605.15687](https://arxiv.org/abs/2605.15687)\n- **Abstract**:\n> Multimodal large language models (MLLMs) may memorize sensitive cross-modal information during pretraining, making machine unlearning (MU) crucial. Existing methods typically evaluate unlearning effectiveness based on output deviations, while overlooking the generation quality after unlearning. This can easily lead to hallucinated or rigid responses, thereby affecting the usability and safety of the unlearned model. To address this issue, we propose ASRU, a controllable multimodal unlearning framework that incorporates generation quality as a core evaluation objective. ASRU first induces initial refusal behavior through activation redirection, and then optimizes fine-grained refusal boundaries using a customized reward function, thereby achieving a better trade-off between target knowledge unlearning and model utility. Experiments on Qwen3-VL show that ASRU significantly improves unlearning effectiveness (+24.6%) on average and generation quality (5.8x) on average while effectively preserving model utility, using only a small amount of retained supervision data.\n\n---\n\n## 126. Interaction-Aware Influence Functions for Group Attribution\n- **Authors**: Jaeseung Heo , Kyeongheung Yun , Youngbin Choi , Sehyun Hwang , Jungseul Ok , Dongwoo Kim\n- **URL**: [https://arxiv.org/abs/2605.15675](https://arxiv.org/abs/2605.15675)\n- **Abstract**:\n> Influence functions approximate how removing a training example changes a quantity of interest, called the target function, such as a held-out loss. To estimate the influence of a group of examples, the standard practice is to sum the individual influences of its members. However, this sum does not capture how examples jointly affect the target: a pair of examples may be redundant or complementary, but the sum cannot distinguish these cases. We propose an interaction-aware influence function that characterizes how interactions between examples influence the target. By expanding the target to second order around the trained parameters, we obtain an estimator that augments the standard sum with a pairwise interaction term that captures the alignment between two examples' effects on the target. We empirically evaluate our estimator in two settings. First, on six dataset-model pairs spanning logistic regression, MLPs, and ResNet-9, our estimator tracks leave-group-out retraining substantially better than first-order influence across all settings. Second, when used as a greedy selection rule for instruction-tuning data on Llama-3.1-8B, it beats prior influence-based and representation-similarity baselines on five of seven downstream tasks, in a regime where standard influence-based selection underperforms random selection.\n\n---\n\n## 127. VLMs Trace Without Tracking: Diagnosing Failures in Visual Path Following\n- **Authors**: Hyesoo Hong , Minsoo Kim , Wonje Jeung , Sangyeon Yoon , Dongjae Jeon , Albert No\n- **URL**: [https://arxiv.org/abs/2605.15672](https://arxiv.org/abs/2605.15672)\n- **Abstract**:\n> Vision-language models (VLMs) achieve strong performance on multimodal benchmarks, but may still lack robust control over basic visual operations. We study \\textit{line tracing}, where a model must follow a selected visual path through successive local continuations. To isolate this ability, we design controlled tracing tasks that introduce nearby competitors while reducing semantic and topological ambiguity such as crossings and overlaps. Across these tasks, even state-of-the-art VLMs frequently lose the target path and switch to nearby alternatives, especially when those alternatives look locally similar to the target. Behavioral interventions and internal analyses indicate that these failures arise from local competition: nearby similar distractors pull the model away from the true continuation. Standard remedies do not remove this bottleneck: model-size scaling provides only limited gains, reasoning partially compensates through costly substitute strategies, and explicit tracing instructions fail to recover stable path following. Finally, tests on tangled-cable scenes and metro maps with richer visual complexity show that the same path-switching failure persists beyond our controlled settings.\n\n---\n\n## 128. VAGS: Velocity Adaptive Guidance Scale for Image Editing and Generation\n- **Authors**: Yan Luo , Ahmadou Aidara , Jingyi Lu , Jeremy Moebel , Kai Han , Mengyu Wang\n- **URL**: [https://arxiv.org/abs/2605.15661](https://arxiv.org/abs/2605.15661)\n- **Abstract**:\n> Classifier-free guidance (CFG) is the primary control over how strongly text semantics move a flow-based sampler, yet standard practice holds its scale fixed across the entire ODE trajectory. This is a fundamental mismatch: early steps are noise-dominated and carry weak semantic signal, while late steps commit image structure and demand stronger directional commitment; more critically, the value of any guidance strength depends on whether the guided velocity is consistent with the model's current dynamics or working against them. We propose \\textit{Velocity-Adaptive Guidance Scale} (VAGS), a training-free replacement that multiplies the nominal scale by a bounded factor combining a temporal signal-level term with the cosine similarity between task-relevant velocity fields. For inversion-free editing, VAGS measures the alignment between source- and target-guided velocities, so edit strength at each step reflects local compatibility between preservation and transformation. For generation, VAGS-Gen uses the alignment between unconditional and conditional velocities as the analogous signal. Neither variant requires fine-tuning, auxiliary networks, or extra forward passes, and fixed CFG is recovered as a special case. On PIE-Bench and DIV2K for editing, and COCO17, CUB-200, and Flickr30K for generation, VAGS consistently improves structural fidelity and generation quality over fixed CFG and recent training-free guidance variants. The code is publicly available at this https URL .\n\n---\n\n## 129. TFZ-Tree: An Ultra-Lightweight Waveform Classification Framework for Resource-Constrained Devices\n- **Authors**: Hao Wang , Kuang Zhang , Yonggang Chi , Tianqi Zhao , Yanbo Fu , Jiaxing Guo\n- **URL**: [https://arxiv.org/abs/2605.15656](https://arxiv.org/abs/2605.15656)\n- **Abstract**:\n> Under the trend of multi-waveform coexistence in 6G IoT, intelligent receivers must first identify physical-layer waveform types before performing correct demodulation and resource scheduling. However, existing signal identification research largely focuses on symbol-level modulation classification. Research directly targeting physical-layer waveform types (e.g., OFDM, OTFS, LoRa) is not only extremely scarce but also heavily reliant on deep neural networks and complex time-frequency transforms, making deployment on resource-constrained terminals difficult. Symbol modulation classification methods themselves cannot circumvent the prerequisite of ``waveform identification first.'' To address this dual gap, we propose an ultra-lightweight waveform classification framework based on time-frequency multidimensional features with a cooperative Z-test tree (ZTree). The framework employs low-complexity time-domain feature extraction, and the classification backend adopts a ZTree optimized by Z-statistical testing, which uses hypothesis testing confidence to automatically control decision tree splitting and size, ensuring efficient execution on resource-limited processors. Tested on ten 6G candidate waveforms including OFDM, OTFS, DSSS, LoRa, and NB-IoT, the method achieves 99.5\\% average accuracy under AWGN and 87.4\\% under TDL-C multipath channels, with main confusion between OTFS and LoRa. Implemented in C on an x86 platform, single inference latency is under 4~ms. To the best of our knowledge, this is the first work achieving real-time recognition of ten IoT waveform types. Future work will target deployment acceleration on embedded MCUs. Code and dataset are open-sourced at: this https URL .\n\n---\n\n## 130. Bridging Silicon and the Hippocampus: Algebro-Deterministic Memory \"VaCoAl\" as a Substrate for Vector-HaSH and TEM\n- **Authors**: Hiroyuki Chuma , Kanji Otsuka , Yoichi Sato\n- **URL**: [https://arxiv.org/abs/2605.15652](https://arxiv.org/abs/2605.15652)\n- **Abstract**:\n> Vector-HaSH and the Tolman-Eichenbaum Machine (TEM) propose that the hippocampal-entorhinal circuit factorizes content from a prestructured grid-cell scaffold and supports compositional memory via ripple-mediated replay. Human iEEG shows that hippocampal sharp-wave ripples (SWRs) gate episodic recall, ripple-locked cortical reactivation recapitulates encoding-time patterns, and multi-hop replay fidelity decays multiplicatively along sequence length. These literatures have advanced in parallel without a shared algebraic object. We show that VaCoAl, an algebro-deterministic hyperdimensional memory architecture built from Galois-field LFSRs, supplies that object. Specifically, deterministic Galois-field diffusion provides a substrate-level alternative to Vector-HaSH's random scaffold-to-hippocampus projection that satisfies the same quasi-orthogonality requirement, with matched second-moment statistics, stronger avalanche behavior, and bit-exact reproducibility. The path-integral Confidence Ratio $CR_2$, the product of per-step $CR_1$ values along an $n$-hop chain, is the natural functional form for multi-hop replay-fidelity decay under conditional independence of per-step reactivation, providing the first algebraically tractable model of reported multiplicative decay. STDP-like path selection in VaCoAl follows from architectural demands -- similarity preservation, compositional reversibility, and bounded-frontier search -- that also constrain hippocampal computation. We further argue that VaCoAl operating regimes share architectural commitments with the EC--CA3 and EC--DG--CA3 pathways, motivating an energy-capacity-plasticity reading of why both are conserved across $>$520 Myr of evolution and primate dentate-gyrus elaboration. We prove formal correspondences, derive testable iEEG predictions, and bridge computational neuroscience and hyperdimensional engineering.\n\n---\n\n## 131. Sharp Spectral Thresholds for Logit Fixed Points\n- **Authors**: Tongxi Wang\n- **URL**: [https://arxiv.org/abs/2605.15651](https://arxiv.org/abs/2605.15651)\n- **Abstract**:\n> Softmax feedback systems are a common mathematical core of entropy-regularized reinforcement learning, logit game dynamics, population choice, and mean-field variational updates. Their central stability question is simple: when does a self-reinforcing softmax system produce a unique and globally predictable outcome? Classical theory gives a conservative answer. By treating softmax as a unit-scale response, it certifies stability only in a strongly randomized regime. We prove that the classical approach misses an entire stable regime and does not identify the point at which the qualitative change truly occurs. For finite-dimensional affine logit systems, the sharp dimension-free Euclidean threshold is $$\\beta\\|\\Pi W\\Pi\\|_{\\mathcal T\\to\\mathcal T}<2,$$ rather than the previously used condition, which certifies stability only while the softmax system remains safely over-regularized. Our theorem fills the previously missing pre-bifurcation regime, extending stability guarantees for affine softmax feedback systems to reward-responsive yet globally predictable systems. It enlarges the certified stability boundary for these systems and identifies where the model genuinely undergoes a phase transition.\n\n---\n\n## 132. Latent Video Prediction Learns Better World Models\n- **Authors**: Ali J Alrasheed , Aryan Yazdan Parast , Basim Azam , James Bailey , Naveed Akhtar\n- **URL**: [https://arxiv.org/abs/2605.15618](https://arxiv.org/abs/2605.15618)\n- **Abstract**:\n> Self-supervised video models are increasingly framed as world models, yet their evaluation remains largely confined to a single top-1 accuracy score on clean benchmarks. This leaves a major gap in comprehending their potential as world models. We present the first systematic study addressing this gap, analyzing four matched-capacity frontier video foundation models, V-JEPA 2.1, V-JEPA 2, VideoPrism, and VideoMAEv2, across five robustness axes relevant to their deployment as video world models: feature discriminability, corruption robustness, fine-grained discrimination, occlusion robustness, and sensitivity to temporal direction. Our evaluations establish that across all five axes, latent-prediction models form a distinct and consistent profile. They degrade more gracefully under pixel corruption, preserve usable class structure rather than mere geometric stability under occlusion, capture fine-grained physical contact cues without reconstructing pixels, and uniquely encode the arrow of time. These advantages can even survive task adaptation: a frozen V-JEPA 2 backbone with a lightweight attentive probe outperforms a fully fine-tuned VideoMAE and a supervised TimeSformer on corruption and occlusion robustness. Our extensive results offer concrete new evidence in favor of latent prediction for robust world modeling.\n\n---\n\n## 133. A Few GPUs, A Whole Lotta Scale: Faithful LLM Training Emulation with PrismLLM\n- **Authors**: Shaoke Xi , ChonLam Lao , Boyi Jia , Jiaqi Gao , Zhipeng Zhang , Jiamin Cao , Brian Sutioso , Erci Xu , Minlan Yu , Kui Ren , Yong Li , Zhengping Qian , Ennan Zhai , Jingren Zhou\n- **URL**: [https://arxiv.org/abs/2605.15617](https://arxiv.org/abs/2605.15617)\n- **Abstract**:\n> Large language model (LLM) training today runs on clusters spanning thousands of GPUs. While this scale enables rapid model advances, developing, debugging, and performance-tuning the training framework inevitably becomes complex and costly. This is because engineers often need to reproduce production behaviors to diagnose failures or evaluate optimizations, thereby demanding frequent and even exclusive access to production-scale clusters -- which becomes increasingly hard given that the majority of GPUs are already committed to production workloads. Simulation relies on complex performance models that are difficult to maintain, and downscaled experiments often fail to capture scale-dependent behaviors. We present PrismLLM to decouple large-scale execution from the need to access large clusters, enabling engineers to run and observe ranks of interest under faithful large-scale behavior using only a few GPUs. PrismLLM constructs a high-fidelity execution graph via a slicing-based approach that captures computation, communication, and dependencies of the target scale. Then, PrismLLM performs hybrid emulation where selected ranks execute the original program while the remaining ranks are replayed as virtual participants. Experiments on large-scale LLM training workloads show that PrismLLM accurately reproduces performance and memory behavior, achieving only 0.58\\% average error in iteration time and less than 0.01\\% error in peak GPU memory usage. PrismLLM can emulate clusters of up to 8192 GPUs using fewer than 1\\% of the physical GPUs required by the original deployment.\n\n---\n\n## 134. Offline Reinforcement Learning with Universal Horizon Models\n- **Authors**: Hojun Chung , Junseo Lee , Songhwai Oh\n- **URL**: [https://arxiv.org/abs/2605.15603](https://arxiv.org/abs/2605.15603)\n- **Abstract**:\n> Model-based reinforcement learning (RL) offers a compelling approach to offline RL by enabling value learning on imagined on-policy trajectories. However, it often suffers from compounding errors due to repeated model inference on self-generated states. While geometric horizon models (GHM) alleviate this issue through direct prediction over a discounted infinite-horizon future, they remain challenged in accurately modeling distant future states. To this end, we introduce universal horizon models (UHM), a generalization of GHM that directly predicts future states under arbitrary horizons. Leveraging this flexibility, we propose a scalable value learning method that employs a winsorized horizon distribution to stabilize training by capping excessively large horizons. Experimental results on 100 challenging OGBench tasks demonstrate that the proposed method outperforms competitive baselines, particularly on tasks with highly suboptimal datasets and those requiring long-horizon reasoning. Project page: this https URL\n\n---\n\n## 135. Pretraining Objective Matters in Extreme Low-Data FGVC: A Backbone-Controlled Study\n- **Authors**: Alexander Hackett , Srikanth Thudumu , Ginny Fisher , Mahule Roy , Aisha Sartaj , Jason Fisher\n- **URL**: [https://arxiv.org/abs/2605.15599](https://arxiv.org/abs/2605.15599)\n- **Abstract**:\n> Extreme low-data fine-grained classification is common in expert domains where labeling is expensive, yet practitioners still need principled guidance for selecting pretrained encoders. We study emerald inclusion grading with a custom dataset of labeled images across three classes and ask: under matched backbone capacity, how does pretraining objective affect downstream representation quality? We compare four frozen ViT-B/16 encoders trained with supervised classification, contrastive learning (SigLIP2), masked reconstruction (MAE), and self-distillation (DINOv3), and evaluate them with leave-one-out cross-validation using linear and nonlinear probes. To control statistical noise in the low-N regime, we use permutation testing (N=1000) on macro one-vs-rest AUC. Supervised and contrastive encoders provide the strongest linear separability (logistic AUC: 0.768 and 0.735; SVM AUC: 0.739 and 0.697), while MAE improves under nonlinear probes (XGBoost AUC: 0.713). We find that DINOv3 underperforms across probe families in this domain. These results support a practical recommendation for extreme low-data FGVC: prioritize margin-enforcing pretraining objectives when data scarcity restricts probing to linear decision rules, and consider reconstruction-style encoders when nonlinear classifiers are feasible given dataset constraints.\n\n---\n\n## 136. Embracing Biased Transition Matrices for Complementary-Label Learning with Many Classes\n- **Authors**: Tan-Ha Mai , Chao-Kai Chiang , Han-Hwa Shih , Gang Niu , Masashi Sugiyama , Hsuan-Tien Lin\n- **URL**: [https://arxiv.org/abs/2605.15586](https://arxiv.org/abs/2605.15586)\n- **Abstract**:\n> Complementary-label learning (CLL) is a weakly supervised paradigm where instances are labeled with classes they do not belong to. Despite a decade of research, CLL methods remain competitive mainly on 10-class classification, with scaling to large label spaces continuing to be an enduring bottleneck. This limitation stems from the common assumption of uniform label generation in traditional methods, which fatally dilutes the learning signal in many-class settings. In this paper, we demonstrate that this long-standing barrier can be overcome by deliberately designing a biased (non-uniform) generation process that restricts complementary labels to a subset of classes. This finding motivates us to propose Bias-Induced Constrained Labeling (BICL), a principled framework spanning data collection to training that leverages this bias. BICL enables effective learning on CIFAR-100 and TinyImageNet-200, achieving more than sevenfold accuracy improvements over traditional methods. Our findings establish a new trajectory for making CLL feasible for many classes in real-world applications.\n\n---\n\n## 137. Detecting Privilege Escalation in Polyglot Microservices via Agentic Program Analysis\n- **Authors**: Penghui Li , Hong Yau Chong , Yinzhi Cao , Junfeng Yang\n- **URL**: [https://arxiv.org/abs/2605.15569](https://arxiv.org/abs/2605.15569)\n- **Abstract**:\n> Microservices are widely adopted in modern cloud systems due to their scalability and fault tolerance. However, microservice architectures introduce significant complexity in privilege and permission control, creating risks of privilege escalation where attackers can gain unauthorized access to resources or operations. Detecting such vulnerabilities is challenging due to complex cross-service interactions, polyglot codebases, and diverse privileged operations and permission checks. We present Neo, an agentic program analysis framework that combines large language models (LLMs) with classic program analysis to address these challenges. Neo leverages an LLM-based agent that dynamically generates analysis plans, adapts code search strategies, and validates semantics. We develop code search primitives that enable Neo to perform scalable and flexible code exploration across services and languages. We evaluated Neo on 25 open-source microservice applications spanning 7 programming languages and 6.2 million lines of code. Neo uncovered 24 zero-day privilege escalation vulnerabilities and achieved 81.0% precision and 85.0% recall on a ground-truth dataset. Compared to existing program analysis and agentic solutions, Neo demonstrated significant improvements in both detection accuracy and scalability. We further showcased Neo's extensibility by applying it to other application domains and vulnerability types, uncovering 18 additional zero-day vulnerabilities.\n\n---\n\n## 138. AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs\n- **Authors**: Haizhong Zheng , Yizhuo Di , Jiahui Wang , Shuowei Jin , Xueshen Liu , Yongji Wu , Z. Morley Mao , Ion Stoica , Jiawei Zhao , Beidi Chen\n- **URL**: [https://arxiv.org/abs/2605.15565](https://arxiv.org/abs/2605.15565)\n- **Abstract**:\n> Reinforcement learning (RL) is increasingly used to improve the reasoning, coding, and tool-use capabilities of large language models, but agentic RL remains prohibitively expensive. Scaling RL to agentic LLMs requires supporting complex workloads, including multi-policy collaborative training, while efficiently using elastic, heterogeneous, and cross-region compute resources. Existing LLM RL systems support some of these capabilities, but each new extension often requires dedicated system engineering. This burden arises from trainer-centered control architectures and the lack of principled abstractions for RL system components. To address these limitations, we propose AstraFlow, a dataflow-oriented RL system that replaces conventional trainer-centered control with principled component abstractions. In AstraFlow, rollout services, dataflow management, and training are decoupled into autonomous components, enabling the system to natively support complex multi-policy agentic RL workloads and efficiently exploit diverse compute resources. We evaluate AstraFlow across math, code, search, and AgentBench workloads, showing that the same system supports multi-policy training, elastic scaling, heterogeneous cross-region execution, and composable data algorithms without system-level code changes. In multi-policy collaborative training, AstraFlow achieves comparable or better accuracy than existing RL systems while speeding up training time by 2.7x.\n\n---\n\n## 139. Operator-Controlled 6G: From Connectivity Infrastructure to Guaranteed Digital Services\n- **Authors**: David Soldani\n- **URL**: [https://arxiv.org/abs/2605.15553](https://arxiv.org/abs/2605.15553)\n- **Abstract**:\n> Sixth-generation mobile networks (6G) are approaching a structural inflection point. Five generations of vendor-led architectures have left operators procuring and operating networks they do not own, on platforms they cannot modify, with AI layers they cannot audit. This paper argues that 6G must reverse this trajectory by reordering operator priorities: Control First, Customer First, Business First, Operations First, and Technology Last. Technology should serve operator control, customer outcomes, monetizable guarantees, and software-driven operations, not dictate this http URL contributions operationalize this thesis. The 6G Control Compact defines a three-layer ownership taxonomy--own, federate, and consume--that allocates architectural sovereignty according to strategic value. The Guarantee Economy defines a five-tier, outcome-priced commercial model that converts operator control into enforceable service-level objectives. The framework is grounded in operational evidence from Rakuten Mobile, the world's first national-scale, fully cloud-native, fully Open RAN deployment, which reached full-year EBITDA profitability in FY2025. It is aligned with the ITU-R IMT-2030 framework, 3GPP 6G use cases and service requirements, NGMN recommendations, ETSI standards, O-RAN Alliance and AI-RAN Alliance specifications, IOWN Global Forum sustainability metrics, Linux Foundation initiatives, and leading industry and academic programs. A three-phase roadmap covering 2025-2027, 2027-2029, and 2029-2032 and beyond, together with seven stakeholder-specific calls to action, translates the architecture into industry commitments. The central claim is that Rakuten Mobile's deployment demonstrates the feasibility of operator-controlled 6G. Decisions made during 2026-2028 will determine whether 6G becomes a platform for guaranteed digital services or another vendor-dependent infrastructure cycle.\n\n---\n\n## 140. CTF4Nuclear: Common Task Framework for Nuclear Fission and Fusion Models\n- **Authors**: Stefano Riva , Carolina Introini , Antonio Cammi , Dean Price , Alexey Yermakov , Yue Zhao , Philippe M. Wyder , Judah Goldfeder , Jan Williams , Amy Sara Rude , Matteo Tomasetto , Joe Germany , Joseph Bakarji , Georg Maierhofer , Miles Cranmer , J. Nathan Kutz\n- **URL**: [https://arxiv.org/abs/2605.15549](https://arxiv.org/abs/2605.15549)\n- **Abstract**:\n> The demand for clean energy is ever increasing, with new nuclear technologies presenting a complementary solution to renewable energies. However, designing and operating these systems is exceptionally difficult, given the complexity of the physical phenomena that interact to form the system dynamics. While high-fidelity simulations help to understand the non-linear, multi-physics interactions within a reactor, they are computationally expensive and rarely suitable for real-time applications. Furthermore, model-based approaches are inherently sensitive to simplifying assumptions required to derive their governing equations and parameters, leading to inevitable discrepancies with real-world measurements. In contrast, Machine Learning (ML) methods have the potential to generate reliable surrogate models which may be able to quickly predict the system's behaviour. However, the number of data-driven methods that can potentially be used for this task is large and diverse. In a safety-critical setting such as nuclear engineering, a fair comparison of different ML methods, and a clear understanding of their advantages and limitations, is of paramount importance. To address this, we introduce a Common Task Framework (CTF) for ML in nuclear engineering, building upon previous efforts in dynamical systems and seismology. This CTF considers a curated set of datasets from different nuclear and nuclear-adjacent systems. The CTF evaluates the performance of a method on 12 established metrics, alongside a new paradigm focused on system monitoring from sparse measurements only. We illustrate the framework by benchmarking standard ML baselines against these datasets, revealing current method limitations. Our vision is to replace ad hoc comparisons with standardized evaluations on hidden test sets, raising the bar for rigour and reproducibility in scientific ML for the nuclear industry.\n\n---\n\n## 141. Domain-Independent Game Abstraction using Word Embedding Techniques\n- **Authors**: Juho Kim , Tuomas Sandholm\n- **URL**: [https://arxiv.org/abs/2605.15543](https://arxiv.org/abs/2605.15543)\n- **Abstract**:\n> Many games of interest in the real world are often intractably large, thereby necessitating the use of game abstraction to shrink them in size, typically by many magnitudes. Over the last two decades, there have been significant advances in game abstraction; however, the domain-specific nature (usually poker) of much of the prior work prevents those techniques from being easily generalized to other settings without extensively analyzing the game at hand. In this paper, we propose a domain-independent approach to game abstraction, which applies word embedding techniques from the field of natural language processing. Treating each action as a word and gameplay data as a corpus, word vectors can be trained to represent each action as a real-valued vector, which can then be clustered to facilitate game abstraction. We also explore the use of foundational embedding models and show that action embeddings obtained this way can capture a surprising amount of information about the underlying game. Experimental results demonstrate that our proposed game abstraction technique is effective, although it does not outperform specialized algorithms tailored to specific games.\n\n---\n\n## 142. SkiP: When to Skip and When to Refine for Efficient Robot Manipulation\n- **Authors**: Mingtong Dai , Guanqi Peng , Yongjie Bai , Feng Yan , Chunjie Chen , Lingbo Liu , Liang Lin , Xinyu Wu\n- **URL**: [https://arxiv.org/abs/2605.15536](https://arxiv.org/abs/2605.15536)\n- **Abstract**:\n> Previous imitation learning policies predict future actions at every control step, whether in smooth motion phases or precise, contact-rich operation phases. This uniform treatment is wasteful: most steps in a manipulation trajectory traverse free space and carry little task-relevant information, while a small fraction of \\emph{key} steps around contacts, grasps, and alignment demand dense, high-resolution prediction. We propose a novel \\emph{action relabeling} mechanism: at each timestep in a skip segment, we replace the behavior cloning target with the action at the entrance of the next key segment, enabling the policy to leap over redundant steps in a single decision. The resulting \\textbf{Skip Policy (SkiP)} dynamically leaps over skip segments and intensively refines actions in key segments, within a single unified network requiring no learned skip planner or hierarchical structure. To automatically partition demonstrations into key and skip segments without manual annotation, we introduce \\emph{Motion Spectrum Keying} (MSK), a fast, task-agnostic procedure that detects local motion complexity from action signals. Extensive experiments across 72 simulated manipulation tasks and three real-robot tasks show that SkiP reduces executed steps by $15$--$40\\%$ while matching or improving success rates across various policy backbones. Project page: \\texttt{ this https URL }.\n\n---\n\n## 143. Tuning-free Instruction-based Video Editing Via Structural Noise Initialization and Guidance\n- **Authors**: Song Wu , Xinyu Chen , Qian Wang , Liang Li , Zili Yi , Junlan Feng\n- **URL**: [https://arxiv.org/abs/2605.15533](https://arxiv.org/abs/2605.15533)\n- **Abstract**:\n> Video editing poses a significant challenge. While a series of tuning-free methods circumvent the need for extensive data collection and model training, they often underutilize the rich information embedded within noisy latent, leading to unsatisfactory results. To address this, we propose a \\textit{tuning-free, instruction-based} video editing framework. We approach video editing from the perspective of noisy latent: we design a Structural Noise Initialization Strategy (SNIS) to secure a superior editing starting point by assigning higher noise levels to edited regions (to facilitate content change) and lower noise levels to unedited regions (to maintain content consistency). We introduce a Noise Guidance Mechanism (NGM), which leverages the video prior in the generative model and effectively integrates rich information within the noisy latent to guide the denoising process, thereby preserving unedited content and overall visual coherence. Experiments show that our proposed method achieves better visual quality and state-of-the-art performance.\n\n---\n\n## 144. DeltaPrompts: Escaping the Zero-Delta Trap in Multimodal Distillation\n- **Authors**: Jaehun Jung , Hyunwoo Kim , Brandon Cui , Ximing Lu , David Acuna , Prithviraj Ammanabrolu , Yejin Choi\n- **URL**: [https://arxiv.org/abs/2605.15532](https://arxiv.org/abs/2605.15532)\n- **Abstract**:\n> Distillation enables compact Vision-Language Models (VLMs) to obtain strong reasoning capabilities, yet the prompts driving this process are typically chosen via simple heuristics or aggregated from off-the-shelf datasets. We reveal a critical inefficiency in this approach: up to 69% of the prompts in standard chart / document reasoning datasets are effectively zero-delta, meaning the teacher and student already induce the exact same answer distribution. Training on these prompts provides minimal learning signal, causing student improvement to rapidly saturate regardless of data scale. To escape the zero-delta trap, we return to first principles: distillation fundamentally minimizes distributional divergence, and thus a prompt is valuable only if it exposes a functional capability gap between the teacher and student. We quantify this gap through answer divergence ($\\Delta$), demonstrating that non-zero divergence is critical for effective scaling. Building on this insight, we propose a staged synthesis pipeline that repurposes existing datasets as seeds, actively targeting student failure modes to produce better prompts. The result is DeltaPrompts, a diverse dataset of 200k synthetic, high-divergence reasoning problems. We evaluate DeltaPrompts across three distinct settings: on-policy distillation with the target teacher-student pair, transfer to a novel model family without regenerating the data, and off-policy fine-tuning of a non-reasoning model. Across all scenarios, DeltaPrompts drives substantial gains, yielding up to 15% relative improvement even on top of a highly-optimized reasoning model (e.g., Qwen3-VL-8B-Thinking) -- averaged over 10 benchmarks spanning chart, document and perception-centric reasoning.\n\n---\n\n## 145. Process Rewards with Learned Reliability\n- **Authors**: Jinyuan Li , Langlin Huang , Chengsong Huang , Shaoyang Xu , Donghong Cai , Yuyi Yang , Wenxuan Zhang , Jiaxin Huang\n- **URL**: [https://arxiv.org/abs/2605.15529](https://arxiv.org/abs/2605.15529)\n- **Abstract**:\n> Process Reward Models (PRMs) provide step-level feedback for reasoning, but current PRMs usually output only a single reward score for each step. Downstream methods must therefore treat imperfect step-level reward predictions as reliable decision signals, with no indication of when these predictions should be trusted. We propose BetaPRM, a distributional PRM that predicts both a step-level success probability and the reliability of that prediction. Given step-success supervision from Monte Carlo continuations, BetaPRM learns a Beta belief that explains the observed number of successful continuations through a Beta-Binomial likelihood, rather than regressing to the finite-sample success ratio as a point target. This learned reliability signal indicates when a step reward should be trusted, enabling downstream applications to distinguish reliable rewards from uncertain ones. As one application, we introduce Adaptive Computation Allocation (ACA) for PRM-guided Best-of-N reasoning. ACA uses the learned reliability signal to stop when a high-reward solution is reliable and to spend additional computation on uncertain candidate prefixes. Experiments across four backbones and four reasoning benchmarks show that BetaPRM improves PRM-guided Best-of-N selection while preserving standard step-level error detection. Built on this signal, ACA improves the accuracy--token tradeoff over fixed-budget Best-of-16, reducing token usage by up to 33.57% while improving final-answer accuracy.\n\n---\n\n## 146. Neural Point-Forms\n- **Authors**: Bruno Trentini , Jacob Hume , Vincenzo Antonio Isoldi , Philipp Misof , Ekaterina S. Ivshina , Kelly Maggs\n- **URL**: [https://arxiv.org/abs/2605.15524](https://arxiv.org/abs/2605.15524)\n- **Abstract**:\n> Point cloud learning often rests on the premise that observed samples are noisy traces of an underlying geometric object, such as a manifold embedded in a high-dimensional feature space. Yet much of this geometry is not captured directly by coordinates, pairwise distances, or learned graph neighborhoods alone. In the smooth setting, differential forms are devices to encode higher order tangency information. In this work, we introduce a new family of principled learnable geometric features for point clouds called neural point-forms (NPFs). In the absence of a natural tangency structure, we instead use Laplacian-based techniques from Diffusion Geometry to build a discrete model for comparing differential forms on point clouds via inner products. In the continuum, submanifolds of a shared ambient feature space are represented as comparison matrices, whose entries describe how pairs of feature forms interact with extrinsic tangency information. We make this intuition precise by proving the long-run consistency of comparison matrices under standard sampling, bandwidth, density, and manifold-hypothesis assumptions. This yields a compact, efficient and permutation-invariant neural layer whose output is a learned form-comparison matrix. Across synthetic and biologically relevant experiments, we show that NPFs provide a competitive, and interpretable representation, with the strongest benefits appearing when labels depend on sampling density, manifold-like structure, or response-relevant population geometry.\n\n---\n\n## 147. On the Fragility of Data Attribution When Learning Is Distributed\n- **Authors**: Xian Gao , Bo Hui , Min-Te Sun , Wei-Shinn Ku\n- **URL**: [https://arxiv.org/abs/2605.15520](https://arxiv.org/abs/2605.15520)\n- **Abstract**:\n> Data attribution has become an important component of pricing, auditing, and governance in machine learning pipelines, yet most attribution methods implicitly assume that attribution values faithfully reflect participants' contributions. We show that this assumption can fail: a single participant in a standard distributed training workflow can substantially inflate its measured attribution value while preserving global utility. Our attribution-first attack uses latent optimization to inject small synthetic batches that preserve utility while exploiting non-IID label coverage and evaluator sensitivities. Across datasets, models, and multiple marginal-utility evaluators, the attack consistently increases the adversary's attribution value and reshapes the relative attribution structure among benign clients without degrading accuracy or triggering geometry-based defenses. These results show that attribution itself forms a new attack surface and motivate the development of attribution-robust and incentive-compatible scoring mechanisms.\n\n---\n\n## 148. DiffVAS: Diffusion-Guided Visual Active Search in Partially Observable Environments\n- **Authors**: Anindya Sarkar , Srikumar Sastry , Aleksis Pirinen , Nathan Jacobs , Yevgeniy Vorobeychik\n- **URL**: [https://arxiv.org/abs/2605.15519](https://arxiv.org/abs/2605.15519)\n- **Abstract**:\n> Visual active search (VAS) has been introduced as a modeling framework that leverages visual cues to direct aerial (e.g., UAV-based) exploration and pinpoint areas of interest within extensive geospatial regions. Potential applications of VAS include detecting hotspots for rare wildlife poaching, aiding search-and-rescue missions, and uncovering illegal trafficking of weapons, among other uses. Previous VAS approaches assume that the entire search space is known upfront, which is often unrealistic due to constraints such as a restricted field of view and high acquisition costs, and they typically learn policies tailored to specific target objects, which limits their ability to search for multiple target categories simultaneously. In this work, we propose DiffVAS, a target-conditioned policy that searches for diverse objects simultaneously according to task requirements in partially observable environments, which advances the deployment of visual active search policies in real-world applications. DiffVAS leverages a diffusion model to reconstruct the entire geospatial area from sequentially observed partial glimpses, which enables a target-conditioned reinforcement learning-based planning module to effectively reason and guide subsequent search steps. Extensive experiments demonstrate that DiffVAS excels in searching diverse objects in partially observable environments, significantly surpassing state-of-the-art methods on several datasets.\n\n---\n\n## 149. RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably\n- **Authors**: Yufeng Du , Phillip Harris , Minyang Tian , Eliu A Huerta , Srikanth Ronanki , Subendhu Rongali , Aram Galstyan , Hao Peng\n- **URL**: [https://arxiv.org/abs/2605.15514](https://arxiv.org/abs/2605.15514)\n- **Abstract**:\n> We identify intrinsic limitations of Rotary Positional Embeddings (RoPE) in Transformer-based long-context language models. Our theoretical analysis abstracts away from the specific content of the context and depends only on its length. We prove that as context length increases, RoPE-based attention becomes unpredictable and loses two properties that are central to its effectiveness. First, it loses its locality bias: RoPE is no more likely to favor nearer positions than substantially farther ones. Second, it loses consistency in token relevance: a key vector that receives a higher attention score than an alternative at one position may receive a lower score at another. In both cases, the probability of failure approaches 0.5, no better than random guessing. We further prove that the attention score can remain unchanged when a key token is moved to a different position, or even replaced by a different token, indicating a failure to distinguish positions or tokens. Adjusting the RoPE base trades off distinguishing positions against distinguishing tokens but cannot preserve both at the same time. Increasing the RoPE base hyperparameter, a common practice in today's long-context models, helps distinguish different tokens, but inevitably sacrifices the ability to distinguish positions. Our empirical analysis shows that multi-head, multi-layer architectures are insufficient to overcome these limitations. Our findings suggest that fundamentally new mechanisms for encoding position and token order may be needed in future Transformer long-context language models.\n\n---\n\n## 150. PrismQuant: Rate-Distortion-Optimal Vector Quantization for Gaussian-Mixture Sources\n- **Authors**: Bumsu Park , Chanho Park , Youngmok Park , Namyoon Lee\n- **URL**: [https://arxiv.org/abs/2605.15507](https://arxiv.org/abs/2605.15507)\n- **Abstract**:\n> For a Gaussian source under mean-squared error (MSE), classical transform coding is rate--distortion (RD) optimal: the Karhunen--Loeve transform (KLT) diagonalizes the covariance, reverse waterfilling allocates the bits, and scalar quantization closes the loop. This elegant story breaks down for multimodal sources, where no single covariance can capture heterogeneous local geometries, and the RD function loses its closed form. We revisit this problem through Gaussian-mixture sources and develop a constructive RD theory for them. Our key finding is that the mixture structure incurs only a component label cost. Conditioned on the active mixture component, each branch is Gaussian; the challenge is allocating bits across heterogeneous branches. We prove that the genie-aided conditional RD function is governed by a single global reverse-waterfilling level shared across all components and eigenmodes. Building on this result, we introduce PrismQuant, which transmits the component label losslessly and encodes the residual using the component-matched KLT, followed by scalar quantization, achieving a rate of H(C)/n bits per source dimension of the converse, with a vanishing asymptotic gap. We further develop a practical implementation based on EM-driven Gaussian-mixture learning, component-adaptive KLTs, and entropy-constrained scalar quantization (ECSQ). Experiments on synthetic Gaussian mixtures show that PrismQuant closely approaches the theoretical RD bound, while experiments on real-world channel-state-information (CSI) data demonstrate competitive or superior performance compared with transformer-based learned codecs at more than one order of magnitude smaller model size.\n\n---\n\n## 151. Learning with Conflicts of Interest\n- **Authors**: Nischal Aryal , Arash Termehchy , Ali Vakilian , Marianne Winslett\n- **URL**: [https://arxiv.org/abs/2605.15504](https://arxiv.org/abs/2605.15504)\n- **Abstract**:\n> Financial, social, and political factors often prevent the interests of the owners of ML systems and services and their users from being perfectly aligned. ML systems often produce biased information that can influence users to make decisions that are not in their best interest. Current solution approaches require ML systems to implement protocols to mitigate their biases. However, ML system owners usually do not have any incentive to implement these protocols and often argue that it limits their freedom of expression or business. We believe that a successful solution to this problem must recognize the conflict of interest between the ML systems and their users, and use this information to protect users against information that adversely influences their decisions while allowing users to safely benefit from these systems. To this end, we propose a game-theoretic framework that models the interaction between ML systems and users with conflicts of interest. We present scalable algorithms with theoretical guarantees that maximize the amount of desired information and actions and minimize the amount of biased and manipulative actions in interaction with ML systems.\n\n---\n\n## 152. Ghosted Layers: Unconstrained Activation Alignment for Recovering Layer-Pruned LLMs\n- **Authors**: Vincent-Daniel Yun , Junhyuk Jo , Sai Praneeth Karimireddy , Sunwoo Lee\n- **URL**: [https://arxiv.org/abs/2605.15491](https://arxiv.org/abs/2605.15491)\n- **Abstract**:\n> Layer pruning removes entire Transformer decoder blocks from large language models, but introduces a mismatch between the hidden state received by the next surviving layer and the distribution it was trained to process, leading to significant performance degradation. We propose Ghosted Layers, a training-free recovery module that addresses this issue by solving a boundary activation alignment problem. Our method derives a closed-form optimal linear operator from a small calibration set to reconstruct the activation discrepancy introduced by the pruned layers. We show that this solution corresponds to the unconstrained optimum of the alignment objective, whereas existing methods are restricted to constrained solutions over limited operator subspaces. Experiments across multiple LLM backbones and pruning strategies demonstrate that our method consistently improves accuracy and perplexity over prior training-free baselines, while preserving the efficiency gains of layer pruning.\n\n---\n\n## 153. Hybrid LLM-based Intelligent Framework for Robot Task Scheduling\n- **Authors**: Swayamjit Saha , Subhabrata Das , Haonan Duan , Xiao-Yang Liu\n- **URL**: [https://arxiv.org/abs/2605.15486](https://arxiv.org/abs/2605.15486)\n- **Abstract**:\n> This study introduces intelligent frameworks that use Large Language Models (LLMs) to improve task scheduling for construction robots. The LLM is fed with key data about the desired task, such as agent action abilities, and the desired end goal to be achieved. A well-balanced allocation strategy is developed, optimizing both time efficiency and resource utilization. Our system utilizes a Natural Language Processing interface to streamline communication with construction professionals and adapt in real-time to unexpected site conditions. We concurrently use two LLM agents, specifically generator (GPT-4) and supervisor (Gemma 3/Llama 4/Mistral 7b) LLM agents to provide a more precise task schedule. We evaluate the proposed methodology using a straightforward scenario and provide metric scores to prove the efficacy of the frameworks. Our results highlight that the implementation of LLMs is crucial in construction operational tasks including robots.\n\n---\n\n## 154. Residual Reinforcement Learning for Robot Teleoperation under Stochastic Delays\n- **Authors**: Kaize Deng , Zewen Yang\n- **URL**: [https://arxiv.org/abs/2605.15480](https://arxiv.org/abs/2605.15480)\n- **Abstract**:\n> Stochastic communication delays in teleoperation introduce signal discontinuities that undermine control stability and degrade control performance. Consequently, the conventional reinforcement learning (RL) methods struggle with the delayed observations due to the delay-induced observations, leading to high-frequency chattering. To address this, we propose a hybrid control framework, delay-resilient RL, integrating a state estimator utilizing Long Short-Term Memory (LSTM) with a residual RL policy, which is resilient to stochastic delays. The LSTM reconstructs smooth, continuous state estimates from delayed observations, enabling the RL agent to learn a residual torque compensation policy that balances tracking accuracy with velocity smoothness. Experimental validation on Franka Panda robots demonstrates that our approach significantly outperforms the state-of-the-art baselines, ensuring robust and stable teleoperation even under high-variance stochastic delays.\n\n---\n\n## 155. Retrieval-Augmented Large Language Models for Schema-Constrained Clinical Information Extraction\n- **Authors**: A H M Rezaul Karim , Ozlem Uzuner\n- **URL**: [https://arxiv.org/abs/2605.15467](https://arxiv.org/abs/2605.15467)\n- **Abstract**:\n> Conversational nurse-patient transcripts contain actionable observations, but converting these transcripts into structured representations at scale remains challenging. Documentation burden is substantial, with prior studies showing clinicians spend large portions of their workday on documentation and related desk work rather than direct patient care. MEDIQA-SYNUR focuses on observation extraction from conversational nurse-patient transcripts, requiring systems to normalize these narratives into a predefined schema with value-type constraints. We propose a modular retrieval-augmented generation (RAG) pipeline that uses the training set as an exemplar corpus, combines schema-constrained prompting (full schema vs. pruned candidate schema), deterministic schema-based postprocessing, and a second-pass audit, with two LLM backbones: Llama-4-Scout-17B-16E-Instruct and GPT-5.2 with corresponding embedding models for RAG. Our best configuration uses GPT-5.2 with full schema, RAG, and a second-pass auditing, achieving 80.36% F1 score. Overall, our results show that RAG consistently improves performance, while the optimal degree of schema constraint depends on the model, and second-pass auditing yields modest additional gains by correcting residual schema-adherence errors.\n\n---\n\n## 156. GRLO: Towards Generalizable Reinforcement Learning in Open-Ended Environments from Zero\n- **Authors**: Shangjian Yin , Yu Fu , Yue Dong , Zhouxing Shi\n- **URL**: [https://arxiv.org/abs/2605.15464](https://arxiv.org/abs/2605.15464)\n- **Abstract**:\n> Post-training has become a crucial step for unlocking the capabilities of large language models, with reinforcement learning (RL) emerging as a critical paradigm. Recent RL-based post-training has increasingly split into two paradigms: reinforcement learning from human feedback (RLHF), which optimizes models using human preference signals in target domains, and reinforcement learning from verifiable rewards (RLVR), which operates in verifier-backed environments. The latter has dominated recent reasoning-oriented post-training because it delivers stronger gains and higher efficiency on domain-specific tasks (e.g., reasoning). However, although in-domain RL training achieves promising performance, it still requires a substantial amount of GPU compute, which remains a major barrier to broad adoption. In this work, we study the generalization ability of RLHF learned from scratch from a small set of interactions in open-ended environments, and investigate whether the conversational abilities it explicitly acquires can implicitly transfer to downstream tasks such as mathematical reasoning and code generation, namely GRLO. Specifically, on Qwen3-4B-Base backbone, GRLO improves the average performance across all domains from 24.1 to 63.1 with only 5K prompts and 22.7 GPU hours, requiring about $46\\times$ less data and $68\\times$ less compute than a strong in-domain RLVR baseline. The resulting model is even competitive with Qwen's released post-trained models which required a much larger training cost. Notably, a subsequent in-domain RLVR stage brings only selective gains, mainly on harder competition-math benchmarks. We hope GRLO offers a simple and efficient recipe for building broadly capable post-trained models. Our code and data will be available at: \\href{ this https URL }{ this https URL }.\n\n---\n\n## 157. DrugSAGE:Self-evolving Agent Experience for Efficient State-of-the-Art Drug Discovery\n- **Authors**: Yikun Zhang , Xiwei Cheng , Tianyu Liu , Yuanqi Du , Wengong Jin\n- **URL**: [https://arxiv.org/abs/2605.15461](https://arxiv.org/abs/2605.15461)\n- **Abstract**:\n> Building state-of-the-art (SOTA) predictive models for drug discovery requires expensive search over tools, architectures, and training strategies. Current LLM-based agents can find SOTA solutions through extensive trial and error, but they do not retain the experience accumulated along the way and therefore pay the full search cost on every new task. We propose \\method (Self-evolving Agent Experience), a framework that accumulates and reuses experience across tasks to build SOTA drug discovery models efficiently. \\method maintains a cross-task memory of verified skills, statistical evidence about effective strategies, and a record of recurring errors and their fixes. In some cases, \\method transfers a working solution directly without test-time search. In 33 molecular property prediction tasks, \\method ranks first among nine SOTA agents in a single-task setting. With memory accumulated from 16 smaller tasks, \\method achieves an averaged normalized score of 0.935 on 17 held-out tasks in a cross-task evaluation setting and outperforms all baseline agents by 10-30\\% in a zero-test-time search regime. In summary, our work shows the advantage of cross-task memory for efficient SOTA model development in drug discovery.\n\n---\n\n## 158. Differentially Private Motif-Preserving Multi-modal Hashing\n- **Authors**: Zehua Cheng , Wei Dai , Jiahao Sun\n- **URL**: [https://arxiv.org/abs/2605.15460](https://arxiv.org/abs/2605.15460)\n- **Abstract**:\n> Cross-modal hashing enables efficient retrieval by encoding images and text into compact binary codes. State-of-the-art methods rely on semantic similarity graphs derived from user interactions for supervision, yet these graphs encode sensitive behavioral patterns vulnerable to link reconstruction attacks. Existing privacy-preserving approaches fail on graph-structured data: Differentially Private SGD destroys relational motifs by treating samples independently, while graph synthesis methods suffer from unbounded local sensitivity in scale-free networks, hub nodes cause single-edge modifications to alter triangle counts by $\\mathcal{O}(N)$, necessitating prohibitive noise injection. We term this phenomenon Hubness Explosion. We propose DMP-MH, a Sanitize-then-Distill framework that decouples privacy from representation learning. Our approach first bounds sensitivity by deterministically clipping node degrees, capping the $L_2$-sensitivity of triangle motifs independently of dataset size. A sanitized synthetic graph is then generated via Noisy Mirror Descent under $(\\epsilon,\\delta)$-Edge Differential Privacy. Finally, dual-stream hashing networks distill this topology using a holistic structural loss that enforces cross-modal alignment. Evaluated on MIRFlickr-25K and NUS-WIDE under a strict inductive protocol, DMP-MH outperforms private baselines by up to 11.4 mAP points while retaining up to 92.5% of non-private performance.\n\n---\n\n## 159. RIDE: Retinex-Informed Decoupling for Exposing Concealed Objects\n- **Authors**: Chunming He , Rihan Zhang , Dingming Zhang , Chengyu Fang , Longxiang Tang , Jingjia Feng , Fengyang Xiao , Sina Farsiu\n- **URL**: [https://arxiv.org/abs/2605.15450](https://arxiv.org/abs/2605.15450)\n- **Abstract**:\n> Concealed Object Segmentation (COS) encompasses a family of dense-prediction tasks, including camouflaged object detection, polyp segmentation, transparent object detection, and industrial defect inspection, where targets are visually entangled with their surroundings through different physical mechanisms. Existing methods either operate directly on RGB images or employ \\emph{heterogeneous} decompositions (\\eg, Fourier, wavelet) that redistribute spatial evidence across scale/frequency coefficients, making pixel-aligned cues less direct. We introduce a fundamentally different perspective: \\textbf{homogeneous image decomposition} via Retinex theory, which factorizes an image into illumination and reflectance components within the \\emph{same} spatial domain. Our key insight is that visual entanglement enforces appearance matching in the composite space, but this does \\emph{not} necessitate simultaneous matching in both component spaces, a phenomenon we formalize as the \\textbf{Discriminability Gap Theorem}. Crucially, we show that across diverse COS sub-tasks, the underlying physical processes systematically anti-correlate illumination and reflectance differences, yielding theoretical guarantees that Retinex decomposition preserves or strictly improves total foreground--background discriminability across the full physical regime, with anti-correlation maximizing the gain. Building on this, we propose \\textbf{RIDE} comprising: (i) a Task-Driven Retinex Decomposition module that learns segmentation-optimal factorizations end-to-end; (ii) a Discriminability Gap Attention mechanism that adaptively exploits where decomposition helps; and (iii) a Camouflage-Breaking Contrastive loss operating in reflectance feature space.\n\n---\n\n## 160. Runtime-Structured Task Decomposition for Agentic Coding Systems\n- **Authors**: Shubhi Asthana , Bing Zhang , Chad DeLuca , Hima Patel , Ruchi Mahindru\n- **URL**: [https://arxiv.org/abs/2605.15425](https://arxiv.org/abs/2605.15425)\n- **Abstract**:\n> Agentic coding systems increasingly use large language models (LLMs) for software engineering tasks such as debugging, root cause analysis, and code review. However, many existing systems encode task logic, execution flow, and output generation inside monolithic prompts. This design creates brittle behavior, limited debuggability, and high retry costs because failures often require rerunning the full workflow. We present runtime-structured task decomposition, an architectural approach in which task partitioning and execution flow are managed through executable control logic rather than prompt structure alone. LLMs are used only for focused judgment tasks, and outputs are validated against predefined schemas before downstream execution. We evaluate this approach on two software engineering workloads using three configurations: monolithic execution, static decomposition with fixed subtasks and no runtime branching, and runtime-structured decomposition. Each configuration was evaluated across 10 runs. Our results show that decomposition alone does not necessarily reduce retry cost. In the Kubernetes root cause analysis workload, the static decomposition baseline produced a retry cost of 1,632 +/- 145 tokens versus 904 +/- 17 tokens for the monolithic baseline because failures forced reruns of downstream subtasks. A similar pattern appeared in the multi-file debugging workload, where the static baseline consumed 933 tokens compared to 703 tokens for the monolithic system. The runtime-structured approach reran only failed subtasks, reducing retry costs to 436 +/- 132 tokens for root cause analysis and 460 tokens for debugging. Overall, the approach achieved up to 51.7% lower retry cost than monolithic systems and 73.2% lower retry cost than static decomposition baselines, improving efficiency, debuggability, and operational reliability in agentic coding systems.\n\n---\n\n## 161. MR2-ByteTrack: CNN and Transformer-based Video Object Detection for AI-augmented Embedded Vision Sensor Nodes\n- **Authors**: Luca Bompani , Manuele Rusci , Luca Benini , Daniele Palossi , Francesco Conti\n- **URL**: [https://arxiv.org/abs/2605.15423](https://arxiv.org/abs/2605.15423)\n- **Abstract**:\n> Modern smart vision sensors need on-device intelligence to process video streams, as cloud computing is often impractical due to bandwidth, latency, and privacy constraints. However, these sensory systems typically rely on ultra-low-power microcontrollers (MCUs) with limited memory and compute, making conventional video object detection methods, which require feature storage or multi-frame buffering, unfeasible. To address this challenge, we introduce Multi-Resolution Rescored ByteTrack (MR2-ByteTrack), a Video Object Detection (VOD) method tailored for MCU-based embedded vision nodes. MR2-ByteTrack reduces computational cost by alternating between full- and low-resolution inference, while linking detections across frames via ByteTrack and correcting misclassifications through the Rescore algorithm, which applies probability union rules to aggregate detection confidence scores across frames. We apply our approach to both a CNN-based detector and a Transformer-based model, demonstrating its generality across architectures with fundamentally different spatial processing. Experiments on ImageNetVID demonstrate that MR2-ByteTrack maintains accuracy, achieving mAP scores of up to 49.0 for the CNN-based models and 48.7 for the Transformer, while reducing multiply-accumulate operations by as much as 53\\% for the CNNs and 32\\% for the Transformer. When deployed on GAP9, an ultra-low-power RISC-V multicore MCU, our method yields up to 55\\% energy savings compared to processing only full-resolution images, enabling the first real-time Transformer-based VOD on an MCU-class embedded vision node. Code available at this https URL\n\n---\n\n## 162. $f$-Trajectory Balance: A Loss Family for Tuning GFlowNets, Generative Models, and LLMs with Off- and On-Policy Data\n- **Authors**: Jake Fawkes , Jason Hartford\n- **URL**: [https://arxiv.org/abs/2605.15417](https://arxiv.org/abs/2605.15417)\n- **Abstract**:\n> In GFlowNets and variational inference, it has been shown that the mean square error between target and model log probabilities is an effective, low variance, surrogate loss for training generative models. This loss has the property that when evaluated \\emph{on-policy} its gradients correspond to those of the KL divergence, while \\emph{off-policy} it remains a valid loss with the same global minimizer. In this work, we demonstrate that this construction can be extended to the whole family of $f$-divergences, leading to a family of losses whose on-policy gradients are that of the corresponding $f$-divergence, but retain the same global minimizer off-policy. Specifically, we show that the on-policy gradients lead to a one to one correspondence between translation invariant loss functions on the target and model log probabilities, and $f$-divergences. This equivalence allows us to design new surrogate loss functions for tuning a wide class of generative models that inherit the properties of the corresponding $f$-divergence, such as being more mode covering, whilst being applicable to off-policy data. We apply our losses on a range of tasks, including classic synthetic examples, SynFlowNets for molecule discovery, and asynchronous large language model (LLM) tuning, demonstrating that our models retain their predicted properties on- and off-policy in a wide class of generative models.\n\n---\n\n## 163. Margin-Adaptive Confidence Ranking for Reliable LLM Judgement\n- **Authors**: Gaojie Jin , Yong Tao , Lijia Yu , Tianjin Huang\n- **URL**: [https://arxiv.org/abs/2605.15416](https://arxiv.org/abs/2605.15416)\n- **Abstract**:\n> Jung et al. (2025) introduce a hypothesis testing framework for guaranteeing agreement between large language models (LLMs) and human judgments, relying on the assumption that the model's estimated confidence is monotonic with respect to human-disagreement risk. In practice, however, this assumption may be violated, and the generalization behavior of the confidence estimator is not explicitly analyzed. We mitigate these issues by learning a dedicated confidence estimator instead of relying on heuristic confidence signals. Our approach leverages simulated annotator diversity and a margin-based ranking formulation to explicitly model how confidently an LLM distinguishes between human-agreement and human-disagreement cases. We further derive generalization guarantees for this estimator, revealing a margin-dependent trade-off that informs the design of an adaptive estimator training procedure. When integrated into fixed-sequence testing, the learned confidence estimator yields improved ranking accuracy and empirically strengthens the monotonic relationship between confidence and disagreement risk, leading to higher success rates in satisfying target agreement levels across multiple datasets and judge models.\n\n---\n\n## 164. From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery\n- **Authors**: Lingzhe Zhang , Tong Jia , Yunpeng Zhai , Zixuan Xie , Chiming Duan , Minghua He , Philip S. Yu , Ying Li\n- **URL**: [https://arxiv.org/abs/2605.15412](https://arxiv.org/abs/2605.15412)\n- **Abstract**:\n> Modern quantitative trading increasingly relies on systematic models to extract predictive signals from large-scale financial data, where alpha factor discovery plays a central role in transforming market observations into tradable signals. Recent LLM-based methods have shown promise in automating factor generation, but most of them still rely on prompt-level generation--evaluation--feedback loops for iterative optimization. As the loop becomes longer, repeatedly appended historical candidates and feedback can cause context explosion, increase inference cost, dilute useful information, and introduce feedback drift. Moreover, these methods often depend on very large LLMs whose stable generation preferences may lead to structurally similar expressions, redundant candidates, and search stagnation. To address these limitations, we propose \\textsc{QuantEvolver}, a self-evolving alpha factor discovery framework based on reinforcement fine-tuning. Instead of accumulating feedback in the prompt, \\textsc{QuantEvolver} converts executable quantitative evaluation into policy updates, enabling a Miner LLM to internalize historical optimization experience through parameter learning. Specifically, \\textsc{QuantEvolver} constructs high-quality seed factors, builds diverse seed--time-window training tasks, generates executable Factor DSL expressions, evaluates them through Regime Backtest, and optimizes the Miner LLM with Diversity-Complementarity Reward. During training, high-quality factors are continuously accumulated in a Mined Factor Database, which serves as the final discovered factor library. Extensive experiments on three realistic market benchmarks demonstrate the effectiveness of \\textsc{QuantEvolver}, which consistently improves the primary evaluation metric of each task over existing LLM-based alpha factor discovery baselines, produces higher-quality and more complementary factor pools.\n\n---\n\n## 165. Diagonal Adaptive Non-local Observables on Quantum Neural Networks\n- **Authors**: Huan-Hsin Tseng , Yan Li , Hsin-Yi Lin , Samuel Yen-Chi Chen\n- **URL**: [https://arxiv.org/abs/2605.15410](https://arxiv.org/abs/2605.15410)\n- **Abstract**:\n> Adaptive Non-local Observables (ANOs) have shown that making quantum observables dynamic can substantially enlarge the function space of Variational Quantum Algorithms, partly shifting hardware demands from circuit synthesis to measurement design. However, this advantage is accompanied by a steep increase in the number of parameters, as well as the classical optimization cost for varying general Hermitian observables. We propose a special form of ANO that significantly reduces this burden by considering only diagonal observables paired with quantum circuits. Mathematically, this is equivalent to the full ANO of a large parameter space since diagonal matrices are canonical representatives of the ANO space modulo unitary similarity. As a result, Diagonal ANO retains the same capability of full ANO while reducing $k$-local observable complexity from $O(4^k)$ to $O(2^k)$ and lowering the corresponding measurement-side classical computation. In this sense, diagonal ANO preserves much of the benefit of full ANO while encompassing conventional VQCs as a special case.\n\n---\n\n## 166. Amortized Energy-Based Bayesian Inference\n- **Authors**: Hojjat Kaveh , Ricardo Baptista , Andrew M. Stuart\n- **URL**: [https://arxiv.org/abs/2605.15407](https://arxiv.org/abs/2605.15407)\n- **Abstract**:\n> We consider amortized Bayesian inference for nonlinear inverse problems in settings where only samples from the joint distribution of parameters and observations are available. Classical methods such as Markov chain Monte Carlo require solving a new inference problem for each observation, which can be computationally prohibitive when inference must be repeated many times. We propose a transport-based approach that learns an observation-dependent map pushing forward a reference measure to approximate the posterior distribution. The map is trained by minimizing an averaged energy-distance objective between the true posterior and the learned pushforward. This formulation is likelihood-free, requiring only joint samples, and avoids density evaluation, invertibility constraints, and Jacobian determinant computations. For function-space inverse problems with Gaussian priors, we parameterize the transport map as the identity plus a perturbation in the Cameron-Martin space of the prior, preserving absolute continuity with respect to the prior. In infinite-dimensional settings, the map is represented using neural operators. We illustrate the method on a finite-dimensional nonlinear inverse problem and two PDE-constrained inverse problems arising in porous medium flow and seismic inversion. The results show that the learned transport captures posterior structure, including multimodality and dominant modes, while enabling fast posterior sampling for new observations.\n\n---\n\n## 167. Breakeven complexity: A new perspective on neural partial differential equation solvers\n- **Authors**: Yijing Zhang , Nicholas Roberts , Tanya Marwah , Mikhail Khodak\n- **URL**: [https://arxiv.org/abs/2605.15399](https://arxiv.org/abs/2605.15399)\n- **Abstract**:\n> Neural surrogate solvers of partial differential equations (PDEs) promise dramatic speedups over numerical methods, especially in scenarios requiring many solves. However, current accuracy-based evaluations do not fully consider two central issues: (1) neural solvers incur substantial up-front costs for data generation, training, and tuning; and (2) classical solvers can also generate low-fidelity solutions at a sufficiently low simulation cost. To explicitly account for these realities and fully incorporate end-to-end costs, we propose an evaluation framework centered on breakeven complexity, a metric that counts the forward solves before a learned solver is cost-effective relative to an error-equivalent traditional solver. To evaluate this measure, we apply scaling laws to determine how much training budget to allocate to data generation and discuss how to achieve smooth error-matching in diverse settings. We evaluate the breakeven complexity of multiple neural PDE solvers on three PDEs on 2D periodic domains from APEBench and a novel benchmark of flows past multiple obstacles generated by the GPU-native PyFR code. Among other findings, our results suggest that neural PDE solvers become more effective as problems get harder in terms of cost, dimension, rollout, physics regime (e.g. higher Reynolds number), etc.\n\n---\n\n## 168. Representation Without Reward: A JEPA Audit for LLM Fine-Tuning\n- **Authors**: Biswa Sengupta\n- **URL**: [https://arxiv.org/abs/2605.15394](https://arxiv.org/abs/2605.15394)\n- **Abstract**:\n> Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outputs. For autoregressive language-model fine-tuning the principle entails a stricter requirement: the induced hidden-state geometry must reach the language-model head \\emph{and} improve the decoded task metric. We test that requirement under a fixed Llama-3.2-1B-Instruct LoRA harness on natural-language-to-regex generation, comparing twenty-two training-time auxiliaries across trajectory-shape regularisation, distributional constraints, predictor/target asymmetry, Fisher-metric Jacobi residuals, and a decoder-visible JEPA objective constructed to lie in cross-entropy's positive cone. The empirical answer is a structured null: several auxiliaries clear single-cell paired $\\alpha = 0.10$ without correction (T3-Local at $\\Delta = +2.53$~pp, $p = 0.003$ being the strongest), but none survives Bonferroni or Holm--Bonferroni at the relevant family-wise threshold, even though many change curvature, anisotropy, variance, and gradient direction. Decoder-visible JEPA yields the first positive auxiliary--cross-entropy gradient cosine in the study, yet exact match remains inside seed noise; a full-fine-tuning replication of the same auxiliary at $n = 5$ seeds reproduces the null on both benchmarks (TURK: $\\Delta = +0.04$~pp, $p_{\\text{paired} } = 0.96$; SYNTH: $\\Delta = +0.52$~pp, $p_{\\text{paired} } = 0.28$), so the null is robust across LoRA and full fine-tuning for the decoder-visible construction. Hidden-state representation work and decoded-task accuracy are therefore weakly coupled in this regime; we accordingly reframe LLM-domain JEPA evaluation as a coupling problem, in which the operative question is under which metrics useful hidden geometry becomes decoder-visible task signal.\n\n---\n\n## 169. PanoWorld: Geometry-Consistent Panoramic Video World Modeling\n- **Authors**: Le Jiang , Xiangyu Bai , Bishoy Galoaa , Shayda Moezzi , Caleb James Lee , Tooba Imtiaz , Edmund Yeh , Jennifer Dy , Yanzhi Wang , Sarah Ostadabbas\n- **URL**: [https://arxiv.org/abs/2605.15391](https://arxiv.org/abs/2605.15391)\n- **Abstract**:\n> We present PanoWorld, a panoramic video world model that generates geometry-consistent 360$\\degree$ video from a single image and a caption. Existing panoramic video methods optimize primarily for visual realism and do not explicitly constrain the underlying 3D scene state, producing outputs that appear plausible yet exhibit inconsistent depth, broken correspondences, and implausible motion across the spherical surface. We address this gap by framing panoramic video generation as a geometry- and dynamics-consistent latent state modeling problem rather than pure visual synthesis. Building on a pre-trained perspective video world model, we introduce two lightweight regularizers: a depth consistency loss against pseudo ground-truth panoramic depth, and a trajectory consistency loss that supervises the 3D world-frame positions of tracked points across time. We further apply spherical-geometry-aware adaptation to the conditioning and positional encoding. We additionally introduce PanoGeo, a unified geometry-aware panoramic video dataset with consistent depth, trajectory, and prompt annotations across diverse real and synthetic sources, used for both training and stratified evaluation. Experiments show that PanoWorld improves geometric consistency over prior panoramic generation methods while maintaining competitive visual realism, establishing that panoramic video generation must be treated as a geometric modeling problem to support the holistic spatial understanding requirements of embodied AI applications. Code is available at this https URL .\n\n---\n\n## 170. Is One Score Enough? Rethinking the Evaluation of Sequentially Evolving LLM Memory\n- **Authors**: Songwei Dong , Zihan Chen , Chengshuai Shi , Peng Wang , Jundong Li , Cong Shen\n- **URL**: [https://arxiv.org/abs/2605.15384](https://arxiv.org/abs/2605.15384)\n- **Abstract**:\n> Memory plays a central role in enabling large language models (LLMs) to operate over sequential tasks by accumulating and reusing experience over time. However, existing evaluations of LLM memory mostly rely on aggregate metrics such as final hold-out accuracy or cumulative online performance, which can obscure critical failure modes such as forgetting and negative transfer. In this paper, we introduce SeqMem-Eval, a diagnostic evaluation framework for sequentially evolving LLM memory. Drawing inspiration from continual learning, it targets a test-time setting in which memory is external, prompt-mediated, and updated without modifying model parameters. Rather than focusing only on final performance, SeqMem-Eval evaluates how memory states evolve, generalize, consolidate experience, and retain useful information during sequential inference. Specifically, it measures online utility, hold-out generalization, backward transfer, and forgetting, providing a finer-grained view of memory quality. Through extensive experiments across diverse tasks and memory methods, we show that higher final or cumulative accuracy does not necessarily imply better memory quality: many methods exhibit strong performance gains while suffering from substantial forgetting or negative transfer. Moreover, different memory designs exhibit distinct trade-offs between adaptability and stability that remain invisible under standard evaluation metrics.\n\n---\n\n## 171. ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing\n- **Authors**: Blaž Rolih , Matic Fučka , Filip Wolf , Luka Čehovin Zajc\n- **URL**: [https://arxiv.org/abs/2605.15375](https://arxiv.org/abs/2605.15375)\n- **Abstract**:\n> Remote sensing change detection (RSCD) aims to localise changes between two images of the same geographic region. In practice, change masks often follow region-level annotation conventions rather than purely local appearance differences, making them context-dependent and occasionally ambiguous. Most state-of-the-art methods utilise per-pixel discriminative classification, which produces a single prediction per input and fails to explicitly model the changed region as a coherent whole. A natural alternative is generative formulation, which can model a distribution of plausible masks, enabling sampling to capture ambiguity and encourage global consistency. However, existing generative RSCD approaches typically lag behind strong discriminative baselines due to the high computational cost of pixel-space generation and the complexity of their conditioning mechanisms. To address the limitations of prior discriminative and generative methods, we propose ChangeFlow, a generative framework that reformulates change detection as the synthesis of a change mask in latent space via rectified flow. ChangeFlow is guided by a structured yet lightweight conditioning signal, and its stochastic design naturally supports sampling-based prediction ensembling. Namely, aggregating multiple predicted change masks improves robustness, while sample agreement provides a practical confidence estimation that highlights ambiguous regions. Across four benchmarks, ChangeFlow achieves an average F1 of 80.4\\%, improving by 1.3 points on average over the previous best method, while maintaining inference speed comparable to recent strong baselines. Project page: this https URL\n\n---\n\n## 172. PACER: Acyclic Causal Discovery from Large-Scale Interventional Data\n- **Authors**: Ramon Viñas Torné , Sílvia Fàbregas Salazar , Soyon Park , Ivo Alexander Ban , Artyom Gadetsky , Nikita Doikov , Maria Brbić\n- **URL**: [https://arxiv.org/abs/2605.15353](https://arxiv.org/abs/2605.15353)\n- **Abstract**:\n> Inferring the structure of directed acyclic graphs (DAGs) from data is a central challenge in causal discovery, particularly in modern high-dimensional settings where large-scale interventional data are increasingly available. While interventional data can improve identifiability, existing methods remain limited by soft acyclicity constraints, leading to optimization over invalid cyclic graphs, numerical instability, and reduced scalability. We introduce PACER (Perturbation-driven Acyclic Causal Edge Recovery), a scalable framework for causal discovery that guarantees acyclicity by construction. PACER parameterizes a distribution over DAGs through a joint model of variable permutations and edge probabilities, enabling direct optimization over valid causal structures without surrogate penalties. The framework supports a unified likelihood-based treatment of observational and interventional data, flexible conditional density models, and the incorporation of structural prior knowledge. For linear-Gaussian mechanisms, we derive closed-form expressions for the expected interventional log-likelihood and its gradients, yielding substantial computational gains. Empirically, PACER matches or exceeds state-of-the-art methods on protein signaling and large-scale genetic perturbation benchmarks, while scaling efficiently to networks with thousands of variables and achieving up to two orders of magnitude speedups over penalty-based differentiable approaches. These results demonstrate that exact and scalable causal discovery from high-dimensional perturbation data is achievable through principled search space design.\n\n---\n\n## 173. LEAP: Trajectory-Level Evaluation of LLMs in Iterative Scientific Design\n- **Authors**: Marilyn Zhang , Tianfeng Chen , Fabián Barzuna , Ankita Rathod , Mark E. Whiting\n- **URL**: [https://arxiv.org/abs/2605.15341](https://arxiv.org/abs/2605.15341)\n- **Abstract**:\n> LLMs are increasingly deployed in autonomous laboratories, under the assumption that their domain priors and reasoning over iterative feedback let them converge on good designs in fewer iterations than feedback-only baselines. Current iterative scientific design benchmarks, however, score only outcome snapshots at fixed horizons. This leaves the learning trajectory unmeasured, even though the trajectory is what captures learning efficiency, where each iteration saved is a real saving in cost and time. Motivated by this, we examine three evaluation choices that change the conclusions one draws about LLM learning efficiency in iterative scientific design: what to measure, what baseline to compare against, and what to ground against. We introduce LEAPBench, Learning Efficiency in Adaptive Processes, a 55-task framework that pairs a best-so-far area under the curve (AUC) trajectory metric with a classical Bayesian-optimization reference and an audit grounded in published literature. Applied to eight contemporary LLMs, switching from final-outcome to trajectory scoring changes the best-model decision on 53% of tasks at matched horizons, and exposes efficiency gains overlooked by outcome-based scoring. LLMs do not outperform a classical Bayesian baseline. On 16 biology tasks where the oracle's reward signal is aligned with configurations from the published-best design, domain-aware prompting leads to LLM choices that match the published-best's approximately 10 percentage points less often than domain-agnostic prompting at iteration 30. The pattern is sharpest on 6 tasks where the literature-typical and published-best configurations diverge, and domain-agnostic prompting matches the published-best more often on all 6. The trajectory metric also doubles as a tractable training target. Offline reinforcement learning with the metric as a reward improves performance on 14 of 21 held-out tasks.\n\n---\n\n## 174. Hidden in Memory: Sleeper Memory Poisoning in LLM Agents\n- **Authors**: Sidharth Pulipaka , Stanislau Hlebik , Leonidas Raghav , Sahar Abdelnabi , Vyas Raina , Ivaxi Sheth , Mario Fritz\n- **URL**: [https://arxiv.org/abs/2605.15338](https://arxiv.org/abs/2605.15338)\n- **Abstract**:\n> Large language models are increasingly augmented with persistent memory, allowing assistants to store user-specific information across sessions for personalization and continuity. This statefulness introduces a new security risk: adversarial content can corrupt what an assistant remembers and thereby influence future interactions. We propose and study sleeper memory poisoning, a delayed attack in which an adversary manipulates external context, such as a document, webpage, or repository, to cause the assistant to store a fabricated memory about the user. Unlike conventional prompt injection, the attack can remain dormant and re-emerge across multiple later conversations. We evaluate the full attack pipeline: whether poisoned memories are written, later retrieved, and ultimately used to steer the following conversations. Across stateful LLM assistants, poisoned memories were added up to 99.8% on GPT-5.5 and 95% on Kimi-K2.6. Crucially, among successful retrievals, poisoned memories cause attacker-intended agentic actions in 60-89% of evaluations across models. These results show that persistent memory can act as a long-term attack surface across multiple future conversations.\n\n---\n\n## 175. HoloMotion-1 Technical Report\n- **Authors**: Maiyue Chen , Kaihui Wang , Bo Zhang , Xihan Ma , Zhiyuan Yang , Yi Ren , Qijun Huang , Zihao Zhu , Yucheng Wang , Zhizhong Su\n- **URL**: [https://arxiv.org/abs/2605.15336](https://arxiv.org/abs/2605.15336)\n- **Abstract**:\n> In this report, we present HoloMotion-1, a humanoid motion foundation model for zero-shot whole-body motion tracking. A key innovation of HoloMotion-1 is to scale control-policy training with a large-scale hybrid motion corpus, where video-reconstructed motions from in-the-wild videos provide the dominant source of motion diversity, while curated motion-capture and in-house motion data provide higher-fidelity supervision and deployment-oriented coverage. This data regime enables HoloMotion-1 to move beyond conventional MoCap-only training and exposes the policy to substantially broader behaviors, capture conditions, and motion styles. Learning from such heterogeneous data introduces new challenges, including reconstruction noise, source-domain mismatch, uneven motion quality, and the need for temporal modeling under large behavioral variation. To address these challenges, HoloMotion-1 integrates large-capacity temporal modeling, a sparsely activated Mixture-of-Experts Transformer with KV-cache inference for real-time control, and a sequence-level training strategy that improves learning efficiency on extended motion sequences. Extensive experiments on multiple unseen motion benchmarks show that HoloMotion-1 generalizes robustly across diverse motion types and capture conditions, significantly improves tracking accuracy over prior methods, and transfers directly to a real humanoid robot without task-specific fine-tuning.\n\n---\n\n## 176. From I/O to Code with Discovery Agent\n- **Authors**: Yihong Dong , Jiaru Qian , Haoran Zhang , Peixu Wang , Binhua Li , Zhi Jin , Yongbin Li , Ge Li , Xiaokang Yang , Xue Jiang\n- **URL**: [https://arxiv.org/abs/2605.15334](https://arxiv.org/abs/2605.15334)\n- **Abstract**:\n> The automatic synthesis of a program from any form of specification is regarded as a holy grail of computer science. Fueled by LLMs, NL2Code has achieved tremendous success, yet the fundamentally more challenging task of synthesizing programs from input-output behavior, which we refer to as IO2Code, remains largely unsolved. Whereas NL2Code can exploit the semantic alignment between natural language and code acquired during pretraining, IO2Code requires recovering underlying principles from concrete computational behavior, navigating a vast and underspecified hypothesis space. To address this, we propose DIO-Agent, a discovery agent for IO2Code. Our method frames IO2Code as an evolutionary search over discrete program space, in which an LLM serves as the mutation operator and concrete error signals from execution guide each mutation. To prevent the search from wandering into structurally complex yet incorrect dead ends, we introduce the Transformation Priority Premise as a mutation prior that biases the LLM toward the simplest hypothesis consistent with current evidence, progressively escalating from constants to conditionals to iteration only when simpler constructs are insufficient. To facilitate systematic study, we further construct an IO2CodeBench spanning multiple difficulty levels. Extensive experiments show that DIO-Agent consistently outperforms both traditional program-by-example method and SOTA evolution-agent baselines across all difficulty levels and various LLMs, while substantially surpassing test-time scaling strategies with equivalent sampling budgets.\n\n---\n\n## 177. Fortress: A Case Study in Stabilizing Search Recommendations via Temporal Data Augmentation and Feature Pruning\n- **Authors**: Milind Pandurang Jagre , Jia Huang , Dayvid V. R. Oliveira , Zhinan Cheng , Babak Seyed Aghazadeh , Puja Das , Chris Alvino , Jinda Han , Kailash Thiyagarajan\n- **URL**: [https://arxiv.org/abs/2605.15299](https://arxiv.org/abs/2605.15299)\n- **Abstract**:\n> In search and recommendation systems, predictive models often suffer from temporal instability when certain input features introduce volatility in output scores. This instability can degrade model reliability and user experience especially in multi-stage systems where consistent predictions are critical for downstream decision making. We introduce Fortress, a general framework for enhancing model stability and accuracy by identifying and pruning features that contribute to inconsistent prediction scores over time. Fortress leverages historical snapshots temporally partitioned datasets capturing score fluctuations for the same entity across periods and follows a four-step process: (1) collect historical snapshots, (2) identify samples with unstable predictions, (3) isolate and remove instability-inducing features, and (4) retrain models using only stable features. While semantic features from LLMs and BERT-based models improve generalization, they often lack full query or entity coverage. Engagement-based features offer strong predictive power but tend to introduce temporal instability. Fortress mitigates this trade-off by suppressing the volatility of engagement signals while retaining their predictive value leading to more stable and accurate models. We validate Fortress on a query-to-app relevance model in a large-scale app marketplace. Offline experiments demonstrate notable improvements in prediction stability (measured by Coefficient of Variation) and classification performance (measured by PR-AUC).\n\n---\n\n## 178. PhysBrain 1.0 Technical Report\n- **Authors**: Shijie Lian , Bin Yu , Xiaopeng Lin , Changti Wu , Hang Yuan , Xiaolin Hu , Zhaolong Shen , Yuzhuo Miao , Haishan Liu , Yuxuan Tian , Yukun Shi , Cong Huang , Kai Chen\n- **URL**: [https://arxiv.org/abs/2605.15298](https://arxiv.org/abs/2605.15298)\n- **Abstract**:\n> Vision-language-action models have advanced rapidly, but robot trajectories alone provide limited coverage for learning broad physical understanding. PhysBrain 1.0 studies a complementary route: converting large-scale human egocentric video into structured physical commonsense supervision before robot adaptation. Our data engine extracts scene elements, spatial dynamics, action execution, and depth-aware relations, then turns them into question-answer supervision for training PhysBrain VLMs. The resulting physical priors are further transferred to VLA policies through a capability-preserving and language-sensitive adaptation design. Across multimodal QA benchmarks and embodied control benchmarks, including ERQA, PhysBench, SimplerEnv-WidowX, LIBERO, and RoboCasa, PhysBrain 1.0 achieves SOTA results and shows especially strong out-of-domain performance on SimplerEnv. These results suggest that scaling physical commonsense from human interaction video can provide an effective bridge from multimodal understanding to robot action.\n\n---\n\n## 179. GESD: Beyond Outcome-Oriented Fairness\n- **Authors**: Gideon Popoola , John Sheppard\n- **URL**: [https://arxiv.org/abs/2605.15295](https://arxiv.org/abs/2605.15295)\n- **Abstract**:\n> Machine learning (ML) algorithms are increasingly deployed in high-stakes decision-making domains such as loan approvals, hiring, and recidivism predictions. While existing fairness metrics (e.g., statistical parity, equal opportunity) effectively quantify outcome-oriented disparities, they offer limited insight into the procedure or explanation behind biased decisions. To address this gap, we propose Group-level Explanation Stability Disparity (GESD), a \\textit{procedural-oriented} fairness metric that measures disparities in the stability, robustness, and sensitivity of model explanations across different subgroups in a protected category. %GESD is explainer-agnostic, model-agnostic, and extends the scope of fairness analyses to the level of explainability. We further integrate GESD into a multi-objective optimization framework that jointly optimizes for utility, outcome-based fairness, and explanation-based fairness called FEU (Fairness--Explainability--Utility). Empirical results on multiple benchmark datasets show that GESD effectively captures group-wise discrepancies in explanation quality, and that FEU improves both utility and fairness over state-of-the-art methods. By bridging outcome-based and explanation-based fairness, GESD offers a comprehensive tool for diagnosing and mitigating bias in predictive modeling. Our code and datasets are available on GitHub {\\hyperlink{ this https URL }{ this https URL } }\n\n---\n\n## 180. GQA-μP: The maximal parameterization update for grouped query attention\n- **Authors**: Kyle R. Chickering , Huijuan Wang , Mengxi Wu , Alexander Moreno , Muhao Chen , Xuezhe Ma , Daria Soboleva , Joel Hestness , Zhengzhong Liu , Eric Xing\n- **URL**: [https://arxiv.org/abs/2605.15290](https://arxiv.org/abs/2605.15290)\n- **Abstract**:\n> Hyperparameter transfer across model architectures dramatically reduces the amount of compute necessary for tuning large language models (LLMs). The maximal update parameterization ({\\mu}P) ensures transfer through principled mathematical analysis but can be challenging to derive for new model architectures. Building on the spectral feature-learning view of Yang et al. (2023a), we make two advances. First, we promote spectral norm conditions on the weights from a heuristic to the definition of feature learning, and as a consequence arrive at the Complete-P depth and weight-decay scalings without recourse to lazy-learning. Second, we consider a modified spectral norm that preserves the valid scaling law of network weights when weight matrices are not full rank. This enables (to our knowledge, the first) derivation of {\\mu}P scalings for grouped-query attention (GQA). We demonstrate the efficacy of our theoretical derivations by showing learning rate transfer across the GQA repetition hyperparameter as well as experiments regarding transfer over weight decay.\n\n---\n\n## 181. Universal Approximation of Nonlinear Operators and Their Derivatives\n- **Authors**: Filippo de Feo\n- **URL**: [https://arxiv.org/abs/2605.15285](https://arxiv.org/abs/2605.15285)\n- **Abstract**:\n> Derivative-Informed Operator Learning (DIOL), i.e. learning a (nonlinear) operator and its derivatives, is an open research frontier at the foundations of the influential field of Operator Learning (OL). In particular, Universal Approximation Theorems (UATs) of nonlinear operators and their derivatives are foundational open questions and delicate problems in nonlinear functional analysis. In this manuscript, we prove the first UATs of non-linear $k$-times differentiable operators between Banach spaces and their derivatives, uniformly on compact sets and in weighted Sobolev norms for general finite input measures, via OL architectures. Our results are the first complete generalizations of the corresponding influential classical results in [Hornik, 1991] to infinite-dimensional settings and OL. We discuss several open areas where DIOL and our UATs find applications: high-order accuracy in OL, fast constrained optimization in Banach spaces (e.g. optimal control of PDEs, inverse problems) and numerical methods for infinite-dimensional PDEs (e.g. HJB PDEs on Banach spaces from optimal control of PDEs, SPDEs, path-dependent systems, partially observed systems, mean-field control). We parameterize nonlinear operators via Encoder-Decoder Architectures, renowned classes in OL due to their generality, including classical architectures, such as DeepONets, Deep-H-ONets, PCA-Nets. Our results are based on four key features that allow us to prove UATs in full generality: (i) Approximation Properties of Banach spaces. (ii) $k$-times continuous differentiability in the sense of Bastiani (weaker than $k$-times continuous Fréchet differentiability). (iii) Natural compact-open topologies for UA; indeed, we show that UA in standard compact-open topologies induced by operator norms is violated even for Fréchet derivatives. (iv) Construction of novel weighted Sobolev spaces for the UA.\n\n---\n\n## 182. Autonomous Intelligent Agents for Natural-Language-Driven Web Execution with Integrated Security Assurance\n- **Authors**: Vinil Pasupuleti , Siva Rama Krishna Varma Bayyavarapu , Shrey Tyagi\n- **URL**: [https://arxiv.org/abs/2605.15281](https://arxiv.org/abs/2605.15281)\n- **Abstract**:\n> Modern web test suites rot. A UI refactor breaks locators, a timing change causes race conditions, and within weeks developers abandon the suite entirely. This paper presents an AI-driven autonomous testing framework that addresses these failure modes through five integrated strategies - navigation reliability, context-aware selector generation, post-generation validation, smart wait injection, and failure learning - implemented over a containerised worker architecture that decouples orchestration from long-running browser execution. Evaluated across four production applications and 176 scenarios, the framework improves script generation success from 55% to 93%, achieves an 8x reduction in navigation failures, eliminates 80% of timing-related race conditions, and reduces test creation time by 75% compared to manual Selenium authoring. The framework extends naturally to security validation: testers describe attack scenarios in plain English - \"try accessing another user's invoice\" - which the agent converts to OWASP Top 10-aligned browser probes, detecting 85% of authentication bypass vulnerabilities and 95% of input validation flaws with false positive rates below 12%. Natural-language-driven security testing of this kind represents, to our knowledge, a novel contribution to the field.\n\n---\n\n## 183. PDRNN: Modular Data-driven Pedestrian Dead Reckoning on Loosely Coupled Radio- and Inertial-Signalstreams\n- **Authors**: Peter Bauer , Andreas Porada , Felix Ott , Christopher Mutschler , Tobias Feigl\n- **URL**: [https://arxiv.org/abs/2605.15252](https://arxiv.org/abs/2605.15252)\n- **Abstract**:\n> Modern pedestrian dead reckoning (PDR) systems rely on fusing noisy and biased estimates of position, velocity, and calibrated orientation derived from loosely coupled sensors to determine the current pose of a localized object. However, discrepancies in the sampling rates of sensor-specific estimation methods and unreliable transmission pose significant challenges. And traditional methods often fail to effectively fuse multimodal sensor data during dynamic movements characterized by high accelerations, velocities, and rapidly varying orientations. To address these limitations, we propose a simple recurrent neural network (RNN) architecture capable of implicitly forecasting asynchronous sensor data streams from diverse estimation methods along reference trajectories. The proposed approach introduces PDRNN, a modular hybrid AI-assisted PDR system that handles each component as an independent ensemble of machine learning (ML) models to estimate both key parameter means and variances. Separate ML-based models are employed to estimate orientation, (un)directed velocity or distance from acceleration and gyroscope data, with optional absolute positioning from synchronized radio systems such as 5G for stabilization. A final fusion model combines these outputs, position, velocity, and orientation, while using uncertainty estimates to enhance system robustness. The modular design allows individual components to be updated, fine-tuned, or replaced without affecting the entire system. Experiments on dynamic sports movement data show that PDRNN achieves superior accuracy and precision compared to classic and ML-based methods, effectively avoiding error accumulation common in black-box approaches. And PDRNN offers forecast capabilities and better component control despite increased system complexity.\n\n---\n\n## 184. GQLA: Group-Query Latent Attention for Hardware-Adaptive Large Language Model Decoding\n- **Authors**: Fanxu Meng\n- **URL**: [https://arxiv.org/abs/2605.15250](https://arxiv.org/abs/2605.15250)\n- **Abstract**:\n> Multi-head Latent Attention (MLA), the attention used in DeepSeek-V2/V3, jointly compresses keys and values into a low-rank latent and matches the H100 roofline almost perfectly. Its trained weights, however, expose only one decoding path - an absorbed MQA form - which ties efficient inference to H100-class compute-bandwidth ratios, forfeits tensor parallelism along the head axis, and yields no Multi-Token Prediction (MTP) gain on commodity inference GPUs such as the export-restricted H20. We propose Group-Query Latent Attention (GQLA), a minimal modification of MLA whose trained weights expose two algebraically equivalent decoding paths over the same parameters: an MQA-absorb path identical to MLA's, and a GQA path with a per-group expanded cache. The runtime picks the path that matches the target hardware - no retraining, no custom kernels - so a single set of GQLA weights pins the rooflines of both H100 (MQA-absorb, s_q=1) and H20 (GQA + MTP, s_q=2), while supporting up to 8-way zero-redundancy tensor parallelism on the GQA path. To avoid pretraining from scratch we extend TransMLA into TransGQLA, which converts a pretrained GQA checkpoint into a GQLA model; on LLaMA-3-8B it compresses the per-token KV cache to 28.125% of the GQA baseline on the MQA-absorb path while structurally preserving GQA-level traffic on the per-group path.\n\n---\n\n## 185. Reading the Cell, Designing the Cure: Perturbation-Conditioned Molecular Diffusion for Function-Oriented Drug Design\n- **Authors**: Ziyu Xu , Zijian Zhang , Liang Wang , Zhiyuan Liu , Qiang Liu , Shu Wu , Liang Wang\n- **URL**: [https://arxiv.org/abs/2605.15243](https://arxiv.org/abs/2605.15243)\n- **Abstract**:\n> When reliable target structures are unavailable at scale or phenotypes arise from dysregulated pathways, transcriptomic perturbations provide a system-level functional readout for drug action. In this work, we formalize \\emph{Transcriptome-based Drug Design (TBDD)} as a generative inverse problem: designing drug molecules conditioned on desired transcriptomic state transitions. We analyze the inherently ill-posed nature of this task, which is further complicated by the profound domain gap between biology and chemistry and by the sparsity of transcriptomic signals. To address these challenges, we propose \\textbf{\\themodel{} } (A \\textbf{C}ell\\textbf{U}lar \\textbf{R}esponse \\textbf{E}ngine), a multi-resolution transcriptome-guided diffusion framework. \\themodel{} features a specialized \\textbf{Transcriptome Perturbation Functional Feature Extractor (TFE)} that (1) distills function-oriented perturbation embeddings from pre/post states, (2) aligns these signatures to dual chemical views to bridge the cross-modal gap, and (3) performs heterogeneity-aware aggregation to extract robust state-specific signals from noisy transcriptomic data. Extensive evaluations on both standard benchmarks and rigorous out-of-distribution protocols demonstrate that \\themodel{} consistently outperforms strong baselines in structural quality and functional consistency. Furthermore, we validate its practical utility via a zero-shot gene-inhibitor design task, highlighting the potential of phenotype-driven generative discovery.\n\n---\n\n## 186. Hydra: Efficient, Correct Code Generation via Checkpoint-and-Rollback Support\n- **Authors**: Alexander Du , Jianjun Ou , Danyang Zhuo , Matthew Lentz\n- **URL**: [https://arxiv.org/abs/2605.15238](https://arxiv.org/abs/2605.15238)\n- **Abstract**:\n> Large language models are increasingly used for code generation, but many generated programs fail to compile, a prerequisite for further correctness checks such as unit tests. Existing solutions for repairing static errors are costly in both latency and token consumption. Post-hoc repair delays error detection until generation completes and commonly regenerates large regions of previously valid code. Constrained semantic decoding checks after each token, incurring per-token overhead while limiting repair to the current token even when the root cause lies earlier. We present Hydra, a system for efficient recovery from static errors during code generation. Hydra allows checking to proceed asynchronously with generation, avoiding checker overhead when the generated code is semantically correct. In addition, it provides checkpoint-and-rollback support for targeted repair, avoiding regeneration and rechecking of valid prefixes. We retrofit the Clang C/C++ compiler to support Hydra with modest modifications. Paired with a token-efficient repair strategy, Hydra reduces latency by up to 71% and token consumption by up to 70% relative to post-hoc repair on C/C++ code generation tasks that encounter static errors.\n\n---\n\n## 187. A3D: Agentic AI flow for autonomous Accelerator Design\n- **Authors**: Abinand Nallathambi , Christopher Knight , Shantanu Ganguly , Wilfried Haensch , Anand Raghunathan\n- **URL**: [https://arxiv.org/abs/2605.15237](https://arxiv.org/abs/2605.15237)\n- **Abstract**:\n> Accelerating applications through the design of hardware accelerators can significantly enhance system performance and energy efficiency. Despite advances, such as high-level synthesis (HLS), designing accelerators for complex applications still remains highly labor-intensive, demanding considerable expertise in understanding workloads to be accelerated, hardware design, micro-architecture, and EDA tool usage, posing challenges for application domain experts. Therefore, most accelerator solutions are limited to applications with a regular predictable dataflow. Advances in AI have enabled agents that perform autonomous planning, reasoning, execution and reflection, leading to unprecedented potential for automation through agentic AI. We present A3D, an Agentic AI flow for end-to-end Automation of hardware Accelerator Design. A3D automates workload analysis, performance bottleneck identification, code refactoring for HLS compatibility and micro-architecture generation. A3D also generates diverse accelerator designs by automatically exploring the speed-area tradeoff space. Recent efforts have explored the use of AI for specific tasks such as design space exploration in HLS, leaving several tasks to still be performed manually. A3D addresses the challenges in applying modern LLMs to accelerator design by judiciously partitioning tasks among specialist agents, orchestrating process loops with specialist and verifier agents, utilizing pre-existing and custom tools, and employing agentic RAG for codebase and proprietary EDA tool documentation exploration. Our implementation of A3D, using commercial components like Claude Sonnet 4.5 and the Catapult HLS tool, demonstrates its effectiveness by generating accelerator designs with no human intervention from complex scientific applications like LAMMPS (molecular dynamics simulation) and QMCPACK (quantum chemistry).\n\n---\n\n## 188. Learning Selective Merge Policies for Deadline-Constrained Coded Caching via Deep Reinforcement Learning\n- **Authors**: Amirhossein Yousefiramandi\n- **URL**: [https://arxiv.org/abs/2605.15236](https://arxiv.org/abs/2605.15236)\n- **Abstract**:\n> With the coded caching, the server can use the information the users have cached to serve multiple users at a time by sending a single coded multi-casting message, i.e., the merged message, thereby relieving the peak network loads. However, for the delay-sensitive applications of the users, like the video streaming services, it becomes essential to choose which messages to merge online, considering the strict deadlines for each request. The problem, however, is that while the merge is helpful for the formation of the current coded multi-casting message, it can be harmful for the subsequent ones. We proposed a DRL-based solution that formulates the deadline-constrained coded delivery as a masked discrete-action queue-state control problem, while we trained a graph-attention policy network via proximal policy optimization. The policy network reduces the broadcast-packet expiration ratio $\\rho$ by $40.9%$ ($0.208$ vs. $0.352$) with respect to the best coded multi-casting baseline (SACM++) on the uniform-demand benchmark, while also attaining the best broadcast-efficiency score $\\sigma$ across the Track A battery among the coded multi-casting methods. The interesting fact we observed is that for the applications of the users with tight deadlines, the method of selective merging is better than the method of aggressive merging, i.e., the policy network learns to merge at only $\\approx 31.8%$ rate, even though the same observation holds across the variations within the same simulator family.\n\n---\n\n## 189. PBT-Bench: Benchmarking AI Agents on Property-Based Testing\n- **Authors**: Lucas Jing , Xinqi Wang , Liao Zhang , Simon S. Du\n- **URL**: [https://arxiv.org/abs/2605.15229](https://arxiv.org/abs/2605.15229)\n- **Abstract**:\n> Existing code benchmarks measure whether an agent can produce any test that reproduces a known bug, or whether it can produce a patch that fixes a described issue. Neither isolates the distinct skill of property-based testing: deriving a semantic invariant from documentation, and then constructing an input-generation strategy precise enough to make a random search reveal the violation. We introduce PBT-Bench, a benchmark of 100 curated property-based testing problems across 40 real Python libraries. Each problem injects one or more semantic bugs (365 in total, mean 3.65 per problem) designed so that default-strategy random inputs almost never trigger them; the agent must read the library's documentation, identify the relevant invariant, and specify a Hypothesis @given strategy that concentrates mass in the trigger region. Bugs are stratified across three difficulty levels (L1-L3) spanning single-constraint boundary bugs to stateful, cross-function protocol violations. We evaluate eight contemporary LLMs under two prompting regimes (open-ended baseline vs. explicit Hypothesis scaffolding) for three independent runs per configuration. Bug recall under the PBT-guided prompt ranges from 42.1% to 83.4% across models; under the open-ended baseline, from 31.4% to 76.7%. Hypothesis scaffolding lifts mid-capability models by over 20 percentage points, but yields smaller gains for the strongest models, with two exceptions showing degradation, suggesting the structured prompt can interfere with certain model behaviours rather than complementing them. The hardest bugs prove model-specific: different architectures fail on different problems, leaving persistent gaps that no single model closes. We release the benchmark, harness, and full evaluation corpus to support downstream work on documentation-grounded semantic reasoning.\n\n---\n\n## 190. Is Agentic AI Ready for Real-World Hardware Engineering? A Deep Dive with Phoenix-bench\n- **Authors**: Qingyun Zou , Feng Yu , Hongshi Tan , Bingsheng He , WengFai Wong\n- **URL**: [https://arxiv.org/abs/2605.15226](https://arxiv.org/abs/2605.15226)\n- **Abstract**:\n> We ask whether agentic AI systems built for software engineering transfer to realistic hardware engineering. Existing hardware LLM benchmarks isolate sub-tasks but none jointly requires repository navigation, hierarchy-aware localization, Electronic Design Automation (EDA) executable verification, and maintenance-style patching. We introduce \\textbf{Phoenix-bench}, a synchronized corpus of 511 verified Verilator instances from 114 GitHub repositories, each shipped with the developer patch, design-flow labels, fail-to-pass and pass-to-pass testbenches, and a Docker-pinned EDA environment so resolved-rate differences reflect agent behavior rather than toolchain availability. Using Phoenix-bench we run a uniform evaluation of four commercial agents and eight open-source agentic structures across four LLM backbones, plus two diagnostic interventions (file-level oracle localization and one round of testbench-log feedback). Three findings emerge. (i)~Software and hardware are fundamentally different engineering tasks: the same agent loses 37\\% to 58\\% from SWE-bench Verified to Phoenix-bench because hardware bugs propagate across parallel instantiated modules through signal flow rather than along a software-style call graph, and software-tuned agents stop at the symptom file instead of tracing back through the instantiation chain. (ii)~Failures concentrate on design control-flow / finite state machine (FSM) bugs, verification testbench bugs, and hard cases that demand cross-hierarchy signal-flow tracking and coordinated multi-file edits. (iii)~Localization granularity matters far more than localization itself: a perfect file-level oracle yields only $+1.4$\\% because the agent then breaks files that did not need editing, while a single round of test case feedback lifts resolved rate by $42$\\% to $45$\\% because the test case tells \\emph{where} the bug is and \\emph{what} the fix has to look like.\n\n---\n\n## 191. Do Biological Structural Guarantees Earn Their Complexity?\n- **Authors**: Bogdan Banu\n- **URL**: [https://arxiv.org/abs/2605.15225](https://arxiv.org/abs/2605.15225)\n- **Abstract**:\n> Biologically-inspired AI agent frameworks claim reliability benefits through structural guarantees adapted from gene regulatory networks, immune systems, and metabolic control. These claims are rarely tested empirically against simpler alternatives. We present three deep benchmarks: metabolic priority gating, autoinducer-based quorum sensing, and Bayesian stagnation detection, each comparing a biologically-grounded implementation against a naive non-biological alternative and an ablated control, across 1,000 trials per seed and 10 seeds (10M+ data points total).\n\n---\n\n## 192. GenAI-Driven Approach to RISC-V Supply Chain Exploration\n- **Authors**: Nenad Petrovic , Andre Schamschurko , Yingjie Xu , Alois Knoll\n- **URL**: [https://arxiv.org/abs/2605.15223](https://arxiv.org/abs/2605.15223)\n- **Abstract**:\n> This paper presents an LLM-empowered workflow for RISC-V supply chain analysis, integrating Vision-Language Models (VLMs) and Model-Driven Engineering (MDE) to enable comprehensive, multimodal data-driven insights. The proposed approach addresses the challenges of heterogeneous and unstructured supply chain data by leveraging LLMs for textual understanding and VLMs for extracting information from visual artifacts such as diagrams, tables, and scanned documents. These models collaboratively identify key entities and relationships, which are then organized into a knowledge graph representing supply chain components and their interdependencies. For analytical reasoning, the workflow incorporates MDE techniques and constraint-based modeling to enable formal validation of dependencies, detection of bottlenecks, and assessment of risks. The synergy between LLM- and VLM-based semantic understanding and MDE-based formal analysis supports both exploratory and systematic evaluation of supply chain resilience. A human-in-the-loop mechanism further enables interactive querying and expert validation. The approach is evaluated in RISC-V ecosystem scenarios, demonstrating its effectiveness in generating actionable insights, enhancing transparency, and supporting decision-making in complex semiconductor supply chains.\n\n---\n\n## 193. Effective Harness Engineering for Algorithm Discovery with Coding Agents\n- **Authors**: Yoichi Ishibashi , Taro Yano , Masafumi Oyamada\n- **URL**: [https://arxiv.org/abs/2605.15221](https://arxiv.org/abs/2605.15221)\n- **Abstract**:\n> AlphaEvolve and FunSearch have demonstrated the potential of combining large language models (LLMs) with evolutionary search for automated algorithm discovery. However, discovery success is shaped not only by model capability but also significantly by the design of the execution infrastructure, i.e., the harness. This paper investigates effective harness design through three questions: under a fixed token budget, is it better to produce many algorithms with brief thought or fewer algorithms with deeper thought? How should the harness handle evaluation hacks, where generated programs exploit the scoring function? And how can agents that require full filesystem access execute safely in parallel? Using Vesper, an algorithm discovery framework that incorporates harness improvements addressing these questions, we evaluate on Circle Packing under the same token budget. Interestingly, generating fewer algorithms while thinking more deeply about each one achieved higher scores. That is, scaling the quality of each individual is more budget-efficient than scaling the number of evolutionary generations. Surprisingly, more capable models produced evaluation hacks at higher rates, making hack detection increasingly necessary as models scale.\n\n---\n\n## 194. Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time\n- **Authors**: Michael Y. Hu , Apurva Gandhi , Kyunghyun Cho , Tal Linzen , Pratyusha Sharma\n- **URL**: [https://arxiv.org/abs/2605.15220](https://arxiv.org/abs/2605.15220)\n- **Abstract**:\n> Data mixing decides how to combine different sources or types of data and is a consequential problem throughout language model training. In pretraining, data composition is a key determinant of model quality; in continual learning and adaptation, it governs what is retained and acquired. Yet existing data mixing methods address only one phase of this lifecycle at a time: some require smaller proxy models tied to a single training phase, others assume a fixed domain set, and continual learning lacks principled guidance altogether. We argue that data mixing is fundamentally an online decision making problem -- one that recurs throughout training and demands a single, unified solution. We introduce OP-Mix (On-Policy Mix), a data mixing algorithm that operates across the entire language model training lifecycle. Our main insight is that candidate data mixtures can be cheaply simulated by interpolating between low-rank adapters trained directly on the current model, eliminating separate proxy models and ensuring the search is always grounded in the model's actual learning dynamics. Across pretraining, continual midtraining, and continual instruction tuning, OP-Mix consistently finds near-optimal mixtures while using a fraction of the compute of the baselines. In pretraining, OP-Mix improves upon training without mixing by 6.3% in average perplexity. For continual learning, OP-Mix matches the performance of both retraining and on-policy distillation while using 66% and 95% less overall compute, respectively. OP-Mix suggests a different view of language model training: not a sequence of distinct phases, but a single continuous process of learning from data.\n\n---\n\n## 195. An LLM-RAG Approach for Healthy Eating Index-Informed Personalized Food Recommendations\n- **Authors**: Yibin Wang , Yanjie Yang , Grace Melo Guerrero , Rodolfo M. Nayga Jr. , Azlan Zahid\n- **URL**: [https://arxiv.org/abs/2605.15213](https://arxiv.org/abs/2605.15213)\n- **Abstract**:\n> Diet quality is a leading determinant of chronic disease risk. Advances in artificial intelligence (AI) have enabled food recommendation systems to adapt suggestions to user preferences and health goals. However, most current systems rely on loosely curated food databases and provide limited connection to a validated index. In this study, we propose a Healthy Eating Index (HEI) informed retrieval-augmented generation (RAG) framework that combines standardized nutrition databases with large language models (LLMs) for personalized food recommendations. Our proposed method anchors retrieval in the National Health and Nutrition Examination Survey (NHANES) and the Food Patterns Equivalents Database (FPED). A food-level embedding space is constructed from FPED-derived textual descriptions. For each entity, the system computes baseline HEI scores, retrieves candidate foods for intake recommendations, and estimates the HEI impact of simple substitutions or additions. A constrained RAG pipeline instantiated with a pretrained OpenAI LLM generates personalized recommendations and sources based on nutrient profiles and HEI contributions. The simulation results showed a mean HEI improvement of 6.45, with the proportion of users HEI over 50 increasing from 45.12 to 61.26. Quantile analysis revealed consistent improved shifts across the HEI distribution. Our findings suggest that the proposed LLM-RAG-based AI systems can support more precise, explainable, and personalized nutrition guidance to improve diet quality.\n\n---\n\n## 196. Fault tolerance estimation in digital circuits with visualised generative networks\n- **Authors**: Sascha Biel , Carl Alexander Gaede , Amiel Glaser , Jan Wolter , Alexej Schelle\n- **URL**: [https://arxiv.org/abs/2605.15212](https://arxiv.org/abs/2605.15212)\n- **Abstract**:\n> We propose a new numerical method to estimate the fault tolerance of failure modes in digital circuit structures with a generative network sampling technique. From a random input of generated bitwise configurations of ideally digitalised analog currents in the digital circuit design with classical logical gates, expected output currents are compared to the realistic signals of a numerical experiment at the discriminator part of the Generative Adversarial Network (GAN) to calculate the deviation from ideal digital electronic signals, including various error modes, such as missing or interchanged logical devices. From the present analysis of a representation of the GAN in terms of complex variables, it is possible to evaluate the robustness in electronic designs by differentiating the impact of failure modes associated with different classical logical elements in the circuit.\n\n---\n\n## 197. Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels\n- **Authors**: Plawan Kumar Rath , Rahul Maliakkal\n- **URL**: [https://arxiv.org/abs/2605.15208](https://arxiv.org/abs/2605.15208)\n- **Abstract**:\n> Large Language Models are routinely compressed via post-training quantization to reduce inference costs and memory footprint for cloud and edge deployment, yet the impact of this compression on model quality remains poorly understood. Existing studies typically compare only two conditions (full-precision vs. a single quantized variant), rely on aggregate bias metrics, and evaluate a single model family, making it impossible to distinguish gradual degradation from threshold-dependent safety failures. We conduct a controlled empirical study of three instruction-tuned models (Qwen2.5-7B, Mistral-7B, Phi-3.5-mini) at five precision levels (BF16 through 3-bit) on 12,148 BBQ bias benchmark items across 5 random seeds, totaling 911,100 inference records. Our results reveal that 3-bit quantization causes 6-21% of previously unbiased items to develop new stereotypical behaviors, following a clear dose-response pattern confirmed via logistic regression, while models' willingness to select \"unknown\" answers declines by 17.4%. Crucially, these item-level changes are invisible to standard quality metrics: perplexity increases by less than 0.5% at 8-bit and under 3% at 4-bit across all three models, yet 2.5-5.6% of items already develop new biases at 4-bit. These findings demonstrate that aggregate evaluation metrics systematically miss fairness-critical degradation, underscoring the need for quality-aware compression protocols that explicitly test for bias emergence before deployment.\n\n---\n\n## 198. AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices\n- **Authors**: Dzung Pham , Kleomenis Katevas , Ali Shahin Shamsabadi , Hamed Haddadi\n- **URL**: [https://arxiv.org/abs/2605.15206](https://arxiv.org/abs/2605.15206)\n- **Abstract**:\n> Autonomous agents powered by large language models (LLMs) are increasingly used to automate complex, multi-step tasks such as coding or web-based question answering. While remote, cloud-based agents offer scalability and ease of deployment, they raise privacy concerns, depend on network connectivity, and incur recurring API costs. Deploying agents locally on user devices mitigates these issues by preserving data privacy and eliminating usage-based fees. However, agentic workflows are far more resource-intensive than typical LLM interactions. Iterative reasoning, tool use, and failure retries substantially increase token consumption, often expending significant compute without successfully completing tasks. In this work, we investigate the time, token, and energy overhead of locally deployed LLM-based agents on consumer hardware. Our measurements show that agentic execution increases GPU power draw, temperature, and battery drain compared to single-inference workloads. To address this inefficiency, we introduce AgentStop, a lightweight efficiency supervisor that predicts and preemptively terminates trajectories unlikely to succeed. Leveraging low-cost execution signals, such as token-level log probabilities, AgentStop can reduce wasted energy by 15-20% with minimal impact on task performance (<5% utility drop) for challenging web-based question answering and coding benchmarks. These findings position predictive early termination as a practical mechanism for enabling sustainable, privacy-preserving LLM agents on user devices. Our project code and data are available at this https URL .\n\n---\n\n## 199. Agent4POI: Agentic Context-Conditioned Affordance Reasoning for Multimodal Point-of-Interest Recommendation\n- **Authors**: Jinze Wang , Yangchen Zeng , Tiehua Zhang , Lu Zhang , Yuze Liu , Yongchao Liu , Xingjun Ma , Zhu Sun\n- **URL**: [https://arxiv.org/abs/2605.15203](https://arxiv.org/abs/2605.15203)\n- **Abstract**:\n> We introduce Agent4POI, the first POI recommendation framework that generates context-conditioned multimodal representations at recommendation time, rather than relying on static POI embeddings pre-computed independently of context. Existing multimodal systems encode each POI once as a static embedding, a design that precludes reasoning about why the same cafe affords solo work on Monday but group celebration on Friday evening. We formally prove that no pre-computed encoder can satisfy context-sensitive ranking under standard bilinear scoring, motivating inference-time item-side representation. Agent4POI inverts this computation: given a situational context, a four-phase LLM agent generates dynamic, context-specific affordance queries (Phase 1) and executes a five-step cross-modal chain-of-thought over image, review, and metadata evidence (Phase 2). The resulting uncertainty-aware affordance representation is grounded in Gibsonian affordance theory. These cross-modal verdicts form a structured, uncertainty-adjusted affordance representation (Phase 3), which is aligned with user preferences via a semantic caching system for low-latency ranking (Phase 4). On three POI benchmarks and three evaluation configurations (standard, cold-start, context-shift), Agent4POI achieves a 23.2% relative gain over the strongest baseline and degrades by only 7.5% under context-shift versus 16--17\\% for the strongest baselines. In cold-start scenarios, Agent4POI outperforms the best content-based baseline by up to 2.4x, whereas ID-based methods fail to generalize.\n\n---\n\n## 200. Ensuring Logic in the Fog: Sound POMDP Synthesis with LTL Objectives\n- **Authors**: Can Zhou , Yulong Gao , Pian Yu\n- **URL**: [https://arxiv.org/abs/2605.12581](https://arxiv.org/abs/2605.12581)\n- **Abstract**:\n> Synthesising autonomous agents that can navigate uncertain environments while adhering to complex temporal constraints remains a fundamental challenge. While Linear Temporal Logic (LTL) provides a rigorous language for specifying such tasks, the inherent undecidability of qualitatively verifying LTL satisfaction in partially observable Markov decision processes renders quantitative synthesis difficult, especially when designing reliable reward signals for approximate solvers. In this paper, we bridge this gap with a novel, sound reward-shaping mechanism that dynamically generates belief-dependent rewards grounded in certified LTL satisfaction. By integrating this mechanism into an enhanced Monte Carlo Planning framework, we empower agents to navigate the `fog' of partial observability with a search process focused on maximising verifiable success. Our experiments demonstrate that this approach not only thrives in scenarios where existing solvers fail but also maintains effectiveness and scalability across diverse benchmark domains.\n\n---\n\n{% endraw %}\n"}, "edd467a86c941fe5df7a48dbfa5a7a07f2952440": {"t": 1785770987.527317, "n": 340278, "h": "05c8c40e431f029e", "c": "---\nlayout: default\ntitle: \"전체 AI 논문 - 2026-05-18\"\n---\n\n{% raw %}\n# 전체 AI 논문 - 2026-05-18\n\n## 1. Prospective multi-pathogen disease forecasting using autonomous LLM-guided tree search\n- **Authors**: Sarah Martinson , Michael P. Brenner , Martyna Plomecka , Brian P. Williams , Nicholas G. Reich , Zahra Shamsi\n- **URL**: [https://arxiv.org/abs/2605.16238](https://arxiv.org/abs/2605.16238)\n- **Abstract**:\n> Probabilistic forecasting of infectious diseases is crucial for public health but relies on labor-intensive manual model curation by expert modeling teams. This bespoke development bottlenecks scalability to granular geographic resolutions or emerging pathogens. Here, we present an autonomous system using Large Language Model (LLM)-guided tree search to iteratively generate, evaluate, and optimize executable forecasting software. In a fully prospective, real-time evaluation during the 2025-2026 US respiratory season, the system autonomously discovered methodologically diverse models for influenza, COVID-19, and respiratory syncytial virus (RSV). Aggregating these machine-generated models yielded an ensemble that consistently matched or outperformed the gold-standard, human-curated Centers for Disease Control and Prevention (CDC) hub ensembles out-of-sample. The system successfully navigated data-scarce \"cold start\" scenarios for RSV. Moreover, controlled retrospective ablations revealed that optimizing log-scale distance metrics prevents reward hacking, while an automated judge-in-the-loop ensures structural fidelity to complex scientific theories. By autonomously translating epidemiological theory into accurate, transparent code, this framework overcomes the modeling labor bottleneck, enabling rapid deployment of expert-level disease forecasting at unprecedented scales.\n\n---\n\n## 2. FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast\n- **Authors**: Igor Bogdanov , Chung-Horng Lung , Thomas Kunz , Jie Gao , Adrian Taylor , Marzia Zaman\n- **URL**: [https://arxiv.org/abs/2605.16233](https://arxiv.org/abs/2605.16233)\n- **Abstract**:\n> Can LLM agents improve decision-making through self-generated memory without gradient updates? We propose FORGE (Failure-Optimized Reflective Graduation and Evolution), a staged, population-based protocol that evolves prompt-injected natural-language memory for hierarchical ReAct agents. FORGE wraps a Reflexion-style inner loop, where a dedicated reflection agent (using the same underlying LLM, no distillation from a stronger model) converts failed trajectories into reusable knowledge artifacts: textual heuristics (Rules), few-shot demonstrations (Examples), or both (Mixed), with an outer loop that propagates the best-performing instance's memory to the population between stages and freezes converged instances via a graduation criterion. We evaluate on CybORG CAGE-2, a stochastic network-defense POMDP at a 30-step horizon against the B-line attacker, where all four tested LLM families (Gemini-2.5-Flash-Lite, Grok-4-Fast, Llama-4-Maverick, Qwen3-235B) exhibit strongly negative, heavy-tailed zero-shot rewards. Compared against both a zero-shot baseline and a Reflexion baseline (isolated single-stream learning), FORGE improves average evaluation return by 1.7-7.7$\\times$ over zero-shot and by 29-72% over Reflexion in all 12 model-representation conditions, reducing major-failure rates (below $-100$) to as low as $\\sim$1%. We find that (1) population broadcast is critical mechanism, with a no-graduation ablation confirming that broadcast carries the performance gains while graduation primarily saves compute; (2) Examples achieves the strongest returns for three of four models, Rules offers the best cost-reliability profile with $\\sim$40% fewer tokens; and (3) weaker baseline models benefit disproportionately, suggesting FORGE may mitigate capability gaps rather than amplify strong models. All evidence is confined to CAGE-2 B-line; cross-family findings are directional evidence.\n\n---\n\n## 3. Fully Open Meditron: An Auditable Pipeline for Clinical LLMs\n- **Authors**: Xavier Theimer-Lienhard , Mushtaha El-Amin , Fay Elhassan , Sahaj Vaidya , Victor Cartier-Negadi , David Sasu , Lars Klein , Mary-Anne Hartley\n- **URL**: [https://arxiv.org/abs/2605.16215](https://arxiv.org/abs/2605.16215)\n- **Abstract**:\n> Clinical decision support systems (CDSS) require scrutable, auditable pipelines that enable rigorous, reproducible validation. Yet current LLM-based CDSS remain largely opaque. Most \"open\" models are open-weight only, releasing parameters while withholding the data provenance, curation procedures, and generation pipelines that determine model behavior. Fully Open (FO) models, which expose the complete training stack end-to-end, do not currently exist in medicine. We introduce Fully Open Meditron, the first fully open pipeline for building LLM-CDSS, comprising a clinician-audited training corpus, a reproducible data construction and training framework, and a use-aligned evaluation protocol. The corpus unifies eight public medical QA datasets into a normalized conversational format and expands coverage with three clinician-vetted synthetic extensions: exam-style QA, guideline-grounded QA derived from 46,469 clinical practice guidelines, and clinical vignettes. The pipeline enforces system-wide decontamination, gold-label resampling of teacher generations, and end-to-end validation by a four-physician panel. We evaluate using an LLM-as-a-judge protocol over expert-written clinical vignettes, calibrated against 204 human raters. We apply the recipe to five FO base models (Apertus-70B/8B-Instruct, OLMo-2-32B-SFT, EuroLLM-22B/9B-Instruct). All MeditronFO variants are preferred over their bases. Apertus-70B-MeditronFO improves +6.6 points over its base (47.2% to 53.8%) on aggregate medical benchmarks, establishing a new FO SoTA. Gemma-3-27B-MeditronFO is preferred over MedGemma in 58.6% of LLM-as-a-judge comparisons and outperforms it on HealthBench (58% vs 55.9%). These results show that fully open pipelines can achieve state-of-the-art domain-specific performance without sacrificing auditability or reproducibility.\n\n---\n\n## 4. Confirming Correct, Missing the Rest: LLM Tutoring Agents Struggle Where Feedback Matters Most\n- **Authors**: Tahreem Yasir , Wenbo Li , Sam Gilson , Sutapa Dey Tithi , Xiaoyi Tian , Tiffany Barnes\n- **URL**: [https://arxiv.org/abs/2605.16207](https://arxiv.org/abs/2605.16207)\n- **Abstract**:\n> Effective tutoring requires distinguishing optimal, valid but suboptimal, and incorrect student solutions, a distinction central to intelligent tutoring systems (ITS) but untested for LLM-based tutors. As LLMs are increasingly explored as conversational complements to ITS, evaluating their diagnostic precision is essential. We present a benchmark of seven LLM feedback agents in propositional logic using knowledge-graph-derived ground truth across 10,836 solution--feedback pairs and three feedback conditions. Models achieved near-ceiling performance on optimal steps but systematically over-rejected valid but suboptimal reasoning and over-validated incorrect solutions, precisely where adaptive tutoring matters most. These failures persisted across models regardless of solution context, suggesting architectural rather than informational limits. Moreover, accurate diagnosis did not reliably produce pedagogically actionable feedback, revealing a gap between diagnostic judgment and instructional effectiveness. Our findings suggest that LLMs are better suited for hybrid architectures where KG-grounded models handle diagnosis while LLMs support open-ended scaffolding and dialogue.\n\n---\n\n## 5. Context, Reasoning, and Hierarchy: A Cost-Performance Study of Compound LLM Agent Design in an Adversarial POMDP\n- **Authors**: Igor Bogdanov , Chung-Horng Lung , Thomas Kunz , Jie Gao , Adrian Taylor , Marzia Zaman\n- **URL**: [https://arxiv.org/abs/2605.16205](https://arxiv.org/abs/2605.16205)\n- **Abstract**:\n> Deploying compound LLM agents in adversarial, partially observable sequential environments requires navigating several design dimensions: (1) what the agent sees, (2) how it reasons, and (3) how tasks are decomposed across components. Yet practitioners lack guidance on which design choices improve performance versus merely increase inference costs. We present a controlled study of compound LLM agent design in CybORG CAGE-2, a cyber defense environment modeled as a Partially Observable Markov Decision Process (POMDP). Reward is non-positive, so all configurations operate in a failure-mitigation mode. Our evaluation spans five model families, six models, and twelve configurations (3,475 episodes) with token-level cost accounting. We vary context representation (raw observations vs. a deterministic state-tracking layer with compressed history), deliberation (self-questioning, self-critique, and self-improvement tools, with optional chain-of-thought prompting), and hierarchical decomposition (monolithic ReAct vs. delegation to specialized sub-agents). We find that: (1) Programmatic state abstraction delivers the largest returns per token spent (RPTS), improving mean return by up to 76% over raw observations. (2) Distributing deliberation tools across a hierarchy degrades performance relative to hierarchy alone for all five model families, reaching up to 3.4$\\times$ worse mean return while using 1.8-2.7$\\times$ more tokens. We call this destructive pattern a deliberation cascade. (3) Hierarchical decomposition without deliberation achieves the best absolute performance for most models, and context engineering is generally more cost-effective than deliberation. These findings suggest a design principle for structured adversarial POMDPs: invest in programmatic infrastructure and clean task decomposition rather than deeper per-agent reasoning, as these strategies can interfere when combined.\n\n---\n\n## 6. Formal Methods Meet LLMs: Auditing, Monitoring, and Intervention for Compliance of Advanced AI Systems\n- **Authors**: Parand A. Alamdari , Toryn Q. Klassen , Sheila A. McIlraith\n- **URL**: [https://arxiv.org/abs/2605.16198](https://arxiv.org/abs/2605.16198)\n- **Abstract**:\n> We examine one particular dimension of AI governance: how to monitor and audit AI-enabled products and services throughout the AI development lifecycle, from pre-deployment testing to post-deployment auditing. Combining principles from formal methods with SoTA machine learning, we propose techniques that enable AI-enabled product and service developers, as well as third party AI developers and evaluators, to perform offline auditing and online (runtime) monitoring of product-specific (temporally extended) behavioral constraints such as safety constraints, norms, rules and regulations with respect to black-box advanced AI systems, notably LLMs. We further provide practical techniques for predictive monitoring, such as sampling-based methods, and we introduce intervening monitors that act at runtime to preempt and potentially mitigate predicted violations. Experimental results show that by exploiting the formal syntax and semantics of Linear Temporal Logic (LTL), our proposed auditing and monitoring techniques are superior to LLM baseline methods in detecting violations of temporally extended behavioral constraints; with our approach, even small-model labelers match or exceed frontier LLM judges. Our predictive and intervening monitors significantly reduce the violation rates of LLM-based agents while largely preserving task performance. We further show through controlled experiments that LLMs' temporal reasoning shows a pronounced degradation in accuracy with increasing event distance, number of constraints, and number of propositions.\n\n---\n\n## 7. An Algebraic Exposition of the Theory of Dyadic Morality\n- **Authors**: Kush R. Varshney\n- **URL**: [https://arxiv.org/abs/2605.16153](https://arxiv.org/abs/2605.16153)\n- **Abstract**:\n> This paper provides an algebraic exposition of the theory of dyadic morality (TDM), a psychological model of moral judgment grounded in a simple two-node template: an intentional agent causing harm to a vulnerable patient. We formalize TDM using structural causal modeling (SCM) notation and identify three psychological operators (typecasting operator, completion operator, and valence-dependent inference mechanism) that extend standard SCM to capture how people compute moral judgments under constraints. We address scalability challenges arising from TDM's dyadic limitation, showing how moral cognition compresses multi-node scenarios through node collapse and sequential processing. Drawing on this algebraic framework, we demonstrate concrete applications to AI policy design: detecting conflicting obligations, structuring helpfulness policies to preserve user agency, and designing post-failure communication as causal interventions. Finally, we recommend scoped, contextual measurement of mind perception over universal averaging to operationalize the theory empirically. This algebraic formalization enables neurosymbolic AI systems to compute morality in a way that is both mathematically rigorous and faithful to human moral cognition.\n\n---\n\n## 8. Look Before You Leap: Autonomous Exploration for LLM Agents\n- **Authors**: Ziang Ye , Wentao Shi , Yuxin Liu , Yu Wang , Zhengzhou Cai , Yaorui Shi , Qi Gu , Xunliang Cai , Fuli Feng\n- **URL**: [https://arxiv.org/abs/2605.16143](https://arxiv.org/abs/2605.16143)\n- **Abstract**:\n> Large language model based agents often fail in unfamiliar environments due to premature exploitation: a tendency to act on prior knowledge before acquiring sufficient environment-specific information. We identify autonomous exploration as a critical yet underexplored capability for building adaptive agents. To formalize and quantify this capability, we introduce Exploration Checkpoint Coverage, a verifiable metric that measures how broadly an agent discovers key states, objects, and affordances. Our systematic evaluation reveals that agents trained with standard task-oriented reinforcement learning consistently exhibit narrow and repetitive behaviors that impede downstream performance. To address this limitation, we develop a training strategy that interleaves task-execution rollouts and exploration rollouts, with each type of rollout optimized by its corresponding verifiable reward. Building on this training strategy, we propose the Explore-then-Act paradigm, which decouples information-gathering from task execution: agents first utilize an interaction budget to acquire grounded environmental knowledge, then leverage it for task resolution. Our results demonstrate that learning to systematically explore is imperative for building generalizable and real-world-ready agents.\n\n---\n\n## 9. Property-Guided LLM Program Synthesis for Planning\n- **Authors**: Augusto B. Corrêa , André G. Pereira , Jendrik Seipp\n- **URL**: [https://arxiv.org/abs/2605.16142](https://arxiv.org/abs/2605.16142)\n- **Abstract**:\n> LLMs have shown impressive success in program synthesis, discovering programs that surpass prior solutions. However, these approaches rely on simple numeric scores to signal program quality, such as the value of the solution or the number of passed tests. Because a score offers no guidance on why a program failed, the system must generate and evaluate many candidates hoping some succeed, increasing LLM inference and evaluation costs. We study a different approach: property-guided LLM program synthesis. Instead of scoring programs after evaluation, we check whether a candidate satisfies a formally defined property. When the property is violated, we stop the evaluation early and provide the LLM with a concrete counterexample showing exactly how the program failed. This feedback drastically reduces both the number of program generations and the evaluation cost, and can guide the LLM to generate stronger programs. We evaluate this approach on PDDL planning domains, asking the LLM to synthesize direct heuristic functions: every state reachable by strictly improving transitions has a strictly improving successor. A heuristic with this property leads hill-climbing algorithm directly to a goal state. A counterexample-guided repair loop generates one candidate program, checks the property over a training set, and returns the first case that violates the property. We evaluate our approach on ten planning domains with an out-of-distribution test set. The synthesized heuristics are effectively direct on virtually all test tasks, and compared to the best prior generation method our approach generates seven times fewer programs per domain on average, solves more tasks without using search, and requires several orders of magnitude less computation to evaluate candidates. Whenever a problem admits a verifiable property, property-guided LLM synthesis can reduce cost and improve program quality.\n\n---\n\n## 10. ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents\n- **Authors**: Chinmay Savadikar , Mingyu Zhao , Yuanzheng Zhu , Han Li , Shuang Xie , Alberto Castelo , Tianfu Wu , Lingyun Wang\n- **URL**: [https://arxiv.org/abs/2605.16116](https://arxiv.org/abs/2605.16116)\n- **Abstract**:\n> Developing and evaluating e-commerce web agents requires environments that preserve meaningful task structure while enabling controllable, reproducible, and scalable scientific comparison. Existing methodologies force a tradeoff: live storefronts provide realism but are non-stationary, difficult to inspect, and irreproducible, while hand-built sandbox benchmarks provide control but cover only a narrow range of layouts, catalogs, policies, and interaction patterns. We argue that the core bottleneck is methodological: the field lacks a scalable way to construct evaluation settings that are simultaneously realistic, diverse, controllable, inspectable, and reproducible. We introduce ShopGym, an integrated framework for realistic simulation and scalable benchmarking of e-commerce web agents. ShopGym is a framework for constructing e-commerce simulation environments and grounded benchmark tasks. Its simulation layer, ShopArena, converts live seed storefronts into self-contained sandbox shops through anonymized shop specifications and a staged, validated generation process. On top of these simulated storefronts, ShopGuru synthesizes benchmark tasks across seven skill categories, grounding each task in the shop's catalog, navigation structure, policies, and interaction affordances. Together, ShopArena and ShopGuru produce self-contained, resettable, inspectable, and stable evaluation artifacts that preserve structural properties and agent-evaluation signals relevant to shopping tasks. We validate the framework through graph-based structural analysis and agent-based behavioral evaluation with 224 generated tasks across six sandbox shops: three constructed with synthetic data and three with real data. Our results show that the synthetic shops preserve key structural properties of live storefronts, with agent performance on synthetic shops positively correlated with performance on live storefronts.\n\n---\n\n## 11. Sign-Separated Finite-Time Error Analysis of Q-Learning\n- **Authors**: Donghwan Lee\n- **URL**: [https://arxiv.org/abs/2605.16103](https://arxiv.org/abs/2605.16103)\n- **Abstract**:\n> This paper develops a sign-separated finite-time error analysis for constant step-size Q-learning. Starting from the switching-system representation, the error is decomposed into its componentwise negative and positive parts. The negative part is dominated by a lower comparison linear time-invariant (LTI) system associated with a fixed optimal policy, whereas the positive part is controlled by a linear switching system. The resulting bounds show that the negative-side LTI certificate is no slower than the positive-side switching certificate and may produce a faster exponential envelope. The analysis identifies a max-induced asymmetry in Q-learning error dynamics. This asymmetry is connected to overestimation: positive action-wise errors can be selected and propagated by the Bellman maximum, whereas negative errors admit an optimal-policy lower comparison. Finite-time bounds are provided for both deterministic and stochastic constant-step-size recursions.\n\n---\n\n## 12. Reasoners or Translators? Contamination-aware Evaluation and Neuro-Symbolic Robustness in Tax Law\n- **Authors**: Parisa Kordjamshidi , Samer Aslan , Madhavan Seshadri , Leslie Barrett , Enrico Santus\n- **URL**: [https://arxiv.org/abs/2605.16052](https://arxiv.org/abs/2605.16052)\n- **Abstract**:\n> Recent advances in large language models (LLMs) have significantly enhanced automated legal reasoning. Yet, it remains unclear whether their performance reflects genuine legal reasoning ability or artifacts of data contamination. We present a comprehensive empirical study of tax law reasoning approaches and implement a contamination detection protocol to rigorously assess LLM reliability. We show that performance can be inflated by contamination. Building on this analysis, we conduct a systematic evaluation, comparing monolithic LLMs with hybrid systems that translate statutory text into formal representations and delegate inference to symbolic solvers. We build a novel test suite designed to probe generalization to unseen documents via case and rule variations. Our findings indicate that legal reasoning is inherently compositional and that neuro-symbolic frameworks offer a more reliable and robust foundation for legal AI, as well as improved generalization to unobserved situations.\n\n---\n\n## 13. ScreenSearch: Uncertainty-Aware OS Exploration\n- **Authors**: Michael Solodko , Justin Wagle\n- **URL**: [https://arxiv.org/abs/2605.16024](https://arxiv.org/abs/2605.16024)\n- **Abstract**:\n> Desktop GUI agents operate under partial observability: visually similar screens can correspond to different underlying workflow states, so locally plausible actions can lead to sharply different outcomes. We frame this as a problem of computer/OS state exploration, where effective behavior requires both expanding the reachable frontier and reducing ambiguity before committing. We present ScreenSearch, a system that combines structural screen retrieval and deduplication with an ambiguity-aware PUCT graph-bandit for large-scale desktop exploration. The retrieval layer converts UIA trees into location-aware structural features, indexes related screens through sparse token search and metadata filters, and maintains a shared deduplicated state graph across VM workers. On top of this graph, we define a scalable ambiguity signal based on matched-action outcome dispersion. If similar screens produce different next states under the same action signature, the state should be probed further rather than treated as resolved. We use this signal together with frontier rewards to drive large-scale exploration and replay-start policy evaluation over the shared graph. Across 11 desktop applications, ScreenSearch collects over 1M screenshots and over 30K deduplicated states, yielding large exploration corpora with substantial cross-application and within-application diversity. On a fixed replay-start slice, we observe a clear novelty--ambiguity trade-off: some policies reduce ambiguity quickly while discovering little frontier. Ambiguity reduction alone is therefore not a sufficient exploration objective. Appendix ablations show that stronger proposal priors can materially improve unique-state discovery during corpus building. These results suggest that state identity, proposal quality, and ambiguity-aware search all matter when deciding when to probe and when to commit.\n\n---\n\n## 14. Petri Net Induced Heuristic Search for Resource Constrained Scheduling\n- **Authors**: Ido Lublin , Dor Atzmon , Izack Cohen\n- **URL**: [https://arxiv.org/abs/2605.15983](https://arxiv.org/abs/2605.15983)\n- **Abstract**:\n> We formulate the Resource-Constrained Project Scheduling Problem (RCPSP) as optimal search over the reachability graph of a Timed Transition Petri Net with Resources, using relative-delay tokens so that scheduling decisions correspond to transition firings in the induced state space. We solve the resulting problem with $A^*$ guided by a heuristic that combines Critical Path and resource-based lower bounds, and prove that it is consistent under our token-based time semantics. Experiments on the PSPLIB benchmarks show that the approach outperforms strong exact Mixed-Integer Linear Programming (MIP) baselines (SCIP, CBC) in both success rate and solve time. Per-instance analysis shows that heuristic search and MIP degrade along independent axes, resource tightness for $A^*$ and formulation size for MIP, with resource strength mediating which solver benefits from scale.\n\n---\n\n## 15. Learning Bilevel Policies over Symbolic World Models for Long-Horizon Planning\n- **Authors**: Dillon Z. Chen , Till Hofmann , Toryn Q. Klassen , Sheila A. McIlraith\n- **URL**: [https://arxiv.org/abs/2605.15975](https://arxiv.org/abs/2605.15975)\n- **Abstract**:\n> We tackle the challenge of building embodied AI agents that can reliably solve long-horizon planning problems. Imitation learning from demonstrations has shown itself to be effective in training robots to solve a diversity of complex tasks requiring fine motor control and manipulation over low-level (LL), continuous environments. Yet, it remains a difficult endeavour to generate long-horizon plans from imitation learning alone. In contrast, high-level (HL), symbolic abstractions facilitate efficient and interpretable long-horizon planning. We propose to combine the strengths of LL imitation learning for manipulation and control, and HL symbolic abstractions for long-horizon planning. We realise this idea via \\emph{bilevel policies} of the form $(\\pi^{\\mathrm{hl} }, \\pi^{\\mathrm{ll} })$, consisting of a neural policy $\\pi^{\\mathrm{ll} }$ learned from LL demonstrations, and an HL symbolic policy $\\pi^{\\mathrm{hl} }$ that is constructed from symbolic abstractions of the LL demonstrations combined with inductive generalisation. We implement these ideas in the BISON system. Experiments on extended MetaWorld benchmarks demonstrate that BISON generalises to long horizons and problems with greater numbers of objects than those solved by VLA and end-to-end methods, and is more time and memory efficient in training and inference. Notably, when ignoring LL execution, BISON's HL policies can solve HL problems with 10,000 relevant objects in under a minute. Project page: this https URL\n\n---\n\n## 16. Deterministic Event-Graph Substrates as World Models for Counterfactual Reasoning\n- **Authors**: Fabio Rovai\n- **URL**: [https://arxiv.org/abs/2605.15967](https://arxiv.org/abs/2605.15967)\n- **Abstract**:\n> We study event-graph substrates: a class of world models that represent agent state as an append-only log of typed RDF triples and answer counterfactual queries by forking the log under a structured intervention vocabulary. Substrates are inspectable at the triple level, support exact counterfactuals, and transfer across domains without learned components. We formalize the class, prove a duality between explanatory and counterfactual queries that reduces both to the same causal-ancestor traversal, and evaluate a 1,400-line CLEVRER-DSL interpreter atop a domain-agnostic substrate runtime at full CLEVRER validation scale (n=75,618). The substrate exceeds the NS-DR symbolic oracle on all four per-question categories (by 9.89, 20.26, 17.65, and 0.80 percentage points), and exceeds the parametric ALOE baseline on descriptive and explanatory while lagging on predictive and counterfactual. We also introduce twin-EventLog, a 500-specification Park-canonical Smallville counterfactual benchmark on which the substrate exceeds Llama-3.1-8B with full context by 18.80 points joint accuracy.\n\n---\n\n## 17. PAGER: Bridging the Semantic-Execution Gap in Point-Precise Geometric GUI Control\n- **Authors**: Jingxuan Wei , Xi Bai , Shan Liu , Caijun Jia , Zheng Sun , Xinglong Xu , Siyuan Li , Linzhuang Sun , Bihui Yu , Conghui He , Cheng Tan\n- **URL**: [https://arxiv.org/abs/2605.15963](https://arxiv.org/abs/2605.15963)\n- **Abstract**:\n> Large vision-language models have significantly advanced GUI agents, enabling executable interaction across web, mobile, and desktop interfaces. Yet these gains largely rely on a forgiving region-tolerant paradigm, where many nearby pixels inside the same component remain valid. Precise geometric construction breaks this assumption: actions must land on points in continuous canvas space rather than tolerant regions. Because geometric primitives carry ontological dependencies, a local coordinate error can induce cascading topological failures that distort downstream objects and invalidate the final construction. We identify this regime as precision-sensitive GUI tasks, requiring point-level accuracy, geometry-aware verification, and robustness to dependency-driven error propagation. To benchmark it, we introduce PAGE Bench, with 4,906 problems and over 224K process-supervised, pixel-level GUI actions. We further propose PAGER, a topology-aware agent that decomposes construction into dependency-structured planning and pixel-level execution. Pixel-grounded supervised tuning establishes executable action grammar, while precision-aligned reinforcement learning mitigates rollout-induced exposure bias through state-conditioned geometric feedback. Experiments reveal a pronounced Semantic-Execution Gap: general multimodal models can exceed 88% action type accuracy yet remain below 6% task success. PAGER closes this gap, delivering 4.1x higher task success than the strongest evaluated general baseline and raising step success rate from below 9% for GUI-specialized agents to over 62%, establishing a new state of the art for point-precise GUI control.\n\n---\n\n## 18. Imperfect World Models are Exploitable\n- **Authors**: Logan Mondal Bhamidipaty (University of Edinburgh), Esmeralda S. Whitammer (University of Edinburgh), David Abel (University of Edinburgh), Mykel J. Kochenderfer (Stanford University), Subramanian Ramamoorthy (University of Edinburgh)\n- **URL**: [https://arxiv.org/abs/2605.15960](https://arxiv.org/abs/2605.15960)\n- **Abstract**:\n> We propose a novel definition of model exploitation in reinforcement learning. Informally, a world model is exploitable if it implies that one policy should be strictly preferred over another while the environment's true transition model implies the reverse. We analogize our definition with a prior characterization of reward hacking but show that the associated proof of inevitability does not transfer to exploitation. To overcome this obstruction, we develop a general theory of reward hacking and model exploitation that proves that exploitation is essentially unavoidable on large policy sets and yields the corresponding claim for hacking as a special case. Unfortunately, we also find that the conditions that guarantee unhackability in finite policy sets have no counterpart that precludes exploitation. Consequently, we introduce a relaxed notion of exploitation and derive a safe horizon within which it can be avoided. Taken together, our results establish a formal bridge between reward hacking and model exploitation and elucidate the limits of safe planning in world models.\n\n---\n\n## 19. Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design\n- **Authors**: Alberto Pepe , Chien-Yu Lin , Despoina Magka , Bilge Acun , Yannan Nellie Wu , Anton Protopopov , Carole-Jean Wu , Yoram Bachrach\n- **URL**: [https://arxiv.org/abs/2605.15871](https://arxiv.org/abs/2605.15871)\n- **Abstract**:\n> Toward recursive self-improvement, we investigate LLM agents autonomously designing foundation models beyond standard Transformers. We introduce a dual-framework approach: AIRA-Compose for high-level architecture search, and AIRA-Design for low-level mechanistic implementation. AIRA-Compose uses 11 agents to explore fundamental computational primitives under a 24-hour budget. Agents evaluate million-parameter candidates, extrapolating top designs to 350M, 1B, and 3B scales. This yields 14 architectures across two families: AIRAformers (Transformer-based) and AIRAhybrids (Transformer-Mamba). Pre-trained at 1B scale, these consistently outperform Llama 3.2 and Composer-found baselines. On downstream tasks, AIRAformer-D and AIRAhybrid-D improve accuracy by 2.4% and 3.8% over Llama 3.2. Furthermore, AIRA-Compose finds models with highly efficient scaling frontiers: AIRAformer-C scales 54% and 71% faster than Llama 3.2 and Composer's best Transformer, while AIRAhybrid-C outscales Nemotron-2 by 23% and Composer's best hybrid by 37%. AIRA-Design tasks 20 agents with writing novel attention mechanisms for long-range dependencies and high-performing training scripts. On the Long Range Arena benchmark, agent-designed architectures reach within 2.3% and 2.6% of human state-of-the-art on document matching and text classification. On the Autoresearch benchmark, Greedy Opus 4.5 achieves 0.968 validation bits-per-byte under a fixed time budget, surpassing the published minimum. Together, these frameworks show AI agents can autonomously discover architectures and algorithmic optimizations matching or surpassing hand-designed baselines. This establishes a powerful paradigm for discovering next-generation foundation models, marking a clear step toward recursive self-improvement.\n\n---\n\n## 20. SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows?\n- **Authors**: Kean Shi , Zihang Li , Tianyi Ma , Zengji Tu , Jialong Wu , Xinbo Xu , Qingyao Yang , Ruoyu Wu , Weichu Xie , Ming Wu , Jason Zeng , Michael Heinrich , Elvis Zhang , Liang Chen , Kuan Li , Baobao Chang\n- **URL**: [https://arxiv.org/abs/2605.15777](https://arxiv.org/abs/2605.15777)\n- **Abstract**:\n> Computer-Using Agents (CUAs) are rapidly extending large language models (LLMs) beyond text-based reasoning toward action execution in more complex environments, such as web browsers and graphical user interfaces (GUIs). However, existing web and GUI agent benchmarks often rely on simplified settings, isolated tasks, or short-horizon interactions, making it difficult to assess capabilities of agents in realistic professional workflows. Software-as-a-Service (SaaS) environments are a natural choice for CUA evaluation, as they host a large share of modern digital work and naturally involve dynamic system states, cross-application coordination, domain-specific knowledge, and long-horizon dependencies. To this end, we introduce SaaS-Bench, a benchmark built on 23 deployable SaaS systems across six professional domains, containing 106 tasks grounded in realistic work scenarios. These tasks require long-horizon execution, cover both text-only and multimodal settings, and are evaluated with weighted verification checkpoints that measure strict task completion and partial progress. Experiments show that representative LLM-based agents struggle on SaaS-Bench, with even the strongest model completing fewer than 4% of tasks end-to-end, exposing limitations in planning, state tracking, cross-application context maintenance, and error recovery. Code are available at this https URL for reproduction.\n\n---\n\n## 21. ALSO: Adversarial Online Strategy Optimization for Social Agents\n- **Authors**: Xiang Li , Liping Yi , Mingze Kong , Min Zhang , Zhongxiang Dai , QingHua Hu\n- **URL**: [https://arxiv.org/abs/2605.15768](https://arxiv.org/abs/2605.15768)\n- **Abstract**:\n> Social simulation provides a compelling testbed for studying social intelligence, where agents interact through multi-turn dialogues under evolving contexts and strategically adapting opponents. Such environments are inherently non-stationary, requiring agents to dynamically adjust their strategies over time. However, most Large Language Model (LLM) based social agents rely on static personas, while existing approaches for enhancing social intelligence, such as offline reinforcement learning or external planners, are ill-suited to these settings, typically assuming stationarity and incurring substantial training overhead. To bridge this gap, we propose \\textbf{ALSO} (\\textbf{A}dversarial on\\textbf{L}ine \\textbf{S}trategy \\textbf{O}ptimization), the first framework for online strategy optimization in multi-agent social simulation. ALSO advances social adaptation through two key contributions. (1) ALSO formulates multi-turn interaction as an adversarial bandit problem, where combinations of static personas and dynamic strategy instructions are treated as arms, providing a principled solution to non-stationarity without relying on environmental stability assumptions. (2) To predict rewards and generalize sparse feedback in multi-turn dialogues, ALSO introduces a lightweight neural surrogate to predict rewards from interaction histories, enabling sample-efficient exploration and continuous online adaptation. Experiments on the Sotopia benchmark demonstrate that ALSO consistently outperforms static baselines and existing optimization methods in dynamic environments, validating the effectiveness of adversarial online strategy optimization for building robust social agents.\n\n---\n\n## 22. Can We Trust AI-Inferred User States. A Psychometric Framework for Validating the Reliability of Users States Classification by LLMs in Operational Environments\n- **Authors**: Izabella Krzeminska , Michal Butkiewicz , Ewa Komkowska\n- **URL**: [https://arxiv.org/abs/2605.15734](https://arxiv.org/abs/2605.15734)\n- **Abstract**:\n> The use of large language models to assess user states in conversational and adaptive systems is based on the assumption that the metrics used for such assessment are stable and interpretable at the level of individual scores. This paper empirically tests this assumption, focusing on the psychometric reliability of artificial intelligence (AI) measures of user states. This study employed replication evaluation procedures to assess the repeatability of a broad set of metrics across three different bimodal large language models (GPT-4o audio, Gemini 2.0 Flash, Gemini 2.5 Flash). Analyses include both individual score reliability and aggregated reliability, allowing us to distinguish metrics potentially useful for real-time adaptation from those that retain their value only in aggregated analyses. The results demonstrate that metric reliability cannot be considered a default property in interpretive domains. The lack of stability at the level of individual scores precludes the interpretation of such scores as indicators of user state in real-time adaptive systems, even if these metrics demonstrate stability after aggregation. At the same time, the study indicates that individually unstable metrics can retain analytical utility in post-hoc studies, identifying rules governing interactions and their relationships with user experience parameters such as satisfaction, trust, and engagement. The main contribution of this work, besides quantifying the severity of the problem (only 31 of 213 metrics met the criteria), is the proposal of a replicable evaluation framework, enabling measurable evaluations of metric applicability. This approach supports more responsible AI design of adaptive systems, in which the interpretation of results requires explicit validation of reliability and monitoring for violations over time.\n\n---\n\n## 23. Nudging Beyond the Comfort Zone: Efficient Strategy-Guided Exploration for RLVR\n- **Authors**: Chanuk Lee , Sangwoo Park , Minki Kang , Sung Ju Hwang\n- **URL**: [https://arxiv.org/abs/2605.15726](https://arxiv.org/abs/2605.15726)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has emerged as a scalable paradigm for improving the reasoning capabilities of large language models. However, its effectiveness is fundamentally limited by exploration: the policy can only improve on trajectories it has already sampled. While increasing the number of rollouts alleviates this issue, such brute-force scaling is computationally expensive, and existing approaches that modify the optimization objective provide limited control over what is explored. In this work, we propose NudgeRL, a framework for structured and diversity-driven exploration in RLVR. Our approach introduces Strategy Nudging, which conditions each rollout on lightweight, strategy-level contexts to induce diverse reasoning trajectories without relying on expensive oracle supervision. To effectively learn from such structured exploration, we further propose a unified objective, which decomposes the reward signal into inter- and intra-context components and incorporates a distillation objective to transfer discovered behaviors back to the base policy. Empirically, NudgeRL outperforms standard GRPO with up to 8 times larger rollout budgets, while outperforming oracle-guided RL baseline on average across five challenging math benchmarks. These results demonstrate that structured, context-driven exploration can serve as an efficient and scalable alternative to both brute-force rollout scaling and feasibility-oriented methods based on privileged information. Our code is available at this https URL .\n\n---\n\n## 24. PRISM: Prompt Reliability via Iterative Simulation and Monitoring for Enterprise Conversational AI\n- **Authors**: Keshava Chaitanya , Jahnavi Gundakaram\n- **URL**: [https://arxiv.org/abs/2605.15665](https://arxiv.org/abs/2605.15665)\n- **Abstract**:\n> Deploying large language model (LLM)-driven conversational agents in enterprise settings requires prompts that are simultaneously correct at launch and resilient to the non-deterministic behavioral drift that characterizes production LLM deployments. Existing prompt optimization frameworks address prompt quality as a one-time compile-time problem, leaving open the equally critical question of how to detect and repair prompt regressions caused by silent LLM behavior changes over time. We present PRISM (Prompt Reliability via Iterative Simulation and Monitoring), a closed-loop framework that treats prompt engineering as a continuous reliability engineering problem rather than a one-time authorship task. PRISM takes as input plain-language agent requirements, a set of configured tools and memory variables, and an initial draft prompt. It automatically generates test cases from requirements, simulates full multi-turn conversations against a platform-faithful LLM environment, evaluates pass/fail using an LLM-as-judge, diagnoses root causes of failures, and surgically repairs the prompt -- iterating until all tests pass. Critically, PRISM is designed to run on a scheduled basis (daily), treating LLM behavioral drift as a first-class reliability concern. We evaluate PRISM across 35 enterprise conversational agents over a three-week deployment period on the this http URL V3 platform. PRISM reduces median prompt authoring time from 2 days to under 30 minutes, achieves 99% production reliability across all evaluated agents, and successfully identifies and repairs production regressions caused by LLM behavioral drift within a 24-hour detection window. Our results suggest that continuous, simulation-driven prompt optimization is both tractable and necessary for reliable enterprise conversational AI at scale.\n\n---\n\n## 25. ColPackAgent: Agent-Skill-Guided Hard-Particle Monte Carlo Workflows for Colloidal Packing\n- **Authors**: Lijie Ding , Changwoo Do\n- **URL**: [https://arxiv.org/abs/2605.15625](https://arxiv.org/abs/2605.15625)\n- **Abstract**:\n> We introduce ColPackAgent, an agent framework that autonomously runs Monte Carlo simulations of colloidal packing through a Model Context Protocol (MCP) tool server and an agent skill, whether as a standalone agent or inside an existing agent system. By harnessing the MCP server and agent skill, ColPackAgent executes a structured workflow for colloidal packing simulations, which are central to studies of phase behavior, self-assembly, and materials design. Without dedicated simulation tools and workflow instructions, general-purpose Large Language Model (LLM) agents tend to describe such workflows rather than execute them reliably. The MCP server exposes a custom-built colpack Python package that wraps HOOMD-blue hard-particle Monte Carlo, and the skill encodes a four-stage workflow contract. ColPackAgent can carry out the workflow interactively with human feedback, autonomously from an end-to-end prompt, or as autoresearch following a provided program file. We demonstrate the system in different modes with several colloidal packing simulation examples such as cube particles in 3D, a binary system of disks and capsules in 2D, and the 2D hard-disk freezing transition using autoresearch. We also compare model performance on this workflow across a panel of LLMs with 17 stage-specific prompts. This benchmark provides a stage-level check of how reliably different models follow the setup, planning, and analysis workflow. Together, these results show that pairing a domain Python package with MCP tools and a portable agent skill provides a practical route for turning a simulation toolkit into an agent-assisted research workflow.\n\n---\n\n## 26. TopoEvo: A Topology-Aware Self-Evolving Multi-Agent Framework for Root Cause Analysis in Microservices\n- **Authors**: Junle Wang , Xingchuang Liao , Wenjun Wu\n- **URL**: [https://arxiv.org/abs/2605.15611](https://arxiv.org/abs/2605.15611)\n- **Abstract**:\n> Root cause analysis (RCA) in microservices is challenging due to (i) noisy and heterogeneous multimodal observability (metrics, logs, traces), (ii) cascading failure propagation that amplifies downstream symptoms, and (iii) non-stationary topology drift induced by autoscaling and rolling updates. Recent LLM-based RCA agents can generate tool-grounded explanations, yet they often remain topology-agnostic and suffer from \\emph{symptom-amplification bias}, misattributing the root cause to salient downstream victims. We propose \\textbf{TopoEvo}, a topology-aware self-evolving multi-agent framework that couples graph representation learning with structured, topology-constrained reasoning. TopoEvo first introduces \\emph{Metric-orthogonal Multimodal Alignment} (MOMA), which decomposes metric embeddings into complementary subspaces and contrastively aligns logs and traces to reduce modality redundancy and sparsity, yielding stable node representations for graph encoding. It then applies \\emph{Vector Quantization} (VQ) to discretize topology-enhanced states into auditable \\emph{symptom tokens} with a symptom lexicon, enabling reliable retrieval and token-level evidence grounding. On top of these discrete topology cues, TopoEvo performs a multi-agent \\emph{Hypothesis--Evidence--Test} (HET) workflow to explicitly verify propagation-consistent explanations and separate initiating anomalies from amplified downstream symptoms. Finally, a \\emph{Self-Evolving Mechanism} refreshes hierarchical incident memory and performs conservative test-time adaptation with high-confidence pseudo-labels to maintain robustness under drift.\n\n---\n\n## 27. See Before You Code: Learning Visual Priors for Spatially Aware Educational Animation Generation\n- **Authors**: Yuejia Li , Ke He , Junheng Li , Shutong Chen , Jingkang Xia , Zhiyue Su , Junchi Zhang , Mang Ye\n- **URL**: [https://arxiv.org/abs/2605.15585](https://arxiv.org/abs/2605.15585)\n- **Abstract**:\n> Large language models can generate executable code for educational animations, but the resulting renders often exhibit visual defects, including element overlap, misalignment, and broken animation continuity. These defects cannot be reliably detected from the code alone and become apparent only after execution. We formalize this problem as render-feedback-aware constrained code generation: given a natural language specification, the model must generate executable code whose rendered output satisfies structured quality criteria that can be evaluated only after rendering. To address this problem, we introduce OmniManim, a render-feedback-aware educational animation generation framework built around a shared scene state, explicit visual planning, structured post-render diagnostics, and localized repair. Within OmniManim, the Vision Agent is a task-specific visual planning module: it predicts sparse keyframe layouts with coarse-to-fine bounding-box denoising and optimizes an interpolation-aware objective to reduce intermediate-frame failures induced by downstream animation interpolation. We further construct two datasets, ManimLayout-1K and EduRequire-500, and provide a reproducible evaluation protocol covering executability, instructional quality, visual quality, and efficiency. On EduRequire-500, OmniManim improves measured render quality over both single-model baselines and existing multi-agent frameworks. Systematic ablation studies further verify that explicit visual planning, especially its coarse spatial prior, bounding-box refinement, and interpolation-aware optimization, is central to these gains.\n\n---\n\n## 28. STAR: A Stage-attributed Triage and Repair framework for RCA Agents in Microservices\n- **Authors**: Junle Wang , Xingchuang Liao , Wenjun Wu\n- **URL**: [https://arxiv.org/abs/2605.15581](https://arxiv.org/abs/2605.15581)\n- **Abstract**:\n> LLM-based root cause analysis (RCA) agents have recently emerged as a promising paradigm for incident diagnosis in microservice AIOps. However, their reliability remains fragile: an error in early evidence collection, hypothesis formulation, or causal analysis can propagate through the reasoning trace and eventually corrupt the final diagnosis. In this paper, we present \\textbf{STAR}, a \\emph{Stage-attributed Triage and Repair} framework for repairing erroneous RCA traces. STAR explicitly decomposes an RCA workflow into four structured stages, namely \\emph{Evidence Package} (EP), \\emph{Hypothesis Set} (HS), \\emph{Analysis Structure} (AS), and \\emph{Decision Report} (DR), and treats agent failure as a stage-localizable reasoning bug rather than a monolithic end-to-end error. Built on top of LangGraph, STAR performs stage-wise auditing, budget-aware \\emph{Fast/Slow Routing}, \\emph{decisive stage localization via counterfactual candidate evaluation}, and stage-specific patch-and-replay repair. We evaluate STAR on a public large-scale benchmark and a real-world production dataset, using two RCA agent workflows and three foundation models. Experimental results show that STAR consistently improves both root cause localization and fault type classification over strong baselines. Moreover, STAR identifies the decisive faulty stage with high accuracy, repairs most initially incorrect traces within one or two replay rounds, and benefits substantially from both Fast/Slow Routing and counterfactual stage evaluation. These results suggest that explicitly modeling \\emph{where} an RCA agent fails is an effective path toward reliable, debuggable, and self-repairing agentic RCA systems.\n\n---\n\n## 29. Position: Artificial Intelligence Needs Meta Intelligence -- the Case for Metacognitive AI\n- **Authors**: Sergei Chuprov , Richard D. Lange , Leon Reznik , Paulo Shakarian , Raman Zatsarenko , Dmitrii Korobeinikov\n- **URL**: [https://arxiv.org/abs/2605.15567](https://arxiv.org/abs/2605.15567)\n- **Abstract**:\n> This position paper argues for metacognition as a general design principle for creating more accurate, secure, and efficient AI. The metacognitive solution involves systems monitoring their own states and judiciously allocating resources depending on each problem instance's difficulty or cost of mistakes. Drawing inspiration both from past work on resource-rational AI and from well-documented metacognitive strategies in psychology and cognitive science, we identify specific challenges in embedding these strategies into AI design and highlight open theoretical and implementation problems. We showcase these principles through a tangible example of improved learning efficiency, effectiveness, and security in a Federated Learning (FL) case study. We show how these principles can be translated into practice with a novel software framework developed specifically to allow the community to design, deploy, and experiment with metacognition-enabled AI applications.\n\n---\n\n## 30. DRS-GUI: Dynamic Region Search for Training-Free GUI Grounding\n- **Authors**: Yichao Liu , Huawen Shen , Liu Yu , Shiyu Liu , Zeyu Chen , Yu Zhou\n- **URL**: [https://arxiv.org/abs/2605.15542](https://arxiv.org/abs/2605.15542)\n- **Abstract**:\n> GUI agents powered by Multimodal Large Language Models (MLLMs) have demonstrated impressive capability in understanding and executing user instructions. However, accurately grounding instruction-relevant elements from high-resolution screenshots cluttered with irrelevant UI components remains challenging for existing approaches. Inspired by how humans dynamically adjust their perceptual scope to locate task-related regions on complex screens, we propose DRS-GUI, a training-free dynamic region search framework for GUI grounding that can be seamlessly integrated into existing MLLMs. DRS-GUI introduces a lightweight UI Perceptor that performs three human-like perceptual actions (Focus, Shift, and Scatter) to progressively explore the interface and generate region proposals. To dynamically schedule these actions, we further design an Action Planner based on Monte Carlo Tree Search (MCTS). A region quality reward is employed to evaluate and select the highly instruction-relevant region, efficiently pruning redundant UI elements. Experiments demonstrate that DRS-GUI yields a 14\\% improvement on ScreenSpot-Pro for general and GUI-specific MLLMs (Qwen2.5-VL-7B and UGround-V1-7B), significantly enhancing grounding performance and generalization.\n\n---\n\n## 31. RTL-BenchMT: Dynamic Maintenance of RTL Generation Benchmark Through Agent-Assisted Analysis and Revision\n- **Authors**: Jing Wang , Shang Liu , Hangan Zhou , Zhiyao Xie\n- **URL**: [https://arxiv.org/abs/2605.15537](https://arxiv.org/abs/2605.15537)\n- **Abstract**:\n> This paper introduces RTL-BenchMT, an agentic framework for dynamically maintaining RTL generation benchmarks. Large Language Models (LLMs) assisted automated RTL generation is one of the most important directions in EDA research. However, current RTL benchmarks face two critical challenges: (1) flawed cases in the benchmarks and (2) overfitting to the benchmarks. Both challenges are difficult to resolve purely by manual engineering effort. To address these issues and systematically reduce human maintenance costs, we propose an automated agentic framework, RTL-BenchMT. RTL-BenchMT focuses on two key applications: (1) automatically identifying and revising flawed benchmark cases and (2) automatically detecting and updating overfitting cases. With the assistance of RTL-BenchMT, we conduct a thorough, in-depth analysis of flawed and overfitting cases and produce a refined benchmark suite that will be open-sourced to the community.\n\n---\n\n## 32. CAPS: Cascaded Adaptive Pairwise Selection for Efficient Parallel Reasoning\n- **Authors**: Fangzhou Lin , Shuo Xing , Peiran Li , Siyuan Yang , Qianwen Ge , Kazunori Yamada , Ziming Zhang , Haichong Zhang , Zhengzhong Tu\n- **URL**: [https://arxiv.org/abs/2605.15513](https://arxiv.org/abs/2605.15513)\n- **Abstract**:\n> Parallel reasoning, where a generator samples many candidate solutions and an aggregator selects the best, is one of the most effective forms of test-time scaling in large language models, and pairwise self-verification has become its strongest aggregation primitive. Yet pairwise verification carries a heavy cost: each judgment reads two complete solutions in full, and existing methods perform tens of such judgments per problem regardless of whether the comparison is informative. We introduce CAPS (Cascaded Adaptive Pairwise Selection), an inference-only framework that allocates verifier compute non-uniformly along two orthogonal axes: an evidence axis that adapts how much of each candidate the judge sees, and a distribution axis that adapts how comparisons are spread across the pool. CAPS instantiates these into a four-stage cascade with an optional rescue subroutine, and admits a closed-form verifier-token cost in which the per-candidate marginal cost is roughly halved relative to uniform full-evidence schedules. On four self-verifying models (Qwen3-14B, GPT-OSS-20B, Qwen3-4B-Instruct/Thinking) and five reasoning benchmarks spanning code (LiveCodeBench-v5/v6, CodeContests) and math (AIME 2025, HMMT 2025), CAPS outperforms the leading pairwise verifier on 14 of 20 suites while using 25.4% of its verifier-token budget on code, and outperforms pointwise self-verification on all 20. The trade-off suites admit an interpretable diagnostic in terms of the verifier's accuracy at partial versus full evidence, providing a concrete pre-deployment check for cascade suitability.\n\n---\n\n## 33. X-SYNTH: Beyond Retrieval -- Enterprise Context Synthesis from Observed Human Attention\n- **Authors**: Guruprasad Raghavan , George Nychis , Rohan Narayana Murthy\n- **URL**: [https://arxiv.org/abs/2605.15505](https://arxiv.org/abs/2605.15505)\n- **Abstract**:\n> In enterprise operations, the context required for an AI agent task is scattered across systems of record, static information stores, and communication channels. What is stored is system state, a lossy representation of the work that actually happened [2, 52]. The prevailing approach [17, 31, 34, 36] retrieves by matching request content to what is stored; for narrow requests this works well. But synthesis quality depends on knowing what to surface and how to interpret it: knowledge specific to each organization, team, and individual [5, 57, 61], present in behavioral patterns, absent from any retrieval index. For complex agentic tasks it breaks down: True Lead Rate is low, False Lead Rate is high, and the model has no mechanism to improve. We present X-SYNTH, a framework for enterprise context synthesis grounded in human attention, the digitally observable interaction signatures of each worker, encoding not just what they did but the sequence in which they did it, along with implicit reward signals. Behavioral traces preceding positive outcomes are distinguishable from those that did not, without external labeling. X-SYNTH models each individual's behavioral baseline as a Digital Twin Signature (DTS) and selects among seven qualitatively distinct attention filters: Proportional, Inverse, Differential, Recurrent, Comparative, Sequential, and Collective, per individual and per query, to identify causally relevant activity signatures. A four-stage pipeline assembles ranked context grounded in behavioral patterns rather than query embeddings. On a sales lead identification task, a frontier model unaided achieves 9.5% True Lead Rate (TLR) with 90.5% False Lead Rate (FLR). Augmented with X-SYNTH, TLR rises to 61.9% (6.5x) while FLR falls to 18.8%. Enterprise context synthesis is not a retrieval problem. It is a relevance problem, and human attention is its most reliable ground truth.\n\n---\n\n## 34. From LLM-Generated Conjectures to Lean Formalizations: Automated Polynomial Inequality Proving via Sum-of-Squares Certificates\n- **Authors**: Ruobing Zuo , Hanrui Zhao , Gaolei He , Zhengfeng Yang , Jianlin Wang\n- **URL**: [https://arxiv.org/abs/2605.15445](https://arxiv.org/abs/2605.15445)\n- **Abstract**:\n> Automated proving of polynomial inequalities is a fundamental challenge in automated mathematical reasoning, where rich algebraic structure and a rapidly growing certificate search space hinder scalability. Purely symbolic approaches provide strong guarantees but often scale poorly as the number of variables or the degree increases, due to expensive algebraic manipulations and rapidly growing intermediate expressions. In parallel, LLM-guided methods have made notable progress, particularly on competition-style inequalities with a small number of variables. To address the remaining scalability challenges, we propose NSPI, a neuro-symbolic framework that combines the complementary strengths of LLMs and symbolic computation for polynomial-inequality proving. Concretely, an LLM proposes a conjecture in the form of an approximate polynomial Sum-Of-Squares (SOS) decomposition; we refine it via symbolic computation to obtain an exact polynomial SOS representation, which directly proves the target inequality, and we further certify the proof in Lean, yielding an end-to-end pipeline from heuristic discovery to machine-checked proof. Experiments on challenging benchmarks involving polynomials with up to 10 variables demonstrate the effectiveness and scalability of the proposed method.\n\n---\n\n## 35. Beyond Partner Diversity: An Influence-Based Team Steering Framework for Zero-Shot Human-Machine Teaming\n- **Authors**: Wei Sheng , Rohan Paleja\n- **URL**: [https://arxiv.org/abs/2605.15400](https://arxiv.org/abs/2605.15400)\n- **Abstract**:\n> While AI agents are rapidly advancing from isolated tools to interactive collaborators, data-driven human-machine teaming (HMT) methods remain costly in their reliance on human interaction data across domains, teammates, and team sizes. Zero-shot coordination (ZSC) addresses this bottleneck by simulating diverse partner populations to approximate how unseen partners might behave. However, partner coverage alone is insufficient as team settings scale and communication becomes degraded. To remedy this deficiency, we propose Influence-Based Team Steering (IBTS), a framework that uses influence shaping to incentivize agents to discover diverse, high-performing team interaction patterns and further steers ongoing trajectories toward stronger learned coordination modes. We assess IBTS on Overcooked-AI in both two-agent and three-agent settings, allowing us to test whether learned coordination structure transfers beyond dyadic interaction. Our evaluation includes simulated partners, synthetic partner-style variation, and, to our knowledge, the first 30-subject Overcooked-AI HMT study involving two real human teammates and one machine teammate. Across these evaluations, IBTS improves team performance against competing baselines, highlighting the need for scaled ZSC to combine sparse-reward coordination mechanisms with partner-variation coverage rather than relying on diversity alone.\n\n---\n\n## 36. Ensemble Monitoring for AI Control: Diverse Signals Outweigh More Compute\n- **Authors**: Eugene Koran , Yejun Yun , Samantha Tetef , Benjamin Arnav , Pablo Bernabeu-Pérez\n- **URL**: [https://arxiv.org/abs/2605.15377](https://arxiv.org/abs/2605.15377)\n- **Abstract**:\n> As AI systems are increasingly deployed in autonomous agentic settings at scale, it is important to ensure the actions they take are safe and aligned with user intent. Monitoring agent actions is a key safety mechanism, yet reliable monitors remain difficult to build and the scale of these systems makes human oversight impractical. We show that combining signals from diverse monitors into an ensemble improves detection of misaligned actions. We build 12 GPT-4.1-Mini monitors using both prompting and fine-tuning strategies. We evaluate them on coding tasks where candidate solutions pass standard tests but fail on adversarial inputs. In this setting, diverse ensembles outperform both individual monitors and homogeneous ensembles. Our best 3-monitor ensemble achieves 2.4x greater detection performance gain compared to an ensemble composed of three identical monitors, with the same ensemble performing strongly on an independent dataset. We contend that these results show that diversity - not scale - drives gains. The best ensembles combine strong individual performance with low correlation between monitors. Furthermore, fine-tuned monitors appear in every top-performing ensemble and maintain this advantage on out-of-distribution attack types, suggesting that fine-tuning enables detection capabilities that prompting alone does not elicit. These results support ensemble monitoring as a practical AI control strategy for safety gains at reasonable inference costs.\n\n---\n\n## 37. Belief Engine: Configurable and Inspectable Stance Dynamics in Multi-Agent LLM Deliberation\n- **Authors**: Joshua C. Yang , Maurice Flechtner , Damian Dailisan , Michiel A. Bakker\n- **URL**: [https://arxiv.org/abs/2605.15343](https://arxiv.org/abs/2605.15343)\n- **Abstract**:\n> LLM-based agents are increasingly used to simulate deliberative interactions such as negotiation, conflict resolution, and multi-turn opinion exchange. Yet generated transcripts often do not reveal why an agent's stance changes: movement may reflect evidence uptake, anchoring, role drift, echoing, or changed prompt and retrieval context. We introduce the Belief Engine (BE), an auditable belief-update layer that treats \"belief\" as an evidential state over a proposition and exposes it as scalar stance. BE extracts arguments into structured memory and updates stance with a log-odds rule controlled by evidence uptake u and prior anchoring a. Across multiple base LLMs, parameter sweeps show that these controls reliably shape stance dynamics while preserving an evidence-level update trail. On DEBATE, a human deliberation dataset with pre/post opinions, BE best reconstructs participants whose final stance follows extracted evidence; stable and evidence-opposed cases instead point to anchoring or factors outside the extracted evidence stream. BE provides configurable infrastructure for studying evidence-grounded deliberation, where openness, commitment, convergence, and disagreement can be tied to explicit update assumptions rather than hidden prompt effects.\n\n---\n\n## 38. Zero-Shot Goal Recognition with Large Language Models\n- **Authors**: Kin Max Piamolini Gusmão , Nathan Gavenski , Nir Oren , Felipe Meneguzzi\n- **URL**: [https://arxiv.org/abs/2605.15333](https://arxiv.org/abs/2605.15333)\n- **Abstract**:\n> Large language models have recently reached near-parity with classical planners on well-known planning domains, yet this competence relies on world-knowledge exploitation rather than genuine symbolic reasoning. Goal recognition is a complementary abductive task structurally better suited to LLM strengths: it consists of evaluating consistency with world knowledge rather than generating novel action sequences. This paper provides the first systematic zero-shot evaluation of frontier LLMs as goal recognisers on key classical PDDL benchmarks. Our results show that LLM competence on goal recognition is uneven: some models scale with evidence and approach landmark-based accuracy at full observations, while others remain anchored to world-knowledge priors regardless of how much evidence accumulates. Qualitative analysis of model reasoning traces reveals that this divergence reflects a fundamental difference in evidence integration rather than domain familiarity. These findings position goal recognition as a principled benchmark for the foundational planning knowledge of LLMs.\n\n---\n\n## 39. Context Pruning for Coding Agents via Multi-Rubric Latent Reasoning\n- **Authors**: Jingjing Wang , Xiwen Chen , Wenhui Zhu , Huayu Li , Zhengxiao He , Feiyang Cai , Ana S. Carreon-Rascon , Xuanzhao Dong , Feng Luo\n- **URL**: [https://arxiv.org/abs/2605.15315](https://arxiv.org/abs/2605.15315)\n- **Abstract**:\n> LLM-powered coding agents spend the majority of their token budget reading repository files, yet much of the retrieved code is irrelevant to the task at hand. Existing learned pruners compress this context with a single-objective sequence labeler, collapsing all facets of code relevance into one score and one transition matrix. We show that this formulation creates a modeling bottleneck: a single CRF transition prior must serve heterogeneous retention patterns, including contiguous semantic spans and sparse structural support lines. We propose LaMR (Latent Multi-Rubric), a structured pruning framework that decomposes code relevance into two interpretable quality dimensions, semantic evidence and dependency support, each modeled by a dedicated CRF with dimension-specific transition dynamics. A mixture-of-experts gating network dynamically weights the per-rubric emissions conditioned on the query, and a final CRF layer on the fused emissions produces the aggregate keep-or-prune decision. To supervise each dimension without additional annotation cost, we derive multi-rubric labels from the existing training corpus via AST-based program analysis, simultaneously denoising the teacher's binary labels. By effectively filtering distracting noise, LaMR frequently matches or even outperforms unpruned full-context baselines. Experiments on four benchmarks (SWE-Bench Verified, SWE-QA, LCC, LongCodeQA) show that LaMR wins 12 of 16 head-to-head multi-turn comparisons. It saves up to 31% more tokens on multi-turn agent tasks and improves Exact Match by up to +3.5 on single-turn tasks, while performance is frequently enhanced by denoising the context, and any remaining drops are marginal.\n\n---\n\n## 40. SMCEvolve: Principled Scientific Discovery via Sequential Monte Carlo Evolution\n- **Authors**: Jiachen Jiang , Huminhao Zhu , Zhihui Zhu\n- **URL**: [https://arxiv.org/abs/2605.15308](https://arxiv.org/abs/2605.15308)\n- **Abstract**:\n> LLM-driven program evolution has emerged as a powerful tool for automated scientific discovery, yet existing frameworks offer no principled guide for designing their individual components and provide no guarantee that the search converges. We introduce SMCEvolve, which recasts program search as sampling from a reward-tilted target distribution and approximates it with a Sequential Monte Carlo (SMC) sampler. From this view, three core mechanisms emerge as principled components: adaptive parent resampling, mixture of mutation with acceptance, and automatic convergence control. We further provide a finite-sample complexity analysis that bounds the LLM-call budget required to reach a target approximation error. Across math, algorithm efficiency, symbolic regression, and end-to-end ML research benchmarks, SMCEvolve surpasses state-of-the-art evolving systems while using fewer LLM calls under self-determined termination. The code is available at this https URL .\n\n---\n\n## 41. Solvita: Enhancing Large Language Models for Competitive Programming via Agentic Evolution\n- **Authors**: Han Li , Jinyu Tian , Rili Feng , Yuqiao Du , Chong Zheng , Chenyu Wang , Chenchen Liu , Shihao Li , Xinping Lei , Yifan Yao , Weihao Xie , Letian Zhu , Jiaheng Liu\n- **URL**: [https://arxiv.org/abs/2605.15301](https://arxiv.org/abs/2605.15301)\n- **Abstract**:\n> Large language models (LLMs) still struggle with the rigorous reasoning demands of hard competitive programming. While recent multi-agent frameworks attempt to bridge this reliability gap, they remain fundamentally stateless: they rely on static retrieval and discard the valuable problem-solving and debugging experience gained from previous tasks. To address this, we present Solvita, an agentic evolution framework that enables continuous learning without requiring weight updates to the underlying LLM. Solvita reorganizes problem-solving into a closed-loop system of strategy selection, program synthesis, certified supervision, and targeted hacking, executed by four specialized agents: Planner, Solver, Oracle, and Hacker. Crucially, each agent is paired with a trainable, graph-structured knowledge network. As the system operates, outcome signals, such as pass/fail verdicts, test certification quality, and adversarial vulnerabilities discovered by the Hacker, are recast as reinforcement learning updates to these network weights. This allows the agents to dynamically route future queries based on past successes and failures, effectively accumulating transferable reasoning experience over time. Evaluated across CodeContests, APPS, AetherCode, and live Codeforces rounds, Solvita establishes a new state-of-the-art among code-generation agents, outperforming existing multi-agent pipelines and nearly doubling the accuracy of single-pass baselines.\n\n---\n\n## 42. Verifiable Agentic Infrastructure: Proof-Derived Authorization for Sovereign AI Systems\n- **Authors**: Jun He , Deying Yu\n- **URL**: [https://arxiv.org/abs/2605.15228](https://arxiv.org/abs/2605.15228)\n- **Abstract**:\n> Modern cloud and enterprise systems rely on identity-centric authorization, assuming that callers possessing valid credentials are safe to execute commands. The emergence of autonomous AI agents invalidates this assumption: agents can generate syntactically valid but semantically unsafe actions, making standing privileges a significant operational risk. This risk becomes especially acute in sovereign AI systems, where autonomous agents may interact with cloud infrastructure, regulated data, financial workflows, and national-scale digital services. Governed mutation substrates reduce this risk by interposing on agent actions: agents submit intents, infrastructure evaluates context and policy, and execution is mediated. However, this shifts the trust boundary: how can the decision to authorize an intent be made verifiable, distributed, and replayable? We introduce a Distributed Trust Framework (DTF), a verification framework for governed mutation systems that computes execution authority from structured, verifiable artifacts. DTF introduces a Justification Proof to encode the admissibility basis of an action, a consensus model for independent evaluation, an ephemeral Execution Identity derived from the approved proof, and an append-only Evidence Chain that preserves the authorization lifecycle. Under stated substrate assumptions, this architecture enforces a compact authorization invariant: no high-stakes execution without a proof object, no derived authority without consensus, and no valid mutation detached from evidence. We define the model, instantiate it over an OpenKedge-based governed mutation substrate, and show how it maps onto cloud-native environments. By shifting authorization from standing identity to proof-derived authority, DTF provides an infrastructure foundation for making agentic execution governable, auditable, and bounded in sovereign AI deployments.\n\n---\n\n## 43. NIMO Controller: a self-driving laboratory orchestrator based on the Model Context Protocol\n- **Authors**: Naruki Yoshikawa , Ryo Tamura\n- **URL**: [https://arxiv.org/abs/2605.15227](https://arxiv.org/abs/2605.15227)\n- **Abstract**:\n> Self-driving laboratories (SDLs) have attracted increasing attention as a means of accelerating scientific discovery; however, developing SDL software remains technically demanding. To improve accessibility, orchestration software frameworks have been proposed to coordinate SDL components. Nevertheless, existing frameworks are primarily designed for human interaction and do not provide standardized interfaces suitable for AI agents. In this work, we propose an SDL software architecture based on the Model Context Protocol (MCP), in which all SDL functionalities are exposed through MCP servers. Following this design principle, we introduce an MCP-based SDL orchestrator, named NIMO Controller. It provides a visual programming interface automatically generated through MCP-based tool discovery, allowing human users to design experimental workflows without writing code. The same MCP backend can also be accessed by AI agents, providing a unified interface for both human users and AI agents. We demonstrate the proposed system through a case study on a color-matching SDL. The results validate the usability of the proposed MCP-based SDL architecture.\n\n---\n\n## 44. ICRL: Learning to Internalize Self-Critique with Reinforcement Learning\n- **Authors**: Jianbo Lin , Xiaomin Yu , Yi Xin , Yifu Guo , Zhuosong Jiang , Zhongqi Yue , Weishi Wang , Heqing Zou , Chengwei Qin , Hui Xiong\n- **URL**: [https://arxiv.org/abs/2605.15224](https://arxiv.org/abs/2605.15224)\n- **Abstract**:\n> Large language model-based agents make mistakes, yet critique can often guide the same model toward correct behavior. However, when critique is removed, the model may fail again on the same query, indicating that it has not internalized the critique's guidance into its underlying capability. Meanwhile, a frozen critic cannot improve its feedback quality over time, limiting the potential for iterative self-improvement. To address this, we propose learning to internalize self-critique with reinforcement learning(ICRL), a novel framework that jointly trains a solver and a critic from a shared backbone to convert critique-induced success into unassisted solver ability. The critic is rewarded based on the solver's subsequent performance gain, incentivizing actionable feedback. To address the distribution shift between critique-conditioned and critique-free behavior, ICRL introduces a distribution-calibration re-weighting ratio that selectively transfers critique-guided improvements compatible with the solver's own prompt distribution. Additionally, a role-wise group advantage estimation stabilizes joint optimization across the two roles. Together, these mechanisms ensure that the solver learns to improve itself without external critique, rather than becoming dependent on critique-conditioned behavior. We evaluate ICRL on diverse benchmarks spanning agentic and mathematical reasoning tasks, using Qwen3-4B and Qwen3-8B as backbones. Results show consistent improvements, with average gains of 6.4 points over GRPO on agentic tasks, and 7.0 points on mathematical reasoning. Notably, the learned 8B critic is comparable to 32B critics while using substantially fewer tokens. The code is available at this https URL .\n\n---\n\n## 45. NOVA: Fundamental Limits of Knowledge Discovery Through AI\n- **Authors**: Salman Avestimehr , Ken Duffy , Muriel Médard\n- **URL**: [https://arxiv.org/abs/2605.15219](https://arxiv.org/abs/2605.15219)\n- **Abstract**:\n> Can AI systems discover genuinely new knowledge through iterative self improvement, and if so, at what cost? We introduce the NOVA framework, which models the common ``generate, verify, accumulate, retrain'' loop as an adaptive sampling process over a knowledge space. We identify sufficient conditions under which accumulated genuine knowledge eventually covers a finite domain, and show how their violations produce distinct failure modes: contamination, forgetting, exploration failure, and acceptance failure. We then analyze imperfect verification and identify a contamination trap: as easy-to-find knowledge is exhausted, the model mass assigned to new valid artifacts shrinks, so even small false-positive rates can cause invalid artifacts to enter the knowledge base faster than genuine discoveries. We clarify that Good--Turing estimation is a local batch-diversity diagnostic, not an estimator of the historically undiscovered valid mass that governs long-term discovery. Under a separate tail-equivalence assumption relating the model's effective discovery distribution to a Zipf law with exponent $\\alpha>1$, we prove that the cumulative generation cost required to obtain $D$ distinct genuine discoveries satisfies $R_{\\mathrm{cum} }(D)=\\Theta(c_{\\mathrm{gen} }D^\\alpha)$, where $c_{\\mathrm{gen} }$ is the per-candidate generation cost. This scaling law quantifies asymptotic diminishing returns as the discovery frontier advances. Finally, we formalize human amplification through guidance, generation, and verification, explaining why expert input is most valuable near autonomous exploration barriers.\n\n---\n\n## 46. CAX-Agent: A Lightweight Agent Harness for Reliable APDL Automation\n- **Authors**: Chenying Lin , Yichen Hai , Yi He , Ran Wang , Haiyan Qiang , Liang Yu\n- **URL**: [https://arxiv.org/abs/2605.15218](https://arxiv.org/abs/2605.15218)\n- **Abstract**:\n> Large language models deployed for MAPDL finite-element simulation face practical reliability challenges: without structured execution control, tool encapsulation, and fault recovery, outputs may be inconsistent and task failures are common. The Agent Harness paradigm addresses this by inserting domain-specific orchestration middleware that manages tool lifecycles, workflow state, and recovery escalation. This paper presents the architecture of CAX-Agent, a lightweight agent harness purpose-built for MAPDL automation, and empirically evaluates one of its core components -- the recovery this http URL -Agent organizes execution into three layers -- LLM service, agent harness, and solver backend -- with a recovery ladder that escalates from deterministic rule patching through model-driven regeneration to context enrichment and human intervention. We evaluate three recovery strategies (no_recovery, rule_only, and model_only) on 50 standard structural benchmarks with three repeated runs per strategy (450 case-runs total). Two independent human raters score task completion under blind conditions; inter-rater agreement is strong (quadratic weighted Cohen's kappa = 0.84, 96 percent of score pairs within one point). Model_only achieves the best completion rate (0.9267), task score (3.59/4), total score (9.16/10), and zero-intervention rate (0.84), outperforming rule_only (0.7733, 3.17/4, 7.03/10, 0.00) and no_recovery (0.6933, 2.74/4, 5.60/10, 0.00) with large effect sizes (Cliff's delta = 0.81-0.87). The benchmark uses deliberately simple geometries to isolate recovery-policy effects; we discuss the scope of these findings and directions for broader validation.\n\n---\n\n## 47. Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions\n- **Authors**: Jagdish Tripathy , Marcus Buckmann\n- **URL**: [https://arxiv.org/abs/2605.15217](https://arxiv.org/abs/2605.15217)\n- **Abstract**:\n> Instruction-tuned language models exhibit behavioural fairness in high-stakes decisions while retaining biased associations in their internal representations. However, whether these suppressed representations can affect model outputs - and whether such causal potency is symmetric across demographic groups - remains unknown. We investigate the use of open-weight models for mortgage underwriting using matched applications that differ only in racially-associated names and reveal a critical disconnect: models show no output-level bias, yet retain and amplify demographic representations across model layers. Through activation steering and novel cross-layer interventions, we demonstrate that this suppressed information is decision-relevant: when reinjected at critical layers, it produces near-complete decision reversals. Critically, this latent bias is asymmetric - steering interventions affect decisions in one demographic direction, while producing minimal effects in reverse - and susceptible to adversarial prompt engineering and parameter-efficient fine-tuning. These findings demonstrate that behavioural audits focused on outputs are insufficient: fair outputs can mask exploitable internal biases. They also motivate dual-layer testing frameworks combining output evaluation with representational analysis for AI governance in high-stakes decisions.\n\n---\n\n## 48. SkillSmith: Compiling Agent Skills into Boundary-Guided Runtime Interfaces\n- **Authors**: Duling Xu , Zheng Chen , Zaifeng Pan , Jiawei Guan , Dong Dong , Jialin Li , Bangzheng Pu\n- **URL**: [https://arxiv.org/abs/2605.15215](https://arxiv.org/abs/2605.15215)\n- **Abstract**:\n> Recently, skills have been widely adopted in large language model (LLM)-based agent systems across various domains. In existing frameworks, skills are typically injected into the agent reasoning loop as contextual guidance once matched to a runtime task, enabling specialized task-solving capabilities. We find that this execution paradigm introduces two major sources of redundancy: irrelevant context injection and repeated skill-specific reasoning and planning. To this end, we propose SkillSmith, a boundary-first compiler-runtime framework that compiles skill packages offline into minimal executable interfaces. By extracting fine-grained operational boundaries from skills, SkillSmith enables agents to dynamically access and execute only the relevant components at runtime, thereby minimizing unnecessary context injection and redundant reasoning overhead. In the evaluation on SkillsBench benchmark, SkillSmith reduces solve-stage token usage by 57.44%, thinking iterations by 42.99%, solve time by 50.57% (2.02x faster), and token-proportional monetary cost by 57.44% compared with using raw-skills. Moreover, compiled artifacts produced by a stronger model can be reused by a smaller or more efficient runtime model, improving task accuracy in cases where raw skill interpretation fails. The source code and data are available at this https URL .\n\n---\n\n## 49. Does Theory of Mind Improvement Really Benefit Human-AI Interactions? Empirical Findings from Interactive Evaluations\n- **Authors**: Nanxu Gong , Zixin Chen , Haotian Li , Zishu Zhao , Jianxun Lian , Huamin Qu , Yanjie Fu , Xing Xie\n- **URL**: [https://arxiv.org/abs/2605.15205](https://arxiv.org/abs/2605.15205)\n- **Abstract**:\n> Improving the Theory of Mind (ToM) capability of Large Language Models (LLMs) is crucial for effective social interactions between these AI models and humans. However, the existing benchmarks often measure ToM capability improvement through story-reading, multiple-choice questions from a third-person perspective, while ignoring the first-person, dynamic, and open-ended nature of human-AI (HAI) interactions. To directly examine how ToM improvement techniques benefit HAI interactions, we first proposed the new paradigm of interactive ToM evaluation with both perspective and metric shifts. Next, following the paradigm, we conducted a systematic study of four representative ToM enhancement techniques using both four real-world datasets and a user study, covering both goal-oriented tasks (e.g., coding, math) and experience-oriented tasks (e.g., counseling). Our findings reveal that improvements on static benchmarks do not always translate to better performance in dynamic HAI interactions. This paper offers critical insights into ToM evaluation, showing the necessity of interaction-based assessments in developing next-generation, socially aware LLMs for HAI symbiosis.\n\n---\n\n## 50. SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch\n- **Authors**: Zhantao Wang\n- **URL**: [https://arxiv.org/abs/2605.15204](https://arxiv.org/abs/2605.15204)\n- **Abstract**:\n> Multi-agent orchestration frameworks such as LangChain, LangGraph, and CrewAI route tasks through graph-based pipelines but do not enforce the stage constraints that govern real business processes. We present SDOF, a framework that treats multi-agent execution as a constrained state machine. SDOF operates through two primary defensive layers, implemented by three components: (1) an Online-RLHF Specialized Intent Router trained via Generative Reward Modeling (GRPO) and (2) a StateAwareDispatcher with GoalStage finite-automaton checks and precondition/postcondition SkillRegistry validation for auditable execution control. On a recruitment system backed by the Beisen iTalent platform (6000+ enterprises), 185 expert-curated scenarios trigger 1671 live API calls. Our GSPO-aligned 7B Intent Router achieves higher joint accuracy than zero-shot GPT-4o on this FSM-constrained adversarial routing benchmark (80.9% versus 48.9%). In end-to-end execution, SDOF reaches 86.5% task completion (95% confidence interval 80.8 to 90.7) and blocks all 22 operations in the injection, illegal HR subset. Under a broader message-level blocking audit, SDOF attains precision 100% and recall 88%, expert agreement kappa=0.94. A separate evaluation on 960 SGD-derived dialogues spanning 8 service domains surfaces 201 stage-order conflicts under our FSM mapping, 41 of which arise in the normal split. This arXiv version reports the current validated scope; extended multi-seed training comparisons and deeper workflow evaluations will be released in a subsequent update.\n\n---\n\n## 51. DeepSlide: From Artifacts to Presentation Delivery\n- **Authors**: Ming Yang , Zhiwei Zhang , Jiahang Li , Haoseng Liu , Yuzheng Cai , Weiguo Zheng\n- **URL**: [https://arxiv.org/abs/2605.15202](https://arxiv.org/abs/2605.15202)\n- **Abstract**:\n> Presentations are a primary medium for scholarly communication, yet most AI slide generators optimize the artifact (a visually plausible deck) while under-optimizing the delivery process (pacing, narrative, and presentation preparation). We present DeepSlide, a human-in-the-loop multi-agent system that supports preparing the full presentation process, from requirement elicitation and time-budgeted narrative planning, to evidence-grounded slide--script generation, attention augmentation, and rehearsal support. DeepSlide integrates (i) a controllable logical-chain planner with per-node time budgets, (ii) a lightweight content-tree retriever for grounding, (iii) Markov-style sequential rendering with style inheritance, and (iv) sandboxed execution with minimal repair to ensure renderability. We further introduce a dual-scoreboard benchmark that cleanly separates static artifact quality from dynamic delivery excellence. Across 20 domains and diverse audience profiles, DeepSlide matches strong baselines on artifact quality while consistently achieving larger gains on delivery metrics, improving narrative flow, pacing precision, and slide--script synergy with clearer attention guidance.\n\n---\n\n## 52. IVGT: Implicit Visual Geometry Transformer for Neural Scene Representation\n- **Authors**: Yuqi Wu , Tianyu Hu , Wenzhao Zheng , Yuanhui Huang , Haowen Sun , Jie Zhou , Jiwen Lu\n- **URL**: [https://arxiv.org/abs/2605.16258](https://arxiv.org/abs/2605.16258)\n- **Abstract**:\n> Reconstructing coherent 3D geometry and appearance from unposed multi-view images is a fundamental yet challenging problem in computer vision. Most existing visual geometry foundation models predict explicit geometry by regressing pixel-aligned pointmaps, often suffering from redundancy and limited geometric continuity. We propose IVGT, an Implicit Visual Geometry Transformer that implicitly models continuous and coherent geometry from pose-free multi-view images. This formulation learns a continuous neural scene representation in a canonical coordinate system and supports continuous spatial queries at any 3D positions, retrieving local features to predict signed distance (SDF) values and colors using lightweight decoders. It allows direct extraction of continuous and coherent surface geometry, enabling rendering of RGB images, depth maps, and surface normal maps from arbitrary viewpoints. We train IVGT via multi-dataset joint optimization with 2D supervision and 3D geometric regularization. IVGT demonstrates generalization across scenes and achieves strong performance on various tasks, including mesh and point cloud reconstruction, novel view synthesis, depth and surface normal estimation, and camera pose estimation.\n\n---\n\n## 53. Designing Datacenter Power Delivery Hierarchies for the AI Era\n- **Authors**: Grant Wilkins , Fiodar Kazhamiaka , Alok Gautam Kumbhare , Chaojie Zhang , Ricardo Bianchini\n- **URL**: [https://arxiv.org/abs/2605.16255](https://arxiv.org/abs/2605.16255)\n- **Abstract**:\n> Demand for AI accelerators is rapidly increasing rack power density, with projections approaching 1MW per deployment by 2027. This poses a major challenge for datacenter power delivery designers. As power densities increase, a datacenter designed for a different target density may strand power, i.e., may be unable to use all the power that its delivery hierarchy has provisioned. Designs must remain efficient over long datacenter lifetimes and multiple hardware generations. Power utilization is particularly important as grid power capacity is a scarce resource in the AI era. Designing an efficient power delivery hierarchy for the long run is difficult because rack placement feasibility, workload impact, and cost depend jointly on electrical topology, deployment granularity, placement policy, power oversubscription, and workload mix. Moreover, each of these factors evolve over time, have inter-dependencies across multiple resource dimensions, and generally do not lend themselves to closed-form analysis. To address this challenge, we develop a framework for evaluating datacenter power delivery designs using throughput, power, and cost metrics over realistic arrival, oversubscription, and decommissioning sequences. The framework combines projection models for GPU, compute, and storage deployments with operational factors grounded in production data from Microsoft Azure. Our results show that multi-resource stranding materially changes deployable capacity, effective capital expenditure, and delivered performance, and quantify how rising density from rack- and pod-scale AI systems shapes these outcomes. For AI datacenter design, the relevant planning objective is not installed megawatts, but deployable capacity over time.\n\n---\n\n## 54. A Generative AI Framework for Intelligent Utility Billing CO 2 Analytics and Sustainable Resource Optimisation\n- **Authors**: Pavan Manjunath , Thomas Pruefer\n- **URL**: [https://arxiv.org/abs/2605.16250](https://arxiv.org/abs/2605.16250)\n- **Abstract**:\n> Distribution utilities are now expected to deliver bills that customers can actually read attach a defensible carbon number to every kWh sold and schedule load against grid stress and emissions constraints We propose an end-to-end framework that unifies four production-grade capabilities under one architectural roof a generative-AI agent that drafts each customers natural-language billing statement from structured numeric inputs under a constrained decoding policy a transformer-based forecaster that supplies the day-ahead consumption estimate with calibrated quantile bands\n\n---\n\n## 55. AI-Mediated Communication Can Steer Collective Opinion\n- **Authors**: Stratis Tsirtsis , Kai Rawal , Chris Russell , Brent Mittelstadt , Sandra Wachter\n- **URL**: [https://arxiv.org/abs/2605.16245](https://arxiv.org/abs/2605.16245)\n- **Abstract**:\n> Generative artificial intelligence (AI) is increasingly integrated into the online platforms where humans exchange opinions; large language models (LLMs) now polish users' posts on LinkedIn and provide context for content shared on X. While prior work has shown that AI can express biased opinions and shape individuals' opinions during human-AI interactions, less attention has been paid to its influence on collective opinion formation when mediating human-to-human communication. We address this gap via a combination of empirical and theoretical analyses. We show empirically that LLMs from multiple popular families introduce directional biases when instructed to edit human-written texts on contested topics, for example, nudging texts in favor of gun control and against atheism. Building on this observation, we introduce a mathematical model of opinion dynamics in which an AI system sits between users on a social network, transforming the opinions they express and perceive. By analytically characterizing the equilibrium of this model and performing simulations on real social network data, we show that biases introduced by AI in human-to-human communication can be amplified through the network and shift collective opinion in their direction. In light of these findings, we investigate whether such biases are controllable by online platforms. We audit the \"Explain this post\" feature on X and find evidence of pro-life bias in Grok's outputs on abortion-related content, which we trace back to specific design choices. We conclude with a discussion of the broader implications of our findings in relation to ongoing legislative efforts in the European Union.\n\n---\n\n## 56. Offline Semantic Guidance for Efficient Vision-Language-Action Policy Distillation\n- **Authors**: Jin Shi , Brady Zhang , Yishun Lu\n- **URL**: [https://arxiv.org/abs/2605.16241](https://arxiv.org/abs/2605.16241)\n- **Abstract**:\n> Billion-parameter Vision-Language-Action (VLA) policies have recently shown impressive performance in robotic manipulation, yet their size and inference cost remain major obstacles for real-time closed-loop control. We introduce \\textbf{VLA-AD}, a distillation framework that uses a Vision-Language Model as an offline semantic supervisor to transfer large VLA teachers into lightweight student policies. Instead of relying only on low-level action imitation, VLA-AD augments teacher-provided 7-DoF action targets with high-level semantic guidance, including task phase anchors and multi-frame operating-direction descriptions. These auxiliary signals are used only during training: at test time, the student policy runs independently, with neither the VLA teacher nor the VLM required. We evaluate VLA-AD on three LIBERO benchmark suites. Using OpenVLA-7B as the teacher, our method produces a 158M-parameter student, yielding a $44\\times$ reduction in model size while matching the teacher with only a $0.27\\%$ average relative gap. The resulting policy runs at 12.5 Hz on an RTX 4090, achieving a $3.28\\times$ inference speedup over OpenVLA-7B. We further show that the same semantic distillation pipeline generalizes to a different $\\pi_{0.5}$-4B teacher, where the student outperforms the teacher on two suites and remains within $0.53\\%$ on \\texttt{libero\\_goal}. Additional analysis indicates that phase-level supervision and multi-frame directional cues make the student less sensitive to noisy teacher actions, such as erroneous high-frequency gripper changes. Overall, VLA-AD demonstrates that offline semantic guidance from VLMs can substantially improve the efficiency, robustness, and deployability of VLA policy distillation.\n\n---\n\n## 57. Layer Equivalence Is Not a Property of Layers Alone: How You Test Redundancy Changes What You Find\n- **Authors**: Gabriel Garcia\n- **URL**: [https://arxiv.org/abs/2605.16234](https://arxiv.org/abs/2605.16234)\n- **Abstract**:\n> When researchers ask whether two transformer layers are \"equivalent\" for compression, they often conflate distinct tests. Replacement asks whether one layer's map can substitute for another's in place; interchange asks whether two layers approximately commute when their positions are swapped. Both are output-grounded swap-KL probes, but they need not agree: on pretrained transformers the protocol gap can change which layers look safe to prune by several-fold under the same evaluator, especially when replacement distances are high. We measure both protocols across checkpoints and architectures. On a Pythia training trajectory (410M and 1.4B), the replacement-interchange gap grows from initialization to convergence. Under one matched WikiText-2 contract at 8B scale, Qwen3-8B enters a divergent regime: interchange-guided removal is several-fold safer than replacement-guided at the same layer budgets, while Llama-3.1-8B ties the two protocols for pruning cost even though interchange KL is lower, showing metric gaps need not map one-to-one to removal. Before layer removal or merging, score both swap-KLs on the target checkpoint; the diagnostic requires only unlabeled forward passes.\n\n---\n\n## 58. A Unified Generative-AI Framework for Smart Energy Infrastructure: Intelligent Gas Distribution, Utility Billing, Carbon Analytics, and Quantum-Inspired Optimisation\n- **Authors**: Pavan Manjunath , Thomas pruefer\n- **URL**: [https://arxiv.org/abs/2605.16232](https://arxiv.org/abs/2605.16232)\n- **Abstract**:\n> The accelerating convergence of smart metering, generative artificial intelligence, and quantum-inspired combinatorial optimisation is reshaping how energy utilities manage physical infrastructure, customer engagement, and environmental accountability\n\n---\n\n## 59. Evaluating Design Video Generation: Metrics for Compositional Fidelity\n- **Authors**: Adrienne Deganutti , Dingning Cao , Jaejung Seol , Elad Hirsch , Purvanshi Mehta\n- **URL**: [https://arxiv.org/abs/2605.16223](https://arxiv.org/abs/2605.16223)\n- **Abstract**:\n> Generative video models are increasingly used in design animation tasks, yet no standardized evaluation framework exists for this domain. Unlike natural video generation, design animation imposes structured constraints: specific components shall animate with prescribed motion types, directions, speed and timing, while non-animated regions must remain stable and layout structure must be preserved. This paper provides a fully automated evaluation framework organized across four dimensions: layout fidelity, motion correctness, temporal quality, and content fidelity. This eliminates the reliance on subjective human evaluation and establishes a common basis for benchmarking progress in the field.\n\n---\n\n## 60. Argus: Evidence Assembly for Scalable Deep Research Agents\n- **Authors**: Zhen Zhang , Liangcai Su , Zhuo Chen , Xiang Lin , Haotian Xu , Simon Shaolei Du , Kaiyu Yang , Bo An , Lidong Bing , Xinyu Wang\n- **URL**: [https://arxiv.org/abs/2605.16217](https://arxiv.org/abs/2605.16217)\n- **Abstract**:\n> Deep research agents have achieved remarkable progress on complex information seeking tasks. Even long ReAct style rollouts explore only a single trajectory, while recent state of the art systems scale inference time compute via parallel search and aggregation. Yet deep research answers are composed of complementary pieces of evidence, which parallel rollouts often duplicate rather than complete, yielding diminishing returns while pushing the aggregation context toward the model's limit. We propose Argus, an agentic system in which a Searcher and a Navigator cooperate to treat deep research as assembling a jigsaw from complementary evidence pieces, rather than brute forcing the whole answer in parallel. The Searcher collects evidence traces for a given sub-query through ReAct-style interaction. The Navigator maintains a shared evidence graph, verifying which pieces are still missing, dispatching Searchers to gather them, and reasoning over the completed graph to produce a source-traced final answer. We train the Navigator with reinforcement learning to verify, dispatch, and synthesize, while independently training the Searcher to remain a standard ReAct agent. The resulting Navigator supports rollouts with a single Searcher or many in parallel without retraining. With both Searcher and Navigator built on a 35B-A3B MoE backbone, Argus gains 5.5 points with a single Searcher and 12.7 points with 8 parallel Searchers, averaged over eight benchmarks. With 64 Searchers it reaches 86.2 on BrowseComp, surpassing every proprietary agent we benchmark, while the Navigator's reasoning context stays under 21.5K tokens.\n\n---\n\n## 61. paper.json: A Coordination Convention for LLM-Agent-Actionable Papers\n- **Authors**: Arquimedes Canedo\n- **URL**: [https://arxiv.org/abs/2605.16194](https://arxiv.org/abs/2605.16194)\n- **Abstract**:\n> LLM agents routinely serve as first (and sometimes only) readers of academic papers, skimming for sub-claims, extracting reproducibility steps, and generalizing scope. Standard prose papers produce recurring failures in this role: sub-claims that cannot be cited at sub-paper granularity, scope overextension beyond what the paper tests, and figure commands buried in codebases rather than the paper itself. We propose ` this http URL `, a companion JSON file that travels with the PDF and addresses each failure with a lightweight convention: stable claim IDs (C1), an explicit does-not-claim list (C2), exact per-figure shell commands (C3), and stable definition IDs (C5). A fifth convention (C4) holds that minimum viable compliance, hand-written JSON alongside the PDF, is achievable in under an hour for a finished paper without touching the human-readable output. C1, C2, C3, and C5 are open invitations: an agent that reads a compliant paper and acts on it produces evidence for or against them. This paper is itself compliant: `uv run this http URL this http URL --against this http URL ` passes. Repo: this https URL\n\n---\n\n## 62. Second-Order Multi-Level Variance Correction for Modality Competition in Multimodal Models\n- **Authors**: Yishun Lu , Wes Armour\n- **URL**: [https://arxiv.org/abs/2605.16165](https://arxiv.org/abs/2605.16165)\n- **Abstract**:\n> Autoregressive next-token training offers a unified formulation for image generation and text understanding, but it also creates strong modality competition that destabilizes optimization and limits large-batch scaling. We show that first-order optimizers such as AdamW are vulnerable to cross-modality gradient heterogeneity, while second-order preconditioning, particularly SOAP, provides a more stable basis for multimodal alignment. Building on this insight, we propose \\emph{ML-FOP-SOAP}, a second-order optimization framework with Multi-Level Variance Correction. Our Fisher-Orthogonal Projection suppresses variance-induced modality conflicts, reducing the trade-off between visual generation and textual understanding. To make this practical under large gradient accumulation, we introduce a hierarchical folding strategy that captures fine-grained variance with low micro-step overhead. Experiments on Janus and Emu3 show consistent gains across both modalities and stable training at batch size 8192. Compared with AdamW, our method improves sample efficiency by up to $1.4\\times$ and accelerates wall-clock training by up to $1.5\\times$, offering a robust optimizer for scaling multimodal foundation models.\n\n---\n\n## 63. Surrogate Neural Architecture Codesign Package (SNAC-Pack)\n- **Authors**: Jason Weitz , Dmitri Demler , Benjamin Hawks , Aaron Wang , Nhan Tran , Javier Duarte\n- **URL**: [https://arxiv.org/abs/2605.16138](https://arxiv.org/abs/2605.16138)\n- **Abstract**:\n> Neural architecture search (NAS) is a powerful approach for automating model design, but existing methods often optimize for accuracy alone or rely on proxy metrics such as bit operations (BOPs) that correlate poorly with hardware cost. This gap is particularly large for FPGA deployment, where cost is dominated by a multi-dimensional budget of lookup tables, DSPs, flip-flops, BRAM, and latency. We present the Surrogate Neural Architecture Codesign Package (SNAC-Pack), an open-source AutoML framework for hardware-aware neural architecture codesign and end-to-end FPGA deployment. SNAC-Pack runs a multi-objective global search with Optuna and NSGA-II, loading trials to a shared SQLite store that enables parallel workers across compute nodes. A hardware surrogate model outputs per-trial resource and latency estimates, avoiding the synthesis cost that would otherwise dominate the search loop. A local search stage then applies quantization-aware training (QAT) together with iterative magnitude pruning in a combined compression loop, after which the final model is synthesized to FPGA firmware via the hls4ml Python library. A YAML configuration and an optional agentic frontend let users run the pipeline on new datasets without modifying the framework. We demonstrate SNAC-Pack on jet classification at the Large Hadron Collider and superconducting qubit readout, discovering compact architectures that match or exceed strong baselines on the task metric while reducing FPGA resource utilization and, in the qubit readout case, reducing the design space exploration process from months of manual fine-tuning to hours of automated search.\n\n---\n\n## 64. Navigating Potholes with Geometry-Aware Sharpness Minimization\n- **Authors**: Simon Dufort-Labbé , Mehrab Hamidi , Razvan Pascanu , Ioannis Mitliagkas , Damien Scieur , Aristide Baratin\n- **URL**: [https://arxiv.org/abs/2605.16134](https://arxiv.org/abs/2605.16134)\n- **Abstract**:\n> Sharpness-aware minimization (SAM) encourages flat minima by perturbing parameters along directions of high loss curvature, but treats all parameter directions uniformly, ignoring the underlying loss geometry. We introduce LLQR+SAM, which combines SAM with a learned preconditioner obtained from the recently proposed LLQR framework, a second-order method that recasts steepest descent as a layerwise linear-quadratic regulator problem. The preconditioner is updated sparsely and maintained as a slow exponential moving average, so it captures a smoothed, low-resolution picture of the loss landscape geometry. The SAM perturbation then operates on top of this learned geometry, probing curvature at a faster timescale. We show that this two-timescale structure is not merely a computational convenience: theoretically, the preconditioner amplifies the SAM escape signal in directions that are flat under the average geometry but locally sharp (potholes). Wide, flat basins, by contrast, remain stable. Empirically, LLQR+SAM gives consistent gains over both SAM and LLQR alone across standard vision and sequence modeling benchmarks, supporting the view that slow learned geometry and fast sharpness correction are genuinely complementary.\n\n---\n\n## 65. Entropy Across the Bridge: Conditional-Marginal Discretization for Flow and Schrödinger Samplers\n- **Authors**: Bruno Trentini , Dejan Stancevic , Michael M. Bronstein , Alexander Tong , Luca Ambrogioni\n- **URL**: [https://arxiv.org/abs/2605.16126](https://arxiv.org/abs/2605.16126)\n- **Abstract**:\n> For a fixed flow-based generative model under a small inference budget, sample quality can depend strongly on where the sampler spends its few function evaluations. Flow matching and Schrödinger bridges define probability paths, yet their inference grids are usually heuristic or inherited from one-endpoint diffusion. We derive a conditional-marginal entropy-rate objective for bridge-aware discretization, separating endpoint-conditioned bridge geometry from marginal flow evolution, and use it to build a training-free entropic inference-time scheduler from first principles. For Gaussian Brownian bridges this rate is closed-form and U-shaped, motivating boundary-heavy nonuniform grids. On trained two-dimensional bridge/flow models, the estimated profile recovers the predicted shape and improves 10-step ODE-Heun MMD over linear by 18.1%, with a paired 22.7% SDE-Heun improvement in the same low-NFE sweep. On EDM/CIFAR-10, the entropic time-discretization gives the best tested five-step FID (186.3 \\pm 4.0 versus 200.5 \\pm 2.9 for linear and 238.0 \\pm 5.3 for cosine). On AlphaFlow protein generation, entropic conditional-marginal (cond-marg) scheduling shows advantage in low-NFE regimes on both CAMEO22 and ATLAS benchmarks. These results support entropy-rate scheduling as a practical low-budget allocation signal for high-dimensional bridge and flow samplers.\n\n---\n\n## 66. GenShield: Unified Detection and Artifact Correction for AI-Generated Images\n- **Authors**: Zhipei Xu , Xuanyu Zhang , Youmin Xu , Qing Huang , Shen Chen , Taiping Yao , Shouhong Ding , Jian Zhang\n- **URL**: [https://arxiv.org/abs/2605.16122](https://arxiv.org/abs/2605.16122)\n- **Abstract**:\n> Diffusion-based image synthesis has made AI-generated images (AIGI) increasingly photorealistic, raising urgent concerns about authenticity in applications such as misinformation detection, digital forensics, and content moderation. Despite the substantial advances in AIGI detection, how to correct detected AI-generated images with visible artifacts and restore realistic appearance remains largely underexplored. Moreover, few existing work has established the connection between AIGI detection and artifact correction. To fill this gap, we propose GenShield, a unified autoregressive framework that jointly performs explainable AIGI detection and controllable artifact correction in a closed loop from diagnosis to restoration, revealing a mutually reinforcing relationship between these two tasks. We further introduce a Visual Chain-of-Thought based curriculum learning strategy that enables self-explained, multi-step ``diagnose-then-repair'' correction with an explicit stopping criterion. A high-quality dataset with large-scale ``artifact-restored'' pairs is also constructed alongside a unified evaluation pipeline. Extensive experiments on our correction benchmark and mainstream AIGI detection benchmarks demonstrate state-of-the-art performance and strong generalization of our method. The code is available at this https URL .\n\n---\n\n## 67. DebiasRAG: A Tuning-Free Path to Fair Generation in Large Language Models through Retrieval-Augmented Generation\n- **Authors**: Rui Chu , Bingyin Zhao , Thanh Quoc Hung Le , Duy Cao Hoang , Huawei Lin , Ping Li , Weijie Zhao , Khoa D Doan , Yingjie Lao\n- **URL**: [https://arxiv.org/abs/2605.16113](https://arxiv.org/abs/2605.16113)\n- **Abstract**:\n> Large language models (LLMs) have achieved unprecedented success due to their exceptional generative capabilities. However, because they depend on knowledge encapsulated from training corpora, they may produce hallucinations, stereotypes, and socially biased content. In particular, LLMs are prone to prejudiced responses involving race, gender, and age, which are collectively referred to as social biases. Prior studies have used fine-tuning and prompt engineering to mitigate such biases in LLMs, but these methods require additional training resources or domain knowledge to design the framework. Moreover, they may degrade the original capabilities of LLMs and often overlook the need for dynamic debiasing contexts for fairer inference. In this paper, we propose DebiasRAG, a novel tuning-free and dynamic query-specific debiasing framework based on retrieval-augmented generation (RAG). DebiasRAG improves fairness while preserving the intrinsic properties of LLMs, such as representation ability. DebiasRAG consists of three stages: (1) query-specific debiasing candidate generation; (2) context candidate pool construction; and (3) gradient-updated debiasing-guided context piece reranking. First, DebiasRAG leverages self-diagnosed bias contexts relevant to the query through regular retrieval, where the bias contexts are prepared offline by the DebiasRAG provider. Given the query-specific bias contexts, DebiasRAG reversely produces debiasing contexts, which are provided as additional fairness constraints for LLM outputs. Second, a regular RAG retrieval process produces query-related contexts from the regular RAG document database, such as a chunked Wikipedia dataset.\n\n---\n\n## 68. Attention Dispersion in Dynamic Graph Transformers: Diagnosis and a Transferable Fix\n- **Authors**: Jinhao Zhang , Kangfei Zhao , Qiuhao Zeng , Long-Kai Huang\n- **URL**: [https://arxiv.org/abs/2605.16112](https://arxiv.org/abs/2605.16112)\n- **Abstract**:\n> Transformer-based architectures have become the dominant paradigm for Continuous-Time Dynamic Graph (CTDG) learning, yet their performance remains limited on temporally shifted datasets. In this work, we identify attention dispersion as a shared failure mode of dynamic graph Transformers under temporal distribution shift. Through controlled ablation contrasting structurally and temporally distinguished historical neighbors against random ones, we show that prediction depends on a class of critical nodes that carry consistently more predictive signal than arbitrary neighbors. However, existing Transformers fail to focus on these nodes even when they are present in the input, as temporal shift weakens attention contrast and produces overly dispersed attention distributions. This diagnosis suggests a simple and transferable fix: replace standard attention with differential attention, which suppresses common-mode attention and amplifies distinctive token-level signals. When added to three representative CTDG Transformer baselines, differential attention consistently improves performance, with gains concentrated on high-shift datasets. Attention-level measurements further confirm the mechanism, showing reduced attention entropy and increased attention mass on critical nodes. Building on these findings, we introduce DiffDyG, a reference implementation combining differential attention with standard input encodings. Across 9 benchmarks and three negative sampling protocols, DiffDyG achieves SOTA performance, with especially large gains on the most shifted datasets.\n\n---\n\n## 69. Federated Imputation under Heterogeneous Feature Spaces\n- **Authors**: Imane Hocine , Chaimaa Medjadji , Sylvain Kubler , Gregoire Danoy , Yves Le Traon\n- **URL**: [https://arxiv.org/abs/2605.16099](https://arxiv.org/abs/2605.16099)\n- **Abstract**:\n> Federated Learning (FL) enables collaborative training across decentralized clients, but most methods assume aligned feature schemas, an assumption that rarely holds in tabular settings where clients observe only partially overlapping feature subsets. In these heterogeneous feature spaces, parameter-averaging methods (e.g., FedAvg) transfer little information across weakly overlapping or disjoint feature groups, limiting their effectiveness for federated imputation. To overcome this, we propose \\textbf{FedHF-Impute}, a federated imputation framework that separates structural feature unavailability from conventional missingness and uses a shared global feature graph to propagate information across statistically related features through message passing. This enables indirect cross-client knowledge transfer, even when features are never jointly observed locally, while preserving standard federated communication. Under simulated partial schema overlap on the SECOM and AirQuality datasets, FedHF-Impute improves imputation accuracy (RMSE) over FL baselines by 26.9\\%, and 8.4\\% respectively, while achieving comparable performance on PhysioNET, with only a 0.3\\% difference relative to the best baseline.\n\n---\n\n## 70. GeoGS-CE: Learning Delay--Beam Channel Priors with 3D Gaussians for High-Mobility Scenarios\n- **Authors**: Yumeng Zhang , Jiajia Guo , Chaozheng Wen , Chenghong Bian , Jun Zhang\n- **URL**: [https://arxiv.org/abs/2605.16094](https://arxiv.org/abs/2605.16094)\n- **Abstract**:\n> Wideband channel estimation (CE) in high-mobility scenarios remains challenging because channel responses vary rapidly, while practical systems can allocate only sparse pilots to accommodate dense users. Fortunately, many high-mobility environments, such as high-speed railways, exhibit scheduled trajectories, predictable velocities, and a limited number of dominant propagation paths. These properties induce a delay--beam power spectrum that is more stable than the instantaneous complex channel frequency response (CFR), less sensitive to the random phase coherence, and rich in geometric information. To exploit such environmental properties, we propose GeoGS-CE, a two-stage channel estimation framework for sparse-pilot high-mobility scenarios. In the offline stage, GeoGS-CE jointly models: 1) a scene-level 3D Gaussian representation that captures the non-line-of-sight (NLoS) geometric scattering support, and 2) a leakage-aware differentiable wireless rendering process that maps the NLoS Gaussians, together with an explicit virtual line-of-sight (LoS) component, to the measured delay--beam power spectrum, while accounting for practical OFDM delay and array leakage effects. In the online stage, the delay--beam power spectrum is predicted for each user location and used as a strong covariance prior, enabling accurate full-band and full-array CFR reconstruction and tracking through a linear MMSE estimator. Simulations based on channels generated from a segment of the Guangshen high-speed railway show that the proposed geometric prior substantially improves CFR reconstruction over pilot-only and non-geometric baselines.\n\n---\n\n## 71. Centralized vs Decentralized Federated Learning: A trade-off performance analysis\n- **Authors**: Chaimaa Medjadji , Guilain Leduc , Sylvain Kubler , Yves Le Traon\n- **URL**: [https://arxiv.org/abs/2605.16089](https://arxiv.org/abs/2605.16089)\n- **Abstract**:\n> Federated Learning (FL) has emerged as a promising paradigm for collaborative model training across distributed edge devices while preserving data privacy especially with the huge increase amount of data due to the adoption of technologies which contributes to the growing number of IoT devices. Storing this amount of data centrally is challenging due to issues like limited communication, privacy, and regulations. FL can be Centralized (CFL), Decentralized (DFL), and Semi-decentralized (SDFL). Choosing the right FL architecture depends on the application's needs. However, very few research studies have experimentally compared these three types of architectures to not only understand the respective strengths and limitations, but also trade-offs between different performance indicators. This paper overcome this lack of analysis, conducting experimental analyses using the Fedstellar simulator, MNIST dataset, and MLP classifier.\n\n---\n\n## 72. Multi-level Self-supervised Pretraining on Compositional Hierarchical Graph for Molecular Property Prediction\n- **Authors**: Xiayu Liu , Zhengyi Lu , Hou-biao Li\n- **URL**: [https://arxiv.org/abs/2605.16088](https://arxiv.org/abs/2605.16088)\n- **Abstract**:\n> Self-supervised pretraining on molecular graphs has emerged as a promising approach for molecular property prediction, yet most existing methods operate at a single structural granularity and treat bond information as auxiliary edge attributes rather than as an independent semantic layer. In this work, we propose MolCHG, a multi-level self-supervised pretraining framework built upon a novel Compositional Hierarchical Graph that organizes molecular structure into four types of nodes across three semantic levels. By introducing a bond graph that operates in parallel with the atom graph, our architecture elevates bond-level information to independently evolving node representations, enabling fragment nodes to aggregate atom-level and bond-level semantics on an equal footing. We design three level-specific pretraining objectives: an atom-bond cross-view contrastive task that aligns the atom-view and bond-view representations within each fragment, a fragment-level functional group prediction task to inject domain-relevant chemical knowledge, and graph-level structure prediction tasks to encode global molecular topology. Experiments on nine MoleculeNet benchmarks demonstrate that MolCHG achieves the best performance on seven datasets across both classification and regression tasks, remaining competitive with the strongest baselines on the rest. Ablation studies further confirm that the multi-level supervision signals are complementary and that each component contributes to the overall performance.\n\n---\n\n## 73. Towards Trustworthy and Explainable AI for Perception Models: From Concept to Prototype Vehicle Deployment\n- **Authors**: Till Beemelmanns , Shayan Sharifi , Manas Mehrotra , Ayushman Choudhuri , Lutz Eckstein\n- **URL**: [https://arxiv.org/abs/2605.16087](https://arxiv.org/abs/2605.16087)\n- **Abstract**:\n> Deep Neural Networks have become the dominant solution for Autonomous Driving perception, but their opacity conflicts with emerging Trustworthy AI guidelines and complicates safety assurance, debugging, and human oversight. While theoretical frameworks for safe and Explainable AI (XAI) exist, concrete implementations of Trustworthy AI for 3D scene understanding remain scarce. We address this gap by proposing a Trustworthy AI perception module that is remarkably robust, integrates faithful explainability, and calibrated uncertainty estimates. Building on a transformer-based detector, we derive explanation from the attention mechanism at inference time and validate their faithfulness using perturbation-based consistency tests. We further integrate an uncertainty estimation and calibration module, and apply robustness-enhancing training methods. Experiments show faithful saliency behavior, improved robustness, and well-calibrated uncertainty estimates. Finally, we deploy these Trustworthy AI elements in a prototype vehicle and provide an XAI Interface that visualizes documentation artifacts, model uncertainty state, and saliency maps, demonstrating the feasibility of trustworthy perception monitoring in real time. Supplementary materials are available at this https URL .\n\n---\n\n## 74. Towards Foundation Models for Relational Databases with Language Models and Graph Neural Networks\n- **Authors**: Jingcheng Wu , Ratan Bahadur Thapa , Mojtaba Nayyeri , Lucas Etteldorf , Max Finkenbeiner , Fabian Leeske , Steffen Staab\n- **URL**: [https://arxiv.org/abs/2605.16085](https://arxiv.org/abs/2605.16085)\n- **Abstract**:\n> Relational databases store much of the world's structured information, and they are essential for driving complex predictive applications. However, deep learning progress on relational data remains limited, as conventional approaches flatten databases into single tables via manual feature engineering, discarding relational context. Relational deep learning (RDL) addresses this by modeling databases as relational entity graphs (REGs) for graph neural networks (GNNs), but remains task- and database-specific. To combine the strengths of both paradigms, we propose a hybrid architecture combining a fine-tuned BART encoder to capture intra-row semantics with a GraphSAGE-based GNN over REGs to inject relational context. Experiments on RelBench show that the GNN substantially enriches BART's row embeddings, achieving a ROC-AUC of 67.40 on the driver-dnf task from the rel-f1 dataset. This performance is competitive with supervised baselines such as LightGBM (68.86) and narrows the gap to RDL (72.62) to within 5.22 points, though a substantial gap remains to state-of-the-art foundation models such as KumoRFM (82.63). These results suggest that lightweight hybrid LM-GNN architectures offer a promising and resource-efficient path towards foundation models for relational databases.\n\n---\n\n## 75. VideoSeeker: Incentivizing Instance-level Video Understanding via Native Agentic Tool Invocation\n- **Authors**: Yiming Zhao , Yu Zeng , Wenxuan Huang , Zhen Fang , Qing Miao , Qisheng Su , Jiawei Zhao , Jiayin Cai , Lin Chen , Zehui Chen , Yukun Qi , Yao Hu , Xiaolong Jiang , Feng Zhao\n- **URL**: [https://arxiv.org/abs/2605.16079](https://arxiv.org/abs/2605.16079)\n- **Abstract**:\n> Large Vision-Language Models (LVLMs) have shown significant progress in video understanding, yet they face substantial challenges in tasks requiring precise spatiotemporal localization at the instance level. Existing methods primarily rely on text prompts for human-model interaction, but these prompts struggle to provide precise spatial and temporal references, resulting in poor user experience. Furthermore, current approaches typically decouple visual perception from language reasoning, centering reasoning around language rather than visual content, which limits the model's ability to proactively perceive fine-grained visual evidence. To address these challenges, we propose VideoSeeker, a novel paradigm for instance-level video understanding through visual prompts. VideoSeeker seamlessly integrates agentic reasoning with instance-level video understanding tasks, enabling the model to proactively perceive and retrieve relevant video segments on demand. We construct a four-stage fully automated data synthesis pipeline to efficiently generate large-scale, high-quality instance-level video data. We internalize tool-calling and proactive perception capabilities into the model via cold-start supervision and RL training, building a powerful video understanding model. Experiments demonstrate that our model achieves an average improvement of +13.7% over baselines on instance-level video understanding tasks, surpassing powerful closed-source models such as GPT-4o and Gemini-2.5-Pro, while also showing effective transferability on general video understanding benchmarks. The relevant datasets and code will be released publicly.\n\n---\n\n## 76. AgriMind: An Ensemble Deep Learning Framework for Multi-Class Plant Disease Classification\n- **Authors**: Salma Hoque Talukdar Koli , Fahima Haque Talukder Jely\n- **URL**: [https://arxiv.org/abs/2605.16076](https://arxiv.org/abs/2605.16076)\n- **Abstract**:\n> Plant disease detection is still largely manual in Bangladesh, where extension workers eyeball leaf samples across millions of smallholdings. We built AgriMind to automate this: an ensemble of ResNet50, EfficientNet-B0, and DenseNet121 trained on 20,638 PlantVillage images across 15 pepper, potato, and tomato disease classes. Transfer learning with frozen ImageNet backbones and 10 epochs of head-only training keeps the pipeline lightweight. Individual models hit 96--97% on the held-out test set, but averaging their softmax outputs pushes the ensemble to 99.23% -- a two-thirds cut in error rate. We tried biasing the average toward the best validation model; it backfired. Dropping any single model also hurt. Pepper and potato classify perfectly; tomato, with ten visually similar classes, still reaches 99.01%. On an NVIDIA T4 GPU the full ensemble runs at 53 FPS. Whether that translates to real-time mobile use depends on TensorFlow Lite optimization -- work we have not yet completed.\n\n---\n\n## 77. Robust Prior-Guided Segmentation for Editable 3D Gaussian Splatting\n- **Authors**: Raushan Joshi , Jean-Yves Guillemaut\n- **URL**: [https://arxiv.org/abs/2605.16065](https://arxiv.org/abs/2605.16065)\n- **Abstract**:\n> 3D Gaussian Splatting (3D-GS) enables real-time 3D scene reconstruction but lacks robust segmentation for editing tasks such as object removal, extraction, and recoloring. Existing approaches that lift 2D segmentations to the 3D domain suffer from view inconsistencies and coarse masks. In this paper, we propose a novel framework that leverages the Segment Anything Model High Quality (SAM-HQ) to generate accurate 2D masks, addressing the limitations of the standard SAM in boundary fidelity and fine-structure preservation. To achieve robust 3D segmentation of any target object in a given scene, we introduce a prior-guided label reassignment method that assigns labels to 3D Gaussians by enforcing multiview consistency with learned priors. Our approach achieves state-of-the-art segmentation accuracy and enables interactive, real-time object editing while maintaining high visual fidelity. Qualitative results demonstrate superior boundary preservation and practical utility in Virtual Reality (VR) and robotics, advancing 3D scene editing.\n\n---\n\n## 78. Misspecified Explore-then-Exploit Leads to Supra-Competitive Prices\n- **Authors**: Jackie Baek , Vivek F. Farias , Farrell Wu\n- **URL**: [https://arxiv.org/abs/2605.16064](https://arxiv.org/abs/2605.16064)\n- **Abstract**:\n> We study whether simple algorithmic pricing systems can systematically produce collusive-like prices in multi-firm markets. We consider firms using an explore-then-exploit pipeline: they randomize prices during an initial exploration phase, then estimate demand from their own historical data and set prices myopically thereafter. The estimation step relies on a misspecified, monopoly-style model that omits competitors' prices. We characterize when this pipeline converges to supra-competitive prices above the Nash equilibrium, via a fluid-limit ordinary differential equation analysis. We show that supra-competitive prices arise when firms explore within similar price ranges on the same side of the Nash price. Moreover, prices can be substantially above the Nash price; we show that prices can reach monopoly levels under symmetric exploration. Simulations calibrated to a real multifamily rental market confirm that supra-competitive outcomes arise robustly beyond our theoretical assumptions, including under finite horizons, heterogeneous products, and nonlinear logit demand.\n\n---\n\n## 79. Ada-Diffuser: Latent-Aware Adaptive Diffusion for Decision-Making\n- **Authors**: Fan Feng , Selena Ge , Minghao Fu , Zijian Li , Yujia Zheng , Zeyu Tang , Yingyao Hu , Biwei Huang , Kun Zhang\n- **URL**: [https://arxiv.org/abs/2605.16054](https://arxiv.org/abs/2605.16054)\n- **Abstract**:\n> Recent work has framed decision-making as a sequence modeling problem using generative models such as diffusion models. Although promising, these approaches often overlook latent factors that exhibit evolving dynamics, elements that are fundamental to environment transitions, reward structures, and high-level agent behavior. Explicitly modeling these hidden processes is essential for both precise dynamics modeling and effective decision-making. In this paper, we propose a unified framework that explicitly incorporates latent dynamic inference into generative decision-making from minimal yet sufficient observations. We theoretically show that under mild conditions, the latent process can be identified from small temporal blocks of observations. Building on this insight, we introduce Ada-Diffuser, a causal diffusion model that learns the temporal structure of observed interactions and the underlying latent dynamics simultaneously, and furthermore, leverages them for planning and control. With a modular design, Ada-Diffuser supports both planning and policy learning tasks, enabling adaptation to latent variations in dynamics, rewards, and latent actions. Experiments on simulated control and robotic benchmarks demonstrate its effectiveness in accurate latent inference and adaptive policy learning.\n\n---\n\n## 80. Looped SSMs: Depth-Recurrence and Input Reshaping for Time Series Classification\n- **Authors**: Mónika Farsang , Ramin Hasani , Daniela Rus , Radu Grosu\n- **URL**: [https://arxiv.org/abs/2605.16048](https://arxiv.org/abs/2605.16048)\n- **Abstract**:\n> State Space Models (SSMs) are inherently recurrent along the sequence dimension, yet depth-recurrence - reusing the same block repeatedly across layers, as recently applied in looped transformers - has not been explored in this model family. We show that a looped SSM with $k$ parameters iterated $L$ times consistently closely matches or outperforms a standard SSM with $k \\cdot L$ independent parameters across four architectures (LRU, S5, LinOSS, LrcSSM) and six time series classification benchmarks, despite operating within a strictly smaller hypothesis space, as we formally establish. Since the larger model contains the looped model as a special case, this dominance cannot be explained by expressivity and instead points to parameter sharing across depth as a beneficial inductive bias that simplifies optimization. These results demonstrate that depth-recurrence is orthogonal to sequence-recurrence and independently beneficial. We further show that input reshaping is an equally neglected design axis: concatenating timesteps for low-dimensional inputs, or flattening and rechunking the joint feature-time dimension for high-dimensional ones, yields accuracy gains of 1-6% across all models, confirmed over 5 random seeds. Both techniques provide standalone improvements that compound when combined, suggesting that depth and input reshaping are two independent and underexplored design axes for SSMs on time series.\n\n---\n\n## 81. XSearch: Explainable Code Search via Concept-to-Code Alignment\n- **Authors**: Yiming Liu , Ruofan Liu , Yun Lin , Zicong Zhang , Weiyu Kong , Pengnian Qi , Xiao Cheng , Weinan Zhang , Qianxiang Wang , Linpeng Huang\n- **URL**: [https://arxiv.org/abs/2605.16046](https://arxiv.org/abs/2605.16046)\n- **Abstract**:\n> Semantic code search has been widely adopted in both academia and industry. These approaches embed natural-language queries and code snippets into a shared embedding space and retrieve results based on vector similarity. Despit strong performance on benchmark datasets, they often suffer from poor explainability and generalization. Retrieved code may appear semantically similar yet miss critical functional requirements of the query, while providing no explanation of why the result was retrieved. Moreover, such failures become more severe under distribution shift, where models struggle to generalize to unseen benchmarks. In this work, we propose XSearch, an intrinsically explainable code search framework. Our key insight is that by relying on global embedding similarity, existing retrievers inherently take an inductive view. They learn statistical patterns rather than truly understanding the query's functional requirements. We address this problem by reformulating code search as a deductive concept alignment problem. XSearch (i) identifies functional concepts in the query and (ii) explicitly aligns them with corresponding code statements. This explain-then-predict design produces inherent concept-level explanations and mitigates shortcut learning that harms out-of-distribution generalization. We train an encoder with explicit concept-alignment objectives and perform retrieval through explicit matching between query concepts and code statements. Experiments show that, trained on CodeSearchNet using GraphCodeBERT (125M parameters), XSearch improves performance on out-of-distribution benchmarks from 0.02 to 0.33 (15x) over eight state-of-the-art retrievers, and consistently outperforms both encoder- and decoder-based baselines with up to 7B parameters. A user study demonstrates that concept-alignment explanations enable users to evaluate retrieved results faster and more accurately.\n\n---\n\n## 82. RecMem: Recurrence-based Memory Consolidation for Efficient and Effective Long-Running LLM Agents\n- **Authors**: Zijie Dai , Shiyuan Deng , Sheng Guan , Yizhou Tian , Xin Yao , Xiao Yan , James Cheng\n- **URL**: [https://arxiv.org/abs/2605.16045](https://arxiv.org/abs/2605.16045)\n- **Abstract**:\n> Memory systems often organize user-agent interactions as retrievable external memory and are crucial for long-running agents by overcoming the limited context windows of LLMs. However, existing memory systems invoke LLMs to process every incoming interaction for memory extraction, and such an eager memory consolidation scheme leads to substantial token consumption. To tackle this problem, we propose RecMem by rethinking when memory consolidation should be conducted. RecMem stores incoming interactions in a subconscious memory layer and encode them using lightweight embedding models for retrieval. LLMs are only invoked to extract episodic and semantic memory when sustained recurrence are observed for semantically similar interactions. Such recurrence-based consolidation works because these interactions correspond to a semantic cluster with rich information and thus are worth extraction and summarization. To improve accuracy, RecMem also incorporates a semantic refinement mechanism that recovers the fine-grained facts omitted by memory extraction. Experiments show that RecMem reduces the memory construction token cost of three SOTA memory systems by up to 87% while exceeding their accuracy.\n\n---\n\n## 83. Learning Sim-Grounded Policies for Bimanual Rope Manipulation from Human Teleoperation Data\n- **Authors**: Gina Wigginghaus , Tim Missal , Berk Guler , Simon Manschitz , Jan Peters\n- **URL**: [https://arxiv.org/abs/2605.16043](https://arxiv.org/abs/2605.16043)\n- **Abstract**:\n> Deformable Linear Objects (DLOs) such as ropes and cables are widely encountered in both household and industrial applications, yet remain challenging to manipulate due to their infinite-dimensional configuration space and frequent self-occlusion. Imitation learning from teleoperation offers a practical path to bimanual DLO manipulation, but its scalability is limited by human effort, making the choice of observation space critical for generalization from small datasets. In this study, we investigate whether the lack of generalization in egocentric visual policies for the knot-untangling task stems from the observation space itself, rather than from the policy architecture or data scale. We compare two Action Chunking with Transformers policies trained on the same bimanual teleoperation data: a vision-based policy conditioned on two egocentric RGB streams from wrist-mounted cameras, and a state-based policy conditioned on the DLO's 3D particle state, extracted from an initial observation via multi-view fusion and evolved in a particle-based eXtended Position-Based Dynamics simulation. Evaluated open-loop on an unseen rope configuration, the state-based policy outperforms its visual counterpart with a 30.8% reduction in L1 error when predicting the initial grasp-and-pull action, quantifying the observability gap between pixels and physics-consistent state, and pointing toward more data-efficient robot learning for the DLO manipulation task from limited human demonstrations.\n\n---\n\n## 84. Who Owns This Agent? Tracing AI Agents Back to Their Owners\n- **Authors**: Ruben Chocron , Doron Jonathan Ben Chayim , Eyal Lenga , Gilad Gressel , Alina Oprea , Yisroel Mirsky\n- **URL**: [https://arxiv.org/abs/2605.16035](https://arxiv.org/abs/2605.16035)\n- **Abstract**:\n> AI agents are increasingly deployed to act autonomously in the world, yet there is still no reliable way to trace a harmful agent back to the account that deployed it. This creates the same accountability gap across both ends of the intent spectrum: benign operators may deploy misconfigured or overbroad agents that cause harm unintentionally, while malicious operators may deliberately weaponize agents for scams, harassment, or cyber attacks. In many cases, these agents are powered by vendor-hosted models, a dependency that holds even for sophisticated adversaries such as state actors conducting cyber operations. In either case, affected parties can observe the behavior but cannot notify the responsible operator, stop the session, or identify the account for investigation. We formalize this gap as the problem of agent attribution: linking an observed agent interaction to the responsible account at the hosting vendor. To our knowledge, this is the first work to define the problem and present a practical solution. Our protocol is canary-based: an authorized party injects a canary into the agent's interaction stream, and the vendor searches a narrow window of session logs to recover the originating session and account. Simple canaries suffice in non-adversarial settings. For adversarial operators who filter or paraphrase incoming content, we develop robust canary constructions that cannot be suppressed without degrading the agent's own task performance, yielding a formal asymmetry in the defender's favor. We evaluate a variety of scenarios including real-world agents and show that our attribution method is reliable, robust, and scalable for vendor-side deployment.\n\n---\n\n## 85. From Flat Language Labels to Typological Priors: Structured Language Conditioning for Multilingual Speech-to-Speech Translation\n- **Authors**: Yu Pan , Yang Hou , Xiongfei Wu , Liang Zhang , Yves Le Traon , Lei Ma , Jianjun Zhao\n- **URL**: [https://arxiv.org/abs/2605.16026](https://arxiv.org/abs/2605.16026)\n- **Abstract**:\n> Compositional speech-to-speech translation (S2ST) systems built upon speech large language models (SpeechLLMs) have recently shown promising performance. However, existing S2ST systems often either neglect source-language information or encode it through a language-as-label paradigm, representing each source language as an independent flat embedding. Such a design overlooks systematic linguistic structure shared across languages, which may limit data-efficient multilingual adaptation when supervised S2ST data are scarce. To address this issue, we propose S2ST-Omni 2, a many-to-one compositional S2ST framework that systematically reformulates multilingual language conditioning from flat language labels to structured typological priors. Specifically, S2ST-Omni 2 revisits language conditioning at three levels: typology-informed hierarchical language encoding for structured source-language representation, dynamically-gated language-aware Dual-CTC for content-adaptive acoustic modulation, and typology-aware LLM prompting for decoder-side linguistic guidance. Experiments on CVSS-C show that S2ST-Omni 2 achieves superior average performance among representative S2ST approaches across BLEU, COMET, ASR-BLEU, and BLASER 2.0 under the adopted evaluation protocol. Ablation studies indicate that the proposed representation-level, acoustic-level, and decoding-level strategies provide complementary benefits. Moreover, controlled data-budget analyses and a Japanese-to-English evaluation using only approximately 3 hours of supervised training data suggest that explicit typological priors provide useful inductive biases for data-efficient multilingual S2ST.\n\n---\n\n## 86. Can Vision Language Models Be Adaptive in Mathematics Education? A Learner Model-based Rubric Study\n- **Authors**: Jie Gao , Yongan Yu , Junzhu Su , Yiran Lin , Adam K. Dube , Jackie Chi Kit Cheung\n- **URL**: [https://arxiv.org/abs/2605.16011](https://arxiv.org/abs/2605.16011)\n- **Abstract**:\n> Adaptive learning refers to educational technologies that track learners' learning progress and adapt the instructional process based on individual learners' learning performance. It is increasingly recognized as critical for developing an effective learning support tool. Vision language models (VLMs) have seen adoption in mathematics education, and students have been using them as learning aids for personalized instruction. However, it is unknown whether VLMs have the ability to adapt to different learner profiles when providing mathematical instructions. Current VLMs lack a systematic evaluation framework for this adaptivity to different learner profiles in mathematics tutoring tasks. To address this gap, we draw on the learner model from the adaptive learning framework (Shute and Towle, 2018) and propose a learner model-based rubric. Our rubric formalizes adaptivity assessment into three aspects: cognitive aspects, motivational aspects, and complexity. We also evaluate two additional dimensions of VLM responses: correctness (of answers and solutions) and quality (of the response itself). Our experimental results show measurable differences in adaptivity across models and also reveal that current VLMs struggle to consistently produce learner model-based instructional responses, especially when receiving limited learner information.\n\n---\n\n## 87. CitePrism: Human-in-the-Loop AI for Citation Auditing and Editorial Integrity\n- **Authors**: Gowrika Mahesh , Budanur Madappa Darshan Gowda , Kavana Gopladevarahalli Papegowda , Prajwal Basavaraj , Binh Vu , Swati Chandna , Mehrdad Jalali\n- **URL**: [https://arxiv.org/abs/2605.16000](https://arxiv.org/abs/2605.16000)\n- **Abstract**:\n> Editors and reviewers are expected to ensure that manuscripts cite relevant, accurate, current, and ethically appropriate literature, yet manuscript-level citation auditing remains largely manual, fragmented, and difficult to scale. Citation context, metadata quality, self-citation patterns, and bibliographic integrity all affect whether a reference appropriately supports a local claim. We present CitePrism, a transparent hybrid decision-support framework for editorial citation auditing that combines LLM-assisted contextual reasoning, embedding-based semantic similarity, metadata verification, integrity-oriented flags, and human-in-the-loop analyst review. CitePrism extracts citation neighborhoods, enriches reference metadata, computes fused relevance scores, surfaces metadata and self-citation review prompts, and supports configurable threshold-based triage. In a preliminary validation on a single case-study manuscript with 104 references from pavement engineering, agreement with human binary relevance labels reached Cohen's kappa = 0.429. At operating threshold tau = 17, CitePrism flagged all human-labeled irrelevant citations, while also producing false positives requiring analyst review. These results suggest that CitePrism may support conservative editorial screening and citation-quality triage, but they do not establish general editorial performance. CitePrism is intended as pilot-stage decision support, not as an autonomous misconduct detector or automated editorial decision system. Broader validation across manuscripts, domains, annotators, baselines, and deployment settings is required before operational use.\n\n---\n\n## 88. Constrained latent state modeling: A unifying perspective on representation learning under competing constraints\n- **Authors**: Gwenolé Quellec\n- **URL**: [https://arxiv.org/abs/2605.15995](https://arxiv.org/abs/2605.15995)\n- **Abstract**:\n> Learning latent representations from complex data is central to modern machine learning, spanning temporal, multimodal, and partially observed systems. In such settings, representations are better understood as latent states capturing underlying system dynamics, rather than as mere compressed summaries of observations. Yet current approaches remain fragmented, relying on distinct -- and often implicit -- assumptions about what these states should represent. We argue that this fragmentation reflects a more fundamental limitation: latent representations are typically learned from underconstrained objectives that fail to specify the properties that meaningful latent states should satisfy. As a result, multiple representations can satisfy the same objective, leading to ambiguity in their structure and interpretation. While many of the underlying principles have been explored in isolation, their interactions have not been explicitly formalized. In this work, we propose constrained latent state modeling (CLSM) as a unifying perspective. We identify a set of core properties -- predictive sufficiency, minimality, temporal coherence, observation compatibility, invariance to nuisance factors, and structural constraints -- and show that they are intrinsically coupled through fundamental trade-offs. Revisiting major modeling families through this lens, we show that existing approaches can be interpreted as enforcing different subsets of constraints, thereby occupying distinct regions of a common design space. This perspective reframes persistent challenges such as lack of identifiability as consequences of underconstrained formulations, rather than isolated technical limitations. More broadly, CLSM provides a principled framework to make design choices explicit, to analyze trade-offs, and to guide the development of more interpretable, robust, and task-aligned latent state models.\n\n---\n\n## 89. Beyond Content: A Comprehensive Speech Toxicity Dataset and Detection Framework Incorporating Paralinguistic Cues\n- **Authors**: Zhongjie Ba , Liang Yi , Peng Cheng , Qingcao Li , Qinglong Wang , Li Lu\n- **URL**: [https://arxiv.org/abs/2605.15984](https://arxiv.org/abs/2605.15984)\n- **Abstract**:\n> Toxic speech detection has become a crucial challenge in maintaining safe online communication environments. However, existing approaches to toxic speech detection often neglect the contribution of paralinguistic cues, such as emotion, intonation, and speech rate, which are key to detecting speech toxicity. Moreover, current toxic speech datasets are predominantly text-based, limiting the development of models that can capture paralinguistic this http URL address these challenges, we present ToxiAlert-Bench, a large-scale audio dataset comprising over 30,000 audio clips annotated with seven major toxic categories and twenty fine-grained toxic labels. Uniquely, our dataset annotates toxicity sources -- distinguishing between textual content and paralinguistic origins -- for comprehensive toxic speech this http URL , we propose a dual-head neural network with a multi-stage training strategy tailored for toxic speech detection. This architecture features two task-specific classification headers: one for identifying the source of sensitivity (textual or paralinguistic), and the other for categorizing the specific toxic type. The training process involves independent head training followed by joint fine-tuning to reduce task interference. To mitigate data class imbalance, we incorporate class-balanced sampling and weighted loss this http URL experimental results show that leveraging paralinguistic features significantly improves detection performance. Our method consistently outperforms existing baselines across multiple evaluation metrics, with a 21.1% relative improvement in Macro-F1 score and a 13.0% relative gain in accuracy over the strongest baseline, highlighting its enhanced effectiveness and practical applicability.\n\n---\n\n## 90. Ontology for Policing: Conceptual Knowledge Learning for Semantic Understanding and Reasoning in Law Enforcement Reports\n- **Authors**: Anita Srbinovska , Jansen Orfan , Adrian Martin , Ernest Fokoué\n- **URL**: [https://arxiv.org/abs/2605.15978](https://arxiv.org/abs/2605.15978)\n- **Abstract**:\n> Law enforcement reports contain structured fields and written narratives. However, many incident facts that are needed for review, police training, and investigations are in natural language and require manual reading. We propose a framework using symbolic methods for converting narratives into evidence-linked facts. Our objective is to measure the value of narratives to recover incident details only from the unstructured text and build temporal graphs with time cues and domain axioms. We achieve this by redacting personal identifiers, semantic parsing, predicate mapping to ontology, and reasoning. We evaluate the symbolic approach on 450 property crime reports and a short human review. Of the extracted events from the system, 54.1% had a confidence score of at least 0.80 and 93.7% were mapped through the PropBank--VerbNet--WordNet semantic path. 100% agreement was reached on incident initiation, stolen items, and temporal cues and lower agreement for forced entry interpretation.\n\n---\n\n## 91. Reference-Free Reinforcement Learning Fine-Tuning for MT: A Seq2Seq Perspective\n- **Authors**: Ernesto Garcia-Estrada , Carlos Escolano , José A. R. Fonallosa\n- **URL**: [https://arxiv.org/abs/2605.15976](https://arxiv.org/abs/2605.15976)\n- **Abstract**:\n> Production machine translation relies overwhelmingly on encoder-decoder Seq2Seq models, yet reinforcement learning approaches to MT fine-tuning have largely targeted decoder-only LLMs at $\\geq$7B parameters, with limited systematic study of encoder-decoder architectures. We apply Group Relative Policy Optimization to NLLB-200 (600M and 1.3B) using a hybrid reference-free reward (LaBSE and COMET-Kiwi) that requires no parallel data at fine-tuning time, evaluating across 13 typologically diverse languages. GRPO yields consistent improvements on all 13 languages, up to $+$5.03 chrF++ for Traditional Chinese, and, without any target-language data, competes with 3-epoch supervised fine-tuning on morphologically complex languages . We identify a consistent empirical pattern in which gains are largest where baseline performance is weakest and reward discriminability is highest, making this approach most effective precisely where parallel data is scarcest, and replicate this pattern across English and Spanish source languages.\n\n---\n\n## 92. When and Why Adversarial Training Improves PINNs: A Neural Tangent Kernel Perspective\n- **Authors**: Yuan-dong Cao , Chi Chiu SO , Jun-Min Wang , He Wang\n- **URL**: [https://arxiv.org/abs/2605.15959](https://arxiv.org/abs/2605.15959)\n- **Abstract**:\n> Physics-informed neural networks (PINNs) are powerful surrogates for differential equations but are notoriously difficult to train due to spectral bias, stiffness, and poor accuracy on high-frequency or multiscale solutions. Adversarial training based on generative adversarial networks (GANs) has recently gained surprisingly strong empirical results in improving training, but the underlying mechanisms remain elusive. To this end, we propose a new analysis framework for adversarially trained PINNs, based on the key observation of how the discriminator in GANs can influence the training dynamics of PINNs. The framework first provides a much needed theoretical grounding to why and when adversarial training is effective in PINNs, then presents a unified analysis of GANs variants in such training, and finally leads to a new, practical, efficient training algorithm for PINNs. Empirical results demonstrate that our method can significantly reduce the pathology of PINNs training, thereby providing better models with superior performances, often several magnitudes more accurate than alternative methods.\n\n---\n\n## 93. Decomposed Vision-Language Alignment for Fine-Grained Open-Vocabulary Segmentation\n- **Authors**: Chenhao Wang , Yingrui Ji , Yu Meng , Yao Zhu\n- **URL**: [https://arxiv.org/abs/2605.15942](https://arxiv.org/abs/2605.15942)\n- **Abstract**:\n> Open-vocabulary segmentation models often struggle to generalize to unseen combinations of object categories and attributes, because fine-grained descriptions are typically encoded as holistic sentences that entangle multiple semantic units. We propose a Decomposed Vision-Language Alignment framework that explicitly factorizes textual prompts into a concept token and multiple attribute tokens, enabling separate cross-modal interactions for each semantic unit. At the feature level, we introduce a Feature-Gated Cross-Attention module that generates attribute-specific gating maps to fuse information in a multiplicative manner, effectively enforcing compositional semantics. At the scoring level, per-token similarities are aggregated in log-space, producing a stable and interpretable compositional matching. The method can be seamlessly integrated into existing transformer-based segmentation architectures and significantly improves generalization to unseen attribute-category compositions in fine-grained open-vocabulary segmentation benchmarks.\n\n---\n\n## 94. LoCO: Low-rank Compositional Rotation Fine-tuning\n- **Authors**: An Nguyen , Jaesik Choi , Anh Tong\n- **URL**: [https://arxiv.org/abs/2605.15916](https://arxiv.org/abs/2605.15916)\n- **Abstract**:\n> Parameter-efficient fine-tuning (PEFT) has emerged as an critical technique for adapting large-scale foundation models across natural language processing and computer vision. While existing methods such as low-rank adaptations achieve parameter efficiency via low-rank weight updates, they are limited in their ability to preserve the geometric structure of pretrained representations. We introduce Low-rank Compositional Orthogonal fine-tuning (LoCO), a novel PEFT method that constructs orthogonal transformations through low-rank skew-symmetric matrices and compositional rotation chains. We propose an approximation scheme that enables fully parallel computation of compositional rotations, making the approach practical for high-dimensional feature spaces. Our method maintains low computational complexity while maintaining orthogonality with controlled approximation error. We validate LoCO across diverse domains, including diffusion transformer fine-tuning, vision transformer adaptation, and language model adaptation. Our method demonstrates superior or competitive performance compared to both existing orthogonal and non-orthogonal methods.\n\n---\n\n## 95. SLIP & ETHICS: Graduated Intervention for AI Emotional Companions\n- **Authors**: Minseo Kim\n- **URL**: [https://arxiv.org/abs/2605.15915](https://arxiv.org/abs/2605.15915)\n- **Abstract**:\n> AI emotional companions face a safety-rapport paradox: restrictive safeguards can damage supportive alliance, while permissive systems risk user harm. We present SLIP (Staged Layers of Intervention Protocol), a four-stage graduated methodology deriving interventions (none, soft, hard) from structured qualitative indicators -- affect intensity (a) and narrative dynamism (m) -- alongside ETHICS (Emergent Taxonomy for Human-AI Interaction Context Signals), a \"signals not labels\" taxonomy. An evaluation combining a small-scale production deployment (N=68 entries, 10 users, 10 weeks) with a synthetic persona battery (N=91, 5 behavioral-risk profiles) achieved 0% false positives for the flow persona and showed expected escalation patterns in crisis-oriented personas. However, initial results showed that 8 consecutive days of high-energy elevation produced zero interventions (0/8), exposing a boundary where the \"do not pathologize\" principle conflicts with safety. A subsequent three-model stress test demonstrated that increased model capability improves detection from 0/8 to 6/8 while preserving 0/10 flow false positives in the largest model. Read as preliminary, these findings position graduated intervention as a design direction for navigating -- not resolving -- the safety-rapport tension in affective computing.\n\n---\n\n## 96. Towards Generalization of Block Attention via Automatic Segmentation and Block Distillation\n- **Authors**: Shuaiyi Li , Zhisong Zhang , Yan Wang , Lei Zhu , Dongyang Ma , Chenlong Deng , Yang Deng , Wai Lam\n- **URL**: [https://arxiv.org/abs/2605.15913](https://arxiv.org/abs/2605.15913)\n- **Abstract**:\n> Block attention, which processes the input as separate blocks that cannot attend to one another, offers significant potential to improve KV cache reuse in long-context scenarios such as Retrieval-Augmented Generation (RAG). However, its broader application is hindered by two key challenges: the difficulty of segmenting input text into meaningful, self-contained blocks, and the inefficiency of existing block fine-tuning methods that risk degrading performance. To address these, we first construct SemanticSeg, a large and diverse semantic segmentation dataset containing over 30k instances across 16 categories-including books, code, web text, and conversations with text lengths ranging from 2k to 32k. Using this dataset, we train a lightweight segmenter to automatically partition text into human-instinct-aligned blocks with controllable granularity. Second, we propose block distillation, a training framework that is more efficient than block fine-tuning, which uses a frozen full-attention teacher model to guide the block-attention student. This framework integrates three novel components: block sink tokens to mitigate information loss at block boundaries, block dropout to leverage training signals from all blocks, and token-level loss weighting to focus learning on block-attention-sensitive tokens. Experiments across multiple models and benchmarks demonstrate that our segmenter outperforms heuristic and statistical baselines, and block distillation achieves near-full-attention performance under block attention, establishing a practical and scalable pathway for deploying block attention.\n\n---\n\n## 97. RaPD: Resolution-Agnostic Pixel Diffusion via Semantics-Enriched Implicit Representations\n- **Authors**: Yanhao Ge , Shanyan Guan , Weihao Wang , Ying Tai , Mingyu You\n- **URL**: [https://arxiv.org/abs/2605.15908](https://arxiv.org/abs/2605.15908)\n- **Abstract**:\n> Natural images are continuous, yet most generative models synthesize them on discrete grids, limiting resolution-flexible generation. Continuous neural fields enable resolution-free rendering, but prior methods introduce continuity only at the decoding stage as an interpolation module, leaving the generative latent space discretized and reconstruction-oriented. We propose RaPD (Resolution-agnostic Pixel Diffusion), which performs diffusion in a continuous Neural Image Field (NIF) latent space. RaPD bridges this reconstruction-generation gap with Semantic Representation Guidance for generation-aware latent learning and a Coordinate-Queried Attention Renderer for coordinate-conditioned, scale-aware rendering. A single denoised latent can be rendered at arbitrary resolutions by changing only the query coordinates, keeping diffusion cost fixed. Experiments demonstrate superior generation quality and resolution scalability.\n\n---\n\n## 98. Generative Long-term User Interest Modeling for Click-Through Rate Prediction\n- **Authors**: Jiangli Shao , Kaifu Zheng , Hao Fang , Huimu Ye , Zhiwei Liu , Bo Zhang , Shu Han , Xingxing Wang\n- **URL**: [https://arxiv.org/abs/2605.15905](https://arxiv.org/abs/2605.15905)\n- **Abstract**:\n> Modeling long-term user interests with massive historical user behaviors enhances click-through rate (CTR) prediction performance in advertising and recommendation systems. Typically, a two-stage framework is widely adopted, where a general search unit (GSU) first retrieves top-$k$ relevant behaviors towards the target item, and an exact search unit (ESU) generates interest features via tailored attention. However, current target-centered GSU would ignore other latent user interests, leading to incomplete and biased interest features. Additionally, the matching-based retrieval process in GSUs depends on the pairwise similarity score between target item and each historical behavior, which not only becomes time-consuming for online services as user behaviors continue to grow, but also overlooks the interaction information among user behaviors. To combat these problems, we propose a \\textbf{Gen}erative \\textbf{L}ong-term user \\textbf{I}nterest model named GenLI for CTR prediction. GenLI consists of an interest generation module (IGM), a behavior retrieval module (BRM), and an interest fusion module (IFM). The IGM generates multiple interest distributions to indicate different aspects of real-time user interests, which is target-independent and incorporates interaction information among behaviors, ensuring complete and diverse interest features. The BRM selects related behaviors via a simple lookup operation, reducing the time complexity for weighting each behavior to $O(1)$. Finally, the IFM uses delicate gating mechanisms to generate interest features. Based on the generation process, GenLI improves the diversity of user interests and avoids complex matching-based behavioral retrieval, achieving a better balance between accuracy and efficiency for CTR prediction.\n\n---\n\n## 99. Uncertainty-Aware Wildfire Smoke Density Classification from Satellite Imagery via CBAM-Augmented EfficientNet with Evidential Deep Learning\n- **Authors**: Ranjith Chodavarapu\n- **URL**: [https://arxiv.org/abs/2605.15894](https://arxiv.org/abs/2605.15894)\n- **Abstract**:\n> Rapid and accurate wildfire smoke severity assessment from satellite images is essential for emergency response, air quality modeling, and human health risk management. Existing deep learning approaches treat smoke detection as a binary task, producing point estimates without any measure of prediction confidence. We propose a probabilistic framework to categorize a satellite patch into Light, Moderate, and Heavy severity classes and to provide decomposed epistemic and aleatoric uncertainty in a single forward pass. Our architecture uses the backbone of a pre-trained EfficientNet-B3 and a CBAM module with an evidential deep learning head that predicts Dirichlet concentration parameters, directly estimating vacuity (epistemic) and dissonance (aleatoric) without Monte Carlo sampling. Evaluated on 16,298 real satellite patches derived from the Wildfire Detection dataset, our model achieves 93.8% weighted test accuracy (91.1% unweighted) with ECE=0.0274. Selective prediction retaining the most certain 50% of patches achieves 96.7% accuracy. As image quality degrades, uncertainty increases monotonically, and vacuity is a practical scan quality measure. The Moderate class represents transitional smoke conditions that exhibit the highest epistemic uncertainty (mean vacuity = 0.187), confirming the model correctly identifies ambiguous smoke boundary regions. CBAM spatial attention maps localize to structurally distinctive scene regions, and t-SNE demonstrates the clear cluster separation of Light and Heavy smoke.\n\n---\n\n## 100. CHoE: Cross-Domain Heterogeneous Graph Prompt Learning via Structure-Conditioned Experts\n- **Authors**: Peiyuan Li , Yongqi Huang , Jitao Zhao , Dongxiao He , Di Jin , Weixiong Zhang\n- **URL**: [https://arxiv.org/abs/2605.15888](https://arxiv.org/abs/2605.15888)\n- **Abstract**:\n> Heterogeneous Graph Prompt Learning (HGPL)has emerged as a promising paradigm for bridging the gap between the objectives of pre-training foundation models and their downstream applications in heterogeneous graph settings. However, existing HGPL methods are primarily designed for in-domain scenarios, whereas real-world deployments often span multiple domains, and the data used for pre-training and downstream tasks may originate from different distributions. Consequently, the applicability of current HGPL approaches is limited to in-domain settings, and their performance typically degrades when application domains shift. To address this serious limitation, we develop CHoE, a cross-domain HGPL method built upon an expert network. During pre-training, we introduce and train structure-conditioned experts, and during prompt tuning, we adopt a structure-aware expert routing and load balancing mechanism to select structurally compatible experts for each meta-path view. In addition, we design a prompt-based semantic fusion module to integrate representations across multiple views for downstream prediction. Extensive experiments show that CHoE consistently improves performance in few-shot cross-domain applications, outperforming all baseline approaches.\n\n---\n\n## 101. Symplectic Neural Operators for Learning Infinite Dimensional Hamiltonian Systems\n- **Authors**: Yeang Makara , Yusuke Tanaka , Takashi Matsubara , Takaharu Yaguchi\n- **URL**: [https://arxiv.org/abs/2605.15881](https://arxiv.org/abs/2605.15881)\n- **Abstract**:\n> The modeling and simulation of infinite-dimensional Hamiltonian systems are central problems in mathematical physics and engineering, however they pose significant computational and structural challenges for standard data-driven architectures. In this work, we introduce the Symplectic Neural Operator, a neural operator architecture designed to preserve the symplectic structure intrinsic to Hamiltonian PDEs. We provide a theoretical characterization of their symplecticity and establish a rigorous long-term stability result based on the combination of symplectic structure preservation and learning accuracy. Numerical experiments on canonical Hamiltonian PDEs corroborate this theoretical result and show that SNOs exhibit improved energy behavior compared with non-structure-preserving neural operators.\n\n---\n\n## 102. FSCM: Frequency-Enhanced Spatial-Spectral Coupled Mamba for Infrared Hyperspectral Image Colorization\n- **Authors**: Tingting Liu , Yuan Liu , Guiping Chen , Xiubao Sui , Qian Chen\n- **URL**: [https://arxiv.org/abs/2605.15880](https://arxiv.org/abs/2605.15880)\n- **Abstract**:\n> Thermal infrared imaging is robust to illumination variations and smoke interference, making it important for all-weather perception. However, the lack of natural color and fine texture limits target recognition, human visual interpretation, and the transfer of visible-light models. Existing infrared colorization methods mainly rely on single-band images, where insufficient spectral cues may lead to structural distortion and semantic confusion. Although infrared hyperspectral images provide rich spectral responses and material information, existing single-band frameworks remain limited in modeling spatial-spectral coupling and weak texture details. To address these issues, this paper presents FSCM, a spectral-information-guided GAN framework. Within FSCM, a frequency-enhanced spatial-spectral state-space generator composed of cascaded FSB units is constructed. Each FSB integrates three complementary components: state-space modeling captures global spatial-spectral dependencies; the frequency enhancement module (FEM) combines multi-level wavelet decomposition and Fourier gating to recover structural contours, directional high-frequency details, and global frequency responses; and the dual-stream hybrid gating module (DGM) integrates deformation-aware sampling with sparse attention to enhance effective local structures and suppress background interference. Additionally, an online semantic segmentation-guided loss is introduced to constrain the generated results, improving semantic consistency in complex road scenes. Experiments show that FSCM outperforms existing infrared colorization methods in visual quality and semantic fidelity.\n\n---\n\n## 103. Shapley Neuron Values for Continual Learning: Which Neurons Matter Most?\n- **Authors**: Mohammad Ali Vahedifar , Abhisek Ray , Qi Zhang\n- **URL**: [https://arxiv.org/abs/2605.15877](https://arxiv.org/abs/2605.15877)\n- **Abstract**:\n> Continual learning enables neural networks to learn tasks sequentially without forgetting previously acquired knowledge. However, neural networks suffer from catastrophic forgetting, where learning new tasks degrades performance on earlier ones. We address this problem with Shapley Neuron Valuation (SNV), a principled framework that quantifies Neuron importance in continual learning, grounded in cooperative game theory. SNV selectively freezes important Neurons while keeping others plastic, enabling buffer-free continual learning without expanding architecture. Experiments on ImageNet-1k show that SNV consistently outperforms existing buffer-free methods. In particular, SNV improves accuracy by +2.88% in the class incremental learning and +6.46% in the task incremental learning scenarios compared to the second baseline.\n\n---\n\n## 104. Access Timing as Scaffolding: A Reinforcement Learning Approach to GenAI in Education\n- **Authors**: Janne Rotter , Pau Benazet i Montobbio , Davinia Hernández-Leo\n- **URL**: [https://arxiv.org/abs/2605.15850](https://arxiv.org/abs/2605.15850)\n- **Abstract**:\n> In recent years, generative AI (GenAI) in educational settings has become ubiquitous in students' daily lives, despite its potential to induce over-reliance, metacognitive disengagement, and diminished learning when used unrestrictedly. While most prior research has thus focused on how to pedagogically scaffold its usage, the question of when to allow off-the-shelf GenAI remains understudied and lacks pedagogically grounded empirical investigation. We treat access timing itself as a form of implicit scaffolding and operationalize it through a reinforcement learning (RL) agent that decides when students should access GenAI, with a reward function grounded in metacognitive theory, cognitive load theory, and productive failure. In a mixed-methods controlled lab study with N=105 participants, we compared the agent's effect on learning gains and metacognitive engagement to unrestricted and fully restricted use. Results show that strategically timed GenAI access under the reinforcement learning condition improved objective post-test performance and metacognitive accuracy compared with unrestricted access, while reducing task errors and time on task relative to complete withholding, all without the need for explicit metacognitive prompts or structured scaffolding. However, no between-condition differences emerged on self-reported metacognitive awareness. Overall, timing of GenAI access therefore is a tractable, theoretically grounded, and scalable pedagogical paradigm that improves over completely unrestricted and withheld access, compatible with off-the-shelf tools and potentially low adoption barrier. This opens up a new research area that explores how access timing can be facilitated by educators and implemented in human-AI learning system design.\n\n---\n\n## 105. RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades\n- **Authors**: Xinbo Xu , Ruihan Yang , Haiyang Shen , Wendong Xu , Bofei Gao , Ruoyu Wu , Kean Shi , Weichu Xie , Xuanzhong Chen , Ming Wu , Jason Zeng , Michael Heinrich , Elvis Zhang , Liang Chen , Kuan Li , Baobao Chang\n- **URL**: [https://arxiv.org/abs/2605.15846](https://arxiv.org/abs/2605.15846)\n- **Abstract**:\n> Coding agents are increasingly deployed in real software development, where a single version iteration requires months of coordinated work across many files. However, most existing benchmarks focus predominantly on single-issue bug fixes from Python repositories, with coarse pass/fail evaluation outcomes, and thus fail to capture long-horizon, multi-target development at real engineering scale. To address this gap, we present RoadmapBench, a benchmark of 115 long-horizon coding tasks grounded in real open-source version upgrades across 17 repositories and 5 programming languages. Each task places the agent on a source-version code snapshot and provides a multi-target roadmap instruction requiring it to implement the functionality introduced in the target version, with a median modification of 3,700 lines across 51 files. We conduct a systematic evaluation on thirteen frontier models and find that even the strongest, Claude-Opus-4.7, resolves only 39.1% of tasks, while the weakest achieves merely 5.2%, in stark contrast to existing bug-fix benchmarks, suggesting that long-horizon software development remains a largely unsolved problem.\n\n---\n\n## 106. GAP: Geometric Anchor Pre-training for Data-Efficient Visuomotor Learning of Manipulation Tasks\n- **Authors**: Davide Buoso , Andrea Protopapa , Stefano Di Carlo , Francesca Pistilli , Giuseppe Averta\n- **URL**: [https://arxiv.org/abs/2605.15836](https://arxiv.org/abs/2605.15836)\n- **Abstract**:\n> Learning visuomotor policies from scarce expert demonstrations remains a core challenge in robotic manipulation. A primary hurdle lies in distilling high-dimensional RGB representations into control-relevant geometry without overfitting. While using frozen pre-trained Vision Foundation Models (VFMs) improves data efficiency, it also shifts most task adaptation onto a small spatial pooling module, which can latch onto task-irrelevant shortcuts and lose geometric grounding when finetuned with few data samples. More broadly, pre-trained visual representations used for policy learning have been observed to struggle under even minor scene perturbations, highlighting the need for robustness-oriented inductive biases. We propose Geometric Anchor Pre-training (GAP), a simple, action-free warm-up stage that regularizes the spatial adapter before downstream imitation learning. GAP pre-trains the pooling layer on a lightweight simulated proxy task where object masks are available at no cost, encouraging the adapter to produce keypoints that lie on the object, cover its spatial extent, and remain sharp and repeatable over time. This yields stable geometric anchors that provide a reliable coordinate interface for few-shot policy learning, while keeping the VFM frozen. We evaluate GAP on RoboMimic and ManiSkill under severe data scarcity (15-50 demonstrations) and domain shift. A simple adapter regularized with GAP consistently outperforms stronger attention-based poolers and end-to-end fine-tuning, achieving 62% success on RoboMimic Can with 15 demonstrations (+16% over AFA), 63% on the long-horizon high-precision Tool Hang task with 50 demonstrations, and 61% on ManiSkill StackCube with 30 demonstrations (+11% over full fine-tuning). The proxy stage is lightweight and fully decoupled from downstream tasks, making it practical to reuse across environments and manipulation skills.\n\n---\n\n## 107. Modeling Music as a Time-Frequency Image: A 2D Tokenizer for Music Generation\n- **Authors**: Yuqing Cheng , Xingyu Ma , Guochen Yu , Xiaotao Gu\n- **URL**: [https://arxiv.org/abs/2605.15831](https://arxiv.org/abs/2605.15831)\n- **Abstract**:\n> Autoregressive music generation depends strongly on the audio tokenizer. Existing high-fidelity codecs often use residual multi-codebook quantization, which preserves reconstruction quality but complicates language modeling after sequence flattening, as the residual hierarchy imposes strong sequential dependencies and can amplify error accumulation. We propose BandTok, a generation-oriented 2D Mel-spectrogram tokenizer that represents each frame with Mel-frequency band tokens from a single shared codebook. This design yields a physically interpretable time-frequency token grid with a more independent token structure, making it better suited for autoregressive modeling. BandTok improves reconstruction with a multi-scale PatchGAN objective and EMA codebook updates. We further introduce an autoregressive language model with 2D Rotary Position Embedding (2D RoPE) to preserve temporal and frequency-band structure during generation. Experiments show that BandTok improves over residual-codebook tokenizers and achieves strong results in a data-limited setting. The source code and generation demos for this work are publicly available.\n\n---\n\n## 108. Toward Natural and Companionable Virtual Agents via Cross-Temporal Emotional Modeling\n- **Authors**: Feier Qin , Xiao Li , Yi Zheng , Haibin Huang , Hanyao Wang , Xiaoyu Wang , Yan Lu , Yuan Zhang\n- **URL**: [https://arxiv.org/abs/2605.15812](https://arxiv.org/abs/2605.15812)\n- **Abstract**:\n> Recent advances in foundation models have enabled conversational agents that aim for sustained companionship rather than mere task completion. Yet most still remain unable to support natural, long-term companion-like interactions, resulting in experiences that feel episodic and inauthentic. We argue that current agents overlooked cross-temporal modeling of agents' social behaviors and internal emotions: generated behaviors rarely influence an agent's emotional state, and emotional states seldom shape subsequent behaviors. We present Cross-Temporal Emotion Modeling (CTEM), a framework that links long-term behavioral history to moment-to-moment emotional expression. CTEM establishes a closed loop where past experiences update an evolving emotional state; this state conditions immediate interactions; and user feedback continually revises both memory and emotional state, enabling reflection and anticipation. We instantiate CTEM as Auri, a companion agent on an instant-messaging platform, and report a 21-day in-the-wild study showing that CTEM shows improvements in perceived naturalness, coherence, and emotional harmony.\n\n---\n\n## 109. Grokking as Structural Inference: Transformers Need Bayesian Lottery Tickets\n- **Authors**: Kai Hidajat , Solden Stoll , Joseph An\n- **URL**: [https://arxiv.org/abs/2605.15787](https://arxiv.org/abs/2605.15787)\n- **Abstract**:\n> Why does a Transformer that has memorized its training set wait thousands of steps before it generalizes? Existing accounts locate this delay in norm minimization, feature emergence, or the late discovery of sparse subnetworks. These explanations capture important parts of the transition, but ignore a constraint unique to attention-based models: if attention discards an informative token, no bounded downstream computation can recover it. We formalize attention as an implicit Bayesian posterior over the task dependency graph and prove that generalization requires two separable conditions: a familiar Goldilocks bound on MLP capacity, coinciding with norm-based theories of grokking, and a novel Bayesian structural condition requiring attention to place sufficient mass on every informative token. This decoupling explains delayed generalization as delayed structural inference. Early in training, the MLP memorizes through unaligned features, drives the cross-entropy loss near zero, and thereby starves attention of structural gradient. Weight decay must then erode memorization before the missing graph becomes learnable, yielding the known inverse-weight-decay delay, which we derive as a structural waiting time. We then prove that this explaining-away delay can be bypassed by a KL-based structural intervention, yielding an inverse-intervention-strength scaling law for the grokking time. Experiments on algorithmic sequence tasks isolate structure from capacity and show that this Bayesian ticket matches or outperforms lottery-ticket transfer.\n\n---\n\n## 110. A Topology-Aware Spatiotemporal Handover Framework for Continuous Multi-UAV Tracking\n- **Authors**: Jianlin Ye , Christos Kyrkou , Panayiotis Kolios\n- **URL**: [https://arxiv.org/abs/2605.15779](https://arxiv.org/abs/2605.15779)\n- **Abstract**:\n> The integration of Unmanned Aerial Vehicles(UAVs) into Intelligent Transportation Systems (ITS) offers synoptic visibility for traffic monitoring, yet scalable deployment is hindered by trajectory fragmentation, where vehicle identity persistence is lost across multi-UAV Fields of View (FOV). While state-of-the-art frameworks excel in optimizing local trajectory extraction and stability for single-drone imagery, they often function as isolated data silos that generate disjointed trajectories, thereby precluding network-level analysis such as Origin-Destination estimation. This paper presents a real-time Multi-Camera Multi-Vehicle Tracking (MCMT) system designed to handle global identity persistence. Addressing the visual ambiguity and computational cost of appearance-based Re-Identification (Re-ID) in nadir views, we introduce a lightweight Topology-Based Spatiotemporal Handover mechanism. We implement a high-throughput parallel pipeline leveraging YOLO11 and ByteTrack to process concurrent 4K streams. Our core contribution is a deterministic queue-based matching algorithm that utilizes geometric overlaps and virtual lane discretization to predictively manage identity handover via FIFO queues. Experimental results on complex urban environments, including intersections and merging traffic, demonstrate a Handover Success Rate (HOSR) of 99.8% in continuous traffic flows, significantly outperforming Re-ID baselines (74.1%) while validating edge deployment feasibility. The source code is available at this https URL .\n\n---\n\n## 111. Lamarckian Inheritance in Dynamic Environments: How Key Variables Affect Evolutionary Dynamics\n- **Authors**: K. Ege de Bruin , Kyrre Glette , Kai Olav Ellefsen\n- **URL**: [https://arxiv.org/abs/2605.15769](https://arxiv.org/abs/2605.15769)\n- **Abstract**:\n> The co-optimization of a robot's body and brain presents a coupled challenge: the morphology constrains which control strategies are effective, while the control determines how well the morphology performs. To address this, we combine morphology optimization as evolution with controller optimization as lifetime learning, utilizing Lamarckian inheritance to transfer learned controller parameters from parent to offspring. In dynamic environments, existing literature presents conflicting evidence: while traditional evolutionary theory often suggests Lamarckian inheritance lacks benefit, recent studies in evolutionary robotics indicate it can improve performance. We hypothesize that this is because previous works have not included all relevant variables with dynamic environments. In this work, we show that the benefit of Lamarckian inheritance depends on two variables: how conflicting the environmental changes are to robot control, and the predictability of those changes for the robotic agent. Using virtual soft robots and two different learning approaches, Bayesian optimization and reinforcement learning, we show that Lamarckian inheritance only underperforms Darwinian inheritance when the changes are both conflicting and unpredictable. We find that adding a sensor to detect environmental changes restores the benefits for Lamarckian inheritance in conflicting environments, by allowing robotic agents to predict the need for a different behavior, thereby generalizing their control.\n\n---\n\n## 112. GRASP: Learning to Ground Social Reasoning in Multi-Person Non-Verbal Interactions\n- **Authors**: Junho Kim , Xu Cao , Houze Yang , Bikram Boote , Ana Jojic , Fiona Ryan , Bolin Lai , Sangmin Lee , James M. Rehg\n- **URL**: [https://arxiv.org/abs/2605.15764](https://arxiv.org/abs/2605.15764)\n- **Abstract**:\n> Understanding social interactions requires reasoning over subtle non-verbal cues, yet current multimodal large language models (MLLMs) often fail to identify who interacts with whom in multi-person videos. We introduce GRASP, a large-scale social reasoning dataset that connects high-level social QA with fine-grained gaze and deictic gesture events. GRASP contains 290K question--answer pairs over 46K videos totaling 749 hours, organized by a 16-category taxonomy spanning gaze, gesture, and joint gaze--gesture reasoning, together with GRASP-Bench for evaluation. Unlike prior resources that focus on either isolated cues or high-level social QA, GRASP builds questions from identity-consistent gaze trajectories, deictic gestures, and their joint compositions into social events. Moreover, we propose Social Grounding Reward (SGR), a learning signal that uses these social events to encourage models to reason about the participants involved in each interaction. Experiments show that SGR improves performance on GRASP-Bench while maintaining zero-shot performance on related social video QA benchmarks.\n\n---\n\n## 113. CompactQE: Interpretable Translation Quality Estimation via Small Open-Weight LLMs\n- **Authors**: Kamil Guttmann , Zofia Fraś , Artur Nowakowski , Krzysztof Jassem\n- **URL**: [https://arxiv.org/abs/2605.15763](https://arxiv.org/abs/2605.15763)\n- **Abstract**:\n> Current state-of-the-art Quality Estimation (QE) in machine translation relies on massive, proprietary LLMs, raising data privacy concerns. We demonstrate that smaller, open-source LLMs (<30B parameters) are a viable, cost-effective and privacy-preserving alternative. Using a single-pass prompting strategy, our models simultaneously generate quality scores, MQM error annotations, suggested error corrections, and full post-editions. Our analysis shows these models achieve highly competitive system-level correlations with human judgments that outperform traditional neural metrics, fine-tuned models, and human inter-annotator agreement, effectively approximating the capabilities of much larger proprietary LLMs.\n\n---\n\n## 114. BiomedAP: A Vision-Informed Dual-Anchor Framework with Gated Cross-Modal Fusion for Robust Medical Vision-Language Adaptation\n- **Authors**: Huanyang Tong , Kai Liu , Fangjun Kuang , Huiling Chen\n- **URL**: [https://arxiv.org/abs/2605.15736](https://arxiv.org/abs/2605.15736)\n- **Abstract**:\n> Biomedical Vision--Language Models (VLMs) have shown remarkable promise in few-shot medical diagnosis but face a critical bottleneck: \\textit{fragility to prompt variations}.Existing adaptation frameworks typically optimize visual and textual prompts as independent streams, relying on ideal ``Golden Prompts''. In clinical reality, where descriptions are often noisy and heterogeneous, this modality isolation leads to unstable cross-modal alignment. To address this, we propose BiomedAP, a vision-informed dual-anchor framework with gated cross-modal this http URL enforces synergistic alignment through two mechanisms: (1) Gated Cross-Modal Fusion, which enables layer-wise interaction between modalities, acting as a dynamic noise regulator to suppress irrelevant textual cues; and (2) a Dual-Anchor Constraint that regularizes learnable prompts toward stable semantic centroids derived from both expert templates (High Anchors) and few-shot visual prototypes (Low Anchors). Extensive experiments across 11 benchmarks demonstrate that BiomedAP consistently surpasses baselines, achieving competitive few-shot accuracy and markedly enhanced robustness under prompt perturbations. Our code is available at: this https URL . Keywords: Vision-Language Models; Prompt Learning; Parameter-Efficient Fine-Tuning; Few-shot Learning\n\n---\n\n## 115. UAM: A Dual-Stream Perspective on Forgetting in VLA Training\n- **Authors**: Jianke Zhang , Yuanfei Luo , Yucheng Hu , Xiaoyu Chen , Yanjiang Guo , Ziyang Liu , Hongbin Xu , Tian Lan , Jianyu Chen\n- **URL**: [https://arxiv.org/abs/2605.15735](https://arxiv.org/abs/2605.15735)\n- **Abstract**:\n> Vision--language--action (VLA) models are typically built by fine-tuning a pretrained vision--language model (VLM) on action data. However, we show that this standard recipe systematically erodes the VLM's multimodal competence, a side effect we call the embodiment tax. But do VLAs have to forget? Inspired by the two-stream organization of biological vision, we trace this degradation to a structural bottleneck: current VLAs ask a single encoder to support both language-grounded semantics and control-relevant visual features, whereas biological vision separates recognition and visuomotor control into distinct pathways. Building on this view, we propose the Unified Action Model (UAM), which adds a parallel Dorsal Expert, an analog of the brain's dorsal pathway. To make the Dorsal Expert an effective second pathway and reduce the control-learning burden on the VLM, we initialize it from a pretrained generative model and train it with a mid-level reasoning objective that predicts visual dynamics. This design allows us to train the whole VLA end-to-end on action data alone: with no parameter freezing, no gradient stopping, and no auxiliary VL co-training, UAM retains over $95\\%$ of the underlying VLM's multimodal capability and at the same time achieves the highest average success rate among baselines on a variety of manipulation tasks that probe out-of-distribution generalization, including unseen objects, novel object--target compositions, and instruction variation. Together, these results suggest that semantic preservation in VLAs can emerge from architectural separation itself, rather than being enforced by frozen weights or auxiliary data replay, and that this preserved semantic capability can naturally transfer from VLMs to semantic generalization in actions.\n\n---\n\n## 116. Structure Abstraction and Generalization in a Hippocampal-Entorhinal Inspired World Model\n- **Authors**: Tianqiu Zhang , Muyang Lyu , Xiao Liu , Si Wu\n- **URL**: [https://arxiv.org/abs/2605.15733](https://arxiv.org/abs/2605.15733)\n- **Abstract**:\n> Humans abstract experiences into structured representations to facilitate pattern inference and knowledge transfer. While the hippocampal-entorhinal (HPC-MEC) circuit is known to represent both spatial and conceptual spaces, the mechanisms for concurrently extracting abstract structures from continuous, high-dimensional dynamics remain poorly understood. We propose a brain-inspired hierarchical model that simultaneously infers latent transitions and constructs a predictive visual world model. Our architecture employs an inverse model for structural extraction alongside an HPC-MEC coupling model that dissociates relational structures (MEC) from integrated episodic scenes (HPC). Using primitive transformation dynamics as a benchmark, we demonstrate the model's capacity for structural abstraction. By leveraging velocity-driven path integration, the framework enables robust prediction and structural reuse across diverse contexts, thereby achieving structural generalization. This work provides a novel computational framework for understanding how brain-inspired, self-supervised learning of world models facilitates the acquisition of reusable abstract knowledge.\n\n---\n\n## 117. DecomPose: Disentangling Cross-Category Optimization Contention for Category-Level 6D Object Pose Estimation\n- **Authors**: Yifan Gao , Lu Zou , Zhangjin Huang , Guoping Wang\n- **URL**: [https://arxiv.org/abs/2605.15728](https://arxiv.org/abs/2605.15728)\n- **Abstract**:\n> Category-level 6D object pose estimation is typically formulated as a multi-category joint learning problem with fully shared model parameters. However, pronounced geometric heterogeneity across categories entangles incompatible optimization signals in shared modules, resulting in gradient conflicts and negative transfer during training. To address this challenge, we first introduce gradient-based diagnostics to quantify module-level cross-category contention. Building on results of diagnostics, we propose DecomPose, a difficulty-aware decomposition framework that mitigates optimization contention via: (1) difficulty-aware gradient decoupling, which groups categories using a data-driven difficulty proxy and routes each instance to a group-specific correspondence branch to isolate incompatible updates; and (2) stability-driven asymmetric branching, which assigns higher-capacity branches to structurally simple categories as stable optimization anchors while constraining complex categories with lightweight branches to suppress noisy updates and alleviate negative transfer. Extensive experiments on REAL275, CAMERA25, and HouseCat6D demonstrate that DecomPose effectively reduces cross-category optimization contention and delivers superior pose estimation performance across multiple benchmarks.\n\n---\n\n## 118. DiLA: Disentangled Latent Action World Models\n- **Authors**: Tianqiu Zhang , Muyang Lyu , Yufan Zhang , Fang Fang , Si Wu\n- **URL**: [https://arxiv.org/abs/2605.15725](https://arxiv.org/abs/2605.15725)\n- **Abstract**:\n> Latent Action Models (LAMs) enable the learning of world models from unlabeled video by inferring abstract actions between consecutive frames. However, LAMs face a fundamental trade-off between action abstraction and generation fidelity. Existing methods typically circumvent this issue by using two-stage training with pre-trained world models or by limiting predictions to optical flow. In this paper, we introduce DiLA, a novel Disentangled Latent Action world model that aims to resolve this trade-off via content-structure disentanglement. Our key insight is that disentanglement and latent action learning are co-evolving: the predictive bottleneck inherent in latent action learning serves as a driving force for disentanglement, compelling the model to distill spatial layouts into the structure pathway while offloading visual details to a separate content pathway for generation. This synergy yields a continuous, semantically structured latent action space without compromising generative quality. DiLA achieves superior results in video generation quality, action transfer, visual planning, and manifold interpretability. These findings establish DiLA as a unified framework that simultaneously achieves high-level action abstraction and high-fidelity generation, advancing the frontier of self-supervised world model learning.\n\n---\n\n## 119. Bidirectional Fusion Guided by Cardiac Patterns for Semi-Supervised ECG Segmentation\n- **Authors**: Jeonghwa Lim , Minje Park , Sunghoon Joo\n- **URL**: [https://arxiv.org/abs/2605.15722](https://arxiv.org/abs/2605.15722)\n- **Abstract**:\n> Accurate delineation of electrocardiogram (ECG), the segmentation of meaningful waveform features, is crucial for cardiovascular diagnostics. However, the scarcity of annotated data poses a significant challenge for training deep learning models. Conventional semi-supervised semantic segmentation (SemiSeg) methods primarily focus on consistency from unlabeled data, underutilizing the information exchange possible between labeled and unlabeled sets. To address this, we introduce CardioMix, a framework built on a bidirectional CutMix strategy guided by cardiac patterns for ECG segmentation. This approach enriches the labeled set with realistic variations from unlabeled data while simultaneously applying stronger supervisory signals to the unlabeled set, as the cardiac pattern-guided mixing ensures all augmented samples remain physiologically meaningful. Our framework is designed as a plug-and-play module, demonstrating high compatibility with various SemiSeg algorithms. Extensive experiments on SemiSegECG, a public multi-dataset benchmark for ECG delineation, demonstrate that CardioMix consistently outperforms existing CutMix-based fusion strategies across diverse datasets and labeled ratios as a plug-and-play module compatible with various SemiSeg algorithms.\n\n---\n\n## 120. Position: Early-Stage Quality Assurance in Annotation Pipelines Is More Cost-Effective Than Late-Stage Validation\n- **Authors**: Sunil Kothari , Sumukha Sharma Thoppanahalli Chandramouli , Naman Khandelwal , Parth Kulshreshtha , Ashi Jain , Kriti Banka , Tanuja Chintada , Venkata Triveni , Gulipalli Praveen Kumar , Manish Mehta , Tao Liu\n- **URL**: [https://arxiv.org/abs/2605.15714](https://arxiv.org/abs/2605.15714)\n- **Abstract**:\n> This position paper argues that the machine learning community should prioritize early-stage quality assurance in annotation pipelines over the prevailing practice of late-stage validation. Data quality bottlenecks increasingly limit foundation model improvement, yet quality assurance research focuses almost exclusively on validation methods rather than validation timing. When validation occurs, not merely what methods are employed, fundamentally determines both error rates and annotation costs. This temporal neglect is puzzling given the well-established \"shift-left\" principle from software engineering, where empirical studies demonstrate 4--100x cost multipliers for defects detected in later stages (Boehm, 1981; Shull et al., 2002). Annotation pipelines exhibit analogous dynamics: errors caught before annotation begins cost a fraction of those discovered after review cycles complete. We propose a taxonomy of three QA trigger points, namely pre-annotation (T0), post-annotation (T1), and post-review (T2), that decompose annotation workflows into discrete validation opportunities. A parametric error-propagation model formalizes when timing affects final error rates versus only economics, making timing a measurable design variable rather than a configuration afterthought. A survey of 47 recent papers reveals that only 4% report when validation occurs, a striking gap given timing's demonstrated impact in adjacent fields. Without explicit attention to QA timing, the community risks optimizing validation methods while ignoring the structural variable that may matter most. Acting on this position requires three steps: researchers should report QA timing configurations alongside validation methods; annotation platforms should expose timing as a first-class parameter; and the community should run controlled experiments that measure stage-specific detection rates directly.\n\n---\n\n## 121. Learning Dynamic Pick-and-Place for a Legged Manipulator\n- **Authors**: Moonkyu Jung , Jiseong Lee , Zhengmao He , Donghoon Youm , Juhyeok Mun , HyeongJun Kim , Hyunsik Oh , Donghyuk Choi , Jungwoo Hur , Jie Song , Jemin Hwangbo\n- **URL**: [https://arxiv.org/abs/2605.15713](https://arxiv.org/abs/2605.15713)\n- **Abstract**:\n> Legged manipulators extend robotic capabilities beyond static manipulation by integrating agile locomotion with versatile arm control. However, achieving precise manipulation while maintaining coordinated locomotion remains a major challenge. This work presents a hierarchical reinforcement learning framework for dynamic pick-and-place tasks using a quadruped equipped with a 6-DOF robotic arm. The framework incorporates an explicit mass estimation module enabling adaptive whole-body control for objects with varying weights. In simulation, the system achieves an 86.05% success rate with payloads up to 2.3 kg. The approach is further validated through real-world experiments across six representative scenarios with controlled variations in object physical properties (size and mass) and task heights. Specifically, within a wide vertical workspace ranging from ground level to 1.1~m-high tabletops, the system demonstrates an average success rate of 73.3% for payloads up to 1.3 kg, with an average execution time of 4.06 s. Unlike prior works that handle lightweight objects and execute pick-and-place motions with slow, piecewise motions, the proposed framework exploits concurrent locomotion and manipulation for dynamic, continuous execution. These results demonstrate the potential of quadrupedal mobile manipulators for adaptive, whole-body pick-and-place with heavier payloads and extended workspaces.\n\n---\n\n## 122. Feedback World Model Enables Precise Guidance of Diffusion Policy\n- **Authors**: Tuo An , Jindou Jia , Gen Li , Jingliang Li , Chuhao Zhou , Pengfei Liu , Bofan Lyu , Jiaqi Bai , Xinying Guo , Geng Li , Jianfei Yang\n- **URL**: [https://arxiv.org/abs/2605.15705](https://arxiv.org/abs/2605.15705)\n- **Abstract**:\n> World models aim to improve robotic decision making by predicting the consequences of actions. However, in practice, their predictions often become unreliable once the robot encounters states outside the training distribution, limiting their effectiveness at deployment. We observe that execution itself provides a natural but underutilized signal: after each action, the robot directly observes the true next state, revealing the mismatch between predicted and actual outcomes. Building on this insight, we propose feedback world model, a new paradigm that closes the loop between prediction and observation at inference time. Instead of treating the world model as a static open-loop predictor, our method maintains a lightweight feedback state that is updated online to iteratively correct future predictions, compensating for model errors using real-time observations without additional training data or parameter updates. We show that this process can be interpreted as a latent-space observer and admits convergence guarantees under mild conditions. We further introduce action-aware guidance to better translate corrected predictions into control by emphasizing action-controllable components while suppressing irrelevant variations. Experiments on LIBERO-Plus, Robomimic, and real-world manipulation tasks demonstrate that our method substantially improves both prediction accuracy and policy performance under distribution shift. In particular, it reduces world model prediction error by up to 76.4% and improves out-of-distribution (OOD) success rate by 30%. These results show that incorporating real-time feedback at inference time provides a simple yet powerful alternative to static world modeling.\n\n---\n\n## 123. H-Mem: A Novel Memory Mechanism for Evolving and Retrieving Agent Memory via a Hybrid Structure\n- **Authors**: Jiawei Yu , Yixiang Fang , Xilin Liu , Yuchi Ma\n- **URL**: [https://arxiv.org/abs/2605.15701](https://arxiv.org/abs/2605.15701)\n- **Abstract**:\n> Memory data are ubiquitous in Large Language Model (LLM)-based agents (e.g., OpenClaw and Manus). A few recent works have attempted to exploit agents'memory for improving their performance on the question-answering (QA) task, but they lack a principled mechanism for effectively modeling how memory data evolves over time and retrieving memory data effectively, leading to poor performance in memory utilization. To fill this gap, we present H-Mem, a novel memory mechanism via a hybrid structure that can not only effectively model the evolution of agent memory over a long period of time, but also provide an efficient memory retrieval approach. Particularly, H-Mem builds a temporal and semantic tree structure that allows the short-term memory data to evolve progressively into long-term memory data, where the latter provides summarized information about the former, while simultaneously constructing a knowledge graph to capture the relationships between entities in memory. Moreover, it offers an effective memory retrieval approach by exploiting the hybrid structure of the tree and graph structures. Extensive experiments on three agent memory benchmarks show that H-Mem achieves state-of-the-art performance on the QA task.\n\n---\n\n## 124. $α$-TCAV: A Unified Framework for Testing with Concept Activation Vectors\n- **Authors**: Ekkehard Schnoor , Jawher Said , Malik Tiomoko , Wojciech Samek , Alexander Jung\n- **URL**: [https://arxiv.org/abs/2605.15688](https://arxiv.org/abs/2605.15688)\n- **Abstract**:\n> Concept Activation Vectors (CAVs) are a fundamental tool for concept-based explainability in deep learning, yet their practical utility is limited by statistical instability. We analyze the stochastic nature of CAVs and the Testing with CAVs (TCAV) method, deriving the distributions of major CAV classes including PatternCAV, FastCAV, and ridge regression-based CAVs. We then identify a fundamental flaw in the standard TCAV score: its reliance on a discontinuous indicator function induces non-decaying variance in critical regimes. To address this, we introduce $\\alpha$-TCAV, a generalized framework that replaces the indicator with a parameterized smooth function, yielding a unified probabilistic formulation that subsumes both TCAV and Multi-TCAV. We characterize the induced distributions of sensitivity scores and different TCAV variants, showing that established state-of-the-art choices lack theoretical justification. We provide principled guidance on tuning the parameter in $\\alpha$-TCAV -- either to imitate Multi-TCAV at substantially lower computational cost, or to obtain a calibrated Bayes-optimal probabilistic measure of a concept's influence. Finally, our analysis yields practical recommendations that challenge established routines: most notably, allocating the full sampling budget to a single CAV rather than splitting it across several.\n\n---\n\n## 125. ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models\n- **Authors**: Jiahui Guang , Yingjie Zhu , Cuiyun Gao , Haiyan Wang , Jing Li , Di Shao , Zhaoquan Gu\n- **URL**: [https://arxiv.org/abs/2605.15687](https://arxiv.org/abs/2605.15687)\n- **Abstract**:\n> Multimodal large language models (MLLMs) may memorize sensitive cross-modal information during pretraining, making machine unlearning (MU) crucial. Existing methods typically evaluate unlearning effectiveness based on output deviations, while overlooking the generation quality after unlearning. This can easily lead to hallucinated or rigid responses, thereby affecting the usability and safety of the unlearned model. To address this issue, we propose ASRU, a controllable multimodal unlearning framework that incorporates generation quality as a core evaluation objective. ASRU first induces initial refusal behavior through activation redirection, and then optimizes fine-grained refusal boundaries using a customized reward function, thereby achieving a better trade-off between target knowledge unlearning and model utility. Experiments on Qwen3-VL show that ASRU significantly improves unlearning effectiveness (+24.6%) on average and generation quality (5.8x) on average while effectively preserving model utility, using only a small amount of retained supervision data.\n\n---\n\n## 126. Interaction-Aware Influence Functions for Group Attribution\n- **Authors**: Jaeseung Heo , Kyeongheung Yun , Youngbin Choi , Sehyun Hwang , Jungseul Ok , Dongwoo Kim\n- **URL**: [https://arxiv.org/abs/2605.15675](https://arxiv.org/abs/2605.15675)\n- **Abstract**:\n> Influence functions approximate how removing a training example changes a quantity of interest, called the target function, such as a held-out loss. To estimate the influence of a group of examples, the standard practice is to sum the individual influences of its members. However, this sum does not capture how examples jointly affect the target: a pair of examples may be redundant or complementary, but the sum cannot distinguish these cases. We propose an interaction-aware influence function that characterizes how interactions between examples influence the target. By expanding the target to second order around the trained parameters, we obtain an estimator that augments the standard sum with a pairwise interaction term that captures the alignment between two examples' effects on the target. We empirically evaluate our estimator in two settings. First, on six dataset-model pairs spanning logistic regression, MLPs, and ResNet-9, our estimator tracks leave-group-out retraining substantially better than first-order influence across all settings. Second, when used as a greedy selection rule for instruction-tuning data on Llama-3.1-8B, it beats prior influence-based and representation-similarity baselines on five of seven downstream tasks, in a regime where standard influence-based selection underperforms random selection.\n\n---\n\n## 127. VLMs Trace Without Tracking: Diagnosing Failures in Visual Path Following\n- **Authors**: Hyesoo Hong , Minsoo Kim , Wonje Jeung , Sangyeon Yoon , Dongjae Jeon , Albert No\n- **URL**: [https://arxiv.org/abs/2605.15672](https://arxiv.org/abs/2605.15672)\n- **Abstract**:\n> Vision-language models (VLMs) achieve strong performance on multimodal benchmarks, but may still lack robust control over basic visual operations. We study \\textit{line tracing}, where a model must follow a selected visual path through successive local continuations. To isolate this ability, we design controlled tracing tasks that introduce nearby competitors while reducing semantic and topological ambiguity such as crossings and overlaps. Across these tasks, even state-of-the-art VLMs frequently lose the target path and switch to nearby alternatives, especially when those alternatives look locally similar to the target. Behavioral interventions and internal analyses indicate that these failures arise from local competition: nearby similar distractors pull the model away from the true continuation. Standard remedies do not remove this bottleneck: model-size scaling provides only limited gains, reasoning partially compensates through costly substitute strategies, and explicit tracing instructions fail to recover stable path following. Finally, tests on tangled-cable scenes and metro maps with richer visual complexity show that the same path-switching failure persists beyond our controlled settings.\n\n---\n\n## 128. VAGS: Velocity Adaptive Guidance Scale for Image Editing and Generation\n- **Authors**: Yan Luo , Ahmadou Aidara , Jingyi Lu , Jeremy Moebel , Kai Han , Mengyu Wang\n- **URL**: [https://arxiv.org/abs/2605.15661](https://arxiv.org/abs/2605.15661)\n- **Abstract**:\n> Classifier-free guidance (CFG) is the primary control over how strongly text semantics move a flow-based sampler, yet standard practice holds its scale fixed across the entire ODE trajectory. This is a fundamental mismatch: early steps are noise-dominated and carry weak semantic signal, while late steps commit image structure and demand stronger directional commitment; more critically, the value of any guidance strength depends on whether the guided velocity is consistent with the model's current dynamics or working against them. We propose \\textit{Velocity-Adaptive Guidance Scale} (VAGS), a training-free replacement that multiplies the nominal scale by a bounded factor combining a temporal signal-level term with the cosine similarity between task-relevant velocity fields. For inversion-free editing, VAGS measures the alignment between source- and target-guided velocities, so edit strength at each step reflects local compatibility between preservation and transformation. For generation, VAGS-Gen uses the alignment between unconditional and conditional velocities as the analogous signal. Neither variant requires fine-tuning, auxiliary networks, or extra forward passes, and fixed CFG is recovered as a special case. On PIE-Bench and DIV2K for editing, and COCO17, CUB-200, and Flickr30K for generation, VAGS consistently improves structural fidelity and generation quality over fixed CFG and recent training-free guidance variants. The code is publicly available at this https URL .\n\n---\n\n## 129. TFZ-Tree: An Ultra-Lightweight Waveform Classification Framework for Resource-Constrained Devices\n- **Authors**: Hao Wang , Kuang Zhang , Yonggang Chi , Tianqi Zhao , Yanbo Fu , Jiaxing Guo\n- **URL**: [https://arxiv.org/abs/2605.15656](https://arxiv.org/abs/2605.15656)\n- **Abstract**:\n> Under the trend of multi-waveform coexistence in 6G IoT, intelligent receivers must first identify physical-layer waveform types before performing correct demodulation and resource scheduling. However, existing signal identification research largely focuses on symbol-level modulation classification. Research directly targeting physical-layer waveform types (e.g., OFDM, OTFS, LoRa) is not only extremely scarce but also heavily reliant on deep neural networks and complex time-frequency transforms, making deployment on resource-constrained terminals difficult. Symbol modulation classification methods themselves cannot circumvent the prerequisite of ``waveform identification first.'' To address this dual gap, we propose an ultra-lightweight waveform classification framework based on time-frequency multidimensional features with a cooperative Z-test tree (ZTree). The framework employs low-complexity time-domain feature extraction, and the classification backend adopts a ZTree optimized by Z-statistical testing, which uses hypothesis testing confidence to automatically control decision tree splitting and size, ensuring efficient execution on resource-limited processors. Tested on ten 6G candidate waveforms including OFDM, OTFS, DSSS, LoRa, and NB-IoT, the method achieves 99.5\\% average accuracy under AWGN and 87.4\\% under TDL-C multipath channels, with main confusion between OTFS and LoRa. Implemented in C on an x86 platform, single inference latency is under 4~ms. To the best of our knowledge, this is the first work achieving real-time recognition of ten IoT waveform types. Future work will target deployment acceleration on embedded MCUs. Code and dataset are open-sourced at: this https URL .\n\n---\n\n## 130. Bridging Silicon and the Hippocampus: Algebro-Deterministic Memory \"VaCoAl\" as a Substrate for Vector-HaSH and TEM\n- **Authors**: Hiroyuki Chuma , Kanji Otsuka , Yoichi Sato\n- **URL**: [https://arxiv.org/abs/2605.15652](https://arxiv.org/abs/2605.15652)\n- **Abstract**:\n> Vector-HaSH and the Tolman-Eichenbaum Machine (TEM) propose that the hippocampal-entorhinal circuit factorizes content from a prestructured grid-cell scaffold and supports compositional memory via ripple-mediated replay. Human iEEG shows that hippocampal sharp-wave ripples (SWRs) gate episodic recall, ripple-locked cortical reactivation recapitulates encoding-time patterns, and multi-hop replay fidelity decays multiplicatively along sequence length. These literatures have advanced in parallel without a shared algebraic object. We show that VaCoAl, an algebro-deterministic hyperdimensional memory architecture built from Galois-field LFSRs, supplies that object. Specifically, deterministic Galois-field diffusion provides a substrate-level alternative to Vector-HaSH's random scaffold-to-hippocampus projection that satisfies the same quasi-orthogonality requirement, with matched second-moment statistics, stronger avalanche behavior, and bit-exact reproducibility. The path-integral Confidence Ratio $CR_2$, the product of per-step $CR_1$ values along an $n$-hop chain, is the natural functional form for multi-hop replay-fidelity decay under conditional independence of per-step reactivation, providing the first algebraically tractable model of reported multiplicative decay. STDP-like path selection in VaCoAl follows from architectural demands -- similarity preservation, compositional reversibility, and bounded-frontier search -- that also constrain hippocampal computation. We further argue that VaCoAl operating regimes share architectural commitments with the EC--CA3 and EC--DG--CA3 pathways, motivating an energy-capacity-plasticity reading of why both are conserved across $>$520 Myr of evolution and primate dentate-gyrus elaboration. We prove formal correspondences, derive testable iEEG predictions, and bridge computational neuroscience and hyperdimensional engineering.\n\n---\n\n## 131. Sharp Spectral Thresholds for Logit Fixed Points\n- **Authors**: Tongxi Wang\n- **URL**: [https://arxiv.org/abs/2605.15651](https://arxiv.org/abs/2605.15651)\n- **Abstract**:\n> Softmax feedback systems are a common mathematical core of entropy-regularized reinforcement learning, logit game dynamics, population choice, and mean-field variational updates. Their central stability question is simple: when does a self-reinforcing softmax system produce a unique and globally predictable outcome? Classical theory gives a conservative answer. By treating softmax as a unit-scale response, it certifies stability only in a strongly randomized regime. We prove that the classical approach misses an entire stable regime and does not identify the point at which the qualitative change truly occurs. For finite-dimensional affine logit systems, the sharp dimension-free Euclidean threshold is $$\\beta\\|\\Pi W\\Pi\\|_{\\mathcal T\\to\\mathcal T}<2,$$ rather than the previously used condition, which certifies stability only while the softmax system remains safely over-regularized. Our theorem fills the previously missing pre-bifurcation regime, extending stability guarantees for affine softmax feedback systems to reward-responsive yet globally predictable systems. It enlarges the certified stability boundary for these systems and identifies where the model genuinely undergoes a phase transition.\n\n---\n\n## 132. Latent Video Prediction Learns Better World Models\n- **Authors**: Ali J Alrasheed , Aryan Yazdan Parast , Basim Azam , James Bailey , Naveed Akhtar\n- **URL**: [https://arxiv.org/abs/2605.15618](https://arxiv.org/abs/2605.15618)\n- **Abstract**:\n> Self-supervised video models are increasingly framed as world models, yet their evaluation remains largely confined to a single top-1 accuracy score on clean benchmarks. This leaves a major gap in comprehending their potential as world models. We present the first systematic study addressing this gap, analyzing four matched-capacity frontier video foundation models, V-JEPA 2.1, V-JEPA 2, VideoPrism, and VideoMAEv2, across five robustness axes relevant to their deployment as video world models: feature discriminability, corruption robustness, fine-grained discrimination, occlusion robustness, and sensitivity to temporal direction. Our evaluations establish that across all five axes, latent-prediction models form a distinct and consistent profile. They degrade more gracefully under pixel corruption, preserve usable class structure rather than mere geometric stability under occlusion, capture fine-grained physical contact cues without reconstructing pixels, and uniquely encode the arrow of time. These advantages can even survive task adaptation: a frozen V-JEPA 2 backbone with a lightweight attentive probe outperforms a fully fine-tuned VideoMAE and a supervised TimeSformer on corruption and occlusion robustness. Our extensive results offer concrete new evidence in favor of latent prediction for robust world modeling.\n\n---\n\n## 133. A Few GPUs, A Whole Lotta Scale: Faithful LLM Training Emulation with PrismLLM\n- **Authors**: Shaoke Xi , ChonLam Lao , Boyi Jia , Jiaqi Gao , Zhipeng Zhang , Jiamin Cao , Brian Sutioso , Erci Xu , Minlan Yu , Kui Ren , Yong Li , Zhengping Qian , Ennan Zhai , Jingren Zhou\n- **URL**: [https://arxiv.org/abs/2605.15617](https://arxiv.org/abs/2605.15617)\n- **Abstract**:\n> Large language model (LLM) training today runs on clusters spanning thousands of GPUs. While this scale enables rapid model advances, developing, debugging, and performance-tuning the training framework inevitably becomes complex and costly. This is because engineers often need to reproduce production behaviors to diagnose failures or evaluate optimizations, thereby demanding frequent and even exclusive access to production-scale clusters -- which becomes increasingly hard given that the majority of GPUs are already committed to production workloads. Simulation relies on complex performance models that are difficult to maintain, and downscaled experiments often fail to capture scale-dependent behaviors. We present PrismLLM to decouple large-scale execution from the need to access large clusters, enabling engineers to run and observe ranks of interest under faithful large-scale behavior using only a few GPUs. PrismLLM constructs a high-fidelity execution graph via a slicing-based approach that captures computation, communication, and dependencies of the target scale. Then, PrismLLM performs hybrid emulation where selected ranks execute the original program while the remaining ranks are replayed as virtual participants. Experiments on large-scale LLM training workloads show that PrismLLM accurately reproduces performance and memory behavior, achieving only 0.58\\% average error in iteration time and less than 0.01\\% error in peak GPU memory usage. PrismLLM can emulate clusters of up to 8192 GPUs using fewer than 1\\% of the physical GPUs required by the original deployment.\n\n---\n\n## 134. Offline Reinforcement Learning with Universal Horizon Models\n- **Authors**: Hojun Chung , Junseo Lee , Songhwai Oh\n- **URL**: [https://arxiv.org/abs/2605.15603](https://arxiv.org/abs/2605.15603)\n- **Abstract**:\n> Model-based reinforcement learning (RL) offers a compelling approach to offline RL by enabling value learning on imagined on-policy trajectories. However, it often suffers from compounding errors due to repeated model inference on self-generated states. While geometric horizon models (GHM) alleviate this issue through direct prediction over a discounted infinite-horizon future, they remain challenged in accurately modeling distant future states. To this end, we introduce universal horizon models (UHM), a generalization of GHM that directly predicts future states under arbitrary horizons. Leveraging this flexibility, we propose a scalable value learning method that employs a winsorized horizon distribution to stabilize training by capping excessively large horizons. Experimental results on 100 challenging OGBench tasks demonstrate that the proposed method outperforms competitive baselines, particularly on tasks with highly suboptimal datasets and those requiring long-horizon reasoning. Project page: this https URL\n\n---\n\n## 135. Pretraining Objective Matters in Extreme Low-Data FGVC: A Backbone-Controlled Study\n- **Authors**: Alexander Hackett , Srikanth Thudumu , Ginny Fisher , Mahule Roy , Aisha Sartaj , Jason Fisher\n- **URL**: [https://arxiv.org/abs/2605.15599](https://arxiv.org/abs/2605.15599)\n- **Abstract**:\n> Extreme low-data fine-grained classification is common in expert domains where labeling is expensive, yet practitioners still need principled guidance for selecting pretrained encoders. We study emerald inclusion grading with a custom dataset of labeled images across three classes and ask: under matched backbone capacity, how does pretraining objective affect downstream representation quality? We compare four frozen ViT-B/16 encoders trained with supervised classification, contrastive learning (SigLIP2), masked reconstruction (MAE), and self-distillation (DINOv3), and evaluate them with leave-one-out cross-validation using linear and nonlinear probes. To control statistical noise in the low-N regime, we use permutation testing (N=1000) on macro one-vs-rest AUC. Supervised and contrastive encoders provide the strongest linear separability (logistic AUC: 0.768 and 0.735; SVM AUC: 0.739 and 0.697), while MAE improves under nonlinear probes (XGBoost AUC: 0.713). We find that DINOv3 underperforms across probe families in this domain. These results support a practical recommendation for extreme low-data FGVC: prioritize margin-enforcing pretraining objectives when data scarcity restricts probing to linear decision rules, and consider reconstruction-style encoders when nonlinear classifiers are feasible given dataset constraints.\n\n---\n\n## 136. Embracing Biased Transition Matrices for Complementary-Label Learning with Many Classes\n- **Authors**: Tan-Ha Mai , Chao-Kai Chiang , Han-Hwa Shih , Gang Niu , Masashi Sugiyama , Hsuan-Tien Lin\n- **URL**: [https://arxiv.org/abs/2605.15586](https://arxiv.org/abs/2605.15586)\n- **Abstract**:\n> Complementary-label learning (CLL) is a weakly supervised paradigm where instances are labeled with classes they do not belong to. Despite a decade of research, CLL methods remain competitive mainly on 10-class classification, with scaling to large label spaces continuing to be an enduring bottleneck. This limitation stems from the common assumption of uniform label generation in traditional methods, which fatally dilutes the learning signal in many-class settings. In this paper, we demonstrate that this long-standing barrier can be overcome by deliberately designing a biased (non-uniform) generation process that restricts complementary labels to a subset of classes. This finding motivates us to propose Bias-Induced Constrained Labeling (BICL), a principled framework spanning data collection to training that leverages this bias. BICL enables effective learning on CIFAR-100 and TinyImageNet-200, achieving more than sevenfold accuracy improvements over traditional methods. Our findings establish a new trajectory for making CLL feasible for many classes in real-world applications.\n\n---\n\n## 137. Detecting Privilege Escalation in Polyglot Microservices via Agentic Program Analysis\n- **Authors**: Penghui Li , Hong Yau Chong , Yinzhi Cao , Junfeng Yang\n- **URL**: [https://arxiv.org/abs/2605.15569](https://arxiv.org/abs/2605.15569)\n- **Abstract**:\n> Microservices are widely adopted in modern cloud systems due to their scalability and fault tolerance. However, microservice architectures introduce significant complexity in privilege and permission control, creating risks of privilege escalation where attackers can gain unauthorized access to resources or operations. Detecting such vulnerabilities is challenging due to complex cross-service interactions, polyglot codebases, and diverse privileged operations and permission checks. We present Neo, an agentic program analysis framework that combines large language models (LLMs) with classic program analysis to address these challenges. Neo leverages an LLM-based agent that dynamically generates analysis plans, adapts code search strategies, and validates semantics. We develop code search primitives that enable Neo to perform scalable and flexible code exploration across services and languages. We evaluated Neo on 25 open-source microservice applications spanning 7 programming languages and 6.2 million lines of code. Neo uncovered 24 zero-day privilege escalation vulnerabilities and achieved 81.0% precision and 85.0% recall on a ground-truth dataset. Compared to existing program analysis and agentic solutions, Neo demonstrated significant improvements in both detection accuracy and scalability. We further showcased Neo's extensibility by applying it to other application domains and vulnerability types, uncovering 18 additional zero-day vulnerabilities.\n\n---\n\n## 138. AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs\n- **Authors**: Haizhong Zheng , Yizhuo Di , Jiahui Wang , Shuowei Jin , Xueshen Liu , Yongji Wu , Z. Morley Mao , Ion Stoica , Jiawei Zhao , Beidi Chen\n- **URL**: [https://arxiv.org/abs/2605.15565](https://arxiv.org/abs/2605.15565)\n- **Abstract**:\n> Reinforcement learning (RL) is increasingly used to improve the reasoning, coding, and tool-use capabilities of large language models, but agentic RL remains prohibitively expensive. Scaling RL to agentic LLMs requires supporting complex workloads, including multi-policy collaborative training, while efficiently using elastic, heterogeneous, and cross-region compute resources. Existing LLM RL systems support some of these capabilities, but each new extension often requires dedicated system engineering. This burden arises from trainer-centered control architectures and the lack of principled abstractions for RL system components. To address these limitations, we propose AstraFlow, a dataflow-oriented RL system that replaces conventional trainer-centered control with principled component abstractions. In AstraFlow, rollout services, dataflow management, and training are decoupled into autonomous components, enabling the system to natively support complex multi-policy agentic RL workloads and efficiently exploit diverse compute resources. We evaluate AstraFlow across math, code, search, and AgentBench workloads, showing that the same system supports multi-policy training, elastic scaling, heterogeneous cross-region execution, and composable data algorithms without system-level code changes. In multi-policy collaborative training, AstraFlow achieves comparable or better accuracy than existing RL systems while speeding up training time by 2.7x.\n\n---\n\n## 139. Operator-Controlled 6G: From Connectivity Infrastructure to Guaranteed Digital Services\n- **Authors**: David Soldani\n- **URL**: [https://arxiv.org/abs/2605.15553](https://arxiv.org/abs/2605.15553)\n- **Abstract**:\n> Sixth-generation mobile networks (6G) are approaching a structural inflection point. Five generations of vendor-led architectures have left operators procuring and operating networks they do not own, on platforms they cannot modify, with AI layers they cannot audit. This paper argues that 6G must reverse this trajectory by reordering operator priorities: Control First, Customer First, Business First, Operations First, and Technology Last. Technology should serve operator control, customer outcomes, monetizable guarantees, and software-driven operations, not dictate this http URL contributions operationalize this thesis. The 6G Control Compact defines a three-layer ownership taxonomy--own, federate, and consume--that allocates architectural sovereignty according to strategic value. The Guarantee Economy defines a five-tier, outcome-priced commercial model that converts operator control into enforceable service-level objectives. The framework is grounded in operational evidence from Rakuten Mobile, the world's first national-scale, fully cloud-native, fully Open RAN deployment, which reached full-year EBITDA profitability in FY2025. It is aligned with the ITU-R IMT-2030 framework, 3GPP 6G use cases and service requirements, NGMN recommendations, ETSI standards, O-RAN Alliance and AI-RAN Alliance specifications, IOWN Global Forum sustainability metrics, Linux Foundation initiatives, and leading industry and academic programs. A three-phase roadmap covering 2025-2027, 2027-2029, and 2029-2032 and beyond, together with seven stakeholder-specific calls to action, translates the architecture into industry commitments. The central claim is that Rakuten Mobile's deployment demonstrates the feasibility of operator-controlled 6G. Decisions made during 2026-2028 will determine whether 6G becomes a platform for guaranteed digital services or another vendor-dependent infrastructure cycle.\n\n---\n\n## 140. CTF4Nuclear: Common Task Framework for Nuclear Fission and Fusion Models\n- **Authors**: Stefano Riva , Carolina Introini , Antonio Cammi , Dean Price , Alexey Yermakov , Yue Zhao , Philippe M. Wyder , Judah Goldfeder , Jan Williams , Amy Sara Rude , Matteo Tomasetto , Joe Germany , Joseph Bakarji , Georg Maierhofer , Miles Cranmer , J. Nathan Kutz\n- **URL**: [https://arxiv.org/abs/2605.15549](https://arxiv.org/abs/2605.15549)\n- **Abstract**:\n> The demand for clean energy is ever increasing, with new nuclear technologies presenting a complementary solution to renewable energies. However, designing and operating these systems is exceptionally difficult, given the complexity of the physical phenomena that interact to form the system dynamics. While high-fidelity simulations help to understand the non-linear, multi-physics interactions within a reactor, they are computationally expensive and rarely suitable for real-time applications. Furthermore, model-based approaches are inherently sensitive to simplifying assumptions required to derive their governing equations and parameters, leading to inevitable discrepancies with real-world measurements. In contrast, Machine Learning (ML) methods have the potential to generate reliable surrogate models which may be able to quickly predict the system's behaviour. However, the number of data-driven methods that can potentially be used for this task is large and diverse. In a safety-critical setting such as nuclear engineering, a fair comparison of different ML methods, and a clear understanding of their advantages and limitations, is of paramount importance. To address this, we introduce a Common Task Framework (CTF) for ML in nuclear engineering, building upon previous efforts in dynamical systems and seismology. This CTF considers a curated set of datasets from different nuclear and nuclear-adjacent systems. The CTF evaluates the performance of a method on 12 established metrics, alongside a new paradigm focused on system monitoring from sparse measurements only. We illustrate the framework by benchmarking standard ML baselines against these datasets, revealing current method limitations. Our vision is to replace ad hoc comparisons with standardized evaluations on hidden test sets, raising the bar for rigour and reproducibility in scientific ML for the nuclear industry.\n\n---\n\n## 141. Domain-Independent Game Abstraction using Word Embedding Techniques\n- **Authors**: Juho Kim , Tuomas Sandholm\n- **URL**: [https://arxiv.org/abs/2605.15543](https://arxiv.org/abs/2605.15543)\n- **Abstract**:\n> Many games of interest in the real world are often intractably large, thereby necessitating the use of game abstraction to shrink them in size, typically by many magnitudes. Over the last two decades, there have been significant advances in game abstraction; however, the domain-specific nature (usually poker) of much of the prior work prevents those techniques from being easily generalized to other settings without extensively analyzing the game at hand. In this paper, we propose a domain-independent approach to game abstraction, which applies word embedding techniques from the field of natural language processing. Treating each action as a word and gameplay data as a corpus, word vectors can be trained to represent each action as a real-valued vector, which can then be clustered to facilitate game abstraction. We also explore the use of foundational embedding models and show that action embeddings obtained this way can capture a surprising amount of information about the underlying game. Experimental results demonstrate that our proposed game abstraction technique is effective, although it does not outperform specialized algorithms tailored to specific games.\n\n---\n\n## 142. SkiP: When to Skip and When to Refine for Efficient Robot Manipulation\n- **Authors**: Mingtong Dai , Guanqi Peng , Yongjie Bai , Feng Yan , Chunjie Chen , Lingbo Liu , Liang Lin , Xinyu Wu\n- **URL**: [https://arxiv.org/abs/2605.15536](https://arxiv.org/abs/2605.15536)\n- **Abstract**:\n> Previous imitation learning policies predict future actions at every control step, whether in smooth motion phases or precise, contact-rich operation phases. This uniform treatment is wasteful: most steps in a manipulation trajectory traverse free space and carry little task-relevant information, while a small fraction of \\emph{key} steps around contacts, grasps, and alignment demand dense, high-resolution prediction. We propose a novel \\emph{action relabeling} mechanism: at each timestep in a skip segment, we replace the behavior cloning target with the action at the entrance of the next key segment, enabling the policy to leap over redundant steps in a single decision. The resulting \\textbf{Skip Policy (SkiP)} dynamically leaps over skip segments and intensively refines actions in key segments, within a single unified network requiring no learned skip planner or hierarchical structure. To automatically partition demonstrations into key and skip segments without manual annotation, we introduce \\emph{Motion Spectrum Keying} (MSK), a fast, task-agnostic procedure that detects local motion complexity from action signals. Extensive experiments across 72 simulated manipulation tasks and three real-robot tasks show that SkiP reduces executed steps by $15$--$40\\%$ while matching or improving success rates across various policy backbones. Project page: \\texttt{ this https URL }.\n\n---\n\n## 143. Tuning-free Instruction-based Video Editing Via Structural Noise Initialization and Guidance\n- **Authors**: Song Wu , Xinyu Chen , Qian Wang , Liang Li , Zili Yi , Junlan Feng\n- **URL**: [https://arxiv.org/abs/2605.15533](https://arxiv.org/abs/2605.15533)\n- **Abstract**:\n> Video editing poses a significant challenge. While a series of tuning-free methods circumvent the need for extensive data collection and model training, they often underutilize the rich information embedded within noisy latent, leading to unsatisfactory results. To address this, we propose a \\textit{tuning-free, instruction-based} video editing framework. We approach video editing from the perspective of noisy latent: we design a Structural Noise Initialization Strategy (SNIS) to secure a superior editing starting point by assigning higher noise levels to edited regions (to facilitate content change) and lower noise levels to unedited regions (to maintain content consistency). We introduce a Noise Guidance Mechanism (NGM), which leverages the video prior in the generative model and effectively integrates rich information within the noisy latent to guide the denoising process, thereby preserving unedited content and overall visual coherence. Experiments show that our proposed method achieves better visual quality and state-of-the-art performance.\n\n---\n\n## 144. DeltaPrompts: Escaping the Zero-Delta Trap in Multimodal Distillation\n- **Authors**: Jaehun Jung , Hyunwoo Kim , Brandon Cui , Ximing Lu , David Acuna , Prithviraj Ammanabrolu , Yejin Choi\n- **URL**: [https://arxiv.org/abs/2605.15532](https://arxiv.org/abs/2605.15532)\n- **Abstract**:\n> Distillation enables compact Vision-Language Models (VLMs) to obtain strong reasoning capabilities, yet the prompts driving this process are typically chosen via simple heuristics or aggregated from off-the-shelf datasets. We reveal a critical inefficiency in this approach: up to 69% of the prompts in standard chart / document reasoning datasets are effectively zero-delta, meaning the teacher and student already induce the exact same answer distribution. Training on these prompts provides minimal learning signal, causing student improvement to rapidly saturate regardless of data scale. To escape the zero-delta trap, we return to first principles: distillation fundamentally minimizes distributional divergence, and thus a prompt is valuable only if it exposes a functional capability gap between the teacher and student. We quantify this gap through answer divergence ($\\Delta$), demonstrating that non-zero divergence is critical for effective scaling. Building on this insight, we propose a staged synthesis pipeline that repurposes existing datasets as seeds, actively targeting student failure modes to produce better prompts. The result is DeltaPrompts, a diverse dataset of 200k synthetic, high-divergence reasoning problems. We evaluate DeltaPrompts across three distinct settings: on-policy distillation with the target teacher-student pair, transfer to a novel model family without regenerating the data, and off-policy fine-tuning of a non-reasoning model. Across all scenarios, DeltaPrompts drives substantial gains, yielding up to 15% relative improvement even on top of a highly-optimized reasoning model (e.g., Qwen3-VL-8B-Thinking) -- averaged over 10 benchmarks spanning chart, document and perception-centric reasoning.\n\n---\n\n## 145. Process Rewards with Learned Reliability\n- **Authors**: Jinyuan Li , Langlin Huang , Chengsong Huang , Shaoyang Xu , Donghong Cai , Yuyi Yang , Wenxuan Zhang , Jiaxin Huang\n- **URL**: [https://arxiv.org/abs/2605.15529](https://arxiv.org/abs/2605.15529)\n- **Abstract**:\n> Process Reward Models (PRMs) provide step-level feedback for reasoning, but current PRMs usually output only a single reward score for each step. Downstream methods must therefore treat imperfect step-level reward predictions as reliable decision signals, with no indication of when these predictions should be trusted. We propose BetaPRM, a distributional PRM that predicts both a step-level success probability and the reliability of that prediction. Given step-success supervision from Monte Carlo continuations, BetaPRM learns a Beta belief that explains the observed number of successful continuations through a Beta-Binomial likelihood, rather than regressing to the finite-sample success ratio as a point target. This learned reliability signal indicates when a step reward should be trusted, enabling downstream applications to distinguish reliable rewards from uncertain ones. As one application, we introduce Adaptive Computation Allocation (ACA) for PRM-guided Best-of-N reasoning. ACA uses the learned reliability signal to stop when a high-reward solution is reliable and to spend additional computation on uncertain candidate prefixes. Experiments across four backbones and four reasoning benchmarks show that BetaPRM improves PRM-guided Best-of-N selection while preserving standard step-level error detection. Built on this signal, ACA improves the accuracy--token tradeoff over fixed-budget Best-of-16, reducing token usage by up to 33.57% while improving final-answer accuracy.\n\n---\n\n## 146. Neural Point-Forms\n- **Authors**: Bruno Trentini , Jacob Hume , Vincenzo Antonio Isoldi , Philipp Misof , Ekaterina S. Ivshina , Kelly Maggs\n- **URL**: [https://arxiv.org/abs/2605.15524](https://arxiv.org/abs/2605.15524)\n- **Abstract**:\n> Point cloud learning often rests on the premise that observed samples are noisy traces of an underlying geometric object, such as a manifold embedded in a high-dimensional feature space. Yet much of this geometry is not captured directly by coordinates, pairwise distances, or learned graph neighborhoods alone. In the smooth setting, differential forms are devices to encode higher order tangency information. In this work, we introduce a new family of principled learnable geometric features for point clouds called neural point-forms (NPFs). In the absence of a natural tangency structure, we instead use Laplacian-based techniques from Diffusion Geometry to build a discrete model for comparing differential forms on point clouds via inner products. In the continuum, submanifolds of a shared ambient feature space are represented as comparison matrices, whose entries describe how pairs of feature forms interact with extrinsic tangency information. We make this intuition precise by proving the long-run consistency of comparison matrices under standard sampling, bandwidth, density, and manifold-hypothesis assumptions. This yields a compact, efficient and permutation-invariant neural layer whose output is a learned form-comparison matrix. Across synthetic and biologically relevant experiments, we show that NPFs provide a competitive, and interpretable representation, with the strongest benefits appearing when labels depend on sampling density, manifold-like structure, or response-relevant population geometry.\n\n---\n\n## 147. On the Fragility of Data Attribution When Learning Is Distributed\n- **Authors**: Xian Gao , Bo Hui , Min-Te Sun , Wei-Shinn Ku\n- **URL**: [https://arxiv.org/abs/2605.15520](https://arxiv.org/abs/2605.15520)\n- **Abstract**:\n> Data attribution has become an important component of pricing, auditing, and governance in machine learning pipelines, yet most attribution methods implicitly assume that attribution values faithfully reflect participants' contributions. We show that this assumption can fail: a single participant in a standard distributed training workflow can substantially inflate its measured attribution value while preserving global utility. Our attribution-first attack uses latent optimization to inject small synthetic batches that preserve utility while exploiting non-IID label coverage and evaluator sensitivities. Across datasets, models, and multiple marginal-utility evaluators, the attack consistently increases the adversary's attribution value and reshapes the relative attribution structure among benign clients without degrading accuracy or triggering geometry-based defenses. These results show that attribution itself forms a new attack surface and motivate the development of attribution-robust and incentive-compatible scoring mechanisms.\n\n---\n\n## 148. DiffVAS: Diffusion-Guided Visual Active Search in Partially Observable Environments\n- **Authors**: Anindya Sarkar , Srikumar Sastry , Aleksis Pirinen , Nathan Jacobs , Yevgeniy Vorobeychik\n- **URL**: [https://arxiv.org/abs/2605.15519](https://arxiv.org/abs/2605.15519)\n- **Abstract**:\n> Visual active search (VAS) has been introduced as a modeling framework that leverages visual cues to direct aerial (e.g., UAV-based) exploration and pinpoint areas of interest within extensive geospatial regions. Potential applications of VAS include detecting hotspots for rare wildlife poaching, aiding search-and-rescue missions, and uncovering illegal trafficking of weapons, among other uses. Previous VAS approaches assume that the entire search space is known upfront, which is often unrealistic due to constraints such as a restricted field of view and high acquisition costs, and they typically learn policies tailored to specific target objects, which limits their ability to search for multiple target categories simultaneously. In this work, we propose DiffVAS, a target-conditioned policy that searches for diverse objects simultaneously according to task requirements in partially observable environments, which advances the deployment of visual active search policies in real-world applications. DiffVAS leverages a diffusion model to reconstruct the entire geospatial area from sequentially observed partial glimpses, which enables a target-conditioned reinforcement learning-based planning module to effectively reason and guide subsequent search steps. Extensive experiments demonstrate that DiffVAS excels in searching diverse objects in partially observable environments, significantly surpassing state-of-the-art methods on several datasets.\n\n---\n\n## 149. RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably\n- **Authors**: Yufeng Du , Phillip Harris , Minyang Tian , Eliu A Huerta , Srikanth Ronanki , Subendhu Rongali , Aram Galstyan , Hao Peng\n- **URL**: [https://arxiv.org/abs/2605.15514](https://arxiv.org/abs/2605.15514)\n- **Abstract**:\n> We identify intrinsic limitations of Rotary Positional Embeddings (RoPE) in Transformer-based long-context language models. Our theoretical analysis abstracts away from the specific content of the context and depends only on its length. We prove that as context length increases, RoPE-based attention becomes unpredictable and loses two properties that are central to its effectiveness. First, it loses its locality bias: RoPE is no more likely to favor nearer positions than substantially farther ones. Second, it loses consistency in token relevance: a key vector that receives a higher attention score than an alternative at one position may receive a lower score at another. In both cases, the probability of failure approaches 0.5, no better than random guessing. We further prove that the attention score can remain unchanged when a key token is moved to a different position, or even replaced by a different token, indicating a failure to distinguish positions or tokens. Adjusting the RoPE base trades off distinguishing positions against distinguishing tokens but cannot preserve both at the same time. Increasing the RoPE base hyperparameter, a common practice in today's long-context models, helps distinguish different tokens, but inevitably sacrifices the ability to distinguish positions. Our empirical analysis shows that multi-head, multi-layer architectures are insufficient to overcome these limitations. Our findings suggest that fundamentally new mechanisms for encoding position and token order may be needed in future Transformer long-context language models.\n\n---\n\n## 150. PrismQuant: Rate-Distortion-Optimal Vector Quantization for Gaussian-Mixture Sources\n- **Authors**: Bumsu Park , Chanho Park , Youngmok Park , Namyoon Lee\n- **URL**: [https://arxiv.org/abs/2605.15507](https://arxiv.org/abs/2605.15507)\n- **Abstract**:\n> For a Gaussian source under mean-squared error (MSE), classical transform coding is rate--distortion (RD) optimal: the Karhunen--Loeve transform (KLT) diagonalizes the covariance, reverse waterfilling allocates the bits, and scalar quantization closes the loop. This elegant story breaks down for multimodal sources, where no single covariance can capture heterogeneous local geometries, and the RD function loses its closed form. We revisit this problem through Gaussian-mixture sources and develop a constructive RD theory for them. Our key finding is that the mixture structure incurs only a component label cost. Conditioned on the active mixture component, each branch is Gaussian; the challenge is allocating bits across heterogeneous branches. We prove that the genie-aided conditional RD function is governed by a single global reverse-waterfilling level shared across all components and eigenmodes. Building on this result, we introduce PrismQuant, which transmits the component label losslessly and encodes the residual using the component-matched KLT, followed by scalar quantization, achieving a rate of H(C)/n bits per source dimension of the converse, with a vanishing asymptotic gap. We further develop a practical implementation based on EM-driven Gaussian-mixture learning, component-adaptive KLTs, and entropy-constrained scalar quantization (ECSQ). Experiments on synthetic Gaussian mixtures show that PrismQuant closely approaches the theoretical RD bound, while experiments on real-world channel-state-information (CSI) data demonstrate competitive or superior performance compared with transformer-based learned codecs at more than one order of magnitude smaller model size.\n\n---\n\n## 151. Learning with Conflicts of Interest\n- **Authors**: Nischal Aryal , Arash Termehchy , Ali Vakilian , Marianne Winslett\n- **URL**: [https://arxiv.org/abs/2605.15504](https://arxiv.org/abs/2605.15504)\n- **Abstract**:\n> Financial, social, and political factors often prevent the interests of the owners of ML systems and services and their users from being perfectly aligned. ML systems often produce biased information that can influence users to make decisions that are not in their best interest. Current solution approaches require ML systems to implement protocols to mitigate their biases. However, ML system owners usually do not have any incentive to implement these protocols and often argue that it limits their freedom of expression or business. We believe that a successful solution to this problem must recognize the conflict of interest between the ML systems and their users, and use this information to protect users against information that adversely influences their decisions while allowing users to safely benefit from these systems. To this end, we propose a game-theoretic framework that models the interaction between ML systems and users with conflicts of interest. We present scalable algorithms with theoretical guarantees that maximize the amount of desired information and actions and minimize the amount of biased and manipulative actions in interaction with ML systems.\n\n---\n\n## 152. Ghosted Layers: Unconstrained Activation Alignment for Recovering Layer-Pruned LLMs\n- **Authors**: Vincent-Daniel Yun , Junhyuk Jo , Sai Praneeth Karimireddy , Sunwoo Lee\n- **URL**: [https://arxiv.org/abs/2605.15491](https://arxiv.org/abs/2605.15491)\n- **Abstract**:\n> Layer pruning removes entire Transformer decoder blocks from large language models, but introduces a mismatch between the hidden state received by the next surviving layer and the distribution it was trained to process, leading to significant performance degradation. We propose Ghosted Layers, a training-free recovery module that addresses this issue by solving a boundary activation alignment problem. Our method derives a closed-form optimal linear operator from a small calibration set to reconstruct the activation discrepancy introduced by the pruned layers. We show that this solution corresponds to the unconstrained optimum of the alignment objective, whereas existing methods are restricted to constrained solutions over limited operator subspaces. Experiments across multiple LLM backbones and pruning strategies demonstrate that our method consistently improves accuracy and perplexity over prior training-free baselines, while preserving the efficiency gains of layer pruning.\n\n---\n\n## 153. Hybrid LLM-based Intelligent Framework for Robot Task Scheduling\n- **Authors**: Swayamjit Saha , Subhabrata Das , Haonan Duan , Xiao-Yang Liu\n- **URL**: [https://arxiv.org/abs/2605.15486](https://arxiv.org/abs/2605.15486)\n- **Abstract**:\n> This study introduces intelligent frameworks that use Large Language Models (LLMs) to improve task scheduling for construction robots. The LLM is fed with key data about the desired task, such as agent action abilities, and the desired end goal to be achieved. A well-balanced allocation strategy is developed, optimizing both time efficiency and resource utilization. Our system utilizes a Natural Language Processing interface to streamline communication with construction professionals and adapt in real-time to unexpected site conditions. We concurrently use two LLM agents, specifically generator (GPT-4) and supervisor (Gemma 3/Llama 4/Mistral 7b) LLM agents to provide a more precise task schedule. We evaluate the proposed methodology using a straightforward scenario and provide metric scores to prove the efficacy of the frameworks. Our results highlight that the implementation of LLMs is crucial in construction operational tasks including robots.\n\n---\n\n## 154. Residual Reinforcement Learning for Robot Teleoperation under Stochastic Delays\n- **Authors**: Kaize Deng , Zewen Yang\n- **URL**: [https://arxiv.org/abs/2605.15480](https://arxiv.org/abs/2605.15480)\n- **Abstract**:\n> Stochastic communication delays in teleoperation introduce signal discontinuities that undermine control stability and degrade control performance. Consequently, the conventional reinforcement learning (RL) methods struggle with the delayed observations due to the delay-induced observations, leading to high-frequency chattering. To address this, we propose a hybrid control framework, delay-resilient RL, integrating a state estimator utilizing Long Short-Term Memory (LSTM) with a residual RL policy, which is resilient to stochastic delays. The LSTM reconstructs smooth, continuous state estimates from delayed observations, enabling the RL agent to learn a residual torque compensation policy that balances tracking accuracy with velocity smoothness. Experimental validation on Franka Panda robots demonstrates that our approach significantly outperforms the state-of-the-art baselines, ensuring robust and stable teleoperation even under high-variance stochastic delays.\n\n---\n\n## 155. Retrieval-Augmented Large Language Models for Schema-Constrained Clinical Information Extraction\n- **Authors**: A H M Rezaul Karim , Ozlem Uzuner\n- **URL**: [https://arxiv.org/abs/2605.15467](https://arxiv.org/abs/2605.15467)\n- **Abstract**:\n> Conversational nurse-patient transcripts contain actionable observations, but converting these transcripts into structured representations at scale remains challenging. Documentation burden is substantial, with prior studies showing clinicians spend large portions of their workday on documentation and related desk work rather than direct patient care. MEDIQA-SYNUR focuses on observation extraction from conversational nurse-patient transcripts, requiring systems to normalize these narratives into a predefined schema with value-type constraints. We propose a modular retrieval-augmented generation (RAG) pipeline that uses the training set as an exemplar corpus, combines schema-constrained prompting (full schema vs. pruned candidate schema), deterministic schema-based postprocessing, and a second-pass audit, with two LLM backbones: Llama-4-Scout-17B-16E-Instruct and GPT-5.2 with corresponding embedding models for RAG. Our best configuration uses GPT-5.2 with full schema, RAG, and a second-pass auditing, achieving 80.36% F1 score. Overall, our results show that RAG consistently improves performance, while the optimal degree of schema constraint depends on the model, and second-pass auditing yields modest additional gains by correcting residual schema-adherence errors.\n\n---\n\n## 156. GRLO: Towards Generalizable Reinforcement Learning in Open-Ended Environments from Zero\n- **Authors**: Shangjian Yin , Yu Fu , Yue Dong , Zhouxing Shi\n- **URL**: [https://arxiv.org/abs/2605.15464](https://arxiv.org/abs/2605.15464)\n- **Abstract**:\n> Post-training has become a crucial step for unlocking the capabilities of large language models, with reinforcement learning (RL) emerging as a critical paradigm. Recent RL-based post-training has increasingly split into two paradigms: reinforcement learning from human feedback (RLHF), which optimizes models using human preference signals in target domains, and reinforcement learning from verifiable rewards (RLVR), which operates in verifier-backed environments. The latter has dominated recent reasoning-oriented post-training because it delivers stronger gains and higher efficiency on domain-specific tasks (e.g., reasoning). However, although in-domain RL training achieves promising performance, it still requires a substantial amount of GPU compute, which remains a major barrier to broad adoption. In this work, we study the generalization ability of RLHF learned from scratch from a small set of interactions in open-ended environments, and investigate whether the conversational abilities it explicitly acquires can implicitly transfer to downstream tasks such as mathematical reasoning and code generation, namely GRLO. Specifically, on Qwen3-4B-Base backbone, GRLO improves the average performance across all domains from 24.1 to 63.1 with only 5K prompts and 22.7 GPU hours, requiring about $46\\times$ less data and $68\\times$ less compute than a strong in-domain RLVR baseline. The resulting model is even competitive with Qwen's released post-trained models which required a much larger training cost. Notably, a subsequent in-domain RLVR stage brings only selective gains, mainly on harder competition-math benchmarks. We hope GRLO offers a simple and efficient recipe for building broadly capable post-trained models. Our code and data will be available at: \\href{ this https URL }{ this https URL }.\n\n---\n\n## 157. DrugSAGE:Self-evolving Agent Experience for Efficient State-of-the-Art Drug Discovery\n- **Authors**: Yikun Zhang , Xiwei Cheng , Tianyu Liu , Yuanqi Du , Wengong Jin\n- **URL**: [https://arxiv.org/abs/2605.15461](https://arxiv.org/abs/2605.15461)\n- **Abstract**:\n> Building state-of-the-art (SOTA) predictive models for drug discovery requires expensive search over tools, architectures, and training strategies. Current LLM-based agents can find SOTA solutions through extensive trial and error, but they do not retain the experience accumulated along the way and therefore pay the full search cost on every new task. We propose \\method (Self-evolving Agent Experience), a framework that accumulates and reuses experience across tasks to build SOTA drug discovery models efficiently. \\method maintains a cross-task memory of verified skills, statistical evidence about effective strategies, and a record of recurring errors and their fixes. In some cases, \\method transfers a working solution directly without test-time search. In 33 molecular property prediction tasks, \\method ranks first among nine SOTA agents in a single-task setting. With memory accumulated from 16 smaller tasks, \\method achieves an averaged normalized score of 0.935 on 17 held-out tasks in a cross-task evaluation setting and outperforms all baseline agents by 10-30\\% in a zero-test-time search regime. In summary, our work shows the advantage of cross-task memory for efficient SOTA model development in drug discovery.\n\n---\n\n## 158. Differentially Private Motif-Preserving Multi-modal Hashing\n- **Authors**: Zehua Cheng , Wei Dai , Jiahao Sun\n- **URL**: [https://arxiv.org/abs/2605.15460](https://arxiv.org/abs/2605.15460)\n- **Abstract**:\n> Cross-modal hashing enables efficient retrieval by encoding images and text into compact binary codes. State-of-the-art methods rely on semantic similarity graphs derived from user interactions for supervision, yet these graphs encode sensitive behavioral patterns vulnerable to link reconstruction attacks. Existing privacy-preserving approaches fail on graph-structured data: Differentially Private SGD destroys relational motifs by treating samples independently, while graph synthesis methods suffer from unbounded local sensitivity in scale-free networks, hub nodes cause single-edge modifications to alter triangle counts by $\\mathcal{O}(N)$, necessitating prohibitive noise injection. We term this phenomenon Hubness Explosion. We propose DMP-MH, a Sanitize-then-Distill framework that decouples privacy from representation learning. Our approach first bounds sensitivity by deterministically clipping node degrees, capping the $L_2$-sensitivity of triangle motifs independently of dataset size. A sanitized synthetic graph is then generated via Noisy Mirror Descent under $(\\epsilon,\\delta)$-Edge Differential Privacy. Finally, dual-stream hashing networks distill this topology using a holistic structural loss that enforces cross-modal alignment. Evaluated on MIRFlickr-25K and NUS-WIDE under a strict inductive protocol, DMP-MH outperforms private baselines by up to 11.4 mAP points while retaining up to 92.5% of non-private performance.\n\n---\n\n## 159. RIDE: Retinex-Informed Decoupling for Exposing Concealed Objects\n- **Authors**: Chunming He , Rihan Zhang , Dingming Zhang , Chengyu Fang , Longxiang Tang , Jingjia Feng , Fengyang Xiao , Sina Farsiu\n- **URL**: [https://arxiv.org/abs/2605.15450](https://arxiv.org/abs/2605.15450)\n- **Abstract**:\n> Concealed Object Segmentation (COS) encompasses a family of dense-prediction tasks, including camouflaged object detection, polyp segmentation, transparent object detection, and industrial defect inspection, where targets are visually entangled with their surroundings through different physical mechanisms. Existing methods either operate directly on RGB images or employ \\emph{heterogeneous} decompositions (\\eg, Fourier, wavelet) that redistribute spatial evidence across scale/frequency coefficients, making pixel-aligned cues less direct. We introduce a fundamentally different perspective: \\textbf{homogeneous image decomposition} via Retinex theory, which factorizes an image into illumination and reflectance components within the \\emph{same} spatial domain. Our key insight is that visual entanglement enforces appearance matching in the composite space, but this does \\emph{not} necessitate simultaneous matching in both component spaces, a phenomenon we formalize as the \\textbf{Discriminability Gap Theorem}. Crucially, we show that across diverse COS sub-tasks, the underlying physical processes systematically anti-correlate illumination and reflectance differences, yielding theoretical guarantees that Retinex decomposition preserves or strictly improves total foreground--background discriminability across the full physical regime, with anti-correlation maximizing the gain. Building on this, we propose \\textbf{RIDE} comprising: (i) a Task-Driven Retinex Decomposition module that learns segmentation-optimal factorizations end-to-end; (ii) a Discriminability Gap Attention mechanism that adaptively exploits where decomposition helps; and (iii) a Camouflage-Breaking Contrastive loss operating in reflectance feature space.\n\n---\n\n## 160. Runtime-Structured Task Decomposition for Agentic Coding Systems\n- **Authors**: Shubhi Asthana , Bing Zhang , Chad DeLuca , Hima Patel , Ruchi Mahindru\n- **URL**: [https://arxiv.org/abs/2605.15425](https://arxiv.org/abs/2605.15425)\n- **Abstract**:\n> Agentic coding systems increasingly use large language models (LLMs) for software engineering tasks such as debugging, root cause analysis, and code review. However, many existing systems encode task logic, execution flow, and output generation inside monolithic prompts. This design creates brittle behavior, limited debuggability, and high retry costs because failures often require rerunning the full workflow. We present runtime-structured task decomposition, an architectural approach in which task partitioning and execution flow are managed through executable control logic rather than prompt structure alone. LLMs are used only for focused judgment tasks, and outputs are validated against predefined schemas before downstream execution. We evaluate this approach on two software engineering workloads using three configurations: monolithic execution, static decomposition with fixed subtasks and no runtime branching, and runtime-structured decomposition. Each configuration was evaluated across 10 runs. Our results show that decomposition alone does not necessarily reduce retry cost. In the Kubernetes root cause analysis workload, the static decomposition baseline produced a retry cost of 1,632 +/- 145 tokens versus 904 +/- 17 tokens for the monolithic baseline because failures forced reruns of downstream subtasks. A similar pattern appeared in the multi-file debugging workload, where the static baseline consumed 933 tokens compared to 703 tokens for the monolithic system. The runtime-structured approach reran only failed subtasks, reducing retry costs to 436 +/- 132 tokens for root cause analysis and 460 tokens for debugging. Overall, the approach achieved up to 51.7% lower retry cost than monolithic systems and 73.2% lower retry cost than static decomposition baselines, improving efficiency, debuggability, and operational reliability in agentic coding systems.\n\n---\n\n## 161. MR2-ByteTrack: CNN and Transformer-based Video Object Detection for AI-augmented Embedded Vision Sensor Nodes\n- **Authors**: Luca Bompani , Manuele Rusci , Luca Benini , Daniele Palossi , Francesco Conti\n- **URL**: [https://arxiv.org/abs/2605.15423](https://arxiv.org/abs/2605.15423)\n- **Abstract**:\n> Modern smart vision sensors need on-device intelligence to process video streams, as cloud computing is often impractical due to bandwidth, latency, and privacy constraints. However, these sensory systems typically rely on ultra-low-power microcontrollers (MCUs) with limited memory and compute, making conventional video object detection methods, which require feature storage or multi-frame buffering, unfeasible. To address this challenge, we introduce Multi-Resolution Rescored ByteTrack (MR2-ByteTrack), a Video Object Detection (VOD) method tailored for MCU-based embedded vision nodes. MR2-ByteTrack reduces computational cost by alternating between full- and low-resolution inference, while linking detections across frames via ByteTrack and correcting misclassifications through the Rescore algorithm, which applies probability union rules to aggregate detection confidence scores across frames. We apply our approach to both a CNN-based detector and a Transformer-based model, demonstrating its generality across architectures with fundamentally different spatial processing. Experiments on ImageNetVID demonstrate that MR2-ByteTrack maintains accuracy, achieving mAP scores of up to 49.0 for the CNN-based models and 48.7 for the Transformer, while reducing multiply-accumulate operations by as much as 53\\% for the CNNs and 32\\% for the Transformer. When deployed on GAP9, an ultra-low-power RISC-V multicore MCU, our method yields up to 55\\% energy savings compared to processing only full-resolution images, enabling the first real-time Transformer-based VOD on an MCU-class embedded vision node. Code available at this https URL\n\n---\n\n## 162. $f$-Trajectory Balance: A Loss Family for Tuning GFlowNets, Generative Models, and LLMs with Off- and On-Policy Data\n- **Authors**: Jake Fawkes , Jason Hartford\n- **URL**: [https://arxiv.org/abs/2605.15417](https://arxiv.org/abs/2605.15417)\n- **Abstract**:\n> In GFlowNets and variational inference, it has been shown that the mean square error between target and model log probabilities is an effective, low variance, surrogate loss for training generative models. This loss has the property that when evaluated \\emph{on-policy} its gradients correspond to those of the KL divergence, while \\emph{off-policy} it remains a valid loss with the same global minimizer. In this work, we demonstrate that this construction can be extended to the whole family of $f$-divergences, leading to a family of losses whose on-policy gradients are that of the corresponding $f$-divergence, but retain the same global minimizer off-policy. Specifically, we show that the on-policy gradients lead to a one to one correspondence between translation invariant loss functions on the target and model log probabilities, and $f$-divergences. This equivalence allows us to design new surrogate loss functions for tuning a wide class of generative models that inherit the properties of the corresponding $f$-divergence, such as being more mode covering, whilst being applicable to off-policy data. We apply our losses on a range of tasks, including classic synthetic examples, SynFlowNets for molecule discovery, and asynchronous large language model (LLM) tuning, demonstrating that our models retain their predicted properties on- and off-policy in a wide class of generative models.\n\n---\n\n## 163. Margin-Adaptive Confidence Ranking for Reliable LLM Judgement\n- **Authors**: Gaojie Jin , Yong Tao , Lijia Yu , Tianjin Huang\n- **URL**: [https://arxiv.org/abs/2605.15416](https://arxiv.org/abs/2605.15416)\n- **Abstract**:\n> Jung et al. (2025) introduce a hypothesis testing framework for guaranteeing agreement between large language models (LLMs) and human judgments, relying on the assumption that the model's estimated confidence is monotonic with respect to human-disagreement risk. In practice, however, this assumption may be violated, and the generalization behavior of the confidence estimator is not explicitly analyzed. We mitigate these issues by learning a dedicated confidence estimator instead of relying on heuristic confidence signals. Our approach leverages simulated annotator diversity and a margin-based ranking formulation to explicitly model how confidently an LLM distinguishes between human-agreement and human-disagreement cases. We further derive generalization guarantees for this estimator, revealing a margin-dependent trade-off that informs the design of an adaptive estimator training procedure. When integrated into fixed-sequence testing, the learned confidence estimator yields improved ranking accuracy and empirically strengthens the monotonic relationship between confidence and disagreement risk, leading to higher success rates in satisfying target agreement levels across multiple datasets and judge models.\n\n---\n\n## 164. From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery\n- **Authors**: Lingzhe Zhang , Tong Jia , Yunpeng Zhai , Zixuan Xie , Chiming Duan , Minghua He , Philip S. Yu , Ying Li\n- **URL**: [https://arxiv.org/abs/2605.15412](https://arxiv.org/abs/2605.15412)\n- **Abstract**:\n> Modern quantitative trading increasingly relies on systematic models to extract predictive signals from large-scale financial data, where alpha factor discovery plays a central role in transforming market observations into tradable signals. Recent LLM-based methods have shown promise in automating factor generation, but most of them still rely on prompt-level generation--evaluation--feedback loops for iterative optimization. As the loop becomes longer, repeatedly appended historical candidates and feedback can cause context explosion, increase inference cost, dilute useful information, and introduce feedback drift. Moreover, these methods often depend on very large LLMs whose stable generation preferences may lead to structurally similar expressions, redundant candidates, and search stagnation. To address these limitations, we propose \\textsc{QuantEvolver}, a self-evolving alpha factor discovery framework based on reinforcement fine-tuning. Instead of accumulating feedback in the prompt, \\textsc{QuantEvolver} converts executable quantitative evaluation into policy updates, enabling a Miner LLM to internalize historical optimization experience through parameter learning. Specifically, \\textsc{QuantEvolver} constructs high-quality seed factors, builds diverse seed--time-window training tasks, generates executable Factor DSL expressions, evaluates them through Regime Backtest, and optimizes the Miner LLM with Diversity-Complementarity Reward. During training, high-quality factors are continuously accumulated in a Mined Factor Database, which serves as the final discovered factor library. Extensive experiments on three realistic market benchmarks demonstrate the effectiveness of \\textsc{QuantEvolver}, which consistently improves the primary evaluation metric of each task over existing LLM-based alpha factor discovery baselines, produces higher-quality and more complementary factor pools.\n\n---\n\n## 165. Diagonal Adaptive Non-local Observables on Quantum Neural Networks\n- **Authors**: Huan-Hsin Tseng , Yan Li , Hsin-Yi Lin , Samuel Yen-Chi Chen\n- **URL**: [https://arxiv.org/abs/2605.15410](https://arxiv.org/abs/2605.15410)\n- **Abstract**:\n> Adaptive Non-local Observables (ANOs) have shown that making quantum observables dynamic can substantially enlarge the function space of Variational Quantum Algorithms, partly shifting hardware demands from circuit synthesis to measurement design. However, this advantage is accompanied by a steep increase in the number of parameters, as well as the classical optimization cost for varying general Hermitian observables. We propose a special form of ANO that significantly reduces this burden by considering only diagonal observables paired with quantum circuits. Mathematically, this is equivalent to the full ANO of a large parameter space since diagonal matrices are canonical representatives of the ANO space modulo unitary similarity. As a result, Diagonal ANO retains the same capability of full ANO while reducing $k$-local observable complexity from $O(4^k)$ to $O(2^k)$ and lowering the corresponding measurement-side classical computation. In this sense, diagonal ANO preserves much of the benefit of full ANO while encompassing conventional VQCs as a special case.\n\n---\n\n## 166. Amortized Energy-Based Bayesian Inference\n- **Authors**: Hojjat Kaveh , Ricardo Baptista , Andrew M. Stuart\n- **URL**: [https://arxiv.org/abs/2605.15407](https://arxiv.org/abs/2605.15407)\n- **Abstract**:\n> We consider amortized Bayesian inference for nonlinear inverse problems in settings where only samples from the joint distribution of parameters and observations are available. Classical methods such as Markov chain Monte Carlo require solving a new inference problem for each observation, which can be computationally prohibitive when inference must be repeated many times. We propose a transport-based approach that learns an observation-dependent map pushing forward a reference measure to approximate the posterior distribution. The map is trained by minimizing an averaged energy-distance objective between the true posterior and the learned pushforward. This formulation is likelihood-free, requiring only joint samples, and avoids density evaluation, invertibility constraints, and Jacobian determinant computations. For function-space inverse problems with Gaussian priors, we parameterize the transport map as the identity plus a perturbation in the Cameron-Martin space of the prior, preserving absolute continuity with respect to the prior. In infinite-dimensional settings, the map is represented using neural operators. We illustrate the method on a finite-dimensional nonlinear inverse problem and two PDE-constrained inverse problems arising in porous medium flow and seismic inversion. The results show that the learned transport captures posterior structure, including multimodality and dominant modes, while enabling fast posterior sampling for new observations.\n\n---\n\n## 167. Breakeven complexity: A new perspective on neural partial differential equation solvers\n- **Authors**: Yijing Zhang , Nicholas Roberts , Tanya Marwah , Mikhail Khodak\n- **URL**: [https://arxiv.org/abs/2605.15399](https://arxiv.org/abs/2605.15399)\n- **Abstract**:\n> Neural surrogate solvers of partial differential equations (PDEs) promise dramatic speedups over numerical methods, especially in scenarios requiring many solves. However, current accuracy-based evaluations do not fully consider two central issues: (1) neural solvers incur substantial up-front costs for data generation, training, and tuning; and (2) classical solvers can also generate low-fidelity solutions at a sufficiently low simulation cost. To explicitly account for these realities and fully incorporate end-to-end costs, we propose an evaluation framework centered on breakeven complexity, a metric that counts the forward solves before a learned solver is cost-effective relative to an error-equivalent traditional solver. To evaluate this measure, we apply scaling laws to determine how much training budget to allocate to data generation and discuss how to achieve smooth error-matching in diverse settings. We evaluate the breakeven complexity of multiple neural PDE solvers on three PDEs on 2D periodic domains from APEBench and a novel benchmark of flows past multiple obstacles generated by the GPU-native PyFR code. Among other findings, our results suggest that neural PDE solvers become more effective as problems get harder in terms of cost, dimension, rollout, physics regime (e.g. higher Reynolds number), etc.\n\n---\n\n## 168. Representation Without Reward: A JEPA Audit for LLM Fine-Tuning\n- **Authors**: Biswa Sengupta\n- **URL**: [https://arxiv.org/abs/2605.15394](https://arxiv.org/abs/2605.15394)\n- **Abstract**:\n> Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outputs. For autoregressive language-model fine-tuning the principle entails a stricter requirement: the induced hidden-state geometry must reach the language-model head \\emph{and} improve the decoded task metric. We test that requirement under a fixed Llama-3.2-1B-Instruct LoRA harness on natural-language-to-regex generation, comparing twenty-two training-time auxiliaries across trajectory-shape regularisation, distributional constraints, predictor/target asymmetry, Fisher-metric Jacobi residuals, and a decoder-visible JEPA objective constructed to lie in cross-entropy's positive cone. The empirical answer is a structured null: several auxiliaries clear single-cell paired $\\alpha = 0.10$ without correction (T3-Local at $\\Delta = +2.53$~pp, $p = 0.003$ being the strongest), but none survives Bonferroni or Holm--Bonferroni at the relevant family-wise threshold, even though many change curvature, anisotropy, variance, and gradient direction. Decoder-visible JEPA yields the first positive auxiliary--cross-entropy gradient cosine in the study, yet exact match remains inside seed noise; a full-fine-tuning replication of the same auxiliary at $n = 5$ seeds reproduces the null on both benchmarks (TURK: $\\Delta = +0.04$~pp, $p_{\\text{paired} } = 0.96$; SYNTH: $\\Delta = +0.52$~pp, $p_{\\text{paired} } = 0.28$), so the null is robust across LoRA and full fine-tuning for the decoder-visible construction. Hidden-state representation work and decoded-task accuracy are therefore weakly coupled in this regime; we accordingly reframe LLM-domain JEPA evaluation as a coupling problem, in which the operative question is under which metrics useful hidden geometry becomes decoder-visible task signal.\n\n---\n\n## 169. PanoWorld: Geometry-Consistent Panoramic Video World Modeling\n- **Authors**: Le Jiang , Xiangyu Bai , Bishoy Galoaa , Shayda Moezzi , Caleb James Lee , Tooba Imtiaz , Edmund Yeh , Jennifer Dy , Yanzhi Wang , Sarah Ostadabbas\n- **URL**: [https://arxiv.org/abs/2605.15391](https://arxiv.org/abs/2605.15391)\n- **Abstract**:\n> We present PanoWorld, a panoramic video world model that generates geometry-consistent 360$\\degree$ video from a single image and a caption. Existing panoramic video methods optimize primarily for visual realism and do not explicitly constrain the underlying 3D scene state, producing outputs that appear plausible yet exhibit inconsistent depth, broken correspondences, and implausible motion across the spherical surface. We address this gap by framing panoramic video generation as a geometry- and dynamics-consistent latent state modeling problem rather than pure visual synthesis. Building on a pre-trained perspective video world model, we introduce two lightweight regularizers: a depth consistency loss against pseudo ground-truth panoramic depth, and a trajectory consistency loss that supervises the 3D world-frame positions of tracked points across time. We further apply spherical-geometry-aware adaptation to the conditioning and positional encoding. We additionally introduce PanoGeo, a unified geometry-aware panoramic video dataset with consistent depth, trajectory, and prompt annotations across diverse real and synthetic sources, used for both training and stratified evaluation. Experiments show that PanoWorld improves geometric consistency over prior panoramic generation methods while maintaining competitive visual realism, establishing that panoramic video generation must be treated as a geometric modeling problem to support the holistic spatial understanding requirements of embodied AI applications. Code is available at this https URL .\n\n---\n\n## 170. Is One Score Enough? Rethinking the Evaluation of Sequentially Evolving LLM Memory\n- **Authors**: Songwei Dong , Zihan Chen , Chengshuai Shi , Peng Wang , Jundong Li , Cong Shen\n- **URL**: [https://arxiv.org/abs/2605.15384](https://arxiv.org/abs/2605.15384)\n- **Abstract**:\n> Memory plays a central role in enabling large language models (LLMs) to operate over sequential tasks by accumulating and reusing experience over time. However, existing evaluations of LLM memory mostly rely on aggregate metrics such as final hold-out accuracy or cumulative online performance, which can obscure critical failure modes such as forgetting and negative transfer. In this paper, we introduce SeqMem-Eval, a diagnostic evaluation framework for sequentially evolving LLM memory. Drawing inspiration from continual learning, it targets a test-time setting in which memory is external, prompt-mediated, and updated without modifying model parameters. Rather than focusing only on final performance, SeqMem-Eval evaluates how memory states evolve, generalize, consolidate experience, and retain useful information during sequential inference. Specifically, it measures online utility, hold-out generalization, backward transfer, and forgetting, providing a finer-grained view of memory quality. Through extensive experiments across diverse tasks and memory methods, we show that higher final or cumulative accuracy does not necessarily imply better memory quality: many methods exhibit strong performance gains while suffering from substantial forgetting or negative transfer. Moreover, different memory designs exhibit distinct trade-offs between adaptability and stability that remain invisible under standard evaluation metrics.\n\n---\n\n## 171. ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing\n- **Authors**: Blaž Rolih , Matic Fučka , Filip Wolf , Luka Čehovin Zajc\n- **URL**: [https://arxiv.org/abs/2605.15375](https://arxiv.org/abs/2605.15375)\n- **Abstract**:\n> Remote sensing change detection (RSCD) aims to localise changes between two images of the same geographic region. In practice, change masks often follow region-level annotation conventions rather than purely local appearance differences, making them context-dependent and occasionally ambiguous. Most state-of-the-art methods utilise per-pixel discriminative classification, which produces a single prediction per input and fails to explicitly model the changed region as a coherent whole. A natural alternative is generative formulation, which can model a distribution of plausible masks, enabling sampling to capture ambiguity and encourage global consistency. However, existing generative RSCD approaches typically lag behind strong discriminative baselines due to the high computational cost of pixel-space generation and the complexity of their conditioning mechanisms. To address the limitations of prior discriminative and generative methods, we propose ChangeFlow, a generative framework that reformulates change detection as the synthesis of a change mask in latent space via rectified flow. ChangeFlow is guided by a structured yet lightweight conditioning signal, and its stochastic design naturally supports sampling-based prediction ensembling. Namely, aggregating multiple predicted change masks improves robustness, while sample agreement provides a practical confidence estimation that highlights ambiguous regions. Across four benchmarks, ChangeFlow achieves an average F1 of 80.4\\%, improving by 1.3 points on average over the previous best method, while maintaining inference speed comparable to recent strong baselines. Project page: this https URL\n\n---\n\n## 172. PACER: Acyclic Causal Discovery from Large-Scale Interventional Data\n- **Authors**: Ramon Viñas Torné , Sílvia Fàbregas Salazar , Soyon Park , Ivo Alexander Ban , Artyom Gadetsky , Nikita Doikov , Maria Brbić\n- **URL**: [https://arxiv.org/abs/2605.15353](https://arxiv.org/abs/2605.15353)\n- **Abstract**:\n> Inferring the structure of directed acyclic graphs (DAGs) from data is a central challenge in causal discovery, particularly in modern high-dimensional settings where large-scale interventional data are increasingly available. While interventional data can improve identifiability, existing methods remain limited by soft acyclicity constraints, leading to optimization over invalid cyclic graphs, numerical instability, and reduced scalability. We introduce PACER (Perturbation-driven Acyclic Causal Edge Recovery), a scalable framework for causal discovery that guarantees acyclicity by construction. PACER parameterizes a distribution over DAGs through a joint model of variable permutations and edge probabilities, enabling direct optimization over valid causal structures without surrogate penalties. The framework supports a unified likelihood-based treatment of observational and interventional data, flexible conditional density models, and the incorporation of structural prior knowledge. For linear-Gaussian mechanisms, we derive closed-form expressions for the expected interventional log-likelihood and its gradients, yielding substantial computational gains. Empirically, PACER matches or exceeds state-of-the-art methods on protein signaling and large-scale genetic perturbation benchmarks, while scaling efficiently to networks with thousands of variables and achieving up to two orders of magnitude speedups over penalty-based differentiable approaches. These results demonstrate that exact and scalable causal discovery from high-dimensional perturbation data is achievable through principled search space design.\n\n---\n\n## 173. LEAP: Trajectory-Level Evaluation of LLMs in Iterative Scientific Design\n- **Authors**: Marilyn Zhang , Tianfeng Chen , Fabián Barzuna , Ankita Rathod , Mark E. Whiting\n- **URL**: [https://arxiv.org/abs/2605.15341](https://arxiv.org/abs/2605.15341)\n- **Abstract**:\n> LLMs are increasingly deployed in autonomous laboratories, under the assumption that their domain priors and reasoning over iterative feedback let them converge on good designs in fewer iterations than feedback-only baselines. Current iterative scientific design benchmarks, however, score only outcome snapshots at fixed horizons. This leaves the learning trajectory unmeasured, even though the trajectory is what captures learning efficiency, where each iteration saved is a real saving in cost and time. Motivated by this, we examine three evaluation choices that change the conclusions one draws about LLM learning efficiency in iterative scientific design: what to measure, what baseline to compare against, and what to ground against. We introduce LEAPBench, Learning Efficiency in Adaptive Processes, a 55-task framework that pairs a best-so-far area under the curve (AUC) trajectory metric with a classical Bayesian-optimization reference and an audit grounded in published literature. Applied to eight contemporary LLMs, switching from final-outcome to trajectory scoring changes the best-model decision on 53% of tasks at matched horizons, and exposes efficiency gains overlooked by outcome-based scoring. LLMs do not outperform a classical Bayesian baseline. On 16 biology tasks where the oracle's reward signal is aligned with configurations from the published-best design, domain-aware prompting leads to LLM choices that match the published-best's approximately 10 percentage points less often than domain-agnostic prompting at iteration 30. The pattern is sharpest on 6 tasks where the literature-typical and published-best configurations diverge, and domain-agnostic prompting matches the published-best more often on all 6. The trajectory metric also doubles as a tractable training target. Offline reinforcement learning with the metric as a reward improves performance on 14 of 21 held-out tasks.\n\n---\n\n## 174. Hidden in Memory: Sleeper Memory Poisoning in LLM Agents\n- **Authors**: Sidharth Pulipaka , Stanislau Hlebik , Leonidas Raghav , Sahar Abdelnabi , Vyas Raina , Ivaxi Sheth , Mario Fritz\n- **URL**: [https://arxiv.org/abs/2605.15338](https://arxiv.org/abs/2605.15338)\n- **Abstract**:\n> Large language models are increasingly augmented with persistent memory, allowing assistants to store user-specific information across sessions for personalization and continuity. This statefulness introduces a new security risk: adversarial content can corrupt what an assistant remembers and thereby influence future interactions. We propose and study sleeper memory poisoning, a delayed attack in which an adversary manipulates external context, such as a document, webpage, or repository, to cause the assistant to store a fabricated memory about the user. Unlike conventional prompt injection, the attack can remain dormant and re-emerge across multiple later conversations. We evaluate the full attack pipeline: whether poisoned memories are written, later retrieved, and ultimately used to steer the following conversations. Across stateful LLM assistants, poisoned memories were added up to 99.8% on GPT-5.5 and 95% on Kimi-K2.6. Crucially, among successful retrievals, poisoned memories cause attacker-intended agentic actions in 60-89% of evaluations across models. These results show that persistent memory can act as a long-term attack surface across multiple future conversations.\n\n---\n\n## 175. HoloMotion-1 Technical Report\n- **Authors**: Maiyue Chen , Kaihui Wang , Bo Zhang , Xihan Ma , Zhiyuan Yang , Yi Ren , Qijun Huang , Zihao Zhu , Yucheng Wang , Zhizhong Su\n- **URL**: [https://arxiv.org/abs/2605.15336](https://arxiv.org/abs/2605.15336)\n- **Abstract**:\n> In this report, we present HoloMotion-1, a humanoid motion foundation model for zero-shot whole-body motion tracking. A key innovation of HoloMotion-1 is to scale control-policy training with a large-scale hybrid motion corpus, where video-reconstructed motions from in-the-wild videos provide the dominant source of motion diversity, while curated motion-capture and in-house motion data provide higher-fidelity supervision and deployment-oriented coverage. This data regime enables HoloMotion-1 to move beyond conventional MoCap-only training and exposes the policy to substantially broader behaviors, capture conditions, and motion styles. Learning from such heterogeneous data introduces new challenges, including reconstruction noise, source-domain mismatch, uneven motion quality, and the need for temporal modeling under large behavioral variation. To address these challenges, HoloMotion-1 integrates large-capacity temporal modeling, a sparsely activated Mixture-of-Experts Transformer with KV-cache inference for real-time control, and a sequence-level training strategy that improves learning efficiency on extended motion sequences. Extensive experiments on multiple unseen motion benchmarks show that HoloMotion-1 generalizes robustly across diverse motion types and capture conditions, significantly improves tracking accuracy over prior methods, and transfers directly to a real humanoid robot without task-specific fine-tuning.\n\n---\n\n## 176. From I/O to Code with Discovery Agent\n- **Authors**: Yihong Dong , Jiaru Qian , Haoran Zhang , Peixu Wang , Binhua Li , Zhi Jin , Yongbin Li , Ge Li , Xiaokang Yang , Xue Jiang\n- **URL**: [https://arxiv.org/abs/2605.15334](https://arxiv.org/abs/2605.15334)\n- **Abstract**:\n> The automatic synthesis of a program from any form of specification is regarded as a holy grail of computer science. Fueled by LLMs, NL2Code has achieved tremendous success, yet the fundamentally more challenging task of synthesizing programs from input-output behavior, which we refer to as IO2Code, remains largely unsolved. Whereas NL2Code can exploit the semantic alignment between natural language and code acquired during pretraining, IO2Code requires recovering underlying principles from concrete computational behavior, navigating a vast and underspecified hypothesis space. To address this, we propose DIO-Agent, a discovery agent for IO2Code. Our method frames IO2Code as an evolutionary search over discrete program space, in which an LLM serves as the mutation operator and concrete error signals from execution guide each mutation. To prevent the search from wandering into structurally complex yet incorrect dead ends, we introduce the Transformation Priority Premise as a mutation prior that biases the LLM toward the simplest hypothesis consistent with current evidence, progressively escalating from constants to conditionals to iteration only when simpler constructs are insufficient. To facilitate systematic study, we further construct an IO2CodeBench spanning multiple difficulty levels. Extensive experiments show that DIO-Agent consistently outperforms both traditional program-by-example method and SOTA evolution-agent baselines across all difficulty levels and various LLMs, while substantially surpassing test-time scaling strategies with equivalent sampling budgets.\n\n---\n\n## 177. Fortress: A Case Study in Stabilizing Search Recommendations via Temporal Data Augmentation and Feature Pruning\n- **Authors**: Milind Pandurang Jagre , Jia Huang , Dayvid V. R. Oliveira , Zhinan Cheng , Babak Seyed Aghazadeh , Puja Das , Chris Alvino , Jinda Han , Kailash Thiyagarajan\n- **URL**: [https://arxiv.org/abs/2605.15299](https://arxiv.org/abs/2605.15299)\n- **Abstract**:\n> In search and recommendation systems, predictive models often suffer from temporal instability when certain input features introduce volatility in output scores. This instability can degrade model reliability and user experience especially in multi-stage systems where consistent predictions are critical for downstream decision making. We introduce Fortress, a general framework for enhancing model stability and accuracy by identifying and pruning features that contribute to inconsistent prediction scores over time. Fortress leverages historical snapshots temporally partitioned datasets capturing score fluctuations for the same entity across periods and follows a four-step process: (1) collect historical snapshots, (2) identify samples with unstable predictions, (3) isolate and remove instability-inducing features, and (4) retrain models using only stable features. While semantic features from LLMs and BERT-based models improve generalization, they often lack full query or entity coverage. Engagement-based features offer strong predictive power but tend to introduce temporal instability. Fortress mitigates this trade-off by suppressing the volatility of engagement signals while retaining their predictive value leading to more stable and accurate models. We validate Fortress on a query-to-app relevance model in a large-scale app marketplace. Offline experiments demonstrate notable improvements in prediction stability (measured by Coefficient of Variation) and classification performance (measured by PR-AUC).\n\n---\n\n## 178. PhysBrain 1.0 Technical Report\n- **Authors**: Shijie Lian , Bin Yu , Xiaopeng Lin , Changti Wu , Hang Yuan , Xiaolin Hu , Zhaolong Shen , Yuzhuo Miao , Haishan Liu , Yuxuan Tian , Yukun Shi , Cong Huang , Kai Chen\n- **URL**: [https://arxiv.org/abs/2605.15298](https://arxiv.org/abs/2605.15298)\n- **Abstract**:\n> Vision-language-action models have advanced rapidly, but robot trajectories alone provide limited coverage for learning broad physical understanding. PhysBrain 1.0 studies a complementary route: converting large-scale human egocentric video into structured physical commonsense supervision before robot adaptation. Our data engine extracts scene elements, spatial dynamics, action execution, and depth-aware relations, then turns them into question-answer supervision for training PhysBrain VLMs. The resulting physical priors are further transferred to VLA policies through a capability-preserving and language-sensitive adaptation design. Across multimodal QA benchmarks and embodied control benchmarks, including ERQA, PhysBench, SimplerEnv-WidowX, LIBERO, and RoboCasa, PhysBrain 1.0 achieves SOTA results and shows especially strong out-of-domain performance on SimplerEnv. These results suggest that scaling physical commonsense from human interaction video can provide an effective bridge from multimodal understanding to robot action.\n\n---\n\n## 179. GESD: Beyond Outcome-Oriented Fairness\n- **Authors**: Gideon Popoola , John Sheppard\n- **URL**: [https://arxiv.org/abs/2605.15295](https://arxiv.org/abs/2605.15295)\n- **Abstract**:\n> Machine learning (ML) algorithms are increasingly deployed in high-stakes decision-making domains such as loan approvals, hiring, and recidivism predictions. While existing fairness metrics (e.g., statistical parity, equal opportunity) effectively quantify outcome-oriented disparities, they offer limited insight into the procedure or explanation behind biased decisions. To address this gap, we propose Group-level Explanation Stability Disparity (GESD), a \\textit{procedural-oriented} fairness metric that measures disparities in the stability, robustness, and sensitivity of model explanations across different subgroups in a protected category. %GESD is explainer-agnostic, model-agnostic, and extends the scope of fairness analyses to the level of explainability. We further integrate GESD into a multi-objective optimization framework that jointly optimizes for utility, outcome-based fairness, and explanation-based fairness called FEU (Fairness--Explainability--Utility). Empirical results on multiple benchmark datasets show that GESD effectively captures group-wise discrepancies in explanation quality, and that FEU improves both utility and fairness over state-of-the-art methods. By bridging outcome-based and explanation-based fairness, GESD offers a comprehensive tool for diagnosing and mitigating bias in predictive modeling. Our code and datasets are available on GitHub {\\hyperlink{ this https URL }{ this https URL } }\n\n---\n\n## 180. GQA-μP: The maximal parameterization update for grouped query attention\n- **Authors**: Kyle R. Chickering , Huijuan Wang , Mengxi Wu , Alexander Moreno , Muhao Chen , Xuezhe Ma , Daria Soboleva , Joel Hestness , Zhengzhong Liu , Eric Xing\n- **URL**: [https://arxiv.org/abs/2605.15290](https://arxiv.org/abs/2605.15290)\n- **Abstract**:\n> Hyperparameter transfer across model architectures dramatically reduces the amount of compute necessary for tuning large language models (LLMs). The maximal update parameterization ({\\mu}P) ensures transfer through principled mathematical analysis but can be challenging to derive for new model architectures. Building on the spectral feature-learning view of Yang et al. (2023a), we make two advances. First, we promote spectral norm conditions on the weights from a heuristic to the definition of feature learning, and as a consequence arrive at the Complete-P depth and weight-decay scalings without recourse to lazy-learning. Second, we consider a modified spectral norm that preserves the valid scaling law of network weights when weight matrices are not full rank. This enables (to our knowledge, the first) derivation of {\\mu}P scalings for grouped-query attention (GQA). We demonstrate the efficacy of our theoretical derivations by showing learning rate transfer across the GQA repetition hyperparameter as well as experiments regarding transfer over weight decay.\n\n---\n\n## 181. Universal Approximation of Nonlinear Operators and Their Derivatives\n- **Authors**: Filippo de Feo\n- **URL**: [https://arxiv.org/abs/2605.15285](https://arxiv.org/abs/2605.15285)\n- **Abstract**:\n> Derivative-Informed Operator Learning (DIOL), i.e. learning a (nonlinear) operator and its derivatives, is an open research frontier at the foundations of the influential field of Operator Learning (OL). In particular, Universal Approximation Theorems (UATs) of nonlinear operators and their derivatives are foundational open questions and delicate problems in nonlinear functional analysis. In this manuscript, we prove the first UATs of non-linear $k$-times differentiable operators between Banach spaces and their derivatives, uniformly on compact sets and in weighted Sobolev norms for general finite input measures, via OL architectures. Our results are the first complete generalizations of the corresponding influential classical results in [Hornik, 1991] to infinite-dimensional settings and OL. We discuss several open areas where DIOL and our UATs find applications: high-order accuracy in OL, fast constrained optimization in Banach spaces (e.g. optimal control of PDEs, inverse problems) and numerical methods for infinite-dimensional PDEs (e.g. HJB PDEs on Banach spaces from optimal control of PDEs, SPDEs, path-dependent systems, partially observed systems, mean-field control). We parameterize nonlinear operators via Encoder-Decoder Architectures, renowned classes in OL due to their generality, including classical architectures, such as DeepONets, Deep-H-ONets, PCA-Nets. Our results are based on four key features that allow us to prove UATs in full generality: (i) Approximation Properties of Banach spaces. (ii) $k$-times continuous differentiability in the sense of Bastiani (weaker than $k$-times continuous Fréchet differentiability). (iii) Natural compact-open topologies for UA; indeed, we show that UA in standard compact-open topologies induced by operator norms is violated even for Fréchet derivatives. (iv) Construction of novel weighted Sobolev spaces for the UA.\n\n---\n\n## 182. Autonomous Intelligent Agents for Natural-Language-Driven Web Execution with Integrated Security Assurance\n- **Authors**: Vinil Pasupuleti , Siva Rama Krishna Varma Bayyavarapu , Shrey Tyagi\n- **URL**: [https://arxiv.org/abs/2605.15281](https://arxiv.org/abs/2605.15281)\n- **Abstract**:\n> Modern web test suites rot. A UI refactor breaks locators, a timing change causes race conditions, and within weeks developers abandon the suite entirely. This paper presents an AI-driven autonomous testing framework that addresses these failure modes through five integrated strategies - navigation reliability, context-aware selector generation, post-generation validation, smart wait injection, and failure learning - implemented over a containerised worker architecture that decouples orchestration from long-running browser execution. Evaluated across four production applications and 176 scenarios, the framework improves script generation success from 55% to 93%, achieves an 8x reduction in navigation failures, eliminates 80% of timing-related race conditions, and reduces test creation time by 75% compared to manual Selenium authoring. The framework extends naturally to security validation: testers describe attack scenarios in plain English - \"try accessing another user's invoice\" - which the agent converts to OWASP Top 10-aligned browser probes, detecting 85% of authentication bypass vulnerabilities and 95% of input validation flaws with false positive rates below 12%. Natural-language-driven security testing of this kind represents, to our knowledge, a novel contribution to the field.\n\n---\n\n## 183. PDRNN: Modular Data-driven Pedestrian Dead Reckoning on Loosely Coupled Radio- and Inertial-Signalstreams\n- **Authors**: Peter Bauer , Andreas Porada , Felix Ott , Christopher Mutschler , Tobias Feigl\n- **URL**: [https://arxiv.org/abs/2605.15252](https://arxiv.org/abs/2605.15252)\n- **Abstract**:\n> Modern pedestrian dead reckoning (PDR) systems rely on fusing noisy and biased estimates of position, velocity, and calibrated orientation derived from loosely coupled sensors to determine the current pose of a localized object. However, discrepancies in the sampling rates of sensor-specific estimation methods and unreliable transmission pose significant challenges. And traditional methods often fail to effectively fuse multimodal sensor data during dynamic movements characterized by high accelerations, velocities, and rapidly varying orientations. To address these limitations, we propose a simple recurrent neural network (RNN) architecture capable of implicitly forecasting asynchronous sensor data streams from diverse estimation methods along reference trajectories. The proposed approach introduces PDRNN, a modular hybrid AI-assisted PDR system that handles each component as an independent ensemble of machine learning (ML) models to estimate both key parameter means and variances. Separate ML-based models are employed to estimate orientation, (un)directed velocity or distance from acceleration and gyroscope data, with optional absolute positioning from synchronized radio systems such as 5G for stabilization. A final fusion model combines these outputs, position, velocity, and orientation, while using uncertainty estimates to enhance system robustness. The modular design allows individual components to be updated, fine-tuned, or replaced without affecting the entire system. Experiments on dynamic sports movement data show that PDRNN achieves superior accuracy and precision compared to classic and ML-based methods, effectively avoiding error accumulation common in black-box approaches. And PDRNN offers forecast capabilities and better component control despite increased system complexity.\n\n---\n\n## 184. GQLA: Group-Query Latent Attention for Hardware-Adaptive Large Language Model Decoding\n- **Authors**: Fanxu Meng\n- **URL**: [https://arxiv.org/abs/2605.15250](https://arxiv.org/abs/2605.15250)\n- **Abstract**:\n> Multi-head Latent Attention (MLA), the attention used in DeepSeek-V2/V3, jointly compresses keys and values into a low-rank latent and matches the H100 roofline almost perfectly. Its trained weights, however, expose only one decoding path - an absorbed MQA form - which ties efficient inference to H100-class compute-bandwidth ratios, forfeits tensor parallelism along the head axis, and yields no Multi-Token Prediction (MTP) gain on commodity inference GPUs such as the export-restricted H20. We propose Group-Query Latent Attention (GQLA), a minimal modification of MLA whose trained weights expose two algebraically equivalent decoding paths over the same parameters: an MQA-absorb path identical to MLA's, and a GQA path with a per-group expanded cache. The runtime picks the path that matches the target hardware - no retraining, no custom kernels - so a single set of GQLA weights pins the rooflines of both H100 (MQA-absorb, s_q=1) and H20 (GQA + MTP, s_q=2), while supporting up to 8-way zero-redundancy tensor parallelism on the GQA path. To avoid pretraining from scratch we extend TransMLA into TransGQLA, which converts a pretrained GQA checkpoint into a GQLA model; on LLaMA-3-8B it compresses the per-token KV cache to 28.125% of the GQA baseline on the MQA-absorb path while structurally preserving GQA-level traffic on the per-group path.\n\n---\n\n## 185. Reading the Cell, Designing the Cure: Perturbation-Conditioned Molecular Diffusion for Function-Oriented Drug Design\n- **Authors**: Ziyu Xu , Zijian Zhang , Liang Wang , Zhiyuan Liu , Qiang Liu , Shu Wu , Liang Wang\n- **URL**: [https://arxiv.org/abs/2605.15243](https://arxiv.org/abs/2605.15243)\n- **Abstract**:\n> When reliable target structures are unavailable at scale or phenotypes arise from dysregulated pathways, transcriptomic perturbations provide a system-level functional readout for drug action. In this work, we formalize \\emph{Transcriptome-based Drug Design (TBDD)} as a generative inverse problem: designing drug molecules conditioned on desired transcriptomic state transitions. We analyze the inherently ill-posed nature of this task, which is further complicated by the profound domain gap between biology and chemistry and by the sparsity of transcriptomic signals. To address these challenges, we propose \\textbf{\\themodel{} } (A \\textbf{C}ell\\textbf{U}lar \\textbf{R}esponse \\textbf{E}ngine), a multi-resolution transcriptome-guided diffusion framework. \\themodel{} features a specialized \\textbf{Transcriptome Perturbation Functional Feature Extractor (TFE)} that (1) distills function-oriented perturbation embeddings from pre/post states, (2) aligns these signatures to dual chemical views to bridge the cross-modal gap, and (3) performs heterogeneity-aware aggregation to extract robust state-specific signals from noisy transcriptomic data. Extensive evaluations on both standard benchmarks and rigorous out-of-distribution protocols demonstrate that \\themodel{} consistently outperforms strong baselines in structural quality and functional consistency. Furthermore, we validate its practical utility via a zero-shot gene-inhibitor design task, highlighting the potential of phenotype-driven generative discovery.\n\n---\n\n## 186. Hydra: Efficient, Correct Code Generation via Checkpoint-and-Rollback Support\n- **Authors**: Alexander Du , Jianjun Ou , Danyang Zhuo , Matthew Lentz\n- **URL**: [https://arxiv.org/abs/2605.15238](https://arxiv.org/abs/2605.15238)\n- **Abstract**:\n> Large language models are increasingly used for code generation, but many generated programs fail to compile, a prerequisite for further correctness checks such as unit tests. Existing solutions for repairing static errors are costly in both latency and token consumption. Post-hoc repair delays error detection until generation completes and commonly regenerates large regions of previously valid code. Constrained semantic decoding checks after each token, incurring per-token overhead while limiting repair to the current token even when the root cause lies earlier. We present Hydra, a system for efficient recovery from static errors during code generation. Hydra allows checking to proceed asynchronously with generation, avoiding checker overhead when the generated code is semantically correct. In addition, it provides checkpoint-and-rollback support for targeted repair, avoiding regeneration and rechecking of valid prefixes. We retrofit the Clang C/C++ compiler to support Hydra with modest modifications. Paired with a token-efficient repair strategy, Hydra reduces latency by up to 71% and token consumption by up to 70% relative to post-hoc repair on C/C++ code generation tasks that encounter static errors.\n\n---\n\n## 187. A3D: Agentic AI flow for autonomous Accelerator Design\n- **Authors**: Abinand Nallathambi , Christopher Knight , Shantanu Ganguly , Wilfried Haensch , Anand Raghunathan\n- **URL**: [https://arxiv.org/abs/2605.15237](https://arxiv.org/abs/2605.15237)\n- **Abstract**:\n> Accelerating applications through the design of hardware accelerators can significantly enhance system performance and energy efficiency. Despite advances, such as high-level synthesis (HLS), designing accelerators for complex applications still remains highly labor-intensive, demanding considerable expertise in understanding workloads to be accelerated, hardware design, micro-architecture, and EDA tool usage, posing challenges for application domain experts. Therefore, most accelerator solutions are limited to applications with a regular predictable dataflow. Advances in AI have enabled agents that perform autonomous planning, reasoning, execution and reflection, leading to unprecedented potential for automation through agentic AI. We present A3D, an Agentic AI flow for end-to-end Automation of hardware Accelerator Design. A3D automates workload analysis, performance bottleneck identification, code refactoring for HLS compatibility and micro-architecture generation. A3D also generates diverse accelerator designs by automatically exploring the speed-area tradeoff space. Recent efforts have explored the use of AI for specific tasks such as design space exploration in HLS, leaving several tasks to still be performed manually. A3D addresses the challenges in applying modern LLMs to accelerator design by judiciously partitioning tasks among specialist agents, orchestrating process loops with specialist and verifier agents, utilizing pre-existing and custom tools, and employing agentic RAG for codebase and proprietary EDA tool documentation exploration. Our implementation of A3D, using commercial components like Claude Sonnet 4.5 and the Catapult HLS tool, demonstrates its effectiveness by generating accelerator designs with no human intervention from complex scientific applications like LAMMPS (molecular dynamics simulation) and QMCPACK (quantum chemistry).\n\n---\n\n## 188. Learning Selective Merge Policies for Deadline-Constrained Coded Caching via Deep Reinforcement Learning\n- **Authors**: Amirhossein Yousefiramandi\n- **URL**: [https://arxiv.org/abs/2605.15236](https://arxiv.org/abs/2605.15236)\n- **Abstract**:\n> With the coded caching, the server can use the information the users have cached to serve multiple users at a time by sending a single coded multi-casting message, i.e., the merged message, thereby relieving the peak network loads. However, for the delay-sensitive applications of the users, like the video streaming services, it becomes essential to choose which messages to merge online, considering the strict deadlines for each request. The problem, however, is that while the merge is helpful for the formation of the current coded multi-casting message, it can be harmful for the subsequent ones. We proposed a DRL-based solution that formulates the deadline-constrained coded delivery as a masked discrete-action queue-state control problem, while we trained a graph-attention policy network via proximal policy optimization. The policy network reduces the broadcast-packet expiration ratio $\\rho$ by $40.9%$ ($0.208$ vs. $0.352$) with respect to the best coded multi-casting baseline (SACM++) on the uniform-demand benchmark, while also attaining the best broadcast-efficiency score $\\sigma$ across the Track A battery among the coded multi-casting methods. The interesting fact we observed is that for the applications of the users with tight deadlines, the method of selective merging is better than the method of aggressive merging, i.e., the policy network learns to merge at only $\\approx 31.8%$ rate, even though the same observation holds across the variations within the same simulator family.\n\n---\n\n## 189. PBT-Bench: Benchmarking AI Agents on Property-Based Testing\n- **Authors**: Lucas Jing , Xinqi Wang , Liao Zhang , Simon S. Du\n- **URL**: [https://arxiv.org/abs/2605.15229](https://arxiv.org/abs/2605.15229)\n- **Abstract**:\n> Existing code benchmarks measure whether an agent can produce any test that reproduces a known bug, or whether it can produce a patch that fixes a described issue. Neither isolates the distinct skill of property-based testing: deriving a semantic invariant from documentation, and then constructing an input-generation strategy precise enough to make a random search reveal the violation. We introduce PBT-Bench, a benchmark of 100 curated property-based testing problems across 40 real Python libraries. Each problem injects one or more semantic bugs (365 in total, mean 3.65 per problem) designed so that default-strategy random inputs almost never trigger them; the agent must read the library's documentation, identify the relevant invariant, and specify a Hypothesis @given strategy that concentrates mass in the trigger region. Bugs are stratified across three difficulty levels (L1-L3) spanning single-constraint boundary bugs to stateful, cross-function protocol violations. We evaluate eight contemporary LLMs under two prompting regimes (open-ended baseline vs. explicit Hypothesis scaffolding) for three independent runs per configuration. Bug recall under the PBT-guided prompt ranges from 42.1% to 83.4% across models; under the open-ended baseline, from 31.4% to 76.7%. Hypothesis scaffolding lifts mid-capability models by over 20 percentage points, but yields smaller gains for the strongest models, with two exceptions showing degradation, suggesting the structured prompt can interfere with certain model behaviours rather than complementing them. The hardest bugs prove model-specific: different architectures fail on different problems, leaving persistent gaps that no single model closes. We release the benchmark, harness, and full evaluation corpus to support downstream work on documentation-grounded semantic reasoning.\n\n---\n\n## 190. Is Agentic AI Ready for Real-World Hardware Engineering? A Deep Dive with Phoenix-bench\n- **Authors**: Qingyun Zou , Feng Yu , Hongshi Tan , Bingsheng He , WengFai Wong\n- **URL**: [https://arxiv.org/abs/2605.15226](https://arxiv.org/abs/2605.15226)\n- **Abstract**:\n> We ask whether agentic AI systems built for software engineering transfer to realistic hardware engineering. Existing hardware LLM benchmarks isolate sub-tasks but none jointly requires repository navigation, hierarchy-aware localization, Electronic Design Automation (EDA) executable verification, and maintenance-style patching. We introduce \\textbf{Phoenix-bench}, a synchronized corpus of 511 verified Verilator instances from 114 GitHub repositories, each shipped with the developer patch, design-flow labels, fail-to-pass and pass-to-pass testbenches, and a Docker-pinned EDA environment so resolved-rate differences reflect agent behavior rather than toolchain availability. Using Phoenix-bench we run a uniform evaluation of four commercial agents and eight open-source agentic structures across four LLM backbones, plus two diagnostic interventions (file-level oracle localization and one round of testbench-log feedback). Three findings emerge. (i)~Software and hardware are fundamentally different engineering tasks: the same agent loses 37\\% to 58\\% from SWE-bench Verified to Phoenix-bench because hardware bugs propagate across parallel instantiated modules through signal flow rather than along a software-style call graph, and software-tuned agents stop at the symptom file instead of tracing back through the instantiation chain. (ii)~Failures concentrate on design control-flow / finite state machine (FSM) bugs, verification testbench bugs, and hard cases that demand cross-hierarchy signal-flow tracking and coordinated multi-file edits. (iii)~Localization granularity matters far more than localization itself: a perfect file-level oracle yields only $+1.4$\\% because the agent then breaks files that did not need editing, while a single round of test case feedback lifts resolved rate by $42$\\% to $45$\\% because the test case tells \\emph{where} the bug is and \\emph{what} the fix has to look like.\n\n---\n\n## 191. Do Biological Structural Guarantees Earn Their Complexity?\n- **Authors**: Bogdan Banu\n- **URL**: [https://arxiv.org/abs/2605.15225](https://arxiv.org/abs/2605.15225)\n- **Abstract**:\n> Biologically-inspired AI agent frameworks claim reliability benefits through structural guarantees adapted from gene regulatory networks, immune systems, and metabolic control. These claims are rarely tested empirically against simpler alternatives. We present three deep benchmarks: metabolic priority gating, autoinducer-based quorum sensing, and Bayesian stagnation detection, each comparing a biologically-grounded implementation against a naive non-biological alternative and an ablated control, across 1,000 trials per seed and 10 seeds (10M+ data points total).\n\n---\n\n## 192. GenAI-Driven Approach to RISC-V Supply Chain Exploration\n- **Authors**: Nenad Petrovic , Andre Schamschurko , Yingjie Xu , Alois Knoll\n- **URL**: [https://arxiv.org/abs/2605.15223](https://arxiv.org/abs/2605.15223)\n- **Abstract**:\n> This paper presents an LLM-empowered workflow for RISC-V supply chain analysis, integrating Vision-Language Models (VLMs) and Model-Driven Engineering (MDE) to enable comprehensive, multimodal data-driven insights. The proposed approach addresses the challenges of heterogeneous and unstructured supply chain data by leveraging LLMs for textual understanding and VLMs for extracting information from visual artifacts such as diagrams, tables, and scanned documents. These models collaboratively identify key entities and relationships, which are then organized into a knowledge graph representing supply chain components and their interdependencies. For analytical reasoning, the workflow incorporates MDE techniques and constraint-based modeling to enable formal validation of dependencies, detection of bottlenecks, and assessment of risks. The synergy between LLM- and VLM-based semantic understanding and MDE-based formal analysis supports both exploratory and systematic evaluation of supply chain resilience. A human-in-the-loop mechanism further enables interactive querying and expert validation. The approach is evaluated in RISC-V ecosystem scenarios, demonstrating its effectiveness in generating actionable insights, enhancing transparency, and supporting decision-making in complex semiconductor supply chains.\n\n---\n\n## 193. Effective Harness Engineering for Algorithm Discovery with Coding Agents\n- **Authors**: Yoichi Ishibashi , Taro Yano , Masafumi Oyamada\n- **URL**: [https://arxiv.org/abs/2605.15221](https://arxiv.org/abs/2605.15221)\n- **Abstract**:\n> AlphaEvolve and FunSearch have demonstrated the potential of combining large language models (LLMs) with evolutionary search for automated algorithm discovery. However, discovery success is shaped not only by model capability but also significantly by the design of the execution infrastructure, i.e., the harness. This paper investigates effective harness design through three questions: under a fixed token budget, is it better to produce many algorithms with brief thought or fewer algorithms with deeper thought? How should the harness handle evaluation hacks, where generated programs exploit the scoring function? And how can agents that require full filesystem access execute safely in parallel? Using Vesper, an algorithm discovery framework that incorporates harness improvements addressing these questions, we evaluate on Circle Packing under the same token budget. Interestingly, generating fewer algorithms while thinking more deeply about each one achieved higher scores. That is, scaling the quality of each individual is more budget-efficient than scaling the number of evolutionary generations. Surprisingly, more capable models produced evaluation hacks at higher rates, making hack detection increasingly necessary as models scale.\n\n---\n\n## 194. Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time\n- **Authors**: Michael Y. Hu , Apurva Gandhi , Kyunghyun Cho , Tal Linzen , Pratyusha Sharma\n- **URL**: [https://arxiv.org/abs/2605.15220](https://arxiv.org/abs/2605.15220)\n- **Abstract**:\n> Data mixing decides how to combine different sources or types of data and is a consequential problem throughout language model training. In pretraining, data composition is a key determinant of model quality; in continual learning and adaptation, it governs what is retained and acquired. Yet existing data mixing methods address only one phase of this lifecycle at a time: some require smaller proxy models tied to a single training phase, others assume a fixed domain set, and continual learning lacks principled guidance altogether. We argue that data mixing is fundamentally an online decision making problem -- one that recurs throughout training and demands a single, unified solution. We introduce OP-Mix (On-Policy Mix), a data mixing algorithm that operates across the entire language model training lifecycle. Our main insight is that candidate data mixtures can be cheaply simulated by interpolating between low-rank adapters trained directly on the current model, eliminating separate proxy models and ensuring the search is always grounded in the model's actual learning dynamics. Across pretraining, continual midtraining, and continual instruction tuning, OP-Mix consistently finds near-optimal mixtures while using a fraction of the compute of the baselines. In pretraining, OP-Mix improves upon training without mixing by 6.3% in average perplexity. For continual learning, OP-Mix matches the performance of both retraining and on-policy distillation while using 66% and 95% less overall compute, respectively. OP-Mix suggests a different view of language model training: not a sequence of distinct phases, but a single continuous process of learning from data.\n\n---\n\n## 195. An LLM-RAG Approach for Healthy Eating Index-Informed Personalized Food Recommendations\n- **Authors**: Yibin Wang , Yanjie Yang , Grace Melo Guerrero , Rodolfo M. Nayga Jr. , Azlan Zahid\n- **URL**: [https://arxiv.org/abs/2605.15213](https://arxiv.org/abs/2605.15213)\n- **Abstract**:\n> Diet quality is a leading determinant of chronic disease risk. Advances in artificial intelligence (AI) have enabled food recommendation systems to adapt suggestions to user preferences and health goals. However, most current systems rely on loosely curated food databases and provide limited connection to a validated index. In this study, we propose a Healthy Eating Index (HEI) informed retrieval-augmented generation (RAG) framework that combines standardized nutrition databases with large language models (LLMs) for personalized food recommendations. Our proposed method anchors retrieval in the National Health and Nutrition Examination Survey (NHANES) and the Food Patterns Equivalents Database (FPED). A food-level embedding space is constructed from FPED-derived textual descriptions. For each entity, the system computes baseline HEI scores, retrieves candidate foods for intake recommendations, and estimates the HEI impact of simple substitutions or additions. A constrained RAG pipeline instantiated with a pretrained OpenAI LLM generates personalized recommendations and sources based on nutrient profiles and HEI contributions. The simulation results showed a mean HEI improvement of 6.45, with the proportion of users HEI over 50 increasing from 45.12 to 61.26. Quantile analysis revealed consistent improved shifts across the HEI distribution. Our findings suggest that the proposed LLM-RAG-based AI systems can support more precise, explainable, and personalized nutrition guidance to improve diet quality.\n\n---\n\n## 196. Fault tolerance estimation in digital circuits with visualised generative networks\n- **Authors**: Sascha Biel , Carl Alexander Gaede , Amiel Glaser , Jan Wolter , Alexej Schelle\n- **URL**: [https://arxiv.org/abs/2605.15212](https://arxiv.org/abs/2605.15212)\n- **Abstract**:\n> We propose a new numerical method to estimate the fault tolerance of failure modes in digital circuit structures with a generative network sampling technique. From a random input of generated bitwise configurations of ideally digitalised analog currents in the digital circuit design with classical logical gates, expected output currents are compared to the realistic signals of a numerical experiment at the discriminator part of the Generative Adversarial Network (GAN) to calculate the deviation from ideal digital electronic signals, including various error modes, such as missing or interchanged logical devices. From the present analysis of a representation of the GAN in terms of complex variables, it is possible to evaluate the robustness in electronic designs by differentiating the impact of failure modes associated with different classical logical elements in the circuit.\n\n---\n\n## 197. Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels\n- **Authors**: Plawan Kumar Rath , Rahul Maliakkal\n- **URL**: [https://arxiv.org/abs/2605.15208](https://arxiv.org/abs/2605.15208)\n- **Abstract**:\n> Large Language Models are routinely compressed via post-training quantization to reduce inference costs and memory footprint for cloud and edge deployment, yet the impact of this compression on model quality remains poorly understood. Existing studies typically compare only two conditions (full-precision vs. a single quantized variant), rely on aggregate bias metrics, and evaluate a single model family, making it impossible to distinguish gradual degradation from threshold-dependent safety failures. We conduct a controlled empirical study of three instruction-tuned models (Qwen2.5-7B, Mistral-7B, Phi-3.5-mini) at five precision levels (BF16 through 3-bit) on 12,148 BBQ bias benchmark items across 5 random seeds, totaling 911,100 inference records. Our results reveal that 3-bit quantization causes 6-21% of previously unbiased items to develop new stereotypical behaviors, following a clear dose-response pattern confirmed via logistic regression, while models' willingness to select \"unknown\" answers declines by 17.4%. Crucially, these item-level changes are invisible to standard quality metrics: perplexity increases by less than 0.5% at 8-bit and under 3% at 4-bit across all three models, yet 2.5-5.6% of items already develop new biases at 4-bit. These findings demonstrate that aggregate evaluation metrics systematically miss fairness-critical degradation, underscoring the need for quality-aware compression protocols that explicitly test for bias emergence before deployment.\n\n---\n\n## 198. AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices\n- **Authors**: Dzung Pham , Kleomenis Katevas , Ali Shahin Shamsabadi , Hamed Haddadi\n- **URL**: [https://arxiv.org/abs/2605.15206](https://arxiv.org/abs/2605.15206)\n- **Abstract**:\n> Autonomous agents powered by large language models (LLMs) are increasingly used to automate complex, multi-step tasks such as coding or web-based question answering. While remote, cloud-based agents offer scalability and ease of deployment, they raise privacy concerns, depend on network connectivity, and incur recurring API costs. Deploying agents locally on user devices mitigates these issues by preserving data privacy and eliminating usage-based fees. However, agentic workflows are far more resource-intensive than typical LLM interactions. Iterative reasoning, tool use, and failure retries substantially increase token consumption, often expending significant compute without successfully completing tasks. In this work, we investigate the time, token, and energy overhead of locally deployed LLM-based agents on consumer hardware. Our measurements show that agentic execution increases GPU power draw, temperature, and battery drain compared to single-inference workloads. To address this inefficiency, we introduce AgentStop, a lightweight efficiency supervisor that predicts and preemptively terminates trajectories unlikely to succeed. Leveraging low-cost execution signals, such as token-level log probabilities, AgentStop can reduce wasted energy by 15-20% with minimal impact on task performance (<5% utility drop) for challenging web-based question answering and coding benchmarks. These findings position predictive early termination as a practical mechanism for enabling sustainable, privacy-preserving LLM agents on user devices. Our project code and data are available at this https URL .\n\n---\n\n## 199. Agent4POI: Agentic Context-Conditioned Affordance Reasoning for Multimodal Point-of-Interest Recommendation\n- **Authors**: Jinze Wang , Yangchen Zeng , Tiehua Zhang , Lu Zhang , Yuze Liu , Yongchao Liu , Xingjun Ma , Zhu Sun\n- **URL**: [https://arxiv.org/abs/2605.15203](https://arxiv.org/abs/2605.15203)\n- **Abstract**:\n> We introduce Agent4POI, the first POI recommendation framework that generates context-conditioned multimodal representations at recommendation time, rather than relying on static POI embeddings pre-computed independently of context. Existing multimodal systems encode each POI once as a static embedding, a design that precludes reasoning about why the same cafe affords solo work on Monday but group celebration on Friday evening. We formally prove that no pre-computed encoder can satisfy context-sensitive ranking under standard bilinear scoring, motivating inference-time item-side representation. Agent4POI inverts this computation: given a situational context, a four-phase LLM agent generates dynamic, context-specific affordance queries (Phase 1) and executes a five-step cross-modal chain-of-thought over image, review, and metadata evidence (Phase 2). The resulting uncertainty-aware affordance representation is grounded in Gibsonian affordance theory. These cross-modal verdicts form a structured, uncertainty-adjusted affordance representation (Phase 3), which is aligned with user preferences via a semantic caching system for low-latency ranking (Phase 4). On three POI benchmarks and three evaluation configurations (standard, cold-start, context-shift), Agent4POI achieves a 23.2% relative gain over the strongest baseline and degrades by only 7.5% under context-shift versus 16--17\\% for the strongest baselines. In cold-start scenarios, Agent4POI outperforms the best content-based baseline by up to 2.4x, whereas ID-based methods fail to generalize.\n\n---\n\n## 200. Ensuring Logic in the Fog: Sound POMDP Synthesis with LTL Objectives\n- **Authors**: Can Zhou , Yulong Gao , Pian Yu\n- **URL**: [https://arxiv.org/abs/2605.12581](https://arxiv.org/abs/2605.12581)\n- **Abstract**:\n> Synthesising autonomous agents that can navigate uncertain environments while adhering to complex temporal constraints remains a fundamental challenge. While Linear Temporal Logic (LTL) provides a rigorous language for specifying such tasks, the inherent undecidability of qualitatively verifying LTL satisfaction in partially observable Markov decision processes renders quantitative synthesis difficult, especially when designing reliable reward signals for approximate solvers. In this paper, we bridge this gap with a novel, sound reward-shaping mechanism that dynamically generates belief-dependent rewards grounded in certified LTL satisfaction. By integrating this mechanism into an enhanced Monte Carlo Planning framework, we empower agents to navigate the `fog' of partial observability with a search process focused on maximising verifiable success. Our experiments demonstrate that this approach not only thrives in scenarios where existing solvers fail but also maintains effectiveness and scalability across diverse benchmark domains.\n\n---\n\n{% endraw %}\n"}, "c234ff9921f9be8d378a7558752264c977290646": {"t": 1786068100.771258, "n": 340278, "h": "05c8c40e431f029e", "c": "---\nlayout: default\ntitle: \"전체 AI 논문 - 2026-05-18\"\n---\n\n{% raw %}\n# 전체 AI 논문 - 2026-05-18\n\n## 1. Prospective multi-pathogen disease forecasting using autonomous LLM-guided tree search\n- **Authors**: Sarah Martinson , Michael P. Brenner , Martyna Plomecka , Brian P. Williams , Nicholas G. Reich , Zahra Shamsi\n- **URL**: [https://arxiv.org/abs/2605.16238](https://arxiv.org/abs/2605.16238)\n- **Abstract**:\n> Probabilistic forecasting of infectious diseases is crucial for public health but relies on labor-intensive manual model curation by expert modeling teams. This bespoke development bottlenecks scalability to granular geographic resolutions or emerging pathogens. Here, we present an autonomous system using Large Language Model (LLM)-guided tree search to iteratively generate, evaluate, and optimize executable forecasting software. In a fully prospective, real-time evaluation during the 2025-2026 US respiratory season, the system autonomously discovered methodologically diverse models for influenza, COVID-19, and respiratory syncytial virus (RSV). Aggregating these machine-generated models yielded an ensemble that consistently matched or outperformed the gold-standard, human-curated Centers for Disease Control and Prevention (CDC) hub ensembles out-of-sample. The system successfully navigated data-scarce \"cold start\" scenarios for RSV. Moreover, controlled retrospective ablations revealed that optimizing log-scale distance metrics prevents reward hacking, while an automated judge-in-the-loop ensures structural fidelity to complex scientific theories. By autonomously translating epidemiological theory into accurate, transparent code, this framework overcomes the modeling labor bottleneck, enabling rapid deployment of expert-level disease forecasting at unprecedented scales.\n\n---\n\n## 2. FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast\n- **Authors**: Igor Bogdanov , Chung-Horng Lung , Thomas Kunz , Jie Gao , Adrian Taylor , Marzia Zaman\n- **URL**: [https://arxiv.org/abs/2605.16233](https://arxiv.org/abs/2605.16233)\n- **Abstract**:\n> Can LLM agents improve decision-making through self-generated memory without gradient updates? We propose FORGE (Failure-Optimized Reflective Graduation and Evolution), a staged, population-based protocol that evolves prompt-injected natural-language memory for hierarchical ReAct agents. FORGE wraps a Reflexion-style inner loop, where a dedicated reflection agent (using the same underlying LLM, no distillation from a stronger model) converts failed trajectories into reusable knowledge artifacts: textual heuristics (Rules), few-shot demonstrations (Examples), or both (Mixed), with an outer loop that propagates the best-performing instance's memory to the population between stages and freezes converged instances via a graduation criterion. We evaluate on CybORG CAGE-2, a stochastic network-defense POMDP at a 30-step horizon against the B-line attacker, where all four tested LLM families (Gemini-2.5-Flash-Lite, Grok-4-Fast, Llama-4-Maverick, Qwen3-235B) exhibit strongly negative, heavy-tailed zero-shot rewards. Compared against both a zero-shot baseline and a Reflexion baseline (isolated single-stream learning), FORGE improves average evaluation return by 1.7-7.7$\\times$ over zero-shot and by 29-72% over Reflexion in all 12 model-representation conditions, reducing major-failure rates (below $-100$) to as low as $\\sim$1%. We find that (1) population broadcast is critical mechanism, with a no-graduation ablation confirming that broadcast carries the performance gains while graduation primarily saves compute; (2) Examples achieves the strongest returns for three of four models, Rules offers the best cost-reliability profile with $\\sim$40% fewer tokens; and (3) weaker baseline models benefit disproportionately, suggesting FORGE may mitigate capability gaps rather than amplify strong models. All evidence is confined to CAGE-2 B-line; cross-family findings are directional evidence.\n\n---\n\n## 3. Fully Open Meditron: An Auditable Pipeline for Clinical LLMs\n- **Authors**: Xavier Theimer-Lienhard , Mushtaha El-Amin , Fay Elhassan , Sahaj Vaidya , Victor Cartier-Negadi , David Sasu , Lars Klein , Mary-Anne Hartley\n- **URL**: [https://arxiv.org/abs/2605.16215](https://arxiv.org/abs/2605.16215)\n- **Abstract**:\n> Clinical decision support systems (CDSS) require scrutable, auditable pipelines that enable rigorous, reproducible validation. Yet current LLM-based CDSS remain largely opaque. Most \"open\" models are open-weight only, releasing parameters while withholding the data provenance, curation procedures, and generation pipelines that determine model behavior. Fully Open (FO) models, which expose the complete training stack end-to-end, do not currently exist in medicine. We introduce Fully Open Meditron, the first fully open pipeline for building LLM-CDSS, comprising a clinician-audited training corpus, a reproducible data construction and training framework, and a use-aligned evaluation protocol. The corpus unifies eight public medical QA datasets into a normalized conversational format and expands coverage with three clinician-vetted synthetic extensions: exam-style QA, guideline-grounded QA derived from 46,469 clinical practice guidelines, and clinical vignettes. The pipeline enforces system-wide decontamination, gold-label resampling of teacher generations, and end-to-end validation by a four-physician panel. We evaluate using an LLM-as-a-judge protocol over expert-written clinical vignettes, calibrated against 204 human raters. We apply the recipe to five FO base models (Apertus-70B/8B-Instruct, OLMo-2-32B-SFT, EuroLLM-22B/9B-Instruct). All MeditronFO variants are preferred over their bases. Apertus-70B-MeditronFO improves +6.6 points over its base (47.2% to 53.8%) on aggregate medical benchmarks, establishing a new FO SoTA. Gemma-3-27B-MeditronFO is preferred over MedGemma in 58.6% of LLM-as-a-judge comparisons and outperforms it on HealthBench (58% vs 55.9%). These results show that fully open pipelines can achieve state-of-the-art domain-specific performance without sacrificing auditability or reproducibility.\n\n---\n\n## 4. Confirming Correct, Missing the Rest: LLM Tutoring Agents Struggle Where Feedback Matters Most\n- **Authors**: Tahreem Yasir , Wenbo Li , Sam Gilson , Sutapa Dey Tithi , Xiaoyi Tian , Tiffany Barnes\n- **URL**: [https://arxiv.org/abs/2605.16207](https://arxiv.org/abs/2605.16207)\n- **Abstract**:\n> Effective tutoring requires distinguishing optimal, valid but suboptimal, and incorrect student solutions, a distinction central to intelligent tutoring systems (ITS) but untested for LLM-based tutors. As LLMs are increasingly explored as conversational complements to ITS, evaluating their diagnostic precision is essential. We present a benchmark of seven LLM feedback agents in propositional logic using knowledge-graph-derived ground truth across 10,836 solution--feedback pairs and three feedback conditions. Models achieved near-ceiling performance on optimal steps but systematically over-rejected valid but suboptimal reasoning and over-validated incorrect solutions, precisely where adaptive tutoring matters most. These failures persisted across models regardless of solution context, suggesting architectural rather than informational limits. Moreover, accurate diagnosis did not reliably produce pedagogically actionable feedback, revealing a gap between diagnostic judgment and instructional effectiveness. Our findings suggest that LLMs are better suited for hybrid architectures where KG-grounded models handle diagnosis while LLMs support open-ended scaffolding and dialogue.\n\n---\n\n## 5. Context, Reasoning, and Hierarchy: A Cost-Performance Study of Compound LLM Agent Design in an Adversarial POMDP\n- **Authors**: Igor Bogdanov , Chung-Horng Lung , Thomas Kunz , Jie Gao , Adrian Taylor , Marzia Zaman\n- **URL**: [https://arxiv.org/abs/2605.16205](https://arxiv.org/abs/2605.16205)\n- **Abstract**:\n> Deploying compound LLM agents in adversarial, partially observable sequential environments requires navigating several design dimensions: (1) what the agent sees, (2) how it reasons, and (3) how tasks are decomposed across components. Yet practitioners lack guidance on which design choices improve performance versus merely increase inference costs. We present a controlled study of compound LLM agent design in CybORG CAGE-2, a cyber defense environment modeled as a Partially Observable Markov Decision Process (POMDP). Reward is non-positive, so all configurations operate in a failure-mitigation mode. Our evaluation spans five model families, six models, and twelve configurations (3,475 episodes) with token-level cost accounting. We vary context representation (raw observations vs. a deterministic state-tracking layer with compressed history), deliberation (self-questioning, self-critique, and self-improvement tools, with optional chain-of-thought prompting), and hierarchical decomposition (monolithic ReAct vs. delegation to specialized sub-agents). We find that: (1) Programmatic state abstraction delivers the largest returns per token spent (RPTS), improving mean return by up to 76% over raw observations. (2) Distributing deliberation tools across a hierarchy degrades performance relative to hierarchy alone for all five model families, reaching up to 3.4$\\times$ worse mean return while using 1.8-2.7$\\times$ more tokens. We call this destructive pattern a deliberation cascade. (3) Hierarchical decomposition without deliberation achieves the best absolute performance for most models, and context engineering is generally more cost-effective than deliberation. These findings suggest a design principle for structured adversarial POMDPs: invest in programmatic infrastructure and clean task decomposition rather than deeper per-agent reasoning, as these strategies can interfere when combined.\n\n---\n\n## 6. Formal Methods Meet LLMs: Auditing, Monitoring, and Intervention for Compliance of Advanced AI Systems\n- **Authors**: Parand A. Alamdari , Toryn Q. Klassen , Sheila A. McIlraith\n- **URL**: [https://arxiv.org/abs/2605.16198](https://arxiv.org/abs/2605.16198)\n- **Abstract**:\n> We examine one particular dimension of AI governance: how to monitor and audit AI-enabled products and services throughout the AI development lifecycle, from pre-deployment testing to post-deployment auditing. Combining principles from formal methods with SoTA machine learning, we propose techniques that enable AI-enabled product and service developers, as well as third party AI developers and evaluators, to perform offline auditing and online (runtime) monitoring of product-specific (temporally extended) behavioral constraints such as safety constraints, norms, rules and regulations with respect to black-box advanced AI systems, notably LLMs. We further provide practical techniques for predictive monitoring, such as sampling-based methods, and we introduce intervening monitors that act at runtime to preempt and potentially mitigate predicted violations. Experimental results show that by exploiting the formal syntax and semantics of Linear Temporal Logic (LTL), our proposed auditing and monitoring techniques are superior to LLM baseline methods in detecting violations of temporally extended behavioral constraints; with our approach, even small-model labelers match or exceed frontier LLM judges. Our predictive and intervening monitors significantly reduce the violation rates of LLM-based agents while largely preserving task performance. We further show through controlled experiments that LLMs' temporal reasoning shows a pronounced degradation in accuracy with increasing event distance, number of constraints, and number of propositions.\n\n---\n\n## 7. An Algebraic Exposition of the Theory of Dyadic Morality\n- **Authors**: Kush R. Varshney\n- **URL**: [https://arxiv.org/abs/2605.16153](https://arxiv.org/abs/2605.16153)\n- **Abstract**:\n> This paper provides an algebraic exposition of the theory of dyadic morality (TDM), a psychological model of moral judgment grounded in a simple two-node template: an intentional agent causing harm to a vulnerable patient. We formalize TDM using structural causal modeling (SCM) notation and identify three psychological operators (typecasting operator, completion operator, and valence-dependent inference mechanism) that extend standard SCM to capture how people compute moral judgments under constraints. We address scalability challenges arising from TDM's dyadic limitation, showing how moral cognition compresses multi-node scenarios through node collapse and sequential processing. Drawing on this algebraic framework, we demonstrate concrete applications to AI policy design: detecting conflicting obligations, structuring helpfulness policies to preserve user agency, and designing post-failure communication as causal interventions. Finally, we recommend scoped, contextual measurement of mind perception over universal averaging to operationalize the theory empirically. This algebraic formalization enables neurosymbolic AI systems to compute morality in a way that is both mathematically rigorous and faithful to human moral cognition.\n\n---\n\n## 8. Look Before You Leap: Autonomous Exploration for LLM Agents\n- **Authors**: Ziang Ye , Wentao Shi , Yuxin Liu , Yu Wang , Zhengzhou Cai , Yaorui Shi , Qi Gu , Xunliang Cai , Fuli Feng\n- **URL**: [https://arxiv.org/abs/2605.16143](https://arxiv.org/abs/2605.16143)\n- **Abstract**:\n> Large language model based agents often fail in unfamiliar environments due to premature exploitation: a tendency to act on prior knowledge before acquiring sufficient environment-specific information. We identify autonomous exploration as a critical yet underexplored capability for building adaptive agents. To formalize and quantify this capability, we introduce Exploration Checkpoint Coverage, a verifiable metric that measures how broadly an agent discovers key states, objects, and affordances. Our systematic evaluation reveals that agents trained with standard task-oriented reinforcement learning consistently exhibit narrow and repetitive behaviors that impede downstream performance. To address this limitation, we develop a training strategy that interleaves task-execution rollouts and exploration rollouts, with each type of rollout optimized by its corresponding verifiable reward. Building on this training strategy, we propose the Explore-then-Act paradigm, which decouples information-gathering from task execution: agents first utilize an interaction budget to acquire grounded environmental knowledge, then leverage it for task resolution. Our results demonstrate that learning to systematically explore is imperative for building generalizable and real-world-ready agents.\n\n---\n\n## 9. Property-Guided LLM Program Synthesis for Planning\n- **Authors**: Augusto B. Corrêa , André G. Pereira , Jendrik Seipp\n- **URL**: [https://arxiv.org/abs/2605.16142](https://arxiv.org/abs/2605.16142)\n- **Abstract**:\n> LLMs have shown impressive success in program synthesis, discovering programs that surpass prior solutions. However, these approaches rely on simple numeric scores to signal program quality, such as the value of the solution or the number of passed tests. Because a score offers no guidance on why a program failed, the system must generate and evaluate many candidates hoping some succeed, increasing LLM inference and evaluation costs. We study a different approach: property-guided LLM program synthesis. Instead of scoring programs after evaluation, we check whether a candidate satisfies a formally defined property. When the property is violated, we stop the evaluation early and provide the LLM with a concrete counterexample showing exactly how the program failed. This feedback drastically reduces both the number of program generations and the evaluation cost, and can guide the LLM to generate stronger programs. We evaluate this approach on PDDL planning domains, asking the LLM to synthesize direct heuristic functions: every state reachable by strictly improving transitions has a strictly improving successor. A heuristic with this property leads hill-climbing algorithm directly to a goal state. A counterexample-guided repair loop generates one candidate program, checks the property over a training set, and returns the first case that violates the property. We evaluate our approach on ten planning domains with an out-of-distribution test set. The synthesized heuristics are effectively direct on virtually all test tasks, and compared to the best prior generation method our approach generates seven times fewer programs per domain on average, solves more tasks without using search, and requires several orders of magnitude less computation to evaluate candidates. Whenever a problem admits a verifiable property, property-guided LLM synthesis can reduce cost and improve program quality.\n\n---\n\n## 10. ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents\n- **Authors**: Chinmay Savadikar , Mingyu Zhao , Yuanzheng Zhu , Han Li , Shuang Xie , Alberto Castelo , Tianfu Wu , Lingyun Wang\n- **URL**: [https://arxiv.org/abs/2605.16116](https://arxiv.org/abs/2605.16116)\n- **Abstract**:\n> Developing and evaluating e-commerce web agents requires environments that preserve meaningful task structure while enabling controllable, reproducible, and scalable scientific comparison. Existing methodologies force a tradeoff: live storefronts provide realism but are non-stationary, difficult to inspect, and irreproducible, while hand-built sandbox benchmarks provide control but cover only a narrow range of layouts, catalogs, policies, and interaction patterns. We argue that the core bottleneck is methodological: the field lacks a scalable way to construct evaluation settings that are simultaneously realistic, diverse, controllable, inspectable, and reproducible. We introduce ShopGym, an integrated framework for realistic simulation and scalable benchmarking of e-commerce web agents. ShopGym is a framework for constructing e-commerce simulation environments and grounded benchmark tasks. Its simulation layer, ShopArena, converts live seed storefronts into self-contained sandbox shops through anonymized shop specifications and a staged, validated generation process. On top of these simulated storefronts, ShopGuru synthesizes benchmark tasks across seven skill categories, grounding each task in the shop's catalog, navigation structure, policies, and interaction affordances. Together, ShopArena and ShopGuru produce self-contained, resettable, inspectable, and stable evaluation artifacts that preserve structural properties and agent-evaluation signals relevant to shopping tasks. We validate the framework through graph-based structural analysis and agent-based behavioral evaluation with 224 generated tasks across six sandbox shops: three constructed with synthetic data and three with real data. Our results show that the synthetic shops preserve key structural properties of live storefronts, with agent performance on synthetic shops positively correlated with performance on live storefronts.\n\n---\n\n## 11. Sign-Separated Finite-Time Error Analysis of Q-Learning\n- **Authors**: Donghwan Lee\n- **URL**: [https://arxiv.org/abs/2605.16103](https://arxiv.org/abs/2605.16103)\n- **Abstract**:\n> This paper develops a sign-separated finite-time error analysis for constant step-size Q-learning. Starting from the switching-system representation, the error is decomposed into its componentwise negative and positive parts. The negative part is dominated by a lower comparison linear time-invariant (LTI) system associated with a fixed optimal policy, whereas the positive part is controlled by a linear switching system. The resulting bounds show that the negative-side LTI certificate is no slower than the positive-side switching certificate and may produce a faster exponential envelope. The analysis identifies a max-induced asymmetry in Q-learning error dynamics. This asymmetry is connected to overestimation: positive action-wise errors can be selected and propagated by the Bellman maximum, whereas negative errors admit an optimal-policy lower comparison. Finite-time bounds are provided for both deterministic and stochastic constant-step-size recursions.\n\n---\n\n## 12. Reasoners or Translators? Contamination-aware Evaluation and Neuro-Symbolic Robustness in Tax Law\n- **Authors**: Parisa Kordjamshidi , Samer Aslan , Madhavan Seshadri , Leslie Barrett , Enrico Santus\n- **URL**: [https://arxiv.org/abs/2605.16052](https://arxiv.org/abs/2605.16052)\n- **Abstract**:\n> Recent advances in large language models (LLMs) have significantly enhanced automated legal reasoning. Yet, it remains unclear whether their performance reflects genuine legal reasoning ability or artifacts of data contamination. We present a comprehensive empirical study of tax law reasoning approaches and implement a contamination detection protocol to rigorously assess LLM reliability. We show that performance can be inflated by contamination. Building on this analysis, we conduct a systematic evaluation, comparing monolithic LLMs with hybrid systems that translate statutory text into formal representations and delegate inference to symbolic solvers. We build a novel test suite designed to probe generalization to unseen documents via case and rule variations. Our findings indicate that legal reasoning is inherently compositional and that neuro-symbolic frameworks offer a more reliable and robust foundation for legal AI, as well as improved generalization to unobserved situations.\n\n---\n\n## 13. ScreenSearch: Uncertainty-Aware OS Exploration\n- **Authors**: Michael Solodko , Justin Wagle\n- **URL**: [https://arxiv.org/abs/2605.16024](https://arxiv.org/abs/2605.16024)\n- **Abstract**:\n> Desktop GUI agents operate under partial observability: visually similar screens can correspond to different underlying workflow states, so locally plausible actions can lead to sharply different outcomes. We frame this as a problem of computer/OS state exploration, where effective behavior requires both expanding the reachable frontier and reducing ambiguity before committing. We present ScreenSearch, a system that combines structural screen retrieval and deduplication with an ambiguity-aware PUCT graph-bandit for large-scale desktop exploration. The retrieval layer converts UIA trees into location-aware structural features, indexes related screens through sparse token search and metadata filters, and maintains a shared deduplicated state graph across VM workers. On top of this graph, we define a scalable ambiguity signal based on matched-action outcome dispersion. If similar screens produce different next states under the same action signature, the state should be probed further rather than treated as resolved. We use this signal together with frontier rewards to drive large-scale exploration and replay-start policy evaluation over the shared graph. Across 11 desktop applications, ScreenSearch collects over 1M screenshots and over 30K deduplicated states, yielding large exploration corpora with substantial cross-application and within-application diversity. On a fixed replay-start slice, we observe a clear novelty--ambiguity trade-off: some policies reduce ambiguity quickly while discovering little frontier. Ambiguity reduction alone is therefore not a sufficient exploration objective. Appendix ablations show that stronger proposal priors can materially improve unique-state discovery during corpus building. These results suggest that state identity, proposal quality, and ambiguity-aware search all matter when deciding when to probe and when to commit.\n\n---\n\n## 14. Petri Net Induced Heuristic Search for Resource Constrained Scheduling\n- **Authors**: Ido Lublin , Dor Atzmon , Izack Cohen\n- **URL**: [https://arxiv.org/abs/2605.15983](https://arxiv.org/abs/2605.15983)\n- **Abstract**:\n> We formulate the Resource-Constrained Project Scheduling Problem (RCPSP) as optimal search over the reachability graph of a Timed Transition Petri Net with Resources, using relative-delay tokens so that scheduling decisions correspond to transition firings in the induced state space. We solve the resulting problem with $A^*$ guided by a heuristic that combines Critical Path and resource-based lower bounds, and prove that it is consistent under our token-based time semantics. Experiments on the PSPLIB benchmarks show that the approach outperforms strong exact Mixed-Integer Linear Programming (MIP) baselines (SCIP, CBC) in both success rate and solve time. Per-instance analysis shows that heuristic search and MIP degrade along independent axes, resource tightness for $A^*$ and formulation size for MIP, with resource strength mediating which solver benefits from scale.\n\n---\n\n## 15. Learning Bilevel Policies over Symbolic World Models for Long-Horizon Planning\n- **Authors**: Dillon Z. Chen , Till Hofmann , Toryn Q. Klassen , Sheila A. McIlraith\n- **URL**: [https://arxiv.org/abs/2605.15975](https://arxiv.org/abs/2605.15975)\n- **Abstract**:\n> We tackle the challenge of building embodied AI agents that can reliably solve long-horizon planning problems. Imitation learning from demonstrations has shown itself to be effective in training robots to solve a diversity of complex tasks requiring fine motor control and manipulation over low-level (LL), continuous environments. Yet, it remains a difficult endeavour to generate long-horizon plans from imitation learning alone. In contrast, high-level (HL), symbolic abstractions facilitate efficient and interpretable long-horizon planning. We propose to combine the strengths of LL imitation learning for manipulation and control, and HL symbolic abstractions for long-horizon planning. We realise this idea via \\emph{bilevel policies} of the form $(\\pi^{\\mathrm{hl} }, \\pi^{\\mathrm{ll} })$, consisting of a neural policy $\\pi^{\\mathrm{ll} }$ learned from LL demonstrations, and an HL symbolic policy $\\pi^{\\mathrm{hl} }$ that is constructed from symbolic abstractions of the LL demonstrations combined with inductive generalisation. We implement these ideas in the BISON system. Experiments on extended MetaWorld benchmarks demonstrate that BISON generalises to long horizons and problems with greater numbers of objects than those solved by VLA and end-to-end methods, and is more time and memory efficient in training and inference. Notably, when ignoring LL execution, BISON's HL policies can solve HL problems with 10,000 relevant objects in under a minute. Project page: this https URL\n\n---\n\n## 16. Deterministic Event-Graph Substrates as World Models for Counterfactual Reasoning\n- **Authors**: Fabio Rovai\n- **URL**: [https://arxiv.org/abs/2605.15967](https://arxiv.org/abs/2605.15967)\n- **Abstract**:\n> We study event-graph substrates: a class of world models that represent agent state as an append-only log of typed RDF triples and answer counterfactual queries by forking the log under a structured intervention vocabulary. Substrates are inspectable at the triple level, support exact counterfactuals, and transfer across domains without learned components. We formalize the class, prove a duality between explanatory and counterfactual queries that reduces both to the same causal-ancestor traversal, and evaluate a 1,400-line CLEVRER-DSL interpreter atop a domain-agnostic substrate runtime at full CLEVRER validation scale (n=75,618). The substrate exceeds the NS-DR symbolic oracle on all four per-question categories (by 9.89, 20.26, 17.65, and 0.80 percentage points), and exceeds the parametric ALOE baseline on descriptive and explanatory while lagging on predictive and counterfactual. We also introduce twin-EventLog, a 500-specification Park-canonical Smallville counterfactual benchmark on which the substrate exceeds Llama-3.1-8B with full context by 18.80 points joint accuracy.\n\n---\n\n## 17. PAGER: Bridging the Semantic-Execution Gap in Point-Precise Geometric GUI Control\n- **Authors**: Jingxuan Wei , Xi Bai , Shan Liu , Caijun Jia , Zheng Sun , Xinglong Xu , Siyuan Li , Linzhuang Sun , Bihui Yu , Conghui He , Cheng Tan\n- **URL**: [https://arxiv.org/abs/2605.15963](https://arxiv.org/abs/2605.15963)\n- **Abstract**:\n> Large vision-language models have significantly advanced GUI agents, enabling executable interaction across web, mobile, and desktop interfaces. Yet these gains largely rely on a forgiving region-tolerant paradigm, where many nearby pixels inside the same component remain valid. Precise geometric construction breaks this assumption: actions must land on points in continuous canvas space rather than tolerant regions. Because geometric primitives carry ontological dependencies, a local coordinate error can induce cascading topological failures that distort downstream objects and invalidate the final construction. We identify this regime as precision-sensitive GUI tasks, requiring point-level accuracy, geometry-aware verification, and robustness to dependency-driven error propagation. To benchmark it, we introduce PAGE Bench, with 4,906 problems and over 224K process-supervised, pixel-level GUI actions. We further propose PAGER, a topology-aware agent that decomposes construction into dependency-structured planning and pixel-level execution. Pixel-grounded supervised tuning establishes executable action grammar, while precision-aligned reinforcement learning mitigates rollout-induced exposure bias through state-conditioned geometric feedback. Experiments reveal a pronounced Semantic-Execution Gap: general multimodal models can exceed 88% action type accuracy yet remain below 6% task success. PAGER closes this gap, delivering 4.1x higher task success than the strongest evaluated general baseline and raising step success rate from below 9% for GUI-specialized agents to over 62%, establishing a new state of the art for point-precise GUI control.\n\n---\n\n## 18. Imperfect World Models are Exploitable\n- **Authors**: Logan Mondal Bhamidipaty (University of Edinburgh), Esmeralda S. Whitammer (University of Edinburgh), David Abel (University of Edinburgh), Mykel J. Kochenderfer (Stanford University), Subramanian Ramamoorthy (University of Edinburgh)\n- **URL**: [https://arxiv.org/abs/2605.15960](https://arxiv.org/abs/2605.15960)\n- **Abstract**:\n> We propose a novel definition of model exploitation in reinforcement learning. Informally, a world model is exploitable if it implies that one policy should be strictly preferred over another while the environment's true transition model implies the reverse. We analogize our definition with a prior characterization of reward hacking but show that the associated proof of inevitability does not transfer to exploitation. To overcome this obstruction, we develop a general theory of reward hacking and model exploitation that proves that exploitation is essentially unavoidable on large policy sets and yields the corresponding claim for hacking as a special case. Unfortunately, we also find that the conditions that guarantee unhackability in finite policy sets have no counterpart that precludes exploitation. Consequently, we introduce a relaxed notion of exploitation and derive a safe horizon within which it can be avoided. Taken together, our results establish a formal bridge between reward hacking and model exploitation and elucidate the limits of safe planning in world models.\n\n---\n\n## 19. Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design\n- **Authors**: Alberto Pepe , Chien-Yu Lin , Despoina Magka , Bilge Acun , Yannan Nellie Wu , Anton Protopopov , Carole-Jean Wu , Yoram Bachrach\n- **URL**: [https://arxiv.org/abs/2605.15871](https://arxiv.org/abs/2605.15871)\n- **Abstract**:\n> Toward recursive self-improvement, we investigate LLM agents autonomously designing foundation models beyond standard Transformers. We introduce a dual-framework approach: AIRA-Compose for high-level architecture search, and AIRA-Design for low-level mechanistic implementation. AIRA-Compose uses 11 agents to explore fundamental computational primitives under a 24-hour budget. Agents evaluate million-parameter candidates, extrapolating top designs to 350M, 1B, and 3B scales. This yields 14 architectures across two families: AIRAformers (Transformer-based) and AIRAhybrids (Transformer-Mamba). Pre-trained at 1B scale, these consistently outperform Llama 3.2 and Composer-found baselines. On downstream tasks, AIRAformer-D and AIRAhybrid-D improve accuracy by 2.4% and 3.8% over Llama 3.2. Furthermore, AIRA-Compose finds models with highly efficient scaling frontiers: AIRAformer-C scales 54% and 71% faster than Llama 3.2 and Composer's best Transformer, while AIRAhybrid-C outscales Nemotron-2 by 23% and Composer's best hybrid by 37%. AIRA-Design tasks 20 agents with writing novel attention mechanisms for long-range dependencies and high-performing training scripts. On the Long Range Arena benchmark, agent-designed architectures reach within 2.3% and 2.6% of human state-of-the-art on document matching and text classification. On the Autoresearch benchmark, Greedy Opus 4.5 achieves 0.968 validation bits-per-byte under a fixed time budget, surpassing the published minimum. Together, these frameworks show AI agents can autonomously discover architectures and algorithmic optimizations matching or surpassing hand-designed baselines. This establishes a powerful paradigm for discovering next-generation foundation models, marking a clear step toward recursive self-improvement.\n\n---\n\n## 20. SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows?\n- **Authors**: Kean Shi , Zihang Li , Tianyi Ma , Zengji Tu , Jialong Wu , Xinbo Xu , Qingyao Yang , Ruoyu Wu , Weichu Xie , Ming Wu , Jason Zeng , Michael Heinrich , Elvis Zhang , Liang Chen , Kuan Li , Baobao Chang\n- **URL**: [https://arxiv.org/abs/2605.15777](https://arxiv.org/abs/2605.15777)\n- **Abstract**:\n> Computer-Using Agents (CUAs) are rapidly extending large language models (LLMs) beyond text-based reasoning toward action execution in more complex environments, such as web browsers and graphical user interfaces (GUIs). However, existing web and GUI agent benchmarks often rely on simplified settings, isolated tasks, or short-horizon interactions, making it difficult to assess capabilities of agents in realistic professional workflows. Software-as-a-Service (SaaS) environments are a natural choice for CUA evaluation, as they host a large share of modern digital work and naturally involve dynamic system states, cross-application coordination, domain-specific knowledge, and long-horizon dependencies. To this end, we introduce SaaS-Bench, a benchmark built on 23 deployable SaaS systems across six professional domains, containing 106 tasks grounded in realistic work scenarios. These tasks require long-horizon execution, cover both text-only and multimodal settings, and are evaluated with weighted verification checkpoints that measure strict task completion and partial progress. Experiments show that representative LLM-based agents struggle on SaaS-Bench, with even the strongest model completing fewer than 4% of tasks end-to-end, exposing limitations in planning, state tracking, cross-application context maintenance, and error recovery. Code are available at this https URL for reproduction.\n\n---\n\n## 21. ALSO: Adversarial Online Strategy Optimization for Social Agents\n- **Authors**: Xiang Li , Liping Yi , Mingze Kong , Min Zhang , Zhongxiang Dai , QingHua Hu\n- **URL**: [https://arxiv.org/abs/2605.15768](https://arxiv.org/abs/2605.15768)\n- **Abstract**:\n> Social simulation provides a compelling testbed for studying social intelligence, where agents interact through multi-turn dialogues under evolving contexts and strategically adapting opponents. Such environments are inherently non-stationary, requiring agents to dynamically adjust their strategies over time. However, most Large Language Model (LLM) based social agents rely on static personas, while existing approaches for enhancing social intelligence, such as offline reinforcement learning or external planners, are ill-suited to these settings, typically assuming stationarity and incurring substantial training overhead. To bridge this gap, we propose \\textbf{ALSO} (\\textbf{A}dversarial on\\textbf{L}ine \\textbf{S}trategy \\textbf{O}ptimization), the first framework for online strategy optimization in multi-agent social simulation. ALSO advances social adaptation through two key contributions. (1) ALSO formulates multi-turn interaction as an adversarial bandit problem, where combinations of static personas and dynamic strategy instructions are treated as arms, providing a principled solution to non-stationarity without relying on environmental stability assumptions. (2) To predict rewards and generalize sparse feedback in multi-turn dialogues, ALSO introduces a lightweight neural surrogate to predict rewards from interaction histories, enabling sample-efficient exploration and continuous online adaptation. Experiments on the Sotopia benchmark demonstrate that ALSO consistently outperforms static baselines and existing optimization methods in dynamic environments, validating the effectiveness of adversarial online strategy optimization for building robust social agents.\n\n---\n\n## 22. Can We Trust AI-Inferred User States. A Psychometric Framework for Validating the Reliability of Users States Classification by LLMs in Operational Environments\n- **Authors**: Izabella Krzeminska , Michal Butkiewicz , Ewa Komkowska\n- **URL**: [https://arxiv.org/abs/2605.15734](https://arxiv.org/abs/2605.15734)\n- **Abstract**:\n> The use of large language models to assess user states in conversational and adaptive systems is based on the assumption that the metrics used for such assessment are stable and interpretable at the level of individual scores. This paper empirically tests this assumption, focusing on the psychometric reliability of artificial intelligence (AI) measures of user states. This study employed replication evaluation procedures to assess the repeatability of a broad set of metrics across three different bimodal large language models (GPT-4o audio, Gemini 2.0 Flash, Gemini 2.5 Flash). Analyses include both individual score reliability and aggregated reliability, allowing us to distinguish metrics potentially useful for real-time adaptation from those that retain their value only in aggregated analyses. The results demonstrate that metric reliability cannot be considered a default property in interpretive domains. The lack of stability at the level of individual scores precludes the interpretation of such scores as indicators of user state in real-time adaptive systems, even if these metrics demonstrate stability after aggregation. At the same time, the study indicates that individually unstable metrics can retain analytical utility in post-hoc studies, identifying rules governing interactions and their relationships with user experience parameters such as satisfaction, trust, and engagement. The main contribution of this work, besides quantifying the severity of the problem (only 31 of 213 metrics met the criteria), is the proposal of a replicable evaluation framework, enabling measurable evaluations of metric applicability. This approach supports more responsible AI design of adaptive systems, in which the interpretation of results requires explicit validation of reliability and monitoring for violations over time.\n\n---\n\n## 23. Nudging Beyond the Comfort Zone: Efficient Strategy-Guided Exploration for RLVR\n- **Authors**: Chanuk Lee , Sangwoo Park , Minki Kang , Sung Ju Hwang\n- **URL**: [https://arxiv.org/abs/2605.15726](https://arxiv.org/abs/2605.15726)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has emerged as a scalable paradigm for improving the reasoning capabilities of large language models. However, its effectiveness is fundamentally limited by exploration: the policy can only improve on trajectories it has already sampled. While increasing the number of rollouts alleviates this issue, such brute-force scaling is computationally expensive, and existing approaches that modify the optimization objective provide limited control over what is explored. In this work, we propose NudgeRL, a framework for structured and diversity-driven exploration in RLVR. Our approach introduces Strategy Nudging, which conditions each rollout on lightweight, strategy-level contexts to induce diverse reasoning trajectories without relying on expensive oracle supervision. To effectively learn from such structured exploration, we further propose a unified objective, which decomposes the reward signal into inter- and intra-context components and incorporates a distillation objective to transfer discovered behaviors back to the base policy. Empirically, NudgeRL outperforms standard GRPO with up to 8 times larger rollout budgets, while outperforming oracle-guided RL baseline on average across five challenging math benchmarks. These results demonstrate that structured, context-driven exploration can serve as an efficient and scalable alternative to both brute-force rollout scaling and feasibility-oriented methods based on privileged information. Our code is available at this https URL .\n\n---\n\n## 24. PRISM: Prompt Reliability via Iterative Simulation and Monitoring for Enterprise Conversational AI\n- **Authors**: Keshava Chaitanya , Jahnavi Gundakaram\n- **URL**: [https://arxiv.org/abs/2605.15665](https://arxiv.org/abs/2605.15665)\n- **Abstract**:\n> Deploying large language model (LLM)-driven conversational agents in enterprise settings requires prompts that are simultaneously correct at launch and resilient to the non-deterministic behavioral drift that characterizes production LLM deployments. Existing prompt optimization frameworks address prompt quality as a one-time compile-time problem, leaving open the equally critical question of how to detect and repair prompt regressions caused by silent LLM behavior changes over time. We present PRISM (Prompt Reliability via Iterative Simulation and Monitoring), a closed-loop framework that treats prompt engineering as a continuous reliability engineering problem rather than a one-time authorship task. PRISM takes as input plain-language agent requirements, a set of configured tools and memory variables, and an initial draft prompt. It automatically generates test cases from requirements, simulates full multi-turn conversations against a platform-faithful LLM environment, evaluates pass/fail using an LLM-as-judge, diagnoses root causes of failures, and surgically repairs the prompt -- iterating until all tests pass. Critically, PRISM is designed to run on a scheduled basis (daily), treating LLM behavioral drift as a first-class reliability concern. We evaluate PRISM across 35 enterprise conversational agents over a three-week deployment period on the this http URL V3 platform. PRISM reduces median prompt authoring time from 2 days to under 30 minutes, achieves 99% production reliability across all evaluated agents, and successfully identifies and repairs production regressions caused by LLM behavioral drift within a 24-hour detection window. Our results suggest that continuous, simulation-driven prompt optimization is both tractable and necessary for reliable enterprise conversational AI at scale.\n\n---\n\n## 25. ColPackAgent: Agent-Skill-Guided Hard-Particle Monte Carlo Workflows for Colloidal Packing\n- **Authors**: Lijie Ding , Changwoo Do\n- **URL**: [https://arxiv.org/abs/2605.15625](https://arxiv.org/abs/2605.15625)\n- **Abstract**:\n> We introduce ColPackAgent, an agent framework that autonomously runs Monte Carlo simulations of colloidal packing through a Model Context Protocol (MCP) tool server and an agent skill, whether as a standalone agent or inside an existing agent system. By harnessing the MCP server and agent skill, ColPackAgent executes a structured workflow for colloidal packing simulations, which are central to studies of phase behavior, self-assembly, and materials design. Without dedicated simulation tools and workflow instructions, general-purpose Large Language Model (LLM) agents tend to describe such workflows rather than execute them reliably. The MCP server exposes a custom-built colpack Python package that wraps HOOMD-blue hard-particle Monte Carlo, and the skill encodes a four-stage workflow contract. ColPackAgent can carry out the workflow interactively with human feedback, autonomously from an end-to-end prompt, or as autoresearch following a provided program file. We demonstrate the system in different modes with several colloidal packing simulation examples such as cube particles in 3D, a binary system of disks and capsules in 2D, and the 2D hard-disk freezing transition using autoresearch. We also compare model performance on this workflow across a panel of LLMs with 17 stage-specific prompts. This benchmark provides a stage-level check of how reliably different models follow the setup, planning, and analysis workflow. Together, these results show that pairing a domain Python package with MCP tools and a portable agent skill provides a practical route for turning a simulation toolkit into an agent-assisted research workflow.\n\n---\n\n## 26. TopoEvo: A Topology-Aware Self-Evolving Multi-Agent Framework for Root Cause Analysis in Microservices\n- **Authors**: Junle Wang , Xingchuang Liao , Wenjun Wu\n- **URL**: [https://arxiv.org/abs/2605.15611](https://arxiv.org/abs/2605.15611)\n- **Abstract**:\n> Root cause analysis (RCA) in microservices is challenging due to (i) noisy and heterogeneous multimodal observability (metrics, logs, traces), (ii) cascading failure propagation that amplifies downstream symptoms, and (iii) non-stationary topology drift induced by autoscaling and rolling updates. Recent LLM-based RCA agents can generate tool-grounded explanations, yet they often remain topology-agnostic and suffer from \\emph{symptom-amplification bias}, misattributing the root cause to salient downstream victims. We propose \\textbf{TopoEvo}, a topology-aware self-evolving multi-agent framework that couples graph representation learning with structured, topology-constrained reasoning. TopoEvo first introduces \\emph{Metric-orthogonal Multimodal Alignment} (MOMA), which decomposes metric embeddings into complementary subspaces and contrastively aligns logs and traces to reduce modality redundancy and sparsity, yielding stable node representations for graph encoding. It then applies \\emph{Vector Quantization} (VQ) to discretize topology-enhanced states into auditable \\emph{symptom tokens} with a symptom lexicon, enabling reliable retrieval and token-level evidence grounding. On top of these discrete topology cues, TopoEvo performs a multi-agent \\emph{Hypothesis--Evidence--Test} (HET) workflow to explicitly verify propagation-consistent explanations and separate initiating anomalies from amplified downstream symptoms. Finally, a \\emph{Self-Evolving Mechanism} refreshes hierarchical incident memory and performs conservative test-time adaptation with high-confidence pseudo-labels to maintain robustness under drift.\n\n---\n\n## 27. See Before You Code: Learning Visual Priors for Spatially Aware Educational Animation Generation\n- **Authors**: Yuejia Li , Ke He , Junheng Li , Shutong Chen , Jingkang Xia , Zhiyue Su , Junchi Zhang , Mang Ye\n- **URL**: [https://arxiv.org/abs/2605.15585](https://arxiv.org/abs/2605.15585)\n- **Abstract**:\n> Large language models can generate executable code for educational animations, but the resulting renders often exhibit visual defects, including element overlap, misalignment, and broken animation continuity. These defects cannot be reliably detected from the code alone and become apparent only after execution. We formalize this problem as render-feedback-aware constrained code generation: given a natural language specification, the model must generate executable code whose rendered output satisfies structured quality criteria that can be evaluated only after rendering. To address this problem, we introduce OmniManim, a render-feedback-aware educational animation generation framework built around a shared scene state, explicit visual planning, structured post-render diagnostics, and localized repair. Within OmniManim, the Vision Agent is a task-specific visual planning module: it predicts sparse keyframe layouts with coarse-to-fine bounding-box denoising and optimizes an interpolation-aware objective to reduce intermediate-frame failures induced by downstream animation interpolation. We further construct two datasets, ManimLayout-1K and EduRequire-500, and provide a reproducible evaluation protocol covering executability, instructional quality, visual quality, and efficiency. On EduRequire-500, OmniManim improves measured render quality over both single-model baselines and existing multi-agent frameworks. Systematic ablation studies further verify that explicit visual planning, especially its coarse spatial prior, bounding-box refinement, and interpolation-aware optimization, is central to these gains.\n\n---\n\n## 28. STAR: A Stage-attributed Triage and Repair framework for RCA Agents in Microservices\n- **Authors**: Junle Wang , Xingchuang Liao , Wenjun Wu\n- **URL**: [https://arxiv.org/abs/2605.15581](https://arxiv.org/abs/2605.15581)\n- **Abstract**:\n> LLM-based root cause analysis (RCA) agents have recently emerged as a promising paradigm for incident diagnosis in microservice AIOps. However, their reliability remains fragile: an error in early evidence collection, hypothesis formulation, or causal analysis can propagate through the reasoning trace and eventually corrupt the final diagnosis. In this paper, we present \\textbf{STAR}, a \\emph{Stage-attributed Triage and Repair} framework for repairing erroneous RCA traces. STAR explicitly decomposes an RCA workflow into four structured stages, namely \\emph{Evidence Package} (EP), \\emph{Hypothesis Set} (HS), \\emph{Analysis Structure} (AS), and \\emph{Decision Report} (DR), and treats agent failure as a stage-localizable reasoning bug rather than a monolithic end-to-end error. Built on top of LangGraph, STAR performs stage-wise auditing, budget-aware \\emph{Fast/Slow Routing}, \\emph{decisive stage localization via counterfactual candidate evaluation}, and stage-specific patch-and-replay repair. We evaluate STAR on a public large-scale benchmark and a real-world production dataset, using two RCA agent workflows and three foundation models. Experimental results show that STAR consistently improves both root cause localization and fault type classification over strong baselines. Moreover, STAR identifies the decisive faulty stage with high accuracy, repairs most initially incorrect traces within one or two replay rounds, and benefits substantially from both Fast/Slow Routing and counterfactual stage evaluation. These results suggest that explicitly modeling \\emph{where} an RCA agent fails is an effective path toward reliable, debuggable, and self-repairing agentic RCA systems.\n\n---\n\n## 29. Position: Artificial Intelligence Needs Meta Intelligence -- the Case for Metacognitive AI\n- **Authors**: Sergei Chuprov , Richard D. Lange , Leon Reznik , Paulo Shakarian , Raman Zatsarenko , Dmitrii Korobeinikov\n- **URL**: [https://arxiv.org/abs/2605.15567](https://arxiv.org/abs/2605.15567)\n- **Abstract**:\n> This position paper argues for metacognition as a general design principle for creating more accurate, secure, and efficient AI. The metacognitive solution involves systems monitoring their own states and judiciously allocating resources depending on each problem instance's difficulty or cost of mistakes. Drawing inspiration both from past work on resource-rational AI and from well-documented metacognitive strategies in psychology and cognitive science, we identify specific challenges in embedding these strategies into AI design and highlight open theoretical and implementation problems. We showcase these principles through a tangible example of improved learning efficiency, effectiveness, and security in a Federated Learning (FL) case study. We show how these principles can be translated into practice with a novel software framework developed specifically to allow the community to design, deploy, and experiment with metacognition-enabled AI applications.\n\n---\n\n## 30. DRS-GUI: Dynamic Region Search for Training-Free GUI Grounding\n- **Authors**: Yichao Liu , Huawen Shen , Liu Yu , Shiyu Liu , Zeyu Chen , Yu Zhou\n- **URL**: [https://arxiv.org/abs/2605.15542](https://arxiv.org/abs/2605.15542)\n- **Abstract**:\n> GUI agents powered by Multimodal Large Language Models (MLLMs) have demonstrated impressive capability in understanding and executing user instructions. However, accurately grounding instruction-relevant elements from high-resolution screenshots cluttered with irrelevant UI components remains challenging for existing approaches. Inspired by how humans dynamically adjust their perceptual scope to locate task-related regions on complex screens, we propose DRS-GUI, a training-free dynamic region search framework for GUI grounding that can be seamlessly integrated into existing MLLMs. DRS-GUI introduces a lightweight UI Perceptor that performs three human-like perceptual actions (Focus, Shift, and Scatter) to progressively explore the interface and generate region proposals. To dynamically schedule these actions, we further design an Action Planner based on Monte Carlo Tree Search (MCTS). A region quality reward is employed to evaluate and select the highly instruction-relevant region, efficiently pruning redundant UI elements. Experiments demonstrate that DRS-GUI yields a 14\\% improvement on ScreenSpot-Pro for general and GUI-specific MLLMs (Qwen2.5-VL-7B and UGround-V1-7B), significantly enhancing grounding performance and generalization.\n\n---\n\n## 31. RTL-BenchMT: Dynamic Maintenance of RTL Generation Benchmark Through Agent-Assisted Analysis and Revision\n- **Authors**: Jing Wang , Shang Liu , Hangan Zhou , Zhiyao Xie\n- **URL**: [https://arxiv.org/abs/2605.15537](https://arxiv.org/abs/2605.15537)\n- **Abstract**:\n> This paper introduces RTL-BenchMT, an agentic framework for dynamically maintaining RTL generation benchmarks. Large Language Models (LLMs) assisted automated RTL generation is one of the most important directions in EDA research. However, current RTL benchmarks face two critical challenges: (1) flawed cases in the benchmarks and (2) overfitting to the benchmarks. Both challenges are difficult to resolve purely by manual engineering effort. To address these issues and systematically reduce human maintenance costs, we propose an automated agentic framework, RTL-BenchMT. RTL-BenchMT focuses on two key applications: (1) automatically identifying and revising flawed benchmark cases and (2) automatically detecting and updating overfitting cases. With the assistance of RTL-BenchMT, we conduct a thorough, in-depth analysis of flawed and overfitting cases and produce a refined benchmark suite that will be open-sourced to the community.\n\n---\n\n## 32. CAPS: Cascaded Adaptive Pairwise Selection for Efficient Parallel Reasoning\n- **Authors**: Fangzhou Lin , Shuo Xing , Peiran Li , Siyuan Yang , Qianwen Ge , Kazunori Yamada , Ziming Zhang , Haichong Zhang , Zhengzhong Tu\n- **URL**: [https://arxiv.org/abs/2605.15513](https://arxiv.org/abs/2605.15513)\n- **Abstract**:\n> Parallel reasoning, where a generator samples many candidate solutions and an aggregator selects the best, is one of the most effective forms of test-time scaling in large language models, and pairwise self-verification has become its strongest aggregation primitive. Yet pairwise verification carries a heavy cost: each judgment reads two complete solutions in full, and existing methods perform tens of such judgments per problem regardless of whether the comparison is informative. We introduce CAPS (Cascaded Adaptive Pairwise Selection), an inference-only framework that allocates verifier compute non-uniformly along two orthogonal axes: an evidence axis that adapts how much of each candidate the judge sees, and a distribution axis that adapts how comparisons are spread across the pool. CAPS instantiates these into a four-stage cascade with an optional rescue subroutine, and admits a closed-form verifier-token cost in which the per-candidate marginal cost is roughly halved relative to uniform full-evidence schedules. On four self-verifying models (Qwen3-14B, GPT-OSS-20B, Qwen3-4B-Instruct/Thinking) and five reasoning benchmarks spanning code (LiveCodeBench-v5/v6, CodeContests) and math (AIME 2025, HMMT 2025), CAPS outperforms the leading pairwise verifier on 14 of 20 suites while using 25.4% of its verifier-token budget on code, and outperforms pointwise self-verification on all 20. The trade-off suites admit an interpretable diagnostic in terms of the verifier's accuracy at partial versus full evidence, providing a concrete pre-deployment check for cascade suitability.\n\n---\n\n## 33. X-SYNTH: Beyond Retrieval -- Enterprise Context Synthesis from Observed Human Attention\n- **Authors**: Guruprasad Raghavan , George Nychis , Rohan Narayana Murthy\n- **URL**: [https://arxiv.org/abs/2605.15505](https://arxiv.org/abs/2605.15505)\n- **Abstract**:\n> In enterprise operations, the context required for an AI agent task is scattered across systems of record, static information stores, and communication channels. What is stored is system state, a lossy representation of the work that actually happened [2, 52]. The prevailing approach [17, 31, 34, 36] retrieves by matching request content to what is stored; for narrow requests this works well. But synthesis quality depends on knowing what to surface and how to interpret it: knowledge specific to each organization, team, and individual [5, 57, 61], present in behavioral patterns, absent from any retrieval index. For complex agentic tasks it breaks down: True Lead Rate is low, False Lead Rate is high, and the model has no mechanism to improve. We present X-SYNTH, a framework for enterprise context synthesis grounded in human attention, the digitally observable interaction signatures of each worker, encoding not just what they did but the sequence in which they did it, along with implicit reward signals. Behavioral traces preceding positive outcomes are distinguishable from those that did not, without external labeling. X-SYNTH models each individual's behavioral baseline as a Digital Twin Signature (DTS) and selects among seven qualitatively distinct attention filters: Proportional, Inverse, Differential, Recurrent, Comparative, Sequential, and Collective, per individual and per query, to identify causally relevant activity signatures. A four-stage pipeline assembles ranked context grounded in behavioral patterns rather than query embeddings. On a sales lead identification task, a frontier model unaided achieves 9.5% True Lead Rate (TLR) with 90.5% False Lead Rate (FLR). Augmented with X-SYNTH, TLR rises to 61.9% (6.5x) while FLR falls to 18.8%. Enterprise context synthesis is not a retrieval problem. It is a relevance problem, and human attention is its most reliable ground truth.\n\n---\n\n## 34. From LLM-Generated Conjectures to Lean Formalizations: Automated Polynomial Inequality Proving via Sum-of-Squares Certificates\n- **Authors**: Ruobing Zuo , Hanrui Zhao , Gaolei He , Zhengfeng Yang , Jianlin Wang\n- **URL**: [https://arxiv.org/abs/2605.15445](https://arxiv.org/abs/2605.15445)\n- **Abstract**:\n> Automated proving of polynomial inequalities is a fundamental challenge in automated mathematical reasoning, where rich algebraic structure and a rapidly growing certificate search space hinder scalability. Purely symbolic approaches provide strong guarantees but often scale poorly as the number of variables or the degree increases, due to expensive algebraic manipulations and rapidly growing intermediate expressions. In parallel, LLM-guided methods have made notable progress, particularly on competition-style inequalities with a small number of variables. To address the remaining scalability challenges, we propose NSPI, a neuro-symbolic framework that combines the complementary strengths of LLMs and symbolic computation for polynomial-inequality proving. Concretely, an LLM proposes a conjecture in the form of an approximate polynomial Sum-Of-Squares (SOS) decomposition; we refine it via symbolic computation to obtain an exact polynomial SOS representation, which directly proves the target inequality, and we further certify the proof in Lean, yielding an end-to-end pipeline from heuristic discovery to machine-checked proof. Experiments on challenging benchmarks involving polynomials with up to 10 variables demonstrate the effectiveness and scalability of the proposed method.\n\n---\n\n## 35. Beyond Partner Diversity: An Influence-Based Team Steering Framework for Zero-Shot Human-Machine Teaming\n- **Authors**: Wei Sheng , Rohan Paleja\n- **URL**: [https://arxiv.org/abs/2605.15400](https://arxiv.org/abs/2605.15400)\n- **Abstract**:\n> While AI agents are rapidly advancing from isolated tools to interactive collaborators, data-driven human-machine teaming (HMT) methods remain costly in their reliance on human interaction data across domains, teammates, and team sizes. Zero-shot coordination (ZSC) addresses this bottleneck by simulating diverse partner populations to approximate how unseen partners might behave. However, partner coverage alone is insufficient as team settings scale and communication becomes degraded. To remedy this deficiency, we propose Influence-Based Team Steering (IBTS), a framework that uses influence shaping to incentivize agents to discover diverse, high-performing team interaction patterns and further steers ongoing trajectories toward stronger learned coordination modes. We assess IBTS on Overcooked-AI in both two-agent and three-agent settings, allowing us to test whether learned coordination structure transfers beyond dyadic interaction. Our evaluation includes simulated partners, synthetic partner-style variation, and, to our knowledge, the first 30-subject Overcooked-AI HMT study involving two real human teammates and one machine teammate. Across these evaluations, IBTS improves team performance against competing baselines, highlighting the need for scaled ZSC to combine sparse-reward coordination mechanisms with partner-variation coverage rather than relying on diversity alone.\n\n---\n\n## 36. Ensemble Monitoring for AI Control: Diverse Signals Outweigh More Compute\n- **Authors**: Eugene Koran , Yejun Yun , Samantha Tetef , Benjamin Arnav , Pablo Bernabeu-Pérez\n- **URL**: [https://arxiv.org/abs/2605.15377](https://arxiv.org/abs/2605.15377)\n- **Abstract**:\n> As AI systems are increasingly deployed in autonomous agentic settings at scale, it is important to ensure the actions they take are safe and aligned with user intent. Monitoring agent actions is a key safety mechanism, yet reliable monitors remain difficult to build and the scale of these systems makes human oversight impractical. We show that combining signals from diverse monitors into an ensemble improves detection of misaligned actions. We build 12 GPT-4.1-Mini monitors using both prompting and fine-tuning strategies. We evaluate them on coding tasks where candidate solutions pass standard tests but fail on adversarial inputs. In this setting, diverse ensembles outperform both individual monitors and homogeneous ensembles. Our best 3-monitor ensemble achieves 2.4x greater detection performance gain compared to an ensemble composed of three identical monitors, with the same ensemble performing strongly on an independent dataset. We contend that these results show that diversity - not scale - drives gains. The best ensembles combine strong individual performance with low correlation between monitors. Furthermore, fine-tuned monitors appear in every top-performing ensemble and maintain this advantage on out-of-distribution attack types, suggesting that fine-tuning enables detection capabilities that prompting alone does not elicit. These results support ensemble monitoring as a practical AI control strategy for safety gains at reasonable inference costs.\n\n---\n\n## 37. Belief Engine: Configurable and Inspectable Stance Dynamics in Multi-Agent LLM Deliberation\n- **Authors**: Joshua C. Yang , Maurice Flechtner , Damian Dailisan , Michiel A. Bakker\n- **URL**: [https://arxiv.org/abs/2605.15343](https://arxiv.org/abs/2605.15343)\n- **Abstract**:\n> LLM-based agents are increasingly used to simulate deliberative interactions such as negotiation, conflict resolution, and multi-turn opinion exchange. Yet generated transcripts often do not reveal why an agent's stance changes: movement may reflect evidence uptake, anchoring, role drift, echoing, or changed prompt and retrieval context. We introduce the Belief Engine (BE), an auditable belief-update layer that treats \"belief\" as an evidential state over a proposition and exposes it as scalar stance. BE extracts arguments into structured memory and updates stance with a log-odds rule controlled by evidence uptake u and prior anchoring a. Across multiple base LLMs, parameter sweeps show that these controls reliably shape stance dynamics while preserving an evidence-level update trail. On DEBATE, a human deliberation dataset with pre/post opinions, BE best reconstructs participants whose final stance follows extracted evidence; stable and evidence-opposed cases instead point to anchoring or factors outside the extracted evidence stream. BE provides configurable infrastructure for studying evidence-grounded deliberation, where openness, commitment, convergence, and disagreement can be tied to explicit update assumptions rather than hidden prompt effects.\n\n---\n\n## 38. Zero-Shot Goal Recognition with Large Language Models\n- **Authors**: Kin Max Piamolini Gusmão , Nathan Gavenski , Nir Oren , Felipe Meneguzzi\n- **URL**: [https://arxiv.org/abs/2605.15333](https://arxiv.org/abs/2605.15333)\n- **Abstract**:\n> Large language models have recently reached near-parity with classical planners on well-known planning domains, yet this competence relies on world-knowledge exploitation rather than genuine symbolic reasoning. Goal recognition is a complementary abductive task structurally better suited to LLM strengths: it consists of evaluating consistency with world knowledge rather than generating novel action sequences. This paper provides the first systematic zero-shot evaluation of frontier LLMs as goal recognisers on key classical PDDL benchmarks. Our results show that LLM competence on goal recognition is uneven: some models scale with evidence and approach landmark-based accuracy at full observations, while others remain anchored to world-knowledge priors regardless of how much evidence accumulates. Qualitative analysis of model reasoning traces reveals that this divergence reflects a fundamental difference in evidence integration rather than domain familiarity. These findings position goal recognition as a principled benchmark for the foundational planning knowledge of LLMs.\n\n---\n\n## 39. Context Pruning for Coding Agents via Multi-Rubric Latent Reasoning\n- **Authors**: Jingjing Wang , Xiwen Chen , Wenhui Zhu , Huayu Li , Zhengxiao He , Feiyang Cai , Ana S. Carreon-Rascon , Xuanzhao Dong , Feng Luo\n- **URL**: [https://arxiv.org/abs/2605.15315](https://arxiv.org/abs/2605.15315)\n- **Abstract**:\n> LLM-powered coding agents spend the majority of their token budget reading repository files, yet much of the retrieved code is irrelevant to the task at hand. Existing learned pruners compress this context with a single-objective sequence labeler, collapsing all facets of code relevance into one score and one transition matrix. We show that this formulation creates a modeling bottleneck: a single CRF transition prior must serve heterogeneous retention patterns, including contiguous semantic spans and sparse structural support lines. We propose LaMR (Latent Multi-Rubric), a structured pruning framework that decomposes code relevance into two interpretable quality dimensions, semantic evidence and dependency support, each modeled by a dedicated CRF with dimension-specific transition dynamics. A mixture-of-experts gating network dynamically weights the per-rubric emissions conditioned on the query, and a final CRF layer on the fused emissions produces the aggregate keep-or-prune decision. To supervise each dimension without additional annotation cost, we derive multi-rubric labels from the existing training corpus via AST-based program analysis, simultaneously denoising the teacher's binary labels. By effectively filtering distracting noise, LaMR frequently matches or even outperforms unpruned full-context baselines. Experiments on four benchmarks (SWE-Bench Verified, SWE-QA, LCC, LongCodeQA) show that LaMR wins 12 of 16 head-to-head multi-turn comparisons. It saves up to 31% more tokens on multi-turn agent tasks and improves Exact Match by up to +3.5 on single-turn tasks, while performance is frequently enhanced by denoising the context, and any remaining drops are marginal.\n\n---\n\n## 40. SMCEvolve: Principled Scientific Discovery via Sequential Monte Carlo Evolution\n- **Authors**: Jiachen Jiang , Huminhao Zhu , Zhihui Zhu\n- **URL**: [https://arxiv.org/abs/2605.15308](https://arxiv.org/abs/2605.15308)\n- **Abstract**:\n> LLM-driven program evolution has emerged as a powerful tool for automated scientific discovery, yet existing frameworks offer no principled guide for designing their individual components and provide no guarantee that the search converges. We introduce SMCEvolve, which recasts program search as sampling from a reward-tilted target distribution and approximates it with a Sequential Monte Carlo (SMC) sampler. From this view, three core mechanisms emerge as principled components: adaptive parent resampling, mixture of mutation with acceptance, and automatic convergence control. We further provide a finite-sample complexity analysis that bounds the LLM-call budget required to reach a target approximation error. Across math, algorithm efficiency, symbolic regression, and end-to-end ML research benchmarks, SMCEvolve surpasses state-of-the-art evolving systems while using fewer LLM calls under self-determined termination. The code is available at this https URL .\n\n---\n\n## 41. Solvita: Enhancing Large Language Models for Competitive Programming via Agentic Evolution\n- **Authors**: Han Li , Jinyu Tian , Rili Feng , Yuqiao Du , Chong Zheng , Chenyu Wang , Chenchen Liu , Shihao Li , Xinping Lei , Yifan Yao , Weihao Xie , Letian Zhu , Jiaheng Liu\n- **URL**: [https://arxiv.org/abs/2605.15301](https://arxiv.org/abs/2605.15301)\n- **Abstract**:\n> Large language models (LLMs) still struggle with the rigorous reasoning demands of hard competitive programming. While recent multi-agent frameworks attempt to bridge this reliability gap, they remain fundamentally stateless: they rely on static retrieval and discard the valuable problem-solving and debugging experience gained from previous tasks. To address this, we present Solvita, an agentic evolution framework that enables continuous learning without requiring weight updates to the underlying LLM. Solvita reorganizes problem-solving into a closed-loop system of strategy selection, program synthesis, certified supervision, and targeted hacking, executed by four specialized agents: Planner, Solver, Oracle, and Hacker. Crucially, each agent is paired with a trainable, graph-structured knowledge network. As the system operates, outcome signals, such as pass/fail verdicts, test certification quality, and adversarial vulnerabilities discovered by the Hacker, are recast as reinforcement learning updates to these network weights. This allows the agents to dynamically route future queries based on past successes and failures, effectively accumulating transferable reasoning experience over time. Evaluated across CodeContests, APPS, AetherCode, and live Codeforces rounds, Solvita establishes a new state-of-the-art among code-generation agents, outperforming existing multi-agent pipelines and nearly doubling the accuracy of single-pass baselines.\n\n---\n\n## 42. Verifiable Agentic Infrastructure: Proof-Derived Authorization for Sovereign AI Systems\n- **Authors**: Jun He , Deying Yu\n- **URL**: [https://arxiv.org/abs/2605.15228](https://arxiv.org/abs/2605.15228)\n- **Abstract**:\n> Modern cloud and enterprise systems rely on identity-centric authorization, assuming that callers possessing valid credentials are safe to execute commands. The emergence of autonomous AI agents invalidates this assumption: agents can generate syntactically valid but semantically unsafe actions, making standing privileges a significant operational risk. This risk becomes especially acute in sovereign AI systems, where autonomous agents may interact with cloud infrastructure, regulated data, financial workflows, and national-scale digital services. Governed mutation substrates reduce this risk by interposing on agent actions: agents submit intents, infrastructure evaluates context and policy, and execution is mediated. However, this shifts the trust boundary: how can the decision to authorize an intent be made verifiable, distributed, and replayable? We introduce a Distributed Trust Framework (DTF), a verification framework for governed mutation systems that computes execution authority from structured, verifiable artifacts. DTF introduces a Justification Proof to encode the admissibility basis of an action, a consensus model for independent evaluation, an ephemeral Execution Identity derived from the approved proof, and an append-only Evidence Chain that preserves the authorization lifecycle. Under stated substrate assumptions, this architecture enforces a compact authorization invariant: no high-stakes execution without a proof object, no derived authority without consensus, and no valid mutation detached from evidence. We define the model, instantiate it over an OpenKedge-based governed mutation substrate, and show how it maps onto cloud-native environments. By shifting authorization from standing identity to proof-derived authority, DTF provides an infrastructure foundation for making agentic execution governable, auditable, and bounded in sovereign AI deployments.\n\n---\n\n## 43. NIMO Controller: a self-driving laboratory orchestrator based on the Model Context Protocol\n- **Authors**: Naruki Yoshikawa , Ryo Tamura\n- **URL**: [https://arxiv.org/abs/2605.15227](https://arxiv.org/abs/2605.15227)\n- **Abstract**:\n> Self-driving laboratories (SDLs) have attracted increasing attention as a means of accelerating scientific discovery; however, developing SDL software remains technically demanding. To improve accessibility, orchestration software frameworks have been proposed to coordinate SDL components. Nevertheless, existing frameworks are primarily designed for human interaction and do not provide standardized interfaces suitable for AI agents. In this work, we propose an SDL software architecture based on the Model Context Protocol (MCP), in which all SDL functionalities are exposed through MCP servers. Following this design principle, we introduce an MCP-based SDL orchestrator, named NIMO Controller. It provides a visual programming interface automatically generated through MCP-based tool discovery, allowing human users to design experimental workflows without writing code. The same MCP backend can also be accessed by AI agents, providing a unified interface for both human users and AI agents. We demonstrate the proposed system through a case study on a color-matching SDL. The results validate the usability of the proposed MCP-based SDL architecture.\n\n---\n\n## 44. ICRL: Learning to Internalize Self-Critique with Reinforcement Learning\n- **Authors**: Jianbo Lin , Xiaomin Yu , Yi Xin , Yifu Guo , Zhuosong Jiang , Zhongqi Yue , Weishi Wang , Heqing Zou , Chengwei Qin , Hui Xiong\n- **URL**: [https://arxiv.org/abs/2605.15224](https://arxiv.org/abs/2605.15224)\n- **Abstract**:\n> Large language model-based agents make mistakes, yet critique can often guide the same model toward correct behavior. However, when critique is removed, the model may fail again on the same query, indicating that it has not internalized the critique's guidance into its underlying capability. Meanwhile, a frozen critic cannot improve its feedback quality over time, limiting the potential for iterative self-improvement. To address this, we propose learning to internalize self-critique with reinforcement learning(ICRL), a novel framework that jointly trains a solver and a critic from a shared backbone to convert critique-induced success into unassisted solver ability. The critic is rewarded based on the solver's subsequent performance gain, incentivizing actionable feedback. To address the distribution shift between critique-conditioned and critique-free behavior, ICRL introduces a distribution-calibration re-weighting ratio that selectively transfers critique-guided improvements compatible with the solver's own prompt distribution. Additionally, a role-wise group advantage estimation stabilizes joint optimization across the two roles. Together, these mechanisms ensure that the solver learns to improve itself without external critique, rather than becoming dependent on critique-conditioned behavior. We evaluate ICRL on diverse benchmarks spanning agentic and mathematical reasoning tasks, using Qwen3-4B and Qwen3-8B as backbones. Results show consistent improvements, with average gains of 6.4 points over GRPO on agentic tasks, and 7.0 points on mathematical reasoning. Notably, the learned 8B critic is comparable to 32B critics while using substantially fewer tokens. The code is available at this https URL .\n\n---\n\n## 45. NOVA: Fundamental Limits of Knowledge Discovery Through AI\n- **Authors**: Salman Avestimehr , Ken Duffy , Muriel Médard\n- **URL**: [https://arxiv.org/abs/2605.15219](https://arxiv.org/abs/2605.15219)\n- **Abstract**:\n> Can AI systems discover genuinely new knowledge through iterative self improvement, and if so, at what cost? We introduce the NOVA framework, which models the common ``generate, verify, accumulate, retrain'' loop as an adaptive sampling process over a knowledge space. We identify sufficient conditions under which accumulated genuine knowledge eventually covers a finite domain, and show how their violations produce distinct failure modes: contamination, forgetting, exploration failure, and acceptance failure. We then analyze imperfect verification and identify a contamination trap: as easy-to-find knowledge is exhausted, the model mass assigned to new valid artifacts shrinks, so even small false-positive rates can cause invalid artifacts to enter the knowledge base faster than genuine discoveries. We clarify that Good--Turing estimation is a local batch-diversity diagnostic, not an estimator of the historically undiscovered valid mass that governs long-term discovery. Under a separate tail-equivalence assumption relating the model's effective discovery distribution to a Zipf law with exponent $\\alpha>1$, we prove that the cumulative generation cost required to obtain $D$ distinct genuine discoveries satisfies $R_{\\mathrm{cum} }(D)=\\Theta(c_{\\mathrm{gen} }D^\\alpha)$, where $c_{\\mathrm{gen} }$ is the per-candidate generation cost. This scaling law quantifies asymptotic diminishing returns as the discovery frontier advances. Finally, we formalize human amplification through guidance, generation, and verification, explaining why expert input is most valuable near autonomous exploration barriers.\n\n---\n\n## 46. CAX-Agent: A Lightweight Agent Harness for Reliable APDL Automation\n- **Authors**: Chenying Lin , Yichen Hai , Yi He , Ran Wang , Haiyan Qiang , Liang Yu\n- **URL**: [https://arxiv.org/abs/2605.15218](https://arxiv.org/abs/2605.15218)\n- **Abstract**:\n> Large language models deployed for MAPDL finite-element simulation face practical reliability challenges: without structured execution control, tool encapsulation, and fault recovery, outputs may be inconsistent and task failures are common. The Agent Harness paradigm addresses this by inserting domain-specific orchestration middleware that manages tool lifecycles, workflow state, and recovery escalation. This paper presents the architecture of CAX-Agent, a lightweight agent harness purpose-built for MAPDL automation, and empirically evaluates one of its core components -- the recovery this http URL -Agent organizes execution into three layers -- LLM service, agent harness, and solver backend -- with a recovery ladder that escalates from deterministic rule patching through model-driven regeneration to context enrichment and human intervention. We evaluate three recovery strategies (no_recovery, rule_only, and model_only) on 50 standard structural benchmarks with three repeated runs per strategy (450 case-runs total). Two independent human raters score task completion under blind conditions; inter-rater agreement is strong (quadratic weighted Cohen's kappa = 0.84, 96 percent of score pairs within one point). Model_only achieves the best completion rate (0.9267), task score (3.59/4), total score (9.16/10), and zero-intervention rate (0.84), outperforming rule_only (0.7733, 3.17/4, 7.03/10, 0.00) and no_recovery (0.6933, 2.74/4, 5.60/10, 0.00) with large effect sizes (Cliff's delta = 0.81-0.87). The benchmark uses deliberately simple geometries to isolate recovery-policy effects; we discuss the scope of these findings and directions for broader validation.\n\n---\n\n## 47. Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions\n- **Authors**: Jagdish Tripathy , Marcus Buckmann\n- **URL**: [https://arxiv.org/abs/2605.15217](https://arxiv.org/abs/2605.15217)\n- **Abstract**:\n> Instruction-tuned language models exhibit behavioural fairness in high-stakes decisions while retaining biased associations in their internal representations. However, whether these suppressed representations can affect model outputs - and whether such causal potency is symmetric across demographic groups - remains unknown. We investigate the use of open-weight models for mortgage underwriting using matched applications that differ only in racially-associated names and reveal a critical disconnect: models show no output-level bias, yet retain and amplify demographic representations across model layers. Through activation steering and novel cross-layer interventions, we demonstrate that this suppressed information is decision-relevant: when reinjected at critical layers, it produces near-complete decision reversals. Critically, this latent bias is asymmetric - steering interventions affect decisions in one demographic direction, while producing minimal effects in reverse - and susceptible to adversarial prompt engineering and parameter-efficient fine-tuning. These findings demonstrate that behavioural audits focused on outputs are insufficient: fair outputs can mask exploitable internal biases. They also motivate dual-layer testing frameworks combining output evaluation with representational analysis for AI governance in high-stakes decisions.\n\n---\n\n## 48. SkillSmith: Compiling Agent Skills into Boundary-Guided Runtime Interfaces\n- **Authors**: Duling Xu , Zheng Chen , Zaifeng Pan , Jiawei Guan , Dong Dong , Jialin Li , Bangzheng Pu\n- **URL**: [https://arxiv.org/abs/2605.15215](https://arxiv.org/abs/2605.15215)\n- **Abstract**:\n> Recently, skills have been widely adopted in large language model (LLM)-based agent systems across various domains. In existing frameworks, skills are typically injected into the agent reasoning loop as contextual guidance once matched to a runtime task, enabling specialized task-solving capabilities. We find that this execution paradigm introduces two major sources of redundancy: irrelevant context injection and repeated skill-specific reasoning and planning. To this end, we propose SkillSmith, a boundary-first compiler-runtime framework that compiles skill packages offline into minimal executable interfaces. By extracting fine-grained operational boundaries from skills, SkillSmith enables agents to dynamically access and execute only the relevant components at runtime, thereby minimizing unnecessary context injection and redundant reasoning overhead. In the evaluation on SkillsBench benchmark, SkillSmith reduces solve-stage token usage by 57.44%, thinking iterations by 42.99%, solve time by 50.57% (2.02x faster), and token-proportional monetary cost by 57.44% compared with using raw-skills. Moreover, compiled artifacts produced by a stronger model can be reused by a smaller or more efficient runtime model, improving task accuracy in cases where raw skill interpretation fails. The source code and data are available at this https URL .\n\n---\n\n## 49. Does Theory of Mind Improvement Really Benefit Human-AI Interactions? Empirical Findings from Interactive Evaluations\n- **Authors**: Nanxu Gong , Zixin Chen , Haotian Li , Zishu Zhao , Jianxun Lian , Huamin Qu , Yanjie Fu , Xing Xie\n- **URL**: [https://arxiv.org/abs/2605.15205](https://arxiv.org/abs/2605.15205)\n- **Abstract**:\n> Improving the Theory of Mind (ToM) capability of Large Language Models (LLMs) is crucial for effective social interactions between these AI models and humans. However, the existing benchmarks often measure ToM capability improvement through story-reading, multiple-choice questions from a third-person perspective, while ignoring the first-person, dynamic, and open-ended nature of human-AI (HAI) interactions. To directly examine how ToM improvement techniques benefit HAI interactions, we first proposed the new paradigm of interactive ToM evaluation with both perspective and metric shifts. Next, following the paradigm, we conducted a systematic study of four representative ToM enhancement techniques using both four real-world datasets and a user study, covering both goal-oriented tasks (e.g., coding, math) and experience-oriented tasks (e.g., counseling). Our findings reveal that improvements on static benchmarks do not always translate to better performance in dynamic HAI interactions. This paper offers critical insights into ToM evaluation, showing the necessity of interaction-based assessments in developing next-generation, socially aware LLMs for HAI symbiosis.\n\n---\n\n## 50. SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch\n- **Authors**: Zhantao Wang\n- **URL**: [https://arxiv.org/abs/2605.15204](https://arxiv.org/abs/2605.15204)\n- **Abstract**:\n> Multi-agent orchestration frameworks such as LangChain, LangGraph, and CrewAI route tasks through graph-based pipelines but do not enforce the stage constraints that govern real business processes. We present SDOF, a framework that treats multi-agent execution as a constrained state machine. SDOF operates through two primary defensive layers, implemented by three components: (1) an Online-RLHF Specialized Intent Router trained via Generative Reward Modeling (GRPO) and (2) a StateAwareDispatcher with GoalStage finite-automaton checks and precondition/postcondition SkillRegistry validation for auditable execution control. On a recruitment system backed by the Beisen iTalent platform (6000+ enterprises), 185 expert-curated scenarios trigger 1671 live API calls. Our GSPO-aligned 7B Intent Router achieves higher joint accuracy than zero-shot GPT-4o on this FSM-constrained adversarial routing benchmark (80.9% versus 48.9%). In end-to-end execution, SDOF reaches 86.5% task completion (95% confidence interval 80.8 to 90.7) and blocks all 22 operations in the injection, illegal HR subset. Under a broader message-level blocking audit, SDOF attains precision 100% and recall 88%, expert agreement kappa=0.94. A separate evaluation on 960 SGD-derived dialogues spanning 8 service domains surfaces 201 stage-order conflicts under our FSM mapping, 41 of which arise in the normal split. This arXiv version reports the current validated scope; extended multi-seed training comparisons and deeper workflow evaluations will be released in a subsequent update.\n\n---\n\n## 51. DeepSlide: From Artifacts to Presentation Delivery\n- **Authors**: Ming Yang , Zhiwei Zhang , Jiahang Li , Haoseng Liu , Yuzheng Cai , Weiguo Zheng\n- **URL**: [https://arxiv.org/abs/2605.15202](https://arxiv.org/abs/2605.15202)\n- **Abstract**:\n> Presentations are a primary medium for scholarly communication, yet most AI slide generators optimize the artifact (a visually plausible deck) while under-optimizing the delivery process (pacing, narrative, and presentation preparation). We present DeepSlide, a human-in-the-loop multi-agent system that supports preparing the full presentation process, from requirement elicitation and time-budgeted narrative planning, to evidence-grounded slide--script generation, attention augmentation, and rehearsal support. DeepSlide integrates (i) a controllable logical-chain planner with per-node time budgets, (ii) a lightweight content-tree retriever for grounding, (iii) Markov-style sequential rendering with style inheritance, and (iv) sandboxed execution with minimal repair to ensure renderability. We further introduce a dual-scoreboard benchmark that cleanly separates static artifact quality from dynamic delivery excellence. Across 20 domains and diverse audience profiles, DeepSlide matches strong baselines on artifact quality while consistently achieving larger gains on delivery metrics, improving narrative flow, pacing precision, and slide--script synergy with clearer attention guidance.\n\n---\n\n## 52. IVGT: Implicit Visual Geometry Transformer for Neural Scene Representation\n- **Authors**: Yuqi Wu , Tianyu Hu , Wenzhao Zheng , Yuanhui Huang , Haowen Sun , Jie Zhou , Jiwen Lu\n- **URL**: [https://arxiv.org/abs/2605.16258](https://arxiv.org/abs/2605.16258)\n- **Abstract**:\n> Reconstructing coherent 3D geometry and appearance from unposed multi-view images is a fundamental yet challenging problem in computer vision. Most existing visual geometry foundation models predict explicit geometry by regressing pixel-aligned pointmaps, often suffering from redundancy and limited geometric continuity. We propose IVGT, an Implicit Visual Geometry Transformer that implicitly models continuous and coherent geometry from pose-free multi-view images. This formulation learns a continuous neural scene representation in a canonical coordinate system and supports continuous spatial queries at any 3D positions, retrieving local features to predict signed distance (SDF) values and colors using lightweight decoders. It allows direct extraction of continuous and coherent surface geometry, enabling rendering of RGB images, depth maps, and surface normal maps from arbitrary viewpoints. We train IVGT via multi-dataset joint optimization with 2D supervision and 3D geometric regularization. IVGT demonstrates generalization across scenes and achieves strong performance on various tasks, including mesh and point cloud reconstruction, novel view synthesis, depth and surface normal estimation, and camera pose estimation.\n\n---\n\n## 53. Designing Datacenter Power Delivery Hierarchies for the AI Era\n- **Authors**: Grant Wilkins , Fiodar Kazhamiaka , Alok Gautam Kumbhare , Chaojie Zhang , Ricardo Bianchini\n- **URL**: [https://arxiv.org/abs/2605.16255](https://arxiv.org/abs/2605.16255)\n- **Abstract**:\n> Demand for AI accelerators is rapidly increasing rack power density, with projections approaching 1MW per deployment by 2027. This poses a major challenge for datacenter power delivery designers. As power densities increase, a datacenter designed for a different target density may strand power, i.e., may be unable to use all the power that its delivery hierarchy has provisioned. Designs must remain efficient over long datacenter lifetimes and multiple hardware generations. Power utilization is particularly important as grid power capacity is a scarce resource in the AI era. Designing an efficient power delivery hierarchy for the long run is difficult because rack placement feasibility, workload impact, and cost depend jointly on electrical topology, deployment granularity, placement policy, power oversubscription, and workload mix. Moreover, each of these factors evolve over time, have inter-dependencies across multiple resource dimensions, and generally do not lend themselves to closed-form analysis. To address this challenge, we develop a framework for evaluating datacenter power delivery designs using throughput, power, and cost metrics over realistic arrival, oversubscription, and decommissioning sequences. The framework combines projection models for GPU, compute, and storage deployments with operational factors grounded in production data from Microsoft Azure. Our results show that multi-resource stranding materially changes deployable capacity, effective capital expenditure, and delivered performance, and quantify how rising density from rack- and pod-scale AI systems shapes these outcomes. For AI datacenter design, the relevant planning objective is not installed megawatts, but deployable capacity over time.\n\n---\n\n## 54. A Generative AI Framework for Intelligent Utility Billing CO 2 Analytics and Sustainable Resource Optimisation\n- **Authors**: Pavan Manjunath , Thomas Pruefer\n- **URL**: [https://arxiv.org/abs/2605.16250](https://arxiv.org/abs/2605.16250)\n- **Abstract**:\n> Distribution utilities are now expected to deliver bills that customers can actually read attach a defensible carbon number to every kWh sold and schedule load against grid stress and emissions constraints We propose an end-to-end framework that unifies four production-grade capabilities under one architectural roof a generative-AI agent that drafts each customers natural-language billing statement from structured numeric inputs under a constrained decoding policy a transformer-based forecaster that supplies the day-ahead consumption estimate with calibrated quantile bands\n\n---\n\n## 55. AI-Mediated Communication Can Steer Collective Opinion\n- **Authors**: Stratis Tsirtsis , Kai Rawal , Chris Russell , Brent Mittelstadt , Sandra Wachter\n- **URL**: [https://arxiv.org/abs/2605.16245](https://arxiv.org/abs/2605.16245)\n- **Abstract**:\n> Generative artificial intelligence (AI) is increasingly integrated into the online platforms where humans exchange opinions; large language models (LLMs) now polish users' posts on LinkedIn and provide context for content shared on X. While prior work has shown that AI can express biased opinions and shape individuals' opinions during human-AI interactions, less attention has been paid to its influence on collective opinion formation when mediating human-to-human communication. We address this gap via a combination of empirical and theoretical analyses. We show empirically that LLMs from multiple popular families introduce directional biases when instructed to edit human-written texts on contested topics, for example, nudging texts in favor of gun control and against atheism. Building on this observation, we introduce a mathematical model of opinion dynamics in which an AI system sits between users on a social network, transforming the opinions they express and perceive. By analytically characterizing the equilibrium of this model and performing simulations on real social network data, we show that biases introduced by AI in human-to-human communication can be amplified through the network and shift collective opinion in their direction. In light of these findings, we investigate whether such biases are controllable by online platforms. We audit the \"Explain this post\" feature on X and find evidence of pro-life bias in Grok's outputs on abortion-related content, which we trace back to specific design choices. We conclude with a discussion of the broader implications of our findings in relation to ongoing legislative efforts in the European Union.\n\n---\n\n## 56. Offline Semantic Guidance for Efficient Vision-Language-Action Policy Distillation\n- **Authors**: Jin Shi , Brady Zhang , Yishun Lu\n- **URL**: [https://arxiv.org/abs/2605.16241](https://arxiv.org/abs/2605.16241)\n- **Abstract**:\n> Billion-parameter Vision-Language-Action (VLA) policies have recently shown impressive performance in robotic manipulation, yet their size and inference cost remain major obstacles for real-time closed-loop control. We introduce \\textbf{VLA-AD}, a distillation framework that uses a Vision-Language Model as an offline semantic supervisor to transfer large VLA teachers into lightweight student policies. Instead of relying only on low-level action imitation, VLA-AD augments teacher-provided 7-DoF action targets with high-level semantic guidance, including task phase anchors and multi-frame operating-direction descriptions. These auxiliary signals are used only during training: at test time, the student policy runs independently, with neither the VLA teacher nor the VLM required. We evaluate VLA-AD on three LIBERO benchmark suites. Using OpenVLA-7B as the teacher, our method produces a 158M-parameter student, yielding a $44\\times$ reduction in model size while matching the teacher with only a $0.27\\%$ average relative gap. The resulting policy runs at 12.5 Hz on an RTX 4090, achieving a $3.28\\times$ inference speedup over OpenVLA-7B. We further show that the same semantic distillation pipeline generalizes to a different $\\pi_{0.5}$-4B teacher, where the student outperforms the teacher on two suites and remains within $0.53\\%$ on \\texttt{libero\\_goal}. Additional analysis indicates that phase-level supervision and multi-frame directional cues make the student less sensitive to noisy teacher actions, such as erroneous high-frequency gripper changes. Overall, VLA-AD demonstrates that offline semantic guidance from VLMs can substantially improve the efficiency, robustness, and deployability of VLA policy distillation.\n\n---\n\n## 57. Layer Equivalence Is Not a Property of Layers Alone: How You Test Redundancy Changes What You Find\n- **Authors**: Gabriel Garcia\n- **URL**: [https://arxiv.org/abs/2605.16234](https://arxiv.org/abs/2605.16234)\n- **Abstract**:\n> When researchers ask whether two transformer layers are \"equivalent\" for compression, they often conflate distinct tests. Replacement asks whether one layer's map can substitute for another's in place; interchange asks whether two layers approximately commute when their positions are swapped. Both are output-grounded swap-KL probes, but they need not agree: on pretrained transformers the protocol gap can change which layers look safe to prune by several-fold under the same evaluator, especially when replacement distances are high. We measure both protocols across checkpoints and architectures. On a Pythia training trajectory (410M and 1.4B), the replacement-interchange gap grows from initialization to convergence. Under one matched WikiText-2 contract at 8B scale, Qwen3-8B enters a divergent regime: interchange-guided removal is several-fold safer than replacement-guided at the same layer budgets, while Llama-3.1-8B ties the two protocols for pruning cost even though interchange KL is lower, showing metric gaps need not map one-to-one to removal. Before layer removal or merging, score both swap-KLs on the target checkpoint; the diagnostic requires only unlabeled forward passes.\n\n---\n\n## 58. A Unified Generative-AI Framework for Smart Energy Infrastructure: Intelligent Gas Distribution, Utility Billing, Carbon Analytics, and Quantum-Inspired Optimisation\n- **Authors**: Pavan Manjunath , Thomas pruefer\n- **URL**: [https://arxiv.org/abs/2605.16232](https://arxiv.org/abs/2605.16232)\n- **Abstract**:\n> The accelerating convergence of smart metering, generative artificial intelligence, and quantum-inspired combinatorial optimisation is reshaping how energy utilities manage physical infrastructure, customer engagement, and environmental accountability\n\n---\n\n## 59. Evaluating Design Video Generation: Metrics for Compositional Fidelity\n- **Authors**: Adrienne Deganutti , Dingning Cao , Jaejung Seol , Elad Hirsch , Purvanshi Mehta\n- **URL**: [https://arxiv.org/abs/2605.16223](https://arxiv.org/abs/2605.16223)\n- **Abstract**:\n> Generative video models are increasingly used in design animation tasks, yet no standardized evaluation framework exists for this domain. Unlike natural video generation, design animation imposes structured constraints: specific components shall animate with prescribed motion types, directions, speed and timing, while non-animated regions must remain stable and layout structure must be preserved. This paper provides a fully automated evaluation framework organized across four dimensions: layout fidelity, motion correctness, temporal quality, and content fidelity. This eliminates the reliance on subjective human evaluation and establishes a common basis for benchmarking progress in the field.\n\n---\n\n## 60. Argus: Evidence Assembly for Scalable Deep Research Agents\n- **Authors**: Zhen Zhang , Liangcai Su , Zhuo Chen , Xiang Lin , Haotian Xu , Simon Shaolei Du , Kaiyu Yang , Bo An , Lidong Bing , Xinyu Wang\n- **URL**: [https://arxiv.org/abs/2605.16217](https://arxiv.org/abs/2605.16217)\n- **Abstract**:\n> Deep research agents have achieved remarkable progress on complex information seeking tasks. Even long ReAct style rollouts explore only a single trajectory, while recent state of the art systems scale inference time compute via parallel search and aggregation. Yet deep research answers are composed of complementary pieces of evidence, which parallel rollouts often duplicate rather than complete, yielding diminishing returns while pushing the aggregation context toward the model's limit. We propose Argus, an agentic system in which a Searcher and a Navigator cooperate to treat deep research as assembling a jigsaw from complementary evidence pieces, rather than brute forcing the whole answer in parallel. The Searcher collects evidence traces for a given sub-query through ReAct-style interaction. The Navigator maintains a shared evidence graph, verifying which pieces are still missing, dispatching Searchers to gather them, and reasoning over the completed graph to produce a source-traced final answer. We train the Navigator with reinforcement learning to verify, dispatch, and synthesize, while independently training the Searcher to remain a standard ReAct agent. The resulting Navigator supports rollouts with a single Searcher or many in parallel without retraining. With both Searcher and Navigator built on a 35B-A3B MoE backbone, Argus gains 5.5 points with a single Searcher and 12.7 points with 8 parallel Searchers, averaged over eight benchmarks. With 64 Searchers it reaches 86.2 on BrowseComp, surpassing every proprietary agent we benchmark, while the Navigator's reasoning context stays under 21.5K tokens.\n\n---\n\n## 61. paper.json: A Coordination Convention for LLM-Agent-Actionable Papers\n- **Authors**: Arquimedes Canedo\n- **URL**: [https://arxiv.org/abs/2605.16194](https://arxiv.org/abs/2605.16194)\n- **Abstract**:\n> LLM agents routinely serve as first (and sometimes only) readers of academic papers, skimming for sub-claims, extracting reproducibility steps, and generalizing scope. Standard prose papers produce recurring failures in this role: sub-claims that cannot be cited at sub-paper granularity, scope overextension beyond what the paper tests, and figure commands buried in codebases rather than the paper itself. We propose ` this http URL `, a companion JSON file that travels with the PDF and addresses each failure with a lightweight convention: stable claim IDs (C1), an explicit does-not-claim list (C2), exact per-figure shell commands (C3), and stable definition IDs (C5). A fifth convention (C4) holds that minimum viable compliance, hand-written JSON alongside the PDF, is achievable in under an hour for a finished paper without touching the human-readable output. C1, C2, C3, and C5 are open invitations: an agent that reads a compliant paper and acts on it produces evidence for or against them. This paper is itself compliant: `uv run this http URL this http URL --against this http URL ` passes. Repo: this https URL\n\n---\n\n## 62. Second-Order Multi-Level Variance Correction for Modality Competition in Multimodal Models\n- **Authors**: Yishun Lu , Wes Armour\n- **URL**: [https://arxiv.org/abs/2605.16165](https://arxiv.org/abs/2605.16165)\n- **Abstract**:\n> Autoregressive next-token training offers a unified formulation for image generation and text understanding, but it also creates strong modality competition that destabilizes optimization and limits large-batch scaling. We show that first-order optimizers such as AdamW are vulnerable to cross-modality gradient heterogeneity, while second-order preconditioning, particularly SOAP, provides a more stable basis for multimodal alignment. Building on this insight, we propose \\emph{ML-FOP-SOAP}, a second-order optimization framework with Multi-Level Variance Correction. Our Fisher-Orthogonal Projection suppresses variance-induced modality conflicts, reducing the trade-off between visual generation and textual understanding. To make this practical under large gradient accumulation, we introduce a hierarchical folding strategy that captures fine-grained variance with low micro-step overhead. Experiments on Janus and Emu3 show consistent gains across both modalities and stable training at batch size 8192. Compared with AdamW, our method improves sample efficiency by up to $1.4\\times$ and accelerates wall-clock training by up to $1.5\\times$, offering a robust optimizer for scaling multimodal foundation models.\n\n---\n\n## 63. Surrogate Neural Architecture Codesign Package (SNAC-Pack)\n- **Authors**: Jason Weitz , Dmitri Demler , Benjamin Hawks , Aaron Wang , Nhan Tran , Javier Duarte\n- **URL**: [https://arxiv.org/abs/2605.16138](https://arxiv.org/abs/2605.16138)\n- **Abstract**:\n> Neural architecture search (NAS) is a powerful approach for automating model design, but existing methods often optimize for accuracy alone or rely on proxy metrics such as bit operations (BOPs) that correlate poorly with hardware cost. This gap is particularly large for FPGA deployment, where cost is dominated by a multi-dimensional budget of lookup tables, DSPs, flip-flops, BRAM, and latency. We present the Surrogate Neural Architecture Codesign Package (SNAC-Pack), an open-source AutoML framework for hardware-aware neural architecture codesign and end-to-end FPGA deployment. SNAC-Pack runs a multi-objective global search with Optuna and NSGA-II, loading trials to a shared SQLite store that enables parallel workers across compute nodes. A hardware surrogate model outputs per-trial resource and latency estimates, avoiding the synthesis cost that would otherwise dominate the search loop. A local search stage then applies quantization-aware training (QAT) together with iterative magnitude pruning in a combined compression loop, after which the final model is synthesized to FPGA firmware via the hls4ml Python library. A YAML configuration and an optional agentic frontend let users run the pipeline on new datasets without modifying the framework. We demonstrate SNAC-Pack on jet classification at the Large Hadron Collider and superconducting qubit readout, discovering compact architectures that match or exceed strong baselines on the task metric while reducing FPGA resource utilization and, in the qubit readout case, reducing the design space exploration process from months of manual fine-tuning to hours of automated search.\n\n---\n\n## 64. Navigating Potholes with Geometry-Aware Sharpness Minimization\n- **Authors**: Simon Dufort-Labbé , Mehrab Hamidi , Razvan Pascanu , Ioannis Mitliagkas , Damien Scieur , Aristide Baratin\n- **URL**: [https://arxiv.org/abs/2605.16134](https://arxiv.org/abs/2605.16134)\n- **Abstract**:\n> Sharpness-aware minimization (SAM) encourages flat minima by perturbing parameters along directions of high loss curvature, but treats all parameter directions uniformly, ignoring the underlying loss geometry. We introduce LLQR+SAM, which combines SAM with a learned preconditioner obtained from the recently proposed LLQR framework, a second-order method that recasts steepest descent as a layerwise linear-quadratic regulator problem. The preconditioner is updated sparsely and maintained as a slow exponential moving average, so it captures a smoothed, low-resolution picture of the loss landscape geometry. The SAM perturbation then operates on top of this learned geometry, probing curvature at a faster timescale. We show that this two-timescale structure is not merely a computational convenience: theoretically, the preconditioner amplifies the SAM escape signal in directions that are flat under the average geometry but locally sharp (potholes). Wide, flat basins, by contrast, remain stable. Empirically, LLQR+SAM gives consistent gains over both SAM and LLQR alone across standard vision and sequence modeling benchmarks, supporting the view that slow learned geometry and fast sharpness correction are genuinely complementary.\n\n---\n\n## 65. Entropy Across the Bridge: Conditional-Marginal Discretization for Flow and Schrödinger Samplers\n- **Authors**: Bruno Trentini , Dejan Stancevic , Michael M. Bronstein , Alexander Tong , Luca Ambrogioni\n- **URL**: [https://arxiv.org/abs/2605.16126](https://arxiv.org/abs/2605.16126)\n- **Abstract**:\n> For a fixed flow-based generative model under a small inference budget, sample quality can depend strongly on where the sampler spends its few function evaluations. Flow matching and Schrödinger bridges define probability paths, yet their inference grids are usually heuristic or inherited from one-endpoint diffusion. We derive a conditional-marginal entropy-rate objective for bridge-aware discretization, separating endpoint-conditioned bridge geometry from marginal flow evolution, and use it to build a training-free entropic inference-time scheduler from first principles. For Gaussian Brownian bridges this rate is closed-form and U-shaped, motivating boundary-heavy nonuniform grids. On trained two-dimensional bridge/flow models, the estimated profile recovers the predicted shape and improves 10-step ODE-Heun MMD over linear by 18.1%, with a paired 22.7% SDE-Heun improvement in the same low-NFE sweep. On EDM/CIFAR-10, the entropic time-discretization gives the best tested five-step FID (186.3 \\pm 4.0 versus 200.5 \\pm 2.9 for linear and 238.0 \\pm 5.3 for cosine). On AlphaFlow protein generation, entropic conditional-marginal (cond-marg) scheduling shows advantage in low-NFE regimes on both CAMEO22 and ATLAS benchmarks. These results support entropy-rate scheduling as a practical low-budget allocation signal for high-dimensional bridge and flow samplers.\n\n---\n\n## 66. GenShield: Unified Detection and Artifact Correction for AI-Generated Images\n- **Authors**: Zhipei Xu , Xuanyu Zhang , Youmin Xu , Qing Huang , Shen Chen , Taiping Yao , Shouhong Ding , Jian Zhang\n- **URL**: [https://arxiv.org/abs/2605.16122](https://arxiv.org/abs/2605.16122)\n- **Abstract**:\n> Diffusion-based image synthesis has made AI-generated images (AIGI) increasingly photorealistic, raising urgent concerns about authenticity in applications such as misinformation detection, digital forensics, and content moderation. Despite the substantial advances in AIGI detection, how to correct detected AI-generated images with visible artifacts and restore realistic appearance remains largely underexplored. Moreover, few existing work has established the connection between AIGI detection and artifact correction. To fill this gap, we propose GenShield, a unified autoregressive framework that jointly performs explainable AIGI detection and controllable artifact correction in a closed loop from diagnosis to restoration, revealing a mutually reinforcing relationship between these two tasks. We further introduce a Visual Chain-of-Thought based curriculum learning strategy that enables self-explained, multi-step ``diagnose-then-repair'' correction with an explicit stopping criterion. A high-quality dataset with large-scale ``artifact-restored'' pairs is also constructed alongside a unified evaluation pipeline. Extensive experiments on our correction benchmark and mainstream AIGI detection benchmarks demonstrate state-of-the-art performance and strong generalization of our method. The code is available at this https URL .\n\n---\n\n## 67. DebiasRAG: A Tuning-Free Path to Fair Generation in Large Language Models through Retrieval-Augmented Generation\n- **Authors**: Rui Chu , Bingyin Zhao , Thanh Quoc Hung Le , Duy Cao Hoang , Huawei Lin , Ping Li , Weijie Zhao , Khoa D Doan , Yingjie Lao\n- **URL**: [https://arxiv.org/abs/2605.16113](https://arxiv.org/abs/2605.16113)\n- **Abstract**:\n> Large language models (LLMs) have achieved unprecedented success due to their exceptional generative capabilities. However, because they depend on knowledge encapsulated from training corpora, they may produce hallucinations, stereotypes, and socially biased content. In particular, LLMs are prone to prejudiced responses involving race, gender, and age, which are collectively referred to as social biases. Prior studies have used fine-tuning and prompt engineering to mitigate such biases in LLMs, but these methods require additional training resources or domain knowledge to design the framework. Moreover, they may degrade the original capabilities of LLMs and often overlook the need for dynamic debiasing contexts for fairer inference. In this paper, we propose DebiasRAG, a novel tuning-free and dynamic query-specific debiasing framework based on retrieval-augmented generation (RAG). DebiasRAG improves fairness while preserving the intrinsic properties of LLMs, such as representation ability. DebiasRAG consists of three stages: (1) query-specific debiasing candidate generation; (2) context candidate pool construction; and (3) gradient-updated debiasing-guided context piece reranking. First, DebiasRAG leverages self-diagnosed bias contexts relevant to the query through regular retrieval, where the bias contexts are prepared offline by the DebiasRAG provider. Given the query-specific bias contexts, DebiasRAG reversely produces debiasing contexts, which are provided as additional fairness constraints for LLM outputs. Second, a regular RAG retrieval process produces query-related contexts from the regular RAG document database, such as a chunked Wikipedia dataset.\n\n---\n\n## 68. Attention Dispersion in Dynamic Graph Transformers: Diagnosis and a Transferable Fix\n- **Authors**: Jinhao Zhang , Kangfei Zhao , Qiuhao Zeng , Long-Kai Huang\n- **URL**: [https://arxiv.org/abs/2605.16112](https://arxiv.org/abs/2605.16112)\n- **Abstract**:\n> Transformer-based architectures have become the dominant paradigm for Continuous-Time Dynamic Graph (CTDG) learning, yet their performance remains limited on temporally shifted datasets. In this work, we identify attention dispersion as a shared failure mode of dynamic graph Transformers under temporal distribution shift. Through controlled ablation contrasting structurally and temporally distinguished historical neighbors against random ones, we show that prediction depends on a class of critical nodes that carry consistently more predictive signal than arbitrary neighbors. However, existing Transformers fail to focus on these nodes even when they are present in the input, as temporal shift weakens attention contrast and produces overly dispersed attention distributions. This diagnosis suggests a simple and transferable fix: replace standard attention with differential attention, which suppresses common-mode attention and amplifies distinctive token-level signals. When added to three representative CTDG Transformer baselines, differential attention consistently improves performance, with gains concentrated on high-shift datasets. Attention-level measurements further confirm the mechanism, showing reduced attention entropy and increased attention mass on critical nodes. Building on these findings, we introduce DiffDyG, a reference implementation combining differential attention with standard input encodings. Across 9 benchmarks and three negative sampling protocols, DiffDyG achieves SOTA performance, with especially large gains on the most shifted datasets.\n\n---\n\n## 69. Federated Imputation under Heterogeneous Feature Spaces\n- **Authors**: Imane Hocine , Chaimaa Medjadji , Sylvain Kubler , Gregoire Danoy , Yves Le Traon\n- **URL**: [https://arxiv.org/abs/2605.16099](https://arxiv.org/abs/2605.16099)\n- **Abstract**:\n> Federated Learning (FL) enables collaborative training across decentralized clients, but most methods assume aligned feature schemas, an assumption that rarely holds in tabular settings where clients observe only partially overlapping feature subsets. In these heterogeneous feature spaces, parameter-averaging methods (e.g., FedAvg) transfer little information across weakly overlapping or disjoint feature groups, limiting their effectiveness for federated imputation. To overcome this, we propose \\textbf{FedHF-Impute}, a federated imputation framework that separates structural feature unavailability from conventional missingness and uses a shared global feature graph to propagate information across statistically related features through message passing. This enables indirect cross-client knowledge transfer, even when features are never jointly observed locally, while preserving standard federated communication. Under simulated partial schema overlap on the SECOM and AirQuality datasets, FedHF-Impute improves imputation accuracy (RMSE) over FL baselines by 26.9\\%, and 8.4\\% respectively, while achieving comparable performance on PhysioNET, with only a 0.3\\% difference relative to the best baseline.\n\n---\n\n## 70. GeoGS-CE: Learning Delay--Beam Channel Priors with 3D Gaussians for High-Mobility Scenarios\n- **Authors**: Yumeng Zhang , Jiajia Guo , Chaozheng Wen , Chenghong Bian , Jun Zhang\n- **URL**: [https://arxiv.org/abs/2605.16094](https://arxiv.org/abs/2605.16094)\n- **Abstract**:\n> Wideband channel estimation (CE) in high-mobility scenarios remains challenging because channel responses vary rapidly, while practical systems can allocate only sparse pilots to accommodate dense users. Fortunately, many high-mobility environments, such as high-speed railways, exhibit scheduled trajectories, predictable velocities, and a limited number of dominant propagation paths. These properties induce a delay--beam power spectrum that is more stable than the instantaneous complex channel frequency response (CFR), less sensitive to the random phase coherence, and rich in geometric information. To exploit such environmental properties, we propose GeoGS-CE, a two-stage channel estimation framework for sparse-pilot high-mobility scenarios. In the offline stage, GeoGS-CE jointly models: 1) a scene-level 3D Gaussian representation that captures the non-line-of-sight (NLoS) geometric scattering support, and 2) a leakage-aware differentiable wireless rendering process that maps the NLoS Gaussians, together with an explicit virtual line-of-sight (LoS) component, to the measured delay--beam power spectrum, while accounting for practical OFDM delay and array leakage effects. In the online stage, the delay--beam power spectrum is predicted for each user location and used as a strong covariance prior, enabling accurate full-band and full-array CFR reconstruction and tracking through a linear MMSE estimator. Simulations based on channels generated from a segment of the Guangshen high-speed railway show that the proposed geometric prior substantially improves CFR reconstruction over pilot-only and non-geometric baselines.\n\n---\n\n## 71. Centralized vs Decentralized Federated Learning: A trade-off performance analysis\n- **Authors**: Chaimaa Medjadji , Guilain Leduc , Sylvain Kubler , Yves Le Traon\n- **URL**: [https://arxiv.org/abs/2605.16089](https://arxiv.org/abs/2605.16089)\n- **Abstract**:\n> Federated Learning (FL) has emerged as a promising paradigm for collaborative model training across distributed edge devices while preserving data privacy especially with the huge increase amount of data due to the adoption of technologies which contributes to the growing number of IoT devices. Storing this amount of data centrally is challenging due to issues like limited communication, privacy, and regulations. FL can be Centralized (CFL), Decentralized (DFL), and Semi-decentralized (SDFL). Choosing the right FL architecture depends on the application's needs. However, very few research studies have experimentally compared these three types of architectures to not only understand the respective strengths and limitations, but also trade-offs between different performance indicators. This paper overcome this lack of analysis, conducting experimental analyses using the Fedstellar simulator, MNIST dataset, and MLP classifier.\n\n---\n\n## 72. Multi-level Self-supervised Pretraining on Compositional Hierarchical Graph for Molecular Property Prediction\n- **Authors**: Xiayu Liu , Zhengyi Lu , Hou-biao Li\n- **URL**: [https://arxiv.org/abs/2605.16088](https://arxiv.org/abs/2605.16088)\n- **Abstract**:\n> Self-supervised pretraining on molecular graphs has emerged as a promising approach for molecular property prediction, yet most existing methods operate at a single structural granularity and treat bond information as auxiliary edge attributes rather than as an independent semantic layer. In this work, we propose MolCHG, a multi-level self-supervised pretraining framework built upon a novel Compositional Hierarchical Graph that organizes molecular structure into four types of nodes across three semantic levels. By introducing a bond graph that operates in parallel with the atom graph, our architecture elevates bond-level information to independently evolving node representations, enabling fragment nodes to aggregate atom-level and bond-level semantics on an equal footing. We design three level-specific pretraining objectives: an atom-bond cross-view contrastive task that aligns the atom-view and bond-view representations within each fragment, a fragment-level functional group prediction task to inject domain-relevant chemical knowledge, and graph-level structure prediction tasks to encode global molecular topology. Experiments on nine MoleculeNet benchmarks demonstrate that MolCHG achieves the best performance on seven datasets across both classification and regression tasks, remaining competitive with the strongest baselines on the rest. Ablation studies further confirm that the multi-level supervision signals are complementary and that each component contributes to the overall performance.\n\n---\n\n## 73. Towards Trustworthy and Explainable AI for Perception Models: From Concept to Prototype Vehicle Deployment\n- **Authors**: Till Beemelmanns , Shayan Sharifi , Manas Mehrotra , Ayushman Choudhuri , Lutz Eckstein\n- **URL**: [https://arxiv.org/abs/2605.16087](https://arxiv.org/abs/2605.16087)\n- **Abstract**:\n> Deep Neural Networks have become the dominant solution for Autonomous Driving perception, but their opacity conflicts with emerging Trustworthy AI guidelines and complicates safety assurance, debugging, and human oversight. While theoretical frameworks for safe and Explainable AI (XAI) exist, concrete implementations of Trustworthy AI for 3D scene understanding remain scarce. We address this gap by proposing a Trustworthy AI perception module that is remarkably robust, integrates faithful explainability, and calibrated uncertainty estimates. Building on a transformer-based detector, we derive explanation from the attention mechanism at inference time and validate their faithfulness using perturbation-based consistency tests. We further integrate an uncertainty estimation and calibration module, and apply robustness-enhancing training methods. Experiments show faithful saliency behavior, improved robustness, and well-calibrated uncertainty estimates. Finally, we deploy these Trustworthy AI elements in a prototype vehicle and provide an XAI Interface that visualizes documentation artifacts, model uncertainty state, and saliency maps, demonstrating the feasibility of trustworthy perception monitoring in real time. Supplementary materials are available at this https URL .\n\n---\n\n## 74. Towards Foundation Models for Relational Databases with Language Models and Graph Neural Networks\n- **Authors**: Jingcheng Wu , Ratan Bahadur Thapa , Mojtaba Nayyeri , Lucas Etteldorf , Max Finkenbeiner , Fabian Leeske , Steffen Staab\n- **URL**: [https://arxiv.org/abs/2605.16085](https://arxiv.org/abs/2605.16085)\n- **Abstract**:\n> Relational databases store much of the world's structured information, and they are essential for driving complex predictive applications. However, deep learning progress on relational data remains limited, as conventional approaches flatten databases into single tables via manual feature engineering, discarding relational context. Relational deep learning (RDL) addresses this by modeling databases as relational entity graphs (REGs) for graph neural networks (GNNs), but remains task- and database-specific. To combine the strengths of both paradigms, we propose a hybrid architecture combining a fine-tuned BART encoder to capture intra-row semantics with a GraphSAGE-based GNN over REGs to inject relational context. Experiments on RelBench show that the GNN substantially enriches BART's row embeddings, achieving a ROC-AUC of 67.40 on the driver-dnf task from the rel-f1 dataset. This performance is competitive with supervised baselines such as LightGBM (68.86) and narrows the gap to RDL (72.62) to within 5.22 points, though a substantial gap remains to state-of-the-art foundation models such as KumoRFM (82.63). These results suggest that lightweight hybrid LM-GNN architectures offer a promising and resource-efficient path towards foundation models for relational databases.\n\n---\n\n## 75. VideoSeeker: Incentivizing Instance-level Video Understanding via Native Agentic Tool Invocation\n- **Authors**: Yiming Zhao , Yu Zeng , Wenxuan Huang , Zhen Fang , Qing Miao , Qisheng Su , Jiawei Zhao , Jiayin Cai , Lin Chen , Zehui Chen , Yukun Qi , Yao Hu , Xiaolong Jiang , Feng Zhao\n- **URL**: [https://arxiv.org/abs/2605.16079](https://arxiv.org/abs/2605.16079)\n- **Abstract**:\n> Large Vision-Language Models (LVLMs) have shown significant progress in video understanding, yet they face substantial challenges in tasks requiring precise spatiotemporal localization at the instance level. Existing methods primarily rely on text prompts for human-model interaction, but these prompts struggle to provide precise spatial and temporal references, resulting in poor user experience. Furthermore, current approaches typically decouple visual perception from language reasoning, centering reasoning around language rather than visual content, which limits the model's ability to proactively perceive fine-grained visual evidence. To address these challenges, we propose VideoSeeker, a novel paradigm for instance-level video understanding through visual prompts. VideoSeeker seamlessly integrates agentic reasoning with instance-level video understanding tasks, enabling the model to proactively perceive and retrieve relevant video segments on demand. We construct a four-stage fully automated data synthesis pipeline to efficiently generate large-scale, high-quality instance-level video data. We internalize tool-calling and proactive perception capabilities into the model via cold-start supervision and RL training, building a powerful video understanding model. Experiments demonstrate that our model achieves an average improvement of +13.7% over baselines on instance-level video understanding tasks, surpassing powerful closed-source models such as GPT-4o and Gemini-2.5-Pro, while also showing effective transferability on general video understanding benchmarks. The relevant datasets and code will be released publicly.\n\n---\n\n## 76. AgriMind: An Ensemble Deep Learning Framework for Multi-Class Plant Disease Classification\n- **Authors**: Salma Hoque Talukdar Koli , Fahima Haque Talukder Jely\n- **URL**: [https://arxiv.org/abs/2605.16076](https://arxiv.org/abs/2605.16076)\n- **Abstract**:\n> Plant disease detection is still largely manual in Bangladesh, where extension workers eyeball leaf samples across millions of smallholdings. We built AgriMind to automate this: an ensemble of ResNet50, EfficientNet-B0, and DenseNet121 trained on 20,638 PlantVillage images across 15 pepper, potato, and tomato disease classes. Transfer learning with frozen ImageNet backbones and 10 epochs of head-only training keeps the pipeline lightweight. Individual models hit 96--97% on the held-out test set, but averaging their softmax outputs pushes the ensemble to 99.23% -- a two-thirds cut in error rate. We tried biasing the average toward the best validation model; it backfired. Dropping any single model also hurt. Pepper and potato classify perfectly; tomato, with ten visually similar classes, still reaches 99.01%. On an NVIDIA T4 GPU the full ensemble runs at 53 FPS. Whether that translates to real-time mobile use depends on TensorFlow Lite optimization -- work we have not yet completed.\n\n---\n\n## 77. Robust Prior-Guided Segmentation for Editable 3D Gaussian Splatting\n- **Authors**: Raushan Joshi , Jean-Yves Guillemaut\n- **URL**: [https://arxiv.org/abs/2605.16065](https://arxiv.org/abs/2605.16065)\n- **Abstract**:\n> 3D Gaussian Splatting (3D-GS) enables real-time 3D scene reconstruction but lacks robust segmentation for editing tasks such as object removal, extraction, and recoloring. Existing approaches that lift 2D segmentations to the 3D domain suffer from view inconsistencies and coarse masks. In this paper, we propose a novel framework that leverages the Segment Anything Model High Quality (SAM-HQ) to generate accurate 2D masks, addressing the limitations of the standard SAM in boundary fidelity and fine-structure preservation. To achieve robust 3D segmentation of any target object in a given scene, we introduce a prior-guided label reassignment method that assigns labels to 3D Gaussians by enforcing multiview consistency with learned priors. Our approach achieves state-of-the-art segmentation accuracy and enables interactive, real-time object editing while maintaining high visual fidelity. Qualitative results demonstrate superior boundary preservation and practical utility in Virtual Reality (VR) and robotics, advancing 3D scene editing.\n\n---\n\n## 78. Misspecified Explore-then-Exploit Leads to Supra-Competitive Prices\n- **Authors**: Jackie Baek , Vivek F. Farias , Farrell Wu\n- **URL**: [https://arxiv.org/abs/2605.16064](https://arxiv.org/abs/2605.16064)\n- **Abstract**:\n> We study whether simple algorithmic pricing systems can systematically produce collusive-like prices in multi-firm markets. We consider firms using an explore-then-exploit pipeline: they randomize prices during an initial exploration phase, then estimate demand from their own historical data and set prices myopically thereafter. The estimation step relies on a misspecified, monopoly-style model that omits competitors' prices. We characterize when this pipeline converges to supra-competitive prices above the Nash equilibrium, via a fluid-limit ordinary differential equation analysis. We show that supra-competitive prices arise when firms explore within similar price ranges on the same side of the Nash price. Moreover, prices can be substantially above the Nash price; we show that prices can reach monopoly levels under symmetric exploration. Simulations calibrated to a real multifamily rental market confirm that supra-competitive outcomes arise robustly beyond our theoretical assumptions, including under finite horizons, heterogeneous products, and nonlinear logit demand.\n\n---\n\n## 79. Ada-Diffuser: Latent-Aware Adaptive Diffusion for Decision-Making\n- **Authors**: Fan Feng , Selena Ge , Minghao Fu , Zijian Li , Yujia Zheng , Zeyu Tang , Yingyao Hu , Biwei Huang , Kun Zhang\n- **URL**: [https://arxiv.org/abs/2605.16054](https://arxiv.org/abs/2605.16054)\n- **Abstract**:\n> Recent work has framed decision-making as a sequence modeling problem using generative models such as diffusion models. Although promising, these approaches often overlook latent factors that exhibit evolving dynamics, elements that are fundamental to environment transitions, reward structures, and high-level agent behavior. Explicitly modeling these hidden processes is essential for both precise dynamics modeling and effective decision-making. In this paper, we propose a unified framework that explicitly incorporates latent dynamic inference into generative decision-making from minimal yet sufficient observations. We theoretically show that under mild conditions, the latent process can be identified from small temporal blocks of observations. Building on this insight, we introduce Ada-Diffuser, a causal diffusion model that learns the temporal structure of observed interactions and the underlying latent dynamics simultaneously, and furthermore, leverages them for planning and control. With a modular design, Ada-Diffuser supports both planning and policy learning tasks, enabling adaptation to latent variations in dynamics, rewards, and latent actions. Experiments on simulated control and robotic benchmarks demonstrate its effectiveness in accurate latent inference and adaptive policy learning.\n\n---\n\n## 80. Looped SSMs: Depth-Recurrence and Input Reshaping for Time Series Classification\n- **Authors**: Mónika Farsang , Ramin Hasani , Daniela Rus , Radu Grosu\n- **URL**: [https://arxiv.org/abs/2605.16048](https://arxiv.org/abs/2605.16048)\n- **Abstract**:\n> State Space Models (SSMs) are inherently recurrent along the sequence dimension, yet depth-recurrence - reusing the same block repeatedly across layers, as recently applied in looped transformers - has not been explored in this model family. We show that a looped SSM with $k$ parameters iterated $L$ times consistently closely matches or outperforms a standard SSM with $k \\cdot L$ independent parameters across four architectures (LRU, S5, LinOSS, LrcSSM) and six time series classification benchmarks, despite operating within a strictly smaller hypothesis space, as we formally establish. Since the larger model contains the looped model as a special case, this dominance cannot be explained by expressivity and instead points to parameter sharing across depth as a beneficial inductive bias that simplifies optimization. These results demonstrate that depth-recurrence is orthogonal to sequence-recurrence and independently beneficial. We further show that input reshaping is an equally neglected design axis: concatenating timesteps for low-dimensional inputs, or flattening and rechunking the joint feature-time dimension for high-dimensional ones, yields accuracy gains of 1-6% across all models, confirmed over 5 random seeds. Both techniques provide standalone improvements that compound when combined, suggesting that depth and input reshaping are two independent and underexplored design axes for SSMs on time series.\n\n---\n\n## 81. XSearch: Explainable Code Search via Concept-to-Code Alignment\n- **Authors**: Yiming Liu , Ruofan Liu , Yun Lin , Zicong Zhang , Weiyu Kong , Pengnian Qi , Xiao Cheng , Weinan Zhang , Qianxiang Wang , Linpeng Huang\n- **URL**: [https://arxiv.org/abs/2605.16046](https://arxiv.org/abs/2605.16046)\n- **Abstract**:\n> Semantic code search has been widely adopted in both academia and industry. These approaches embed natural-language queries and code snippets into a shared embedding space and retrieve results based on vector similarity. Despit strong performance on benchmark datasets, they often suffer from poor explainability and generalization. Retrieved code may appear semantically similar yet miss critical functional requirements of the query, while providing no explanation of why the result was retrieved. Moreover, such failures become more severe under distribution shift, where models struggle to generalize to unseen benchmarks. In this work, we propose XSearch, an intrinsically explainable code search framework. Our key insight is that by relying on global embedding similarity, existing retrievers inherently take an inductive view. They learn statistical patterns rather than truly understanding the query's functional requirements. We address this problem by reformulating code search as a deductive concept alignment problem. XSearch (i) identifies functional concepts in the query and (ii) explicitly aligns them with corresponding code statements. This explain-then-predict design produces inherent concept-level explanations and mitigates shortcut learning that harms out-of-distribution generalization. We train an encoder with explicit concept-alignment objectives and perform retrieval through explicit matching between query concepts and code statements. Experiments show that, trained on CodeSearchNet using GraphCodeBERT (125M parameters), XSearch improves performance on out-of-distribution benchmarks from 0.02 to 0.33 (15x) over eight state-of-the-art retrievers, and consistently outperforms both encoder- and decoder-based baselines with up to 7B parameters. A user study demonstrates that concept-alignment explanations enable users to evaluate retrieved results faster and more accurately.\n\n---\n\n## 82. RecMem: Recurrence-based Memory Consolidation for Efficient and Effective Long-Running LLM Agents\n- **Authors**: Zijie Dai , Shiyuan Deng , Sheng Guan , Yizhou Tian , Xin Yao , Xiao Yan , James Cheng\n- **URL**: [https://arxiv.org/abs/2605.16045](https://arxiv.org/abs/2605.16045)\n- **Abstract**:\n> Memory systems often organize user-agent interactions as retrievable external memory and are crucial for long-running agents by overcoming the limited context windows of LLMs. However, existing memory systems invoke LLMs to process every incoming interaction for memory extraction, and such an eager memory consolidation scheme leads to substantial token consumption. To tackle this problem, we propose RecMem by rethinking when memory consolidation should be conducted. RecMem stores incoming interactions in a subconscious memory layer and encode them using lightweight embedding models for retrieval. LLMs are only invoked to extract episodic and semantic memory when sustained recurrence are observed for semantically similar interactions. Such recurrence-based consolidation works because these interactions correspond to a semantic cluster with rich information and thus are worth extraction and summarization. To improve accuracy, RecMem also incorporates a semantic refinement mechanism that recovers the fine-grained facts omitted by memory extraction. Experiments show that RecMem reduces the memory construction token cost of three SOTA memory systems by up to 87% while exceeding their accuracy.\n\n---\n\n## 83. Learning Sim-Grounded Policies for Bimanual Rope Manipulation from Human Teleoperation Data\n- **Authors**: Gina Wigginghaus , Tim Missal , Berk Guler , Simon Manschitz , Jan Peters\n- **URL**: [https://arxiv.org/abs/2605.16043](https://arxiv.org/abs/2605.16043)\n- **Abstract**:\n> Deformable Linear Objects (DLOs) such as ropes and cables are widely encountered in both household and industrial applications, yet remain challenging to manipulate due to their infinite-dimensional configuration space and frequent self-occlusion. Imitation learning from teleoperation offers a practical path to bimanual DLO manipulation, but its scalability is limited by human effort, making the choice of observation space critical for generalization from small datasets. In this study, we investigate whether the lack of generalization in egocentric visual policies for the knot-untangling task stems from the observation space itself, rather than from the policy architecture or data scale. We compare two Action Chunking with Transformers policies trained on the same bimanual teleoperation data: a vision-based policy conditioned on two egocentric RGB streams from wrist-mounted cameras, and a state-based policy conditioned on the DLO's 3D particle state, extracted from an initial observation via multi-view fusion and evolved in a particle-based eXtended Position-Based Dynamics simulation. Evaluated open-loop on an unseen rope configuration, the state-based policy outperforms its visual counterpart with a 30.8% reduction in L1 error when predicting the initial grasp-and-pull action, quantifying the observability gap between pixels and physics-consistent state, and pointing toward more data-efficient robot learning for the DLO manipulation task from limited human demonstrations.\n\n---\n\n## 84. Who Owns This Agent? Tracing AI Agents Back to Their Owners\n- **Authors**: Ruben Chocron , Doron Jonathan Ben Chayim , Eyal Lenga , Gilad Gressel , Alina Oprea , Yisroel Mirsky\n- **URL**: [https://arxiv.org/abs/2605.16035](https://arxiv.org/abs/2605.16035)\n- **Abstract**:\n> AI agents are increasingly deployed to act autonomously in the world, yet there is still no reliable way to trace a harmful agent back to the account that deployed it. This creates the same accountability gap across both ends of the intent spectrum: benign operators may deploy misconfigured or overbroad agents that cause harm unintentionally, while malicious operators may deliberately weaponize agents for scams, harassment, or cyber attacks. In many cases, these agents are powered by vendor-hosted models, a dependency that holds even for sophisticated adversaries such as state actors conducting cyber operations. In either case, affected parties can observe the behavior but cannot notify the responsible operator, stop the session, or identify the account for investigation. We formalize this gap as the problem of agent attribution: linking an observed agent interaction to the responsible account at the hosting vendor. To our knowledge, this is the first work to define the problem and present a practical solution. Our protocol is canary-based: an authorized party injects a canary into the agent's interaction stream, and the vendor searches a narrow window of session logs to recover the originating session and account. Simple canaries suffice in non-adversarial settings. For adversarial operators who filter or paraphrase incoming content, we develop robust canary constructions that cannot be suppressed without degrading the agent's own task performance, yielding a formal asymmetry in the defender's favor. We evaluate a variety of scenarios including real-world agents and show that our attribution method is reliable, robust, and scalable for vendor-side deployment.\n\n---\n\n## 85. From Flat Language Labels to Typological Priors: Structured Language Conditioning for Multilingual Speech-to-Speech Translation\n- **Authors**: Yu Pan , Yang Hou , Xiongfei Wu , Liang Zhang , Yves Le Traon , Lei Ma , Jianjun Zhao\n- **URL**: [https://arxiv.org/abs/2605.16026](https://arxiv.org/abs/2605.16026)\n- **Abstract**:\n> Compositional speech-to-speech translation (S2ST) systems built upon speech large language models (SpeechLLMs) have recently shown promising performance. However, existing S2ST systems often either neglect source-language information or encode it through a language-as-label paradigm, representing each source language as an independent flat embedding. Such a design overlooks systematic linguistic structure shared across languages, which may limit data-efficient multilingual adaptation when supervised S2ST data are scarce. To address this issue, we propose S2ST-Omni 2, a many-to-one compositional S2ST framework that systematically reformulates multilingual language conditioning from flat language labels to structured typological priors. Specifically, S2ST-Omni 2 revisits language conditioning at three levels: typology-informed hierarchical language encoding for structured source-language representation, dynamically-gated language-aware Dual-CTC for content-adaptive acoustic modulation, and typology-aware LLM prompting for decoder-side linguistic guidance. Experiments on CVSS-C show that S2ST-Omni 2 achieves superior average performance among representative S2ST approaches across BLEU, COMET, ASR-BLEU, and BLASER 2.0 under the adopted evaluation protocol. Ablation studies indicate that the proposed representation-level, acoustic-level, and decoding-level strategies provide complementary benefits. Moreover, controlled data-budget analyses and a Japanese-to-English evaluation using only approximately 3 hours of supervised training data suggest that explicit typological priors provide useful inductive biases for data-efficient multilingual S2ST.\n\n---\n\n## 86. Can Vision Language Models Be Adaptive in Mathematics Education? A Learner Model-based Rubric Study\n- **Authors**: Jie Gao , Yongan Yu , Junzhu Su , Yiran Lin , Adam K. Dube , Jackie Chi Kit Cheung\n- **URL**: [https://arxiv.org/abs/2605.16011](https://arxiv.org/abs/2605.16011)\n- **Abstract**:\n> Adaptive learning refers to educational technologies that track learners' learning progress and adapt the instructional process based on individual learners' learning performance. It is increasingly recognized as critical for developing an effective learning support tool. Vision language models (VLMs) have seen adoption in mathematics education, and students have been using them as learning aids for personalized instruction. However, it is unknown whether VLMs have the ability to adapt to different learner profiles when providing mathematical instructions. Current VLMs lack a systematic evaluation framework for this adaptivity to different learner profiles in mathematics tutoring tasks. To address this gap, we draw on the learner model from the adaptive learning framework (Shute and Towle, 2018) and propose a learner model-based rubric. Our rubric formalizes adaptivity assessment into three aspects: cognitive aspects, motivational aspects, and complexity. We also evaluate two additional dimensions of VLM responses: correctness (of answers and solutions) and quality (of the response itself). Our experimental results show measurable differences in adaptivity across models and also reveal that current VLMs struggle to consistently produce learner model-based instructional responses, especially when receiving limited learner information.\n\n---\n\n## 87. CitePrism: Human-in-the-Loop AI for Citation Auditing and Editorial Integrity\n- **Authors**: Gowrika Mahesh , Budanur Madappa Darshan Gowda , Kavana Gopladevarahalli Papegowda , Prajwal Basavaraj , Binh Vu , Swati Chandna , Mehrdad Jalali\n- **URL**: [https://arxiv.org/abs/2605.16000](https://arxiv.org/abs/2605.16000)\n- **Abstract**:\n> Editors and reviewers are expected to ensure that manuscripts cite relevant, accurate, current, and ethically appropriate literature, yet manuscript-level citation auditing remains largely manual, fragmented, and difficult to scale. Citation context, metadata quality, self-citation patterns, and bibliographic integrity all affect whether a reference appropriately supports a local claim. We present CitePrism, a transparent hybrid decision-support framework for editorial citation auditing that combines LLM-assisted contextual reasoning, embedding-based semantic similarity, metadata verification, integrity-oriented flags, and human-in-the-loop analyst review. CitePrism extracts citation neighborhoods, enriches reference metadata, computes fused relevance scores, surfaces metadata and self-citation review prompts, and supports configurable threshold-based triage. In a preliminary validation on a single case-study manuscript with 104 references from pavement engineering, agreement with human binary relevance labels reached Cohen's kappa = 0.429. At operating threshold tau = 17, CitePrism flagged all human-labeled irrelevant citations, while also producing false positives requiring analyst review. These results suggest that CitePrism may support conservative editorial screening and citation-quality triage, but they do not establish general editorial performance. CitePrism is intended as pilot-stage decision support, not as an autonomous misconduct detector or automated editorial decision system. Broader validation across manuscripts, domains, annotators, baselines, and deployment settings is required before operational use.\n\n---\n\n## 88. Constrained latent state modeling: A unifying perspective on representation learning under competing constraints\n- **Authors**: Gwenolé Quellec\n- **URL**: [https://arxiv.org/abs/2605.15995](https://arxiv.org/abs/2605.15995)\n- **Abstract**:\n> Learning latent representations from complex data is central to modern machine learning, spanning temporal, multimodal, and partially observed systems. In such settings, representations are better understood as latent states capturing underlying system dynamics, rather than as mere compressed summaries of observations. Yet current approaches remain fragmented, relying on distinct -- and often implicit -- assumptions about what these states should represent. We argue that this fragmentation reflects a more fundamental limitation: latent representations are typically learned from underconstrained objectives that fail to specify the properties that meaningful latent states should satisfy. As a result, multiple representations can satisfy the same objective, leading to ambiguity in their structure and interpretation. While many of the underlying principles have been explored in isolation, their interactions have not been explicitly formalized. In this work, we propose constrained latent state modeling (CLSM) as a unifying perspective. We identify a set of core properties -- predictive sufficiency, minimality, temporal coherence, observation compatibility, invariance to nuisance factors, and structural constraints -- and show that they are intrinsically coupled through fundamental trade-offs. Revisiting major modeling families through this lens, we show that existing approaches can be interpreted as enforcing different subsets of constraints, thereby occupying distinct regions of a common design space. This perspective reframes persistent challenges such as lack of identifiability as consequences of underconstrained formulations, rather than isolated technical limitations. More broadly, CLSM provides a principled framework to make design choices explicit, to analyze trade-offs, and to guide the development of more interpretable, robust, and task-aligned latent state models.\n\n---\n\n## 89. Beyond Content: A Comprehensive Speech Toxicity Dataset and Detection Framework Incorporating Paralinguistic Cues\n- **Authors**: Zhongjie Ba , Liang Yi , Peng Cheng , Qingcao Li , Qinglong Wang , Li Lu\n- **URL**: [https://arxiv.org/abs/2605.15984](https://arxiv.org/abs/2605.15984)\n- **Abstract**:\n> Toxic speech detection has become a crucial challenge in maintaining safe online communication environments. However, existing approaches to toxic speech detection often neglect the contribution of paralinguistic cues, such as emotion, intonation, and speech rate, which are key to detecting speech toxicity. Moreover, current toxic speech datasets are predominantly text-based, limiting the development of models that can capture paralinguistic this http URL address these challenges, we present ToxiAlert-Bench, a large-scale audio dataset comprising over 30,000 audio clips annotated with seven major toxic categories and twenty fine-grained toxic labels. Uniquely, our dataset annotates toxicity sources -- distinguishing between textual content and paralinguistic origins -- for comprehensive toxic speech this http URL , we propose a dual-head neural network with a multi-stage training strategy tailored for toxic speech detection. This architecture features two task-specific classification headers: one for identifying the source of sensitivity (textual or paralinguistic), and the other for categorizing the specific toxic type. The training process involves independent head training followed by joint fine-tuning to reduce task interference. To mitigate data class imbalance, we incorporate class-balanced sampling and weighted loss this http URL experimental results show that leveraging paralinguistic features significantly improves detection performance. Our method consistently outperforms existing baselines across multiple evaluation metrics, with a 21.1% relative improvement in Macro-F1 score and a 13.0% relative gain in accuracy over the strongest baseline, highlighting its enhanced effectiveness and practical applicability.\n\n---\n\n## 90. Ontology for Policing: Conceptual Knowledge Learning for Semantic Understanding and Reasoning in Law Enforcement Reports\n- **Authors**: Anita Srbinovska , Jansen Orfan , Adrian Martin , Ernest Fokoué\n- **URL**: [https://arxiv.org/abs/2605.15978](https://arxiv.org/abs/2605.15978)\n- **Abstract**:\n> Law enforcement reports contain structured fields and written narratives. However, many incident facts that are needed for review, police training, and investigations are in natural language and require manual reading. We propose a framework using symbolic methods for converting narratives into evidence-linked facts. Our objective is to measure the value of narratives to recover incident details only from the unstructured text and build temporal graphs with time cues and domain axioms. We achieve this by redacting personal identifiers, semantic parsing, predicate mapping to ontology, and reasoning. We evaluate the symbolic approach on 450 property crime reports and a short human review. Of the extracted events from the system, 54.1% had a confidence score of at least 0.80 and 93.7% were mapped through the PropBank--VerbNet--WordNet semantic path. 100% agreement was reached on incident initiation, stolen items, and temporal cues and lower agreement for forced entry interpretation.\n\n---\n\n## 91. Reference-Free Reinforcement Learning Fine-Tuning for MT: A Seq2Seq Perspective\n- **Authors**: Ernesto Garcia-Estrada , Carlos Escolano , José A. R. Fonallosa\n- **URL**: [https://arxiv.org/abs/2605.15976](https://arxiv.org/abs/2605.15976)\n- **Abstract**:\n> Production machine translation relies overwhelmingly on encoder-decoder Seq2Seq models, yet reinforcement learning approaches to MT fine-tuning have largely targeted decoder-only LLMs at $\\geq$7B parameters, with limited systematic study of encoder-decoder architectures. We apply Group Relative Policy Optimization to NLLB-200 (600M and 1.3B) using a hybrid reference-free reward (LaBSE and COMET-Kiwi) that requires no parallel data at fine-tuning time, evaluating across 13 typologically diverse languages. GRPO yields consistent improvements on all 13 languages, up to $+$5.03 chrF++ for Traditional Chinese, and, without any target-language data, competes with 3-epoch supervised fine-tuning on morphologically complex languages . We identify a consistent empirical pattern in which gains are largest where baseline performance is weakest and reward discriminability is highest, making this approach most effective precisely where parallel data is scarcest, and replicate this pattern across English and Spanish source languages.\n\n---\n\n## 92. When and Why Adversarial Training Improves PINNs: A Neural Tangent Kernel Perspective\n- **Authors**: Yuan-dong Cao , Chi Chiu SO , Jun-Min Wang , He Wang\n- **URL**: [https://arxiv.org/abs/2605.15959](https://arxiv.org/abs/2605.15959)\n- **Abstract**:\n> Physics-informed neural networks (PINNs) are powerful surrogates for differential equations but are notoriously difficult to train due to spectral bias, stiffness, and poor accuracy on high-frequency or multiscale solutions. Adversarial training based on generative adversarial networks (GANs) has recently gained surprisingly strong empirical results in improving training, but the underlying mechanisms remain elusive. To this end, we propose a new analysis framework for adversarially trained PINNs, based on the key observation of how the discriminator in GANs can influence the training dynamics of PINNs. The framework first provides a much needed theoretical grounding to why and when adversarial training is effective in PINNs, then presents a unified analysis of GANs variants in such training, and finally leads to a new, practical, efficient training algorithm for PINNs. Empirical results demonstrate that our method can significantly reduce the pathology of PINNs training, thereby providing better models with superior performances, often several magnitudes more accurate than alternative methods.\n\n---\n\n## 93. Decomposed Vision-Language Alignment for Fine-Grained Open-Vocabulary Segmentation\n- **Authors**: Chenhao Wang , Yingrui Ji , Yu Meng , Yao Zhu\n- **URL**: [https://arxiv.org/abs/2605.15942](https://arxiv.org/abs/2605.15942)\n- **Abstract**:\n> Open-vocabulary segmentation models often struggle to generalize to unseen combinations of object categories and attributes, because fine-grained descriptions are typically encoded as holistic sentences that entangle multiple semantic units. We propose a Decomposed Vision-Language Alignment framework that explicitly factorizes textual prompts into a concept token and multiple attribute tokens, enabling separate cross-modal interactions for each semantic unit. At the feature level, we introduce a Feature-Gated Cross-Attention module that generates attribute-specific gating maps to fuse information in a multiplicative manner, effectively enforcing compositional semantics. At the scoring level, per-token similarities are aggregated in log-space, producing a stable and interpretable compositional matching. The method can be seamlessly integrated into existing transformer-based segmentation architectures and significantly improves generalization to unseen attribute-category compositions in fine-grained open-vocabulary segmentation benchmarks.\n\n---\n\n## 94. LoCO: Low-rank Compositional Rotation Fine-tuning\n- **Authors**: An Nguyen , Jaesik Choi , Anh Tong\n- **URL**: [https://arxiv.org/abs/2605.15916](https://arxiv.org/abs/2605.15916)\n- **Abstract**:\n> Parameter-efficient fine-tuning (PEFT) has emerged as an critical technique for adapting large-scale foundation models across natural language processing and computer vision. While existing methods such as low-rank adaptations achieve parameter efficiency via low-rank weight updates, they are limited in their ability to preserve the geometric structure of pretrained representations. We introduce Low-rank Compositional Orthogonal fine-tuning (LoCO), a novel PEFT method that constructs orthogonal transformations through low-rank skew-symmetric matrices and compositional rotation chains. We propose an approximation scheme that enables fully parallel computation of compositional rotations, making the approach practical for high-dimensional feature spaces. Our method maintains low computational complexity while maintaining orthogonality with controlled approximation error. We validate LoCO across diverse domains, including diffusion transformer fine-tuning, vision transformer adaptation, and language model adaptation. Our method demonstrates superior or competitive performance compared to both existing orthogonal and non-orthogonal methods.\n\n---\n\n## 95. SLIP & ETHICS: Graduated Intervention for AI Emotional Companions\n- **Authors**: Minseo Kim\n- **URL**: [https://arxiv.org/abs/2605.15915](https://arxiv.org/abs/2605.15915)\n- **Abstract**:\n> AI emotional companions face a safety-rapport paradox: restrictive safeguards can damage supportive alliance, while permissive systems risk user harm. We present SLIP (Staged Layers of Intervention Protocol), a four-stage graduated methodology deriving interventions (none, soft, hard) from structured qualitative indicators -- affect intensity (a) and narrative dynamism (m) -- alongside ETHICS (Emergent Taxonomy for Human-AI Interaction Context Signals), a \"signals not labels\" taxonomy. An evaluation combining a small-scale production deployment (N=68 entries, 10 users, 10 weeks) with a synthetic persona battery (N=91, 5 behavioral-risk profiles) achieved 0% false positives for the flow persona and showed expected escalation patterns in crisis-oriented personas. However, initial results showed that 8 consecutive days of high-energy elevation produced zero interventions (0/8), exposing a boundary where the \"do not pathologize\" principle conflicts with safety. A subsequent three-model stress test demonstrated that increased model capability improves detection from 0/8 to 6/8 while preserving 0/10 flow false positives in the largest model. Read as preliminary, these findings position graduated intervention as a design direction for navigating -- not resolving -- the safety-rapport tension in affective computing.\n\n---\n\n## 96. Towards Generalization of Block Attention via Automatic Segmentation and Block Distillation\n- **Authors**: Shuaiyi Li , Zhisong Zhang , Yan Wang , Lei Zhu , Dongyang Ma , Chenlong Deng , Yang Deng , Wai Lam\n- **URL**: [https://arxiv.org/abs/2605.15913](https://arxiv.org/abs/2605.15913)\n- **Abstract**:\n> Block attention, which processes the input as separate blocks that cannot attend to one another, offers significant potential to improve KV cache reuse in long-context scenarios such as Retrieval-Augmented Generation (RAG). However, its broader application is hindered by two key challenges: the difficulty of segmenting input text into meaningful, self-contained blocks, and the inefficiency of existing block fine-tuning methods that risk degrading performance. To address these, we first construct SemanticSeg, a large and diverse semantic segmentation dataset containing over 30k instances across 16 categories-including books, code, web text, and conversations with text lengths ranging from 2k to 32k. Using this dataset, we train a lightweight segmenter to automatically partition text into human-instinct-aligned blocks with controllable granularity. Second, we propose block distillation, a training framework that is more efficient than block fine-tuning, which uses a frozen full-attention teacher model to guide the block-attention student. This framework integrates three novel components: block sink tokens to mitigate information loss at block boundaries, block dropout to leverage training signals from all blocks, and token-level loss weighting to focus learning on block-attention-sensitive tokens. Experiments across multiple models and benchmarks demonstrate that our segmenter outperforms heuristic and statistical baselines, and block distillation achieves near-full-attention performance under block attention, establishing a practical and scalable pathway for deploying block attention.\n\n---\n\n## 97. RaPD: Resolution-Agnostic Pixel Diffusion via Semantics-Enriched Implicit Representations\n- **Authors**: Yanhao Ge , Shanyan Guan , Weihao Wang , Ying Tai , Mingyu You\n- **URL**: [https://arxiv.org/abs/2605.15908](https://arxiv.org/abs/2605.15908)\n- **Abstract**:\n> Natural images are continuous, yet most generative models synthesize them on discrete grids, limiting resolution-flexible generation. Continuous neural fields enable resolution-free rendering, but prior methods introduce continuity only at the decoding stage as an interpolation module, leaving the generative latent space discretized and reconstruction-oriented. We propose RaPD (Resolution-agnostic Pixel Diffusion), which performs diffusion in a continuous Neural Image Field (NIF) latent space. RaPD bridges this reconstruction-generation gap with Semantic Representation Guidance for generation-aware latent learning and a Coordinate-Queried Attention Renderer for coordinate-conditioned, scale-aware rendering. A single denoised latent can be rendered at arbitrary resolutions by changing only the query coordinates, keeping diffusion cost fixed. Experiments demonstrate superior generation quality and resolution scalability.\n\n---\n\n## 98. Generative Long-term User Interest Modeling for Click-Through Rate Prediction\n- **Authors**: Jiangli Shao , Kaifu Zheng , Hao Fang , Huimu Ye , Zhiwei Liu , Bo Zhang , Shu Han , Xingxing Wang\n- **URL**: [https://arxiv.org/abs/2605.15905](https://arxiv.org/abs/2605.15905)\n- **Abstract**:\n> Modeling long-term user interests with massive historical user behaviors enhances click-through rate (CTR) prediction performance in advertising and recommendation systems. Typically, a two-stage framework is widely adopted, where a general search unit (GSU) first retrieves top-$k$ relevant behaviors towards the target item, and an exact search unit (ESU) generates interest features via tailored attention. However, current target-centered GSU would ignore other latent user interests, leading to incomplete and biased interest features. Additionally, the matching-based retrieval process in GSUs depends on the pairwise similarity score between target item and each historical behavior, which not only becomes time-consuming for online services as user behaviors continue to grow, but also overlooks the interaction information among user behaviors. To combat these problems, we propose a \\textbf{Gen}erative \\textbf{L}ong-term user \\textbf{I}nterest model named GenLI for CTR prediction. GenLI consists of an interest generation module (IGM), a behavior retrieval module (BRM), and an interest fusion module (IFM). The IGM generates multiple interest distributions to indicate different aspects of real-time user interests, which is target-independent and incorporates interaction information among behaviors, ensuring complete and diverse interest features. The BRM selects related behaviors via a simple lookup operation, reducing the time complexity for weighting each behavior to $O(1)$. Finally, the IFM uses delicate gating mechanisms to generate interest features. Based on the generation process, GenLI improves the diversity of user interests and avoids complex matching-based behavioral retrieval, achieving a better balance between accuracy and efficiency for CTR prediction.\n\n---\n\n## 99. Uncertainty-Aware Wildfire Smoke Density Classification from Satellite Imagery via CBAM-Augmented EfficientNet with Evidential Deep Learning\n- **Authors**: Ranjith Chodavarapu\n- **URL**: [https://arxiv.org/abs/2605.15894](https://arxiv.org/abs/2605.15894)\n- **Abstract**:\n> Rapid and accurate wildfire smoke severity assessment from satellite images is essential for emergency response, air quality modeling, and human health risk management. Existing deep learning approaches treat smoke detection as a binary task, producing point estimates without any measure of prediction confidence. We propose a probabilistic framework to categorize a satellite patch into Light, Moderate, and Heavy severity classes and to provide decomposed epistemic and aleatoric uncertainty in a single forward pass. Our architecture uses the backbone of a pre-trained EfficientNet-B3 and a CBAM module with an evidential deep learning head that predicts Dirichlet concentration parameters, directly estimating vacuity (epistemic) and dissonance (aleatoric) without Monte Carlo sampling. Evaluated on 16,298 real satellite patches derived from the Wildfire Detection dataset, our model achieves 93.8% weighted test accuracy (91.1% unweighted) with ECE=0.0274. Selective prediction retaining the most certain 50% of patches achieves 96.7% accuracy. As image quality degrades, uncertainty increases monotonically, and vacuity is a practical scan quality measure. The Moderate class represents transitional smoke conditions that exhibit the highest epistemic uncertainty (mean vacuity = 0.187), confirming the model correctly identifies ambiguous smoke boundary regions. CBAM spatial attention maps localize to structurally distinctive scene regions, and t-SNE demonstrates the clear cluster separation of Light and Heavy smoke.\n\n---\n\n## 100. CHoE: Cross-Domain Heterogeneous Graph Prompt Learning via Structure-Conditioned Experts\n- **Authors**: Peiyuan Li , Yongqi Huang , Jitao Zhao , Dongxiao He , Di Jin , Weixiong Zhang\n- **URL**: [https://arxiv.org/abs/2605.15888](https://arxiv.org/abs/2605.15888)\n- **Abstract**:\n> Heterogeneous Graph Prompt Learning (HGPL)has emerged as a promising paradigm for bridging the gap between the objectives of pre-training foundation models and their downstream applications in heterogeneous graph settings. However, existing HGPL methods are primarily designed for in-domain scenarios, whereas real-world deployments often span multiple domains, and the data used for pre-training and downstream tasks may originate from different distributions. Consequently, the applicability of current HGPL approaches is limited to in-domain settings, and their performance typically degrades when application domains shift. To address this serious limitation, we develop CHoE, a cross-domain HGPL method built upon an expert network. During pre-training, we introduce and train structure-conditioned experts, and during prompt tuning, we adopt a structure-aware expert routing and load balancing mechanism to select structurally compatible experts for each meta-path view. In addition, we design a prompt-based semantic fusion module to integrate representations across multiple views for downstream prediction. Extensive experiments show that CHoE consistently improves performance in few-shot cross-domain applications, outperforming all baseline approaches.\n\n---\n\n## 101. Symplectic Neural Operators for Learning Infinite Dimensional Hamiltonian Systems\n- **Authors**: Yeang Makara , Yusuke Tanaka , Takashi Matsubara , Takaharu Yaguchi\n- **URL**: [https://arxiv.org/abs/2605.15881](https://arxiv.org/abs/2605.15881)\n- **Abstract**:\n> The modeling and simulation of infinite-dimensional Hamiltonian systems are central problems in mathematical physics and engineering, however they pose significant computational and structural challenges for standard data-driven architectures. In this work, we introduce the Symplectic Neural Operator, a neural operator architecture designed to preserve the symplectic structure intrinsic to Hamiltonian PDEs. We provide a theoretical characterization of their symplecticity and establish a rigorous long-term stability result based on the combination of symplectic structure preservation and learning accuracy. Numerical experiments on canonical Hamiltonian PDEs corroborate this theoretical result and show that SNOs exhibit improved energy behavior compared with non-structure-preserving neural operators.\n\n---\n\n## 102. FSCM: Frequency-Enhanced Spatial-Spectral Coupled Mamba for Infrared Hyperspectral Image Colorization\n- **Authors**: Tingting Liu , Yuan Liu , Guiping Chen , Xiubao Sui , Qian Chen\n- **URL**: [https://arxiv.org/abs/2605.15880](https://arxiv.org/abs/2605.15880)\n- **Abstract**:\n> Thermal infrared imaging is robust to illumination variations and smoke interference, making it important for all-weather perception. However, the lack of natural color and fine texture limits target recognition, human visual interpretation, and the transfer of visible-light models. Existing infrared colorization methods mainly rely on single-band images, where insufficient spectral cues may lead to structural distortion and semantic confusion. Although infrared hyperspectral images provide rich spectral responses and material information, existing single-band frameworks remain limited in modeling spatial-spectral coupling and weak texture details. To address these issues, this paper presents FSCM, a spectral-information-guided GAN framework. Within FSCM, a frequency-enhanced spatial-spectral state-space generator composed of cascaded FSB units is constructed. Each FSB integrates three complementary components: state-space modeling captures global spatial-spectral dependencies; the frequency enhancement module (FEM) combines multi-level wavelet decomposition and Fourier gating to recover structural contours, directional high-frequency details, and global frequency responses; and the dual-stream hybrid gating module (DGM) integrates deformation-aware sampling with sparse attention to enhance effective local structures and suppress background interference. Additionally, an online semantic segmentation-guided loss is introduced to constrain the generated results, improving semantic consistency in complex road scenes. Experiments show that FSCM outperforms existing infrared colorization methods in visual quality and semantic fidelity.\n\n---\n\n## 103. Shapley Neuron Values for Continual Learning: Which Neurons Matter Most?\n- **Authors**: Mohammad Ali Vahedifar , Abhisek Ray , Qi Zhang\n- **URL**: [https://arxiv.org/abs/2605.15877](https://arxiv.org/abs/2605.15877)\n- **Abstract**:\n> Continual learning enables neural networks to learn tasks sequentially without forgetting previously acquired knowledge. However, neural networks suffer from catastrophic forgetting, where learning new tasks degrades performance on earlier ones. We address this problem with Shapley Neuron Valuation (SNV), a principled framework that quantifies Neuron importance in continual learning, grounded in cooperative game theory. SNV selectively freezes important Neurons while keeping others plastic, enabling buffer-free continual learning without expanding architecture. Experiments on ImageNet-1k show that SNV consistently outperforms existing buffer-free methods. In particular, SNV improves accuracy by +2.88% in the class incremental learning and +6.46% in the task incremental learning scenarios compared to the second baseline.\n\n---\n\n## 104. Access Timing as Scaffolding: A Reinforcement Learning Approach to GenAI in Education\n- **Authors**: Janne Rotter , Pau Benazet i Montobbio , Davinia Hernández-Leo\n- **URL**: [https://arxiv.org/abs/2605.15850](https://arxiv.org/abs/2605.15850)\n- **Abstract**:\n> In recent years, generative AI (GenAI) in educational settings has become ubiquitous in students' daily lives, despite its potential to induce over-reliance, metacognitive disengagement, and diminished learning when used unrestrictedly. While most prior research has thus focused on how to pedagogically scaffold its usage, the question of when to allow off-the-shelf GenAI remains understudied and lacks pedagogically grounded empirical investigation. We treat access timing itself as a form of implicit scaffolding and operationalize it through a reinforcement learning (RL) agent that decides when students should access GenAI, with a reward function grounded in metacognitive theory, cognitive load theory, and productive failure. In a mixed-methods controlled lab study with N=105 participants, we compared the agent's effect on learning gains and metacognitive engagement to unrestricted and fully restricted use. Results show that strategically timed GenAI access under the reinforcement learning condition improved objective post-test performance and metacognitive accuracy compared with unrestricted access, while reducing task errors and time on task relative to complete withholding, all without the need for explicit metacognitive prompts or structured scaffolding. However, no between-condition differences emerged on self-reported metacognitive awareness. Overall, timing of GenAI access therefore is a tractable, theoretically grounded, and scalable pedagogical paradigm that improves over completely unrestricted and withheld access, compatible with off-the-shelf tools and potentially low adoption barrier. This opens up a new research area that explores how access timing can be facilitated by educators and implemented in human-AI learning system design.\n\n---\n\n## 105. RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades\n- **Authors**: Xinbo Xu , Ruihan Yang , Haiyang Shen , Wendong Xu , Bofei Gao , Ruoyu Wu , Kean Shi , Weichu Xie , Xuanzhong Chen , Ming Wu , Jason Zeng , Michael Heinrich , Elvis Zhang , Liang Chen , Kuan Li , Baobao Chang\n- **URL**: [https://arxiv.org/abs/2605.15846](https://arxiv.org/abs/2605.15846)\n- **Abstract**:\n> Coding agents are increasingly deployed in real software development, where a single version iteration requires months of coordinated work across many files. However, most existing benchmarks focus predominantly on single-issue bug fixes from Python repositories, with coarse pass/fail evaluation outcomes, and thus fail to capture long-horizon, multi-target development at real engineering scale. To address this gap, we present RoadmapBench, a benchmark of 115 long-horizon coding tasks grounded in real open-source version upgrades across 17 repositories and 5 programming languages. Each task places the agent on a source-version code snapshot and provides a multi-target roadmap instruction requiring it to implement the functionality introduced in the target version, with a median modification of 3,700 lines across 51 files. We conduct a systematic evaluation on thirteen frontier models and find that even the strongest, Claude-Opus-4.7, resolves only 39.1% of tasks, while the weakest achieves merely 5.2%, in stark contrast to existing bug-fix benchmarks, suggesting that long-horizon software development remains a largely unsolved problem.\n\n---\n\n## 106. GAP: Geometric Anchor Pre-training for Data-Efficient Visuomotor Learning of Manipulation Tasks\n- **Authors**: Davide Buoso , Andrea Protopapa , Stefano Di Carlo , Francesca Pistilli , Giuseppe Averta\n- **URL**: [https://arxiv.org/abs/2605.15836](https://arxiv.org/abs/2605.15836)\n- **Abstract**:\n> Learning visuomotor policies from scarce expert demonstrations remains a core challenge in robotic manipulation. A primary hurdle lies in distilling high-dimensional RGB representations into control-relevant geometry without overfitting. While using frozen pre-trained Vision Foundation Models (VFMs) improves data efficiency, it also shifts most task adaptation onto a small spatial pooling module, which can latch onto task-irrelevant shortcuts and lose geometric grounding when finetuned with few data samples. More broadly, pre-trained visual representations used for policy learning have been observed to struggle under even minor scene perturbations, highlighting the need for robustness-oriented inductive biases. We propose Geometric Anchor Pre-training (GAP), a simple, action-free warm-up stage that regularizes the spatial adapter before downstream imitation learning. GAP pre-trains the pooling layer on a lightweight simulated proxy task where object masks are available at no cost, encouraging the adapter to produce keypoints that lie on the object, cover its spatial extent, and remain sharp and repeatable over time. This yields stable geometric anchors that provide a reliable coordinate interface for few-shot policy learning, while keeping the VFM frozen. We evaluate GAP on RoboMimic and ManiSkill under severe data scarcity (15-50 demonstrations) and domain shift. A simple adapter regularized with GAP consistently outperforms stronger attention-based poolers and end-to-end fine-tuning, achieving 62% success on RoboMimic Can with 15 demonstrations (+16% over AFA), 63% on the long-horizon high-precision Tool Hang task with 50 demonstrations, and 61% on ManiSkill StackCube with 30 demonstrations (+11% over full fine-tuning). The proxy stage is lightweight and fully decoupled from downstream tasks, making it practical to reuse across environments and manipulation skills.\n\n---\n\n## 107. Modeling Music as a Time-Frequency Image: A 2D Tokenizer for Music Generation\n- **Authors**: Yuqing Cheng , Xingyu Ma , Guochen Yu , Xiaotao Gu\n- **URL**: [https://arxiv.org/abs/2605.15831](https://arxiv.org/abs/2605.15831)\n- **Abstract**:\n> Autoregressive music generation depends strongly on the audio tokenizer. Existing high-fidelity codecs often use residual multi-codebook quantization, which preserves reconstruction quality but complicates language modeling after sequence flattening, as the residual hierarchy imposes strong sequential dependencies and can amplify error accumulation. We propose BandTok, a generation-oriented 2D Mel-spectrogram tokenizer that represents each frame with Mel-frequency band tokens from a single shared codebook. This design yields a physically interpretable time-frequency token grid with a more independent token structure, making it better suited for autoregressive modeling. BandTok improves reconstruction with a multi-scale PatchGAN objective and EMA codebook updates. We further introduce an autoregressive language model with 2D Rotary Position Embedding (2D RoPE) to preserve temporal and frequency-band structure during generation. Experiments show that BandTok improves over residual-codebook tokenizers and achieves strong results in a data-limited setting. The source code and generation demos for this work are publicly available.\n\n---\n\n## 108. Toward Natural and Companionable Virtual Agents via Cross-Temporal Emotional Modeling\n- **Authors**: Feier Qin , Xiao Li , Yi Zheng , Haibin Huang , Hanyao Wang , Xiaoyu Wang , Yan Lu , Yuan Zhang\n- **URL**: [https://arxiv.org/abs/2605.15812](https://arxiv.org/abs/2605.15812)\n- **Abstract**:\n> Recent advances in foundation models have enabled conversational agents that aim for sustained companionship rather than mere task completion. Yet most still remain unable to support natural, long-term companion-like interactions, resulting in experiences that feel episodic and inauthentic. We argue that current agents overlooked cross-temporal modeling of agents' social behaviors and internal emotions: generated behaviors rarely influence an agent's emotional state, and emotional states seldom shape subsequent behaviors. We present Cross-Temporal Emotion Modeling (CTEM), a framework that links long-term behavioral history to moment-to-moment emotional expression. CTEM establishes a closed loop where past experiences update an evolving emotional state; this state conditions immediate interactions; and user feedback continually revises both memory and emotional state, enabling reflection and anticipation. We instantiate CTEM as Auri, a companion agent on an instant-messaging platform, and report a 21-day in-the-wild study showing that CTEM shows improvements in perceived naturalness, coherence, and emotional harmony.\n\n---\n\n## 109. Grokking as Structural Inference: Transformers Need Bayesian Lottery Tickets\n- **Authors**: Kai Hidajat , Solden Stoll , Joseph An\n- **URL**: [https://arxiv.org/abs/2605.15787](https://arxiv.org/abs/2605.15787)\n- **Abstract**:\n> Why does a Transformer that has memorized its training set wait thousands of steps before it generalizes? Existing accounts locate this delay in norm minimization, feature emergence, or the late discovery of sparse subnetworks. These explanations capture important parts of the transition, but ignore a constraint unique to attention-based models: if attention discards an informative token, no bounded downstream computation can recover it. We formalize attention as an implicit Bayesian posterior over the task dependency graph and prove that generalization requires two separable conditions: a familiar Goldilocks bound on MLP capacity, coinciding with norm-based theories of grokking, and a novel Bayesian structural condition requiring attention to place sufficient mass on every informative token. This decoupling explains delayed generalization as delayed structural inference. Early in training, the MLP memorizes through unaligned features, drives the cross-entropy loss near zero, and thereby starves attention of structural gradient. Weight decay must then erode memorization before the missing graph becomes learnable, yielding the known inverse-weight-decay delay, which we derive as a structural waiting time. We then prove that this explaining-away delay can be bypassed by a KL-based structural intervention, yielding an inverse-intervention-strength scaling law for the grokking time. Experiments on algorithmic sequence tasks isolate structure from capacity and show that this Bayesian ticket matches or outperforms lottery-ticket transfer.\n\n---\n\n## 110. A Topology-Aware Spatiotemporal Handover Framework for Continuous Multi-UAV Tracking\n- **Authors**: Jianlin Ye , Christos Kyrkou , Panayiotis Kolios\n- **URL**: [https://arxiv.org/abs/2605.15779](https://arxiv.org/abs/2605.15779)\n- **Abstract**:\n> The integration of Unmanned Aerial Vehicles(UAVs) into Intelligent Transportation Systems (ITS) offers synoptic visibility for traffic monitoring, yet scalable deployment is hindered by trajectory fragmentation, where vehicle identity persistence is lost across multi-UAV Fields of View (FOV). While state-of-the-art frameworks excel in optimizing local trajectory extraction and stability for single-drone imagery, they often function as isolated data silos that generate disjointed trajectories, thereby precluding network-level analysis such as Origin-Destination estimation. This paper presents a real-time Multi-Camera Multi-Vehicle Tracking (MCMT) system designed to handle global identity persistence. Addressing the visual ambiguity and computational cost of appearance-based Re-Identification (Re-ID) in nadir views, we introduce a lightweight Topology-Based Spatiotemporal Handover mechanism. We implement a high-throughput parallel pipeline leveraging YOLO11 and ByteTrack to process concurrent 4K streams. Our core contribution is a deterministic queue-based matching algorithm that utilizes geometric overlaps and virtual lane discretization to predictively manage identity handover via FIFO queues. Experimental results on complex urban environments, including intersections and merging traffic, demonstrate a Handover Success Rate (HOSR) of 99.8% in continuous traffic flows, significantly outperforming Re-ID baselines (74.1%) while validating edge deployment feasibility. The source code is available at this https URL .\n\n---\n\n## 111. Lamarckian Inheritance in Dynamic Environments: How Key Variables Affect Evolutionary Dynamics\n- **Authors**: K. Ege de Bruin , Kyrre Glette , Kai Olav Ellefsen\n- **URL**: [https://arxiv.org/abs/2605.15769](https://arxiv.org/abs/2605.15769)\n- **Abstract**:\n> The co-optimization of a robot's body and brain presents a coupled challenge: the morphology constrains which control strategies are effective, while the control determines how well the morphology performs. To address this, we combine morphology optimization as evolution with controller optimization as lifetime learning, utilizing Lamarckian inheritance to transfer learned controller parameters from parent to offspring. In dynamic environments, existing literature presents conflicting evidence: while traditional evolutionary theory often suggests Lamarckian inheritance lacks benefit, recent studies in evolutionary robotics indicate it can improve performance. We hypothesize that this is because previous works have not included all relevant variables with dynamic environments. In this work, we show that the benefit of Lamarckian inheritance depends on two variables: how conflicting the environmental changes are to robot control, and the predictability of those changes for the robotic agent. Using virtual soft robots and two different learning approaches, Bayesian optimization and reinforcement learning, we show that Lamarckian inheritance only underperforms Darwinian inheritance when the changes are both conflicting and unpredictable. We find that adding a sensor to detect environmental changes restores the benefits for Lamarckian inheritance in conflicting environments, by allowing robotic agents to predict the need for a different behavior, thereby generalizing their control.\n\n---\n\n## 112. GRASP: Learning to Ground Social Reasoning in Multi-Person Non-Verbal Interactions\n- **Authors**: Junho Kim , Xu Cao , Houze Yang , Bikram Boote , Ana Jojic , Fiona Ryan , Bolin Lai , Sangmin Lee , James M. Rehg\n- **URL**: [https://arxiv.org/abs/2605.15764](https://arxiv.org/abs/2605.15764)\n- **Abstract**:\n> Understanding social interactions requires reasoning over subtle non-verbal cues, yet current multimodal large language models (MLLMs) often fail to identify who interacts with whom in multi-person videos. We introduce GRASP, a large-scale social reasoning dataset that connects high-level social QA with fine-grained gaze and deictic gesture events. GRASP contains 290K question--answer pairs over 46K videos totaling 749 hours, organized by a 16-category taxonomy spanning gaze, gesture, and joint gaze--gesture reasoning, together with GRASP-Bench for evaluation. Unlike prior resources that focus on either isolated cues or high-level social QA, GRASP builds questions from identity-consistent gaze trajectories, deictic gestures, and their joint compositions into social events. Moreover, we propose Social Grounding Reward (SGR), a learning signal that uses these social events to encourage models to reason about the participants involved in each interaction. Experiments show that SGR improves performance on GRASP-Bench while maintaining zero-shot performance on related social video QA benchmarks.\n\n---\n\n## 113. CompactQE: Interpretable Translation Quality Estimation via Small Open-Weight LLMs\n- **Authors**: Kamil Guttmann , Zofia Fraś , Artur Nowakowski , Krzysztof Jassem\n- **URL**: [https://arxiv.org/abs/2605.15763](https://arxiv.org/abs/2605.15763)\n- **Abstract**:\n> Current state-of-the-art Quality Estimation (QE) in machine translation relies on massive, proprietary LLMs, raising data privacy concerns. We demonstrate that smaller, open-source LLMs (<30B parameters) are a viable, cost-effective and privacy-preserving alternative. Using a single-pass prompting strategy, our models simultaneously generate quality scores, MQM error annotations, suggested error corrections, and full post-editions. Our analysis shows these models achieve highly competitive system-level correlations with human judgments that outperform traditional neural metrics, fine-tuned models, and human inter-annotator agreement, effectively approximating the capabilities of much larger proprietary LLMs.\n\n---\n\n## 114. BiomedAP: A Vision-Informed Dual-Anchor Framework with Gated Cross-Modal Fusion for Robust Medical Vision-Language Adaptation\n- **Authors**: Huanyang Tong , Kai Liu , Fangjun Kuang , Huiling Chen\n- **URL**: [https://arxiv.org/abs/2605.15736](https://arxiv.org/abs/2605.15736)\n- **Abstract**:\n> Biomedical Vision--Language Models (VLMs) have shown remarkable promise in few-shot medical diagnosis but face a critical bottleneck: \\textit{fragility to prompt variations}.Existing adaptation frameworks typically optimize visual and textual prompts as independent streams, relying on ideal ``Golden Prompts''. In clinical reality, where descriptions are often noisy and heterogeneous, this modality isolation leads to unstable cross-modal alignment. To address this, we propose BiomedAP, a vision-informed dual-anchor framework with gated cross-modal this http URL enforces synergistic alignment through two mechanisms: (1) Gated Cross-Modal Fusion, which enables layer-wise interaction between modalities, acting as a dynamic noise regulator to suppress irrelevant textual cues; and (2) a Dual-Anchor Constraint that regularizes learnable prompts toward stable semantic centroids derived from both expert templates (High Anchors) and few-shot visual prototypes (Low Anchors). Extensive experiments across 11 benchmarks demonstrate that BiomedAP consistently surpasses baselines, achieving competitive few-shot accuracy and markedly enhanced robustness under prompt perturbations. Our code is available at: this https URL . Keywords: Vision-Language Models; Prompt Learning; Parameter-Efficient Fine-Tuning; Few-shot Learning\n\n---\n\n## 115. UAM: A Dual-Stream Perspective on Forgetting in VLA Training\n- **Authors**: Jianke Zhang , Yuanfei Luo , Yucheng Hu , Xiaoyu Chen , Yanjiang Guo , Ziyang Liu , Hongbin Xu , Tian Lan , Jianyu Chen\n- **URL**: [https://arxiv.org/abs/2605.15735](https://arxiv.org/abs/2605.15735)\n- **Abstract**:\n> Vision--language--action (VLA) models are typically built by fine-tuning a pretrained vision--language model (VLM) on action data. However, we show that this standard recipe systematically erodes the VLM's multimodal competence, a side effect we call the embodiment tax. But do VLAs have to forget? Inspired by the two-stream organization of biological vision, we trace this degradation to a structural bottleneck: current VLAs ask a single encoder to support both language-grounded semantics and control-relevant visual features, whereas biological vision separates recognition and visuomotor control into distinct pathways. Building on this view, we propose the Unified Action Model (UAM), which adds a parallel Dorsal Expert, an analog of the brain's dorsal pathway. To make the Dorsal Expert an effective second pathway and reduce the control-learning burden on the VLM, we initialize it from a pretrained generative model and train it with a mid-level reasoning objective that predicts visual dynamics. This design allows us to train the whole VLA end-to-end on action data alone: with no parameter freezing, no gradient stopping, and no auxiliary VL co-training, UAM retains over $95\\%$ of the underlying VLM's multimodal capability and at the same time achieves the highest average success rate among baselines on a variety of manipulation tasks that probe out-of-distribution generalization, including unseen objects, novel object--target compositions, and instruction variation. Together, these results suggest that semantic preservation in VLAs can emerge from architectural separation itself, rather than being enforced by frozen weights or auxiliary data replay, and that this preserved semantic capability can naturally transfer from VLMs to semantic generalization in actions.\n\n---\n\n## 116. Structure Abstraction and Generalization in a Hippocampal-Entorhinal Inspired World Model\n- **Authors**: Tianqiu Zhang , Muyang Lyu , Xiao Liu , Si Wu\n- **URL**: [https://arxiv.org/abs/2605.15733](https://arxiv.org/abs/2605.15733)\n- **Abstract**:\n> Humans abstract experiences into structured representations to facilitate pattern inference and knowledge transfer. While the hippocampal-entorhinal (HPC-MEC) circuit is known to represent both spatial and conceptual spaces, the mechanisms for concurrently extracting abstract structures from continuous, high-dimensional dynamics remain poorly understood. We propose a brain-inspired hierarchical model that simultaneously infers latent transitions and constructs a predictive visual world model. Our architecture employs an inverse model for structural extraction alongside an HPC-MEC coupling model that dissociates relational structures (MEC) from integrated episodic scenes (HPC). Using primitive transformation dynamics as a benchmark, we demonstrate the model's capacity for structural abstraction. By leveraging velocity-driven path integration, the framework enables robust prediction and structural reuse across diverse contexts, thereby achieving structural generalization. This work provides a novel computational framework for understanding how brain-inspired, self-supervised learning of world models facilitates the acquisition of reusable abstract knowledge.\n\n---\n\n## 117. DecomPose: Disentangling Cross-Category Optimization Contention for Category-Level 6D Object Pose Estimation\n- **Authors**: Yifan Gao , Lu Zou , Zhangjin Huang , Guoping Wang\n- **URL**: [https://arxiv.org/abs/2605.15728](https://arxiv.org/abs/2605.15728)\n- **Abstract**:\n> Category-level 6D object pose estimation is typically formulated as a multi-category joint learning problem with fully shared model parameters. However, pronounced geometric heterogeneity across categories entangles incompatible optimization signals in shared modules, resulting in gradient conflicts and negative transfer during training. To address this challenge, we first introduce gradient-based diagnostics to quantify module-level cross-category contention. Building on results of diagnostics, we propose DecomPose, a difficulty-aware decomposition framework that mitigates optimization contention via: (1) difficulty-aware gradient decoupling, which groups categories using a data-driven difficulty proxy and routes each instance to a group-specific correspondence branch to isolate incompatible updates; and (2) stability-driven asymmetric branching, which assigns higher-capacity branches to structurally simple categories as stable optimization anchors while constraining complex categories with lightweight branches to suppress noisy updates and alleviate negative transfer. Extensive experiments on REAL275, CAMERA25, and HouseCat6D demonstrate that DecomPose effectively reduces cross-category optimization contention and delivers superior pose estimation performance across multiple benchmarks.\n\n---\n\n## 118. DiLA: Disentangled Latent Action World Models\n- **Authors**: Tianqiu Zhang , Muyang Lyu , Yufan Zhang , Fang Fang , Si Wu\n- **URL**: [https://arxiv.org/abs/2605.15725](https://arxiv.org/abs/2605.15725)\n- **Abstract**:\n> Latent Action Models (LAMs) enable the learning of world models from unlabeled video by inferring abstract actions between consecutive frames. However, LAMs face a fundamental trade-off between action abstraction and generation fidelity. Existing methods typically circumvent this issue by using two-stage training with pre-trained world models or by limiting predictions to optical flow. In this paper, we introduce DiLA, a novel Disentangled Latent Action world model that aims to resolve this trade-off via content-structure disentanglement. Our key insight is that disentanglement and latent action learning are co-evolving: the predictive bottleneck inherent in latent action learning serves as a driving force for disentanglement, compelling the model to distill spatial layouts into the structure pathway while offloading visual details to a separate content pathway for generation. This synergy yields a continuous, semantically structured latent action space without compromising generative quality. DiLA achieves superior results in video generation quality, action transfer, visual planning, and manifold interpretability. These findings establish DiLA as a unified framework that simultaneously achieves high-level action abstraction and high-fidelity generation, advancing the frontier of self-supervised world model learning.\n\n---\n\n## 119. Bidirectional Fusion Guided by Cardiac Patterns for Semi-Supervised ECG Segmentation\n- **Authors**: Jeonghwa Lim , Minje Park , Sunghoon Joo\n- **URL**: [https://arxiv.org/abs/2605.15722](https://arxiv.org/abs/2605.15722)\n- **Abstract**:\n> Accurate delineation of electrocardiogram (ECG), the segmentation of meaningful waveform features, is crucial for cardiovascular diagnostics. However, the scarcity of annotated data poses a significant challenge for training deep learning models. Conventional semi-supervised semantic segmentation (SemiSeg) methods primarily focus on consistency from unlabeled data, underutilizing the information exchange possible between labeled and unlabeled sets. To address this, we introduce CardioMix, a framework built on a bidirectional CutMix strategy guided by cardiac patterns for ECG segmentation. This approach enriches the labeled set with realistic variations from unlabeled data while simultaneously applying stronger supervisory signals to the unlabeled set, as the cardiac pattern-guided mixing ensures all augmented samples remain physiologically meaningful. Our framework is designed as a plug-and-play module, demonstrating high compatibility with various SemiSeg algorithms. Extensive experiments on SemiSegECG, a public multi-dataset benchmark for ECG delineation, demonstrate that CardioMix consistently outperforms existing CutMix-based fusion strategies across diverse datasets and labeled ratios as a plug-and-play module compatible with various SemiSeg algorithms.\n\n---\n\n## 120. Position: Early-Stage Quality Assurance in Annotation Pipelines Is More Cost-Effective Than Late-Stage Validation\n- **Authors**: Sunil Kothari , Sumukha Sharma Thoppanahalli Chandramouli , Naman Khandelwal , Parth Kulshreshtha , Ashi Jain , Kriti Banka , Tanuja Chintada , Venkata Triveni , Gulipalli Praveen Kumar , Manish Mehta , Tao Liu\n- **URL**: [https://arxiv.org/abs/2605.15714](https://arxiv.org/abs/2605.15714)\n- **Abstract**:\n> This position paper argues that the machine learning community should prioritize early-stage quality assurance in annotation pipelines over the prevailing practice of late-stage validation. Data quality bottlenecks increasingly limit foundation model improvement, yet quality assurance research focuses almost exclusively on validation methods rather than validation timing. When validation occurs, not merely what methods are employed, fundamentally determines both error rates and annotation costs. This temporal neglect is puzzling given the well-established \"shift-left\" principle from software engineering, where empirical studies demonstrate 4--100x cost multipliers for defects detected in later stages (Boehm, 1981; Shull et al., 2002). Annotation pipelines exhibit analogous dynamics: errors caught before annotation begins cost a fraction of those discovered after review cycles complete. We propose a taxonomy of three QA trigger points, namely pre-annotation (T0), post-annotation (T1), and post-review (T2), that decompose annotation workflows into discrete validation opportunities. A parametric error-propagation model formalizes when timing affects final error rates versus only economics, making timing a measurable design variable rather than a configuration afterthought. A survey of 47 recent papers reveals that only 4% report when validation occurs, a striking gap given timing's demonstrated impact in adjacent fields. Without explicit attention to QA timing, the community risks optimizing validation methods while ignoring the structural variable that may matter most. Acting on this position requires three steps: researchers should report QA timing configurations alongside validation methods; annotation platforms should expose timing as a first-class parameter; and the community should run controlled experiments that measure stage-specific detection rates directly.\n\n---\n\n## 121. Learning Dynamic Pick-and-Place for a Legged Manipulator\n- **Authors**: Moonkyu Jung , Jiseong Lee , Zhengmao He , Donghoon Youm , Juhyeok Mun , HyeongJun Kim , Hyunsik Oh , Donghyuk Choi , Jungwoo Hur , Jie Song , Jemin Hwangbo\n- **URL**: [https://arxiv.org/abs/2605.15713](https://arxiv.org/abs/2605.15713)\n- **Abstract**:\n> Legged manipulators extend robotic capabilities beyond static manipulation by integrating agile locomotion with versatile arm control. However, achieving precise manipulation while maintaining coordinated locomotion remains a major challenge. This work presents a hierarchical reinforcement learning framework for dynamic pick-and-place tasks using a quadruped equipped with a 6-DOF robotic arm. The framework incorporates an explicit mass estimation module enabling adaptive whole-body control for objects with varying weights. In simulation, the system achieves an 86.05% success rate with payloads up to 2.3 kg. The approach is further validated through real-world experiments across six representative scenarios with controlled variations in object physical properties (size and mass) and task heights. Specifically, within a wide vertical workspace ranging from ground level to 1.1~m-high tabletops, the system demonstrates an average success rate of 73.3% for payloads up to 1.3 kg, with an average execution time of 4.06 s. Unlike prior works that handle lightweight objects and execute pick-and-place motions with slow, piecewise motions, the proposed framework exploits concurrent locomotion and manipulation for dynamic, continuous execution. These results demonstrate the potential of quadrupedal mobile manipulators for adaptive, whole-body pick-and-place with heavier payloads and extended workspaces.\n\n---\n\n## 122. Feedback World Model Enables Precise Guidance of Diffusion Policy\n- **Authors**: Tuo An , Jindou Jia , Gen Li , Jingliang Li , Chuhao Zhou , Pengfei Liu , Bofan Lyu , Jiaqi Bai , Xinying Guo , Geng Li , Jianfei Yang\n- **URL**: [https://arxiv.org/abs/2605.15705](https://arxiv.org/abs/2605.15705)\n- **Abstract**:\n> World models aim to improve robotic decision making by predicting the consequences of actions. However, in practice, their predictions often become unreliable once the robot encounters states outside the training distribution, limiting their effectiveness at deployment. We observe that execution itself provides a natural but underutilized signal: after each action, the robot directly observes the true next state, revealing the mismatch between predicted and actual outcomes. Building on this insight, we propose feedback world model, a new paradigm that closes the loop between prediction and observation at inference time. Instead of treating the world model as a static open-loop predictor, our method maintains a lightweight feedback state that is updated online to iteratively correct future predictions, compensating for model errors using real-time observations without additional training data or parameter updates. We show that this process can be interpreted as a latent-space observer and admits convergence guarantees under mild conditions. We further introduce action-aware guidance to better translate corrected predictions into control by emphasizing action-controllable components while suppressing irrelevant variations. Experiments on LIBERO-Plus, Robomimic, and real-world manipulation tasks demonstrate that our method substantially improves both prediction accuracy and policy performance under distribution shift. In particular, it reduces world model prediction error by up to 76.4% and improves out-of-distribution (OOD) success rate by 30%. These results show that incorporating real-time feedback at inference time provides a simple yet powerful alternative to static world modeling.\n\n---\n\n## 123. H-Mem: A Novel Memory Mechanism for Evolving and Retrieving Agent Memory via a Hybrid Structure\n- **Authors**: Jiawei Yu , Yixiang Fang , Xilin Liu , Yuchi Ma\n- **URL**: [https://arxiv.org/abs/2605.15701](https://arxiv.org/abs/2605.15701)\n- **Abstract**:\n> Memory data are ubiquitous in Large Language Model (LLM)-based agents (e.g., OpenClaw and Manus). A few recent works have attempted to exploit agents'memory for improving their performance on the question-answering (QA) task, but they lack a principled mechanism for effectively modeling how memory data evolves over time and retrieving memory data effectively, leading to poor performance in memory utilization. To fill this gap, we present H-Mem, a novel memory mechanism via a hybrid structure that can not only effectively model the evolution of agent memory over a long period of time, but also provide an efficient memory retrieval approach. Particularly, H-Mem builds a temporal and semantic tree structure that allows the short-term memory data to evolve progressively into long-term memory data, where the latter provides summarized information about the former, while simultaneously constructing a knowledge graph to capture the relationships between entities in memory. Moreover, it offers an effective memory retrieval approach by exploiting the hybrid structure of the tree and graph structures. Extensive experiments on three agent memory benchmarks show that H-Mem achieves state-of-the-art performance on the QA task.\n\n---\n\n## 124. $α$-TCAV: A Unified Framework for Testing with Concept Activation Vectors\n- **Authors**: Ekkehard Schnoor , Jawher Said , Malik Tiomoko , Wojciech Samek , Alexander Jung\n- **URL**: [https://arxiv.org/abs/2605.15688](https://arxiv.org/abs/2605.15688)\n- **Abstract**:\n> Concept Activation Vectors (CAVs) are a fundamental tool for concept-based explainability in deep learning, yet their practical utility is limited by statistical instability. We analyze the stochastic nature of CAVs and the Testing with CAVs (TCAV) method, deriving the distributions of major CAV classes including PatternCAV, FastCAV, and ridge regression-based CAVs. We then identify a fundamental flaw in the standard TCAV score: its reliance on a discontinuous indicator function induces non-decaying variance in critical regimes. To address this, we introduce $\\alpha$-TCAV, a generalized framework that replaces the indicator with a parameterized smooth function, yielding a unified probabilistic formulation that subsumes both TCAV and Multi-TCAV. We characterize the induced distributions of sensitivity scores and different TCAV variants, showing that established state-of-the-art choices lack theoretical justification. We provide principled guidance on tuning the parameter in $\\alpha$-TCAV -- either to imitate Multi-TCAV at substantially lower computational cost, or to obtain a calibrated Bayes-optimal probabilistic measure of a concept's influence. Finally, our analysis yields practical recommendations that challenge established routines: most notably, allocating the full sampling budget to a single CAV rather than splitting it across several.\n\n---\n\n## 125. ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models\n- **Authors**: Jiahui Guang , Yingjie Zhu , Cuiyun Gao , Haiyan Wang , Jing Li , Di Shao , Zhaoquan Gu\n- **URL**: [https://arxiv.org/abs/2605.15687](https://arxiv.org/abs/2605.15687)\n- **Abstract**:\n> Multimodal large language models (MLLMs) may memorize sensitive cross-modal information during pretraining, making machine unlearning (MU) crucial. Existing methods typically evaluate unlearning effectiveness based on output deviations, while overlooking the generation quality after unlearning. This can easily lead to hallucinated or rigid responses, thereby affecting the usability and safety of the unlearned model. To address this issue, we propose ASRU, a controllable multimodal unlearning framework that incorporates generation quality as a core evaluation objective. ASRU first induces initial refusal behavior through activation redirection, and then optimizes fine-grained refusal boundaries using a customized reward function, thereby achieving a better trade-off between target knowledge unlearning and model utility. Experiments on Qwen3-VL show that ASRU significantly improves unlearning effectiveness (+24.6%) on average and generation quality (5.8x) on average while effectively preserving model utility, using only a small amount of retained supervision data.\n\n---\n\n## 126. Interaction-Aware Influence Functions for Group Attribution\n- **Authors**: Jaeseung Heo , Kyeongheung Yun , Youngbin Choi , Sehyun Hwang , Jungseul Ok , Dongwoo Kim\n- **URL**: [https://arxiv.org/abs/2605.15675](https://arxiv.org/abs/2605.15675)\n- **Abstract**:\n> Influence functions approximate how removing a training example changes a quantity of interest, called the target function, such as a held-out loss. To estimate the influence of a group of examples, the standard practice is to sum the individual influences of its members. However, this sum does not capture how examples jointly affect the target: a pair of examples may be redundant or complementary, but the sum cannot distinguish these cases. We propose an interaction-aware influence function that characterizes how interactions between examples influence the target. By expanding the target to second order around the trained parameters, we obtain an estimator that augments the standard sum with a pairwise interaction term that captures the alignment between two examples' effects on the target. We empirically evaluate our estimator in two settings. First, on six dataset-model pairs spanning logistic regression, MLPs, and ResNet-9, our estimator tracks leave-group-out retraining substantially better than first-order influence across all settings. Second, when used as a greedy selection rule for instruction-tuning data on Llama-3.1-8B, it beats prior influence-based and representation-similarity baselines on five of seven downstream tasks, in a regime where standard influence-based selection underperforms random selection.\n\n---\n\n## 127. VLMs Trace Without Tracking: Diagnosing Failures in Visual Path Following\n- **Authors**: Hyesoo Hong , Minsoo Kim , Wonje Jeung , Sangyeon Yoon , Dongjae Jeon , Albert No\n- **URL**: [https://arxiv.org/abs/2605.15672](https://arxiv.org/abs/2605.15672)\n- **Abstract**:\n> Vision-language models (VLMs) achieve strong performance on multimodal benchmarks, but may still lack robust control over basic visual operations. We study \\textit{line tracing}, where a model must follow a selected visual path through successive local continuations. To isolate this ability, we design controlled tracing tasks that introduce nearby competitors while reducing semantic and topological ambiguity such as crossings and overlaps. Across these tasks, even state-of-the-art VLMs frequently lose the target path and switch to nearby alternatives, especially when those alternatives look locally similar to the target. Behavioral interventions and internal analyses indicate that these failures arise from local competition: nearby similar distractors pull the model away from the true continuation. Standard remedies do not remove this bottleneck: model-size scaling provides only limited gains, reasoning partially compensates through costly substitute strategies, and explicit tracing instructions fail to recover stable path following. Finally, tests on tangled-cable scenes and metro maps with richer visual complexity show that the same path-switching failure persists beyond our controlled settings.\n\n---\n\n## 128. VAGS: Velocity Adaptive Guidance Scale for Image Editing and Generation\n- **Authors**: Yan Luo , Ahmadou Aidara , Jingyi Lu , Jeremy Moebel , Kai Han , Mengyu Wang\n- **URL**: [https://arxiv.org/abs/2605.15661](https://arxiv.org/abs/2605.15661)\n- **Abstract**:\n> Classifier-free guidance (CFG) is the primary control over how strongly text semantics move a flow-based sampler, yet standard practice holds its scale fixed across the entire ODE trajectory. This is a fundamental mismatch: early steps are noise-dominated and carry weak semantic signal, while late steps commit image structure and demand stronger directional commitment; more critically, the value of any guidance strength depends on whether the guided velocity is consistent with the model's current dynamics or working against them. We propose \\textit{Velocity-Adaptive Guidance Scale} (VAGS), a training-free replacement that multiplies the nominal scale by a bounded factor combining a temporal signal-level term with the cosine similarity between task-relevant velocity fields. For inversion-free editing, VAGS measures the alignment between source- and target-guided velocities, so edit strength at each step reflects local compatibility between preservation and transformation. For generation, VAGS-Gen uses the alignment between unconditional and conditional velocities as the analogous signal. Neither variant requires fine-tuning, auxiliary networks, or extra forward passes, and fixed CFG is recovered as a special case. On PIE-Bench and DIV2K for editing, and COCO17, CUB-200, and Flickr30K for generation, VAGS consistently improves structural fidelity and generation quality over fixed CFG and recent training-free guidance variants. The code is publicly available at this https URL .\n\n---\n\n## 129. TFZ-Tree: An Ultra-Lightweight Waveform Classification Framework for Resource-Constrained Devices\n- **Authors**: Hao Wang , Kuang Zhang , Yonggang Chi , Tianqi Zhao , Yanbo Fu , Jiaxing Guo\n- **URL**: [https://arxiv.org/abs/2605.15656](https://arxiv.org/abs/2605.15656)\n- **Abstract**:\n> Under the trend of multi-waveform coexistence in 6G IoT, intelligent receivers must first identify physical-layer waveform types before performing correct demodulation and resource scheduling. However, existing signal identification research largely focuses on symbol-level modulation classification. Research directly targeting physical-layer waveform types (e.g., OFDM, OTFS, LoRa) is not only extremely scarce but also heavily reliant on deep neural networks and complex time-frequency transforms, making deployment on resource-constrained terminals difficult. Symbol modulation classification methods themselves cannot circumvent the prerequisite of ``waveform identification first.'' To address this dual gap, we propose an ultra-lightweight waveform classification framework based on time-frequency multidimensional features with a cooperative Z-test tree (ZTree). The framework employs low-complexity time-domain feature extraction, and the classification backend adopts a ZTree optimized by Z-statistical testing, which uses hypothesis testing confidence to automatically control decision tree splitting and size, ensuring efficient execution on resource-limited processors. Tested on ten 6G candidate waveforms including OFDM, OTFS, DSSS, LoRa, and NB-IoT, the method achieves 99.5\\% average accuracy under AWGN and 87.4\\% under TDL-C multipath channels, with main confusion between OTFS and LoRa. Implemented in C on an x86 platform, single inference latency is under 4~ms. To the best of our knowledge, this is the first work achieving real-time recognition of ten IoT waveform types. Future work will target deployment acceleration on embedded MCUs. Code and dataset are open-sourced at: this https URL .\n\n---\n\n## 130. Bridging Silicon and the Hippocampus: Algebro-Deterministic Memory \"VaCoAl\" as a Substrate for Vector-HaSH and TEM\n- **Authors**: Hiroyuki Chuma , Kanji Otsuka , Yoichi Sato\n- **URL**: [https://arxiv.org/abs/2605.15652](https://arxiv.org/abs/2605.15652)\n- **Abstract**:\n> Vector-HaSH and the Tolman-Eichenbaum Machine (TEM) propose that the hippocampal-entorhinal circuit factorizes content from a prestructured grid-cell scaffold and supports compositional memory via ripple-mediated replay. Human iEEG shows that hippocampal sharp-wave ripples (SWRs) gate episodic recall, ripple-locked cortical reactivation recapitulates encoding-time patterns, and multi-hop replay fidelity decays multiplicatively along sequence length. These literatures have advanced in parallel without a shared algebraic object. We show that VaCoAl, an algebro-deterministic hyperdimensional memory architecture built from Galois-field LFSRs, supplies that object. Specifically, deterministic Galois-field diffusion provides a substrate-level alternative to Vector-HaSH's random scaffold-to-hippocampus projection that satisfies the same quasi-orthogonality requirement, with matched second-moment statistics, stronger avalanche behavior, and bit-exact reproducibility. The path-integral Confidence Ratio $CR_2$, the product of per-step $CR_1$ values along an $n$-hop chain, is the natural functional form for multi-hop replay-fidelity decay under conditional independence of per-step reactivation, providing the first algebraically tractable model of reported multiplicative decay. STDP-like path selection in VaCoAl follows from architectural demands -- similarity preservation, compositional reversibility, and bounded-frontier search -- that also constrain hippocampal computation. We further argue that VaCoAl operating regimes share architectural commitments with the EC--CA3 and EC--DG--CA3 pathways, motivating an energy-capacity-plasticity reading of why both are conserved across $>$520 Myr of evolution and primate dentate-gyrus elaboration. We prove formal correspondences, derive testable iEEG predictions, and bridge computational neuroscience and hyperdimensional engineering.\n\n---\n\n## 131. Sharp Spectral Thresholds for Logit Fixed Points\n- **Authors**: Tongxi Wang\n- **URL**: [https://arxiv.org/abs/2605.15651](https://arxiv.org/abs/2605.15651)\n- **Abstract**:\n> Softmax feedback systems are a common mathematical core of entropy-regularized reinforcement learning, logit game dynamics, population choice, and mean-field variational updates. Their central stability question is simple: when does a self-reinforcing softmax system produce a unique and globally predictable outcome? Classical theory gives a conservative answer. By treating softmax as a unit-scale response, it certifies stability only in a strongly randomized regime. We prove that the classical approach misses an entire stable regime and does not identify the point at which the qualitative change truly occurs. For finite-dimensional affine logit systems, the sharp dimension-free Euclidean threshold is $$\\beta\\|\\Pi W\\Pi\\|_{\\mathcal T\\to\\mathcal T}<2,$$ rather than the previously used condition, which certifies stability only while the softmax system remains safely over-regularized. Our theorem fills the previously missing pre-bifurcation regime, extending stability guarantees for affine softmax feedback systems to reward-responsive yet globally predictable systems. It enlarges the certified stability boundary for these systems and identifies where the model genuinely undergoes a phase transition.\n\n---\n\n## 132. Latent Video Prediction Learns Better World Models\n- **Authors**: Ali J Alrasheed , Aryan Yazdan Parast , Basim Azam , James Bailey , Naveed Akhtar\n- **URL**: [https://arxiv.org/abs/2605.15618](https://arxiv.org/abs/2605.15618)\n- **Abstract**:\n> Self-supervised video models are increasingly framed as world models, yet their evaluation remains largely confined to a single top-1 accuracy score on clean benchmarks. This leaves a major gap in comprehending their potential as world models. We present the first systematic study addressing this gap, analyzing four matched-capacity frontier video foundation models, V-JEPA 2.1, V-JEPA 2, VideoPrism, and VideoMAEv2, across five robustness axes relevant to their deployment as video world models: feature discriminability, corruption robustness, fine-grained discrimination, occlusion robustness, and sensitivity to temporal direction. Our evaluations establish that across all five axes, latent-prediction models form a distinct and consistent profile. They degrade more gracefully under pixel corruption, preserve usable class structure rather than mere geometric stability under occlusion, capture fine-grained physical contact cues without reconstructing pixels, and uniquely encode the arrow of time. These advantages can even survive task adaptation: a frozen V-JEPA 2 backbone with a lightweight attentive probe outperforms a fully fine-tuned VideoMAE and a supervised TimeSformer on corruption and occlusion robustness. Our extensive results offer concrete new evidence in favor of latent prediction for robust world modeling.\n\n---\n\n## 133. A Few GPUs, A Whole Lotta Scale: Faithful LLM Training Emulation with PrismLLM\n- **Authors**: Shaoke Xi , ChonLam Lao , Boyi Jia , Jiaqi Gao , Zhipeng Zhang , Jiamin Cao , Brian Sutioso , Erci Xu , Minlan Yu , Kui Ren , Yong Li , Zhengping Qian , Ennan Zhai , Jingren Zhou\n- **URL**: [https://arxiv.org/abs/2605.15617](https://arxiv.org/abs/2605.15617)\n- **Abstract**:\n> Large language model (LLM) training today runs on clusters spanning thousands of GPUs. While this scale enables rapid model advances, developing, debugging, and performance-tuning the training framework inevitably becomes complex and costly. This is because engineers often need to reproduce production behaviors to diagnose failures or evaluate optimizations, thereby demanding frequent and even exclusive access to production-scale clusters -- which becomes increasingly hard given that the majority of GPUs are already committed to production workloads. Simulation relies on complex performance models that are difficult to maintain, and downscaled experiments often fail to capture scale-dependent behaviors. We present PrismLLM to decouple large-scale execution from the need to access large clusters, enabling engineers to run and observe ranks of interest under faithful large-scale behavior using only a few GPUs. PrismLLM constructs a high-fidelity execution graph via a slicing-based approach that captures computation, communication, and dependencies of the target scale. Then, PrismLLM performs hybrid emulation where selected ranks execute the original program while the remaining ranks are replayed as virtual participants. Experiments on large-scale LLM training workloads show that PrismLLM accurately reproduces performance and memory behavior, achieving only 0.58\\% average error in iteration time and less than 0.01\\% error in peak GPU memory usage. PrismLLM can emulate clusters of up to 8192 GPUs using fewer than 1\\% of the physical GPUs required by the original deployment.\n\n---\n\n## 134. Offline Reinforcement Learning with Universal Horizon Models\n- **Authors**: Hojun Chung , Junseo Lee , Songhwai Oh\n- **URL**: [https://arxiv.org/abs/2605.15603](https://arxiv.org/abs/2605.15603)\n- **Abstract**:\n> Model-based reinforcement learning (RL) offers a compelling approach to offline RL by enabling value learning on imagined on-policy trajectories. However, it often suffers from compounding errors due to repeated model inference on self-generated states. While geometric horizon models (GHM) alleviate this issue through direct prediction over a discounted infinite-horizon future, they remain challenged in accurately modeling distant future states. To this end, we introduce universal horizon models (UHM), a generalization of GHM that directly predicts future states under arbitrary horizons. Leveraging this flexibility, we propose a scalable value learning method that employs a winsorized horizon distribution to stabilize training by capping excessively large horizons. Experimental results on 100 challenging OGBench tasks demonstrate that the proposed method outperforms competitive baselines, particularly on tasks with highly suboptimal datasets and those requiring long-horizon reasoning. Project page: this https URL\n\n---\n\n## 135. Pretraining Objective Matters in Extreme Low-Data FGVC: A Backbone-Controlled Study\n- **Authors**: Alexander Hackett , Srikanth Thudumu , Ginny Fisher , Mahule Roy , Aisha Sartaj , Jason Fisher\n- **URL**: [https://arxiv.org/abs/2605.15599](https://arxiv.org/abs/2605.15599)\n- **Abstract**:\n> Extreme low-data fine-grained classification is common in expert domains where labeling is expensive, yet practitioners still need principled guidance for selecting pretrained encoders. We study emerald inclusion grading with a custom dataset of labeled images across three classes and ask: under matched backbone capacity, how does pretraining objective affect downstream representation quality? We compare four frozen ViT-B/16 encoders trained with supervised classification, contrastive learning (SigLIP2), masked reconstruction (MAE), and self-distillation (DINOv3), and evaluate them with leave-one-out cross-validation using linear and nonlinear probes. To control statistical noise in the low-N regime, we use permutation testing (N=1000) on macro one-vs-rest AUC. Supervised and contrastive encoders provide the strongest linear separability (logistic AUC: 0.768 and 0.735; SVM AUC: 0.739 and 0.697), while MAE improves under nonlinear probes (XGBoost AUC: 0.713). We find that DINOv3 underperforms across probe families in this domain. These results support a practical recommendation for extreme low-data FGVC: prioritize margin-enforcing pretraining objectives when data scarcity restricts probing to linear decision rules, and consider reconstruction-style encoders when nonlinear classifiers are feasible given dataset constraints.\n\n---\n\n## 136. Embracing Biased Transition Matrices for Complementary-Label Learning with Many Classes\n- **Authors**: Tan-Ha Mai , Chao-Kai Chiang , Han-Hwa Shih , Gang Niu , Masashi Sugiyama , Hsuan-Tien Lin\n- **URL**: [https://arxiv.org/abs/2605.15586](https://arxiv.org/abs/2605.15586)\n- **Abstract**:\n> Complementary-label learning (CLL) is a weakly supervised paradigm where instances are labeled with classes they do not belong to. Despite a decade of research, CLL methods remain competitive mainly on 10-class classification, with scaling to large label spaces continuing to be an enduring bottleneck. This limitation stems from the common assumption of uniform label generation in traditional methods, which fatally dilutes the learning signal in many-class settings. In this paper, we demonstrate that this long-standing barrier can be overcome by deliberately designing a biased (non-uniform) generation process that restricts complementary labels to a subset of classes. This finding motivates us to propose Bias-Induced Constrained Labeling (BICL), a principled framework spanning data collection to training that leverages this bias. BICL enables effective learning on CIFAR-100 and TinyImageNet-200, achieving more than sevenfold accuracy improvements over traditional methods. Our findings establish a new trajectory for making CLL feasible for many classes in real-world applications.\n\n---\n\n## 137. Detecting Privilege Escalation in Polyglot Microservices via Agentic Program Analysis\n- **Authors**: Penghui Li , Hong Yau Chong , Yinzhi Cao , Junfeng Yang\n- **URL**: [https://arxiv.org/abs/2605.15569](https://arxiv.org/abs/2605.15569)\n- **Abstract**:\n> Microservices are widely adopted in modern cloud systems due to their scalability and fault tolerance. However, microservice architectures introduce significant complexity in privilege and permission control, creating risks of privilege escalation where attackers can gain unauthorized access to resources or operations. Detecting such vulnerabilities is challenging due to complex cross-service interactions, polyglot codebases, and diverse privileged operations and permission checks. We present Neo, an agentic program analysis framework that combines large language models (LLMs) with classic program analysis to address these challenges. Neo leverages an LLM-based agent that dynamically generates analysis plans, adapts code search strategies, and validates semantics. We develop code search primitives that enable Neo to perform scalable and flexible code exploration across services and languages. We evaluated Neo on 25 open-source microservice applications spanning 7 programming languages and 6.2 million lines of code. Neo uncovered 24 zero-day privilege escalation vulnerabilities and achieved 81.0% precision and 85.0% recall on a ground-truth dataset. Compared to existing program analysis and agentic solutions, Neo demonstrated significant improvements in both detection accuracy and scalability. We further showcased Neo's extensibility by applying it to other application domains and vulnerability types, uncovering 18 additional zero-day vulnerabilities.\n\n---\n\n## 138. AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs\n- **Authors**: Haizhong Zheng , Yizhuo Di , Jiahui Wang , Shuowei Jin , Xueshen Liu , Yongji Wu , Z. Morley Mao , Ion Stoica , Jiawei Zhao , Beidi Chen\n- **URL**: [https://arxiv.org/abs/2605.15565](https://arxiv.org/abs/2605.15565)\n- **Abstract**:\n> Reinforcement learning (RL) is increasingly used to improve the reasoning, coding, and tool-use capabilities of large language models, but agentic RL remains prohibitively expensive. Scaling RL to agentic LLMs requires supporting complex workloads, including multi-policy collaborative training, while efficiently using elastic, heterogeneous, and cross-region compute resources. Existing LLM RL systems support some of these capabilities, but each new extension often requires dedicated system engineering. This burden arises from trainer-centered control architectures and the lack of principled abstractions for RL system components. To address these limitations, we propose AstraFlow, a dataflow-oriented RL system that replaces conventional trainer-centered control with principled component abstractions. In AstraFlow, rollout services, dataflow management, and training are decoupled into autonomous components, enabling the system to natively support complex multi-policy agentic RL workloads and efficiently exploit diverse compute resources. We evaluate AstraFlow across math, code, search, and AgentBench workloads, showing that the same system supports multi-policy training, elastic scaling, heterogeneous cross-region execution, and composable data algorithms without system-level code changes. In multi-policy collaborative training, AstraFlow achieves comparable or better accuracy than existing RL systems while speeding up training time by 2.7x.\n\n---\n\n## 139. Operator-Controlled 6G: From Connectivity Infrastructure to Guaranteed Digital Services\n- **Authors**: David Soldani\n- **URL**: [https://arxiv.org/abs/2605.15553](https://arxiv.org/abs/2605.15553)\n- **Abstract**:\n> Sixth-generation mobile networks (6G) are approaching a structural inflection point. Five generations of vendor-led architectures have left operators procuring and operating networks they do not own, on platforms they cannot modify, with AI layers they cannot audit. This paper argues that 6G must reverse this trajectory by reordering operator priorities: Control First, Customer First, Business First, Operations First, and Technology Last. Technology should serve operator control, customer outcomes, monetizable guarantees, and software-driven operations, not dictate this http URL contributions operationalize this thesis. The 6G Control Compact defines a three-layer ownership taxonomy--own, federate, and consume--that allocates architectural sovereignty according to strategic value. The Guarantee Economy defines a five-tier, outcome-priced commercial model that converts operator control into enforceable service-level objectives. The framework is grounded in operational evidence from Rakuten Mobile, the world's first national-scale, fully cloud-native, fully Open RAN deployment, which reached full-year EBITDA profitability in FY2025. It is aligned with the ITU-R IMT-2030 framework, 3GPP 6G use cases and service requirements, NGMN recommendations, ETSI standards, O-RAN Alliance and AI-RAN Alliance specifications, IOWN Global Forum sustainability metrics, Linux Foundation initiatives, and leading industry and academic programs. A three-phase roadmap covering 2025-2027, 2027-2029, and 2029-2032 and beyond, together with seven stakeholder-specific calls to action, translates the architecture into industry commitments. The central claim is that Rakuten Mobile's deployment demonstrates the feasibility of operator-controlled 6G. Decisions made during 2026-2028 will determine whether 6G becomes a platform for guaranteed digital services or another vendor-dependent infrastructure cycle.\n\n---\n\n## 140. CTF4Nuclear: Common Task Framework for Nuclear Fission and Fusion Models\n- **Authors**: Stefano Riva , Carolina Introini , Antonio Cammi , Dean Price , Alexey Yermakov , Yue Zhao , Philippe M. Wyder , Judah Goldfeder , Jan Williams , Amy Sara Rude , Matteo Tomasetto , Joe Germany , Joseph Bakarji , Georg Maierhofer , Miles Cranmer , J. Nathan Kutz\n- **URL**: [https://arxiv.org/abs/2605.15549](https://arxiv.org/abs/2605.15549)\n- **Abstract**:\n> The demand for clean energy is ever increasing, with new nuclear technologies presenting a complementary solution to renewable energies. However, designing and operating these systems is exceptionally difficult, given the complexity of the physical phenomena that interact to form the system dynamics. While high-fidelity simulations help to understand the non-linear, multi-physics interactions within a reactor, they are computationally expensive and rarely suitable for real-time applications. Furthermore, model-based approaches are inherently sensitive to simplifying assumptions required to derive their governing equations and parameters, leading to inevitable discrepancies with real-world measurements. In contrast, Machine Learning (ML) methods have the potential to generate reliable surrogate models which may be able to quickly predict the system's behaviour. However, the number of data-driven methods that can potentially be used for this task is large and diverse. In a safety-critical setting such as nuclear engineering, a fair comparison of different ML methods, and a clear understanding of their advantages and limitations, is of paramount importance. To address this, we introduce a Common Task Framework (CTF) for ML in nuclear engineering, building upon previous efforts in dynamical systems and seismology. This CTF considers a curated set of datasets from different nuclear and nuclear-adjacent systems. The CTF evaluates the performance of a method on 12 established metrics, alongside a new paradigm focused on system monitoring from sparse measurements only. We illustrate the framework by benchmarking standard ML baselines against these datasets, revealing current method limitations. Our vision is to replace ad hoc comparisons with standardized evaluations on hidden test sets, raising the bar for rigour and reproducibility in scientific ML for the nuclear industry.\n\n---\n\n## 141. Domain-Independent Game Abstraction using Word Embedding Techniques\n- **Authors**: Juho Kim , Tuomas Sandholm\n- **URL**: [https://arxiv.org/abs/2605.15543](https://arxiv.org/abs/2605.15543)\n- **Abstract**:\n> Many games of interest in the real world are often intractably large, thereby necessitating the use of game abstraction to shrink them in size, typically by many magnitudes. Over the last two decades, there have been significant advances in game abstraction; however, the domain-specific nature (usually poker) of much of the prior work prevents those techniques from being easily generalized to other settings without extensively analyzing the game at hand. In this paper, we propose a domain-independent approach to game abstraction, which applies word embedding techniques from the field of natural language processing. Treating each action as a word and gameplay data as a corpus, word vectors can be trained to represent each action as a real-valued vector, which can then be clustered to facilitate game abstraction. We also explore the use of foundational embedding models and show that action embeddings obtained this way can capture a surprising amount of information about the underlying game. Experimental results demonstrate that our proposed game abstraction technique is effective, although it does not outperform specialized algorithms tailored to specific games.\n\n---\n\n## 142. SkiP: When to Skip and When to Refine for Efficient Robot Manipulation\n- **Authors**: Mingtong Dai , Guanqi Peng , Yongjie Bai , Feng Yan , Chunjie Chen , Lingbo Liu , Liang Lin , Xinyu Wu\n- **URL**: [https://arxiv.org/abs/2605.15536](https://arxiv.org/abs/2605.15536)\n- **Abstract**:\n> Previous imitation learning policies predict future actions at every control step, whether in smooth motion phases or precise, contact-rich operation phases. This uniform treatment is wasteful: most steps in a manipulation trajectory traverse free space and carry little task-relevant information, while a small fraction of \\emph{key} steps around contacts, grasps, and alignment demand dense, high-resolution prediction. We propose a novel \\emph{action relabeling} mechanism: at each timestep in a skip segment, we replace the behavior cloning target with the action at the entrance of the next key segment, enabling the policy to leap over redundant steps in a single decision. The resulting \\textbf{Skip Policy (SkiP)} dynamically leaps over skip segments and intensively refines actions in key segments, within a single unified network requiring no learned skip planner or hierarchical structure. To automatically partition demonstrations into key and skip segments without manual annotation, we introduce \\emph{Motion Spectrum Keying} (MSK), a fast, task-agnostic procedure that detects local motion complexity from action signals. Extensive experiments across 72 simulated manipulation tasks and three real-robot tasks show that SkiP reduces executed steps by $15$--$40\\%$ while matching or improving success rates across various policy backbones. Project page: \\texttt{ this https URL }.\n\n---\n\n## 143. Tuning-free Instruction-based Video Editing Via Structural Noise Initialization and Guidance\n- **Authors**: Song Wu , Xinyu Chen , Qian Wang , Liang Li , Zili Yi , Junlan Feng\n- **URL**: [https://arxiv.org/abs/2605.15533](https://arxiv.org/abs/2605.15533)\n- **Abstract**:\n> Video editing poses a significant challenge. While a series of tuning-free methods circumvent the need for extensive data collection and model training, they often underutilize the rich information embedded within noisy latent, leading to unsatisfactory results. To address this, we propose a \\textit{tuning-free, instruction-based} video editing framework. We approach video editing from the perspective of noisy latent: we design a Structural Noise Initialization Strategy (SNIS) to secure a superior editing starting point by assigning higher noise levels to edited regions (to facilitate content change) and lower noise levels to unedited regions (to maintain content consistency). We introduce a Noise Guidance Mechanism (NGM), which leverages the video prior in the generative model and effectively integrates rich information within the noisy latent to guide the denoising process, thereby preserving unedited content and overall visual coherence. Experiments show that our proposed method achieves better visual quality and state-of-the-art performance.\n\n---\n\n## 144. DeltaPrompts: Escaping the Zero-Delta Trap in Multimodal Distillation\n- **Authors**: Jaehun Jung , Hyunwoo Kim , Brandon Cui , Ximing Lu , David Acuna , Prithviraj Ammanabrolu , Yejin Choi\n- **URL**: [https://arxiv.org/abs/2605.15532](https://arxiv.org/abs/2605.15532)\n- **Abstract**:\n> Distillation enables compact Vision-Language Models (VLMs) to obtain strong reasoning capabilities, yet the prompts driving this process are typically chosen via simple heuristics or aggregated from off-the-shelf datasets. We reveal a critical inefficiency in this approach: up to 69% of the prompts in standard chart / document reasoning datasets are effectively zero-delta, meaning the teacher and student already induce the exact same answer distribution. Training on these prompts provides minimal learning signal, causing student improvement to rapidly saturate regardless of data scale. To escape the zero-delta trap, we return to first principles: distillation fundamentally minimizes distributional divergence, and thus a prompt is valuable only if it exposes a functional capability gap between the teacher and student. We quantify this gap through answer divergence ($\\Delta$), demonstrating that non-zero divergence is critical for effective scaling. Building on this insight, we propose a staged synthesis pipeline that repurposes existing datasets as seeds, actively targeting student failure modes to produce better prompts. The result is DeltaPrompts, a diverse dataset of 200k synthetic, high-divergence reasoning problems. We evaluate DeltaPrompts across three distinct settings: on-policy distillation with the target teacher-student pair, transfer to a novel model family without regenerating the data, and off-policy fine-tuning of a non-reasoning model. Across all scenarios, DeltaPrompts drives substantial gains, yielding up to 15% relative improvement even on top of a highly-optimized reasoning model (e.g., Qwen3-VL-8B-Thinking) -- averaged over 10 benchmarks spanning chart, document and perception-centric reasoning.\n\n---\n\n## 145. Process Rewards with Learned Reliability\n- **Authors**: Jinyuan Li , Langlin Huang , Chengsong Huang , Shaoyang Xu , Donghong Cai , Yuyi Yang , Wenxuan Zhang , Jiaxin Huang\n- **URL**: [https://arxiv.org/abs/2605.15529](https://arxiv.org/abs/2605.15529)\n- **Abstract**:\n> Process Reward Models (PRMs) provide step-level feedback for reasoning, but current PRMs usually output only a single reward score for each step. Downstream methods must therefore treat imperfect step-level reward predictions as reliable decision signals, with no indication of when these predictions should be trusted. We propose BetaPRM, a distributional PRM that predicts both a step-level success probability and the reliability of that prediction. Given step-success supervision from Monte Carlo continuations, BetaPRM learns a Beta belief that explains the observed number of successful continuations through a Beta-Binomial likelihood, rather than regressing to the finite-sample success ratio as a point target. This learned reliability signal indicates when a step reward should be trusted, enabling downstream applications to distinguish reliable rewards from uncertain ones. As one application, we introduce Adaptive Computation Allocation (ACA) for PRM-guided Best-of-N reasoning. ACA uses the learned reliability signal to stop when a high-reward solution is reliable and to spend additional computation on uncertain candidate prefixes. Experiments across four backbones and four reasoning benchmarks show that BetaPRM improves PRM-guided Best-of-N selection while preserving standard step-level error detection. Built on this signal, ACA improves the accuracy--token tradeoff over fixed-budget Best-of-16, reducing token usage by up to 33.57% while improving final-answer accuracy.\n\n---\n\n## 146. Neural Point-Forms\n- **Authors**: Bruno Trentini , Jacob Hume , Vincenzo Antonio Isoldi , Philipp Misof , Ekaterina S. Ivshina , Kelly Maggs\n- **URL**: [https://arxiv.org/abs/2605.15524](https://arxiv.org/abs/2605.15524)\n- **Abstract**:\n> Point cloud learning often rests on the premise that observed samples are noisy traces of an underlying geometric object, such as a manifold embedded in a high-dimensional feature space. Yet much of this geometry is not captured directly by coordinates, pairwise distances, or learned graph neighborhoods alone. In the smooth setting, differential forms are devices to encode higher order tangency information. In this work, we introduce a new family of principled learnable geometric features for point clouds called neural point-forms (NPFs). In the absence of a natural tangency structure, we instead use Laplacian-based techniques from Diffusion Geometry to build a discrete model for comparing differential forms on point clouds via inner products. In the continuum, submanifolds of a shared ambient feature space are represented as comparison matrices, whose entries describe how pairs of feature forms interact with extrinsic tangency information. We make this intuition precise by proving the long-run consistency of comparison matrices under standard sampling, bandwidth, density, and manifold-hypothesis assumptions. This yields a compact, efficient and permutation-invariant neural layer whose output is a learned form-comparison matrix. Across synthetic and biologically relevant experiments, we show that NPFs provide a competitive, and interpretable representation, with the strongest benefits appearing when labels depend on sampling density, manifold-like structure, or response-relevant population geometry.\n\n---\n\n## 147. On the Fragility of Data Attribution When Learning Is Distributed\n- **Authors**: Xian Gao , Bo Hui , Min-Te Sun , Wei-Shinn Ku\n- **URL**: [https://arxiv.org/abs/2605.15520](https://arxiv.org/abs/2605.15520)\n- **Abstract**:\n> Data attribution has become an important component of pricing, auditing, and governance in machine learning pipelines, yet most attribution methods implicitly assume that attribution values faithfully reflect participants' contributions. We show that this assumption can fail: a single participant in a standard distributed training workflow can substantially inflate its measured attribution value while preserving global utility. Our attribution-first attack uses latent optimization to inject small synthetic batches that preserve utility while exploiting non-IID label coverage and evaluator sensitivities. Across datasets, models, and multiple marginal-utility evaluators, the attack consistently increases the adversary's attribution value and reshapes the relative attribution structure among benign clients without degrading accuracy or triggering geometry-based defenses. These results show that attribution itself forms a new attack surface and motivate the development of attribution-robust and incentive-compatible scoring mechanisms.\n\n---\n\n## 148. DiffVAS: Diffusion-Guided Visual Active Search in Partially Observable Environments\n- **Authors**: Anindya Sarkar , Srikumar Sastry , Aleksis Pirinen , Nathan Jacobs , Yevgeniy Vorobeychik\n- **URL**: [https://arxiv.org/abs/2605.15519](https://arxiv.org/abs/2605.15519)\n- **Abstract**:\n> Visual active search (VAS) has been introduced as a modeling framework that leverages visual cues to direct aerial (e.g., UAV-based) exploration and pinpoint areas of interest within extensive geospatial regions. Potential applications of VAS include detecting hotspots for rare wildlife poaching, aiding search-and-rescue missions, and uncovering illegal trafficking of weapons, among other uses. Previous VAS approaches assume that the entire search space is known upfront, which is often unrealistic due to constraints such as a restricted field of view and high acquisition costs, and they typically learn policies tailored to specific target objects, which limits their ability to search for multiple target categories simultaneously. In this work, we propose DiffVAS, a target-conditioned policy that searches for diverse objects simultaneously according to task requirements in partially observable environments, which advances the deployment of visual active search policies in real-world applications. DiffVAS leverages a diffusion model to reconstruct the entire geospatial area from sequentially observed partial glimpses, which enables a target-conditioned reinforcement learning-based planning module to effectively reason and guide subsequent search steps. Extensive experiments demonstrate that DiffVAS excels in searching diverse objects in partially observable environments, significantly surpassing state-of-the-art methods on several datasets.\n\n---\n\n## 149. RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably\n- **Authors**: Yufeng Du , Phillip Harris , Minyang Tian , Eliu A Huerta , Srikanth Ronanki , Subendhu Rongali , Aram Galstyan , Hao Peng\n- **URL**: [https://arxiv.org/abs/2605.15514](https://arxiv.org/abs/2605.15514)\n- **Abstract**:\n> We identify intrinsic limitations of Rotary Positional Embeddings (RoPE) in Transformer-based long-context language models. Our theoretical analysis abstracts away from the specific content of the context and depends only on its length. We prove that as context length increases, RoPE-based attention becomes unpredictable and loses two properties that are central to its effectiveness. First, it loses its locality bias: RoPE is no more likely to favor nearer positions than substantially farther ones. Second, it loses consistency in token relevance: a key vector that receives a higher attention score than an alternative at one position may receive a lower score at another. In both cases, the probability of failure approaches 0.5, no better than random guessing. We further prove that the attention score can remain unchanged when a key token is moved to a different position, or even replaced by a different token, indicating a failure to distinguish positions or tokens. Adjusting the RoPE base trades off distinguishing positions against distinguishing tokens but cannot preserve both at the same time. Increasing the RoPE base hyperparameter, a common practice in today's long-context models, helps distinguish different tokens, but inevitably sacrifices the ability to distinguish positions. Our empirical analysis shows that multi-head, multi-layer architectures are insufficient to overcome these limitations. Our findings suggest that fundamentally new mechanisms for encoding position and token order may be needed in future Transformer long-context language models.\n\n---\n\n## 150. PrismQuant: Rate-Distortion-Optimal Vector Quantization for Gaussian-Mixture Sources\n- **Authors**: Bumsu Park , Chanho Park , Youngmok Park , Namyoon Lee\n- **URL**: [https://arxiv.org/abs/2605.15507](https://arxiv.org/abs/2605.15507)\n- **Abstract**:\n> For a Gaussian source under mean-squared error (MSE), classical transform coding is rate--distortion (RD) optimal: the Karhunen--Loeve transform (KLT) diagonalizes the covariance, reverse waterfilling allocates the bits, and scalar quantization closes the loop. This elegant story breaks down for multimodal sources, where no single covariance can capture heterogeneous local geometries, and the RD function loses its closed form. We revisit this problem through Gaussian-mixture sources and develop a constructive RD theory for them. Our key finding is that the mixture structure incurs only a component label cost. Conditioned on the active mixture component, each branch is Gaussian; the challenge is allocating bits across heterogeneous branches. We prove that the genie-aided conditional RD function is governed by a single global reverse-waterfilling level shared across all components and eigenmodes. Building on this result, we introduce PrismQuant, which transmits the component label losslessly and encodes the residual using the component-matched KLT, followed by scalar quantization, achieving a rate of H(C)/n bits per source dimension of the converse, with a vanishing asymptotic gap. We further develop a practical implementation based on EM-driven Gaussian-mixture learning, component-adaptive KLTs, and entropy-constrained scalar quantization (ECSQ). Experiments on synthetic Gaussian mixtures show that PrismQuant closely approaches the theoretical RD bound, while experiments on real-world channel-state-information (CSI) data demonstrate competitive or superior performance compared with transformer-based learned codecs at more than one order of magnitude smaller model size.\n\n---\n\n## 151. Learning with Conflicts of Interest\n- **Authors**: Nischal Aryal , Arash Termehchy , Ali Vakilian , Marianne Winslett\n- **URL**: [https://arxiv.org/abs/2605.15504](https://arxiv.org/abs/2605.15504)\n- **Abstract**:\n> Financial, social, and political factors often prevent the interests of the owners of ML systems and services and their users from being perfectly aligned. ML systems often produce biased information that can influence users to make decisions that are not in their best interest. Current solution approaches require ML systems to implement protocols to mitigate their biases. However, ML system owners usually do not have any incentive to implement these protocols and often argue that it limits their freedom of expression or business. We believe that a successful solution to this problem must recognize the conflict of interest between the ML systems and their users, and use this information to protect users against information that adversely influences their decisions while allowing users to safely benefit from these systems. To this end, we propose a game-theoretic framework that models the interaction between ML systems and users with conflicts of interest. We present scalable algorithms with theoretical guarantees that maximize the amount of desired information and actions and minimize the amount of biased and manipulative actions in interaction with ML systems.\n\n---\n\n## 152. Ghosted Layers: Unconstrained Activation Alignment for Recovering Layer-Pruned LLMs\n- **Authors**: Vincent-Daniel Yun , Junhyuk Jo , Sai Praneeth Karimireddy , Sunwoo Lee\n- **URL**: [https://arxiv.org/abs/2605.15491](https://arxiv.org/abs/2605.15491)\n- **Abstract**:\n> Layer pruning removes entire Transformer decoder blocks from large language models, but introduces a mismatch between the hidden state received by the next surviving layer and the distribution it was trained to process, leading to significant performance degradation. We propose Ghosted Layers, a training-free recovery module that addresses this issue by solving a boundary activation alignment problem. Our method derives a closed-form optimal linear operator from a small calibration set to reconstruct the activation discrepancy introduced by the pruned layers. We show that this solution corresponds to the unconstrained optimum of the alignment objective, whereas existing methods are restricted to constrained solutions over limited operator subspaces. Experiments across multiple LLM backbones and pruning strategies demonstrate that our method consistently improves accuracy and perplexity over prior training-free baselines, while preserving the efficiency gains of layer pruning.\n\n---\n\n## 153. Hybrid LLM-based Intelligent Framework for Robot Task Scheduling\n- **Authors**: Swayamjit Saha , Subhabrata Das , Haonan Duan , Xiao-Yang Liu\n- **URL**: [https://arxiv.org/abs/2605.15486](https://arxiv.org/abs/2605.15486)\n- **Abstract**:\n> This study introduces intelligent frameworks that use Large Language Models (LLMs) to improve task scheduling for construction robots. The LLM is fed with key data about the desired task, such as agent action abilities, and the desired end goal to be achieved. A well-balanced allocation strategy is developed, optimizing both time efficiency and resource utilization. Our system utilizes a Natural Language Processing interface to streamline communication with construction professionals and adapt in real-time to unexpected site conditions. We concurrently use two LLM agents, specifically generator (GPT-4) and supervisor (Gemma 3/Llama 4/Mistral 7b) LLM agents to provide a more precise task schedule. We evaluate the proposed methodology using a straightforward scenario and provide metric scores to prove the efficacy of the frameworks. Our results highlight that the implementation of LLMs is crucial in construction operational tasks including robots.\n\n---\n\n## 154. Residual Reinforcement Learning for Robot Teleoperation under Stochastic Delays\n- **Authors**: Kaize Deng , Zewen Yang\n- **URL**: [https://arxiv.org/abs/2605.15480](https://arxiv.org/abs/2605.15480)\n- **Abstract**:\n> Stochastic communication delays in teleoperation introduce signal discontinuities that undermine control stability and degrade control performance. Consequently, the conventional reinforcement learning (RL) methods struggle with the delayed observations due to the delay-induced observations, leading to high-frequency chattering. To address this, we propose a hybrid control framework, delay-resilient RL, integrating a state estimator utilizing Long Short-Term Memory (LSTM) with a residual RL policy, which is resilient to stochastic delays. The LSTM reconstructs smooth, continuous state estimates from delayed observations, enabling the RL agent to learn a residual torque compensation policy that balances tracking accuracy with velocity smoothness. Experimental validation on Franka Panda robots demonstrates that our approach significantly outperforms the state-of-the-art baselines, ensuring robust and stable teleoperation even under high-variance stochastic delays.\n\n---\n\n## 155. Retrieval-Augmented Large Language Models for Schema-Constrained Clinical Information Extraction\n- **Authors**: A H M Rezaul Karim , Ozlem Uzuner\n- **URL**: [https://arxiv.org/abs/2605.15467](https://arxiv.org/abs/2605.15467)\n- **Abstract**:\n> Conversational nurse-patient transcripts contain actionable observations, but converting these transcripts into structured representations at scale remains challenging. Documentation burden is substantial, with prior studies showing clinicians spend large portions of their workday on documentation and related desk work rather than direct patient care. MEDIQA-SYNUR focuses on observation extraction from conversational nurse-patient transcripts, requiring systems to normalize these narratives into a predefined schema with value-type constraints. We propose a modular retrieval-augmented generation (RAG) pipeline that uses the training set as an exemplar corpus, combines schema-constrained prompting (full schema vs. pruned candidate schema), deterministic schema-based postprocessing, and a second-pass audit, with two LLM backbones: Llama-4-Scout-17B-16E-Instruct and GPT-5.2 with corresponding embedding models for RAG. Our best configuration uses GPT-5.2 with full schema, RAG, and a second-pass auditing, achieving 80.36% F1 score. Overall, our results show that RAG consistently improves performance, while the optimal degree of schema constraint depends on the model, and second-pass auditing yields modest additional gains by correcting residual schema-adherence errors.\n\n---\n\n## 156. GRLO: Towards Generalizable Reinforcement Learning in Open-Ended Environments from Zero\n- **Authors**: Shangjian Yin , Yu Fu , Yue Dong , Zhouxing Shi\n- **URL**: [https://arxiv.org/abs/2605.15464](https://arxiv.org/abs/2605.15464)\n- **Abstract**:\n> Post-training has become a crucial step for unlocking the capabilities of large language models, with reinforcement learning (RL) emerging as a critical paradigm. Recent RL-based post-training has increasingly split into two paradigms: reinforcement learning from human feedback (RLHF), which optimizes models using human preference signals in target domains, and reinforcement learning from verifiable rewards (RLVR), which operates in verifier-backed environments. The latter has dominated recent reasoning-oriented post-training because it delivers stronger gains and higher efficiency on domain-specific tasks (e.g., reasoning). However, although in-domain RL training achieves promising performance, it still requires a substantial amount of GPU compute, which remains a major barrier to broad adoption. In this work, we study the generalization ability of RLHF learned from scratch from a small set of interactions in open-ended environments, and investigate whether the conversational abilities it explicitly acquires can implicitly transfer to downstream tasks such as mathematical reasoning and code generation, namely GRLO. Specifically, on Qwen3-4B-Base backbone, GRLO improves the average performance across all domains from 24.1 to 63.1 with only 5K prompts and 22.7 GPU hours, requiring about $46\\times$ less data and $68\\times$ less compute than a strong in-domain RLVR baseline. The resulting model is even competitive with Qwen's released post-trained models which required a much larger training cost. Notably, a subsequent in-domain RLVR stage brings only selective gains, mainly on harder competition-math benchmarks. We hope GRLO offers a simple and efficient recipe for building broadly capable post-trained models. Our code and data will be available at: \\href{ this https URL }{ this https URL }.\n\n---\n\n## 157. DrugSAGE:Self-evolving Agent Experience for Efficient State-of-the-Art Drug Discovery\n- **Authors**: Yikun Zhang , Xiwei Cheng , Tianyu Liu , Yuanqi Du , Wengong Jin\n- **URL**: [https://arxiv.org/abs/2605.15461](https://arxiv.org/abs/2605.15461)\n- **Abstract**:\n> Building state-of-the-art (SOTA) predictive models for drug discovery requires expensive search over tools, architectures, and training strategies. Current LLM-based agents can find SOTA solutions through extensive trial and error, but they do not retain the experience accumulated along the way and therefore pay the full search cost on every new task. We propose \\method (Self-evolving Agent Experience), a framework that accumulates and reuses experience across tasks to build SOTA drug discovery models efficiently. \\method maintains a cross-task memory of verified skills, statistical evidence about effective strategies, and a record of recurring errors and their fixes. In some cases, \\method transfers a working solution directly without test-time search. In 33 molecular property prediction tasks, \\method ranks first among nine SOTA agents in a single-task setting. With memory accumulated from 16 smaller tasks, \\method achieves an averaged normalized score of 0.935 on 17 held-out tasks in a cross-task evaluation setting and outperforms all baseline agents by 10-30\\% in a zero-test-time search regime. In summary, our work shows the advantage of cross-task memory for efficient SOTA model development in drug discovery.\n\n---\n\n## 158. Differentially Private Motif-Preserving Multi-modal Hashing\n- **Authors**: Zehua Cheng , Wei Dai , Jiahao Sun\n- **URL**: [https://arxiv.org/abs/2605.15460](https://arxiv.org/abs/2605.15460)\n- **Abstract**:\n> Cross-modal hashing enables efficient retrieval by encoding images and text into compact binary codes. State-of-the-art methods rely on semantic similarity graphs derived from user interactions for supervision, yet these graphs encode sensitive behavioral patterns vulnerable to link reconstruction attacks. Existing privacy-preserving approaches fail on graph-structured data: Differentially Private SGD destroys relational motifs by treating samples independently, while graph synthesis methods suffer from unbounded local sensitivity in scale-free networks, hub nodes cause single-edge modifications to alter triangle counts by $\\mathcal{O}(N)$, necessitating prohibitive noise injection. We term this phenomenon Hubness Explosion. We propose DMP-MH, a Sanitize-then-Distill framework that decouples privacy from representation learning. Our approach first bounds sensitivity by deterministically clipping node degrees, capping the $L_2$-sensitivity of triangle motifs independently of dataset size. A sanitized synthetic graph is then generated via Noisy Mirror Descent under $(\\epsilon,\\delta)$-Edge Differential Privacy. Finally, dual-stream hashing networks distill this topology using a holistic structural loss that enforces cross-modal alignment. Evaluated on MIRFlickr-25K and NUS-WIDE under a strict inductive protocol, DMP-MH outperforms private baselines by up to 11.4 mAP points while retaining up to 92.5% of non-private performance.\n\n---\n\n## 159. RIDE: Retinex-Informed Decoupling for Exposing Concealed Objects\n- **Authors**: Chunming He , Rihan Zhang , Dingming Zhang , Chengyu Fang , Longxiang Tang , Jingjia Feng , Fengyang Xiao , Sina Farsiu\n- **URL**: [https://arxiv.org/abs/2605.15450](https://arxiv.org/abs/2605.15450)\n- **Abstract**:\n> Concealed Object Segmentation (COS) encompasses a family of dense-prediction tasks, including camouflaged object detection, polyp segmentation, transparent object detection, and industrial defect inspection, where targets are visually entangled with their surroundings through different physical mechanisms. Existing methods either operate directly on RGB images or employ \\emph{heterogeneous} decompositions (\\eg, Fourier, wavelet) that redistribute spatial evidence across scale/frequency coefficients, making pixel-aligned cues less direct. We introduce a fundamentally different perspective: \\textbf{homogeneous image decomposition} via Retinex theory, which factorizes an image into illumination and reflectance components within the \\emph{same} spatial domain. Our key insight is that visual entanglement enforces appearance matching in the composite space, but this does \\emph{not} necessitate simultaneous matching in both component spaces, a phenomenon we formalize as the \\textbf{Discriminability Gap Theorem}. Crucially, we show that across diverse COS sub-tasks, the underlying physical processes systematically anti-correlate illumination and reflectance differences, yielding theoretical guarantees that Retinex decomposition preserves or strictly improves total foreground--background discriminability across the full physical regime, with anti-correlation maximizing the gain. Building on this, we propose \\textbf{RIDE} comprising: (i) a Task-Driven Retinex Decomposition module that learns segmentation-optimal factorizations end-to-end; (ii) a Discriminability Gap Attention mechanism that adaptively exploits where decomposition helps; and (iii) a Camouflage-Breaking Contrastive loss operating in reflectance feature space.\n\n---\n\n## 160. Runtime-Structured Task Decomposition for Agentic Coding Systems\n- **Authors**: Shubhi Asthana , Bing Zhang , Chad DeLuca , Hima Patel , Ruchi Mahindru\n- **URL**: [https://arxiv.org/abs/2605.15425](https://arxiv.org/abs/2605.15425)\n- **Abstract**:\n> Agentic coding systems increasingly use large language models (LLMs) for software engineering tasks such as debugging, root cause analysis, and code review. However, many existing systems encode task logic, execution flow, and output generation inside monolithic prompts. This design creates brittle behavior, limited debuggability, and high retry costs because failures often require rerunning the full workflow. We present runtime-structured task decomposition, an architectural approach in which task partitioning and execution flow are managed through executable control logic rather than prompt structure alone. LLMs are used only for focused judgment tasks, and outputs are validated against predefined schemas before downstream execution. We evaluate this approach on two software engineering workloads using three configurations: monolithic execution, static decomposition with fixed subtasks and no runtime branching, and runtime-structured decomposition. Each configuration was evaluated across 10 runs. Our results show that decomposition alone does not necessarily reduce retry cost. In the Kubernetes root cause analysis workload, the static decomposition baseline produced a retry cost of 1,632 +/- 145 tokens versus 904 +/- 17 tokens for the monolithic baseline because failures forced reruns of downstream subtasks. A similar pattern appeared in the multi-file debugging workload, where the static baseline consumed 933 tokens compared to 703 tokens for the monolithic system. The runtime-structured approach reran only failed subtasks, reducing retry costs to 436 +/- 132 tokens for root cause analysis and 460 tokens for debugging. Overall, the approach achieved up to 51.7% lower retry cost than monolithic systems and 73.2% lower retry cost than static decomposition baselines, improving efficiency, debuggability, and operational reliability in agentic coding systems.\n\n---\n\n## 161. MR2-ByteTrack: CNN and Transformer-based Video Object Detection for AI-augmented Embedded Vision Sensor Nodes\n- **Authors**: Luca Bompani , Manuele Rusci , Luca Benini , Daniele Palossi , Francesco Conti\n- **URL**: [https://arxiv.org/abs/2605.15423](https://arxiv.org/abs/2605.15423)\n- **Abstract**:\n> Modern smart vision sensors need on-device intelligence to process video streams, as cloud computing is often impractical due to bandwidth, latency, and privacy constraints. However, these sensory systems typically rely on ultra-low-power microcontrollers (MCUs) with limited memory and compute, making conventional video object detection methods, which require feature storage or multi-frame buffering, unfeasible. To address this challenge, we introduce Multi-Resolution Rescored ByteTrack (MR2-ByteTrack), a Video Object Detection (VOD) method tailored for MCU-based embedded vision nodes. MR2-ByteTrack reduces computational cost by alternating between full- and low-resolution inference, while linking detections across frames via ByteTrack and correcting misclassifications through the Rescore algorithm, which applies probability union rules to aggregate detection confidence scores across frames. We apply our approach to both a CNN-based detector and a Transformer-based model, demonstrating its generality across architectures with fundamentally different spatial processing. Experiments on ImageNetVID demonstrate that MR2-ByteTrack maintains accuracy, achieving mAP scores of up to 49.0 for the CNN-based models and 48.7 for the Transformer, while reducing multiply-accumulate operations by as much as 53\\% for the CNNs and 32\\% for the Transformer. When deployed on GAP9, an ultra-low-power RISC-V multicore MCU, our method yields up to 55\\% energy savings compared to processing only full-resolution images, enabling the first real-time Transformer-based VOD on an MCU-class embedded vision node. Code available at this https URL\n\n---\n\n## 162. $f$-Trajectory Balance: A Loss Family for Tuning GFlowNets, Generative Models, and LLMs with Off- and On-Policy Data\n- **Authors**: Jake Fawkes , Jason Hartford\n- **URL**: [https://arxiv.org/abs/2605.15417](https://arxiv.org/abs/2605.15417)\n- **Abstract**:\n> In GFlowNets and variational inference, it has been shown that the mean square error between target and model log probabilities is an effective, low variance, surrogate loss for training generative models. This loss has the property that when evaluated \\emph{on-policy} its gradients correspond to those of the KL divergence, while \\emph{off-policy} it remains a valid loss with the same global minimizer. In this work, we demonstrate that this construction can be extended to the whole family of $f$-divergences, leading to a family of losses whose on-policy gradients are that of the corresponding $f$-divergence, but retain the same global minimizer off-policy. Specifically, we show that the on-policy gradients lead to a one to one correspondence between translation invariant loss functions on the target and model log probabilities, and $f$-divergences. This equivalence allows us to design new surrogate loss functions for tuning a wide class of generative models that inherit the properties of the corresponding $f$-divergence, such as being more mode covering, whilst being applicable to off-policy data. We apply our losses on a range of tasks, including classic synthetic examples, SynFlowNets for molecule discovery, and asynchronous large language model (LLM) tuning, demonstrating that our models retain their predicted properties on- and off-policy in a wide class of generative models.\n\n---\n\n## 163. Margin-Adaptive Confidence Ranking for Reliable LLM Judgement\n- **Authors**: Gaojie Jin , Yong Tao , Lijia Yu , Tianjin Huang\n- **URL**: [https://arxiv.org/abs/2605.15416](https://arxiv.org/abs/2605.15416)\n- **Abstract**:\n> Jung et al. (2025) introduce a hypothesis testing framework for guaranteeing agreement between large language models (LLMs) and human judgments, relying on the assumption that the model's estimated confidence is monotonic with respect to human-disagreement risk. In practice, however, this assumption may be violated, and the generalization behavior of the confidence estimator is not explicitly analyzed. We mitigate these issues by learning a dedicated confidence estimator instead of relying on heuristic confidence signals. Our approach leverages simulated annotator diversity and a margin-based ranking formulation to explicitly model how confidently an LLM distinguishes between human-agreement and human-disagreement cases. We further derive generalization guarantees for this estimator, revealing a margin-dependent trade-off that informs the design of an adaptive estimator training procedure. When integrated into fixed-sequence testing, the learned confidence estimator yields improved ranking accuracy and empirically strengthens the monotonic relationship between confidence and disagreement risk, leading to higher success rates in satisfying target agreement levels across multiple datasets and judge models.\n\n---\n\n## 164. From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery\n- **Authors**: Lingzhe Zhang , Tong Jia , Yunpeng Zhai , Zixuan Xie , Chiming Duan , Minghua He , Philip S. Yu , Ying Li\n- **URL**: [https://arxiv.org/abs/2605.15412](https://arxiv.org/abs/2605.15412)\n- **Abstract**:\n> Modern quantitative trading increasingly relies on systematic models to extract predictive signals from large-scale financial data, where alpha factor discovery plays a central role in transforming market observations into tradable signals. Recent LLM-based methods have shown promise in automating factor generation, but most of them still rely on prompt-level generation--evaluation--feedback loops for iterative optimization. As the loop becomes longer, repeatedly appended historical candidates and feedback can cause context explosion, increase inference cost, dilute useful information, and introduce feedback drift. Moreover, these methods often depend on very large LLMs whose stable generation preferences may lead to structurally similar expressions, redundant candidates, and search stagnation. To address these limitations, we propose \\textsc{QuantEvolver}, a self-evolving alpha factor discovery framework based on reinforcement fine-tuning. Instead of accumulating feedback in the prompt, \\textsc{QuantEvolver} converts executable quantitative evaluation into policy updates, enabling a Miner LLM to internalize historical optimization experience through parameter learning. Specifically, \\textsc{QuantEvolver} constructs high-quality seed factors, builds diverse seed--time-window training tasks, generates executable Factor DSL expressions, evaluates them through Regime Backtest, and optimizes the Miner LLM with Diversity-Complementarity Reward. During training, high-quality factors are continuously accumulated in a Mined Factor Database, which serves as the final discovered factor library. Extensive experiments on three realistic market benchmarks demonstrate the effectiveness of \\textsc{QuantEvolver}, which consistently improves the primary evaluation metric of each task over existing LLM-based alpha factor discovery baselines, produces higher-quality and more complementary factor pools.\n\n---\n\n## 165. Diagonal Adaptive Non-local Observables on Quantum Neural Networks\n- **Authors**: Huan-Hsin Tseng , Yan Li , Hsin-Yi Lin , Samuel Yen-Chi Chen\n- **URL**: [https://arxiv.org/abs/2605.15410](https://arxiv.org/abs/2605.15410)\n- **Abstract**:\n> Adaptive Non-local Observables (ANOs) have shown that making quantum observables dynamic can substantially enlarge the function space of Variational Quantum Algorithms, partly shifting hardware demands from circuit synthesis to measurement design. However, this advantage is accompanied by a steep increase in the number of parameters, as well as the classical optimization cost for varying general Hermitian observables. We propose a special form of ANO that significantly reduces this burden by considering only diagonal observables paired with quantum circuits. Mathematically, this is equivalent to the full ANO of a large parameter space since diagonal matrices are canonical representatives of the ANO space modulo unitary similarity. As a result, Diagonal ANO retains the same capability of full ANO while reducing $k$-local observable complexity from $O(4^k)$ to $O(2^k)$ and lowering the corresponding measurement-side classical computation. In this sense, diagonal ANO preserves much of the benefit of full ANO while encompassing conventional VQCs as a special case.\n\n---\n\n## 166. Amortized Energy-Based Bayesian Inference\n- **Authors**: Hojjat Kaveh , Ricardo Baptista , Andrew M. Stuart\n- **URL**: [https://arxiv.org/abs/2605.15407](https://arxiv.org/abs/2605.15407)\n- **Abstract**:\n> We consider amortized Bayesian inference for nonlinear inverse problems in settings where only samples from the joint distribution of parameters and observations are available. Classical methods such as Markov chain Monte Carlo require solving a new inference problem for each observation, which can be computationally prohibitive when inference must be repeated many times. We propose a transport-based approach that learns an observation-dependent map pushing forward a reference measure to approximate the posterior distribution. The map is trained by minimizing an averaged energy-distance objective between the true posterior and the learned pushforward. This formulation is likelihood-free, requiring only joint samples, and avoids density evaluation, invertibility constraints, and Jacobian determinant computations. For function-space inverse problems with Gaussian priors, we parameterize the transport map as the identity plus a perturbation in the Cameron-Martin space of the prior, preserving absolute continuity with respect to the prior. In infinite-dimensional settings, the map is represented using neural operators. We illustrate the method on a finite-dimensional nonlinear inverse problem and two PDE-constrained inverse problems arising in porous medium flow and seismic inversion. The results show that the learned transport captures posterior structure, including multimodality and dominant modes, while enabling fast posterior sampling for new observations.\n\n---\n\n## 167. Breakeven complexity: A new perspective on neural partial differential equation solvers\n- **Authors**: Yijing Zhang , Nicholas Roberts , Tanya Marwah , Mikhail Khodak\n- **URL**: [https://arxiv.org/abs/2605.15399](https://arxiv.org/abs/2605.15399)\n- **Abstract**:\n> Neural surrogate solvers of partial differential equations (PDEs) promise dramatic speedups over numerical methods, especially in scenarios requiring many solves. However, current accuracy-based evaluations do not fully consider two central issues: (1) neural solvers incur substantial up-front costs for data generation, training, and tuning; and (2) classical solvers can also generate low-fidelity solutions at a sufficiently low simulation cost. To explicitly account for these realities and fully incorporate end-to-end costs, we propose an evaluation framework centered on breakeven complexity, a metric that counts the forward solves before a learned solver is cost-effective relative to an error-equivalent traditional solver. To evaluate this measure, we apply scaling laws to determine how much training budget to allocate to data generation and discuss how to achieve smooth error-matching in diverse settings. We evaluate the breakeven complexity of multiple neural PDE solvers on three PDEs on 2D periodic domains from APEBench and a novel benchmark of flows past multiple obstacles generated by the GPU-native PyFR code. Among other findings, our results suggest that neural PDE solvers become more effective as problems get harder in terms of cost, dimension, rollout, physics regime (e.g. higher Reynolds number), etc.\n\n---\n\n## 168. Representation Without Reward: A JEPA Audit for LLM Fine-Tuning\n- **Authors**: Biswa Sengupta\n- **URL**: [https://arxiv.org/abs/2605.15394](https://arxiv.org/abs/2605.15394)\n- **Abstract**:\n> Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outputs. For autoregressive language-model fine-tuning the principle entails a stricter requirement: the induced hidden-state geometry must reach the language-model head \\emph{and} improve the decoded task metric. We test that requirement under a fixed Llama-3.2-1B-Instruct LoRA harness on natural-language-to-regex generation, comparing twenty-two training-time auxiliaries across trajectory-shape regularisation, distributional constraints, predictor/target asymmetry, Fisher-metric Jacobi residuals, and a decoder-visible JEPA objective constructed to lie in cross-entropy's positive cone. The empirical answer is a structured null: several auxiliaries clear single-cell paired $\\alpha = 0.10$ without correction (T3-Local at $\\Delta = +2.53$~pp, $p = 0.003$ being the strongest), but none survives Bonferroni or Holm--Bonferroni at the relevant family-wise threshold, even though many change curvature, anisotropy, variance, and gradient direction. Decoder-visible JEPA yields the first positive auxiliary--cross-entropy gradient cosine in the study, yet exact match remains inside seed noise; a full-fine-tuning replication of the same auxiliary at $n = 5$ seeds reproduces the null on both benchmarks (TURK: $\\Delta = +0.04$~pp, $p_{\\text{paired} } = 0.96$; SYNTH: $\\Delta = +0.52$~pp, $p_{\\text{paired} } = 0.28$), so the null is robust across LoRA and full fine-tuning for the decoder-visible construction. Hidden-state representation work and decoded-task accuracy are therefore weakly coupled in this regime; we accordingly reframe LLM-domain JEPA evaluation as a coupling problem, in which the operative question is under which metrics useful hidden geometry becomes decoder-visible task signal.\n\n---\n\n## 169. PanoWorld: Geometry-Consistent Panoramic Video World Modeling\n- **Authors**: Le Jiang , Xiangyu Bai , Bishoy Galoaa , Shayda Moezzi , Caleb James Lee , Tooba Imtiaz , Edmund Yeh , Jennifer Dy , Yanzhi Wang , Sarah Ostadabbas\n- **URL**: [https://arxiv.org/abs/2605.15391](https://arxiv.org/abs/2605.15391)\n- **Abstract**:\n> We present PanoWorld, a panoramic video world model that generates geometry-consistent 360$\\degree$ video from a single image and a caption. Existing panoramic video methods optimize primarily for visual realism and do not explicitly constrain the underlying 3D scene state, producing outputs that appear plausible yet exhibit inconsistent depth, broken correspondences, and implausible motion across the spherical surface. We address this gap by framing panoramic video generation as a geometry- and dynamics-consistent latent state modeling problem rather than pure visual synthesis. Building on a pre-trained perspective video world model, we introduce two lightweight regularizers: a depth consistency loss against pseudo ground-truth panoramic depth, and a trajectory consistency loss that supervises the 3D world-frame positions of tracked points across time. We further apply spherical-geometry-aware adaptation to the conditioning and positional encoding. We additionally introduce PanoGeo, a unified geometry-aware panoramic video dataset with consistent depth, trajectory, and prompt annotations across diverse real and synthetic sources, used for both training and stratified evaluation. Experiments show that PanoWorld improves geometric consistency over prior panoramic generation methods while maintaining competitive visual realism, establishing that panoramic video generation must be treated as a geometric modeling problem to support the holistic spatial understanding requirements of embodied AI applications. Code is available at this https URL .\n\n---\n\n## 170. Is One Score Enough? Rethinking the Evaluation of Sequentially Evolving LLM Memory\n- **Authors**: Songwei Dong , Zihan Chen , Chengshuai Shi , Peng Wang , Jundong Li , Cong Shen\n- **URL**: [https://arxiv.org/abs/2605.15384](https://arxiv.org/abs/2605.15384)\n- **Abstract**:\n> Memory plays a central role in enabling large language models (LLMs) to operate over sequential tasks by accumulating and reusing experience over time. However, existing evaluations of LLM memory mostly rely on aggregate metrics such as final hold-out accuracy or cumulative online performance, which can obscure critical failure modes such as forgetting and negative transfer. In this paper, we introduce SeqMem-Eval, a diagnostic evaluation framework for sequentially evolving LLM memory. Drawing inspiration from continual learning, it targets a test-time setting in which memory is external, prompt-mediated, and updated without modifying model parameters. Rather than focusing only on final performance, SeqMem-Eval evaluates how memory states evolve, generalize, consolidate experience, and retain useful information during sequential inference. Specifically, it measures online utility, hold-out generalization, backward transfer, and forgetting, providing a finer-grained view of memory quality. Through extensive experiments across diverse tasks and memory methods, we show that higher final or cumulative accuracy does not necessarily imply better memory quality: many methods exhibit strong performance gains while suffering from substantial forgetting or negative transfer. Moreover, different memory designs exhibit distinct trade-offs between adaptability and stability that remain invisible under standard evaluation metrics.\n\n---\n\n## 171. ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing\n- **Authors**: Blaž Rolih , Matic Fučka , Filip Wolf , Luka Čehovin Zajc\n- **URL**: [https://arxiv.org/abs/2605.15375](https://arxiv.org/abs/2605.15375)\n- **Abstract**:\n> Remote sensing change detection (RSCD) aims to localise changes between two images of the same geographic region. In practice, change masks often follow region-level annotation conventions rather than purely local appearance differences, making them context-dependent and occasionally ambiguous. Most state-of-the-art methods utilise per-pixel discriminative classification, which produces a single prediction per input and fails to explicitly model the changed region as a coherent whole. A natural alternative is generative formulation, which can model a distribution of plausible masks, enabling sampling to capture ambiguity and encourage global consistency. However, existing generative RSCD approaches typically lag behind strong discriminative baselines due to the high computational cost of pixel-space generation and the complexity of their conditioning mechanisms. To address the limitations of prior discriminative and generative methods, we propose ChangeFlow, a generative framework that reformulates change detection as the synthesis of a change mask in latent space via rectified flow. ChangeFlow is guided by a structured yet lightweight conditioning signal, and its stochastic design naturally supports sampling-based prediction ensembling. Namely, aggregating multiple predicted change masks improves robustness, while sample agreement provides a practical confidence estimation that highlights ambiguous regions. Across four benchmarks, ChangeFlow achieves an average F1 of 80.4\\%, improving by 1.3 points on average over the previous best method, while maintaining inference speed comparable to recent strong baselines. Project page: this https URL\n\n---\n\n## 172. PACER: Acyclic Causal Discovery from Large-Scale Interventional Data\n- **Authors**: Ramon Viñas Torné , Sílvia Fàbregas Salazar , Soyon Park , Ivo Alexander Ban , Artyom Gadetsky , Nikita Doikov , Maria Brbić\n- **URL**: [https://arxiv.org/abs/2605.15353](https://arxiv.org/abs/2605.15353)\n- **Abstract**:\n> Inferring the structure of directed acyclic graphs (DAGs) from data is a central challenge in causal discovery, particularly in modern high-dimensional settings where large-scale interventional data are increasingly available. While interventional data can improve identifiability, existing methods remain limited by soft acyclicity constraints, leading to optimization over invalid cyclic graphs, numerical instability, and reduced scalability. We introduce PACER (Perturbation-driven Acyclic Causal Edge Recovery), a scalable framework for causal discovery that guarantees acyclicity by construction. PACER parameterizes a distribution over DAGs through a joint model of variable permutations and edge probabilities, enabling direct optimization over valid causal structures without surrogate penalties. The framework supports a unified likelihood-based treatment of observational and interventional data, flexible conditional density models, and the incorporation of structural prior knowledge. For linear-Gaussian mechanisms, we derive closed-form expressions for the expected interventional log-likelihood and its gradients, yielding substantial computational gains. Empirically, PACER matches or exceeds state-of-the-art methods on protein signaling and large-scale genetic perturbation benchmarks, while scaling efficiently to networks with thousands of variables and achieving up to two orders of magnitude speedups over penalty-based differentiable approaches. These results demonstrate that exact and scalable causal discovery from high-dimensional perturbation data is achievable through principled search space design.\n\n---\n\n## 173. LEAP: Trajectory-Level Evaluation of LLMs in Iterative Scientific Design\n- **Authors**: Marilyn Zhang , Tianfeng Chen , Fabián Barzuna , Ankita Rathod , Mark E. Whiting\n- **URL**: [https://arxiv.org/abs/2605.15341](https://arxiv.org/abs/2605.15341)\n- **Abstract**:\n> LLMs are increasingly deployed in autonomous laboratories, under the assumption that their domain priors and reasoning over iterative feedback let them converge on good designs in fewer iterations than feedback-only baselines. Current iterative scientific design benchmarks, however, score only outcome snapshots at fixed horizons. This leaves the learning trajectory unmeasured, even though the trajectory is what captures learning efficiency, where each iteration saved is a real saving in cost and time. Motivated by this, we examine three evaluation choices that change the conclusions one draws about LLM learning efficiency in iterative scientific design: what to measure, what baseline to compare against, and what to ground against. We introduce LEAPBench, Learning Efficiency in Adaptive Processes, a 55-task framework that pairs a best-so-far area under the curve (AUC) trajectory metric with a classical Bayesian-optimization reference and an audit grounded in published literature. Applied to eight contemporary LLMs, switching from final-outcome to trajectory scoring changes the best-model decision on 53% of tasks at matched horizons, and exposes efficiency gains overlooked by outcome-based scoring. LLMs do not outperform a classical Bayesian baseline. On 16 biology tasks where the oracle's reward signal is aligned with configurations from the published-best design, domain-aware prompting leads to LLM choices that match the published-best's approximately 10 percentage points less often than domain-agnostic prompting at iteration 30. The pattern is sharpest on 6 tasks where the literature-typical and published-best configurations diverge, and domain-agnostic prompting matches the published-best more often on all 6. The trajectory metric also doubles as a tractable training target. Offline reinforcement learning with the metric as a reward improves performance on 14 of 21 held-out tasks.\n\n---\n\n## 174. Hidden in Memory: Sleeper Memory Poisoning in LLM Agents\n- **Authors**: Sidharth Pulipaka , Stanislau Hlebik , Leonidas Raghav , Sahar Abdelnabi , Vyas Raina , Ivaxi Sheth , Mario Fritz\n- **URL**: [https://arxiv.org/abs/2605.15338](https://arxiv.org/abs/2605.15338)\n- **Abstract**:\n> Large language models are increasingly augmented with persistent memory, allowing assistants to store user-specific information across sessions for personalization and continuity. This statefulness introduces a new security risk: adversarial content can corrupt what an assistant remembers and thereby influence future interactions. We propose and study sleeper memory poisoning, a delayed attack in which an adversary manipulates external context, such as a document, webpage, or repository, to cause the assistant to store a fabricated memory about the user. Unlike conventional prompt injection, the attack can remain dormant and re-emerge across multiple later conversations. We evaluate the full attack pipeline: whether poisoned memories are written, later retrieved, and ultimately used to steer the following conversations. Across stateful LLM assistants, poisoned memories were added up to 99.8% on GPT-5.5 and 95% on Kimi-K2.6. Crucially, among successful retrievals, poisoned memories cause attacker-intended agentic actions in 60-89% of evaluations across models. These results show that persistent memory can act as a long-term attack surface across multiple future conversations.\n\n---\n\n## 175. HoloMotion-1 Technical Report\n- **Authors**: Maiyue Chen , Kaihui Wang , Bo Zhang , Xihan Ma , Zhiyuan Yang , Yi Ren , Qijun Huang , Zihao Zhu , Yucheng Wang , Zhizhong Su\n- **URL**: [https://arxiv.org/abs/2605.15336](https://arxiv.org/abs/2605.15336)\n- **Abstract**:\n> In this report, we present HoloMotion-1, a humanoid motion foundation model for zero-shot whole-body motion tracking. A key innovation of HoloMotion-1 is to scale control-policy training with a large-scale hybrid motion corpus, where video-reconstructed motions from in-the-wild videos provide the dominant source of motion diversity, while curated motion-capture and in-house motion data provide higher-fidelity supervision and deployment-oriented coverage. This data regime enables HoloMotion-1 to move beyond conventional MoCap-only training and exposes the policy to substantially broader behaviors, capture conditions, and motion styles. Learning from such heterogeneous data introduces new challenges, including reconstruction noise, source-domain mismatch, uneven motion quality, and the need for temporal modeling under large behavioral variation. To address these challenges, HoloMotion-1 integrates large-capacity temporal modeling, a sparsely activated Mixture-of-Experts Transformer with KV-cache inference for real-time control, and a sequence-level training strategy that improves learning efficiency on extended motion sequences. Extensive experiments on multiple unseen motion benchmarks show that HoloMotion-1 generalizes robustly across diverse motion types and capture conditions, significantly improves tracking accuracy over prior methods, and transfers directly to a real humanoid robot without task-specific fine-tuning.\n\n---\n\n## 176. From I/O to Code with Discovery Agent\n- **Authors**: Yihong Dong , Jiaru Qian , Haoran Zhang , Peixu Wang , Binhua Li , Zhi Jin , Yongbin Li , Ge Li , Xiaokang Yang , Xue Jiang\n- **URL**: [https://arxiv.org/abs/2605.15334](https://arxiv.org/abs/2605.15334)\n- **Abstract**:\n> The automatic synthesis of a program from any form of specification is regarded as a holy grail of computer science. Fueled by LLMs, NL2Code has achieved tremendous success, yet the fundamentally more challenging task of synthesizing programs from input-output behavior, which we refer to as IO2Code, remains largely unsolved. Whereas NL2Code can exploit the semantic alignment between natural language and code acquired during pretraining, IO2Code requires recovering underlying principles from concrete computational behavior, navigating a vast and underspecified hypothesis space. To address this, we propose DIO-Agent, a discovery agent for IO2Code. Our method frames IO2Code as an evolutionary search over discrete program space, in which an LLM serves as the mutation operator and concrete error signals from execution guide each mutation. To prevent the search from wandering into structurally complex yet incorrect dead ends, we introduce the Transformation Priority Premise as a mutation prior that biases the LLM toward the simplest hypothesis consistent with current evidence, progressively escalating from constants to conditionals to iteration only when simpler constructs are insufficient. To facilitate systematic study, we further construct an IO2CodeBench spanning multiple difficulty levels. Extensive experiments show that DIO-Agent consistently outperforms both traditional program-by-example method and SOTA evolution-agent baselines across all difficulty levels and various LLMs, while substantially surpassing test-time scaling strategies with equivalent sampling budgets.\n\n---\n\n## 177. Fortress: A Case Study in Stabilizing Search Recommendations via Temporal Data Augmentation and Feature Pruning\n- **Authors**: Milind Pandurang Jagre , Jia Huang , Dayvid V. R. Oliveira , Zhinan Cheng , Babak Seyed Aghazadeh , Puja Das , Chris Alvino , Jinda Han , Kailash Thiyagarajan\n- **URL**: [https://arxiv.org/abs/2605.15299](https://arxiv.org/abs/2605.15299)\n- **Abstract**:\n> In search and recommendation systems, predictive models often suffer from temporal instability when certain input features introduce volatility in output scores. This instability can degrade model reliability and user experience especially in multi-stage systems where consistent predictions are critical for downstream decision making. We introduce Fortress, a general framework for enhancing model stability and accuracy by identifying and pruning features that contribute to inconsistent prediction scores over time. Fortress leverages historical snapshots temporally partitioned datasets capturing score fluctuations for the same entity across periods and follows a four-step process: (1) collect historical snapshots, (2) identify samples with unstable predictions, (3) isolate and remove instability-inducing features, and (4) retrain models using only stable features. While semantic features from LLMs and BERT-based models improve generalization, they often lack full query or entity coverage. Engagement-based features offer strong predictive power but tend to introduce temporal instability. Fortress mitigates this trade-off by suppressing the volatility of engagement signals while retaining their predictive value leading to more stable and accurate models. We validate Fortress on a query-to-app relevance model in a large-scale app marketplace. Offline experiments demonstrate notable improvements in prediction stability (measured by Coefficient of Variation) and classification performance (measured by PR-AUC).\n\n---\n\n## 178. PhysBrain 1.0 Technical Report\n- **Authors**: Shijie Lian , Bin Yu , Xiaopeng Lin , Changti Wu , Hang Yuan , Xiaolin Hu , Zhaolong Shen , Yuzhuo Miao , Haishan Liu , Yuxuan Tian , Yukun Shi , Cong Huang , Kai Chen\n- **URL**: [https://arxiv.org/abs/2605.15298](https://arxiv.org/abs/2605.15298)\n- **Abstract**:\n> Vision-language-action models have advanced rapidly, but robot trajectories alone provide limited coverage for learning broad physical understanding. PhysBrain 1.0 studies a complementary route: converting large-scale human egocentric video into structured physical commonsense supervision before robot adaptation. Our data engine extracts scene elements, spatial dynamics, action execution, and depth-aware relations, then turns them into question-answer supervision for training PhysBrain VLMs. The resulting physical priors are further transferred to VLA policies through a capability-preserving and language-sensitive adaptation design. Across multimodal QA benchmarks and embodied control benchmarks, including ERQA, PhysBench, SimplerEnv-WidowX, LIBERO, and RoboCasa, PhysBrain 1.0 achieves SOTA results and shows especially strong out-of-domain performance on SimplerEnv. These results suggest that scaling physical commonsense from human interaction video can provide an effective bridge from multimodal understanding to robot action.\n\n---\n\n## 179. GESD: Beyond Outcome-Oriented Fairness\n- **Authors**: Gideon Popoola , John Sheppard\n- **URL**: [https://arxiv.org/abs/2605.15295](https://arxiv.org/abs/2605.15295)\n- **Abstract**:\n> Machine learning (ML) algorithms are increasingly deployed in high-stakes decision-making domains such as loan approvals, hiring, and recidivism predictions. While existing fairness metrics (e.g., statistical parity, equal opportunity) effectively quantify outcome-oriented disparities, they offer limited insight into the procedure or explanation behind biased decisions. To address this gap, we propose Group-level Explanation Stability Disparity (GESD), a \\textit{procedural-oriented} fairness metric that measures disparities in the stability, robustness, and sensitivity of model explanations across different subgroups in a protected category. %GESD is explainer-agnostic, model-agnostic, and extends the scope of fairness analyses to the level of explainability. We further integrate GESD into a multi-objective optimization framework that jointly optimizes for utility, outcome-based fairness, and explanation-based fairness called FEU (Fairness--Explainability--Utility). Empirical results on multiple benchmark datasets show that GESD effectively captures group-wise discrepancies in explanation quality, and that FEU improves both utility and fairness over state-of-the-art methods. By bridging outcome-based and explanation-based fairness, GESD offers a comprehensive tool for diagnosing and mitigating bias in predictive modeling. Our code and datasets are available on GitHub {\\hyperlink{ this https URL }{ this https URL } }\n\n---\n\n## 180. GQA-μP: The maximal parameterization update for grouped query attention\n- **Authors**: Kyle R. Chickering , Huijuan Wang , Mengxi Wu , Alexander Moreno , Muhao Chen , Xuezhe Ma , Daria Soboleva , Joel Hestness , Zhengzhong Liu , Eric Xing\n- **URL**: [https://arxiv.org/abs/2605.15290](https://arxiv.org/abs/2605.15290)\n- **Abstract**:\n> Hyperparameter transfer across model architectures dramatically reduces the amount of compute necessary for tuning large language models (LLMs). The maximal update parameterization ({\\mu}P) ensures transfer through principled mathematical analysis but can be challenging to derive for new model architectures. Building on the spectral feature-learning view of Yang et al. (2023a), we make two advances. First, we promote spectral norm conditions on the weights from a heuristic to the definition of feature learning, and as a consequence arrive at the Complete-P depth and weight-decay scalings without recourse to lazy-learning. Second, we consider a modified spectral norm that preserves the valid scaling law of network weights when weight matrices are not full rank. This enables (to our knowledge, the first) derivation of {\\mu}P scalings for grouped-query attention (GQA). We demonstrate the efficacy of our theoretical derivations by showing learning rate transfer across the GQA repetition hyperparameter as well as experiments regarding transfer over weight decay.\n\n---\n\n## 181. Universal Approximation of Nonlinear Operators and Their Derivatives\n- **Authors**: Filippo de Feo\n- **URL**: [https://arxiv.org/abs/2605.15285](https://arxiv.org/abs/2605.15285)\n- **Abstract**:\n> Derivative-Informed Operator Learning (DIOL), i.e. learning a (nonlinear) operator and its derivatives, is an open research frontier at the foundations of the influential field of Operator Learning (OL). In particular, Universal Approximation Theorems (UATs) of nonlinear operators and their derivatives are foundational open questions and delicate problems in nonlinear functional analysis. In this manuscript, we prove the first UATs of non-linear $k$-times differentiable operators between Banach spaces and their derivatives, uniformly on compact sets and in weighted Sobolev norms for general finite input measures, via OL architectures. Our results are the first complete generalizations of the corresponding influential classical results in [Hornik, 1991] to infinite-dimensional settings and OL. We discuss several open areas where DIOL and our UATs find applications: high-order accuracy in OL, fast constrained optimization in Banach spaces (e.g. optimal control of PDEs, inverse problems) and numerical methods for infinite-dimensional PDEs (e.g. HJB PDEs on Banach spaces from optimal control of PDEs, SPDEs, path-dependent systems, partially observed systems, mean-field control). We parameterize nonlinear operators via Encoder-Decoder Architectures, renowned classes in OL due to their generality, including classical architectures, such as DeepONets, Deep-H-ONets, PCA-Nets. Our results are based on four key features that allow us to prove UATs in full generality: (i) Approximation Properties of Banach spaces. (ii) $k$-times continuous differentiability in the sense of Bastiani (weaker than $k$-times continuous Fréchet differentiability). (iii) Natural compact-open topologies for UA; indeed, we show that UA in standard compact-open topologies induced by operator norms is violated even for Fréchet derivatives. (iv) Construction of novel weighted Sobolev spaces for the UA.\n\n---\n\n## 182. Autonomous Intelligent Agents for Natural-Language-Driven Web Execution with Integrated Security Assurance\n- **Authors**: Vinil Pasupuleti , Siva Rama Krishna Varma Bayyavarapu , Shrey Tyagi\n- **URL**: [https://arxiv.org/abs/2605.15281](https://arxiv.org/abs/2605.15281)\n- **Abstract**:\n> Modern web test suites rot. A UI refactor breaks locators, a timing change causes race conditions, and within weeks developers abandon the suite entirely. This paper presents an AI-driven autonomous testing framework that addresses these failure modes through five integrated strategies - navigation reliability, context-aware selector generation, post-generation validation, smart wait injection, and failure learning - implemented over a containerised worker architecture that decouples orchestration from long-running browser execution. Evaluated across four production applications and 176 scenarios, the framework improves script generation success from 55% to 93%, achieves an 8x reduction in navigation failures, eliminates 80% of timing-related race conditions, and reduces test creation time by 75% compared to manual Selenium authoring. The framework extends naturally to security validation: testers describe attack scenarios in plain English - \"try accessing another user's invoice\" - which the agent converts to OWASP Top 10-aligned browser probes, detecting 85% of authentication bypass vulnerabilities and 95% of input validation flaws with false positive rates below 12%. Natural-language-driven security testing of this kind represents, to our knowledge, a novel contribution to the field.\n\n---\n\n## 183. PDRNN: Modular Data-driven Pedestrian Dead Reckoning on Loosely Coupled Radio- and Inertial-Signalstreams\n- **Authors**: Peter Bauer , Andreas Porada , Felix Ott , Christopher Mutschler , Tobias Feigl\n- **URL**: [https://arxiv.org/abs/2605.15252](https://arxiv.org/abs/2605.15252)\n- **Abstract**:\n> Modern pedestrian dead reckoning (PDR) systems rely on fusing noisy and biased estimates of position, velocity, and calibrated orientation derived from loosely coupled sensors to determine the current pose of a localized object. However, discrepancies in the sampling rates of sensor-specific estimation methods and unreliable transmission pose significant challenges. And traditional methods often fail to effectively fuse multimodal sensor data during dynamic movements characterized by high accelerations, velocities, and rapidly varying orientations. To address these limitations, we propose a simple recurrent neural network (RNN) architecture capable of implicitly forecasting asynchronous sensor data streams from diverse estimation methods along reference trajectories. The proposed approach introduces PDRNN, a modular hybrid AI-assisted PDR system that handles each component as an independent ensemble of machine learning (ML) models to estimate both key parameter means and variances. Separate ML-based models are employed to estimate orientation, (un)directed velocity or distance from acceleration and gyroscope data, with optional absolute positioning from synchronized radio systems such as 5G for stabilization. A final fusion model combines these outputs, position, velocity, and orientation, while using uncertainty estimates to enhance system robustness. The modular design allows individual components to be updated, fine-tuned, or replaced without affecting the entire system. Experiments on dynamic sports movement data show that PDRNN achieves superior accuracy and precision compared to classic and ML-based methods, effectively avoiding error accumulation common in black-box approaches. And PDRNN offers forecast capabilities and better component control despite increased system complexity.\n\n---\n\n## 184. GQLA: Group-Query Latent Attention for Hardware-Adaptive Large Language Model Decoding\n- **Authors**: Fanxu Meng\n- **URL**: [https://arxiv.org/abs/2605.15250](https://arxiv.org/abs/2605.15250)\n- **Abstract**:\n> Multi-head Latent Attention (MLA), the attention used in DeepSeek-V2/V3, jointly compresses keys and values into a low-rank latent and matches the H100 roofline almost perfectly. Its trained weights, however, expose only one decoding path - an absorbed MQA form - which ties efficient inference to H100-class compute-bandwidth ratios, forfeits tensor parallelism along the head axis, and yields no Multi-Token Prediction (MTP) gain on commodity inference GPUs such as the export-restricted H20. We propose Group-Query Latent Attention (GQLA), a minimal modification of MLA whose trained weights expose two algebraically equivalent decoding paths over the same parameters: an MQA-absorb path identical to MLA's, and a GQA path with a per-group expanded cache. The runtime picks the path that matches the target hardware - no retraining, no custom kernels - so a single set of GQLA weights pins the rooflines of both H100 (MQA-absorb, s_q=1) and H20 (GQA + MTP, s_q=2), while supporting up to 8-way zero-redundancy tensor parallelism on the GQA path. To avoid pretraining from scratch we extend TransMLA into TransGQLA, which converts a pretrained GQA checkpoint into a GQLA model; on LLaMA-3-8B it compresses the per-token KV cache to 28.125% of the GQA baseline on the MQA-absorb path while structurally preserving GQA-level traffic on the per-group path.\n\n---\n\n## 185. Reading the Cell, Designing the Cure: Perturbation-Conditioned Molecular Diffusion for Function-Oriented Drug Design\n- **Authors**: Ziyu Xu , Zijian Zhang , Liang Wang , Zhiyuan Liu , Qiang Liu , Shu Wu , Liang Wang\n- **URL**: [https://arxiv.org/abs/2605.15243](https://arxiv.org/abs/2605.15243)\n- **Abstract**:\n> When reliable target structures are unavailable at scale or phenotypes arise from dysregulated pathways, transcriptomic perturbations provide a system-level functional readout for drug action. In this work, we formalize \\emph{Transcriptome-based Drug Design (TBDD)} as a generative inverse problem: designing drug molecules conditioned on desired transcriptomic state transitions. We analyze the inherently ill-posed nature of this task, which is further complicated by the profound domain gap between biology and chemistry and by the sparsity of transcriptomic signals. To address these challenges, we propose \\textbf{\\themodel{} } (A \\textbf{C}ell\\textbf{U}lar \\textbf{R}esponse \\textbf{E}ngine), a multi-resolution transcriptome-guided diffusion framework. \\themodel{} features a specialized \\textbf{Transcriptome Perturbation Functional Feature Extractor (TFE)} that (1) distills function-oriented perturbation embeddings from pre/post states, (2) aligns these signatures to dual chemical views to bridge the cross-modal gap, and (3) performs heterogeneity-aware aggregation to extract robust state-specific signals from noisy transcriptomic data. Extensive evaluations on both standard benchmarks and rigorous out-of-distribution protocols demonstrate that \\themodel{} consistently outperforms strong baselines in structural quality and functional consistency. Furthermore, we validate its practical utility via a zero-shot gene-inhibitor design task, highlighting the potential of phenotype-driven generative discovery.\n\n---\n\n## 186. Hydra: Efficient, Correct Code Generation via Checkpoint-and-Rollback Support\n- **Authors**: Alexander Du , Jianjun Ou , Danyang Zhuo , Matthew Lentz\n- **URL**: [https://arxiv.org/abs/2605.15238](https://arxiv.org/abs/2605.15238)\n- **Abstract**:\n> Large language models are increasingly used for code generation, but many generated programs fail to compile, a prerequisite for further correctness checks such as unit tests. Existing solutions for repairing static errors are costly in both latency and token consumption. Post-hoc repair delays error detection until generation completes and commonly regenerates large regions of previously valid code. Constrained semantic decoding checks after each token, incurring per-token overhead while limiting repair to the current token even when the root cause lies earlier. We present Hydra, a system for efficient recovery from static errors during code generation. Hydra allows checking to proceed asynchronously with generation, avoiding checker overhead when the generated code is semantically correct. In addition, it provides checkpoint-and-rollback support for targeted repair, avoiding regeneration and rechecking of valid prefixes. We retrofit the Clang C/C++ compiler to support Hydra with modest modifications. Paired with a token-efficient repair strategy, Hydra reduces latency by up to 71% and token consumption by up to 70% relative to post-hoc repair on C/C++ code generation tasks that encounter static errors.\n\n---\n\n## 187. A3D: Agentic AI flow for autonomous Accelerator Design\n- **Authors**: Abinand Nallathambi , Christopher Knight , Shantanu Ganguly , Wilfried Haensch , Anand Raghunathan\n- **URL**: [https://arxiv.org/abs/2605.15237](https://arxiv.org/abs/2605.15237)\n- **Abstract**:\n> Accelerating applications through the design of hardware accelerators can significantly enhance system performance and energy efficiency. Despite advances, such as high-level synthesis (HLS), designing accelerators for complex applications still remains highly labor-intensive, demanding considerable expertise in understanding workloads to be accelerated, hardware design, micro-architecture, and EDA tool usage, posing challenges for application domain experts. Therefore, most accelerator solutions are limited to applications with a regular predictable dataflow. Advances in AI have enabled agents that perform autonomous planning, reasoning, execution and reflection, leading to unprecedented potential for automation through agentic AI. We present A3D, an Agentic AI flow for end-to-end Automation of hardware Accelerator Design. A3D automates workload analysis, performance bottleneck identification, code refactoring for HLS compatibility and micro-architecture generation. A3D also generates diverse accelerator designs by automatically exploring the speed-area tradeoff space. Recent efforts have explored the use of AI for specific tasks such as design space exploration in HLS, leaving several tasks to still be performed manually. A3D addresses the challenges in applying modern LLMs to accelerator design by judiciously partitioning tasks among specialist agents, orchestrating process loops with specialist and verifier agents, utilizing pre-existing and custom tools, and employing agentic RAG for codebase and proprietary EDA tool documentation exploration. Our implementation of A3D, using commercial components like Claude Sonnet 4.5 and the Catapult HLS tool, demonstrates its effectiveness by generating accelerator designs with no human intervention from complex scientific applications like LAMMPS (molecular dynamics simulation) and QMCPACK (quantum chemistry).\n\n---\n\n## 188. Learning Selective Merge Policies for Deadline-Constrained Coded Caching via Deep Reinforcement Learning\n- **Authors**: Amirhossein Yousefiramandi\n- **URL**: [https://arxiv.org/abs/2605.15236](https://arxiv.org/abs/2605.15236)\n- **Abstract**:\n> With the coded caching, the server can use the information the users have cached to serve multiple users at a time by sending a single coded multi-casting message, i.e., the merged message, thereby relieving the peak network loads. However, for the delay-sensitive applications of the users, like the video streaming services, it becomes essential to choose which messages to merge online, considering the strict deadlines for each request. The problem, however, is that while the merge is helpful for the formation of the current coded multi-casting message, it can be harmful for the subsequent ones. We proposed a DRL-based solution that formulates the deadline-constrained coded delivery as a masked discrete-action queue-state control problem, while we trained a graph-attention policy network via proximal policy optimization. The policy network reduces the broadcast-packet expiration ratio $\\rho$ by $40.9%$ ($0.208$ vs. $0.352$) with respect to the best coded multi-casting baseline (SACM++) on the uniform-demand benchmark, while also attaining the best broadcast-efficiency score $\\sigma$ across the Track A battery among the coded multi-casting methods. The interesting fact we observed is that for the applications of the users with tight deadlines, the method of selective merging is better than the method of aggressive merging, i.e., the policy network learns to merge at only $\\approx 31.8%$ rate, even though the same observation holds across the variations within the same simulator family.\n\n---\n\n## 189. PBT-Bench: Benchmarking AI Agents on Property-Based Testing\n- **Authors**: Lucas Jing , Xinqi Wang , Liao Zhang , Simon S. Du\n- **URL**: [https://arxiv.org/abs/2605.15229](https://arxiv.org/abs/2605.15229)\n- **Abstract**:\n> Existing code benchmarks measure whether an agent can produce any test that reproduces a known bug, or whether it can produce a patch that fixes a described issue. Neither isolates the distinct skill of property-based testing: deriving a semantic invariant from documentation, and then constructing an input-generation strategy precise enough to make a random search reveal the violation. We introduce PBT-Bench, a benchmark of 100 curated property-based testing problems across 40 real Python libraries. Each problem injects one or more semantic bugs (365 in total, mean 3.65 per problem) designed so that default-strategy random inputs almost never trigger them; the agent must read the library's documentation, identify the relevant invariant, and specify a Hypothesis @given strategy that concentrates mass in the trigger region. Bugs are stratified across three difficulty levels (L1-L3) spanning single-constraint boundary bugs to stateful, cross-function protocol violations. We evaluate eight contemporary LLMs under two prompting regimes (open-ended baseline vs. explicit Hypothesis scaffolding) for three independent runs per configuration. Bug recall under the PBT-guided prompt ranges from 42.1% to 83.4% across models; under the open-ended baseline, from 31.4% to 76.7%. Hypothesis scaffolding lifts mid-capability models by over 20 percentage points, but yields smaller gains for the strongest models, with two exceptions showing degradation, suggesting the structured prompt can interfere with certain model behaviours rather than complementing them. The hardest bugs prove model-specific: different architectures fail on different problems, leaving persistent gaps that no single model closes. We release the benchmark, harness, and full evaluation corpus to support downstream work on documentation-grounded semantic reasoning.\n\n---\n\n## 190. Is Agentic AI Ready for Real-World Hardware Engineering? A Deep Dive with Phoenix-bench\n- **Authors**: Qingyun Zou , Feng Yu , Hongshi Tan , Bingsheng He , WengFai Wong\n- **URL**: [https://arxiv.org/abs/2605.15226](https://arxiv.org/abs/2605.15226)\n- **Abstract**:\n> We ask whether agentic AI systems built for software engineering transfer to realistic hardware engineering. Existing hardware LLM benchmarks isolate sub-tasks but none jointly requires repository navigation, hierarchy-aware localization, Electronic Design Automation (EDA) executable verification, and maintenance-style patching. We introduce \\textbf{Phoenix-bench}, a synchronized corpus of 511 verified Verilator instances from 114 GitHub repositories, each shipped with the developer patch, design-flow labels, fail-to-pass and pass-to-pass testbenches, and a Docker-pinned EDA environment so resolved-rate differences reflect agent behavior rather than toolchain availability. Using Phoenix-bench we run a uniform evaluation of four commercial agents and eight open-source agentic structures across four LLM backbones, plus two diagnostic interventions (file-level oracle localization and one round of testbench-log feedback). Three findings emerge. (i)~Software and hardware are fundamentally different engineering tasks: the same agent loses 37\\% to 58\\% from SWE-bench Verified to Phoenix-bench because hardware bugs propagate across parallel instantiated modules through signal flow rather than along a software-style call graph, and software-tuned agents stop at the symptom file instead of tracing back through the instantiation chain. (ii)~Failures concentrate on design control-flow / finite state machine (FSM) bugs, verification testbench bugs, and hard cases that demand cross-hierarchy signal-flow tracking and coordinated multi-file edits. (iii)~Localization granularity matters far more than localization itself: a perfect file-level oracle yields only $+1.4$\\% because the agent then breaks files that did not need editing, while a single round of test case feedback lifts resolved rate by $42$\\% to $45$\\% because the test case tells \\emph{where} the bug is and \\emph{what} the fix has to look like.\n\n---\n\n## 191. Do Biological Structural Guarantees Earn Their Complexity?\n- **Authors**: Bogdan Banu\n- **URL**: [https://arxiv.org/abs/2605.15225](https://arxiv.org/abs/2605.15225)\n- **Abstract**:\n> Biologically-inspired AI agent frameworks claim reliability benefits through structural guarantees adapted from gene regulatory networks, immune systems, and metabolic control. These claims are rarely tested empirically against simpler alternatives. We present three deep benchmarks: metabolic priority gating, autoinducer-based quorum sensing, and Bayesian stagnation detection, each comparing a biologically-grounded implementation against a naive non-biological alternative and an ablated control, across 1,000 trials per seed and 10 seeds (10M+ data points total).\n\n---\n\n## 192. GenAI-Driven Approach to RISC-V Supply Chain Exploration\n- **Authors**: Nenad Petrovic , Andre Schamschurko , Yingjie Xu , Alois Knoll\n- **URL**: [https://arxiv.org/abs/2605.15223](https://arxiv.org/abs/2605.15223)\n- **Abstract**:\n> This paper presents an LLM-empowered workflow for RISC-V supply chain analysis, integrating Vision-Language Models (VLMs) and Model-Driven Engineering (MDE) to enable comprehensive, multimodal data-driven insights. The proposed approach addresses the challenges of heterogeneous and unstructured supply chain data by leveraging LLMs for textual understanding and VLMs for extracting information from visual artifacts such as diagrams, tables, and scanned documents. These models collaboratively identify key entities and relationships, which are then organized into a knowledge graph representing supply chain components and their interdependencies. For analytical reasoning, the workflow incorporates MDE techniques and constraint-based modeling to enable formal validation of dependencies, detection of bottlenecks, and assessment of risks. The synergy between LLM- and VLM-based semantic understanding and MDE-based formal analysis supports both exploratory and systematic evaluation of supply chain resilience. A human-in-the-loop mechanism further enables interactive querying and expert validation. The approach is evaluated in RISC-V ecosystem scenarios, demonstrating its effectiveness in generating actionable insights, enhancing transparency, and supporting decision-making in complex semiconductor supply chains.\n\n---\n\n## 193. Effective Harness Engineering for Algorithm Discovery with Coding Agents\n- **Authors**: Yoichi Ishibashi , Taro Yano , Masafumi Oyamada\n- **URL**: [https://arxiv.org/abs/2605.15221](https://arxiv.org/abs/2605.15221)\n- **Abstract**:\n> AlphaEvolve and FunSearch have demonstrated the potential of combining large language models (LLMs) with evolutionary search for automated algorithm discovery. However, discovery success is shaped not only by model capability but also significantly by the design of the execution infrastructure, i.e., the harness. This paper investigates effective harness design through three questions: under a fixed token budget, is it better to produce many algorithms with brief thought or fewer algorithms with deeper thought? How should the harness handle evaluation hacks, where generated programs exploit the scoring function? And how can agents that require full filesystem access execute safely in parallel? Using Vesper, an algorithm discovery framework that incorporates harness improvements addressing these questions, we evaluate on Circle Packing under the same token budget. Interestingly, generating fewer algorithms while thinking more deeply about each one achieved higher scores. That is, scaling the quality of each individual is more budget-efficient than scaling the number of evolutionary generations. Surprisingly, more capable models produced evaluation hacks at higher rates, making hack detection increasingly necessary as models scale.\n\n---\n\n## 194. Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time\n- **Authors**: Michael Y. Hu , Apurva Gandhi , Kyunghyun Cho , Tal Linzen , Pratyusha Sharma\n- **URL**: [https://arxiv.org/abs/2605.15220](https://arxiv.org/abs/2605.15220)\n- **Abstract**:\n> Data mixing decides how to combine different sources or types of data and is a consequential problem throughout language model training. In pretraining, data composition is a key determinant of model quality; in continual learning and adaptation, it governs what is retained and acquired. Yet existing data mixing methods address only one phase of this lifecycle at a time: some require smaller proxy models tied to a single training phase, others assume a fixed domain set, and continual learning lacks principled guidance altogether. We argue that data mixing is fundamentally an online decision making problem -- one that recurs throughout training and demands a single, unified solution. We introduce OP-Mix (On-Policy Mix), a data mixing algorithm that operates across the entire language model training lifecycle. Our main insight is that candidate data mixtures can be cheaply simulated by interpolating between low-rank adapters trained directly on the current model, eliminating separate proxy models and ensuring the search is always grounded in the model's actual learning dynamics. Across pretraining, continual midtraining, and continual instruction tuning, OP-Mix consistently finds near-optimal mixtures while using a fraction of the compute of the baselines. In pretraining, OP-Mix improves upon training without mixing by 6.3% in average perplexity. For continual learning, OP-Mix matches the performance of both retraining and on-policy distillation while using 66% and 95% less overall compute, respectively. OP-Mix suggests a different view of language model training: not a sequence of distinct phases, but a single continuous process of learning from data.\n\n---\n\n## 195. An LLM-RAG Approach for Healthy Eating Index-Informed Personalized Food Recommendations\n- **Authors**: Yibin Wang , Yanjie Yang , Grace Melo Guerrero , Rodolfo M. Nayga Jr. , Azlan Zahid\n- **URL**: [https://arxiv.org/abs/2605.15213](https://arxiv.org/abs/2605.15213)\n- **Abstract**:\n> Diet quality is a leading determinant of chronic disease risk. Advances in artificial intelligence (AI) have enabled food recommendation systems to adapt suggestions to user preferences and health goals. However, most current systems rely on loosely curated food databases and provide limited connection to a validated index. In this study, we propose a Healthy Eating Index (HEI) informed retrieval-augmented generation (RAG) framework that combines standardized nutrition databases with large language models (LLMs) for personalized food recommendations. Our proposed method anchors retrieval in the National Health and Nutrition Examination Survey (NHANES) and the Food Patterns Equivalents Database (FPED). A food-level embedding space is constructed from FPED-derived textual descriptions. For each entity, the system computes baseline HEI scores, retrieves candidate foods for intake recommendations, and estimates the HEI impact of simple substitutions or additions. A constrained RAG pipeline instantiated with a pretrained OpenAI LLM generates personalized recommendations and sources based on nutrient profiles and HEI contributions. The simulation results showed a mean HEI improvement of 6.45, with the proportion of users HEI over 50 increasing from 45.12 to 61.26. Quantile analysis revealed consistent improved shifts across the HEI distribution. Our findings suggest that the proposed LLM-RAG-based AI systems can support more precise, explainable, and personalized nutrition guidance to improve diet quality.\n\n---\n\n## 196. Fault tolerance estimation in digital circuits with visualised generative networks\n- **Authors**: Sascha Biel , Carl Alexander Gaede , Amiel Glaser , Jan Wolter , Alexej Schelle\n- **URL**: [https://arxiv.org/abs/2605.15212](https://arxiv.org/abs/2605.15212)\n- **Abstract**:\n> We propose a new numerical method to estimate the fault tolerance of failure modes in digital circuit structures with a generative network sampling technique. From a random input of generated bitwise configurations of ideally digitalised analog currents in the digital circuit design with classical logical gates, expected output currents are compared to the realistic signals of a numerical experiment at the discriminator part of the Generative Adversarial Network (GAN) to calculate the deviation from ideal digital electronic signals, including various error modes, such as missing or interchanged logical devices. From the present analysis of a representation of the GAN in terms of complex variables, it is possible to evaluate the robustness in electronic designs by differentiating the impact of failure modes associated with different classical logical elements in the circuit.\n\n---\n\n## 197. Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels\n- **Authors**: Plawan Kumar Rath , Rahul Maliakkal\n- **URL**: [https://arxiv.org/abs/2605.15208](https://arxiv.org/abs/2605.15208)\n- **Abstract**:\n> Large Language Models are routinely compressed via post-training quantization to reduce inference costs and memory footprint for cloud and edge deployment, yet the impact of this compression on model quality remains poorly understood. Existing studies typically compare only two conditions (full-precision vs. a single quantized variant), rely on aggregate bias metrics, and evaluate a single model family, making it impossible to distinguish gradual degradation from threshold-dependent safety failures. We conduct a controlled empirical study of three instruction-tuned models (Qwen2.5-7B, Mistral-7B, Phi-3.5-mini) at five precision levels (BF16 through 3-bit) on 12,148 BBQ bias benchmark items across 5 random seeds, totaling 911,100 inference records. Our results reveal that 3-bit quantization causes 6-21% of previously unbiased items to develop new stereotypical behaviors, following a clear dose-response pattern confirmed via logistic regression, while models' willingness to select \"unknown\" answers declines by 17.4%. Crucially, these item-level changes are invisible to standard quality metrics: perplexity increases by less than 0.5% at 8-bit and under 3% at 4-bit across all three models, yet 2.5-5.6% of items already develop new biases at 4-bit. These findings demonstrate that aggregate evaluation metrics systematically miss fairness-critical degradation, underscoring the need for quality-aware compression protocols that explicitly test for bias emergence before deployment.\n\n---\n\n## 198. AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices\n- **Authors**: Dzung Pham , Kleomenis Katevas , Ali Shahin Shamsabadi , Hamed Haddadi\n- **URL**: [https://arxiv.org/abs/2605.15206](https://arxiv.org/abs/2605.15206)\n- **Abstract**:\n> Autonomous agents powered by large language models (LLMs) are increasingly used to automate complex, multi-step tasks such as coding or web-based question answering. While remote, cloud-based agents offer scalability and ease of deployment, they raise privacy concerns, depend on network connectivity, and incur recurring API costs. Deploying agents locally on user devices mitigates these issues by preserving data privacy and eliminating usage-based fees. However, agentic workflows are far more resource-intensive than typical LLM interactions. Iterative reasoning, tool use, and failure retries substantially increase token consumption, often expending significant compute without successfully completing tasks. In this work, we investigate the time, token, and energy overhead of locally deployed LLM-based agents on consumer hardware. Our measurements show that agentic execution increases GPU power draw, temperature, and battery drain compared to single-inference workloads. To address this inefficiency, we introduce AgentStop, a lightweight efficiency supervisor that predicts and preemptively terminates trajectories unlikely to succeed. Leveraging low-cost execution signals, such as token-level log probabilities, AgentStop can reduce wasted energy by 15-20% with minimal impact on task performance (<5% utility drop) for challenging web-based question answering and coding benchmarks. These findings position predictive early termination as a practical mechanism for enabling sustainable, privacy-preserving LLM agents on user devices. Our project code and data are available at this https URL .\n\n---\n\n## 199. Agent4POI: Agentic Context-Conditioned Affordance Reasoning for Multimodal Point-of-Interest Recommendation\n- **Authors**: Jinze Wang , Yangchen Zeng , Tiehua Zhang , Lu Zhang , Yuze Liu , Yongchao Liu , Xingjun Ma , Zhu Sun\n- **URL**: [https://arxiv.org/abs/2605.15203](https://arxiv.org/abs/2605.15203)\n- **Abstract**:\n> We introduce Agent4POI, the first POI recommendation framework that generates context-conditioned multimodal representations at recommendation time, rather than relying on static POI embeddings pre-computed independently of context. Existing multimodal systems encode each POI once as a static embedding, a design that precludes reasoning about why the same cafe affords solo work on Monday but group celebration on Friday evening. We formally prove that no pre-computed encoder can satisfy context-sensitive ranking under standard bilinear scoring, motivating inference-time item-side representation. Agent4POI inverts this computation: given a situational context, a four-phase LLM agent generates dynamic, context-specific affordance queries (Phase 1) and executes a five-step cross-modal chain-of-thought over image, review, and metadata evidence (Phase 2). The resulting uncertainty-aware affordance representation is grounded in Gibsonian affordance theory. These cross-modal verdicts form a structured, uncertainty-adjusted affordance representation (Phase 3), which is aligned with user preferences via a semantic caching system for low-latency ranking (Phase 4). On three POI benchmarks and three evaluation configurations (standard, cold-start, context-shift), Agent4POI achieves a 23.2% relative gain over the strongest baseline and degrades by only 7.5% under context-shift versus 16--17\\% for the strongest baselines. In cold-start scenarios, Agent4POI outperforms the best content-based baseline by up to 2.4x, whereas ID-based methods fail to generalize.\n\n---\n\n## 200. Ensuring Logic in the Fog: Sound POMDP Synthesis with LTL Objectives\n- **Authors**: Can Zhou , Yulong Gao , Pian Yu\n- **URL**: [https://arxiv.org/abs/2605.12581](https://arxiv.org/abs/2605.12581)\n- **Abstract**:\n> Synthesising autonomous agents that can navigate uncertain environments while adhering to complex temporal constraints remains a fundamental challenge. While Linear Temporal Logic (LTL) provides a rigorous language for specifying such tasks, the inherent undecidability of qualitatively verifying LTL satisfaction in partially observable Markov decision processes renders quantitative synthesis difficult, especially when designing reliable reward signals for approximate solvers. In this paper, we bridge this gap with a novel, sound reward-shaping mechanism that dynamically generates belief-dependent rewards grounded in certified LTL satisfaction. By integrating this mechanism into an enhanced Monte Carlo Planning framework, we empower agents to navigate the `fog' of partial observability with a search process focused on maximising verifiable success. Our experiments demonstrate that this approach not only thrives in scenarios where existing solvers fail but also maintains effectiveness and scalability across diverse benchmark domains.\n\n---\n\n{% endraw %}\n"}, "ae95e160c1986855bbe27edfadb3713b4c8f4a25": {"t": 1786068570.829035, "n": 340278, "h": "05c8c40e431f029e", "c": "---\nlayout: default\ntitle: \"전체 AI 논문 - 2026-05-18\"\n---\n\n{% raw %}\n# 전체 AI 논문 - 2026-05-18\n\n## 1. Prospective multi-pathogen disease forecasting using autonomous LLM-guided tree search\n- **Authors**: Sarah Martinson , Michael P. Brenner , Martyna Plomecka , Brian P. Williams , Nicholas G. Reich , Zahra Shamsi\n- **URL**: [https://arxiv.org/abs/2605.16238](https://arxiv.org/abs/2605.16238)\n- **Abstract**:\n> Probabilistic forecasting of infectious diseases is crucial for public health but relies on labor-intensive manual model curation by expert modeling teams. This bespoke development bottlenecks scalability to granular geographic resolutions or emerging pathogens. Here, we present an autonomous system using Large Language Model (LLM)-guided tree search to iteratively generate, evaluate, and optimize executable forecasting software. In a fully prospective, real-time evaluation during the 2025-2026 US respiratory season, the system autonomously discovered methodologically diverse models for influenza, COVID-19, and respiratory syncytial virus (RSV). Aggregating these machine-generated models yielded an ensemble that consistently matched or outperformed the gold-standard, human-curated Centers for Disease Control and Prevention (CDC) hub ensembles out-of-sample. The system successfully navigated data-scarce \"cold start\" scenarios for RSV. Moreover, controlled retrospective ablations revealed that optimizing log-scale distance metrics prevents reward hacking, while an automated judge-in-the-loop ensures structural fidelity to complex scientific theories. By autonomously translating epidemiological theory into accurate, transparent code, this framework overcomes the modeling labor bottleneck, enabling rapid deployment of expert-level disease forecasting at unprecedented scales.\n\n---\n\n## 2. FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast\n- **Authors**: Igor Bogdanov , Chung-Horng Lung , Thomas Kunz , Jie Gao , Adrian Taylor , Marzia Zaman\n- **URL**: [https://arxiv.org/abs/2605.16233](https://arxiv.org/abs/2605.16233)\n- **Abstract**:\n> Can LLM agents improve decision-making through self-generated memory without gradient updates? We propose FORGE (Failure-Optimized Reflective Graduation and Evolution), a staged, population-based protocol that evolves prompt-injected natural-language memory for hierarchical ReAct agents. FORGE wraps a Reflexion-style inner loop, where a dedicated reflection agent (using the same underlying LLM, no distillation from a stronger model) converts failed trajectories into reusable knowledge artifacts: textual heuristics (Rules), few-shot demonstrations (Examples), or both (Mixed), with an outer loop that propagates the best-performing instance's memory to the population between stages and freezes converged instances via a graduation criterion. We evaluate on CybORG CAGE-2, a stochastic network-defense POMDP at a 30-step horizon against the B-line attacker, where all four tested LLM families (Gemini-2.5-Flash-Lite, Grok-4-Fast, Llama-4-Maverick, Qwen3-235B) exhibit strongly negative, heavy-tailed zero-shot rewards. Compared against both a zero-shot baseline and a Reflexion baseline (isolated single-stream learning), FORGE improves average evaluation return by 1.7-7.7$\\times$ over zero-shot and by 29-72% over Reflexion in all 12 model-representation conditions, reducing major-failure rates (below $-100$) to as low as $\\sim$1%. We find that (1) population broadcast is critical mechanism, with a no-graduation ablation confirming that broadcast carries the performance gains while graduation primarily saves compute; (2) Examples achieves the strongest returns for three of four models, Rules offers the best cost-reliability profile with $\\sim$40% fewer tokens; and (3) weaker baseline models benefit disproportionately, suggesting FORGE may mitigate capability gaps rather than amplify strong models. All evidence is confined to CAGE-2 B-line; cross-family findings are directional evidence.\n\n---\n\n## 3. Fully Open Meditron: An Auditable Pipeline for Clinical LLMs\n- **Authors**: Xavier Theimer-Lienhard , Mushtaha El-Amin , Fay Elhassan , Sahaj Vaidya , Victor Cartier-Negadi , David Sasu , Lars Klein , Mary-Anne Hartley\n- **URL**: [https://arxiv.org/abs/2605.16215](https://arxiv.org/abs/2605.16215)\n- **Abstract**:\n> Clinical decision support systems (CDSS) require scrutable, auditable pipelines that enable rigorous, reproducible validation. Yet current LLM-based CDSS remain largely opaque. Most \"open\" models are open-weight only, releasing parameters while withholding the data provenance, curation procedures, and generation pipelines that determine model behavior. Fully Open (FO) models, which expose the complete training stack end-to-end, do not currently exist in medicine. We introduce Fully Open Meditron, the first fully open pipeline for building LLM-CDSS, comprising a clinician-audited training corpus, a reproducible data construction and training framework, and a use-aligned evaluation protocol. The corpus unifies eight public medical QA datasets into a normalized conversational format and expands coverage with three clinician-vetted synthetic extensions: exam-style QA, guideline-grounded QA derived from 46,469 clinical practice guidelines, and clinical vignettes. The pipeline enforces system-wide decontamination, gold-label resampling of teacher generations, and end-to-end validation by a four-physician panel. We evaluate using an LLM-as-a-judge protocol over expert-written clinical vignettes, calibrated against 204 human raters. We apply the recipe to five FO base models (Apertus-70B/8B-Instruct, OLMo-2-32B-SFT, EuroLLM-22B/9B-Instruct). All MeditronFO variants are preferred over their bases. Apertus-70B-MeditronFO improves +6.6 points over its base (47.2% to 53.8%) on aggregate medical benchmarks, establishing a new FO SoTA. Gemma-3-27B-MeditronFO is preferred over MedGemma in 58.6% of LLM-as-a-judge comparisons and outperforms it on HealthBench (58% vs 55.9%). These results show that fully open pipelines can achieve state-of-the-art domain-specific performance without sacrificing auditability or reproducibility.\n\n---\n\n## 4. Confirming Correct, Missing the Rest: LLM Tutoring Agents Struggle Where Feedback Matters Most\n- **Authors**: Tahreem Yasir , Wenbo Li , Sam Gilson , Sutapa Dey Tithi , Xiaoyi Tian , Tiffany Barnes\n- **URL**: [https://arxiv.org/abs/2605.16207](https://arxiv.org/abs/2605.16207)\n- **Abstract**:\n> Effective tutoring requires distinguishing optimal, valid but suboptimal, and incorrect student solutions, a distinction central to intelligent tutoring systems (ITS) but untested for LLM-based tutors. As LLMs are increasingly explored as conversational complements to ITS, evaluating their diagnostic precision is essential. We present a benchmark of seven LLM feedback agents in propositional logic using knowledge-graph-derived ground truth across 10,836 solution--feedback pairs and three feedback conditions. Models achieved near-ceiling performance on optimal steps but systematically over-rejected valid but suboptimal reasoning and over-validated incorrect solutions, precisely where adaptive tutoring matters most. These failures persisted across models regardless of solution context, suggesting architectural rather than informational limits. Moreover, accurate diagnosis did not reliably produce pedagogically actionable feedback, revealing a gap between diagnostic judgment and instructional effectiveness. Our findings suggest that LLMs are better suited for hybrid architectures where KG-grounded models handle diagnosis while LLMs support open-ended scaffolding and dialogue.\n\n---\n\n## 5. Context, Reasoning, and Hierarchy: A Cost-Performance Study of Compound LLM Agent Design in an Adversarial POMDP\n- **Authors**: Igor Bogdanov , Chung-Horng Lung , Thomas Kunz , Jie Gao , Adrian Taylor , Marzia Zaman\n- **URL**: [https://arxiv.org/abs/2605.16205](https://arxiv.org/abs/2605.16205)\n- **Abstract**:\n> Deploying compound LLM agents in adversarial, partially observable sequential environments requires navigating several design dimensions: (1) what the agent sees, (2) how it reasons, and (3) how tasks are decomposed across components. Yet practitioners lack guidance on which design choices improve performance versus merely increase inference costs. We present a controlled study of compound LLM agent design in CybORG CAGE-2, a cyber defense environment modeled as a Partially Observable Markov Decision Process (POMDP). Reward is non-positive, so all configurations operate in a failure-mitigation mode. Our evaluation spans five model families, six models, and twelve configurations (3,475 episodes) with token-level cost accounting. We vary context representation (raw observations vs. a deterministic state-tracking layer with compressed history), deliberation (self-questioning, self-critique, and self-improvement tools, with optional chain-of-thought prompting), and hierarchical decomposition (monolithic ReAct vs. delegation to specialized sub-agents). We find that: (1) Programmatic state abstraction delivers the largest returns per token spent (RPTS), improving mean return by up to 76% over raw observations. (2) Distributing deliberation tools across a hierarchy degrades performance relative to hierarchy alone for all five model families, reaching up to 3.4$\\times$ worse mean return while using 1.8-2.7$\\times$ more tokens. We call this destructive pattern a deliberation cascade. (3) Hierarchical decomposition without deliberation achieves the best absolute performance for most models, and context engineering is generally more cost-effective than deliberation. These findings suggest a design principle for structured adversarial POMDPs: invest in programmatic infrastructure and clean task decomposition rather than deeper per-agent reasoning, as these strategies can interfere when combined.\n\n---\n\n## 6. Formal Methods Meet LLMs: Auditing, Monitoring, and Intervention for Compliance of Advanced AI Systems\n- **Authors**: Parand A. Alamdari , Toryn Q. Klassen , Sheila A. McIlraith\n- **URL**: [https://arxiv.org/abs/2605.16198](https://arxiv.org/abs/2605.16198)\n- **Abstract**:\n> We examine one particular dimension of AI governance: how to monitor and audit AI-enabled products and services throughout the AI development lifecycle, from pre-deployment testing to post-deployment auditing. Combining principles from formal methods with SoTA machine learning, we propose techniques that enable AI-enabled product and service developers, as well as third party AI developers and evaluators, to perform offline auditing and online (runtime) monitoring of product-specific (temporally extended) behavioral constraints such as safety constraints, norms, rules and regulations with respect to black-box advanced AI systems, notably LLMs. We further provide practical techniques for predictive monitoring, such as sampling-based methods, and we introduce intervening monitors that act at runtime to preempt and potentially mitigate predicted violations. Experimental results show that by exploiting the formal syntax and semantics of Linear Temporal Logic (LTL), our proposed auditing and monitoring techniques are superior to LLM baseline methods in detecting violations of temporally extended behavioral constraints; with our approach, even small-model labelers match or exceed frontier LLM judges. Our predictive and intervening monitors significantly reduce the violation rates of LLM-based agents while largely preserving task performance. We further show through controlled experiments that LLMs' temporal reasoning shows a pronounced degradation in accuracy with increasing event distance, number of constraints, and number of propositions.\n\n---\n\n## 7. An Algebraic Exposition of the Theory of Dyadic Morality\n- **Authors**: Kush R. Varshney\n- **URL**: [https://arxiv.org/abs/2605.16153](https://arxiv.org/abs/2605.16153)\n- **Abstract**:\n> This paper provides an algebraic exposition of the theory of dyadic morality (TDM), a psychological model of moral judgment grounded in a simple two-node template: an intentional agent causing harm to a vulnerable patient. We formalize TDM using structural causal modeling (SCM) notation and identify three psychological operators (typecasting operator, completion operator, and valence-dependent inference mechanism) that extend standard SCM to capture how people compute moral judgments under constraints. We address scalability challenges arising from TDM's dyadic limitation, showing how moral cognition compresses multi-node scenarios through node collapse and sequential processing. Drawing on this algebraic framework, we demonstrate concrete applications to AI policy design: detecting conflicting obligations, structuring helpfulness policies to preserve user agency, and designing post-failure communication as causal interventions. Finally, we recommend scoped, contextual measurement of mind perception over universal averaging to operationalize the theory empirically. This algebraic formalization enables neurosymbolic AI systems to compute morality in a way that is both mathematically rigorous and faithful to human moral cognition.\n\n---\n\n## 8. Look Before You Leap: Autonomous Exploration for LLM Agents\n- **Authors**: Ziang Ye , Wentao Shi , Yuxin Liu , Yu Wang , Zhengzhou Cai , Yaorui Shi , Qi Gu , Xunliang Cai , Fuli Feng\n- **URL**: [https://arxiv.org/abs/2605.16143](https://arxiv.org/abs/2605.16143)\n- **Abstract**:\n> Large language model based agents often fail in unfamiliar environments due to premature exploitation: a tendency to act on prior knowledge before acquiring sufficient environment-specific information. We identify autonomous exploration as a critical yet underexplored capability for building adaptive agents. To formalize and quantify this capability, we introduce Exploration Checkpoint Coverage, a verifiable metric that measures how broadly an agent discovers key states, objects, and affordances. Our systematic evaluation reveals that agents trained with standard task-oriented reinforcement learning consistently exhibit narrow and repetitive behaviors that impede downstream performance. To address this limitation, we develop a training strategy that interleaves task-execution rollouts and exploration rollouts, with each type of rollout optimized by its corresponding verifiable reward. Building on this training strategy, we propose the Explore-then-Act paradigm, which decouples information-gathering from task execution: agents first utilize an interaction budget to acquire grounded environmental knowledge, then leverage it for task resolution. Our results demonstrate that learning to systematically explore is imperative for building generalizable and real-world-ready agents.\n\n---\n\n## 9. Property-Guided LLM Program Synthesis for Planning\n- **Authors**: Augusto B. Corrêa , André G. Pereira , Jendrik Seipp\n- **URL**: [https://arxiv.org/abs/2605.16142](https://arxiv.org/abs/2605.16142)\n- **Abstract**:\n> LLMs have shown impressive success in program synthesis, discovering programs that surpass prior solutions. However, these approaches rely on simple numeric scores to signal program quality, such as the value of the solution or the number of passed tests. Because a score offers no guidance on why a program failed, the system must generate and evaluate many candidates hoping some succeed, increasing LLM inference and evaluation costs. We study a different approach: property-guided LLM program synthesis. Instead of scoring programs after evaluation, we check whether a candidate satisfies a formally defined property. When the property is violated, we stop the evaluation early and provide the LLM with a concrete counterexample showing exactly how the program failed. This feedback drastically reduces both the number of program generations and the evaluation cost, and can guide the LLM to generate stronger programs. We evaluate this approach on PDDL planning domains, asking the LLM to synthesize direct heuristic functions: every state reachable by strictly improving transitions has a strictly improving successor. A heuristic with this property leads hill-climbing algorithm directly to a goal state. A counterexample-guided repair loop generates one candidate program, checks the property over a training set, and returns the first case that violates the property. We evaluate our approach on ten planning domains with an out-of-distribution test set. The synthesized heuristics are effectively direct on virtually all test tasks, and compared to the best prior generation method our approach generates seven times fewer programs per domain on average, solves more tasks without using search, and requires several orders of magnitude less computation to evaluate candidates. Whenever a problem admits a verifiable property, property-guided LLM synthesis can reduce cost and improve program quality.\n\n---\n\n## 10. ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents\n- **Authors**: Chinmay Savadikar , Mingyu Zhao , Yuanzheng Zhu , Han Li , Shuang Xie , Alberto Castelo , Tianfu Wu , Lingyun Wang\n- **URL**: [https://arxiv.org/abs/2605.16116](https://arxiv.org/abs/2605.16116)\n- **Abstract**:\n> Developing and evaluating e-commerce web agents requires environments that preserve meaningful task structure while enabling controllable, reproducible, and scalable scientific comparison. Existing methodologies force a tradeoff: live storefronts provide realism but are non-stationary, difficult to inspect, and irreproducible, while hand-built sandbox benchmarks provide control but cover only a narrow range of layouts, catalogs, policies, and interaction patterns. We argue that the core bottleneck is methodological: the field lacks a scalable way to construct evaluation settings that are simultaneously realistic, diverse, controllable, inspectable, and reproducible. We introduce ShopGym, an integrated framework for realistic simulation and scalable benchmarking of e-commerce web agents. ShopGym is a framework for constructing e-commerce simulation environments and grounded benchmark tasks. Its simulation layer, ShopArena, converts live seed storefronts into self-contained sandbox shops through anonymized shop specifications and a staged, validated generation process. On top of these simulated storefronts, ShopGuru synthesizes benchmark tasks across seven skill categories, grounding each task in the shop's catalog, navigation structure, policies, and interaction affordances. Together, ShopArena and ShopGuru produce self-contained, resettable, inspectable, and stable evaluation artifacts that preserve structural properties and agent-evaluation signals relevant to shopping tasks. We validate the framework through graph-based structural analysis and agent-based behavioral evaluation with 224 generated tasks across six sandbox shops: three constructed with synthetic data and three with real data. Our results show that the synthetic shops preserve key structural properties of live storefronts, with agent performance on synthetic shops positively correlated with performance on live storefronts.\n\n---\n\n## 11. Sign-Separated Finite-Time Error Analysis of Q-Learning\n- **Authors**: Donghwan Lee\n- **URL**: [https://arxiv.org/abs/2605.16103](https://arxiv.org/abs/2605.16103)\n- **Abstract**:\n> This paper develops a sign-separated finite-time error analysis for constant step-size Q-learning. Starting from the switching-system representation, the error is decomposed into its componentwise negative and positive parts. The negative part is dominated by a lower comparison linear time-invariant (LTI) system associated with a fixed optimal policy, whereas the positive part is controlled by a linear switching system. The resulting bounds show that the negative-side LTI certificate is no slower than the positive-side switching certificate and may produce a faster exponential envelope. The analysis identifies a max-induced asymmetry in Q-learning error dynamics. This asymmetry is connected to overestimation: positive action-wise errors can be selected and propagated by the Bellman maximum, whereas negative errors admit an optimal-policy lower comparison. Finite-time bounds are provided for both deterministic and stochastic constant-step-size recursions.\n\n---\n\n## 12. Reasoners or Translators? Contamination-aware Evaluation and Neuro-Symbolic Robustness in Tax Law\n- **Authors**: Parisa Kordjamshidi , Samer Aslan , Madhavan Seshadri , Leslie Barrett , Enrico Santus\n- **URL**: [https://arxiv.org/abs/2605.16052](https://arxiv.org/abs/2605.16052)\n- **Abstract**:\n> Recent advances in large language models (LLMs) have significantly enhanced automated legal reasoning. Yet, it remains unclear whether their performance reflects genuine legal reasoning ability or artifacts of data contamination. We present a comprehensive empirical study of tax law reasoning approaches and implement a contamination detection protocol to rigorously assess LLM reliability. We show that performance can be inflated by contamination. Building on this analysis, we conduct a systematic evaluation, comparing monolithic LLMs with hybrid systems that translate statutory text into formal representations and delegate inference to symbolic solvers. We build a novel test suite designed to probe generalization to unseen documents via case and rule variations. Our findings indicate that legal reasoning is inherently compositional and that neuro-symbolic frameworks offer a more reliable and robust foundation for legal AI, as well as improved generalization to unobserved situations.\n\n---\n\n## 13. ScreenSearch: Uncertainty-Aware OS Exploration\n- **Authors**: Michael Solodko , Justin Wagle\n- **URL**: [https://arxiv.org/abs/2605.16024](https://arxiv.org/abs/2605.16024)\n- **Abstract**:\n> Desktop GUI agents operate under partial observability: visually similar screens can correspond to different underlying workflow states, so locally plausible actions can lead to sharply different outcomes. We frame this as a problem of computer/OS state exploration, where effective behavior requires both expanding the reachable frontier and reducing ambiguity before committing. We present ScreenSearch, a system that combines structural screen retrieval and deduplication with an ambiguity-aware PUCT graph-bandit for large-scale desktop exploration. The retrieval layer converts UIA trees into location-aware structural features, indexes related screens through sparse token search and metadata filters, and maintains a shared deduplicated state graph across VM workers. On top of this graph, we define a scalable ambiguity signal based on matched-action outcome dispersion. If similar screens produce different next states under the same action signature, the state should be probed further rather than treated as resolved. We use this signal together with frontier rewards to drive large-scale exploration and replay-start policy evaluation over the shared graph. Across 11 desktop applications, ScreenSearch collects over 1M screenshots and over 30K deduplicated states, yielding large exploration corpora with substantial cross-application and within-application diversity. On a fixed replay-start slice, we observe a clear novelty--ambiguity trade-off: some policies reduce ambiguity quickly while discovering little frontier. Ambiguity reduction alone is therefore not a sufficient exploration objective. Appendix ablations show that stronger proposal priors can materially improve unique-state discovery during corpus building. These results suggest that state identity, proposal quality, and ambiguity-aware search all matter when deciding when to probe and when to commit.\n\n---\n\n## 14. Petri Net Induced Heuristic Search for Resource Constrained Scheduling\n- **Authors**: Ido Lublin , Dor Atzmon , Izack Cohen\n- **URL**: [https://arxiv.org/abs/2605.15983](https://arxiv.org/abs/2605.15983)\n- **Abstract**:\n> We formulate the Resource-Constrained Project Scheduling Problem (RCPSP) as optimal search over the reachability graph of a Timed Transition Petri Net with Resources, using relative-delay tokens so that scheduling decisions correspond to transition firings in the induced state space. We solve the resulting problem with $A^*$ guided by a heuristic that combines Critical Path and resource-based lower bounds, and prove that it is consistent under our token-based time semantics. Experiments on the PSPLIB benchmarks show that the approach outperforms strong exact Mixed-Integer Linear Programming (MIP) baselines (SCIP, CBC) in both success rate and solve time. Per-instance analysis shows that heuristic search and MIP degrade along independent axes, resource tightness for $A^*$ and formulation size for MIP, with resource strength mediating which solver benefits from scale.\n\n---\n\n## 15. Learning Bilevel Policies over Symbolic World Models for Long-Horizon Planning\n- **Authors**: Dillon Z. Chen , Till Hofmann , Toryn Q. Klassen , Sheila A. McIlraith\n- **URL**: [https://arxiv.org/abs/2605.15975](https://arxiv.org/abs/2605.15975)\n- **Abstract**:\n> We tackle the challenge of building embodied AI agents that can reliably solve long-horizon planning problems. Imitation learning from demonstrations has shown itself to be effective in training robots to solve a diversity of complex tasks requiring fine motor control and manipulation over low-level (LL), continuous environments. Yet, it remains a difficult endeavour to generate long-horizon plans from imitation learning alone. In contrast, high-level (HL), symbolic abstractions facilitate efficient and interpretable long-horizon planning. We propose to combine the strengths of LL imitation learning for manipulation and control, and HL symbolic abstractions for long-horizon planning. We realise this idea via \\emph{bilevel policies} of the form $(\\pi^{\\mathrm{hl} }, \\pi^{\\mathrm{ll} })$, consisting of a neural policy $\\pi^{\\mathrm{ll} }$ learned from LL demonstrations, and an HL symbolic policy $\\pi^{\\mathrm{hl} }$ that is constructed from symbolic abstractions of the LL demonstrations combined with inductive generalisation. We implement these ideas in the BISON system. Experiments on extended MetaWorld benchmarks demonstrate that BISON generalises to long horizons and problems with greater numbers of objects than those solved by VLA and end-to-end methods, and is more time and memory efficient in training and inference. Notably, when ignoring LL execution, BISON's HL policies can solve HL problems with 10,000 relevant objects in under a minute. Project page: this https URL\n\n---\n\n## 16. Deterministic Event-Graph Substrates as World Models for Counterfactual Reasoning\n- **Authors**: Fabio Rovai\n- **URL**: [https://arxiv.org/abs/2605.15967](https://arxiv.org/abs/2605.15967)\n- **Abstract**:\n> We study event-graph substrates: a class of world models that represent agent state as an append-only log of typed RDF triples and answer counterfactual queries by forking the log under a structured intervention vocabulary. Substrates are inspectable at the triple level, support exact counterfactuals, and transfer across domains without learned components. We formalize the class, prove a duality between explanatory and counterfactual queries that reduces both to the same causal-ancestor traversal, and evaluate a 1,400-line CLEVRER-DSL interpreter atop a domain-agnostic substrate runtime at full CLEVRER validation scale (n=75,618). The substrate exceeds the NS-DR symbolic oracle on all four per-question categories (by 9.89, 20.26, 17.65, and 0.80 percentage points), and exceeds the parametric ALOE baseline on descriptive and explanatory while lagging on predictive and counterfactual. We also introduce twin-EventLog, a 500-specification Park-canonical Smallville counterfactual benchmark on which the substrate exceeds Llama-3.1-8B with full context by 18.80 points joint accuracy.\n\n---\n\n## 17. PAGER: Bridging the Semantic-Execution Gap in Point-Precise Geometric GUI Control\n- **Authors**: Jingxuan Wei , Xi Bai , Shan Liu , Caijun Jia , Zheng Sun , Xinglong Xu , Siyuan Li , Linzhuang Sun , Bihui Yu , Conghui He , Cheng Tan\n- **URL**: [https://arxiv.org/abs/2605.15963](https://arxiv.org/abs/2605.15963)\n- **Abstract**:\n> Large vision-language models have significantly advanced GUI agents, enabling executable interaction across web, mobile, and desktop interfaces. Yet these gains largely rely on a forgiving region-tolerant paradigm, where many nearby pixels inside the same component remain valid. Precise geometric construction breaks this assumption: actions must land on points in continuous canvas space rather than tolerant regions. Because geometric primitives carry ontological dependencies, a local coordinate error can induce cascading topological failures that distort downstream objects and invalidate the final construction. We identify this regime as precision-sensitive GUI tasks, requiring point-level accuracy, geometry-aware verification, and robustness to dependency-driven error propagation. To benchmark it, we introduce PAGE Bench, with 4,906 problems and over 224K process-supervised, pixel-level GUI actions. We further propose PAGER, a topology-aware agent that decomposes construction into dependency-structured planning and pixel-level execution. Pixel-grounded supervised tuning establishes executable action grammar, while precision-aligned reinforcement learning mitigates rollout-induced exposure bias through state-conditioned geometric feedback. Experiments reveal a pronounced Semantic-Execution Gap: general multimodal models can exceed 88% action type accuracy yet remain below 6% task success. PAGER closes this gap, delivering 4.1x higher task success than the strongest evaluated general baseline and raising step success rate from below 9% for GUI-specialized agents to over 62%, establishing a new state of the art for point-precise GUI control.\n\n---\n\n## 18. Imperfect World Models are Exploitable\n- **Authors**: Logan Mondal Bhamidipaty (University of Edinburgh), Esmeralda S. Whitammer (University of Edinburgh), David Abel (University of Edinburgh), Mykel J. Kochenderfer (Stanford University), Subramanian Ramamoorthy (University of Edinburgh)\n- **URL**: [https://arxiv.org/abs/2605.15960](https://arxiv.org/abs/2605.15960)\n- **Abstract**:\n> We propose a novel definition of model exploitation in reinforcement learning. Informally, a world model is exploitable if it implies that one policy should be strictly preferred over another while the environment's true transition model implies the reverse. We analogize our definition with a prior characterization of reward hacking but show that the associated proof of inevitability does not transfer to exploitation. To overcome this obstruction, we develop a general theory of reward hacking and model exploitation that proves that exploitation is essentially unavoidable on large policy sets and yields the corresponding claim for hacking as a special case. Unfortunately, we also find that the conditions that guarantee unhackability in finite policy sets have no counterpart that precludes exploitation. Consequently, we introduce a relaxed notion of exploitation and derive a safe horizon within which it can be avoided. Taken together, our results establish a formal bridge between reward hacking and model exploitation and elucidate the limits of safe planning in world models.\n\n---\n\n## 19. Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design\n- **Authors**: Alberto Pepe , Chien-Yu Lin , Despoina Magka , Bilge Acun , Yannan Nellie Wu , Anton Protopopov , Carole-Jean Wu , Yoram Bachrach\n- **URL**: [https://arxiv.org/abs/2605.15871](https://arxiv.org/abs/2605.15871)\n- **Abstract**:\n> Toward recursive self-improvement, we investigate LLM agents autonomously designing foundation models beyond standard Transformers. We introduce a dual-framework approach: AIRA-Compose for high-level architecture search, and AIRA-Design for low-level mechanistic implementation. AIRA-Compose uses 11 agents to explore fundamental computational primitives under a 24-hour budget. Agents evaluate million-parameter candidates, extrapolating top designs to 350M, 1B, and 3B scales. This yields 14 architectures across two families: AIRAformers (Transformer-based) and AIRAhybrids (Transformer-Mamba). Pre-trained at 1B scale, these consistently outperform Llama 3.2 and Composer-found baselines. On downstream tasks, AIRAformer-D and AIRAhybrid-D improve accuracy by 2.4% and 3.8% over Llama 3.2. Furthermore, AIRA-Compose finds models with highly efficient scaling frontiers: AIRAformer-C scales 54% and 71% faster than Llama 3.2 and Composer's best Transformer, while AIRAhybrid-C outscales Nemotron-2 by 23% and Composer's best hybrid by 37%. AIRA-Design tasks 20 agents with writing novel attention mechanisms for long-range dependencies and high-performing training scripts. On the Long Range Arena benchmark, agent-designed architectures reach within 2.3% and 2.6% of human state-of-the-art on document matching and text classification. On the Autoresearch benchmark, Greedy Opus 4.5 achieves 0.968 validation bits-per-byte under a fixed time budget, surpassing the published minimum. Together, these frameworks show AI agents can autonomously discover architectures and algorithmic optimizations matching or surpassing hand-designed baselines. This establishes a powerful paradigm for discovering next-generation foundation models, marking a clear step toward recursive self-improvement.\n\n---\n\n## 20. SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows?\n- **Authors**: Kean Shi , Zihang Li , Tianyi Ma , Zengji Tu , Jialong Wu , Xinbo Xu , Qingyao Yang , Ruoyu Wu , Weichu Xie , Ming Wu , Jason Zeng , Michael Heinrich , Elvis Zhang , Liang Chen , Kuan Li , Baobao Chang\n- **URL**: [https://arxiv.org/abs/2605.15777](https://arxiv.org/abs/2605.15777)\n- **Abstract**:\n> Computer-Using Agents (CUAs) are rapidly extending large language models (LLMs) beyond text-based reasoning toward action execution in more complex environments, such as web browsers and graphical user interfaces (GUIs). However, existing web and GUI agent benchmarks often rely on simplified settings, isolated tasks, or short-horizon interactions, making it difficult to assess capabilities of agents in realistic professional workflows. Software-as-a-Service (SaaS) environments are a natural choice for CUA evaluation, as they host a large share of modern digital work and naturally involve dynamic system states, cross-application coordination, domain-specific knowledge, and long-horizon dependencies. To this end, we introduce SaaS-Bench, a benchmark built on 23 deployable SaaS systems across six professional domains, containing 106 tasks grounded in realistic work scenarios. These tasks require long-horizon execution, cover both text-only and multimodal settings, and are evaluated with weighted verification checkpoints that measure strict task completion and partial progress. Experiments show that representative LLM-based agents struggle on SaaS-Bench, with even the strongest model completing fewer than 4% of tasks end-to-end, exposing limitations in planning, state tracking, cross-application context maintenance, and error recovery. Code are available at this https URL for reproduction.\n\n---\n\n## 21. ALSO: Adversarial Online Strategy Optimization for Social Agents\n- **Authors**: Xiang Li , Liping Yi , Mingze Kong , Min Zhang , Zhongxiang Dai , QingHua Hu\n- **URL**: [https://arxiv.org/abs/2605.15768](https://arxiv.org/abs/2605.15768)\n- **Abstract**:\n> Social simulation provides a compelling testbed for studying social intelligence, where agents interact through multi-turn dialogues under evolving contexts and strategically adapting opponents. Such environments are inherently non-stationary, requiring agents to dynamically adjust their strategies over time. However, most Large Language Model (LLM) based social agents rely on static personas, while existing approaches for enhancing social intelligence, such as offline reinforcement learning or external planners, are ill-suited to these settings, typically assuming stationarity and incurring substantial training overhead. To bridge this gap, we propose \\textbf{ALSO} (\\textbf{A}dversarial on\\textbf{L}ine \\textbf{S}trategy \\textbf{O}ptimization), the first framework for online strategy optimization in multi-agent social simulation. ALSO advances social adaptation through two key contributions. (1) ALSO formulates multi-turn interaction as an adversarial bandit problem, where combinations of static personas and dynamic strategy instructions are treated as arms, providing a principled solution to non-stationarity without relying on environmental stability assumptions. (2) To predict rewards and generalize sparse feedback in multi-turn dialogues, ALSO introduces a lightweight neural surrogate to predict rewards from interaction histories, enabling sample-efficient exploration and continuous online adaptation. Experiments on the Sotopia benchmark demonstrate that ALSO consistently outperforms static baselines and existing optimization methods in dynamic environments, validating the effectiveness of adversarial online strategy optimization for building robust social agents.\n\n---\n\n## 22. Can We Trust AI-Inferred User States. A Psychometric Framework for Validating the Reliability of Users States Classification by LLMs in Operational Environments\n- **Authors**: Izabella Krzeminska , Michal Butkiewicz , Ewa Komkowska\n- **URL**: [https://arxiv.org/abs/2605.15734](https://arxiv.org/abs/2605.15734)\n- **Abstract**:\n> The use of large language models to assess user states in conversational and adaptive systems is based on the assumption that the metrics used for such assessment are stable and interpretable at the level of individual scores. This paper empirically tests this assumption, focusing on the psychometric reliability of artificial intelligence (AI) measures of user states. This study employed replication evaluation procedures to assess the repeatability of a broad set of metrics across three different bimodal large language models (GPT-4o audio, Gemini 2.0 Flash, Gemini 2.5 Flash). Analyses include both individual score reliability and aggregated reliability, allowing us to distinguish metrics potentially useful for real-time adaptation from those that retain their value only in aggregated analyses. The results demonstrate that metric reliability cannot be considered a default property in interpretive domains. The lack of stability at the level of individual scores precludes the interpretation of such scores as indicators of user state in real-time adaptive systems, even if these metrics demonstrate stability after aggregation. At the same time, the study indicates that individually unstable metrics can retain analytical utility in post-hoc studies, identifying rules governing interactions and their relationships with user experience parameters such as satisfaction, trust, and engagement. The main contribution of this work, besides quantifying the severity of the problem (only 31 of 213 metrics met the criteria), is the proposal of a replicable evaluation framework, enabling measurable evaluations of metric applicability. This approach supports more responsible AI design of adaptive systems, in which the interpretation of results requires explicit validation of reliability and monitoring for violations over time.\n\n---\n\n## 23. Nudging Beyond the Comfort Zone: Efficient Strategy-Guided Exploration for RLVR\n- **Authors**: Chanuk Lee , Sangwoo Park , Minki Kang , Sung Ju Hwang\n- **URL**: [https://arxiv.org/abs/2605.15726](https://arxiv.org/abs/2605.15726)\n- **Abstract**:\n> Reinforcement learning with verifiable rewards (RLVR) has emerged as a scalable paradigm for improving the reasoning capabilities of large language models. However, its effectiveness is fundamentally limited by exploration: the policy can only improve on trajectories it has already sampled. While increasing the number of rollouts alleviates this issue, such brute-force scaling is computationally expensive, and existing approaches that modify the optimization objective provide limited control over what is explored. In this work, we propose NudgeRL, a framework for structured and diversity-driven exploration in RLVR. Our approach introduces Strategy Nudging, which conditions each rollout on lightweight, strategy-level contexts to induce diverse reasoning trajectories without relying on expensive oracle supervision. To effectively learn from such structured exploration, we further propose a unified objective, which decomposes the reward signal into inter- and intra-context components and incorporates a distillation objective to transfer discovered behaviors back to the base policy. Empirically, NudgeRL outperforms standard GRPO with up to 8 times larger rollout budgets, while outperforming oracle-guided RL baseline on average across five challenging math benchmarks. These results demonstrate that structured, context-driven exploration can serve as an efficient and scalable alternative to both brute-force rollout scaling and feasibility-oriented methods based on privileged information. Our code is available at this https URL .\n\n---\n\n## 24. PRISM: Prompt Reliability via Iterative Simulation and Monitoring for Enterprise Conversational AI\n- **Authors**: Keshava Chaitanya , Jahnavi Gundakaram\n- **URL**: [https://arxiv.org/abs/2605.15665](https://arxiv.org/abs/2605.15665)\n- **Abstract**:\n> Deploying large language model (LLM)-driven conversational agents in enterprise settings requires prompts that are simultaneously correct at launch and resilient to the non-deterministic behavioral drift that characterizes production LLM deployments. Existing prompt optimization frameworks address prompt quality as a one-time compile-time problem, leaving open the equally critical question of how to detect and repair prompt regressions caused by silent LLM behavior changes over time. We present PRISM (Prompt Reliability via Iterative Simulation and Monitoring), a closed-loop framework that treats prompt engineering as a continuous reliability engineering problem rather than a one-time authorship task. PRISM takes as input plain-language agent requirements, a set of configured tools and memory variables, and an initial draft prompt. It automatically generates test cases from requirements, simulates full multi-turn conversations against a platform-faithful LLM environment, evaluates pass/fail using an LLM-as-judge, diagnoses root causes of failures, and surgically repairs the prompt -- iterating until all tests pass. Critically, PRISM is designed to run on a scheduled basis (daily), treating LLM behavioral drift as a first-class reliability concern. We evaluate PRISM across 35 enterprise conversational agents over a three-week deployment period on the this http URL V3 platform. PRISM reduces median prompt authoring time from 2 days to under 30 minutes, achieves 99% production reliability across all evaluated agents, and successfully identifies and repairs production regressions caused by LLM behavioral drift within a 24-hour detection window. Our results suggest that continuous, simulation-driven prompt optimization is both tractable and necessary for reliable enterprise conversational AI at scale.\n\n---\n\n## 25. ColPackAgent: Agent-Skill-Guided Hard-Particle Monte Carlo Workflows for Colloidal Packing\n- **Authors**: Lijie Ding , Changwoo Do\n- **URL**: [https://arxiv.org/abs/2605.15625](https://arxiv.org/abs/2605.15625)\n- **Abstract**:\n> We introduce ColPackAgent, an agent framework that autonomously runs Monte Carlo simulations of colloidal packing through a Model Context Protocol (MCP) tool server and an agent skill, whether as a standalone agent or inside an existing agent system. By harnessing the MCP server and agent skill, ColPackAgent executes a structured workflow for colloidal packing simulations, which are central to studies of phase behavior, self-assembly, and materials design. Without dedicated simulation tools and workflow instructions, general-purpose Large Language Model (LLM) agents tend to describe such workflows rather than execute them reliably. The MCP server exposes a custom-built colpack Python package that wraps HOOMD-blue hard-particle Monte Carlo, and the skill encodes a four-stage workflow contract. ColPackAgent can carry out the workflow interactively with human feedback, autonomously from an end-to-end prompt, or as autoresearch following a provided program file. We demonstrate the system in different modes with several colloidal packing simulation examples such as cube particles in 3D, a binary system of disks and capsules in 2D, and the 2D hard-disk freezing transition using autoresearch. We also compare model performance on this workflow across a panel of LLMs with 17 stage-specific prompts. This benchmark provides a stage-level check of how reliably different models follow the setup, planning, and analysis workflow. Together, these results show that pairing a domain Python package with MCP tools and a portable agent skill provides a practical route for turning a simulation toolkit into an agent-assisted research workflow.\n\n---\n\n## 26. TopoEvo: A Topology-Aware Self-Evolving Multi-Agent Framework for Root Cause Analysis in Microservices\n- **Authors**: Junle Wang , Xingchuang Liao , Wenjun Wu\n- **URL**: [https://arxiv.org/abs/2605.15611](https://arxiv.org/abs/2605.15611)\n- **Abstract**:\n> Root cause analysis (RCA) in microservices is challenging due to (i) noisy and heterogeneous multimodal observability (metrics, logs, traces), (ii) cascading failure propagation that amplifies downstream symptoms, and (iii) non-stationary topology drift induced by autoscaling and rolling updates. Recent LLM-based RCA agents can generate tool-grounded explanations, yet they often remain topology-agnostic and suffer from \\emph{symptom-amplification bias}, misattributing the root cause to salient downstream victims. We propose \\textbf{TopoEvo}, a topology-aware self-evolving multi-agent framework that couples graph representation learning with structured, topology-constrained reasoning. TopoEvo first introduces \\emph{Metric-orthogonal Multimodal Alignment} (MOMA), which decomposes metric embeddings into complementary subspaces and contrastively aligns logs and traces to reduce modality redundancy and sparsity, yielding stable node representations for graph encoding. It then applies \\emph{Vector Quantization} (VQ) to discretize topology-enhanced states into auditable \\emph{symptom tokens} with a symptom lexicon, enabling reliable retrieval and token-level evidence grounding. On top of these discrete topology cues, TopoEvo performs a multi-agent \\emph{Hypothesis--Evidence--Test} (HET) workflow to explicitly verify propagation-consistent explanations and separate initiating anomalies from amplified downstream symptoms. Finally, a \\emph{Self-Evolving Mechanism} refreshes hierarchical incident memory and performs conservative test-time adaptation with high-confidence pseudo-labels to maintain robustness under drift.\n\n---\n\n## 27. See Before You Code: Learning Visual Priors for Spatially Aware Educational Animation Generation\n- **Authors**: Yuejia Li , Ke He , Junheng Li , Shutong Chen , Jingkang Xia , Zhiyue Su , Junchi Zhang , Mang Ye\n- **URL**: [https://arxiv.org/abs/2605.15585](https://arxiv.org/abs/2605.15585)\n- **Abstract**:\n> Large language models can generate executable code for educational animations, but the resulting renders often exhibit visual defects, including element overlap, misalignment, and broken animation continuity. These defects cannot be reliably detected from the code alone and become apparent only after execution. We formalize this problem as render-feedback-aware constrained code generation: given a natural language specification, the model must generate executable code whose rendered output satisfies structured quality criteria that can be evaluated only after rendering. To address this problem, we introduce OmniManim, a render-feedback-aware educational animation generation framework built around a shared scene state, explicit visual planning, structured post-render diagnostics, and localized repair. Within OmniManim, the Vision Agent is a task-specific visual planning module: it predicts sparse keyframe layouts with coarse-to-fine bounding-box denoising and optimizes an interpolation-aware objective to reduce intermediate-frame failures induced by downstream animation interpolation. We further construct two datasets, ManimLayout-1K and EduRequire-500, and provide a reproducible evaluation protocol covering executability, instructional quality, visual quality, and efficiency. On EduRequire-500, OmniManim improves measured render quality over both single-model baselines and existing multi-agent frameworks. Systematic ablation studies further verify that explicit visual planning, especially its coarse spatial prior, bounding-box refinement, and interpolation-aware optimization, is central to these gains.\n\n---\n\n## 28. STAR: A Stage-attributed Triage and Repair framework for RCA Agents in Microservices\n- **Authors**: Junle Wang , Xingchuang Liao , Wenjun Wu\n- **URL**: [https://arxiv.org/abs/2605.15581](https://arxiv.org/abs/2605.15581)\n- **Abstract**:\n> LLM-based root cause analysis (RCA) agents have recently emerged as a promising paradigm for incident diagnosis in microservice AIOps. However, their reliability remains fragile: an error in early evidence collection, hypothesis formulation, or causal analysis can propagate through the reasoning trace and eventually corrupt the final diagnosis. In this paper, we present \\textbf{STAR}, a \\emph{Stage-attributed Triage and Repair} framework for repairing erroneous RCA traces. STAR explicitly decomposes an RCA workflow into four structured stages, namely \\emph{Evidence Package} (EP), \\emph{Hypothesis Set} (HS), \\emph{Analysis Structure} (AS), and \\emph{Decision Report} (DR), and treats agent failure as a stage-localizable reasoning bug rather than a monolithic end-to-end error. Built on top of LangGraph, STAR performs stage-wise auditing, budget-aware \\emph{Fast/Slow Routing}, \\emph{decisive stage localization via counterfactual candidate evaluation}, and stage-specific patch-and-replay repair. We evaluate STAR on a public large-scale benchmark and a real-world production dataset, using two RCA agent workflows and three foundation models. Experimental results show that STAR consistently improves both root cause localization and fault type classification over strong baselines. Moreover, STAR identifies the decisive faulty stage with high accuracy, repairs most initially incorrect traces within one or two replay rounds, and benefits substantially from both Fast/Slow Routing and counterfactual stage evaluation. These results suggest that explicitly modeling \\emph{where} an RCA agent fails is an effective path toward reliable, debuggable, and self-repairing agentic RCA systems.\n\n---\n\n## 29. Position: Artificial Intelligence Needs Meta Intelligence -- the Case for Metacognitive AI\n- **Authors**: Sergei Chuprov , Richard D. Lange , Leon Reznik , Paulo Shakarian , Raman Zatsarenko , Dmitrii Korobeinikov\n- **URL**: [https://arxiv.org/abs/2605.15567](https://arxiv.org/abs/2605.15567)\n- **Abstract**:\n> This position paper argues for metacognition as a general design principle for creating more accurate, secure, and efficient AI. The metacognitive solution involves systems monitoring their own states and judiciously allocating resources depending on each problem instance's difficulty or cost of mistakes. Drawing inspiration both from past work on resource-rational AI and from well-documented metacognitive strategies in psychology and cognitive science, we identify specific challenges in embedding these strategies into AI design and highlight open theoretical and implementation problems. We showcase these principles through a tangible example of improved learning efficiency, effectiveness, and security in a Federated Learning (FL) case study. We show how these principles can be translated into practice with a novel software framework developed specifically to allow the community to design, deploy, and experiment with metacognition-enabled AI applications.\n\n---\n\n## 30. DRS-GUI: Dynamic Region Search for Training-Free GUI Grounding\n- **Authors**: Yichao Liu , Huawen Shen , Liu Yu , Shiyu Liu , Zeyu Chen , Yu Zhou\n- **URL**: [https://arxiv.org/abs/2605.15542](https://arxiv.org/abs/2605.15542)\n- **Abstract**:\n> GUI agents powered by Multimodal Large Language Models (MLLMs) have demonstrated impressive capability in understanding and executing user instructions. However, accurately grounding instruction-relevant elements from high-resolution screenshots cluttered with irrelevant UI components remains challenging for existing approaches. Inspired by how humans dynamically adjust their perceptual scope to locate task-related regions on complex screens, we propose DRS-GUI, a training-free dynamic region search framework for GUI grounding that can be seamlessly integrated into existing MLLMs. DRS-GUI introduces a lightweight UI Perceptor that performs three human-like perceptual actions (Focus, Shift, and Scatter) to progressively explore the interface and generate region proposals. To dynamically schedule these actions, we further design an Action Planner based on Monte Carlo Tree Search (MCTS). A region quality reward is employed to evaluate and select the highly instruction-relevant region, efficiently pruning redundant UI elements. Experiments demonstrate that DRS-GUI yields a 14\\% improvement on ScreenSpot-Pro for general and GUI-specific MLLMs (Qwen2.5-VL-7B and UGround-V1-7B), significantly enhancing grounding performance and generalization.\n\n---\n\n## 31. RTL-BenchMT: Dynamic Maintenance of RTL Generation Benchmark Through Agent-Assisted Analysis and Revision\n- **Authors**: Jing Wang , Shang Liu , Hangan Zhou , Zhiyao Xie\n- **URL**: [https://arxiv.org/abs/2605.15537](https://arxiv.org/abs/2605.15537)\n- **Abstract**:\n> This paper introduces RTL-BenchMT, an agentic framework for dynamically maintaining RTL generation benchmarks. Large Language Models (LLMs) assisted automated RTL generation is one of the most important directions in EDA research. However, current RTL benchmarks face two critical challenges: (1) flawed cases in the benchmarks and (2) overfitting to the benchmarks. Both challenges are difficult to resolve purely by manual engineering effort. To address these issues and systematically reduce human maintenance costs, we propose an automated agentic framework, RTL-BenchMT. RTL-BenchMT focuses on two key applications: (1) automatically identifying and revising flawed benchmark cases and (2) automatically detecting and updating overfitting cases. With the assistance of RTL-BenchMT, we conduct a thorough, in-depth analysis of flawed and overfitting cases and produce a refined benchmark suite that will be open-sourced to the community.\n\n---\n\n## 32. CAPS: Cascaded Adaptive Pairwise Selection for Efficient Parallel Reasoning\n- **Authors**: Fangzhou Lin , Shuo Xing , Peiran Li , Siyuan Yang , Qianwen Ge , Kazunori Yamada , Ziming Zhang , Haichong Zhang , Zhengzhong Tu\n- **URL**: [https://arxiv.org/abs/2605.15513](https://arxiv.org/abs/2605.15513)\n- **Abstract**:\n> Parallel reasoning, where a generator samples many candidate solutions and an aggregator selects the best, is one of the most effective forms of test-time scaling in large language models, and pairwise self-verification has become its strongest aggregation primitive. Yet pairwise verification carries a heavy cost: each judgment reads two complete solutions in full, and existing methods perform tens of such judgments per problem regardless of whether the comparison is informative. We introduce CAPS (Cascaded Adaptive Pairwise Selection), an inference-only framework that allocates verifier compute non-uniformly along two orthogonal axes: an evidence axis that adapts how much of each candidate the judge sees, and a distribution axis that adapts how comparisons are spread across the pool. CAPS instantiates these into a four-stage cascade with an optional rescue subroutine, and admits a closed-form verifier-token cost in which the per-candidate marginal cost is roughly halved relative to uniform full-evidence schedules. On four self-verifying models (Qwen3-14B, GPT-OSS-20B, Qwen3-4B-Instruct/Thinking) and five reasoning benchmarks spanning code (LiveCodeBench-v5/v6, CodeContests) and math (AIME 2025, HMMT 2025), CAPS outperforms the leading pairwise verifier on 14 of 20 suites while using 25.4% of its verifier-token budget on code, and outperforms pointwise self-verification on all 20. The trade-off suites admit an interpretable diagnostic in terms of the verifier's accuracy at partial versus full evidence, providing a concrete pre-deployment check for cascade suitability.\n\n---\n\n## 33. X-SYNTH: Beyond Retrieval -- Enterprise Context Synthesis from Observed Human Attention\n- **Authors**: Guruprasad Raghavan , George Nychis , Rohan Narayana Murthy\n- **URL**: [https://arxiv.org/abs/2605.15505](https://arxiv.org/abs/2605.15505)\n- **Abstract**:\n> In enterprise operations, the context required for an AI agent task is scattered across systems of record, static information stores, and communication channels. What is stored is system state, a lossy representation of the work that actually happened [2, 52]. The prevailing approach [17, 31, 34, 36] retrieves by matching request content to what is stored; for narrow requests this works well. But synthesis quality depends on knowing what to surface and how to interpret it: knowledge specific to each organization, team, and individual [5, 57, 61], present in behavioral patterns, absent from any retrieval index. For complex agentic tasks it breaks down: True Lead Rate is low, False Lead Rate is high, and the model has no mechanism to improve. We present X-SYNTH, a framework for enterprise context synthesis grounded in human attention, the digitally observable interaction signatures of each worker, encoding not just what they did but the sequence in which they did it, along with implicit reward signals. Behavioral traces preceding positive outcomes are distinguishable from those that did not, without external labeling. X-SYNTH models each individual's behavioral baseline as a Digital Twin Signature (DTS) and selects among seven qualitatively distinct attention filters: Proportional, Inverse, Differential, Recurrent, Comparative, Sequential, and Collective, per individual and per query, to identify causally relevant activity signatures. A four-stage pipeline assembles ranked context grounded in behavioral patterns rather than query embeddings. On a sales lead identification task, a frontier model unaided achieves 9.5% True Lead Rate (TLR) with 90.5% False Lead Rate (FLR). Augmented with X-SYNTH, TLR rises to 61.9% (6.5x) while FLR falls to 18.8%. Enterprise context synthesis is not a retrieval problem. It is a relevance problem, and human attention is its most reliable ground truth.\n\n---\n\n## 34. From LLM-Generated Conjectures to Lean Formalizations: Automated Polynomial Inequality Proving via Sum-of-Squares Certificates\n- **Authors**: Ruobing Zuo , Hanrui Zhao , Gaolei He , Zhengfeng Yang , Jianlin Wang\n- **URL**: [https://arxiv.org/abs/2605.15445](https://arxiv.org/abs/2605.15445)\n- **Abstract**:\n> Automated proving of polynomial inequalities is a fundamental challenge in automated mathematical reasoning, where rich algebraic structure and a rapidly growing certificate search space hinder scalability. Purely symbolic approaches provide strong guarantees but often scale poorly as the number of variables or the degree increases, due to expensive algebraic manipulations and rapidly growing intermediate expressions. In parallel, LLM-guided methods have made notable progress, particularly on competition-style inequalities with a small number of variables. To address the remaining scalability challenges, we propose NSPI, a neuro-symbolic framework that combines the complementary strengths of LLMs and symbolic computation for polynomial-inequality proving. Concretely, an LLM proposes a conjecture in the form of an approximate polynomial Sum-Of-Squares (SOS) decomposition; we refine it via symbolic computation to obtain an exact polynomial SOS representation, which directly proves the target inequality, and we further certify the proof in Lean, yielding an end-to-end pipeline from heuristic discovery to machine-checked proof. Experiments on challenging benchmarks involving polynomials with up to 10 variables demonstrate the effectiveness and scalability of the proposed method.\n\n---\n\n## 35. Beyond Partner Diversity: An Influence-Based Team Steering Framework for Zero-Shot Human-Machine Teaming\n- **Authors**: Wei Sheng , Rohan Paleja\n- **URL**: [https://arxiv.org/abs/2605.15400](https://arxiv.org/abs/2605.15400)\n- **Abstract**:\n> While AI agents are rapidly advancing from isolated tools to interactive collaborators, data-driven human-machine teaming (HMT) methods remain costly in their reliance on human interaction data across domains, teammates, and team sizes. Zero-shot coordination (ZSC) addresses this bottleneck by simulating diverse partner populations to approximate how unseen partners might behave. However, partner coverage alone is insufficient as team settings scale and communication becomes degraded. To remedy this deficiency, we propose Influence-Based Team Steering (IBTS), a framework that uses influence shaping to incentivize agents to discover diverse, high-performing team interaction patterns and further steers ongoing trajectories toward stronger learned coordination modes. We assess IBTS on Overcooked-AI in both two-agent and three-agent settings, allowing us to test whether learned coordination structure transfers beyond dyadic interaction. Our evaluation includes simulated partners, synthetic partner-style variation, and, to our knowledge, the first 30-subject Overcooked-AI HMT study involving two real human teammates and one machine teammate. Across these evaluations, IBTS improves team performance against competing baselines, highlighting the need for scaled ZSC to combine sparse-reward coordination mechanisms with partner-variation coverage rather than relying on diversity alone.\n\n---\n\n## 36. Ensemble Monitoring for AI Control: Diverse Signals Outweigh More Compute\n- **Authors**: Eugene Koran , Yejun Yun , Samantha Tetef , Benjamin Arnav , Pablo Bernabeu-Pérez\n- **URL**: [https://arxiv.org/abs/2605.15377](https://arxiv.org/abs/2605.15377)\n- **Abstract**:\n> As AI systems are increasingly deployed in autonomous agentic settings at scale, it is important to ensure the actions they take are safe and aligned with user intent. Monitoring agent actions is a key safety mechanism, yet reliable monitors remain difficult to build and the scale of these systems makes human oversight impractical. We show that combining signals from diverse monitors into an ensemble improves detection of misaligned actions. We build 12 GPT-4.1-Mini monitors using both prompting and fine-tuning strategies. We evaluate them on coding tasks where candidate solutions pass standard tests but fail on adversarial inputs. In this setting, diverse ensembles outperform both individual monitors and homogeneous ensembles. Our best 3-monitor ensemble achieves 2.4x greater detection performance gain compared to an ensemble composed of three identical monitors, with the same ensemble performing strongly on an independent dataset. We contend that these results show that diversity - not scale - drives gains. The best ensembles combine strong individual performance with low correlation between monitors. Furthermore, fine-tuned monitors appear in every top-performing ensemble and maintain this advantage on out-of-distribution attack types, suggesting that fine-tuning enables detection capabilities that prompting alone does not elicit. These results support ensemble monitoring as a practical AI control strategy for safety gains at reasonable inference costs.\n\n---\n\n## 37. Belief Engine: Configurable and Inspectable Stance Dynamics in Multi-Agent LLM Deliberation\n- **Authors**: Joshua C. Yang , Maurice Flechtner , Damian Dailisan , Michiel A. Bakker\n- **URL**: [https://arxiv.org/abs/2605.15343](https://arxiv.org/abs/2605.15343)\n- **Abstract**:\n> LLM-based agents are increasingly used to simulate deliberative interactions such as negotiation, conflict resolution, and multi-turn opinion exchange. Yet generated transcripts often do not reveal why an agent's stance changes: movement may reflect evidence uptake, anchoring, role drift, echoing, or changed prompt and retrieval context. We introduce the Belief Engine (BE), an auditable belief-update layer that treats \"belief\" as an evidential state over a proposition and exposes it as scalar stance. BE extracts arguments into structured memory and updates stance with a log-odds rule controlled by evidence uptake u and prior anchoring a. Across multiple base LLMs, parameter sweeps show that these controls reliably shape stance dynamics while preserving an evidence-level update trail. On DEBATE, a human deliberation dataset with pre/post opinions, BE best reconstructs participants whose final stance follows extracted evidence; stable and evidence-opposed cases instead point to anchoring or factors outside the extracted evidence stream. BE provides configurable infrastructure for studying evidence-grounded deliberation, where openness, commitment, convergence, and disagreement can be tied to explicit update assumptions rather than hidden prompt effects.\n\n---\n\n## 38. Zero-Shot Goal Recognition with Large Language Models\n- **Authors**: Kin Max Piamolini Gusmão , Nathan Gavenski , Nir Oren , Felipe Meneguzzi\n- **URL**: [https://arxiv.org/abs/2605.15333](https://arxiv.org/abs/2605.15333)\n- **Abstract**:\n> Large language models have recently reached near-parity with classical planners on well-known planning domains, yet this competence relies on world-knowledge exploitation rather than genuine symbolic reasoning. Goal recognition is a complementary abductive task structurally better suited to LLM strengths: it consists of evaluating consistency with world knowledge rather than generating novel action sequences. This paper provides the first systematic zero-shot evaluation of frontier LLMs as goal recognisers on key classical PDDL benchmarks. Our results show that LLM competence on goal recognition is uneven: some models scale with evidence and approach landmark-based accuracy at full observations, while others remain anchored to world-knowledge priors regardless of how much evidence accumulates. Qualitative analysis of model reasoning traces reveals that this divergence reflects a fundamental difference in evidence integration rather than domain familiarity. These findings position goal recognition as a principled benchmark for the foundational planning knowledge of LLMs.\n\n---\n\n## 39. Context Pruning for Coding Agents via Multi-Rubric Latent Reasoning\n- **Authors**: Jingjing Wang , Xiwen Chen , Wenhui Zhu , Huayu Li , Zhengxiao He , Feiyang Cai , Ana S. Carreon-Rascon , Xuanzhao Dong , Feng Luo\n- **URL**: [https://arxiv.org/abs/2605.15315](https://arxiv.org/abs/2605.15315)\n- **Abstract**:\n> LLM-powered coding agents spend the majority of their token budget reading repository files, yet much of the retrieved code is irrelevant to the task at hand. Existing learned pruners compress this context with a single-objective sequence labeler, collapsing all facets of code relevance into one score and one transition matrix. We show that this formulation creates a modeling bottleneck: a single CRF transition prior must serve heterogeneous retention patterns, including contiguous semantic spans and sparse structural support lines. We propose LaMR (Latent Multi-Rubric), a structured pruning framework that decomposes code relevance into two interpretable quality dimensions, semantic evidence and dependency support, each modeled by a dedicated CRF with dimension-specific transition dynamics. A mixture-of-experts gating network dynamically weights the per-rubric emissions conditioned on the query, and a final CRF layer on the fused emissions produces the aggregate keep-or-prune decision. To supervise each dimension without additional annotation cost, we derive multi-rubric labels from the existing training corpus via AST-based program analysis, simultaneously denoising the teacher's binary labels. By effectively filtering distracting noise, LaMR frequently matches or even outperforms unpruned full-context baselines. Experiments on four benchmarks (SWE-Bench Verified, SWE-QA, LCC, LongCodeQA) show that LaMR wins 12 of 16 head-to-head multi-turn comparisons. It saves up to 31% more tokens on multi-turn agent tasks and improves Exact Match by up to +3.5 on single-turn tasks, while performance is frequently enhanced by denoising the context, and any remaining drops are marginal.\n\n---\n\n## 40. SMCEvolve: Principled Scientific Discovery via Sequential Monte Carlo Evolution\n- **Authors**: Jiachen Jiang , Huminhao Zhu , Zhihui Zhu\n- **URL**: [https://arxiv.org/abs/2605.15308](https://arxiv.org/abs/2605.15308)\n- **Abstract**:\n> LLM-driven program evolution has emerged as a powerful tool for automated scientific discovery, yet existing frameworks offer no principled guide for designing their individual components and provide no guarantee that the search converges. We introduce SMCEvolve, which recasts program search as sampling from a reward-tilted target distribution and approximates it with a Sequential Monte Carlo (SMC) sampler. From this view, three core mechanisms emerge as principled components: adaptive parent resampling, mixture of mutation with acceptance, and automatic convergence control. We further provide a finite-sample complexity analysis that bounds the LLM-call budget required to reach a target approximation error. Across math, algorithm efficiency, symbolic regression, and end-to-end ML research benchmarks, SMCEvolve surpasses state-of-the-art evolving systems while using fewer LLM calls under self-determined termination. The code is available at this https URL .\n\n---\n\n## 41. Solvita: Enhancing Large Language Models for Competitive Programming via Agentic Evolution\n- **Authors**: Han Li , Jinyu Tian , Rili Feng , Yuqiao Du , Chong Zheng , Chenyu Wang , Chenchen Liu , Shihao Li , Xinping Lei , Yifan Yao , Weihao Xie , Letian Zhu , Jiaheng Liu\n- **URL**: [https://arxiv.org/abs/2605.15301](https://arxiv.org/abs/2605.15301)\n- **Abstract**:\n> Large language models (LLMs) still struggle with the rigorous reasoning demands of hard competitive programming. While recent multi-agent frameworks attempt to bridge this reliability gap, they remain fundamentally stateless: they rely on static retrieval and discard the valuable problem-solving and debugging experience gained from previous tasks. To address this, we present Solvita, an agentic evolution framework that enables continuous learning without requiring weight updates to the underlying LLM. Solvita reorganizes problem-solving into a closed-loop system of strategy selection, program synthesis, certified supervision, and targeted hacking, executed by four specialized agents: Planner, Solver, Oracle, and Hacker. Crucially, each agent is paired with a trainable, graph-structured knowledge network. As the system operates, outcome signals, such as pass/fail verdicts, test certification quality, and adversarial vulnerabilities discovered by the Hacker, are recast as reinforcement learning updates to these network weights. This allows the agents to dynamically route future queries based on past successes and failures, effectively accumulating transferable reasoning experience over time. Evaluated across CodeContests, APPS, AetherCode, and live Codeforces rounds, Solvita establishes a new state-of-the-art among code-generation agents, outperforming existing multi-agent pipelines and nearly doubling the accuracy of single-pass baselines.\n\n---\n\n## 42. Verifiable Agentic Infrastructure: Proof-Derived Authorization for Sovereign AI Systems\n- **Authors**: Jun He , Deying Yu\n- **URL**: [https://arxiv.org/abs/2605.15228](https://arxiv.org/abs/2605.15228)\n- **Abstract**:\n> Modern cloud and enterprise systems rely on identity-centric authorization, assuming that callers possessing valid credentials are safe to execute commands. The emergence of autonomous AI agents invalidates this assumption: agents can generate syntactically valid but semantically unsafe actions, making standing privileges a significant operational risk. This risk becomes especially acute in sovereign AI systems, where autonomous agents may interact with cloud infrastructure, regulated data, financial workflows, and national-scale digital services. Governed mutation substrates reduce this risk by interposing on agent actions: agents submit intents, infrastructure evaluates context and policy, and execution is mediated. However, this shifts the trust boundary: how can the decision to authorize an intent be made verifiable, distributed, and replayable? We introduce a Distributed Trust Framework (DTF), a verification framework for governed mutation systems that computes execution authority from structured, verifiable artifacts. DTF introduces a Justification Proof to encode the admissibility basis of an action, a consensus model for independent evaluation, an ephemeral Execution Identity derived from the approved proof, and an append-only Evidence Chain that preserves the authorization lifecycle. Under stated substrate assumptions, this architecture enforces a compact authorization invariant: no high-stakes execution without a proof object, no derived authority without consensus, and no valid mutation detached from evidence. We define the model, instantiate it over an OpenKedge-based governed mutation substrate, and show how it maps onto cloud-native environments. By shifting authorization from standing identity to proof-derived authority, DTF provides an infrastructure foundation for making agentic execution governable, auditable, and bounded in sovereign AI deployments.\n\n---\n\n## 43. NIMO Controller: a self-driving laboratory orchestrator based on the Model Context Protocol\n- **Authors**: Naruki Yoshikawa , Ryo Tamura\n- **URL**: [https://arxiv.org/abs/2605.15227](https://arxiv.org/abs/2605.15227)\n- **Abstract**:\n> Self-driving laboratories (SDLs) have attracted increasing attention as a means of accelerating scientific discovery; however, developing SDL software remains technically demanding. To improve accessibility, orchestration software frameworks have been proposed to coordinate SDL components. Nevertheless, existing frameworks are primarily designed for human interaction and do not provide standardized interfaces suitable for AI agents. In this work, we propose an SDL software architecture based on the Model Context Protocol (MCP), in which all SDL functionalities are exposed through MCP servers. Following this design principle, we introduce an MCP-based SDL orchestrator, named NIMO Controller. It provides a visual programming interface automatically generated through MCP-based tool discovery, allowing human users to design experimental workflows without writing code. The same MCP backend can also be accessed by AI agents, providing a unified interface for both human users and AI agents. We demonstrate the proposed system through a case study on a color-matching SDL. The results validate the usability of the proposed MCP-based SDL architecture.\n\n---\n\n## 44. ICRL: Learning to Internalize Self-Critique with Reinforcement Learning\n- **Authors**: Jianbo Lin , Xiaomin Yu , Yi Xin , Yifu Guo , Zhuosong Jiang , Zhongqi Yue , Weishi Wang , Heqing Zou , Chengwei Qin , Hui Xiong\n- **URL**: [https://arxiv.org/abs/2605.15224](https://arxiv.org/abs/2605.15224)\n- **Abstract**:\n> Large language model-based agents make mistakes, yet critique can often guide the same model toward correct behavior. However, when critique is removed, the model may fail again on the same query, indicating that it has not internalized the critique's guidance into its underlying capability. Meanwhile, a frozen critic cannot improve its feedback quality over time, limiting the potential for iterative self-improvement. To address this, we propose learning to internalize self-critique with reinforcement learning(ICRL), a novel framework that jointly trains a solver and a critic from a shared backbone to convert critique-induced success into unassisted solver ability. The critic is rewarded based on the solver's subsequent performance gain, incentivizing actionable feedback. To address the distribution shift between critique-conditioned and critique-free behavior, ICRL introduces a distribution-calibration re-weighting ratio that selectively transfers critique-guided improvements compatible with the solver's own prompt distribution. Additionally, a role-wise group advantage estimation stabilizes joint optimization across the two roles. Together, these mechanisms ensure that the solver learns to improve itself without external critique, rather than becoming dependent on critique-conditioned behavior. We evaluate ICRL on diverse benchmarks spanning agentic and mathematical reasoning tasks, using Qwen3-4B and Qwen3-8B as backbones. Results show consistent improvements, with average gains of 6.4 points over GRPO on agentic tasks, and 7.0 points on mathematical reasoning. Notably, the learned 8B critic is comparable to 32B critics while using substantially fewer tokens. The code is available at this https URL .\n\n---\n\n## 45. NOVA: Fundamental Limits of Knowledge Discovery Through AI\n- **Authors**: Salman Avestimehr , Ken Duffy , Muriel Médard\n- **URL**: [https://arxiv.org/abs/2605.15219](https://arxiv.org/abs/2605.15219)\n- **Abstract**:\n> Can AI systems discover genuinely new knowledge through iterative self improvement, and if so, at what cost? We introduce the NOVA framework, which models the common ``generate, verify, accumulate, retrain'' loop as an adaptive sampling process over a knowledge space. We identify sufficient conditions under which accumulated genuine knowledge eventually covers a finite domain, and show how their violations produce distinct failure modes: contamination, forgetting, exploration failure, and acceptance failure. We then analyze imperfect verification and identify a contamination trap: as easy-to-find knowledge is exhausted, the model mass assigned to new valid artifacts shrinks, so even small false-positive rates can cause invalid artifacts to enter the knowledge base faster than genuine discoveries. We clarify that Good--Turing estimation is a local batch-diversity diagnostic, not an estimator of the historically undiscovered valid mass that governs long-term discovery. Under a separate tail-equivalence assumption relating the model's effective discovery distribution to a Zipf law with exponent $\\alpha>1$, we prove that the cumulative generation cost required to obtain $D$ distinct genuine discoveries satisfies $R_{\\mathrm{cum} }(D)=\\Theta(c_{\\mathrm{gen} }D^\\alpha)$, where $c_{\\mathrm{gen} }$ is the per-candidate generation cost. This scaling law quantifies asymptotic diminishing returns as the discovery frontier advances. Finally, we formalize human amplification through guidance, generation, and verification, explaining why expert input is most valuable near autonomous exploration barriers.\n\n---\n\n## 46. CAX-Agent: A Lightweight Agent Harness for Reliable APDL Automation\n- **Authors**: Chenying Lin , Yichen Hai , Yi He , Ran Wang , Haiyan Qiang , Liang Yu\n- **URL**: [https://arxiv.org/abs/2605.15218](https://arxiv.org/abs/2605.15218)\n- **Abstract**:\n> Large language models deployed for MAPDL finite-element simulation face practical reliability challenges: without structured execution control, tool encapsulation, and fault recovery, outputs may be inconsistent and task failures are common. The Agent Harness paradigm addresses this by inserting domain-specific orchestration middleware that manages tool lifecycles, workflow state, and recovery escalation. This paper presents the architecture of CAX-Agent, a lightweight agent harness purpose-built for MAPDL automation, and empirically evaluates one of its core components -- the recovery this http URL -Agent organizes execution into three layers -- LLM service, agent harness, and solver backend -- with a recovery ladder that escalates from deterministic rule patching through model-driven regeneration to context enrichment and human intervention. We evaluate three recovery strategies (no_recovery, rule_only, and model_only) on 50 standard structural benchmarks with three repeated runs per strategy (450 case-runs total). Two independent human raters score task completion under blind conditions; inter-rater agreement is strong (quadratic weighted Cohen's kappa = 0.84, 96 percent of score pairs within one point). Model_only achieves the best completion rate (0.9267), task score (3.59/4), total score (9.16/10), and zero-intervention rate (0.84), outperforming rule_only (0.7733, 3.17/4, 7.03/10, 0.00) and no_recovery (0.6933, 2.74/4, 5.60/10, 0.00) with large effect sizes (Cliff's delta = 0.81-0.87). The benchmark uses deliberately simple geometries to isolate recovery-policy effects; we discuss the scope of these findings and directions for broader validation.\n\n---\n\n## 47. Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions\n- **Authors**: Jagdish Tripathy , Marcus Buckmann\n- **URL**: [https://arxiv.org/abs/2605.15217](https://arxiv.org/abs/2605.15217)\n- **Abstract**:\n> Instruction-tuned language models exhibit behavioural fairness in high-stakes decisions while retaining biased associations in their internal representations. However, whether these suppressed representations can affect model outputs - and whether such causal potency is symmetric across demographic groups - remains unknown. We investigate the use of open-weight models for mortgage underwriting using matched applications that differ only in racially-associated names and reveal a critical disconnect: models show no output-level bias, yet retain and amplify demographic representations across model layers. Through activation steering and novel cross-layer interventions, we demonstrate that this suppressed information is decision-relevant: when reinjected at critical layers, it produces near-complete decision reversals. Critically, this latent bias is asymmetric - steering interventions affect decisions in one demographic direction, while producing minimal effects in reverse - and susceptible to adversarial prompt engineering and parameter-efficient fine-tuning. These findings demonstrate that behavioural audits focused on outputs are insufficient: fair outputs can mask exploitable internal biases. They also motivate dual-layer testing frameworks combining output evaluation with representational analysis for AI governance in high-stakes decisions.\n\n---\n\n## 48. SkillSmith: Compiling Agent Skills into Boundary-Guided Runtime Interfaces\n- **Authors**: Duling Xu , Zheng Chen , Zaifeng Pan , Jiawei Guan , Dong Dong , Jialin Li , Bangzheng Pu\n- **URL**: [https://arxiv.org/abs/2605.15215](https://arxiv.org/abs/2605.15215)\n- **Abstract**:\n> Recently, skills have been widely adopted in large language model (LLM)-based agent systems across various domains. In existing frameworks, skills are typically injected into the agent reasoning loop as contextual guidance once matched to a runtime task, enabling specialized task-solving capabilities. We find that this execution paradigm introduces two major sources of redundancy: irrelevant context injection and repeated skill-specific reasoning and planning. To this end, we propose SkillSmith, a boundary-first compiler-runtime framework that compiles skill packages offline into minimal executable interfaces. By extracting fine-grained operational boundaries from skills, SkillSmith enables agents to dynamically access and execute only the relevant components at runtime, thereby minimizing unnecessary context injection and redundant reasoning overhead. In the evaluation on SkillsBench benchmark, SkillSmith reduces solve-stage token usage by 57.44%, thinking iterations by 42.99%, solve time by 50.57% (2.02x faster), and token-proportional monetary cost by 57.44% compared with using raw-skills. Moreover, compiled artifacts produced by a stronger model can be reused by a smaller or more efficient runtime model, improving task accuracy in cases where raw skill interpretation fails. The source code and data are available at this https URL .\n\n---\n\n## 49. Does Theory of Mind Improvement Really Benefit Human-AI Interactions? Empirical Findings from Interactive Evaluations\n- **Authors**: Nanxu Gong , Zixin Chen , Haotian Li , Zishu Zhao , Jianxun Lian , Huamin Qu , Yanjie Fu , Xing Xie\n- **URL**: [https://arxiv.org/abs/2605.15205](https://arxiv.org/abs/2605.15205)\n- **Abstract**:\n> Improving the Theory of Mind (ToM) capability of Large Language Models (LLMs) is crucial for effective social interactions between these AI models and humans. However, the existing benchmarks often measure ToM capability improvement through story-reading, multiple-choice questions from a third-person perspective, while ignoring the first-person, dynamic, and open-ended nature of human-AI (HAI) interactions. To directly examine how ToM improvement techniques benefit HAI interactions, we first proposed the new paradigm of interactive ToM evaluation with both perspective and metric shifts. Next, following the paradigm, we conducted a systematic study of four representative ToM enhancement techniques using both four real-world datasets and a user study, covering both goal-oriented tasks (e.g., coding, math) and experience-oriented tasks (e.g., counseling). Our findings reveal that improvements on static benchmarks do not always translate to better performance in dynamic HAI interactions. This paper offers critical insights into ToM evaluation, showing the necessity of interaction-based assessments in developing next-generation, socially aware LLMs for HAI symbiosis.\n\n---\n\n## 50. SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch\n- **Authors**: Zhantao Wang\n- **URL**: [https://arxiv.org/abs/2605.15204](https://arxiv.org/abs/2605.15204)\n- **Abstract**:\n> Multi-agent orchestration frameworks such as LangChain, LangGraph, and CrewAI route tasks through graph-based pipelines but do not enforce the stage constraints that govern real business processes. We present SDOF, a framework that treats multi-agent execution as a constrained state machine. SDOF operates through two primary defensive layers, implemented by three components: (1) an Online-RLHF Specialized Intent Router trained via Generative Reward Modeling (GRPO) and (2) a StateAwareDispatcher with GoalStage finite-automaton checks and precondition/postcondition SkillRegistry validation for auditable execution control. On a recruitment system backed by the Beisen iTalent platform (6000+ enterprises), 185 expert-curated scenarios trigger 1671 live API calls. Our GSPO-aligned 7B Intent Router achieves higher joint accuracy than zero-shot GPT-4o on this FSM-constrained adversarial routing benchmark (80.9% versus 48.9%). In end-to-end execution, SDOF reaches 86.5% task completion (95% confidence interval 80.8 to 90.7) and blocks all 22 operations in the injection, illegal HR subset. Under a broader message-level blocking audit, SDOF attains precision 100% and recall 88%, expert agreement kappa=0.94. A separate evaluation on 960 SGD-derived dialogues spanning 8 service domains surfaces 201 stage-order conflicts under our FSM mapping, 41 of which arise in the normal split. This arXiv version reports the current validated scope; extended multi-seed training comparisons and deeper workflow evaluations will be released in a subsequent update.\n\n---\n\n## 51. DeepSlide: From Artifacts to Presentation Delivery\n- **Authors**: Ming Yang , Zhiwei Zhang , Jiahang Li , Haoseng Liu , Yuzheng Cai , Weiguo Zheng\n- **URL**: [https://arxiv.org/abs/2605.15202](https://arxiv.org/abs/2605.15202)\n- **Abstract**:\n> Presentations are a primary medium for scholarly communication, yet most AI slide generators optimize the artifact (a visually plausible deck) while under-optimizing the delivery process (pacing, narrative, and presentation preparation). We present DeepSlide, a human-in-the-loop multi-agent system that supports preparing the full presentation process, from requirement elicitation and time-budgeted narrative planning, to evidence-grounded slide--script generation, attention augmentation, and rehearsal support. DeepSlide integrates (i) a controllable logical-chain planner with per-node time budgets, (ii) a lightweight content-tree retriever for grounding, (iii) Markov-style sequential rendering with style inheritance, and (iv) sandboxed execution with minimal repair to ensure renderability. We further introduce a dual-scoreboard benchmark that cleanly separates static artifact quality from dynamic delivery excellence. Across 20 domains and diverse audience profiles, DeepSlide matches strong baselines on artifact quality while consistently achieving larger gains on delivery metrics, improving narrative flow, pacing precision, and slide--script synergy with clearer attention guidance.\n\n---\n\n## 52. IVGT: Implicit Visual Geometry Transformer for Neural Scene Representation\n- **Authors**: Yuqi Wu , Tianyu Hu , Wenzhao Zheng , Yuanhui Huang , Haowen Sun , Jie Zhou , Jiwen Lu\n- **URL**: [https://arxiv.org/abs/2605.16258](https://arxiv.org/abs/2605.16258)\n- **Abstract**:\n> Reconstructing coherent 3D geometry and appearance from unposed multi-view images is a fundamental yet challenging problem in computer vision. Most existing visual geometry foundation models predict explicit geometry by regressing pixel-aligned pointmaps, often suffering from redundancy and limited geometric continuity. We propose IVGT, an Implicit Visual Geometry Transformer that implicitly models continuous and coherent geometry from pose-free multi-view images. This formulation learns a continuous neural scene representation in a canonical coordinate system and supports continuous spatial queries at any 3D positions, retrieving local features to predict signed distance (SDF) values and colors using lightweight decoders. It allows direct extraction of continuous and coherent surface geometry, enabling rendering of RGB images, depth maps, and surface normal maps from arbitrary viewpoints. We train IVGT via multi-dataset joint optimization with 2D supervision and 3D geometric regularization. IVGT demonstrates generalization across scenes and achieves strong performance on various tasks, including mesh and point cloud reconstruction, novel view synthesis, depth and surface normal estimation, and camera pose estimation.\n\n---\n\n## 53. Designing Datacenter Power Delivery Hierarchies for the AI Era\n- **Authors**: Grant Wilkins , Fiodar Kazhamiaka , Alok Gautam Kumbhare , Chaojie Zhang , Ricardo Bianchini\n- **URL**: [https://arxiv.org/abs/2605.16255](https://arxiv.org/abs/2605.16255)\n- **Abstract**:\n> Demand for AI accelerators is rapidly increasing rack power density, with projections approaching 1MW per deployment by 2027. This poses a major challenge for datacenter power delivery designers. As power densities increase, a datacenter designed for a different target density may strand power, i.e., may be unable to use all the power that its delivery hierarchy has provisioned. Designs must remain efficient over long datacenter lifetimes and multiple hardware generations. Power utilization is particularly important as grid power capacity is a scarce resource in the AI era. Designing an efficient power delivery hierarchy for the long run is difficult because rack placement feasibility, workload impact, and cost depend jointly on electrical topology, deployment granularity, placement policy, power oversubscription, and workload mix. Moreover, each of these factors evolve over time, have inter-dependencies across multiple resource dimensions, and generally do not lend themselves to closed-form analysis. To address this challenge, we develop a framework for evaluating datacenter power delivery designs using throughput, power, and cost metrics over realistic arrival, oversubscription, and decommissioning sequences. The framework combines projection models for GPU, compute, and storage deployments with operational factors grounded in production data from Microsoft Azure. Our results show that multi-resource stranding materially changes deployable capacity, effective capital expenditure, and delivered performance, and quantify how rising density from rack- and pod-scale AI systems shapes these outcomes. For AI datacenter design, the relevant planning objective is not installed megawatts, but deployable capacity over time.\n\n---\n\n## 54. A Generative AI Framework for Intelligent Utility Billing CO 2 Analytics and Sustainable Resource Optimisation\n- **Authors**: Pavan Manjunath , Thomas Pruefer\n- **URL**: [https://arxiv.org/abs/2605.16250](https://arxiv.org/abs/2605.16250)\n- **Abstract**:\n> Distribution utilities are now expected to deliver bills that customers can actually read attach a defensible carbon number to every kWh sold and schedule load against grid stress and emissions constraints We propose an end-to-end framework that unifies four production-grade capabilities under one architectural roof a generative-AI agent that drafts each customers natural-language billing statement from structured numeric inputs under a constrained decoding policy a transformer-based forecaster that supplies the day-ahead consumption estimate with calibrated quantile bands\n\n---\n\n## 55. AI-Mediated Communication Can Steer Collective Opinion\n- **Authors**: Stratis Tsirtsis , Kai Rawal , Chris Russell , Brent Mittelstadt , Sandra Wachter\n- **URL**: [https://arxiv.org/abs/2605.16245](https://arxiv.org/abs/2605.16245)\n- **Abstract**:\n> Generative artificial intelligence (AI) is increasingly integrated into the online platforms where humans exchange opinions; large language models (LLMs) now polish users' posts on LinkedIn and provide context for content shared on X. While prior work has shown that AI can express biased opinions and shape individuals' opinions during human-AI interactions, less attention has been paid to its influence on collective opinion formation when mediating human-to-human communication. We address this gap via a combination of empirical and theoretical analyses. We show empirically that LLMs from multiple popular families introduce directional biases when instructed to edit human-written texts on contested topics, for example, nudging texts in favor of gun control and against atheism. Building on this observation, we introduce a mathematical model of opinion dynamics in which an AI system sits between users on a social network, transforming the opinions they express and perceive. By analytically characterizing the equilibrium of this model and performing simulations on real social network data, we show that biases introduced by AI in human-to-human communication can be amplified through the network and shift collective opinion in their direction. In light of these findings, we investigate whether such biases are controllable by online platforms. We audit the \"Explain this post\" feature on X and find evidence of pro-life bias in Grok's outputs on abortion-related content, which we trace back to specific design choices. We conclude with a discussion of the broader implications of our findings in relation to ongoing legislative efforts in the European Union.\n\n---\n\n## 56. Offline Semantic Guidance for Efficient Vision-Language-Action Policy Distillation\n- **Authors**: Jin Shi , Brady Zhang , Yishun Lu\n- **URL**: [https://arxiv.org/abs/2605.16241](https://arxiv.org/abs/2605.16241)\n- **Abstract**:\n> Billion-parameter Vision-Language-Action (VLA) policies have recently shown impressive performance in robotic manipulation, yet their size and inference cost remain major obstacles for real-time closed-loop control. We introduce \\textbf{VLA-AD}, a distillation framework that uses a Vision-Language Model as an offline semantic supervisor to transfer large VLA teachers into lightweight student policies. Instead of relying only on low-level action imitation, VLA-AD augments teacher-provided 7-DoF action targets with high-level semantic guidance, including task phase anchors and multi-frame operating-direction descriptions. These auxiliary signals are used only during training: at test time, the student policy runs independently, with neither the VLA teacher nor the VLM required. We evaluate VLA-AD on three LIBERO benchmark suites. Using OpenVLA-7B as the teacher, our method produces a 158M-parameter student, yielding a $44\\times$ reduction in model size while matching the teacher with only a $0.27\\%$ average relative gap. The resulting policy runs at 12.5 Hz on an RTX 4090, achieving a $3.28\\times$ inference speedup over OpenVLA-7B. We further show that the same semantic distillation pipeline generalizes to a different $\\pi_{0.5}$-4B teacher, where the student outperforms the teacher on two suites and remains within $0.53\\%$ on \\texttt{libero\\_goal}. Additional analysis indicates that phase-level supervision and multi-frame directional cues make the student less sensitive to noisy teacher actions, such as erroneous high-frequency gripper changes. Overall, VLA-AD demonstrates that offline semantic guidance from VLMs can substantially improve the efficiency, robustness, and deployability of VLA policy distillation.\n\n---\n\n## 57. Layer Equivalence Is Not a Property of Layers Alone: How You Test Redundancy Changes What You Find\n- **Authors**: Gabriel Garcia\n- **URL**: [https://arxiv.org/abs/2605.16234](https://arxiv.org/abs/2605.16234)\n- **Abstract**:\n> When researchers ask whether two transformer layers are \"equivalent\" for compression, they often conflate distinct tests. Replacement asks whether one layer's map can substitute for another's in place; interchange asks whether two layers approximately commute when their positions are swapped. Both are output-grounded swap-KL probes, but they need not agree: on pretrained transformers the protocol gap can change which layers look safe to prune by several-fold under the same evaluator, especially when replacement distances are high. We measure both protocols across checkpoints and architectures. On a Pythia training trajectory (410M and 1.4B), the replacement-interchange gap grows from initialization to convergence. Under one matched WikiText-2 contract at 8B scale, Qwen3-8B enters a divergent regime: interchange-guided removal is several-fold safer than replacement-guided at the same layer budgets, while Llama-3.1-8B ties the two protocols for pruning cost even though interchange KL is lower, showing metric gaps need not map one-to-one to removal. Before layer removal or merging, score both swap-KLs on the target checkpoint; the diagnostic requires only unlabeled forward passes.\n\n---\n\n## 58. A Unified Generative-AI Framework for Smart Energy Infrastructure: Intelligent Gas Distribution, Utility Billing, Carbon Analytics, and Quantum-Inspired Optimisation\n- **Authors**: Pavan Manjunath , Thomas pruefer\n- **URL**: [https://arxiv.org/abs/2605.16232](https://arxiv.org/abs/2605.16232)\n- **Abstract**:\n> The accelerating convergence of smart metering, generative artificial intelligence, and quantum-inspired combinatorial optimisation is reshaping how energy utilities manage physical infrastructure, customer engagement, and environmental accountability\n\n---\n\n## 59. Evaluating Design Video Generation: Metrics for Compositional Fidelity\n- **Authors**: Adrienne Deganutti , Dingning Cao , Jaejung Seol , Elad Hirsch , Purvanshi Mehta\n- **URL**: [https://arxiv.org/abs/2605.16223](https://arxiv.org/abs/2605.16223)\n- **Abstract**:\n> Generative video models are increasingly used in design animation tasks, yet no standardized evaluation framework exists for this domain. Unlike natural video generation, design animation imposes structured constraints: specific components shall animate with prescribed motion types, directions, speed and timing, while non-animated regions must remain stable and layout structure must be preserved. This paper provides a fully automated evaluation framework organized across four dimensions: layout fidelity, motion correctness, temporal quality, and content fidelity. This eliminates the reliance on subjective human evaluation and establishes a common basis for benchmarking progress in the field.\n\n---\n\n## 60. Argus: Evidence Assembly for Scalable Deep Research Agents\n- **Authors**: Zhen Zhang , Liangcai Su , Zhuo Chen , Xiang Lin , Haotian Xu , Simon Shaolei Du , Kaiyu Yang , Bo An , Lidong Bing , Xinyu Wang\n- **URL**: [https://arxiv.org/abs/2605.16217](https://arxiv.org/abs/2605.16217)\n- **Abstract**:\n> Deep research agents have achieved remarkable progress on complex information seeking tasks. Even long ReAct style rollouts explore only a single trajectory, while recent state of the art systems scale inference time compute via parallel search and aggregation. Yet deep research answers are composed of complementary pieces of evidence, which parallel rollouts often duplicate rather than complete, yielding diminishing returns while pushing the aggregation context toward the model's limit. We propose Argus, an agentic system in which a Searcher and a Navigator cooperate to treat deep research as assembling a jigsaw from complementary evidence pieces, rather than brute forcing the whole answer in parallel. The Searcher collects evidence traces for a given sub-query through ReAct-style interaction. The Navigator maintains a shared evidence graph, verifying which pieces are still missing, dispatching Searchers to gather them, and reasoning over the completed graph to produce a source-traced final answer. We train the Navigator with reinforcement learning to verify, dispatch, and synthesize, while independently training the Searcher to remain a standard ReAct agent. The resulting Navigator supports rollouts with a single Searcher or many in parallel without retraining. With both Searcher and Navigator built on a 35B-A3B MoE backbone, Argus gains 5.5 points with a single Searcher and 12.7 points with 8 parallel Searchers, averaged over eight benchmarks. With 64 Searchers it reaches 86.2 on BrowseComp, surpassing every proprietary agent we benchmark, while the Navigator's reasoning context stays under 21.5K tokens.\n\n---\n\n## 61. paper.json: A Coordination Convention for LLM-Agent-Actionable Papers\n- **Authors**: Arquimedes Canedo\n- **URL**: [https://arxiv.org/abs/2605.16194](https://arxiv.org/abs/2605.16194)\n- **Abstract**:\n> LLM agents routinely serve as first (and sometimes only) readers of academic papers, skimming for sub-claims, extracting reproducibility steps, and generalizing scope. Standard prose papers produce recurring failures in this role: sub-claims that cannot be cited at sub-paper granularity, scope overextension beyond what the paper tests, and figure commands buried in codebases rather than the paper itself. We propose ` this http URL `, a companion JSON file that travels with the PDF and addresses each failure with a lightweight convention: stable claim IDs (C1), an explicit does-not-claim list (C2), exact per-figure shell commands (C3), and stable definition IDs (C5). A fifth convention (C4) holds that minimum viable compliance, hand-written JSON alongside the PDF, is achievable in under an hour for a finished paper without touching the human-readable output. C1, C2, C3, and C5 are open invitations: an agent that reads a compliant paper and acts on it produces evidence for or against them. This paper is itself compliant: `uv run this http URL this http URL --against this http URL ` passes. Repo: this https URL\n\n---\n\n## 62. Second-Order Multi-Level Variance Correction for Modality Competition in Multimodal Models\n- **Authors**: Yishun Lu , Wes Armour\n- **URL**: [https://arxiv.org/abs/2605.16165](https://arxiv.org/abs/2605.16165)\n- **Abstract**:\n> Autoregressive next-token training offers a unified formulation for image generation and text understanding, but it also creates strong modality competition that destabilizes optimization and limits large-batch scaling. We show that first-order optimizers such as AdamW are vulnerable to cross-modality gradient heterogeneity, while second-order preconditioning, particularly SOAP, provides a more stable basis for multimodal alignment. Building on this insight, we propose \\emph{ML-FOP-SOAP}, a second-order optimization framework with Multi-Level Variance Correction. Our Fisher-Orthogonal Projection suppresses variance-induced modality conflicts, reducing the trade-off between visual generation and textual understanding. To make this practical under large gradient accumulation, we introduce a hierarchical folding strategy that captures fine-grained variance with low micro-step overhead. Experiments on Janus and Emu3 show consistent gains across both modalities and stable training at batch size 8192. Compared with AdamW, our method improves sample efficiency by up to $1.4\\times$ and accelerates wall-clock training by up to $1.5\\times$, offering a robust optimizer for scaling multimodal foundation models.\n\n---\n\n## 63. Surrogate Neural Architecture Codesign Package (SNAC-Pack)\n- **Authors**: Jason Weitz , Dmitri Demler , Benjamin Hawks , Aaron Wang , Nhan Tran , Javier Duarte\n- **URL**: [https://arxiv.org/abs/2605.16138](https://arxiv.org/abs/2605.16138)\n- **Abstract**:\n> Neural architecture search (NAS) is a powerful approach for automating model design, but existing methods often optimize for accuracy alone or rely on proxy metrics such as bit operations (BOPs) that correlate poorly with hardware cost. This gap is particularly large for FPGA deployment, where cost is dominated by a multi-dimensional budget of lookup tables, DSPs, flip-flops, BRAM, and latency. We present the Surrogate Neural Architecture Codesign Package (SNAC-Pack), an open-source AutoML framework for hardware-aware neural architecture codesign and end-to-end FPGA deployment. SNAC-Pack runs a multi-objective global search with Optuna and NSGA-II, loading trials to a shared SQLite store that enables parallel workers across compute nodes. A hardware surrogate model outputs per-trial resource and latency estimates, avoiding the synthesis cost that would otherwise dominate the search loop. A local search stage then applies quantization-aware training (QAT) together with iterative magnitude pruning in a combined compression loop, after which the final model is synthesized to FPGA firmware via the hls4ml Python library. A YAML configuration and an optional agentic frontend let users run the pipeline on new datasets without modifying the framework. We demonstrate SNAC-Pack on jet classification at the Large Hadron Collider and superconducting qubit readout, discovering compact architectures that match or exceed strong baselines on the task metric while reducing FPGA resource utilization and, in the qubit readout case, reducing the design space exploration process from months of manual fine-tuning to hours of automated search.\n\n---\n\n## 64. Navigating Potholes with Geometry-Aware Sharpness Minimization\n- **Authors**: Simon Dufort-Labbé , Mehrab Hamidi , Razvan Pascanu , Ioannis Mitliagkas , Damien Scieur , Aristide Baratin\n- **URL**: [https://arxiv.org/abs/2605.16134](https://arxiv.org/abs/2605.16134)\n- **Abstract**:\n> Sharpness-aware minimization (SAM) encourages flat minima by perturbing parameters along directions of high loss curvature, but treats all parameter directions uniformly, ignoring the underlying loss geometry. We introduce LLQR+SAM, which combines SAM with a learned preconditioner obtained from the recently proposed LLQR framework, a second-order method that recasts steepest descent as a layerwise linear-quadratic regulator problem. The preconditioner is updated sparsely and maintained as a slow exponential moving average, so it captures a smoothed, low-resolution picture of the loss landscape geometry. The SAM perturbation then operates on top of this learned geometry, probing curvature at a faster timescale. We show that this two-timescale structure is not merely a computational convenience: theoretically, the preconditioner amplifies the SAM escape signal in directions that are flat under the average geometry but locally sharp (potholes). Wide, flat basins, by contrast, remain stable. Empirically, LLQR+SAM gives consistent gains over both SAM and LLQR alone across standard vision and sequence modeling benchmarks, supporting the view that slow learned geometry and fast sharpness correction are genuinely complementary.\n\n---\n\n## 65. Entropy Across the Bridge: Conditional-Marginal Discretization for Flow and Schrödinger Samplers\n- **Authors**: Bruno Trentini , Dejan Stancevic , Michael M. Bronstein , Alexander Tong , Luca Ambrogioni\n- **URL**: [https://arxiv.org/abs/2605.16126](https://arxiv.org/abs/2605.16126)\n- **Abstract**:\n> For a fixed flow-based generative model under a small inference budget, sample quality can depend strongly on where the sampler spends its few function evaluations. Flow matching and Schrödinger bridges define probability paths, yet their inference grids are usually heuristic or inherited from one-endpoint diffusion. We derive a conditional-marginal entropy-rate objective for bridge-aware discretization, separating endpoint-conditioned bridge geometry from marginal flow evolution, and use it to build a training-free entropic inference-time scheduler from first principles. For Gaussian Brownian bridges this rate is closed-form and U-shaped, motivating boundary-heavy nonuniform grids. On trained two-dimensional bridge/flow models, the estimated profile recovers the predicted shape and improves 10-step ODE-Heun MMD over linear by 18.1%, with a paired 22.7% SDE-Heun improvement in the same low-NFE sweep. On EDM/CIFAR-10, the entropic time-discretization gives the best tested five-step FID (186.3 \\pm 4.0 versus 200.5 \\pm 2.9 for linear and 238.0 \\pm 5.3 for cosine). On AlphaFlow protein generation, entropic conditional-marginal (cond-marg) scheduling shows advantage in low-NFE regimes on both CAMEO22 and ATLAS benchmarks. These results support entropy-rate scheduling as a practical low-budget allocation signal for high-dimensional bridge and flow samplers.\n\n---\n\n## 66. GenShield: Unified Detection and Artifact Correction for AI-Generated Images\n- **Authors**: Zhipei Xu , Xuanyu Zhang , Youmin Xu , Qing Huang , Shen Chen , Taiping Yao , Shouhong Ding , Jian Zhang\n- **URL**: [https://arxiv.org/abs/2605.16122](https://arxiv.org/abs/2605.16122)\n- **Abstract**:\n> Diffusion-based image synthesis has made AI-generated images (AIGI) increasingly photorealistic, raising urgent concerns about authenticity in applications such as misinformation detection, digital forensics, and content moderation. Despite the substantial advances in AIGI detection, how to correct detected AI-generated images with visible artifacts and restore realistic appearance remains largely underexplored. Moreover, few existing work has established the connection between AIGI detection and artifact correction. To fill this gap, we propose GenShield, a unified autoregressive framework that jointly performs explainable AIGI detection and controllable artifact correction in a closed loop from diagnosis to restoration, revealing a mutually reinforcing relationship between these two tasks. We further introduce a Visual Chain-of-Thought based curriculum learning strategy that enables self-explained, multi-step ``diagnose-then-repair'' correction with an explicit stopping criterion. A high-quality dataset with large-scale ``artifact-restored'' pairs is also constructed alongside a unified evaluation pipeline. Extensive experiments on our correction benchmark and mainstream AIGI detection benchmarks demonstrate state-of-the-art performance and strong generalization of our method. The code is available at this https URL .\n\n---\n\n## 67. DebiasRAG: A Tuning-Free Path to Fair Generation in Large Language Models through Retrieval-Augmented Generation\n- **Authors**: Rui Chu , Bingyin Zhao , Thanh Quoc Hung Le , Duy Cao Hoang , Huawei Lin , Ping Li , Weijie Zhao , Khoa D Doan , Yingjie Lao\n- **URL**: [https://arxiv.org/abs/2605.16113](https://arxiv.org/abs/2605.16113)\n- **Abstract**:\n> Large language models (LLMs) have achieved unprecedented success due to their exceptional generative capabilities. However, because they depend on knowledge encapsulated from training corpora, they may produce hallucinations, stereotypes, and socially biased content. In particular, LLMs are prone to prejudiced responses involving race, gender, and age, which are collectively referred to as social biases. Prior studies have used fine-tuning and prompt engineering to mitigate such biases in LLMs, but these methods require additional training resources or domain knowledge to design the framework. Moreover, they may degrade the original capabilities of LLMs and often overlook the need for dynamic debiasing contexts for fairer inference. In this paper, we propose DebiasRAG, a novel tuning-free and dynamic query-specific debiasing framework based on retrieval-augmented generation (RAG). DebiasRAG improves fairness while preserving the intrinsic properties of LLMs, such as representation ability. DebiasRAG consists of three stages: (1) query-specific debiasing candidate generation; (2) context candidate pool construction; and (3) gradient-updated debiasing-guided context piece reranking. First, DebiasRAG leverages self-diagnosed bias contexts relevant to the query through regular retrieval, where the bias contexts are prepared offline by the DebiasRAG provider. Given the query-specific bias contexts, DebiasRAG reversely produces debiasing contexts, which are provided as additional fairness constraints for LLM outputs. Second, a regular RAG retrieval process produces query-related contexts from the regular RAG document database, such as a chunked Wikipedia dataset.\n\n---\n\n## 68. Attention Dispersion in Dynamic Graph Transformers: Diagnosis and a Transferable Fix\n- **Authors**: Jinhao Zhang , Kangfei Zhao , Qiuhao Zeng , Long-Kai Huang\n- **URL**: [https://arxiv.org/abs/2605.16112](https://arxiv.org/abs/2605.16112)\n- **Abstract**:\n> Transformer-based architectures have become the dominant paradigm for Continuous-Time Dynamic Graph (CTDG) learning, yet their performance remains limited on temporally shifted datasets. In this work, we identify attention dispersion as a shared failure mode of dynamic graph Transformers under temporal distribution shift. Through controlled ablation contrasting structurally and temporally distinguished historical neighbors against random ones, we show that prediction depends on a class of critical nodes that carry consistently more predictive signal than arbitrary neighbors. However, existing Transformers fail to focus on these nodes even when they are present in the input, as temporal shift weakens attention contrast and produces overly dispersed attention distributions. This diagnosis suggests a simple and transferable fix: replace standard attention with differential attention, which suppresses common-mode attention and amplifies distinctive token-level signals. When added to three representative CTDG Transformer baselines, differential attention consistently improves performance, with gains concentrated on high-shift datasets. Attention-level measurements further confirm the mechanism, showing reduced attention entropy and increased attention mass on critical nodes. Building on these findings, we introduce DiffDyG, a reference implementation combining differential attention with standard input encodings. Across 9 benchmarks and three negative sampling protocols, DiffDyG achieves SOTA performance, with especially large gains on the most shifted datasets.\n\n---\n\n## 69. Federated Imputation under Heterogeneous Feature Spaces\n- **Authors**: Imane Hocine , Chaimaa Medjadji , Sylvain Kubler , Gregoire Danoy , Yves Le Traon\n- **URL**: [https://arxiv.org/abs/2605.16099](https://arxiv.org/abs/2605.16099)\n- **Abstract**:\n> Federated Learning (FL) enables collaborative training across decentralized clients, but most methods assume aligned feature schemas, an assumption that rarely holds in tabular settings where clients observe only partially overlapping feature subsets. In these heterogeneous feature spaces, parameter-averaging methods (e.g., FedAvg) transfer little information across weakly overlapping or disjoint feature groups, limiting their effectiveness for federated imputation. To overcome this, we propose \\textbf{FedHF-Impute}, a federated imputation framework that separates structural feature unavailability from conventional missingness and uses a shared global feature graph to propagate information across statistically related features through message passing. This enables indirect cross-client knowledge transfer, even when features are never jointly observed locally, while preserving standard federated communication. Under simulated partial schema overlap on the SECOM and AirQuality datasets, FedHF-Impute improves imputation accuracy (RMSE) over FL baselines by 26.9\\%, and 8.4\\% respectively, while achieving comparable performance on PhysioNET, with only a 0.3\\% difference relative to the best baseline.\n\n---\n\n## 70. GeoGS-CE: Learning Delay--Beam Channel Priors with 3D Gaussians for High-Mobility Scenarios\n- **Authors**: Yumeng Zhang , Jiajia Guo , Chaozheng Wen , Chenghong Bian , Jun Zhang\n- **URL**: [https://arxiv.org/abs/2605.16094](https://arxiv.org/abs/2605.16094)\n- **Abstract**:\n> Wideband channel estimation (CE) in high-mobility scenarios remains challenging because channel responses vary rapidly, while practical systems can allocate only sparse pilots to accommodate dense users. Fortunately, many high-mobility environments, such as high-speed railways, exhibit scheduled trajectories, predictable velocities, and a limited number of dominant propagation paths. These properties induce a delay--beam power spectrum that is more stable than the instantaneous complex channel frequency response (CFR), less sensitive to the random phase coherence, and rich in geometric information. To exploit such environmental properties, we propose GeoGS-CE, a two-stage channel estimation framework for sparse-pilot high-mobility scenarios. In the offline stage, GeoGS-CE jointly models: 1) a scene-level 3D Gaussian representation that captures the non-line-of-sight (NLoS) geometric scattering support, and 2) a leakage-aware differentiable wireless rendering process that maps the NLoS Gaussians, together with an explicit virtual line-of-sight (LoS) component, to the measured delay--beam power spectrum, while accounting for practical OFDM delay and array leakage effects. In the online stage, the delay--beam power spectrum is predicted for each user location and used as a strong covariance prior, enabling accurate full-band and full-array CFR reconstruction and tracking through a linear MMSE estimator. Simulations based on channels generated from a segment of the Guangshen high-speed railway show that the proposed geometric prior substantially improves CFR reconstruction over pilot-only and non-geometric baselines.\n\n---\n\n## 71. Centralized vs Decentralized Federated Learning: A trade-off performance analysis\n- **Authors**: Chaimaa Medjadji , Guilain Leduc , Sylvain Kubler , Yves Le Traon\n- **URL**: [https://arxiv.org/abs/2605.16089](https://arxiv.org/abs/2605.16089)\n- **Abstract**:\n> Federated Learning (FL) has emerged as a promising paradigm for collaborative model training across distributed edge devices while preserving data privacy especially with the huge increase amount of data due to the adoption of technologies which contributes to the growing number of IoT devices. Storing this amount of data centrally is challenging due to issues like limited communication, privacy, and regulations. FL can be Centralized (CFL), Decentralized (DFL), and Semi-decentralized (SDFL). Choosing the right FL architecture depends on the application's needs. However, very few research studies have experimentally compared these three types of architectures to not only understand the respective strengths and limitations, but also trade-offs between different performance indicators. This paper overcome this lack of analysis, conducting experimental analyses using the Fedstellar simulator, MNIST dataset, and MLP classifier.\n\n---\n\n## 72. Multi-level Self-supervised Pretraining on Compositional Hierarchical Graph for Molecular Property Prediction\n- **Authors**: Xiayu Liu , Zhengyi Lu , Hou-biao Li\n- **URL**: [https://arxiv.org/abs/2605.16088](https://arxiv.org/abs/2605.16088)\n- **Abstract**:\n> Self-supervised pretraining on molecular graphs has emerged as a promising approach for molecular property prediction, yet most existing methods operate at a single structural granularity and treat bond information as auxiliary edge attributes rather than as an independent semantic layer. In this work, we propose MolCHG, a multi-level self-supervised pretraining framework built upon a novel Compositional Hierarchical Graph that organizes molecular structure into four types of nodes across three semantic levels. By introducing a bond graph that operates in parallel with the atom graph, our architecture elevates bond-level information to independently evolving node representations, enabling fragment nodes to aggregate atom-level and bond-level semantics on an equal footing. We design three level-specific pretraining objectives: an atom-bond cross-view contrastive task that aligns the atom-view and bond-view representations within each fragment, a fragment-level functional group prediction task to inject domain-relevant chemical knowledge, and graph-level structure prediction tasks to encode global molecular topology. Experiments on nine MoleculeNet benchmarks demonstrate that MolCHG achieves the best performance on seven datasets across both classification and regression tasks, remaining competitive with the strongest baselines on the rest. Ablation studies further confirm that the multi-level supervision signals are complementary and that each component contributes to the overall performance.\n\n---\n\n## 73. Towards Trustworthy and Explainable AI for Perception Models: From Concept to Prototype Vehicle Deployment\n- **Authors**: Till Beemelmanns , Shayan Sharifi , Manas Mehrotra , Ayushman Choudhuri , Lutz Eckstein\n- **URL**: [https://arxiv.org/abs/2605.16087](https://arxiv.org/abs/2605.16087)\n- **Abstract**:\n> Deep Neural Networks have become the dominant solution for Autonomous Driving perception, but their opacity conflicts with emerging Trustworthy AI guidelines and complicates safety assurance, debugging, and human oversight. While theoretical frameworks for safe and Explainable AI (XAI) exist, concrete implementations of Trustworthy AI for 3D scene understanding remain scarce. We address this gap by proposing a Trustworthy AI perception module that is remarkably robust, integrates faithful explainability, and calibrated uncertainty estimates. Building on a transformer-based detector, we derive explanation from the attention mechanism at inference time and validate their faithfulness using perturbation-based consistency tests. We further integrate an uncertainty estimation and calibration module, and apply robustness-enhancing training methods. Experiments show faithful saliency behavior, improved robustness, and well-calibrated uncertainty estimates. Finally, we deploy these Trustworthy AI elements in a prototype vehicle and provide an XAI Interface that visualizes documentation artifacts, model uncertainty state, and saliency maps, demonstrating the feasibility of trustworthy perception monitoring in real time. Supplementary materials are available at this https URL .\n\n---\n\n## 74. Towards Foundation Models for Relational Databases with Language Models and Graph Neural Networks\n- **Authors**: Jingcheng Wu , Ratan Bahadur Thapa , Mojtaba Nayyeri , Lucas Etteldorf , Max Finkenbeiner , Fabian Leeske , Steffen Staab\n- **URL**: [https://arxiv.org/abs/2605.16085](https://arxiv.org/abs/2605.16085)\n- **Abstract**:\n> Relational databases store much of the world's structured information, and they are essential for driving complex predictive applications. However, deep learning progress on relational data remains limited, as conventional approaches flatten databases into single tables via manual feature engineering, discarding relational context. Relational deep learning (RDL) addresses this by modeling databases as relational entity graphs (REGs) for graph neural networks (GNNs), but remains task- and database-specific. To combine the strengths of both paradigms, we propose a hybrid architecture combining a fine-tuned BART encoder to capture intra-row semantics with a GraphSAGE-based GNN over REGs to inject relational context. Experiments on RelBench show that the GNN substantially enriches BART's row embeddings, achieving a ROC-AUC of 67.40 on the driver-dnf task from the rel-f1 dataset. This performance is competitive with supervised baselines such as LightGBM (68.86) and narrows the gap to RDL (72.62) to within 5.22 points, though a substantial gap remains to state-of-the-art foundation models such as KumoRFM (82.63). These results suggest that lightweight hybrid LM-GNN architectures offer a promising and resource-efficient path towards foundation models for relational databases.\n\n---\n\n## 75. VideoSeeker: Incentivizing Instance-level Video Understanding via Native Agentic Tool Invocation\n- **Authors**: Yiming Zhao , Yu Zeng , Wenxuan Huang , Zhen Fang , Qing Miao , Qisheng Su , Jiawei Zhao , Jiayin Cai , Lin Chen , Zehui Chen , Yukun Qi , Yao Hu , Xiaolong Jiang , Feng Zhao\n- **URL**: [https://arxiv.org/abs/2605.16079](https://arxiv.org/abs/2605.16079)\n- **Abstract**:\n> Large Vision-Language Models (LVLMs) have shown significant progress in video understanding, yet they face substantial challenges in tasks requiring precise spatiotemporal localization at the instance level. Existing methods primarily rely on text prompts for human-model interaction, but these prompts struggle to provide precise spatial and temporal references, resulting in poor user experience. Furthermore, current approaches typically decouple visual perception from language reasoning, centering reasoning around language rather than visual content, which limits the model's ability to proactively perceive fine-grained visual evidence. To address these challenges, we propose VideoSeeker, a novel paradigm for instance-level video understanding through visual prompts. VideoSeeker seamlessly integrates agentic reasoning with instance-level video understanding tasks, enabling the model to proactively perceive and retrieve relevant video segments on demand. We construct a four-stage fully automated data synthesis pipeline to efficiently generate large-scale, high-quality instance-level video data. We internalize tool-calling and proactive perception capabilities into the model via cold-start supervision and RL training, building a powerful video understanding model. Experiments demonstrate that our model achieves an average improvement of +13.7% over baselines on instance-level video understanding tasks, surpassing powerful closed-source models such as GPT-4o and Gemini-2.5-Pro, while also showing effective transferability on general video understanding benchmarks. The relevant datasets and code will be released publicly.\n\n---\n\n## 76. AgriMind: An Ensemble Deep Learning Framework for Multi-Class Plant Disease Classification\n- **Authors**: Salma Hoque Talukdar Koli , Fahima Haque Talukder Jely\n- **URL**: [https://arxiv.org/abs/2605.16076](https://arxiv.org/abs/2605.16076)\n- **Abstract**:\n> Plant disease detection is still largely manual in Bangladesh, where extension workers eyeball leaf samples across millions of smallholdings. We built AgriMind to automate this: an ensemble of ResNet50, EfficientNet-B0, and DenseNet121 trained on 20,638 PlantVillage images across 15 pepper, potato, and tomato disease classes. Transfer learning with frozen ImageNet backbones and 10 epochs of head-only training keeps the pipeline lightweight. Individual models hit 96--97% on the held-out test set, but averaging their softmax outputs pushes the ensemble to 99.23% -- a two-thirds cut in error rate. We tried biasing the average toward the best validation model; it backfired. Dropping any single model also hurt. Pepper and potato classify perfectly; tomato, with ten visually similar classes, still reaches 99.01%. On an NVIDIA T4 GPU the full ensemble runs at 53 FPS. Whether that translates to real-time mobile use depends on TensorFlow Lite optimization -- work we have not yet completed.\n\n---\n\n## 77. Robust Prior-Guided Segmentation for Editable 3D Gaussian Splatting\n- **Authors**: Raushan Joshi , Jean-Yves Guillemaut\n- **URL**: [https://arxiv.org/abs/2605.16065](https://arxiv.org/abs/2605.16065)\n- **Abstract**:\n> 3D Gaussian Splatting (3D-GS) enables real-time 3D scene reconstruction but lacks robust segmentation for editing tasks such as object removal, extraction, and recoloring. Existing approaches that lift 2D segmentations to the 3D domain suffer from view inconsistencies and coarse masks. In this paper, we propose a novel framework that leverages the Segment Anything Model High Quality (SAM-HQ) to generate accurate 2D masks, addressing the limitations of the standard SAM in boundary fidelity and fine-structure preservation. To achieve robust 3D segmentation of any target object in a given scene, we introduce a prior-guided label reassignment method that assigns labels to 3D Gaussians by enforcing multiview consistency with learned priors. Our approach achieves state-of-the-art segmentation accuracy and enables interactive, real-time object editing while maintaining high visual fidelity. Qualitative results demonstrate superior boundary preservation and practical utility in Virtual Reality (VR) and robotics, advancing 3D scene editing.\n\n---\n\n## 78. Misspecified Explore-then-Exploit Leads to Supra-Competitive Prices\n- **Authors**: Jackie Baek , Vivek F. Farias , Farrell Wu\n- **URL**: [https://arxiv.org/abs/2605.16064](https://arxiv.org/abs/2605.16064)\n- **Abstract**:\n> We study whether simple algorithmic pricing systems can systematically produce collusive-like prices in multi-firm markets. We consider firms using an explore-then-exploit pipeline: they randomize prices during an initial exploration phase, then estimate demand from their own historical data and set prices myopically thereafter. The estimation step relies on a misspecified, monopoly-style model that omits competitors' prices. We characterize when this pipeline converges to supra-competitive prices above the Nash equilibrium, via a fluid-limit ordinary differential equation analysis. We show that supra-competitive prices arise when firms explore within similar price ranges on the same side of the Nash price. Moreover, prices can be substantially above the Nash price; we show that prices can reach monopoly levels under symmetric exploration. Simulations calibrated to a real multifamily rental market confirm that supra-competitive outcomes arise robustly beyond our theoretical assumptions, including under finite horizons, heterogeneous products, and nonlinear logit demand.\n\n---\n\n## 79. Ada-Diffuser: Latent-Aware Adaptive Diffusion for Decision-Making\n- **Authors**: Fan Feng , Selena Ge , Minghao Fu , Zijian Li , Yujia Zheng , Zeyu Tang , Yingyao Hu , Biwei Huang , Kun Zhang\n- **URL**: [https://arxiv.org/abs/2605.16054](https://arxiv.org/abs/2605.16054)\n- **Abstract**:\n> Recent work has framed decision-making as a sequence modeling problem using generative models such as diffusion models. Although promising, these approaches often overlook latent factors that exhibit evolving dynamics, elements that are fundamental to environment transitions, reward structures, and high-level agent behavior. Explicitly modeling these hidden processes is essential for both precise dynamics modeling and effective decision-making. In this paper, we propose a unified framework that explicitly incorporates latent dynamic inference into generative decision-making from minimal yet sufficient observations. We theoretically show that under mild conditions, the latent process can be identified from small temporal blocks of observations. Building on this insight, we introduce Ada-Diffuser, a causal diffusion model that learns the temporal structure of observed interactions and the underlying latent dynamics simultaneously, and furthermore, leverages them for planning and control. With a modular design, Ada-Diffuser supports both planning and policy learning tasks, enabling adaptation to latent variations in dynamics, rewards, and latent actions. Experiments on simulated control and robotic benchmarks demonstrate its effectiveness in accurate latent inference and adaptive policy learning.\n\n---\n\n## 80. Looped SSMs: Depth-Recurrence and Input Reshaping for Time Series Classification\n- **Authors**: Mónika Farsang , Ramin Hasani , Daniela Rus , Radu Grosu\n- **URL**: [https://arxiv.org/abs/2605.16048](https://arxiv.org/abs/2605.16048)\n- **Abstract**:\n> State Space Models (SSMs) are inherently recurrent along the sequence dimension, yet depth-recurrence - reusing the same block repeatedly across layers, as recently applied in looped transformers - has not been explored in this model family. We show that a looped SSM with $k$ parameters iterated $L$ times consistently closely matches or outperforms a standard SSM with $k \\cdot L$ independent parameters across four architectures (LRU, S5, LinOSS, LrcSSM) and six time series classification benchmarks, despite operating within a strictly smaller hypothesis space, as we formally establish. Since the larger model contains the looped model as a special case, this dominance cannot be explained by expressivity and instead points to parameter sharing across depth as a beneficial inductive bias that simplifies optimization. These results demonstrate that depth-recurrence is orthogonal to sequence-recurrence and independently beneficial. We further show that input reshaping is an equally neglected design axis: concatenating timesteps for low-dimensional inputs, or flattening and rechunking the joint feature-time dimension for high-dimensional ones, yields accuracy gains of 1-6% across all models, confirmed over 5 random seeds. Both techniques provide standalone improvements that compound when combined, suggesting that depth and input reshaping are two independent and underexplored design axes for SSMs on time series.\n\n---\n\n## 81. XSearch: Explainable Code Search via Concept-to-Code Alignment\n- **Authors**: Yiming Liu , Ruofan Liu , Yun Lin , Zicong Zhang , Weiyu Kong , Pengnian Qi , Xiao Cheng , Weinan Zhang , Qianxiang Wang , Linpeng Huang\n- **URL**: [https://arxiv.org/abs/2605.16046](https://arxiv.org/abs/2605.16046)\n- **Abstract**:\n> Semantic code search has been widely adopted in both academia and industry. These approaches embed natural-language queries and code snippets into a shared embedding space and retrieve results based on vector similarity. Despit strong performance on benchmark datasets, they often suffer from poor explainability and generalization. Retrieved code may appear semantically similar yet miss critical functional requirements of the query, while providing no explanation of why the result was retrieved. Moreover, such failures become more severe under distribution shift, where models struggle to generalize to unseen benchmarks. In this work, we propose XSearch, an intrinsically explainable code search framework. Our key insight is that by relying on global embedding similarity, existing retrievers inherently take an inductive view. They learn statistical patterns rather than truly understanding the query's functional requirements. We address this problem by reformulating code search as a deductive concept alignment problem. XSearch (i) identifies functional concepts in the query and (ii) explicitly aligns them with corresponding code statements. This explain-then-predict design produces inherent concept-level explanations and mitigates shortcut learning that harms out-of-distribution generalization. We train an encoder with explicit concept-alignment objectives and perform retrieval through explicit matching between query concepts and code statements. Experiments show that, trained on CodeSearchNet using GraphCodeBERT (125M parameters), XSearch improves performance on out-of-distribution benchmarks from 0.02 to 0.33 (15x) over eight state-of-the-art retrievers, and consistently outperforms both encoder- and decoder-based baselines with up to 7B parameters. A user study demonstrates that concept-alignment explanations enable users to evaluate retrieved results faster and more accurately.\n\n---\n\n## 82. RecMem: Recurrence-based Memory Consolidation for Efficient and Effective Long-Running LLM Agents\n- **Authors**: Zijie Dai , Shiyuan Deng , Sheng Guan , Yizhou Tian , Xin Yao , Xiao Yan , James Cheng\n- **URL**: [https://arxiv.org/abs/2605.16045](https://arxiv.org/abs/2605.16045)\n- **Abstract**:\n> Memory systems often organize user-agent interactions as retrievable external memory and are crucial for long-running agents by overcoming the limited context windows of LLMs. However, existing memory systems invoke LLMs to process every incoming interaction for memory extraction, and such an eager memory consolidation scheme leads to substantial token consumption. To tackle this problem, we propose RecMem by rethinking when memory consolidation should be conducted. RecMem stores incoming interactions in a subconscious memory layer and encode them using lightweight embedding models for retrieval. LLMs are only invoked to extract episodic and semantic memory when sustained recurrence are observed for semantically similar interactions. Such recurrence-based consolidation works because these interactions correspond to a semantic cluster with rich information and thus are worth extraction and summarization. To improve accuracy, RecMem also incorporates a semantic refinement mechanism that recovers the fine-grained facts omitted by memory extraction. Experiments show that RecMem reduces the memory construction token cost of three SOTA memory systems by up to 87% while exceeding their accuracy.\n\n---\n\n## 83. Learning Sim-Grounded Policies for Bimanual Rope Manipulation from Human Teleoperation Data\n- **Authors**: Gina Wigginghaus , Tim Missal , Berk Guler , Simon Manschitz , Jan Peters\n- **URL**: [https://arxiv.org/abs/2605.16043](https://arxiv.org/abs/2605.16043)\n- **Abstract**:\n> Deformable Linear Objects (DLOs) such as ropes and cables are widely encountered in both household and industrial applications, yet remain challenging to manipulate due to their infinite-dimensional configuration space and frequent self-occlusion. Imitation learning from teleoperation offers a practical path to bimanual DLO manipulation, but its scalability is limited by human effort, making the choice of observation space critical for generalization from small datasets. In this study, we investigate whether the lack of generalization in egocentric visual policies for the knot-untangling task stems from the observation space itself, rather than from the policy architecture or data scale. We compare two Action Chunking with Transformers policies trained on the same bimanual teleoperation data: a vision-based policy conditioned on two egocentric RGB streams from wrist-mounted cameras, and a state-based policy conditioned on the DLO's 3D particle state, extracted from an initial observation via multi-view fusion and evolved in a particle-based eXtended Position-Based Dynamics simulation. Evaluated open-loop on an unseen rope configuration, the state-based policy outperforms its visual counterpart with a 30.8% reduction in L1 error when predicting the initial grasp-and-pull action, quantifying the observability gap between pixels and physics-consistent state, and pointing toward more data-efficient robot learning for the DLO manipulation task from limited human demonstrations.\n\n---\n\n## 84. Who Owns This Agent? Tracing AI Agents Back to Their Owners\n- **Authors**: Ruben Chocron , Doron Jonathan Ben Chayim , Eyal Lenga , Gilad Gressel , Alina Oprea , Yisroel Mirsky\n- **URL**: [https://arxiv.org/abs/2605.16035](https://arxiv.org/abs/2605.16035)\n- **Abstract**:\n> AI agents are increasingly deployed to act autonomously in the world, yet there is still no reliable way to trace a harmful agent back to the account that deployed it. This creates the same accountability gap across both ends of the intent spectrum: benign operators may deploy misconfigured or overbroad agents that cause harm unintentionally, while malicious operators may deliberately weaponize agents for scams, harassment, or cyber attacks. In many cases, these agents are powered by vendor-hosted models, a dependency that holds even for sophisticated adversaries such as state actors conducting cyber operations. In either case, affected parties can observe the behavior but cannot notify the responsible operator, stop the session, or identify the account for investigation. We formalize this gap as the problem of agent attribution: linking an observed agent interaction to the responsible account at the hosting vendor. To our knowledge, this is the first work to define the problem and present a practical solution. Our protocol is canary-based: an authorized party injects a canary into the agent's interaction stream, and the vendor searches a narrow window of session logs to recover the originating session and account. Simple canaries suffice in non-adversarial settings. For adversarial operators who filter or paraphrase incoming content, we develop robust canary constructions that cannot be suppressed without degrading the agent's own task performance, yielding a formal asymmetry in the defender's favor. We evaluate a variety of scenarios including real-world agents and show that our attribution method is reliable, robust, and scalable for vendor-side deployment.\n\n---\n\n## 85. From Flat Language Labels to Typological Priors: Structured Language Conditioning for Multilingual Speech-to-Speech Translation\n- **Authors**: Yu Pan , Yang Hou , Xiongfei Wu , Liang Zhang , Yves Le Traon , Lei Ma , Jianjun Zhao\n- **URL**: [https://arxiv.org/abs/2605.16026](https://arxiv.org/abs/2605.16026)\n- **Abstract**:\n> Compositional speech-to-speech translation (S2ST) systems built upon speech large language models (SpeechLLMs) have recently shown promising performance. However, existing S2ST systems often either neglect source-language information or encode it through a language-as-label paradigm, representing each source language as an independent flat embedding. Such a design overlooks systematic linguistic structure shared across languages, which may limit data-efficient multilingual adaptation when supervised S2ST data are scarce. To address this issue, we propose S2ST-Omni 2, a many-to-one compositional S2ST framework that systematically reformulates multilingual language conditioning from flat language labels to structured typological priors. Specifically, S2ST-Omni 2 revisits language conditioning at three levels: typology-informed hierarchical language encoding for structured source-language representation, dynamically-gated language-aware Dual-CTC for content-adaptive acoustic modulation, and typology-aware LLM prompting for decoder-side linguistic guidance. Experiments on CVSS-C show that S2ST-Omni 2 achieves superior average performance among representative S2ST approaches across BLEU, COMET, ASR-BLEU, and BLASER 2.0 under the adopted evaluation protocol. Ablation studies indicate that the proposed representation-level, acoustic-level, and decoding-level strategies provide complementary benefits. Moreover, controlled data-budget analyses and a Japanese-to-English evaluation using only approximately 3 hours of supervised training data suggest that explicit typological priors provide useful inductive biases for data-efficient multilingual S2ST.\n\n---\n\n## 86. Can Vision Language Models Be Adaptive in Mathematics Education? A Learner Model-based Rubric Study\n- **Authors**: Jie Gao , Yongan Yu , Junzhu Su , Yiran Lin , Adam K. Dube , Jackie Chi Kit Cheung\n- **URL**: [https://arxiv.org/abs/2605.16011](https://arxiv.org/abs/2605.16011)\n- **Abstract**:\n> Adaptive learning refers to educational technologies that track learners' learning progress and adapt the instructional process based on individual learners' learning performance. It is increasingly recognized as critical for developing an effective learning support tool. Vision language models (VLMs) have seen adoption in mathematics education, and students have been using them as learning aids for personalized instruction. However, it is unknown whether VLMs have the ability to adapt to different learner profiles when providing mathematical instructions. Current VLMs lack a systematic evaluation framework for this adaptivity to different learner profiles in mathematics tutoring tasks. To address this gap, we draw on the learner model from the adaptive learning framework (Shute and Towle, 2018) and propose a learner model-based rubric. Our rubric formalizes adaptivity assessment into three aspects: cognitive aspects, motivational aspects, and complexity. We also evaluate two additional dimensions of VLM responses: correctness (of answers and solutions) and quality (of the response itself). Our experimental results show measurable differences in adaptivity across models and also reveal that current VLMs struggle to consistently produce learner model-based instructional responses, especially when receiving limited learner information.\n\n---\n\n## 87. CitePrism: Human-in-the-Loop AI for Citation Auditing and Editorial Integrity\n- **Authors**: Gowrika Mahesh , Budanur Madappa Darshan Gowda , Kavana Gopladevarahalli Papegowda , Prajwal Basavaraj , Binh Vu , Swati Chandna , Mehrdad Jalali\n- **URL**: [https://arxiv.org/abs/2605.16000](https://arxiv.org/abs/2605.16000)\n- **Abstract**:\n> Editors and reviewers are expected to ensure that manuscripts cite relevant, accurate, current, and ethically appropriate literature, yet manuscript-level citation auditing remains largely manual, fragmented, and difficult to scale. Citation context, metadata quality, self-citation patterns, and bibliographic integrity all affect whether a reference appropriately supports a local claim. We present CitePrism, a transparent hybrid decision-support framework for editorial citation auditing that combines LLM-assisted contextual reasoning, embedding-based semantic similarity, metadata verification, integrity-oriented flags, and human-in-the-loop analyst review. CitePrism extracts citation neighborhoods, enriches reference metadata, computes fused relevance scores, surfaces metadata and self-citation review prompts, and supports configurable threshold-based triage. In a preliminary validation on a single case-study manuscript with 104 references from pavement engineering, agreement with human binary relevance labels reached Cohen's kappa = 0.429. At operating threshold tau = 17, CitePrism flagged all human-labeled irrelevant citations, while also producing false positives requiring analyst review. These results suggest that CitePrism may support conservative editorial screening and citation-quality triage, but they do not establish general editorial performance. CitePrism is intended as pilot-stage decision support, not as an autonomous misconduct detector or automated editorial decision system. Broader validation across manuscripts, domains, annotators, baselines, and deployment settings is required before operational use.\n\n---\n\n## 88. Constrained latent state modeling: A unifying perspective on representation learning under competing constraints\n- **Authors**: Gwenolé Quellec\n- **URL**: [https://arxiv.org/abs/2605.15995](https://arxiv.org/abs/2605.15995)\n- **Abstract**:\n> Learning latent representations from complex data is central to modern machine learning, spanning temporal, multimodal, and partially observed systems. In such settings, representations are better understood as latent states capturing underlying system dynamics, rather than as mere compressed summaries of observations. Yet current approaches remain fragmented, relying on distinct -- and often implicit -- assumptions about what these states should represent. We argue that this fragmentation reflects a more fundamental limitation: latent representations are typically learned from underconstrained objectives that fail to specify the properties that meaningful latent states should satisfy. As a result, multiple representations can satisfy the same objective, leading to ambiguity in their structure and interpretation. While many of the underlying principles have been explored in isolation, their interactions have not been explicitly formalized. In this work, we propose constrained latent state modeling (CLSM) as a unifying perspective. We identify a set of core properties -- predictive sufficiency, minimality, temporal coherence, observation compatibility, invariance to nuisance factors, and structural constraints -- and show that they are intrinsically coupled through fundamental trade-offs. Revisiting major modeling families through this lens, we show that existing approaches can be interpreted as enforcing different subsets of constraints, thereby occupying distinct regions of a common design space. This perspective reframes persistent challenges such as lack of identifiability as consequences of underconstrained formulations, rather than isolated technical limitations. More broadly, CLSM provides a principled framework to make design choices explicit, to analyze trade-offs, and to guide the development of more interpretable, robust, and task-aligned latent state models.\n\n---\n\n## 89. Beyond Content: A Comprehensive Speech Toxicity Dataset and Detection Framework Incorporating Paralinguistic Cues\n- **Authors**: Zhongjie Ba , Liang Yi , Peng Cheng , Qingcao Li , Qinglong Wang , Li Lu\n- **URL**: [https://arxiv.org/abs/2605.15984](https://arxiv.org/abs/2605.15984)\n- **Abstract**:\n> Toxic speech detection has become a crucial challenge in maintaining safe online communication environments. However, existing approaches to toxic speech detection often neglect the contribution of paralinguistic cues, such as emotion, intonation, and speech rate, which are key to detecting speech toxicity. Moreover, current toxic speech datasets are predominantly text-based, limiting the development of models that can capture paralinguistic this http URL address these challenges, we present ToxiAlert-Bench, a large-scale audio dataset comprising over 30,000 audio clips annotated with seven major toxic categories and twenty fine-grained toxic labels. Uniquely, our dataset annotates toxicity sources -- distinguishing between textual content and paralinguistic origins -- for comprehensive toxic speech this http URL , we propose a dual-head neural network with a multi-stage training strategy tailored for toxic speech detection. This architecture features two task-specific classification headers: one for identifying the source of sensitivity (textual or paralinguistic), and the other for categorizing the specific toxic type. The training process involves independent head training followed by joint fine-tuning to reduce task interference. To mitigate data class imbalance, we incorporate class-balanced sampling and weighted loss this http URL experimental results show that leveraging paralinguistic features significantly improves detection performance. Our method consistently outperforms existing baselines across multiple evaluation metrics, with a 21.1% relative improvement in Macro-F1 score and a 13.0% relative gain in accuracy over the strongest baseline, highlighting its enhanced effectiveness and practical applicability.\n\n---\n\n## 90. Ontology for Policing: Conceptual Knowledge Learning for Semantic Understanding and Reasoning in Law Enforcement Reports\n- **Authors**: Anita Srbinovska , Jansen Orfan , Adrian Martin , Ernest Fokoué\n- **URL**: [https://arxiv.org/abs/2605.15978](https://arxiv.org/abs/2605.15978)\n- **Abstract**:\n> Law enforcement reports contain structured fields and written narratives. However, many incident facts that are needed for review, police training, and investigations are in natural language and require manual reading. We propose a framework using symbolic methods for converting narratives into evidence-linked facts. Our objective is to measure the value of narratives to recover incident details only from the unstructured text and build temporal graphs with time cues and domain axioms. We achieve this by redacting personal identifiers, semantic parsing, predicate mapping to ontology, and reasoning. We evaluate the symbolic approach on 450 property crime reports and a short human review. Of the extracted events from the system, 54.1% had a confidence score of at least 0.80 and 93.7% were mapped through the PropBank--VerbNet--WordNet semantic path. 100% agreement was reached on incident initiation, stolen items, and temporal cues and lower agreement for forced entry interpretation.\n\n---\n\n## 91. Reference-Free Reinforcement Learning Fine-Tuning for MT: A Seq2Seq Perspective\n- **Authors**: Ernesto Garcia-Estrada , Carlos Escolano , José A. R. Fonallosa\n- **URL**: [https://arxiv.org/abs/2605.15976](https://arxiv.org/abs/2605.15976)\n- **Abstract**:\n> Production machine translation relies overwhelmingly on encoder-decoder Seq2Seq models, yet reinforcement learning approaches to MT fine-tuning have largely targeted decoder-only LLMs at $\\geq$7B parameters, with limited systematic study of encoder-decoder architectures. We apply Group Relative Policy Optimization to NLLB-200 (600M and 1.3B) using a hybrid reference-free reward (LaBSE and COMET-Kiwi) that requires no parallel data at fine-tuning time, evaluating across 13 typologically diverse languages. GRPO yields consistent improvements on all 13 languages, up to $+$5.03 chrF++ for Traditional Chinese, and, without any target-language data, competes with 3-epoch supervised fine-tuning on morphologically complex languages . We identify a consistent empirical pattern in which gains are largest where baseline performance is weakest and reward discriminability is highest, making this approach most effective precisely where parallel data is scarcest, and replicate this pattern across English and Spanish source languages.\n\n---\n\n## 92. When and Why Adversarial Training Improves PINNs: A Neural Tangent Kernel Perspective\n- **Authors**: Yuan-dong Cao , Chi Chiu SO , Jun-Min Wang , He Wang\n- **URL**: [https://arxiv.org/abs/2605.15959](https://arxiv.org/abs/2605.15959)\n- **Abstract**:\n> Physics-informed neural networks (PINNs) are powerful surrogates for differential equations but are notoriously difficult to train due to spectral bias, stiffness, and poor accuracy on high-frequency or multiscale solutions. Adversarial training based on generative adversarial networks (GANs) has recently gained surprisingly strong empirical results in improving training, but the underlying mechanisms remain elusive. To this end, we propose a new analysis framework for adversarially trained PINNs, based on the key observation of how the discriminator in GANs can influence the training dynamics of PINNs. The framework first provides a much needed theoretical grounding to why and when adversarial training is effective in PINNs, then presents a unified analysis of GANs variants in such training, and finally leads to a new, practical, efficient training algorithm for PINNs. Empirical results demonstrate that our method can significantly reduce the pathology of PINNs training, thereby providing better models with superior performances, often several magnitudes more accurate than alternative methods.\n\n---\n\n## 93. Decomposed Vision-Language Alignment for Fine-Grained Open-Vocabulary Segmentation\n- **Authors**: Chenhao Wang , Yingrui Ji , Yu Meng , Yao Zhu\n- **URL**: [https://arxiv.org/abs/2605.15942](https://arxiv.org/abs/2605.15942)\n- **Abstract**:\n> Open-vocabulary segmentation models often struggle to generalize to unseen combinations of object categories and attributes, because fine-grained descriptions are typically encoded as holistic sentences that entangle multiple semantic units. We propose a Decomposed Vision-Language Alignment framework that explicitly factorizes textual prompts into a concept token and multiple attribute tokens, enabling separate cross-modal interactions for each semantic unit. At the feature level, we introduce a Feature-Gated Cross-Attention module that generates attribute-specific gating maps to fuse information in a multiplicative manner, effectively enforcing compositional semantics. At the scoring level, per-token similarities are aggregated in log-space, producing a stable and interpretable compositional matching. The method can be seamlessly integrated into existing transformer-based segmentation architectures and significantly improves generalization to unseen attribute-category compositions in fine-grained open-vocabulary segmentation benchmarks.\n\n---\n\n## 94. LoCO: Low-rank Compositional Rotation Fine-tuning\n- **Authors**: An Nguyen , Jaesik Choi , Anh Tong\n- **URL**: [https://arxiv.org/abs/2605.15916](https://arxiv.org/abs/2605.15916)\n- **Abstract**:\n> Parameter-efficient fine-tuning (PEFT) has emerged as an critical technique for adapting large-scale foundation models across natural language processing and computer vision. While existing methods such as low-rank adaptations achieve parameter efficiency via low-rank weight updates, they are limited in their ability to preserve the geometric structure of pretrained representations. We introduce Low-rank Compositional Orthogonal fine-tuning (LoCO), a novel PEFT method that constructs orthogonal transformations through low-rank skew-symmetric matrices and compositional rotation chains. We propose an approximation scheme that enables fully parallel computation of compositional rotations, making the approach practical for high-dimensional feature spaces. Our method maintains low computational complexity while maintaining orthogonality with controlled approximation error. We validate LoCO across diverse domains, including diffusion transformer fine-tuning, vision transformer adaptation, and language model adaptation. Our method demonstrates superior or competitive performance compared to both existing orthogonal and non-orthogonal methods.\n\n---\n\n## 95. SLIP & ETHICS: Graduated Intervention for AI Emotional Companions\n- **Authors**: Minseo Kim\n- **URL**: [https://arxiv.org/abs/2605.15915](https://arxiv.org/abs/2605.15915)\n- **Abstract**:\n> AI emotional companions face a safety-rapport paradox: restrictive safeguards can damage supportive alliance, while permissive systems risk user harm. We present SLIP (Staged Layers of Intervention Protocol), a four-stage graduated methodology deriving interventions (none, soft, hard) from structured qualitative indicators -- affect intensity (a) and narrative dynamism (m) -- alongside ETHICS (Emergent Taxonomy for Human-AI Interaction Context Signals), a \"signals not labels\" taxonomy. An evaluation combining a small-scale production deployment (N=68 entries, 10 users, 10 weeks) with a synthetic persona battery (N=91, 5 behavioral-risk profiles) achieved 0% false positives for the flow persona and showed expected escalation patterns in crisis-oriented personas. However, initial results showed that 8 consecutive days of high-energy elevation produced zero interventions (0/8), exposing a boundary where the \"do not pathologize\" principle conflicts with safety. A subsequent three-model stress test demonstrated that increased model capability improves detection from 0/8 to 6/8 while preserving 0/10 flow false positives in the largest model. Read as preliminary, these findings position graduated intervention as a design direction for navigating -- not resolving -- the safety-rapport tension in affective computing.\n\n---\n\n## 96. Towards Generalization of Block Attention via Automatic Segmentation and Block Distillation\n- **Authors**: Shuaiyi Li , Zhisong Zhang , Yan Wang , Lei Zhu , Dongyang Ma , Chenlong Deng , Yang Deng , Wai Lam\n- **URL**: [https://arxiv.org/abs/2605.15913](https://arxiv.org/abs/2605.15913)\n- **Abstract**:\n> Block attention, which processes the input as separate blocks that cannot attend to one another, offers significant potential to improve KV cache reuse in long-context scenarios such as Retrieval-Augmented Generation (RAG). However, its broader application is hindered by two key challenges: the difficulty of segmenting input text into meaningful, self-contained blocks, and the inefficiency of existing block fine-tuning methods that risk degrading performance. To address these, we first construct SemanticSeg, a large and diverse semantic segmentation dataset containing over 30k instances across 16 categories-including books, code, web text, and conversations with text lengths ranging from 2k to 32k. Using this dataset, we train a lightweight segmenter to automatically partition text into human-instinct-aligned blocks with controllable granularity. Second, we propose block distillation, a training framework that is more efficient than block fine-tuning, which uses a frozen full-attention teacher model to guide the block-attention student. This framework integrates three novel components: block sink tokens to mitigate information loss at block boundaries, block dropout to leverage training signals from all blocks, and token-level loss weighting to focus learning on block-attention-sensitive tokens. Experiments across multiple models and benchmarks demonstrate that our segmenter outperforms heuristic and statistical baselines, and block distillation achieves near-full-attention performance under block attention, establishing a practical and scalable pathway for deploying block attention.\n\n---\n\n## 97. RaPD: Resolution-Agnostic Pixel Diffusion via Semantics-Enriched Implicit Representations\n- **Authors**: Yanhao Ge , Shanyan Guan , Weihao Wang , Ying Tai , Mingyu You\n- **URL**: [https://arxiv.org/abs/2605.15908](https://arxiv.org/abs/2605.15908)\n- **Abstract**:\n> Natural images are continuous, yet most generative models synthesize them on discrete grids, limiting resolution-flexible generation. Continuous neural fields enable resolution-free rendering, but prior methods introduce continuity only at the decoding stage as an interpolation module, leaving the generative latent space discretized and reconstruction-oriented. We propose RaPD (Resolution-agnostic Pixel Diffusion), which performs diffusion in a continuous Neural Image Field (NIF) latent space. RaPD bridges this reconstruction-generation gap with Semantic Representation Guidance for generation-aware latent learning and a Coordinate-Queried Attention Renderer for coordinate-conditioned, scale-aware rendering. A single denoised latent can be rendered at arbitrary resolutions by changing only the query coordinates, keeping diffusion cost fixed. Experiments demonstrate superior generation quality and resolution scalability.\n\n---\n\n## 98. Generative Long-term User Interest Modeling for Click-Through Rate Prediction\n- **Authors**: Jiangli Shao , Kaifu Zheng , Hao Fang , Huimu Ye , Zhiwei Liu , Bo Zhang , Shu Han , Xingxing Wang\n- **URL**: [https://arxiv.org/abs/2605.15905](https://arxiv.org/abs/2605.15905)\n- **Abstract**:\n> Modeling long-term user interests with massive historical user behaviors enhances click-through rate (CTR) prediction performance in advertising and recommendation systems. Typically, a two-stage framework is widely adopted, where a general search unit (GSU) first retrieves top-$k$ relevant behaviors towards the target item, and an exact search unit (ESU) generates interest features via tailored attention. However, current target-centered GSU would ignore other latent user interests, leading to incomplete and biased interest features. Additionally, the matching-based retrieval process in GSUs depends on the pairwise similarity score between target item and each historical behavior, which not only becomes time-consuming for online services as user behaviors continue to grow, but also overlooks the interaction information among user behaviors. To combat these problems, we propose a \\textbf{Gen}erative \\textbf{L}ong-term user \\textbf{I}nterest model named GenLI for CTR prediction. GenLI consists of an interest generation module (IGM), a behavior retrieval module (BRM), and an interest fusion module (IFM). The IGM generates multiple interest distributions to indicate different aspects of real-time user interests, which is target-independent and incorporates interaction information among behaviors, ensuring complete and diverse interest features. The BRM selects related behaviors via a simple lookup operation, reducing the time complexity for weighting each behavior to $O(1)$. Finally, the IFM uses delicate gating mechanisms to generate interest features. Based on the generation process, GenLI improves the diversity of user interests and avoids complex matching-based behavioral retrieval, achieving a better balance between accuracy and efficiency for CTR prediction.\n\n---\n\n## 99. Uncertainty-Aware Wildfire Smoke Density Classification from Satellite Imagery via CBAM-Augmented EfficientNet with Evidential Deep Learning\n- **Authors**: Ranjith Chodavarapu\n- **URL**: [https://arxiv.org/abs/2605.15894](https://arxiv.org/abs/2605.15894)\n- **Abstract**:\n> Rapid and accurate wildfire smoke severity assessment from satellite images is essential for emergency response, air quality modeling, and human health risk management. Existing deep learning approaches treat smoke detection as a binary task, producing point estimates without any measure of prediction confidence. We propose a probabilistic framework to categorize a satellite patch into Light, Moderate, and Heavy severity classes and to provide decomposed epistemic and aleatoric uncertainty in a single forward pass. Our architecture uses the backbone of a pre-trained EfficientNet-B3 and a CBAM module with an evidential deep learning head that predicts Dirichlet concentration parameters, directly estimating vacuity (epistemic) and dissonance (aleatoric) without Monte Carlo sampling. Evaluated on 16,298 real satellite patches derived from the Wildfire Detection dataset, our model achieves 93.8% weighted test accuracy (91.1% unweighted) with ECE=0.0274. Selective prediction retaining the most certain 50% of patches achieves 96.7% accuracy. As image quality degrades, uncertainty increases monotonically, and vacuity is a practical scan quality measure. The Moderate class represents transitional smoke conditions that exhibit the highest epistemic uncertainty (mean vacuity = 0.187), confirming the model correctly identifies ambiguous smoke boundary regions. CBAM spatial attention maps localize to structurally distinctive scene regions, and t-SNE demonstrates the clear cluster separation of Light and Heavy smoke.\n\n---\n\n## 100. CHoE: Cross-Domain Heterogeneous Graph Prompt Learning via Structure-Conditioned Experts\n- **Authors**: Peiyuan Li , Yongqi Huang , Jitao Zhao , Dongxiao He , Di Jin , Weixiong Zhang\n- **URL**: [https://arxiv.org/abs/2605.15888](https://arxiv.org/abs/2605.15888)\n- **Abstract**:\n> Heterogeneous Graph Prompt Learning (HGPL)has emerged as a promising paradigm for bridging the gap between the objectives of pre-training foundation models and their downstream applications in heterogeneous graph settings. However, existing HGPL methods are primarily designed for in-domain scenarios, whereas real-world deployments often span multiple domains, and the data used for pre-training and downstream tasks may originate from different distributions. Consequently, the applicability of current HGPL approaches is limited to in-domain settings, and their performance typically degrades when application domains shift. To address this serious limitation, we develop CHoE, a cross-domain HGPL method built upon an expert network. During pre-training, we introduce and train structure-conditioned experts, and during prompt tuning, we adopt a structure-aware expert routing and load balancing mechanism to select structurally compatible experts for each meta-path view. In addition, we design a prompt-based semantic fusion module to integrate representations across multiple views for downstream prediction. Extensive experiments show that CHoE consistently improves performance in few-shot cross-domain applications, outperforming all baseline approaches.\n\n---\n\n## 101. Symplectic Neural Operators for Learning Infinite Dimensional Hamiltonian Systems\n- **Authors**: Yeang Makara , Yusuke Tanaka , Takashi Matsubara , Takaharu Yaguchi\n- **URL**: [https://arxiv.org/abs/2605.15881](https://arxiv.org/abs/2605.15881)\n- **Abstract**:\n> The modeling and simulation of infinite-dimensional Hamiltonian systems are central problems in mathematical physics and engineering, however they pose significant computational and structural challenges for standard data-driven architectures. In this work, we introduce the Symplectic Neural Operator, a neural operator architecture designed to preserve the symplectic structure intrinsic to Hamiltonian PDEs. We provide a theoretical characterization of their symplecticity and establish a rigorous long-term stability result based on the combination of symplectic structure preservation and learning accuracy. Numerical experiments on canonical Hamiltonian PDEs corroborate this theoretical result and show that SNOs exhibit improved energy behavior compared with non-structure-preserving neural operators.\n\n---\n\n## 102. FSCM: Frequency-Enhanced Spatial-Spectral Coupled Mamba for Infrared Hyperspectral Image Colorization\n- **Authors**: Tingting Liu , Yuan Liu , Guiping Chen , Xiubao Sui , Qian Chen\n- **URL**: [https://arxiv.org/abs/2605.15880](https://arxiv.org/abs/2605.15880)\n- **Abstract**:\n> Thermal infrared imaging is robust to illumination variations and smoke interference, making it important for all-weather perception. However, the lack of natural color and fine texture limits target recognition, human visual interpretation, and the transfer of visible-light models. Existing infrared colorization methods mainly rely on single-band images, where insufficient spectral cues may lead to structural distortion and semantic confusion. Although infrared hyperspectral images provide rich spectral responses and material information, existing single-band frameworks remain limited in modeling spatial-spectral coupling and weak texture details. To address these issues, this paper presents FSCM, a spectral-information-guided GAN framework. Within FSCM, a frequency-enhanced spatial-spectral state-space generator composed of cascaded FSB units is constructed. Each FSB integrates three complementary components: state-space modeling captures global spatial-spectral dependencies; the frequency enhancement module (FEM) combines multi-level wavelet decomposition and Fourier gating to recover structural contours, directional high-frequency details, and global frequency responses; and the dual-stream hybrid gating module (DGM) integrates deformation-aware sampling with sparse attention to enhance effective local structures and suppress background interference. Additionally, an online semantic segmentation-guided loss is introduced to constrain the generated results, improving semantic consistency in complex road scenes. Experiments show that FSCM outperforms existing infrared colorization methods in visual quality and semantic fidelity.\n\n---\n\n## 103. Shapley Neuron Values for Continual Learning: Which Neurons Matter Most?\n- **Authors**: Mohammad Ali Vahedifar , Abhisek Ray , Qi Zhang\n- **URL**: [https://arxiv.org/abs/2605.15877](https://arxiv.org/abs/2605.15877)\n- **Abstract**:\n> Continual learning enables neural networks to learn tasks sequentially without forgetting previously acquired knowledge. However, neural networks suffer from catastrophic forgetting, where learning new tasks degrades performance on earlier ones. We address this problem with Shapley Neuron Valuation (SNV), a principled framework that quantifies Neuron importance in continual learning, grounded in cooperative game theory. SNV selectively freezes important Neurons while keeping others plastic, enabling buffer-free continual learning without expanding architecture. Experiments on ImageNet-1k show that SNV consistently outperforms existing buffer-free methods. In particular, SNV improves accuracy by +2.88% in the class incremental learning and +6.46% in the task incremental learning scenarios compared to the second baseline.\n\n---\n\n## 104. Access Timing as Scaffolding: A Reinforcement Learning Approach to GenAI in Education\n- **Authors**: Janne Rotter , Pau Benazet i Montobbio , Davinia Hernández-Leo\n- **URL**: [https://arxiv.org/abs/2605.15850](https://arxiv.org/abs/2605.15850)\n- **Abstract**:\n> In recent years, generative AI (GenAI) in educational settings has become ubiquitous in students' daily lives, despite its potential to induce over-reliance, metacognitive disengagement, and diminished learning when used unrestrictedly. While most prior research has thus focused on how to pedagogically scaffold its usage, the question of when to allow off-the-shelf GenAI remains understudied and lacks pedagogically grounded empirical investigation. We treat access timing itself as a form of implicit scaffolding and operationalize it through a reinforcement learning (RL) agent that decides when students should access GenAI, with a reward function grounded in metacognitive theory, cognitive load theory, and productive failure. In a mixed-methods controlled lab study with N=105 participants, we compared the agent's effect on learning gains and metacognitive engagement to unrestricted and fully restricted use. Results show that strategically timed GenAI access under the reinforcement learning condition improved objective post-test performance and metacognitive accuracy compared with unrestricted access, while reducing task errors and time on task relative to complete withholding, all without the need for explicit metacognitive prompts or structured scaffolding. However, no between-condition differences emerged on self-reported metacognitive awareness. Overall, timing of GenAI access therefore is a tractable, theoretically grounded, and scalable pedagogical paradigm that improves over completely unrestricted and withheld access, compatible with off-the-shelf tools and potentially low adoption barrier. This opens up a new research area that explores how access timing can be facilitated by educators and implemented in human-AI learning system design.\n\n---\n\n## 105. RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades\n- **Authors**: Xinbo Xu , Ruihan Yang , Haiyang Shen , Wendong Xu , Bofei Gao , Ruoyu Wu , Kean Shi , Weichu Xie , Xuanzhong Chen , Ming Wu , Jason Zeng , Michael Heinrich , Elvis Zhang , Liang Chen , Kuan Li , Baobao Chang\n- **URL**: [https://arxiv.org/abs/2605.15846](https://arxiv.org/abs/2605.15846)\n- **Abstract**:\n> Coding agents are increasingly deployed in real software development, where a single version iteration requires months of coordinated work across many files. However, most existing benchmarks focus predominantly on single-issue bug fixes from Python repositories, with coarse pass/fail evaluation outcomes, and thus fail to capture long-horizon, multi-target development at real engineering scale. To address this gap, we present RoadmapBench, a benchmark of 115 long-horizon coding tasks grounded in real open-source version upgrades across 17 repositories and 5 programming languages. Each task places the agent on a source-version code snapshot and provides a multi-target roadmap instruction requiring it to implement the functionality introduced in the target version, with a median modification of 3,700 lines across 51 files. We conduct a systematic evaluation on thirteen frontier models and find that even the strongest, Claude-Opus-4.7, resolves only 39.1% of tasks, while the weakest achieves merely 5.2%, in stark contrast to existing bug-fix benchmarks, suggesting that long-horizon software development remains a largely unsolved problem.\n\n---\n\n## 106. GAP: Geometric Anchor Pre-training for Data-Efficient Visuomotor Learning of Manipulation Tasks\n- **Authors**: Davide Buoso , Andrea Protopapa , Stefano Di Carlo , Francesca Pistilli , Giuseppe Averta\n- **URL**: [https://arxiv.org/abs/2605.15836](https://arxiv.org/abs/2605.15836)\n- **Abstract**:\n> Learning visuomotor policies from scarce expert demonstrations remains a core challenge in robotic manipulation. A primary hurdle lies in distilling high-dimensional RGB representations into control-relevant geometry without overfitting. While using frozen pre-trained Vision Foundation Models (VFMs) improves data efficiency, it also shifts most task adaptation onto a small spatial pooling module, which can latch onto task-irrelevant shortcuts and lose geometric grounding when finetuned with few data samples. More broadly, pre-trained visual representations used for policy learning have been observed to struggle under even minor scene perturbations, highlighting the need for robustness-oriented inductive biases. We propose Geometric Anchor Pre-training (GAP), a simple, action-free warm-up stage that regularizes the spatial adapter before downstream imitation learning. GAP pre-trains the pooling layer on a lightweight simulated proxy task where object masks are available at no cost, encouraging the adapter to produce keypoints that lie on the object, cover its spatial extent, and remain sharp and repeatable over time. This yields stable geometric anchors that provide a reliable coordinate interface for few-shot policy learning, while keeping the VFM frozen. We evaluate GAP on RoboMimic and ManiSkill under severe data scarcity (15-50 demonstrations) and domain shift. A simple adapter regularized with GAP consistently outperforms stronger attention-based poolers and end-to-end fine-tuning, achieving 62% success on RoboMimic Can with 15 demonstrations (+16% over AFA), 63% on the long-horizon high-precision Tool Hang task with 50 demonstrations, and 61% on ManiSkill StackCube with 30 demonstrations (+11% over full fine-tuning). The proxy stage is lightweight and fully decoupled from downstream tasks, making it practical to reuse across environments and manipulation skills.\n\n---\n\n## 107. Modeling Music as a Time-Frequency Image: A 2D Tokenizer for Music Generation\n- **Authors**: Yuqing Cheng , Xingyu Ma , Guochen Yu , Xiaotao Gu\n- **URL**: [https://arxiv.org/abs/2605.15831](https://arxiv.org/abs/2605.15831)\n- **Abstract**:\n> Autoregressive music generation depends strongly on the audio tokenizer. Existing high-fidelity codecs often use residual multi-codebook quantization, which preserves reconstruction quality but complicates language modeling after sequence flattening, as the residual hierarchy imposes strong sequential dependencies and can amplify error accumulation. We propose BandTok, a generation-oriented 2D Mel-spectrogram tokenizer that represents each frame with Mel-frequency band tokens from a single shared codebook. This design yields a physically interpretable time-frequency token grid with a more independent token structure, making it better suited for autoregressive modeling. BandTok improves reconstruction with a multi-scale PatchGAN objective and EMA codebook updates. We further introduce an autoregressive language model with 2D Rotary Position Embedding (2D RoPE) to preserve temporal and frequency-band structure during generation. Experiments show that BandTok improves over residual-codebook tokenizers and achieves strong results in a data-limited setting. The source code and generation demos for this work are publicly available.\n\n---\n\n## 108. Toward Natural and Companionable Virtual Agents via Cross-Temporal Emotional Modeling\n- **Authors**: Feier Qin , Xiao Li , Yi Zheng , Haibin Huang , Hanyao Wang , Xiaoyu Wang , Yan Lu , Yuan Zhang\n- **URL**: [https://arxiv.org/abs/2605.15812](https://arxiv.org/abs/2605.15812)\n- **Abstract**:\n> Recent advances in foundation models have enabled conversational agents that aim for sustained companionship rather than mere task completion. Yet most still remain unable to support natural, long-term companion-like interactions, resulting in experiences that feel episodic and inauthentic. We argue that current agents overlooked cross-temporal modeling of agents' social behaviors and internal emotions: generated behaviors rarely influence an agent's emotional state, and emotional states seldom shape subsequent behaviors. We present Cross-Temporal Emotion Modeling (CTEM), a framework that links long-term behavioral history to moment-to-moment emotional expression. CTEM establishes a closed loop where past experiences update an evolving emotional state; this state conditions immediate interactions; and user feedback continually revises both memory and emotional state, enabling reflection and anticipation. We instantiate CTEM as Auri, a companion agent on an instant-messaging platform, and report a 21-day in-the-wild study showing that CTEM shows improvements in perceived naturalness, coherence, and emotional harmony.\n\n---\n\n## 109. Grokking as Structural Inference: Transformers Need Bayesian Lottery Tickets\n- **Authors**: Kai Hidajat , Solden Stoll , Joseph An\n- **URL**: [https://arxiv.org/abs/2605.15787](https://arxiv.org/abs/2605.15787)\n- **Abstract**:\n> Why does a Transformer that has memorized its training set wait thousands of steps before it generalizes? Existing accounts locate this delay in norm minimization, feature emergence, or the late discovery of sparse subnetworks. These explanations capture important parts of the transition, but ignore a constraint unique to attention-based models: if attention discards an informative token, no bounded downstream computation can recover it. We formalize attention as an implicit Bayesian posterior over the task dependency graph and prove that generalization requires two separable conditions: a familiar Goldilocks bound on MLP capacity, coinciding with norm-based theories of grokking, and a novel Bayesian structural condition requiring attention to place sufficient mass on every informative token. This decoupling explains delayed generalization as delayed structural inference. Early in training, the MLP memorizes through unaligned features, drives the cross-entropy loss near zero, and thereby starves attention of structural gradient. Weight decay must then erode memorization before the missing graph becomes learnable, yielding the known inverse-weight-decay delay, which we derive as a structural waiting time. We then prove that this explaining-away delay can be bypassed by a KL-based structural intervention, yielding an inverse-intervention-strength scaling law for the grokking time. Experiments on algorithmic sequence tasks isolate structure from capacity and show that this Bayesian ticket matches or outperforms lottery-ticket transfer.\n\n---\n\n## 110. A Topology-Aware Spatiotemporal Handover Framework for Continuous Multi-UAV Tracking\n- **Authors**: Jianlin Ye , Christos Kyrkou , Panayiotis Kolios\n- **URL**: [https://arxiv.org/abs/2605.15779](https://arxiv.org/abs/2605.15779)\n- **Abstract**:\n> The integration of Unmanned Aerial Vehicles(UAVs) into Intelligent Transportation Systems (ITS) offers synoptic visibility for traffic monitoring, yet scalable deployment is hindered by trajectory fragmentation, where vehicle identity persistence is lost across multi-UAV Fields of View (FOV). While state-of-the-art frameworks excel in optimizing local trajectory extraction and stability for single-drone imagery, they often function as isolated data silos that generate disjointed trajectories, thereby precluding network-level analysis such as Origin-Destination estimation. This paper presents a real-time Multi-Camera Multi-Vehicle Tracking (MCMT) system designed to handle global identity persistence. Addressing the visual ambiguity and computational cost of appearance-based Re-Identification (Re-ID) in nadir views, we introduce a lightweight Topology-Based Spatiotemporal Handover mechanism. We implement a high-throughput parallel pipeline leveraging YOLO11 and ByteTrack to process concurrent 4K streams. Our core contribution is a deterministic queue-based matching algorithm that utilizes geometric overlaps and virtual lane discretization to predictively manage identity handover via FIFO queues. Experimental results on complex urban environments, including intersections and merging traffic, demonstrate a Handover Success Rate (HOSR) of 99.8% in continuous traffic flows, significantly outperforming Re-ID baselines (74.1%) while validating edge deployment feasibility. The source code is available at this https URL .\n\n---\n\n## 111. Lamarckian Inheritance in Dynamic Environments: How Key Variables Affect Evolutionary Dynamics\n- **Authors**: K. Ege de Bruin , Kyrre Glette , Kai Olav Ellefsen\n- **URL**: [https://arxiv.org/abs/2605.15769](https://arxiv.org/abs/2605.15769)\n- **Abstract**:\n> The co-optimization of a robot's body and brain presents a coupled challenge: the morphology constrains which control strategies are effective, while the control determines how well the morphology performs. To address this, we combine morphology optimization as evolution with controller optimization as lifetime learning, utilizing Lamarckian inheritance to transfer learned controller parameters from parent to offspring. In dynamic environments, existing literature presents conflicting evidence: while traditional evolutionary theory often suggests Lamarckian inheritance lacks benefit, recent studies in evolutionary robotics indicate it can improve performance. We hypothesize that this is because previous works have not included all relevant variables with dynamic environments. In this work, we show that the benefit of Lamarckian inheritance depends on two variables: how conflicting the environmental changes are to robot control, and the predictability of those changes for the robotic agent. Using virtual soft robots and two different learning approaches, Bayesian optimization and reinforcement learning, we show that Lamarckian inheritance only underperforms Darwinian inheritance when the changes are both conflicting and unpredictable. We find that adding a sensor to detect environmental changes restores the benefits for Lamarckian inheritance in conflicting environments, by allowing robotic agents to predict the need for a different behavior, thereby generalizing their control.\n\n---\n\n## 112. GRASP: Learning to Ground Social Reasoning in Multi-Person Non-Verbal Interactions\n- **Authors**: Junho Kim , Xu Cao , Houze Yang , Bikram Boote , Ana Jojic , Fiona Ryan , Bolin Lai , Sangmin Lee , James M. Rehg\n- **URL**: [https://arxiv.org/abs/2605.15764](https://arxiv.org/abs/2605.15764)\n- **Abstract**:\n> Understanding social interactions requires reasoning over subtle non-verbal cues, yet current multimodal large language models (MLLMs) often fail to identify who interacts with whom in multi-person videos. We introduce GRASP, a large-scale social reasoning dataset that connects high-level social QA with fine-grained gaze and deictic gesture events. GRASP contains 290K question--answer pairs over 46K videos totaling 749 hours, organized by a 16-category taxonomy spanning gaze, gesture, and joint gaze--gesture reasoning, together with GRASP-Bench for evaluation. Unlike prior resources that focus on either isolated cues or high-level social QA, GRASP builds questions from identity-consistent gaze trajectories, deictic gestures, and their joint compositions into social events. Moreover, we propose Social Grounding Reward (SGR), a learning signal that uses these social events to encourage models to reason about the participants involved in each interaction. Experiments show that SGR improves performance on GRASP-Bench while maintaining zero-shot performance on related social video QA benchmarks.\n\n---\n\n## 113. CompactQE: Interpretable Translation Quality Estimation via Small Open-Weight LLMs\n- **Authors**: Kamil Guttmann , Zofia Fraś , Artur Nowakowski , Krzysztof Jassem\n- **URL**: [https://arxiv.org/abs/2605.15763](https://arxiv.org/abs/2605.15763)\n- **Abstract**:\n> Current state-of-the-art Quality Estimation (QE) in machine translation relies on massive, proprietary LLMs, raising data privacy concerns. We demonstrate that smaller, open-source LLMs (<30B parameters) are a viable, cost-effective and privacy-preserving alternative. Using a single-pass prompting strategy, our models simultaneously generate quality scores, MQM error annotations, suggested error corrections, and full post-editions. Our analysis shows these models achieve highly competitive system-level correlations with human judgments that outperform traditional neural metrics, fine-tuned models, and human inter-annotator agreement, effectively approximating the capabilities of much larger proprietary LLMs.\n\n---\n\n## 114. BiomedAP: A Vision-Informed Dual-Anchor Framework with Gated Cross-Modal Fusion for Robust Medical Vision-Language Adaptation\n- **Authors**: Huanyang Tong , Kai Liu , Fangjun Kuang , Huiling Chen\n- **URL**: [https://arxiv.org/abs/2605.15736](https://arxiv.org/abs/2605.15736)\n- **Abstract**:\n> Biomedical Vision--Language Models (VLMs) have shown remarkable promise in few-shot medical diagnosis but face a critical bottleneck: \\textit{fragility to prompt variations}.Existing adaptation frameworks typically optimize visual and textual prompts as independent streams, relying on ideal ``Golden Prompts''. In clinical reality, where descriptions are often noisy and heterogeneous, this modality isolation leads to unstable cross-modal alignment. To address this, we propose BiomedAP, a vision-informed dual-anchor framework with gated cross-modal this http URL enforces synergistic alignment through two mechanisms: (1) Gated Cross-Modal Fusion, which enables layer-wise interaction between modalities, acting as a dynamic noise regulator to suppress irrelevant textual cues; and (2) a Dual-Anchor Constraint that regularizes learnable prompts toward stable semantic centroids derived from both expert templates (High Anchors) and few-shot visual prototypes (Low Anchors). Extensive experiments across 11 benchmarks demonstrate that BiomedAP consistently surpasses baselines, achieving competitive few-shot accuracy and markedly enhanced robustness under prompt perturbations. Our code is available at: this https URL . Keywords: Vision-Language Models; Prompt Learning; Parameter-Efficient Fine-Tuning; Few-shot Learning\n\n---\n\n## 115. UAM: A Dual-Stream Perspective on Forgetting in VLA Training\n- **Authors**: Jianke Zhang , Yuanfei Luo , Yucheng Hu , Xiaoyu Chen , Yanjiang Guo , Ziyang Liu , Hongbin Xu , Tian Lan , Jianyu Chen\n- **URL**: [https://arxiv.org/abs/2605.15735](https://arxiv.org/abs/2605.15735)\n- **Abstract**:\n> Vision--language--action (VLA) models are typically built by fine-tuning a pretrained vision--language model (VLM) on action data. However, we show that this standard recipe systematically erodes the VLM's multimodal competence, a side effect we call the embodiment tax. But do VLAs have to forget? Inspired by the two-stream organization of biological vision, we trace this degradation to a structural bottleneck: current VLAs ask a single encoder to support both language-grounded semantics and control-relevant visual features, whereas biological vision separates recognition and visuomotor control into distinct pathways. Building on this view, we propose the Unified Action Model (UAM), which adds a parallel Dorsal Expert, an analog of the brain's dorsal pathway. To make the Dorsal Expert an effective second pathway and reduce the control-learning burden on the VLM, we initialize it from a pretrained generative model and train it with a mid-level reasoning objective that predicts visual dynamics. This design allows us to train the whole VLA end-to-end on action data alone: with no parameter freezing, no gradient stopping, and no auxiliary VL co-training, UAM retains over $95\\%$ of the underlying VLM's multimodal capability and at the same time achieves the highest average success rate among baselines on a variety of manipulation tasks that probe out-of-distribution generalization, including unseen objects, novel object--target compositions, and instruction variation. Together, these results suggest that semantic preservation in VLAs can emerge from architectural separation itself, rather than being enforced by frozen weights or auxiliary data replay, and that this preserved semantic capability can naturally transfer from VLMs to semantic generalization in actions.\n\n---\n\n## 116. Structure Abstraction and Generalization in a Hippocampal-Entorhinal Inspired World Model\n- **Authors**: Tianqiu Zhang , Muyang Lyu , Xiao Liu , Si Wu\n- **URL**: [https://arxiv.org/abs/2605.15733](https://arxiv.org/abs/2605.15733)\n- **Abstract**:\n> Humans abstract experiences into structured representations to facilitate pattern inference and knowledge transfer. While the hippocampal-entorhinal (HPC-MEC) circuit is known to represent both spatial and conceptual spaces, the mechanisms for concurrently extracting abstract structures from continuous, high-dimensional dynamics remain poorly understood. We propose a brain-inspired hierarchical model that simultaneously infers latent transitions and constructs a predictive visual world model. Our architecture employs an inverse model for structural extraction alongside an HPC-MEC coupling model that dissociates relational structures (MEC) from integrated episodic scenes (HPC). Using primitive transformation dynamics as a benchmark, we demonstrate the model's capacity for structural abstraction. By leveraging velocity-driven path integration, the framework enables robust prediction and structural reuse across diverse contexts, thereby achieving structural generalization. This work provides a novel computational framework for understanding how brain-inspired, self-supervised learning of world models facilitates the acquisition of reusable abstract knowledge.\n\n---\n\n## 117. DecomPose: Disentangling Cross-Category Optimization Contention for Category-Level 6D Object Pose Estimation\n- **Authors**: Yifan Gao , Lu Zou , Zhangjin Huang , Guoping Wang\n- **URL**: [https://arxiv.org/abs/2605.15728](https://arxiv.org/abs/2605.15728)\n- **Abstract**:\n> Category-level 6D object pose estimation is typically formulated as a multi-category joint learning problem with fully shared model parameters. However, pronounced geometric heterogeneity across categories entangles incompatible optimization signals in shared modules, resulting in gradient conflicts and negative transfer during training. To address this challenge, we first introduce gradient-based diagnostics to quantify module-level cross-category contention. Building on results of diagnostics, we propose DecomPose, a difficulty-aware decomposition framework that mitigates optimization contention via: (1) difficulty-aware gradient decoupling, which groups categories using a data-driven difficulty proxy and routes each instance to a group-specific correspondence branch to isolate incompatible updates; and (2) stability-driven asymmetric branching, which assigns higher-capacity branches to structurally simple categories as stable optimization anchors while constraining complex categories with lightweight branches to suppress noisy updates and alleviate negative transfer. Extensive experiments on REAL275, CAMERA25, and HouseCat6D demonstrate that DecomPose effectively reduces cross-category optimization contention and delivers superior pose estimation performance across multiple benchmarks.\n\n---\n\n## 118. DiLA: Disentangled Latent Action World Models\n- **Authors**: Tianqiu Zhang , Muyang Lyu , Yufan Zhang , Fang Fang , Si Wu\n- **URL**: [https://arxiv.org/abs/2605.15725](https://arxiv.org/abs/2605.15725)\n- **Abstract**:\n> Latent Action Models (LAMs) enable the learning of world models from unlabeled video by inferring abstract actions between consecutive frames. However, LAMs face a fundamental trade-off between action abstraction and generation fidelity. Existing methods typically circumvent this issue by using two-stage training with pre-trained world models or by limiting predictions to optical flow. In this paper, we introduce DiLA, a novel Disentangled Latent Action world model that aims to resolve this trade-off via content-structure disentanglement. Our key insight is that disentanglement and latent action learning are co-evolving: the predictive bottleneck inherent in latent action learning serves as a driving force for disentanglement, compelling the model to distill spatial layouts into the structure pathway while offloading visual details to a separate content pathway for generation. This synergy yields a continuous, semantically structured latent action space without compromising generative quality. DiLA achieves superior results in video generation quality, action transfer, visual planning, and manifold interpretability. These findings establish DiLA as a unified framework that simultaneously achieves high-level action abstraction and high-fidelity generation, advancing the frontier of self-supervised world model learning.\n\n---\n\n## 119. Bidirectional Fusion Guided by Cardiac Patterns for Semi-Supervised ECG Segmentation\n- **Authors**: Jeonghwa Lim , Minje Park , Sunghoon Joo\n- **URL**: [https://arxiv.org/abs/2605.15722](https://arxiv.org/abs/2605.15722)\n- **Abstract**:\n> Accurate delineation of electrocardiogram (ECG), the segmentation of meaningful waveform features, is crucial for cardiovascular diagnostics. However, the scarcity of annotated data poses a significant challenge for training deep learning models. Conventional semi-supervised semantic segmentation (SemiSeg) methods primarily focus on consistency from unlabeled data, underutilizing the information exchange possible between labeled and unlabeled sets. To address this, we introduce CardioMix, a framework built on a bidirectional CutMix strategy guided by cardiac patterns for ECG segmentation. This approach enriches the labeled set with realistic variations from unlabeled data while simultaneously applying stronger supervisory signals to the unlabeled set, as the cardiac pattern-guided mixing ensures all augmented samples remain physiologically meaningful. Our framework is designed as a plug-and-play module, demonstrating high compatibility with various SemiSeg algorithms. Extensive experiments on SemiSegECG, a public multi-dataset benchmark for ECG delineation, demonstrate that CardioMix consistently outperforms existing CutMix-based fusion strategies across diverse datasets and labeled ratios as a plug-and-play module compatible with various SemiSeg algorithms.\n\n---\n\n## 120. Position: Early-Stage Quality Assurance in Annotation Pipelines Is More Cost-Effective Than Late-Stage Validation\n- **Authors**: Sunil Kothari , Sumukha Sharma Thoppanahalli Chandramouli , Naman Khandelwal , Parth Kulshreshtha , Ashi Jain , Kriti Banka , Tanuja Chintada , Venkata Triveni , Gulipalli Praveen Kumar , Manish Mehta , Tao Liu\n- **URL**: [https://arxiv.org/abs/2605.15714](https://arxiv.org/abs/2605.15714)\n- **Abstract**:\n> This position paper argues that the machine learning community should prioritize early-stage quality assurance in annotation pipelines over the prevailing practice of late-stage validation. Data quality bottlenecks increasingly limit foundation model improvement, yet quality assurance research focuses almost exclusively on validation methods rather than validation timing. When validation occurs, not merely what methods are employed, fundamentally determines both error rates and annotation costs. This temporal neglect is puzzling given the well-established \"shift-left\" principle from software engineering, where empirical studies demonstrate 4--100x cost multipliers for defects detected in later stages (Boehm, 1981; Shull et al., 2002). Annotation pipelines exhibit analogous dynamics: errors caught before annotation begins cost a fraction of those discovered after review cycles complete. We propose a taxonomy of three QA trigger points, namely pre-annotation (T0), post-annotation (T1), and post-review (T2), that decompose annotation workflows into discrete validation opportunities. A parametric error-propagation model formalizes when timing affects final error rates versus only economics, making timing a measurable design variable rather than a configuration afterthought. A survey of 47 recent papers reveals that only 4% report when validation occurs, a striking gap given timing's demonstrated impact in adjacent fields. Without explicit attention to QA timing, the community risks optimizing validation methods while ignoring the structural variable that may matter most. Acting on this position requires three steps: researchers should report QA timing configurations alongside validation methods; annotation platforms should expose timing as a first-class parameter; and the community should run controlled experiments that measure stage-specific detection rates directly.\n\n---\n\n## 121. Learning Dynamic Pick-and-Place for a Legged Manipulator\n- **Authors**: Moonkyu Jung , Jiseong Lee , Zhengmao He , Donghoon Youm , Juhyeok Mun , HyeongJun Kim , Hyunsik Oh , Donghyuk Choi , Jungwoo Hur , Jie Song , Jemin Hwangbo\n- **URL**: [https://arxiv.org/abs/2605.15713](https://arxiv.org/abs/2605.15713)\n- **Abstract**:\n> Legged manipulators extend robotic capabilities beyond static manipulation by integrating agile locomotion with versatile arm control. However, achieving precise manipulation while maintaining coordinated locomotion remains a major challenge. This work presents a hierarchical reinforcement learning framework for dynamic pick-and-place tasks using a quadruped equipped with a 6-DOF robotic arm. The framework incorporates an explicit mass estimation module enabling adaptive whole-body control for objects with varying weights. In simulation, the system achieves an 86.05% success rate with payloads up to 2.3 kg. The approach is further validated through real-world experiments across six representative scenarios with controlled variations in object physical properties (size and mass) and task heights. Specifically, within a wide vertical workspace ranging from ground level to 1.1~m-high tabletops, the system demonstrates an average success rate of 73.3% for payloads up to 1.3 kg, with an average execution time of 4.06 s. Unlike prior works that handle lightweight objects and execute pick-and-place motions with slow, piecewise motions, the proposed framework exploits concurrent locomotion and manipulation for dynamic, continuous execution. These results demonstrate the potential of quadrupedal mobile manipulators for adaptive, whole-body pick-and-place with heavier payloads and extended workspaces.\n\n---\n\n## 122. Feedback World Model Enables Precise Guidance of Diffusion Policy\n- **Authors**: Tuo An , Jindou Jia , Gen Li , Jingliang Li , Chuhao Zhou , Pengfei Liu , Bofan Lyu , Jiaqi Bai , Xinying Guo , Geng Li , Jianfei Yang\n- **URL**: [https://arxiv.org/abs/2605.15705](https://arxiv.org/abs/2605.15705)\n- **Abstract**:\n> World models aim to improve robotic decision making by predicting the consequences of actions. However, in practice, their predictions often become unreliable once the robot encounters states outside the training distribution, limiting their effectiveness at deployment. We observe that execution itself provides a natural but underutilized signal: after each action, the robot directly observes the true next state, revealing the mismatch between predicted and actual outcomes. Building on this insight, we propose feedback world model, a new paradigm that closes the loop between prediction and observation at inference time. Instead of treating the world model as a static open-loop predictor, our method maintains a lightweight feedback state that is updated online to iteratively correct future predictions, compensating for model errors using real-time observations without additional training data or parameter updates. We show that this process can be interpreted as a latent-space observer and admits convergence guarantees under mild conditions. We further introduce action-aware guidance to better translate corrected predictions into control by emphasizing action-controllable components while suppressing irrelevant variations. Experiments on LIBERO-Plus, Robomimic, and real-world manipulation tasks demonstrate that our method substantially improves both prediction accuracy and policy performance under distribution shift. In particular, it reduces world model prediction error by up to 76.4% and improves out-of-distribution (OOD) success rate by 30%. These results show that incorporating real-time feedback at inference time provides a simple yet powerful alternative to static world modeling.\n\n---\n\n## 123. H-Mem: A Novel Memory Mechanism for Evolving and Retrieving Agent Memory via a Hybrid Structure\n- **Authors**: Jiawei Yu , Yixiang Fang , Xilin Liu , Yuchi Ma\n- **URL**: [https://arxiv.org/abs/2605.15701](https://arxiv.org/abs/2605.15701)\n- **Abstract**:\n> Memory data are ubiquitous in Large Language Model (LLM)-based agents (e.g., OpenClaw and Manus). A few recent works have attempted to exploit agents'memory for improving their performance on the question-answering (QA) task, but they lack a principled mechanism for effectively modeling how memory data evolves over time and retrieving memory data effectively, leading to poor performance in memory utilization. To fill this gap, we present H-Mem, a novel memory mechanism via a hybrid structure that can not only effectively model the evolution of agent memory over a long period of time, but also provide an efficient memory retrieval approach. Particularly, H-Mem builds a temporal and semantic tree structure that allows the short-term memory data to evolve progressively into long-term memory data, where the latter provides summarized information about the former, while simultaneously constructing a knowledge graph to capture the relationships between entities in memory. Moreover, it offers an effective memory retrieval approach by exploiting the hybrid structure of the tree and graph structures. Extensive experiments on three agent memory benchmarks show that H-Mem achieves state-of-the-art performance on the QA task.\n\n---\n\n## 124. $α$-TCAV: A Unified Framework for Testing with Concept Activation Vectors\n- **Authors**: Ekkehard Schnoor , Jawher Said , Malik Tiomoko , Wojciech Samek , Alexander Jung\n- **URL**: [https://arxiv.org/abs/2605.15688](https://arxiv.org/abs/2605.15688)\n- **Abstract**:\n> Concept Activation Vectors (CAVs) are a fundamental tool for concept-based explainability in deep learning, yet their practical utility is limited by statistical instability. We analyze the stochastic nature of CAVs and the Testing with CAVs (TCAV) method, deriving the distributions of major CAV classes including PatternCAV, FastCAV, and ridge regression-based CAVs. We then identify a fundamental flaw in the standard TCAV score: its reliance on a discontinuous indicator function induces non-decaying variance in critical regimes. To address this, we introduce $\\alpha$-TCAV, a generalized framework that replaces the indicator with a parameterized smooth function, yielding a unified probabilistic formulation that subsumes both TCAV and Multi-TCAV. We characterize the induced distributions of sensitivity scores and different TCAV variants, showing that established state-of-the-art choices lack theoretical justification. We provide principled guidance on tuning the parameter in $\\alpha$-TCAV -- either to imitate Multi-TCAV at substantially lower computational cost, or to obtain a calibrated Bayes-optimal probabilistic measure of a concept's influence. Finally, our analysis yields practical recommendations that challenge established routines: most notably, allocating the full sampling budget to a single CAV rather than splitting it across several.\n\n---\n\n## 125. ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models\n- **Authors**: Jiahui Guang , Yingjie Zhu , Cuiyun Gao , Haiyan Wang , Jing Li , Di Shao , Zhaoquan Gu\n- **URL**: [https://arxiv.org/abs/2605.15687](https://arxiv.org/abs/2605.15687)\n- **Abstract**:\n> Multimodal large language models (MLLMs) may memorize sensitive cross-modal information during pretraining, making machine unlearning (MU) crucial. Existing methods typically evaluate unlearning effectiveness based on output deviations, while overlooking the generation quality after unlearning. This can easily lead to hallucinated or rigid responses, thereby affecting the usability and safety of the unlearned model. To address this issue, we propose ASRU, a controllable multimodal unlearning framework that incorporates generation quality as a core evaluation objective. ASRU first induces initial refusal behavior through activation redirection, and then optimizes fine-grained refusal boundaries using a customized reward function, thereby achieving a better trade-off between target knowledge unlearning and model utility. Experiments on Qwen3-VL show that ASRU significantly improves unlearning effectiveness (+24.6%) on average and generation quality (5.8x) on average while effectively preserving model utility, using only a small amount of retained supervision data.\n\n---\n\n## 126. Interaction-Aware Influence Functions for Group Attribution\n- **Authors**: Jaeseung Heo , Kyeongheung Yun , Youngbin Choi , Sehyun Hwang , Jungseul Ok , Dongwoo Kim\n- **URL**: [https://arxiv.org/abs/2605.15675](https://arxiv.org/abs/2605.15675)\n- **Abstract**:\n> Influence functions approximate how removing a training example changes a quantity of interest, called the target function, such as a held-out loss. To estimate the influence of a group of examples, the standard practice is to sum the individual influences of its members. However, this sum does not capture how examples jointly affect the target: a pair of examples may be redundant or complementary, but the sum cannot distinguish these cases. We propose an interaction-aware influence function that characterizes how interactions between examples influence the target. By expanding the target to second order around the trained parameters, we obtain an estimator that augments the standard sum with a pairwise interaction term that captures the alignment between two examples' effects on the target. We empirically evaluate our estimator in two settings. First, on six dataset-model pairs spanning logistic regression, MLPs, and ResNet-9, our estimator tracks leave-group-out retraining substantially better than first-order influence across all settings. Second, when used as a greedy selection rule for instruction-tuning data on Llama-3.1-8B, it beats prior influence-based and representation-similarity baselines on five of seven downstream tasks, in a regime where standard influence-based selection underperforms random selection.\n\n---\n\n## 127. VLMs Trace Without Tracking: Diagnosing Failures in Visual Path Following\n- **Authors**: Hyesoo Hong , Minsoo Kim , Wonje Jeung , Sangyeon Yoon , Dongjae Jeon , Albert No\n- **URL**: [https://arxiv.org/abs/2605.15672](https://arxiv.org/abs/2605.15672)\n- **Abstract**:\n> Vision-language models (VLMs) achieve strong performance on multimodal benchmarks, but may still lack robust control over basic visual operations. We study \\textit{line tracing}, where a model must follow a selected visual path through successive local continuations. To isolate this ability, we design controlled tracing tasks that introduce nearby competitors while reducing semantic and topological ambiguity such as crossings and overlaps. Across these tasks, even state-of-the-art VLMs frequently lose the target path and switch to nearby alternatives, especially when those alternatives look locally similar to the target. Behavioral interventions and internal analyses indicate that these failures arise from local competition: nearby similar distractors pull the model away from the true continuation. Standard remedies do not remove this bottleneck: model-size scaling provides only limited gains, reasoning partially compensates through costly substitute strategies, and explicit tracing instructions fail to recover stable path following. Finally, tests on tangled-cable scenes and metro maps with richer visual complexity show that the same path-switching failure persists beyond our controlled settings.\n\n---\n\n## 128. VAGS: Velocity Adaptive Guidance Scale for Image Editing and Generation\n- **Authors**: Yan Luo , Ahmadou Aidara , Jingyi Lu , Jeremy Moebel , Kai Han , Mengyu Wang\n- **URL**: [https://arxiv.org/abs/2605.15661](https://arxiv.org/abs/2605.15661)\n- **Abstract**:\n> Classifier-free guidance (CFG) is the primary control over how strongly text semantics move a flow-based sampler, yet standard practice holds its scale fixed across the entire ODE trajectory. This is a fundamental mismatch: early steps are noise-dominated and carry weak semantic signal, while late steps commit image structure and demand stronger directional commitment; more critically, the value of any guidance strength depends on whether the guided velocity is consistent with the model's current dynamics or working against them. We propose \\textit{Velocity-Adaptive Guidance Scale} (VAGS), a training-free replacement that multiplies the nominal scale by a bounded factor combining a temporal signal-level term with the cosine similarity between task-relevant velocity fields. For inversion-free editing, VAGS measures the alignment between source- and target-guided velocities, so edit strength at each step reflects local compatibility between preservation and transformation. For generation, VAGS-Gen uses the alignment between unconditional and conditional velocities as the analogous signal. Neither variant requires fine-tuning, auxiliary networks, or extra forward passes, and fixed CFG is recovered as a special case. On PIE-Bench and DIV2K for editing, and COCO17, CUB-200, and Flickr30K for generation, VAGS consistently improves structural fidelity and generation quality over fixed CFG and recent training-free guidance variants. The code is publicly available at this https URL .\n\n---\n\n## 129. TFZ-Tree: An Ultra-Lightweight Waveform Classification Framework for Resource-Constrained Devices\n- **Authors**: Hao Wang , Kuang Zhang , Yonggang Chi , Tianqi Zhao , Yanbo Fu , Jiaxing Guo\n- **URL**: [https://arxiv.org/abs/2605.15656](https://arxiv.org/abs/2605.15656)\n- **Abstract**:\n> Under the trend of multi-waveform coexistence in 6G IoT, intelligent receivers must first identify physical-layer waveform types before performing correct demodulation and resource scheduling. However, existing signal identification research largely focuses on symbol-level modulation classification. Research directly targeting physical-layer waveform types (e.g., OFDM, OTFS, LoRa) is not only extremely scarce but also heavily reliant on deep neural networks and complex time-frequency transforms, making deployment on resource-constrained terminals difficult. Symbol modulation classification methods themselves cannot circumvent the prerequisite of ``waveform identification first.'' To address this dual gap, we propose an ultra-lightweight waveform classification framework based on time-frequency multidimensional features with a cooperative Z-test tree (ZTree). The framework employs low-complexity time-domain feature extraction, and the classification backend adopts a ZTree optimized by Z-statistical testing, which uses hypothesis testing confidence to automatically control decision tree splitting and size, ensuring efficient execution on resource-limited processors. Tested on ten 6G candidate waveforms including OFDM, OTFS, DSSS, LoRa, and NB-IoT, the method achieves 99.5\\% average accuracy under AWGN and 87.4\\% under TDL-C multipath channels, with main confusion between OTFS and LoRa. Implemented in C on an x86 platform, single inference latency is under 4~ms. To the best of our knowledge, this is the first work achieving real-time recognition of ten IoT waveform types. Future work will target deployment acceleration on embedded MCUs. Code and dataset are open-sourced at: this https URL .\n\n---\n\n## 130. Bridging Silicon and the Hippocampus: Algebro-Deterministic Memory \"VaCoAl\" as a Substrate for Vector-HaSH and TEM\n- **Authors**: Hiroyuki Chuma , Kanji Otsuka , Yoichi Sato\n- **URL**: [https://arxiv.org/abs/2605.15652](https://arxiv.org/abs/2605.15652)\n- **Abstract**:\n> Vector-HaSH and the Tolman-Eichenbaum Machine (TEM) propose that the hippocampal-entorhinal circuit factorizes content from a prestructured grid-cell scaffold and supports compositional memory via ripple-mediated replay. Human iEEG shows that hippocampal sharp-wave ripples (SWRs) gate episodic recall, ripple-locked cortical reactivation recapitulates encoding-time patterns, and multi-hop replay fidelity decays multiplicatively along sequence length. These literatures have advanced in parallel without a shared algebraic object. We show that VaCoAl, an algebro-deterministic hyperdimensional memory architecture built from Galois-field LFSRs, supplies that object. Specifically, deterministic Galois-field diffusion provides a substrate-level alternative to Vector-HaSH's random scaffold-to-hippocampus projection that satisfies the same quasi-orthogonality requirement, with matched second-moment statistics, stronger avalanche behavior, and bit-exact reproducibility. The path-integral Confidence Ratio $CR_2$, the product of per-step $CR_1$ values along an $n$-hop chain, is the natural functional form for multi-hop replay-fidelity decay under conditional independence of per-step reactivation, providing the first algebraically tractable model of reported multiplicative decay. STDP-like path selection in VaCoAl follows from architectural demands -- similarity preservation, compositional reversibility, and bounded-frontier search -- that also constrain hippocampal computation. We further argue that VaCoAl operating regimes share architectural commitments with the EC--CA3 and EC--DG--CA3 pathways, motivating an energy-capacity-plasticity reading of why both are conserved across $>$520 Myr of evolution and primate dentate-gyrus elaboration. We prove formal correspondences, derive testable iEEG predictions, and bridge computational neuroscience and hyperdimensional engineering.\n\n---\n\n## 131. Sharp Spectral Thresholds for Logit Fixed Points\n- **Authors**: Tongxi Wang\n- **URL**: [https://arxiv.org/abs/2605.15651](https://arxiv.org/abs/2605.15651)\n- **Abstract**:\n> Softmax feedback systems are a common mathematical core of entropy-regularized reinforcement learning, logit game dynamics, population choice, and mean-field variational updates. Their central stability question is simple: when does a self-reinforcing softmax system produce a unique and globally predictable outcome? Classical theory gives a conservative answer. By treating softmax as a unit-scale response, it certifies stability only in a strongly randomized regime. We prove that the classical approach misses an entire stable regime and does not identify the point at which the qualitative change truly occurs. For finite-dimensional affine logit systems, the sharp dimension-free Euclidean threshold is $$\\beta\\|\\Pi W\\Pi\\|_{\\mathcal T\\to\\mathcal T}<2,$$ rather than the previously used condition, which certifies stability only while the softmax system remains safely over-regularized. Our theorem fills the previously missing pre-bifurcation regime, extending stability guarantees for affine softmax feedback systems to reward-responsive yet globally predictable systems. It enlarges the certified stability boundary for these systems and identifies where the model genuinely undergoes a phase transition.\n\n---\n\n## 132. Latent Video Prediction Learns Better World Models\n- **Authors**: Ali J Alrasheed , Aryan Yazdan Parast , Basim Azam , James Bailey , Naveed Akhtar\n- **URL**: [https://arxiv.org/abs/2605.15618](https://arxiv.org/abs/2605.15618)\n- **Abstract**:\n> Self-supervised video models are increasingly framed as world models, yet their evaluation remains largely confined to a single top-1 accuracy score on clean benchmarks. This leaves a major gap in comprehending their potential as world models. We present the first systematic study addressing this gap, analyzing four matched-capacity frontier video foundation models, V-JEPA 2.1, V-JEPA 2, VideoPrism, and VideoMAEv2, across five robustness axes relevant to their deployment as video world models: feature discriminability, corruption robustness, fine-grained discrimination, occlusion robustness, and sensitivity to temporal direction. Our evaluations establish that across all five axes, latent-prediction models form a distinct and consistent profile. They degrade more gracefully under pixel corruption, preserve usable class structure rather than mere geometric stability under occlusion, capture fine-grained physical contact cues without reconstructing pixels, and uniquely encode the arrow of time. These advantages can even survive task adaptation: a frozen V-JEPA 2 backbone with a lightweight attentive probe outperforms a fully fine-tuned VideoMAE and a supervised TimeSformer on corruption and occlusion robustness. Our extensive results offer concrete new evidence in favor of latent prediction for robust world modeling.\n\n---\n\n## 133. A Few GPUs, A Whole Lotta Scale: Faithful LLM Training Emulation with PrismLLM\n- **Authors**: Shaoke Xi , ChonLam Lao , Boyi Jia , Jiaqi Gao , Zhipeng Zhang , Jiamin Cao , Brian Sutioso , Erci Xu , Minlan Yu , Kui Ren , Yong Li , Zhengping Qian , Ennan Zhai , Jingren Zhou\n- **URL**: [https://arxiv.org/abs/2605.15617](https://arxiv.org/abs/2605.15617)\n- **Abstract**:\n> Large language model (LLM) training today runs on clusters spanning thousands of GPUs. While this scale enables rapid model advances, developing, debugging, and performance-tuning the training framework inevitably becomes complex and costly. This is because engineers often need to reproduce production behaviors to diagnose failures or evaluate optimizations, thereby demanding frequent and even exclusive access to production-scale clusters -- which becomes increasingly hard given that the majority of GPUs are already committed to production workloads. Simulation relies on complex performance models that are difficult to maintain, and downscaled experiments often fail to capture scale-dependent behaviors. We present PrismLLM to decouple large-scale execution from the need to access large clusters, enabling engineers to run and observe ranks of interest under faithful large-scale behavior using only a few GPUs. PrismLLM constructs a high-fidelity execution graph via a slicing-based approach that captures computation, communication, and dependencies of the target scale. Then, PrismLLM performs hybrid emulation where selected ranks execute the original program while the remaining ranks are replayed as virtual participants. Experiments on large-scale LLM training workloads show that PrismLLM accurately reproduces performance and memory behavior, achieving only 0.58\\% average error in iteration time and less than 0.01\\% error in peak GPU memory usage. PrismLLM can emulate clusters of up to 8192 GPUs using fewer than 1\\% of the physical GPUs required by the original deployment.\n\n---\n\n## 134. Offline Reinforcement Learning with Universal Horizon Models\n- **Authors**: Hojun Chung , Junseo Lee , Songhwai Oh\n- **URL**: [https://arxiv.org/abs/2605.15603](https://arxiv.org/abs/2605.15603)\n- **Abstract**:\n> Model-based reinforcement learning (RL) offers a compelling approach to offline RL by enabling value learning on imagined on-policy trajectories. However, it often suffers from compounding errors due to repeated model inference on self-generated states. While geometric horizon models (GHM) alleviate this issue through direct prediction over a discounted infinite-horizon future, they remain challenged in accurately modeling distant future states. To this end, we introduce universal horizon models (UHM), a generalization of GHM that directly predicts future states under arbitrary horizons. Leveraging this flexibility, we propose a scalable value learning method that employs a winsorized horizon distribution to stabilize training by capping excessively large horizons. Experimental results on 100 challenging OGBench tasks demonstrate that the proposed method outperforms competitive baselines, particularly on tasks with highly suboptimal datasets and those requiring long-horizon reasoning. Project page: this https URL\n\n---\n\n## 135. Pretraining Objective Matters in Extreme Low-Data FGVC: A Backbone-Controlled Study\n- **Authors**: Alexander Hackett , Srikanth Thudumu , Ginny Fisher , Mahule Roy , Aisha Sartaj , Jason Fisher\n- **URL**: [https://arxiv.org/abs/2605.15599](https://arxiv.org/abs/2605.15599)\n- **Abstract**:\n> Extreme low-data fine-grained classification is common in expert domains where labeling is expensive, yet practitioners still need principled guidance for selecting pretrained encoders. We study emerald inclusion grading with a custom dataset of labeled images across three classes and ask: under matched backbone capacity, how does pretraining objective affect downstream representation quality? We compare four frozen ViT-B/16 encoders trained with supervised classification, contrastive learning (SigLIP2), masked reconstruction (MAE), and self-distillation (DINOv3), and evaluate them with leave-one-out cross-validation using linear and nonlinear probes. To control statistical noise in the low-N regime, we use permutation testing (N=1000) on macro one-vs-rest AUC. Supervised and contrastive encoders provide the strongest linear separability (logistic AUC: 0.768 and 0.735; SVM AUC: 0.739 and 0.697), while MAE improves under nonlinear probes (XGBoost AUC: 0.713). We find that DINOv3 underperforms across probe families in this domain. These results support a practical recommendation for extreme low-data FGVC: prioritize margin-enforcing pretraining objectives when data scarcity restricts probing to linear decision rules, and consider reconstruction-style encoders when nonlinear classifiers are feasible given dataset constraints.\n\n---\n\n## 136. Embracing Biased Transition Matrices for Complementary-Label Learning with Many Classes\n- **Authors**: Tan-Ha Mai , Chao-Kai Chiang , Han-Hwa Shih , Gang Niu , Masashi Sugiyama , Hsuan-Tien Lin\n- **URL**: [https://arxiv.org/abs/2605.15586](https://arxiv.org/abs/2605.15586)\n- **Abstract**:\n> Complementary-label learning (CLL) is a weakly supervised paradigm where instances are labeled with classes they do not belong to. Despite a decade of research, CLL methods remain competitive mainly on 10-class classification, with scaling to large label spaces continuing to be an enduring bottleneck. This limitation stems from the common assumption of uniform label generation in traditional methods, which fatally dilutes the learning signal in many-class settings. In this paper, we demonstrate that this long-standing barrier can be overcome by deliberately designing a biased (non-uniform) generation process that restricts complementary labels to a subset of classes. This finding motivates us to propose Bias-Induced Constrained Labeling (BICL), a principled framework spanning data collection to training that leverages this bias. BICL enables effective learning on CIFAR-100 and TinyImageNet-200, achieving more than sevenfold accuracy improvements over traditional methods. Our findings establish a new trajectory for making CLL feasible for many classes in real-world applications.\n\n---\n\n## 137. Detecting Privilege Escalation in Polyglot Microservices via Agentic Program Analysis\n- **Authors**: Penghui Li , Hong Yau Chong , Yinzhi Cao , Junfeng Yang\n- **URL**: [https://arxiv.org/abs/2605.15569](https://arxiv.org/abs/2605.15569)\n- **Abstract**:\n> Microservices are widely adopted in modern cloud systems due to their scalability and fault tolerance. However, microservice architectures introduce significant complexity in privilege and permission control, creating risks of privilege escalation where attackers can gain unauthorized access to resources or operations. Detecting such vulnerabilities is challenging due to complex cross-service interactions, polyglot codebases, and diverse privileged operations and permission checks. We present Neo, an agentic program analysis framework that combines large language models (LLMs) with classic program analysis to address these challenges. Neo leverages an LLM-based agent that dynamically generates analysis plans, adapts code search strategies, and validates semantics. We develop code search primitives that enable Neo to perform scalable and flexible code exploration across services and languages. We evaluated Neo on 25 open-source microservice applications spanning 7 programming languages and 6.2 million lines of code. Neo uncovered 24 zero-day privilege escalation vulnerabilities and achieved 81.0% precision and 85.0% recall on a ground-truth dataset. Compared to existing program analysis and agentic solutions, Neo demonstrated significant improvements in both detection accuracy and scalability. We further showcased Neo's extensibility by applying it to other application domains and vulnerability types, uncovering 18 additional zero-day vulnerabilities.\n\n---\n\n## 138. AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs\n- **Authors**: Haizhong Zheng , Yizhuo Di , Jiahui Wang , Shuowei Jin , Xueshen Liu , Yongji Wu , Z. Morley Mao , Ion Stoica , Jiawei Zhao , Beidi Chen\n- **URL**: [https://arxiv.org/abs/2605.15565](https://arxiv.org/abs/2605.15565)\n- **Abstract**:\n> Reinforcement learning (RL) is increasingly used to improve the reasoning, coding, and tool-use capabilities of large language models, but agentic RL remains prohibitively expensive. Scaling RL to agentic LLMs requires supporting complex workloads, including multi-policy collaborative training, while efficiently using elastic, heterogeneous, and cross-region compute resources. Existing LLM RL systems support some of these capabilities, but each new extension often requires dedicated system engineering. This burden arises from trainer-centered control architectures and the lack of principled abstractions for RL system components. To address these limitations, we propose AstraFlow, a dataflow-oriented RL system that replaces conventional trainer-centered control with principled component abstractions. In AstraFlow, rollout services, dataflow management, and training are decoupled into autonomous components, enabling the system to natively support complex multi-policy agentic RL workloads and efficiently exploit diverse compute resources. We evaluate AstraFlow across math, code, search, and AgentBench workloads, showing that the same system supports multi-policy training, elastic scaling, heterogeneous cross-region execution, and composable data algorithms without system-level code changes. In multi-policy collaborative training, AstraFlow achieves comparable or better accuracy than existing RL systems while speeding up training time by 2.7x.\n\n---\n\n## 139. Operator-Controlled 6G: From Connectivity Infrastructure to Guaranteed Digital Services\n- **Authors**: David Soldani\n- **URL**: [https://arxiv.org/abs/2605.15553](https://arxiv.org/abs/2605.15553)\n- **Abstract**:\n> Sixth-generation mobile networks (6G) are approaching a structural inflection point. Five generations of vendor-led architectures have left operators procuring and operating networks they do not own, on platforms they cannot modify, with AI layers they cannot audit. This paper argues that 6G must reverse this trajectory by reordering operator priorities: Control First, Customer First, Business First, Operations First, and Technology Last. Technology should serve operator control, customer outcomes, monetizable guarantees, and software-driven operations, not dictate this http URL contributions operationalize this thesis. The 6G Control Compact defines a three-layer ownership taxonomy--own, federate, and consume--that allocates architectural sovereignty according to strategic value. The Guarantee Economy defines a five-tier, outcome-priced commercial model that converts operator control into enforceable service-level objectives. The framework is grounded in operational evidence from Rakuten Mobile, the world's first national-scale, fully cloud-native, fully Open RAN deployment, which reached full-year EBITDA profitability in FY2025. It is aligned with the ITU-R IMT-2030 framework, 3GPP 6G use cases and service requirements, NGMN recommendations, ETSI standards, O-RAN Alliance and AI-RAN Alliance specifications, IOWN Global Forum sustainability metrics, Linux Foundation initiatives, and leading industry and academic programs. A three-phase roadmap covering 2025-2027, 2027-2029, and 2029-2032 and beyond, together with seven stakeholder-specific calls to action, translates the architecture into industry commitments. The central claim is that Rakuten Mobile's deployment demonstrates the feasibility of operator-controlled 6G. Decisions made during 2026-2028 will determine whether 6G becomes a platform for guaranteed digital services or another vendor-dependent infrastructure cycle.\n\n---\n\n## 140. CTF4Nuclear: Common Task Framework for Nuclear Fission and Fusion Models\n- **Authors**: Stefano Riva , Carolina Introini , Antonio Cammi , Dean Price , Alexey Yermakov , Yue Zhao , Philippe M. Wyder , Judah Goldfeder , Jan Williams , Amy Sara Rude , Matteo Tomasetto , Joe Germany , Joseph Bakarji , Georg Maierhofer , Miles Cranmer , J. Nathan Kutz\n- **URL**: [https://arxiv.org/abs/2605.15549](https://arxiv.org/abs/2605.15549)\n- **Abstract**:\n> The demand for clean energy is ever increasing, with new nuclear technologies presenting a complementary solution to renewable energies. However, designing and operating these systems is exceptionally difficult, given the complexity of the physical phenomena that interact to form the system dynamics. While high-fidelity simulations help to understand the non-linear, multi-physics interactions within a reactor, they are computationally expensive and rarely suitable for real-time applications. Furthermore, model-based approaches are inherently sensitive to simplifying assumptions required to derive their governing equations and parameters, leading to inevitable discrepancies with real-world measurements. In contrast, Machine Learning (ML) methods have the potential to generate reliable surrogate models which may be able to quickly predict the system's behaviour. However, the number of data-driven methods that can potentially be used for this task is large and diverse. In a safety-critical setting such as nuclear engineering, a fair comparison of different ML methods, and a clear understanding of their advantages and limitations, is of paramount importance. To address this, we introduce a Common Task Framework (CTF) for ML in nuclear engineering, building upon previous efforts in dynamical systems and seismology. This CTF considers a curated set of datasets from different nuclear and nuclear-adjacent systems. The CTF evaluates the performance of a method on 12 established metrics, alongside a new paradigm focused on system monitoring from sparse measurements only. We illustrate the framework by benchmarking standard ML baselines against these datasets, revealing current method limitations. Our vision is to replace ad hoc comparisons with standardized evaluations on hidden test sets, raising the bar for rigour and reproducibility in scientific ML for the nuclear industry.\n\n---\n\n## 141. Domain-Independent Game Abstraction using Word Embedding Techniques\n- **Authors**: Juho Kim , Tuomas Sandholm\n- **URL**: [https://arxiv.org/abs/2605.15543](https://arxiv.org/abs/2605.15543)\n- **Abstract**:\n> Many games of interest in the real world are often intractably large, thereby necessitating the use of game abstraction to shrink them in size, typically by many magnitudes. Over the last two decades, there have been significant advances in game abstraction; however, the domain-specific nature (usually poker) of much of the prior work prevents those techniques from being easily generalized to other settings without extensively analyzing the game at hand. In this paper, we propose a domain-independent approach to game abstraction, which applies word embedding techniques from the field of natural language processing. Treating each action as a word and gameplay data as a corpus, word vectors can be trained to represent each action as a real-valued vector, which can then be clustered to facilitate game abstraction. We also explore the use of foundational embedding models and show that action embeddings obtained this way can capture a surprising amount of information about the underlying game. Experimental results demonstrate that our proposed game abstraction technique is effective, although it does not outperform specialized algorithms tailored to specific games.\n\n---\n\n## 142. SkiP: When to Skip and When to Refine for Efficient Robot Manipulation\n- **Authors**: Mingtong Dai , Guanqi Peng , Yongjie Bai , Feng Yan , Chunjie Chen , Lingbo Liu , Liang Lin , Xinyu Wu\n- **URL**: [https://arxiv.org/abs/2605.15536](https://arxiv.org/abs/2605.15536)\n- **Abstract**:\n> Previous imitation learning policies predict future actions at every control step, whether in smooth motion phases or precise, contact-rich operation phases. This uniform treatment is wasteful: most steps in a manipulation trajectory traverse free space and carry little task-relevant information, while a small fraction of \\emph{key} steps around contacts, grasps, and alignment demand dense, high-resolution prediction. We propose a novel \\emph{action relabeling} mechanism: at each timestep in a skip segment, we replace the behavior cloning target with the action at the entrance of the next key segment, enabling the policy to leap over redundant steps in a single decision. The resulting \\textbf{Skip Policy (SkiP)} dynamically leaps over skip segments and intensively refines actions in key segments, within a single unified network requiring no learned skip planner or hierarchical structure. To automatically partition demonstrations into key and skip segments without manual annotation, we introduce \\emph{Motion Spectrum Keying} (MSK), a fast, task-agnostic procedure that detects local motion complexity from action signals. Extensive experiments across 72 simulated manipulation tasks and three real-robot tasks show that SkiP reduces executed steps by $15$--$40\\%$ while matching or improving success rates across various policy backbones. Project page: \\texttt{ this https URL }.\n\n---\n\n## 143. Tuning-free Instruction-based Video Editing Via Structural Noise Initialization and Guidance\n- **Authors**: Song Wu , Xinyu Chen , Qian Wang , Liang Li , Zili Yi , Junlan Feng\n- **URL**: [https://arxiv.org/abs/2605.15533](https://arxiv.org/abs/2605.15533)\n- **Abstract**:\n> Video editing poses a significant challenge. While a series of tuning-free methods circumvent the need for extensive data collection and model training, they often underutilize the rich information embedded within noisy latent, leading to unsatisfactory results. To address this, we propose a \\textit{tuning-free, instruction-based} video editing framework. We approach video editing from the perspective of noisy latent: we design a Structural Noise Initialization Strategy (SNIS) to secure a superior editing starting point by assigning higher noise levels to edited regions (to facilitate content change) and lower noise levels to unedited regions (to maintain content consistency). We introduce a Noise Guidance Mechanism (NGM), which leverages the video prior in the generative model and effectively integrates rich information within the noisy latent to guide the denoising process, thereby preserving unedited content and overall visual coherence. Experiments show that our proposed method achieves better visual quality and state-of-the-art performance.\n\n---\n\n## 144. DeltaPrompts: Escaping the Zero-Delta Trap in Multimodal Distillation\n- **Authors**: Jaehun Jung , Hyunwoo Kim , Brandon Cui , Ximing Lu , David Acuna , Prithviraj Ammanabrolu , Yejin Choi\n- **URL**: [https://arxiv.org/abs/2605.15532](https://arxiv.org/abs/2605.15532)\n- **Abstract**:\n> Distillation enables compact Vision-Language Models (VLMs) to obtain strong reasoning capabilities, yet the prompts driving this process are typically chosen via simple heuristics or aggregated from off-the-shelf datasets. We reveal a critical inefficiency in this approach: up to 69% of the prompts in standard chart / document reasoning datasets are effectively zero-delta, meaning the teacher and student already induce the exact same answer distribution. Training on these prompts provides minimal learning signal, causing student improvement to rapidly saturate regardless of data scale. To escape the zero-delta trap, we return to first principles: distillation fundamentally minimizes distributional divergence, and thus a prompt is valuable only if it exposes a functional capability gap between the teacher and student. We quantify this gap through answer divergence ($\\Delta$), demonstrating that non-zero divergence is critical for effective scaling. Building on this insight, we propose a staged synthesis pipeline that repurposes existing datasets as seeds, actively targeting student failure modes to produce better prompts. The result is DeltaPrompts, a diverse dataset of 200k synthetic, high-divergence reasoning problems. We evaluate DeltaPrompts across three distinct settings: on-policy distillation with the target teacher-student pair, transfer to a novel model family without regenerating the data, and off-policy fine-tuning of a non-reasoning model. Across all scenarios, DeltaPrompts drives substantial gains, yielding up to 15% relative improvement even on top of a highly-optimized reasoning model (e.g., Qwen3-VL-8B-Thinking) -- averaged over 10 benchmarks spanning chart, document and perception-centric reasoning.\n\n---\n\n## 145. Process Rewards with Learned Reliability\n- **Authors**: Jinyuan Li , Langlin Huang , Chengsong Huang , Shaoyang Xu , Donghong Cai , Yuyi Yang , Wenxuan Zhang , Jiaxin Huang\n- **URL**: [https://arxiv.org/abs/2605.15529](https://arxiv.org/abs/2605.15529)\n- **Abstract**:\n> Process Reward Models (PRMs) provide step-level feedback for reasoning, but current PRMs usually output only a single reward score for each step. Downstream methods must therefore treat imperfect step-level reward predictions as reliable decision signals, with no indication of when these predictions should be trusted. We propose BetaPRM, a distributional PRM that predicts both a step-level success probability and the reliability of that prediction. Given step-success supervision from Monte Carlo continuations, BetaPRM learns a Beta belief that explains the observed number of successful continuations through a Beta-Binomial likelihood, rather than regressing to the finite-sample success ratio as a point target. This learned reliability signal indicates when a step reward should be trusted, enabling downstream applications to distinguish reliable rewards from uncertain ones. As one application, we introduce Adaptive Computation Allocation (ACA) for PRM-guided Best-of-N reasoning. ACA uses the learned reliability signal to stop when a high-reward solution is reliable and to spend additional computation on uncertain candidate prefixes. Experiments across four backbones and four reasoning benchmarks show that BetaPRM improves PRM-guided Best-of-N selection while preserving standard step-level error detection. Built on this signal, ACA improves the accuracy--token tradeoff over fixed-budget Best-of-16, reducing token usage by up to 33.57% while improving final-answer accuracy.\n\n---\n\n## 146. Neural Point-Forms\n- **Authors**: Bruno Trentini , Jacob Hume , Vincenzo Antonio Isoldi , Philipp Misof , Ekaterina S. Ivshina , Kelly Maggs\n- **URL**: [https://arxiv.org/abs/2605.15524](https://arxiv.org/abs/2605.15524)\n- **Abstract**:\n> Point cloud learning often rests on the premise that observed samples are noisy traces of an underlying geometric object, such as a manifold embedded in a high-dimensional feature space. Yet much of this geometry is not captured directly by coordinates, pairwise distances, or learned graph neighborhoods alone. In the smooth setting, differential forms are devices to encode higher order tangency information. In this work, we introduce a new family of principled learnable geometric features for point clouds called neural point-forms (NPFs). In the absence of a natural tangency structure, we instead use Laplacian-based techniques from Diffusion Geometry to build a discrete model for comparing differential forms on point clouds via inner products. In the continuum, submanifolds of a shared ambient feature space are represented as comparison matrices, whose entries describe how pairs of feature forms interact with extrinsic tangency information. We make this intuition precise by proving the long-run consistency of comparison matrices under standard sampling, bandwidth, density, and manifold-hypothesis assumptions. This yields a compact, efficient and permutation-invariant neural layer whose output is a learned form-comparison matrix. Across synthetic and biologically relevant experiments, we show that NPFs provide a competitive, and interpretable representation, with the strongest benefits appearing when labels depend on sampling density, manifold-like structure, or response-relevant population geometry.\n\n---\n\n## 147. On the Fragility of Data Attribution When Learning Is Distributed\n- **Authors**: Xian Gao , Bo Hui , Min-Te Sun , Wei-Shinn Ku\n- **URL**: [https://arxiv.org/abs/2605.15520](https://arxiv.org/abs/2605.15520)\n- **Abstract**:\n> Data attribution has become an important component of pricing, auditing, and governance in machine learning pipelines, yet most attribution methods implicitly assume that attribution values faithfully reflect participants' contributions. We show that this assumption can fail: a single participant in a standard distributed training workflow can substantially inflate its measured attribution value while preserving global utility. Our attribution-first attack uses latent optimization to inject small synthetic batches that preserve utility while exploiting non-IID label coverage and evaluator sensitivities. Across datasets, models, and multiple marginal-utility evaluators, the attack consistently increases the adversary's attribution value and reshapes the relative attribution structure among benign clients without degrading accuracy or triggering geometry-based defenses. These results show that attribution itself forms a new attack surface and motivate the development of attribution-robust and incentive-compatible scoring mechanisms.\n\n---\n\n## 148. DiffVAS: Diffusion-Guided Visual Active Search in Partially Observable Environments\n- **Authors**: Anindya Sarkar , Srikumar Sastry , Aleksis Pirinen , Nathan Jacobs , Yevgeniy Vorobeychik\n- **URL**: [https://arxiv.org/abs/2605.15519](https://arxiv.org/abs/2605.15519)\n- **Abstract**:\n> Visual active search (VAS) has been introduced as a modeling framework that leverages visual cues to direct aerial (e.g., UAV-based) exploration and pinpoint areas of interest within extensive geospatial regions. Potential applications of VAS include detecting hotspots for rare wildlife poaching, aiding search-and-rescue missions, and uncovering illegal trafficking of weapons, among other uses. Previous VAS approaches assume that the entire search space is known upfront, which is often unrealistic due to constraints such as a restricted field of view and high acquisition costs, and they typically learn policies tailored to specific target objects, which limits their ability to search for multiple target categories simultaneously. In this work, we propose DiffVAS, a target-conditioned policy that searches for diverse objects simultaneously according to task requirements in partially observable environments, which advances the deployment of visual active search policies in real-world applications. DiffVAS leverages a diffusion model to reconstruct the entire geospatial area from sequentially observed partial glimpses, which enables a target-conditioned reinforcement learning-based planning module to effectively reason and guide subsequent search steps. Extensive experiments demonstrate that DiffVAS excels in searching diverse objects in partially observable environments, significantly surpassing state-of-the-art methods on several datasets.\n\n---\n\n## 149. RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably\n- **Authors**: Yufeng Du , Phillip Harris , Minyang Tian , Eliu A Huerta , Srikanth Ronanki , Subendhu Rongali , Aram Galstyan , Hao Peng\n- **URL**: [https://arxiv.org/abs/2605.15514](https://arxiv.org/abs/2605.15514)\n- **Abstract**:\n> We identify intrinsic limitations of Rotary Positional Embeddings (RoPE) in Transformer-based long-context language models. Our theoretical analysis abstracts away from the specific content of the context and depends only on its length. We prove that as context length increases, RoPE-based attention becomes unpredictable and loses two properties that are central to its effectiveness. First, it loses its locality bias: RoPE is no more likely to favor nearer positions than substantially farther ones. Second, it loses consistency in token relevance: a key vector that receives a higher attention score than an alternative at one position may receive a lower score at another. In both cases, the probability of failure approaches 0.5, no better than random guessing. We further prove that the attention score can remain unchanged when a key token is moved to a different position, or even replaced by a different token, indicating a failure to distinguish positions or tokens. Adjusting the RoPE base trades off distinguishing positions against distinguishing tokens but cannot preserve both at the same time. Increasing the RoPE base hyperparameter, a common practice in today's long-context models, helps distinguish different tokens, but inevitably sacrifices the ability to distinguish positions. Our empirical analysis shows that multi-head, multi-layer architectures are insufficient to overcome these limitations. Our findings suggest that fundamentally new mechanisms for encoding position and token order may be needed in future Transformer long-context language models.\n\n---\n\n## 150. PrismQuant: Rate-Distortion-Optimal Vector Quantization for Gaussian-Mixture Sources\n- **Authors**: Bumsu Park , Chanho Park , Youngmok Park , Namyoon Lee\n- **URL**: [https://arxiv.org/abs/2605.15507](https://arxiv.org/abs/2605.15507)\n- **Abstract**:\n> For a Gaussian source under mean-squared error (MSE), classical transform coding is rate--distortion (RD) optimal: the Karhunen--Loeve transform (KLT) diagonalizes the covariance, reverse waterfilling allocates the bits, and scalar quantization closes the loop. This elegant story breaks down for multimodal sources, where no single covariance can capture heterogeneous local geometries, and the RD function loses its closed form. We revisit this problem through Gaussian-mixture sources and develop a constructive RD theory for them. Our key finding is that the mixture structure incurs only a component label cost. Conditioned on the active mixture component, each branch is Gaussian; the challenge is allocating bits across heterogeneous branches. We prove that the genie-aided conditional RD function is governed by a single global reverse-waterfilling level shared across all components and eigenmodes. Building on this result, we introduce PrismQuant, which transmits the component label losslessly and encodes the residual using the component-matched KLT, followed by scalar quantization, achieving a rate of H(C)/n bits per source dimension of the converse, with a vanishing asymptotic gap. We further develop a practical implementation based on EM-driven Gaussian-mixture learning, component-adaptive KLTs, and entropy-constrained scalar quantization (ECSQ). Experiments on synthetic Gaussian mixtures show that PrismQuant closely approaches the theoretical RD bound, while experiments on real-world channel-state-information (CSI) data demonstrate competitive or superior performance compared with transformer-based learned codecs at more than one order of magnitude smaller model size.\n\n---\n\n## 151. Learning with Conflicts of Interest\n- **Authors**: Nischal Aryal , Arash Termehchy , Ali Vakilian , Marianne Winslett\n- **URL**: [https://arxiv.org/abs/2605.15504](https://arxiv.org/abs/2605.15504)\n- **Abstract**:\n> Financial, social, and political factors often prevent the interests of the owners of ML systems and services and their users from being perfectly aligned. ML systems often produce biased information that can influence users to make decisions that are not in their best interest. Current solution approaches require ML systems to implement protocols to mitigate their biases. However, ML system owners usually do not have any incentive to implement these protocols and often argue that it limits their freedom of expression or business. We believe that a successful solution to this problem must recognize the conflict of interest between the ML systems and their users, and use this information to protect users against information that adversely influences their decisions while allowing users to safely benefit from these systems. To this end, we propose a game-theoretic framework that models the interaction between ML systems and users with conflicts of interest. We present scalable algorithms with theoretical guarantees that maximize the amount of desired information and actions and minimize the amount of biased and manipulative actions in interaction with ML systems.\n\n---\n\n## 152. Ghosted Layers: Unconstrained Activation Alignment for Recovering Layer-Pruned LLMs\n- **Authors**: Vincent-Daniel Yun , Junhyuk Jo , Sai Praneeth Karimireddy , Sunwoo Lee\n- **URL**: [https://arxiv.org/abs/2605.15491](https://arxiv.org/abs/2605.15491)\n- **Abstract**:\n> Layer pruning removes entire Transformer decoder blocks from large language models, but introduces a mismatch between the hidden state received by the next surviving layer and the distribution it was trained to process, leading to significant performance degradation. We propose Ghosted Layers, a training-free recovery module that addresses this issue by solving a boundary activation alignment problem. Our method derives a closed-form optimal linear operator from a small calibration set to reconstruct the activation discrepancy introduced by the pruned layers. We show that this solution corresponds to the unconstrained optimum of the alignment objective, whereas existing methods are restricted to constrained solutions over limited operator subspaces. Experiments across multiple LLM backbones and pruning strategies demonstrate that our method consistently improves accuracy and perplexity over prior training-free baselines, while preserving the efficiency gains of layer pruning.\n\n---\n\n## 153. Hybrid LLM-based Intelligent Framework for Robot Task Scheduling\n- **Authors**: Swayamjit Saha , Subhabrata Das , Haonan Duan , Xiao-Yang Liu\n- **URL**: [https://arxiv.org/abs/2605.15486](https://arxiv.org/abs/2605.15486)\n- **Abstract**:\n> This study introduces intelligent frameworks that use Large Language Models (LLMs) to improve task scheduling for construction robots. The LLM is fed with key data about the desired task, such as agent action abilities, and the desired end goal to be achieved. A well-balanced allocation strategy is developed, optimizing both time efficiency and resource utilization. Our system utilizes a Natural Language Processing interface to streamline communication with construction professionals and adapt in real-time to unexpected site conditions. We concurrently use two LLM agents, specifically generator (GPT-4) and supervisor (Gemma 3/Llama 4/Mistral 7b) LLM agents to provide a more precise task schedule. We evaluate the proposed methodology using a straightforward scenario and provide metric scores to prove the efficacy of the frameworks. Our results highlight that the implementation of LLMs is crucial in construction operational tasks including robots.\n\n---\n\n## 154. Residual Reinforcement Learning for Robot Teleoperation under Stochastic Delays\n- **Authors**: Kaize Deng , Zewen Yang\n- **URL**: [https://arxiv.org/abs/2605.15480](https://arxiv.org/abs/2605.15480)\n- **Abstract**:\n> Stochastic communication delays in teleoperation introduce signal discontinuities that undermine control stability and degrade control performance. Consequently, the conventional reinforcement learning (RL) methods struggle with the delayed observations due to the delay-induced observations, leading to high-frequency chattering. To address this, we propose a hybrid control framework, delay-resilient RL, integrating a state estimator utilizing Long Short-Term Memory (LSTM) with a residual RL policy, which is resilient to stochastic delays. The LSTM reconstructs smooth, continuous state estimates from delayed observations, enabling the RL agent to learn a residual torque compensation policy that balances tracking accuracy with velocity smoothness. Experimental validation on Franka Panda robots demonstrates that our approach significantly outperforms the state-of-the-art baselines, ensuring robust and stable teleoperation even under high-variance stochastic delays.\n\n---\n\n## 155. Retrieval-Augmented Large Language Models for Schema-Constrained Clinical Information Extraction\n- **Authors**: A H M Rezaul Karim , Ozlem Uzuner\n- **URL**: [https://arxiv.org/abs/2605.15467](https://arxiv.org/abs/2605.15467)\n- **Abstract**:\n> Conversational nurse-patient transcripts contain actionable observations, but converting these transcripts into structured representations at scale remains challenging. Documentation burden is substantial, with prior studies showing clinicians spend large portions of their workday on documentation and related desk work rather than direct patient care. MEDIQA-SYNUR focuses on observation extraction from conversational nurse-patient transcripts, requiring systems to normalize these narratives into a predefined schema with value-type constraints. We propose a modular retrieval-augmented generation (RAG) pipeline that uses the training set as an exemplar corpus, combines schema-constrained prompting (full schema vs. pruned candidate schema), deterministic schema-based postprocessing, and a second-pass audit, with two LLM backbones: Llama-4-Scout-17B-16E-Instruct and GPT-5.2 with corresponding embedding models for RAG. Our best configuration uses GPT-5.2 with full schema, RAG, and a second-pass auditing, achieving 80.36% F1 score. Overall, our results show that RAG consistently improves performance, while the optimal degree of schema constraint depends on the model, and second-pass auditing yields modest additional gains by correcting residual schema-adherence errors.\n\n---\n\n## 156. GRLO: Towards Generalizable Reinforcement Learning in Open-Ended Environments from Zero\n- **Authors**: Shangjian Yin , Yu Fu , Yue Dong , Zhouxing Shi\n- **URL**: [https://arxiv.org/abs/2605.15464](https://arxiv.org/abs/2605.15464)\n- **Abstract**:\n> Post-training has become a crucial step for unlocking the capabilities of large language models, with reinforcement learning (RL) emerging as a critical paradigm. Recent RL-based post-training has increasingly split into two paradigms: reinforcement learning from human feedback (RLHF), which optimizes models using human preference signals in target domains, and reinforcement learning from verifiable rewards (RLVR), which operates in verifier-backed environments. The latter has dominated recent reasoning-oriented post-training because it delivers stronger gains and higher efficiency on domain-specific tasks (e.g., reasoning). However, although in-domain RL training achieves promising performance, it still requires a substantial amount of GPU compute, which remains a major barrier to broad adoption. In this work, we study the generalization ability of RLHF learned from scratch from a small set of interactions in open-ended environments, and investigate whether the conversational abilities it explicitly acquires can implicitly transfer to downstream tasks such as mathematical reasoning and code generation, namely GRLO. Specifically, on Qwen3-4B-Base backbone, GRLO improves the average performance across all domains from 24.1 to 63.1 with only 5K prompts and 22.7 GPU hours, requiring about $46\\times$ less data and $68\\times$ less compute than a strong in-domain RLVR baseline. The resulting model is even competitive with Qwen's released post-trained models which required a much larger training cost. Notably, a subsequent in-domain RLVR stage brings only selective gains, mainly on harder competition-math benchmarks. We hope GRLO offers a simple and efficient recipe for building broadly capable post-trained models. Our code and data will be available at: \\href{ this https URL }{ this https URL }.\n\n---\n\n## 157. DrugSAGE:Self-evolving Agent Experience for Efficient State-of-the-Art Drug Discovery\n- **Authors**: Yikun Zhang , Xiwei Cheng , Tianyu Liu , Yuanqi Du , Wengong Jin\n- **URL**: [https://arxiv.org/abs/2605.15461](https://arxiv.org/abs/2605.15461)\n- **Abstract**:\n> Building state-of-the-art (SOTA) predictive models for drug discovery requires expensive search over tools, architectures, and training strategies. Current LLM-based agents can find SOTA solutions through extensive trial and error, but they do not retain the experience accumulated along the way and therefore pay the full search cost on every new task. We propose \\method (Self-evolving Agent Experience), a framework that accumulates and reuses experience across tasks to build SOTA drug discovery models efficiently. \\method maintains a cross-task memory of verified skills, statistical evidence about effective strategies, and a record of recurring errors and their fixes. In some cases, \\method transfers a working solution directly without test-time search. In 33 molecular property prediction tasks, \\method ranks first among nine SOTA agents in a single-task setting. With memory accumulated from 16 smaller tasks, \\method achieves an averaged normalized score of 0.935 on 17 held-out tasks in a cross-task evaluation setting and outperforms all baseline agents by 10-30\\% in a zero-test-time search regime. In summary, our work shows the advantage of cross-task memory for efficient SOTA model development in drug discovery.\n\n---\n\n## 158. Differentially Private Motif-Preserving Multi-modal Hashing\n- **Authors**: Zehua Cheng , Wei Dai , Jiahao Sun\n- **URL**: [https://arxiv.org/abs/2605.15460](https://arxiv.org/abs/2605.15460)\n- **Abstract**:\n> Cross-modal hashing enables efficient retrieval by encoding images and text into compact binary codes. State-of-the-art methods rely on semantic similarity graphs derived from user interactions for supervision, yet these graphs encode sensitive behavioral patterns vulnerable to link reconstruction attacks. Existing privacy-preserving approaches fail on graph-structured data: Differentially Private SGD destroys relational motifs by treating samples independently, while graph synthesis methods suffer from unbounded local sensitivity in scale-free networks, hub nodes cause single-edge modifications to alter triangle counts by $\\mathcal{O}(N)$, necessitating prohibitive noise injection. We term this phenomenon Hubness Explosion. We propose DMP-MH, a Sanitize-then-Distill framework that decouples privacy from representation learning. Our approach first bounds sensitivity by deterministically clipping node degrees, capping the $L_2$-sensitivity of triangle motifs independently of dataset size. A sanitized synthetic graph is then generated via Noisy Mirror Descent under $(\\epsilon,\\delta)$-Edge Differential Privacy. Finally, dual-stream hashing networks distill this topology using a holistic structural loss that enforces cross-modal alignment. Evaluated on MIRFlickr-25K and NUS-WIDE under a strict inductive protocol, DMP-MH outperforms private baselines by up to 11.4 mAP points while retaining up to 92.5% of non-private performance.\n\n---\n\n## 159. RIDE: Retinex-Informed Decoupling for Exposing Concealed Objects\n- **Authors**: Chunming He , Rihan Zhang , Dingming Zhang , Chengyu Fang , Longxiang Tang , Jingjia Feng , Fengyang Xiao , Sina Farsiu\n- **URL**: [https://arxiv.org/abs/2605.15450](https://arxiv.org/abs/2605.15450)\n- **Abstract**:\n> Concealed Object Segmentation (COS) encompasses a family of dense-prediction tasks, including camouflaged object detection, polyp segmentation, transparent object detection, and industrial defect inspection, where targets are visually entangled with their surroundings through different physical mechanisms. Existing methods either operate directly on RGB images or employ \\emph{heterogeneous} decompositions (\\eg, Fourier, wavelet) that redistribute spatial evidence across scale/frequency coefficients, making pixel-aligned cues less direct. We introduce a fundamentally different perspective: \\textbf{homogeneous image decomposition} via Retinex theory, which factorizes an image into illumination and reflectance components within the \\emph{same} spatial domain. Our key insight is that visual entanglement enforces appearance matching in the composite space, but this does \\emph{not} necessitate simultaneous matching in both component spaces, a phenomenon we formalize as the \\textbf{Discriminability Gap Theorem}. Crucially, we show that across diverse COS sub-tasks, the underlying physical processes systematically anti-correlate illumination and reflectance differences, yielding theoretical guarantees that Retinex decomposition preserves or strictly improves total foreground--background discriminability across the full physical regime, with anti-correlation maximizing the gain. Building on this, we propose \\textbf{RIDE} comprising: (i) a Task-Driven Retinex Decomposition module that learns segmentation-optimal factorizations end-to-end; (ii) a Discriminability Gap Attention mechanism that adaptively exploits where decomposition helps; and (iii) a Camouflage-Breaking Contrastive loss operating in reflectance feature space.\n\n---\n\n## 160. Runtime-Structured Task Decomposition for Agentic Coding Systems\n- **Authors**: Shubhi Asthana , Bing Zhang , Chad DeLuca , Hima Patel , Ruchi Mahindru\n- **URL**: [https://arxiv.org/abs/2605.15425](https://arxiv.org/abs/2605.15425)\n- **Abstract**:\n> Agentic coding systems increasingly use large language models (LLMs) for software engineering tasks such as debugging, root cause analysis, and code review. However, many existing systems encode task logic, execution flow, and output generation inside monolithic prompts. This design creates brittle behavior, limited debuggability, and high retry costs because failures often require rerunning the full workflow. We present runtime-structured task decomposition, an architectural approach in which task partitioning and execution flow are managed through executable control logic rather than prompt structure alone. LLMs are used only for focused judgment tasks, and outputs are validated against predefined schemas before downstream execution. We evaluate this approach on two software engineering workloads using three configurations: monolithic execution, static decomposition with fixed subtasks and no runtime branching, and runtime-structured decomposition. Each configuration was evaluated across 10 runs. Our results show that decomposition alone does not necessarily reduce retry cost. In the Kubernetes root cause analysis workload, the static decomposition baseline produced a retry cost of 1,632 +/- 145 tokens versus 904 +/- 17 tokens for the monolithic baseline because failures forced reruns of downstream subtasks. A similar pattern appeared in the multi-file debugging workload, where the static baseline consumed 933 tokens compared to 703 tokens for the monolithic system. The runtime-structured approach reran only failed subtasks, reducing retry costs to 436 +/- 132 tokens for root cause analysis and 460 tokens for debugging. Overall, the approach achieved up to 51.7% lower retry cost than monolithic systems and 73.2% lower retry cost than static decomposition baselines, improving efficiency, debuggability, and operational reliability in agentic coding systems.\n\n---\n\n## 161. MR2-ByteTrack: CNN and Transformer-based Video Object Detection for AI-augmented Embedded Vision Sensor Nodes\n- **Authors**: Luca Bompani , Manuele Rusci , Luca Benini , Daniele Palossi , Francesco Conti\n- **URL**: [https://arxiv.org/abs/2605.15423](https://arxiv.org/abs/2605.15423)\n- **Abstract**:\n> Modern smart vision sensors need on-device intelligence to process video streams, as cloud computing is often impractical due to bandwidth, latency, and privacy constraints. However, these sensory systems typically rely on ultra-low-power microcontrollers (MCUs) with limited memory and compute, making conventional video object detection methods, which require feature storage or multi-frame buffering, unfeasible. To address this challenge, we introduce Multi-Resolution Rescored ByteTrack (MR2-ByteTrack), a Video Object Detection (VOD) method tailored for MCU-based embedded vision nodes. MR2-ByteTrack reduces computational cost by alternating between full- and low-resolution inference, while linking detections across frames via ByteTrack and correcting misclassifications through the Rescore algorithm, which applies probability union rules to aggregate detection confidence scores across frames. We apply our approach to both a CNN-based detector and a Transformer-based model, demonstrating its generality across architectures with fundamentally different spatial processing. Experiments on ImageNetVID demonstrate that MR2-ByteTrack maintains accuracy, achieving mAP scores of up to 49.0 for the CNN-based models and 48.7 for the Transformer, while reducing multiply-accumulate operations by as much as 53\\% for the CNNs and 32\\% for the Transformer. When deployed on GAP9, an ultra-low-power RISC-V multicore MCU, our method yields up to 55\\% energy savings compared to processing only full-resolution images, enabling the first real-time Transformer-based VOD on an MCU-class embedded vision node. Code available at this https URL\n\n---\n\n## 162. $f$-Trajectory Balance: A Loss Family for Tuning GFlowNets, Generative Models, and LLMs with Off- and On-Policy Data\n- **Authors**: Jake Fawkes , Jason Hartford\n- **URL**: [https://arxiv.org/abs/2605.15417](https://arxiv.org/abs/2605.15417)\n- **Abstract**:\n> In GFlowNets and variational inference, it has been shown that the mean square error between target and model log probabilities is an effective, low variance, surrogate loss for training generative models. This loss has the property that when evaluated \\emph{on-policy} its gradients correspond to those of the KL divergence, while \\emph{off-policy} it remains a valid loss with the same global minimizer. In this work, we demonstrate that this construction can be extended to the whole family of $f$-divergences, leading to a family of losses whose on-policy gradients are that of the corresponding $f$-divergence, but retain the same global minimizer off-policy. Specifically, we show that the on-policy gradients lead to a one to one correspondence between translation invariant loss functions on the target and model log probabilities, and $f$-divergences. This equivalence allows us to design new surrogate loss functions for tuning a wide class of generative models that inherit the properties of the corresponding $f$-divergence, such as being more mode covering, whilst being applicable to off-policy data. We apply our losses on a range of tasks, including classic synthetic examples, SynFlowNets for molecule discovery, and asynchronous large language model (LLM) tuning, demonstrating that our models retain their predicted properties on- and off-policy in a wide class of generative models.\n\n---\n\n## 163. Margin-Adaptive Confidence Ranking for Reliable LLM Judgement\n- **Authors**: Gaojie Jin , Yong Tao , Lijia Yu , Tianjin Huang\n- **URL**: [https://arxiv.org/abs/2605.15416](https://arxiv.org/abs/2605.15416)\n- **Abstract**:\n> Jung et al. (2025) introduce a hypothesis testing framework for guaranteeing agreement between large language models (LLMs) and human judgments, relying on the assumption that the model's estimated confidence is monotonic with respect to human-disagreement risk. In practice, however, this assumption may be violated, and the generalization behavior of the confidence estimator is not explicitly analyzed. We mitigate these issues by learning a dedicated confidence estimator instead of relying on heuristic confidence signals. Our approach leverages simulated annotator diversity and a margin-based ranking formulation to explicitly model how confidently an LLM distinguishes between human-agreement and human-disagreement cases. We further derive generalization guarantees for this estimator, revealing a margin-dependent trade-off that informs the design of an adaptive estimator training procedure. When integrated into fixed-sequence testing, the learned confidence estimator yields improved ranking accuracy and empirically strengthens the monotonic relationship between confidence and disagreement risk, leading to higher success rates in satisfying target agreement levels across multiple datasets and judge models.\n\n---\n\n## 164. From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery\n- **Authors**: Lingzhe Zhang , Tong Jia , Yunpeng Zhai , Zixuan Xie , Chiming Duan , Minghua He , Philip S. Yu , Ying Li\n- **URL**: [https://arxiv.org/abs/2605.15412](https://arxiv.org/abs/2605.15412)\n- **Abstract**:\n> Modern quantitative trading increasingly relies on systematic models to extract predictive signals from large-scale financial data, where alpha factor discovery plays a central role in transforming market observations into tradable signals. Recent LLM-based methods have shown promise in automating factor generation, but most of them still rely on prompt-level generation--evaluation--feedback loops for iterative optimization. As the loop becomes longer, repeatedly appended historical candidates and feedback can cause context explosion, increase inference cost, dilute useful information, and introduce feedback drift. Moreover, these methods often depend on very large LLMs whose stable generation preferences may lead to structurally similar expressions, redundant candidates, and search stagnation. To address these limitations, we propose \\textsc{QuantEvolver}, a self-evolving alpha factor discovery framework based on reinforcement fine-tuning. Instead of accumulating feedback in the prompt, \\textsc{QuantEvolver} converts executable quantitative evaluation into policy updates, enabling a Miner LLM to internalize historical optimization experience through parameter learning. Specifically, \\textsc{QuantEvolver} constructs high-quality seed factors, builds diverse seed--time-window training tasks, generates executable Factor DSL expressions, evaluates them through Regime Backtest, and optimizes the Miner LLM with Diversity-Complementarity Reward. During training, high-quality factors are continuously accumulated in a Mined Factor Database, which serves as the final discovered factor library. Extensive experiments on three realistic market benchmarks demonstrate the effectiveness of \\textsc{QuantEvolver}, which consistently improves the primary evaluation metric of each task over existing LLM-based alpha factor discovery baselines, produces higher-quality and more complementary factor pools.\n\n---\n\n## 165. Diagonal Adaptive Non-local Observables on Quantum Neural Networks\n- **Authors**: Huan-Hsin Tseng , Yan Li , Hsin-Yi Lin , Samuel Yen-Chi Chen\n- **URL**: [https://arxiv.org/abs/2605.15410](https://arxiv.org/abs/2605.15410)\n- **Abstract**:\n> Adaptive Non-local Observables (ANOs) have shown that making quantum observables dynamic can substantially enlarge the function space of Variational Quantum Algorithms, partly shifting hardware demands from circuit synthesis to measurement design. However, this advantage is accompanied by a steep increase in the number of parameters, as well as the classical optimization cost for varying general Hermitian observables. We propose a special form of ANO that significantly reduces this burden by considering only diagonal observables paired with quantum circuits. Mathematically, this is equivalent to the full ANO of a large parameter space since diagonal matrices are canonical representatives of the ANO space modulo unitary similarity. As a result, Diagonal ANO retains the same capability of full ANO while reducing $k$-local observable complexity from $O(4^k)$ to $O(2^k)$ and lowering the corresponding measurement-side classical computation. In this sense, diagonal ANO preserves much of the benefit of full ANO while encompassing conventional VQCs as a special case.\n\n---\n\n## 166. Amortized Energy-Based Bayesian Inference\n- **Authors**: Hojjat Kaveh , Ricardo Baptista , Andrew M. Stuart\n- **URL**: [https://arxiv.org/abs/2605.15407](https://arxiv.org/abs/2605.15407)\n- **Abstract**:\n> We consider amortized Bayesian inference for nonlinear inverse problems in settings where only samples from the joint distribution of parameters and observations are available. Classical methods such as Markov chain Monte Carlo require solving a new inference problem for each observation, which can be computationally prohibitive when inference must be repeated many times. We propose a transport-based approach that learns an observation-dependent map pushing forward a reference measure to approximate the posterior distribution. The map is trained by minimizing an averaged energy-distance objective between the true posterior and the learned pushforward. This formulation is likelihood-free, requiring only joint samples, and avoids density evaluation, invertibility constraints, and Jacobian determinant computations. For function-space inverse problems with Gaussian priors, we parameterize the transport map as the identity plus a perturbation in the Cameron-Martin space of the prior, preserving absolute continuity with respect to the prior. In infinite-dimensional settings, the map is represented using neural operators. We illustrate the method on a finite-dimensional nonlinear inverse problem and two PDE-constrained inverse problems arising in porous medium flow and seismic inversion. The results show that the learned transport captures posterior structure, including multimodality and dominant modes, while enabling fast posterior sampling for new observations.\n\n---\n\n## 167. Breakeven complexity: A new perspective on neural partial differential equation solvers\n- **Authors**: Yijing Zhang , Nicholas Roberts , Tanya Marwah , Mikhail Khodak\n- **URL**: [https://arxiv.org/abs/2605.15399](https://arxiv.org/abs/2605.15399)\n- **Abstract**:\n> Neural surrogate solvers of partial differential equations (PDEs) promise dramatic speedups over numerical methods, especially in scenarios requiring many solves. However, current accuracy-based evaluations do not fully consider two central issues: (1) neural solvers incur substantial up-front costs for data generation, training, and tuning; and (2) classical solvers can also generate low-fidelity solutions at a sufficiently low simulation cost. To explicitly account for these realities and fully incorporate end-to-end costs, we propose an evaluation framework centered on breakeven complexity, a metric that counts the forward solves before a learned solver is cost-effective relative to an error-equivalent traditional solver. To evaluate this measure, we apply scaling laws to determine how much training budget to allocate to data generation and discuss how to achieve smooth error-matching in diverse settings. We evaluate the breakeven complexity of multiple neural PDE solvers on three PDEs on 2D periodic domains from APEBench and a novel benchmark of flows past multiple obstacles generated by the GPU-native PyFR code. Among other findings, our results suggest that neural PDE solvers become more effective as problems get harder in terms of cost, dimension, rollout, physics regime (e.g. higher Reynolds number), etc.\n\n---\n\n## 168. Representation Without Reward: A JEPA Audit for LLM Fine-Tuning\n- **Authors**: Biswa Sengupta\n- **URL**: [https://arxiv.org/abs/2605.15394](https://arxiv.org/abs/2605.15394)\n- **Abstract**:\n> Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outputs. For autoregressive language-model fine-tuning the principle entails a stricter requirement: the induced hidden-state geometry must reach the language-model head \\emph{and} improve the decoded task metric. We test that requirement under a fixed Llama-3.2-1B-Instruct LoRA harness on natural-language-to-regex generation, comparing twenty-two training-time auxiliaries across trajectory-shape regularisation, distributional constraints, predictor/target asymmetry, Fisher-metric Jacobi residuals, and a decoder-visible JEPA objective constructed to lie in cross-entropy's positive cone. The empirical answer is a structured null: several auxiliaries clear single-cell paired $\\alpha = 0.10$ without correction (T3-Local at $\\Delta = +2.53$~pp, $p = 0.003$ being the strongest), but none survives Bonferroni or Holm--Bonferroni at the relevant family-wise threshold, even though many change curvature, anisotropy, variance, and gradient direction. Decoder-visible JEPA yields the first positive auxiliary--cross-entropy gradient cosine in the study, yet exact match remains inside seed noise; a full-fine-tuning replication of the same auxiliary at $n = 5$ seeds reproduces the null on both benchmarks (TURK: $\\Delta = +0.04$~pp, $p_{\\text{paired} } = 0.96$; SYNTH: $\\Delta = +0.52$~pp, $p_{\\text{paired} } = 0.28$), so the null is robust across LoRA and full fine-tuning for the decoder-visible construction. Hidden-state representation work and decoded-task accuracy are therefore weakly coupled in this regime; we accordingly reframe LLM-domain JEPA evaluation as a coupling problem, in which the operative question is under which metrics useful hidden geometry becomes decoder-visible task signal.\n\n---\n\n## 169. PanoWorld: Geometry-Consistent Panoramic Video World Modeling\n- **Authors**: Le Jiang , Xiangyu Bai , Bishoy Galoaa , Shayda Moezzi , Caleb James Lee , Tooba Imtiaz , Edmund Yeh , Jennifer Dy , Yanzhi Wang , Sarah Ostadabbas\n- **URL**: [https://arxiv.org/abs/2605.15391](https://arxiv.org/abs/2605.15391)\n- **Abstract**:\n> We present PanoWorld, a panoramic video world model that generates geometry-consistent 360$\\degree$ video from a single image and a caption. Existing panoramic video methods optimize primarily for visual realism and do not explicitly constrain the underlying 3D scene state, producing outputs that appear plausible yet exhibit inconsistent depth, broken correspondences, and implausible motion across the spherical surface. We address this gap by framing panoramic video generation as a geometry- and dynamics-consistent latent state modeling problem rather than pure visual synthesis. Building on a pre-trained perspective video world model, we introduce two lightweight regularizers: a depth consistency loss against pseudo ground-truth panoramic depth, and a trajectory consistency loss that supervises the 3D world-frame positions of tracked points across time. We further apply spherical-geometry-aware adaptation to the conditioning and positional encoding. We additionally introduce PanoGeo, a unified geometry-aware panoramic video dataset with consistent depth, trajectory, and prompt annotations across diverse real and synthetic sources, used for both training and stratified evaluation. Experiments show that PanoWorld improves geometric consistency over prior panoramic generation methods while maintaining competitive visual realism, establishing that panoramic video generation must be treated as a geometric modeling problem to support the holistic spatial understanding requirements of embodied AI applications. Code is available at this https URL .\n\n---\n\n## 170. Is One Score Enough? Rethinking the Evaluation of Sequentially Evolving LLM Memory\n- **Authors**: Songwei Dong , Zihan Chen , Chengshuai Shi , Peng Wang , Jundong Li , Cong Shen\n- **URL**: [https://arxiv.org/abs/2605.15384](https://arxiv.org/abs/2605.15384)\n- **Abstract**:\n> Memory plays a central role in enabling large language models (LLMs) to operate over sequential tasks by accumulating and reusing experience over time. However, existing evaluations of LLM memory mostly rely on aggregate metrics such as final hold-out accuracy or cumulative online performance, which can obscure critical failure modes such as forgetting and negative transfer. In this paper, we introduce SeqMem-Eval, a diagnostic evaluation framework for sequentially evolving LLM memory. Drawing inspiration from continual learning, it targets a test-time setting in which memory is external, prompt-mediated, and updated without modifying model parameters. Rather than focusing only on final performance, SeqMem-Eval evaluates how memory states evolve, generalize, consolidate experience, and retain useful information during sequential inference. Specifically, it measures online utility, hold-out generalization, backward transfer, and forgetting, providing a finer-grained view of memory quality. Through extensive experiments across diverse tasks and memory methods, we show that higher final or cumulative accuracy does not necessarily imply better memory quality: many methods exhibit strong performance gains while suffering from substantial forgetting or negative transfer. Moreover, different memory designs exhibit distinct trade-offs between adaptability and stability that remain invisible under standard evaluation metrics.\n\n---\n\n## 171. ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing\n- **Authors**: Blaž Rolih , Matic Fučka , Filip Wolf , Luka Čehovin Zajc\n- **URL**: [https://arxiv.org/abs/2605.15375](https://arxiv.org/abs/2605.15375)\n- **Abstract**:\n> Remote sensing change detection (RSCD) aims to localise changes between two images of the same geographic region. In practice, change masks often follow region-level annotation conventions rather than purely local appearance differences, making them context-dependent and occasionally ambiguous. Most state-of-the-art methods utilise per-pixel discriminative classification, which produces a single prediction per input and fails to explicitly model the changed region as a coherent whole. A natural alternative is generative formulation, which can model a distribution of plausible masks, enabling sampling to capture ambiguity and encourage global consistency. However, existing generative RSCD approaches typically lag behind strong discriminative baselines due to the high computational cost of pixel-space generation and the complexity of their conditioning mechanisms. To address the limitations of prior discriminative and generative methods, we propose ChangeFlow, a generative framework that reformulates change detection as the synthesis of a change mask in latent space via rectified flow. ChangeFlow is guided by a structured yet lightweight conditioning signal, and its stochastic design naturally supports sampling-based prediction ensembling. Namely, aggregating multiple predicted change masks improves robustness, while sample agreement provides a practical confidence estimation that highlights ambiguous regions. Across four benchmarks, ChangeFlow achieves an average F1 of 80.4\\%, improving by 1.3 points on average over the previous best method, while maintaining inference speed comparable to recent strong baselines. Project page: this https URL\n\n---\n\n## 172. PACER: Acyclic Causal Discovery from Large-Scale Interventional Data\n- **Authors**: Ramon Viñas Torné , Sílvia Fàbregas Salazar , Soyon Park , Ivo Alexander Ban , Artyom Gadetsky , Nikita Doikov , Maria Brbić\n- **URL**: [https://arxiv.org/abs/2605.15353](https://arxiv.org/abs/2605.15353)\n- **Abstract**:\n> Inferring the structure of directed acyclic graphs (DAGs) from data is a central challenge in causal discovery, particularly in modern high-dimensional settings where large-scale interventional data are increasingly available. While interventional data can improve identifiability, existing methods remain limited by soft acyclicity constraints, leading to optimization over invalid cyclic graphs, numerical instability, and reduced scalability. We introduce PACER (Perturbation-driven Acyclic Causal Edge Recovery), a scalable framework for causal discovery that guarantees acyclicity by construction. PACER parameterizes a distribution over DAGs through a joint model of variable permutations and edge probabilities, enabling direct optimization over valid causal structures without surrogate penalties. The framework supports a unified likelihood-based treatment of observational and interventional data, flexible conditional density models, and the incorporation of structural prior knowledge. For linear-Gaussian mechanisms, we derive closed-form expressions for the expected interventional log-likelihood and its gradients, yielding substantial computational gains. Empirically, PACER matches or exceeds state-of-the-art methods on protein signaling and large-scale genetic perturbation benchmarks, while scaling efficiently to networks with thousands of variables and achieving up to two orders of magnitude speedups over penalty-based differentiable approaches. These results demonstrate that exact and scalable causal discovery from high-dimensional perturbation data is achievable through principled search space design.\n\n---\n\n## 173. LEAP: Trajectory-Level Evaluation of LLMs in Iterative Scientific Design\n- **Authors**: Marilyn Zhang , Tianfeng Chen , Fabián Barzuna , Ankita Rathod , Mark E. Whiting\n- **URL**: [https://arxiv.org/abs/2605.15341](https://arxiv.org/abs/2605.15341)\n- **Abstract**:\n> LLMs are increasingly deployed in autonomous laboratories, under the assumption that their domain priors and reasoning over iterative feedback let them converge on good designs in fewer iterations than feedback-only baselines. Current iterative scientific design benchmarks, however, score only outcome snapshots at fixed horizons. This leaves the learning trajectory unmeasured, even though the trajectory is what captures learning efficiency, where each iteration saved is a real saving in cost and time. Motivated by this, we examine three evaluation choices that change the conclusions one draws about LLM learning efficiency in iterative scientific design: what to measure, what baseline to compare against, and what to ground against. We introduce LEAPBench, Learning Efficiency in Adaptive Processes, a 55-task framework that pairs a best-so-far area under the curve (AUC) trajectory metric with a classical Bayesian-optimization reference and an audit grounded in published literature. Applied to eight contemporary LLMs, switching from final-outcome to trajectory scoring changes the best-model decision on 53% of tasks at matched horizons, and exposes efficiency gains overlooked by outcome-based scoring. LLMs do not outperform a classical Bayesian baseline. On 16 biology tasks where the oracle's reward signal is aligned with configurations from the published-best design, domain-aware prompting leads to LLM choices that match the published-best's approximately 10 percentage points less often than domain-agnostic prompting at iteration 30. The pattern is sharpest on 6 tasks where the literature-typical and published-best configurations diverge, and domain-agnostic prompting matches the published-best more often on all 6. The trajectory metric also doubles as a tractable training target. Offline reinforcement learning with the metric as a reward improves performance on 14 of 21 held-out tasks.\n\n---\n\n## 174. Hidden in Memory: Sleeper Memory Poisoning in LLM Agents\n- **Authors**: Sidharth Pulipaka , Stanislau Hlebik , Leonidas Raghav , Sahar Abdelnabi , Vyas Raina , Ivaxi Sheth , Mario Fritz\n- **URL**: [https://arxiv.org/abs/2605.15338](https://arxiv.org/abs/2605.15338)\n- **Abstract**:\n> Large language models are increasingly augmented with persistent memory, allowing assistants to store user-specific information across sessions for personalization and continuity. This statefulness introduces a new security risk: adversarial content can corrupt what an assistant remembers and thereby influence future interactions. We propose and study sleeper memory poisoning, a delayed attack in which an adversary manipulates external context, such as a document, webpage, or repository, to cause the assistant to store a fabricated memory about the user. Unlike conventional prompt injection, the attack can remain dormant and re-emerge across multiple later conversations. We evaluate the full attack pipeline: whether poisoned memories are written, later retrieved, and ultimately used to steer the following conversations. Across stateful LLM assistants, poisoned memories were added up to 99.8% on GPT-5.5 and 95% on Kimi-K2.6. Crucially, among successful retrievals, poisoned memories cause attacker-intended agentic actions in 60-89% of evaluations across models. These results show that persistent memory can act as a long-term attack surface across multiple future conversations.\n\n---\n\n## 175. HoloMotion-1 Technical Report\n- **Authors**: Maiyue Chen , Kaihui Wang , Bo Zhang , Xihan Ma , Zhiyuan Yang , Yi Ren , Qijun Huang , Zihao Zhu , Yucheng Wang , Zhizhong Su\n- **URL**: [https://arxiv.org/abs/2605.15336](https://arxiv.org/abs/2605.15336)\n- **Abstract**:\n> In this report, we present HoloMotion-1, a humanoid motion foundation model for zero-shot whole-body motion tracking. A key innovation of HoloMotion-1 is to scale control-policy training with a large-scale hybrid motion corpus, where video-reconstructed motions from in-the-wild videos provide the dominant source of motion diversity, while curated motion-capture and in-house motion data provide higher-fidelity supervision and deployment-oriented coverage. This data regime enables HoloMotion-1 to move beyond conventional MoCap-only training and exposes the policy to substantially broader behaviors, capture conditions, and motion styles. Learning from such heterogeneous data introduces new challenges, including reconstruction noise, source-domain mismatch, uneven motion quality, and the need for temporal modeling under large behavioral variation. To address these challenges, HoloMotion-1 integrates large-capacity temporal modeling, a sparsely activated Mixture-of-Experts Transformer with KV-cache inference for real-time control, and a sequence-level training strategy that improves learning efficiency on extended motion sequences. Extensive experiments on multiple unseen motion benchmarks show that HoloMotion-1 generalizes robustly across diverse motion types and capture conditions, significantly improves tracking accuracy over prior methods, and transfers directly to a real humanoid robot without task-specific fine-tuning.\n\n---\n\n## 176. From I/O to Code with Discovery Agent\n- **Authors**: Yihong Dong , Jiaru Qian , Haoran Zhang , Peixu Wang , Binhua Li , Zhi Jin , Yongbin Li , Ge Li , Xiaokang Yang , Xue Jiang\n- **URL**: [https://arxiv.org/abs/2605.15334](https://arxiv.org/abs/2605.15334)\n- **Abstract**:\n> The automatic synthesis of a program from any form of specification is regarded as a holy grail of computer science. Fueled by LLMs, NL2Code has achieved tremendous success, yet the fundamentally more challenging task of synthesizing programs from input-output behavior, which we refer to as IO2Code, remains largely unsolved. Whereas NL2Code can exploit the semantic alignment between natural language and code acquired during pretraining, IO2Code requires recovering underlying principles from concrete computational behavior, navigating a vast and underspecified hypothesis space. To address this, we propose DIO-Agent, a discovery agent for IO2Code. Our method frames IO2Code as an evolutionary search over discrete program space, in which an LLM serves as the mutation operator and concrete error signals from execution guide each mutation. To prevent the search from wandering into structurally complex yet incorrect dead ends, we introduce the Transformation Priority Premise as a mutation prior that biases the LLM toward the simplest hypothesis consistent with current evidence, progressively escalating from constants to conditionals to iteration only when simpler constructs are insufficient. To facilitate systematic study, we further construct an IO2CodeBench spanning multiple difficulty levels. Extensive experiments show that DIO-Agent consistently outperforms both traditional program-by-example method and SOTA evolution-agent baselines across all difficulty levels and various LLMs, while substantially surpassing test-time scaling strategies with equivalent sampling budgets.\n\n---\n\n## 177. Fortress: A Case Study in Stabilizing Search Recommendations via Temporal Data Augmentation and Feature Pruning\n- **Authors**: Milind Pandurang Jagre , Jia Huang , Dayvid V. R. Oliveira , Zhinan Cheng , Babak Seyed Aghazadeh , Puja Das , Chris Alvino , Jinda Han , Kailash Thiyagarajan\n- **URL**: [https://arxiv.org/abs/2605.15299](https://arxiv.org/abs/2605.15299)\n- **Abstract**:\n> In search and recommendation systems, predictive models often suffer from temporal instability when certain input features introduce volatility in output scores. This instability can degrade model reliability and user experience especially in multi-stage systems where consistent predictions are critical for downstream decision making. We introduce Fortress, a general framework for enhancing model stability and accuracy by identifying and pruning features that contribute to inconsistent prediction scores over time. Fortress leverages historical snapshots temporally partitioned datasets capturing score fluctuations for the same entity across periods and follows a four-step process: (1) collect historical snapshots, (2) identify samples with unstable predictions, (3) isolate and remove instability-inducing features, and (4) retrain models using only stable features. While semantic features from LLMs and BERT-based models improve generalization, they often lack full query or entity coverage. Engagement-based features offer strong predictive power but tend to introduce temporal instability. Fortress mitigates this trade-off by suppressing the volatility of engagement signals while retaining their predictive value leading to more stable and accurate models. We validate Fortress on a query-to-app relevance model in a large-scale app marketplace. Offline experiments demonstrate notable improvements in prediction stability (measured by Coefficient of Variation) and classification performance (measured by PR-AUC).\n\n---\n\n## 178. PhysBrain 1.0 Technical Report\n- **Authors**: Shijie Lian , Bin Yu , Xiaopeng Lin , Changti Wu , Hang Yuan , Xiaolin Hu , Zhaolong Shen , Yuzhuo Miao , Haishan Liu , Yuxuan Tian , Yukun Shi , Cong Huang , Kai Chen\n- **URL**: [https://arxiv.org/abs/2605.15298](https://arxiv.org/abs/2605.15298)\n- **Abstract**:\n> Vision-language-action models have advanced rapidly, but robot trajectories alone provide limited coverage for learning broad physical understanding. PhysBrain 1.0 studies a complementary route: converting large-scale human egocentric video into structured physical commonsense supervision before robot adaptation. Our data engine extracts scene elements, spatial dynamics, action execution, and depth-aware relations, then turns them into question-answer supervision for training PhysBrain VLMs. The resulting physical priors are further transferred to VLA policies through a capability-preserving and language-sensitive adaptation design. Across multimodal QA benchmarks and embodied control benchmarks, including ERQA, PhysBench, SimplerEnv-WidowX, LIBERO, and RoboCasa, PhysBrain 1.0 achieves SOTA results and shows especially strong out-of-domain performance on SimplerEnv. These results suggest that scaling physical commonsense from human interaction video can provide an effective bridge from multimodal understanding to robot action.\n\n---\n\n## 179. GESD: Beyond Outcome-Oriented Fairness\n- **Authors**: Gideon Popoola , John Sheppard\n- **URL**: [https://arxiv.org/abs/2605.15295](https://arxiv.org/abs/2605.15295)\n- **Abstract**:\n> Machine learning (ML) algorithms are increasingly deployed in high-stakes decision-making domains such as loan approvals, hiring, and recidivism predictions. While existing fairness metrics (e.g., statistical parity, equal opportunity) effectively quantify outcome-oriented disparities, they offer limited insight into the procedure or explanation behind biased decisions. To address this gap, we propose Group-level Explanation Stability Disparity (GESD), a \\textit{procedural-oriented} fairness metric that measures disparities in the stability, robustness, and sensitivity of model explanations across different subgroups in a protected category. %GESD is explainer-agnostic, model-agnostic, and extends the scope of fairness analyses to the level of explainability. We further integrate GESD into a multi-objective optimization framework that jointly optimizes for utility, outcome-based fairness, and explanation-based fairness called FEU (Fairness--Explainability--Utility). Empirical results on multiple benchmark datasets show that GESD effectively captures group-wise discrepancies in explanation quality, and that FEU improves both utility and fairness over state-of-the-art methods. By bridging outcome-based and explanation-based fairness, GESD offers a comprehensive tool for diagnosing and mitigating bias in predictive modeling. Our code and datasets are available on GitHub {\\hyperlink{ this https URL }{ this https URL } }\n\n---\n\n## 180. GQA-μP: The maximal parameterization update for grouped query attention\n- **Authors**: Kyle R. Chickering , Huijuan Wang , Mengxi Wu , Alexander Moreno , Muhao Chen , Xuezhe Ma , Daria Soboleva , Joel Hestness , Zhengzhong Liu , Eric Xing\n- **URL**: [https://arxiv.org/abs/2605.15290](https://arxiv.org/abs/2605.15290)\n- **Abstract**:\n> Hyperparameter transfer across model architectures dramatically reduces the amount of compute necessary for tuning large language models (LLMs). The maximal update parameterization ({\\mu}P) ensures transfer through principled mathematical analysis but can be challenging to derive for new model architectures. Building on the spectral feature-learning view of Yang et al. (2023a), we make two advances. First, we promote spectral norm conditions on the weights from a heuristic to the definition of feature learning, and as a consequence arrive at the Complete-P depth and weight-decay scalings without recourse to lazy-learning. Second, we consider a modified spectral norm that preserves the valid scaling law of network weights when weight matrices are not full rank. This enables (to our knowledge, the first) derivation of {\\mu}P scalings for grouped-query attention (GQA). We demonstrate the efficacy of our theoretical derivations by showing learning rate transfer across the GQA repetition hyperparameter as well as experiments regarding transfer over weight decay.\n\n---\n\n## 181. Universal Approximation of Nonlinear Operators and Their Derivatives\n- **Authors**: Filippo de Feo\n- **URL**: [https://arxiv.org/abs/2605.15285](https://arxiv.org/abs/2605.15285)\n- **Abstract**:\n> Derivative-Informed Operator Learning (DIOL), i.e. learning a (nonlinear) operator and its derivatives, is an open research frontier at the foundations of the influential field of Operator Learning (OL). In particular, Universal Approximation Theorems (UATs) of nonlinear operators and their derivatives are foundational open questions and delicate problems in nonlinear functional analysis. In this manuscript, we prove the first UATs of non-linear $k$-times differentiable operators between Banach spaces and their derivatives, uniformly on compact sets and in weighted Sobolev norms for general finite input measures, via OL architectures. Our results are the first complete generalizations of the corresponding influential classical results in [Hornik, 1991] to infinite-dimensional settings and OL. We discuss several open areas where DIOL and our UATs find applications: high-order accuracy in OL, fast constrained optimization in Banach spaces (e.g. optimal control of PDEs, inverse problems) and numerical methods for infinite-dimensional PDEs (e.g. HJB PDEs on Banach spaces from optimal control of PDEs, SPDEs, path-dependent systems, partially observed systems, mean-field control). We parameterize nonlinear operators via Encoder-Decoder Architectures, renowned classes in OL due to their generality, including classical architectures, such as DeepONets, Deep-H-ONets, PCA-Nets. Our results are based on four key features that allow us to prove UATs in full generality: (i) Approximation Properties of Banach spaces. (ii) $k$-times continuous differentiability in the sense of Bastiani (weaker than $k$-times continuous Fréchet differentiability). (iii) Natural compact-open topologies for UA; indeed, we show that UA in standard compact-open topologies induced by operator norms is violated even for Fréchet derivatives. (iv) Construction of novel weighted Sobolev spaces for the UA.\n\n---\n\n## 182. Autonomous Intelligent Agents for Natural-Language-Driven Web Execution with Integrated Security Assurance\n- **Authors**: Vinil Pasupuleti , Siva Rama Krishna Varma Bayyavarapu , Shrey Tyagi\n- **URL**: [https://arxiv.org/abs/2605.15281](https://arxiv.org/abs/2605.15281)\n- **Abstract**:\n> Modern web test suites rot. A UI refactor breaks locators, a timing change causes race conditions, and within weeks developers abandon the suite entirely. This paper presents an AI-driven autonomous testing framework that addresses these failure modes through five integrated strategies - navigation reliability, context-aware selector generation, post-generation validation, smart wait injection, and failure learning - implemented over a containerised worker architecture that decouples orchestration from long-running browser execution. Evaluated across four production applications and 176 scenarios, the framework improves script generation success from 55% to 93%, achieves an 8x reduction in navigation failures, eliminates 80% of timing-related race conditions, and reduces test creation time by 75% compared to manual Selenium authoring. The framework extends naturally to security validation: testers describe attack scenarios in plain English - \"try accessing another user's invoice\" - which the agent converts to OWASP Top 10-aligned browser probes, detecting 85% of authentication bypass vulnerabilities and 95% of input validation flaws with false positive rates below 12%. Natural-language-driven security testing of this kind represents, to our knowledge, a novel contribution to the field.\n\n---\n\n## 183. PDRNN: Modular Data-driven Pedestrian Dead Reckoning on Loosely Coupled Radio- and Inertial-Signalstreams\n- **Authors**: Peter Bauer , Andreas Porada , Felix Ott , Christopher Mutschler , Tobias Feigl\n- **URL**: [https://arxiv.org/abs/2605.15252](https://arxiv.org/abs/2605.15252)\n- **Abstract**:\n> Modern pedestrian dead reckoning (PDR) systems rely on fusing noisy and biased estimates of position, velocity, and calibrated orientation derived from loosely coupled sensors to determine the current pose of a localized object. However, discrepancies in the sampling rates of sensor-specific estimation methods and unreliable transmission pose significant challenges. And traditional methods often fail to effectively fuse multimodal sensor data during dynamic movements characterized by high accelerations, velocities, and rapidly varying orientations. To address these limitations, we propose a simple recurrent neural network (RNN) architecture capable of implicitly forecasting asynchronous sensor data streams from diverse estimation methods along reference trajectories. The proposed approach introduces PDRNN, a modular hybrid AI-assisted PDR system that handles each component as an independent ensemble of machine learning (ML) models to estimate both key parameter means and variances. Separate ML-based models are employed to estimate orientation, (un)directed velocity or distance from acceleration and gyroscope data, with optional absolute positioning from synchronized radio systems such as 5G for stabilization. A final fusion model combines these outputs, position, velocity, and orientation, while using uncertainty estimates to enhance system robustness. The modular design allows individual components to be updated, fine-tuned, or replaced without affecting the entire system. Experiments on dynamic sports movement data show that PDRNN achieves superior accuracy and precision compared to classic and ML-based methods, effectively avoiding error accumulation common in black-box approaches. And PDRNN offers forecast capabilities and better component control despite increased system complexity.\n\n---\n\n## 184. GQLA: Group-Query Latent Attention for Hardware-Adaptive Large Language Model Decoding\n- **Authors**: Fanxu Meng\n- **URL**: [https://arxiv.org/abs/2605.15250](https://arxiv.org/abs/2605.15250)\n- **Abstract**:\n> Multi-head Latent Attention (MLA), the attention used in DeepSeek-V2/V3, jointly compresses keys and values into a low-rank latent and matches the H100 roofline almost perfectly. Its trained weights, however, expose only one decoding path - an absorbed MQA form - which ties efficient inference to H100-class compute-bandwidth ratios, forfeits tensor parallelism along the head axis, and yields no Multi-Token Prediction (MTP) gain on commodity inference GPUs such as the export-restricted H20. We propose Group-Query Latent Attention (GQLA), a minimal modification of MLA whose trained weights expose two algebraically equivalent decoding paths over the same parameters: an MQA-absorb path identical to MLA's, and a GQA path with a per-group expanded cache. The runtime picks the path that matches the target hardware - no retraining, no custom kernels - so a single set of GQLA weights pins the rooflines of both H100 (MQA-absorb, s_q=1) and H20 (GQA + MTP, s_q=2), while supporting up to 8-way zero-redundancy tensor parallelism on the GQA path. To avoid pretraining from scratch we extend TransMLA into TransGQLA, which converts a pretrained GQA checkpoint into a GQLA model; on LLaMA-3-8B it compresses the per-token KV cache to 28.125% of the GQA baseline on the MQA-absorb path while structurally preserving GQA-level traffic on the per-group path.\n\n---\n\n## 185. Reading the Cell, Designing the Cure: Perturbation-Conditioned Molecular Diffusion for Function-Oriented Drug Design\n- **Authors**: Ziyu Xu , Zijian Zhang , Liang Wang , Zhiyuan Liu , Qiang Liu , Shu Wu , Liang Wang\n- **URL**: [https://arxiv.org/abs/2605.15243](https://arxiv.org/abs/2605.15243)\n- **Abstract**:\n> When reliable target structures are unavailable at scale or phenotypes arise from dysregulated pathways, transcriptomic perturbations provide a system-level functional readout for drug action. In this work, we formalize \\emph{Transcriptome-based Drug Design (TBDD)} as a generative inverse problem: designing drug molecules conditioned on desired transcriptomic state transitions. We analyze the inherently ill-posed nature of this task, which is further complicated by the profound domain gap between biology and chemistry and by the sparsity of transcriptomic signals. To address these challenges, we propose \\textbf{\\themodel{} } (A \\textbf{C}ell\\textbf{U}lar \\textbf{R}esponse \\textbf{E}ngine), a multi-resolution transcriptome-guided diffusion framework. \\themodel{} features a specialized \\textbf{Transcriptome Perturbation Functional Feature Extractor (TFE)} that (1) distills function-oriented perturbation embeddings from pre/post states, (2) aligns these signatures to dual chemical views to bridge the cross-modal gap, and (3) performs heterogeneity-aware aggregation to extract robust state-specific signals from noisy transcriptomic data. Extensive evaluations on both standard benchmarks and rigorous out-of-distribution protocols demonstrate that \\themodel{} consistently outperforms strong baselines in structural quality and functional consistency. Furthermore, we validate its practical utility via a zero-shot gene-inhibitor design task, highlighting the potential of phenotype-driven generative discovery.\n\n---\n\n## 186. Hydra: Efficient, Correct Code Generation via Checkpoint-and-Rollback Support\n- **Authors**: Alexander Du , Jianjun Ou , Danyang Zhuo , Matthew Lentz\n- **URL**: [https://arxiv.org/abs/2605.15238](https://arxiv.org/abs/2605.15238)\n- **Abstract**:\n> Large language models are increasingly used for code generation, but many generated programs fail to compile, a prerequisite for further correctness checks such as unit tests. Existing solutions for repairing static errors are costly in both latency and token consumption. Post-hoc repair delays error detection until generation completes and commonly regenerates large regions of previously valid code. Constrained semantic decoding checks after each token, incurring per-token overhead while limiting repair to the current token even when the root cause lies earlier. We present Hydra, a system for efficient recovery from static errors during code generation. Hydra allows checking to proceed asynchronously with generation, avoiding checker overhead when the generated code is semantically correct. In addition, it provides checkpoint-and-rollback support for targeted repair, avoiding regeneration and rechecking of valid prefixes. We retrofit the Clang C/C++ compiler to support Hydra with modest modifications. Paired with a token-efficient repair strategy, Hydra reduces latency by up to 71% and token consumption by up to 70% relative to post-hoc repair on C/C++ code generation tasks that encounter static errors.\n\n---\n\n## 187. A3D: Agentic AI flow for autonomous Accelerator Design\n- **Authors**: Abinand Nallathambi , Christopher Knight , Shantanu Ganguly , Wilfried Haensch , Anand Raghunathan\n- **URL**: [https://arxiv.org/abs/2605.15237](https://arxiv.org/abs/2605.15237)\n- **Abstract**:\n> Accelerating applications through the design of hardware accelerators can significantly enhance system performance and energy efficiency. Despite advances, such as high-level synthesis (HLS), designing accelerators for complex applications still remains highly labor-intensive, demanding considerable expertise in understanding workloads to be accelerated, hardware design, micro-architecture, and EDA tool usage, posing challenges for application domain experts. Therefore, most accelerator solutions are limited to applications with a regular predictable dataflow. Advances in AI have enabled agents that perform autonomous planning, reasoning, execution and reflection, leading to unprecedented potential for automation through agentic AI. We present A3D, an Agentic AI flow for end-to-end Automation of hardware Accelerator Design. A3D automates workload analysis, performance bottleneck identification, code refactoring for HLS compatibility and micro-architecture generation. A3D also generates diverse accelerator designs by automatically exploring the speed-area tradeoff space. Recent efforts have explored the use of AI for specific tasks such as design space exploration in HLS, leaving several tasks to still be performed manually. A3D addresses the challenges in applying modern LLMs to accelerator design by judiciously partitioning tasks among specialist agents, orchestrating process loops with specialist and verifier agents, utilizing pre-existing and custom tools, and employing agentic RAG for codebase and proprietary EDA tool documentation exploration. Our implementation of A3D, using commercial components like Claude Sonnet 4.5 and the Catapult HLS tool, demonstrates its effectiveness by generating accelerator designs with no human intervention from complex scientific applications like LAMMPS (molecular dynamics simulation) and QMCPACK (quantum chemistry).\n\n---\n\n## 188. Learning Selective Merge Policies for Deadline-Constrained Coded Caching via Deep Reinforcement Learning\n- **Authors**: Amirhossein Yousefiramandi\n- **URL**: [https://arxiv.org/abs/2605.15236](https://arxiv.org/abs/2605.15236)\n- **Abstract**:\n> With the coded caching, the server can use the information the users have cached to serve multiple users at a time by sending a single coded multi-casting message, i.e., the merged message, thereby relieving the peak network loads. However, for the delay-sensitive applications of the users, like the video streaming services, it becomes essential to choose which messages to merge online, considering the strict deadlines for each request. The problem, however, is that while the merge is helpful for the formation of the current coded multi-casting message, it can be harmful for the subsequent ones. We proposed a DRL-based solution that formulates the deadline-constrained coded delivery as a masked discrete-action queue-state control problem, while we trained a graph-attention policy network via proximal policy optimization. The policy network reduces the broadcast-packet expiration ratio $\\rho$ by $40.9%$ ($0.208$ vs. $0.352$) with respect to the best coded multi-casting baseline (SACM++) on the uniform-demand benchmark, while also attaining the best broadcast-efficiency score $\\sigma$ across the Track A battery among the coded multi-casting methods. The interesting fact we observed is that for the applications of the users with tight deadlines, the method of selective merging is better than the method of aggressive merging, i.e., the policy network learns to merge at only $\\approx 31.8%$ rate, even though the same observation holds across the variations within the same simulator family.\n\n---\n\n## 189. PBT-Bench: Benchmarking AI Agents on Property-Based Testing\n- **Authors**: Lucas Jing , Xinqi Wang , Liao Zhang , Simon S. Du\n- **URL**: [https://arxiv.org/abs/2605.15229](https://arxiv.org/abs/2605.15229)\n- **Abstract**:\n> Existing code benchmarks measure whether an agent can produce any test that reproduces a known bug, or whether it can produce a patch that fixes a described issue. Neither isolates the distinct skill of property-based testing: deriving a semantic invariant from documentation, and then constructing an input-generation strategy precise enough to make a random search reveal the violation. We introduce PBT-Bench, a benchmark of 100 curated property-based testing problems across 40 real Python libraries. Each problem injects one or more semantic bugs (365 in total, mean 3.65 per problem) designed so that default-strategy random inputs almost never trigger them; the agent must read the library's documentation, identify the relevant invariant, and specify a Hypothesis @given strategy that concentrates mass in the trigger region. Bugs are stratified across three difficulty levels (L1-L3) spanning single-constraint boundary bugs to stateful, cross-function protocol violations. We evaluate eight contemporary LLMs under two prompting regimes (open-ended baseline vs. explicit Hypothesis scaffolding) for three independent runs per configuration. Bug recall under the PBT-guided prompt ranges from 42.1% to 83.4% across models; under the open-ended baseline, from 31.4% to 76.7%. Hypothesis scaffolding lifts mid-capability models by over 20 percentage points, but yields smaller gains for the strongest models, with two exceptions showing degradation, suggesting the structured prompt can interfere with certain model behaviours rather than complementing them. The hardest bugs prove model-specific: different architectures fail on different problems, leaving persistent gaps that no single model closes. We release the benchmark, harness, and full evaluation corpus to support downstream work on documentation-grounded semantic reasoning.\n\n---\n\n## 190. Is Agentic AI Ready for Real-World Hardware Engineering? A Deep Dive with Phoenix-bench\n- **Authors**: Qingyun Zou , Feng Yu , Hongshi Tan , Bingsheng He , WengFai Wong\n- **URL**: [https://arxiv.org/abs/2605.15226](https://arxiv.org/abs/2605.15226)\n- **Abstract**:\n> We ask whether agentic AI systems built for software engineering transfer to realistic hardware engineering. Existing hardware LLM benchmarks isolate sub-tasks but none jointly requires repository navigation, hierarchy-aware localization, Electronic Design Automation (EDA) executable verification, and maintenance-style patching. We introduce \\textbf{Phoenix-bench}, a synchronized corpus of 511 verified Verilator instances from 114 GitHub repositories, each shipped with the developer patch, design-flow labels, fail-to-pass and pass-to-pass testbenches, and a Docker-pinned EDA environment so resolved-rate differences reflect agent behavior rather than toolchain availability. Using Phoenix-bench we run a uniform evaluation of four commercial agents and eight open-source agentic structures across four LLM backbones, plus two diagnostic interventions (file-level oracle localization and one round of testbench-log feedback). Three findings emerge. (i)~Software and hardware are fundamentally different engineering tasks: the same agent loses 37\\% to 58\\% from SWE-bench Verified to Phoenix-bench because hardware bugs propagate across parallel instantiated modules through signal flow rather than along a software-style call graph, and software-tuned agents stop at the symptom file instead of tracing back through the instantiation chain. (ii)~Failures concentrate on design control-flow / finite state machine (FSM) bugs, verification testbench bugs, and hard cases that demand cross-hierarchy signal-flow tracking and coordinated multi-file edits. (iii)~Localization granularity matters far more than localization itself: a perfect file-level oracle yields only $+1.4$\\% because the agent then breaks files that did not need editing, while a single round of test case feedback lifts resolved rate by $42$\\% to $45$\\% because the test case tells \\emph{where} the bug is and \\emph{what} the fix has to look like.\n\n---\n\n## 191. Do Biological Structural Guarantees Earn Their Complexity?\n- **Authors**: Bogdan Banu\n- **URL**: [https://arxiv.org/abs/2605.15225](https://arxiv.org/abs/2605.15225)\n- **Abstract**:\n> Biologically-inspired AI agent frameworks claim reliability benefits through structural guarantees adapted from gene regulatory networks, immune systems, and metabolic control. These claims are rarely tested empirically against simpler alternatives. We present three deep benchmarks: metabolic priority gating, autoinducer-based quorum sensing, and Bayesian stagnation detection, each comparing a biologically-grounded implementation against a naive non-biological alternative and an ablated control, across 1,000 trials per seed and 10 seeds (10M+ data points total).\n\n---\n\n## 192. GenAI-Driven Approach to RISC-V Supply Chain Exploration\n- **Authors**: Nenad Petrovic , Andre Schamschurko , Yingjie Xu , Alois Knoll\n- **URL**: [https://arxiv.org/abs/2605.15223](https://arxiv.org/abs/2605.15223)\n- **Abstract**:\n> This paper presents an LLM-empowered workflow for RISC-V supply chain analysis, integrating Vision-Language Models (VLMs) and Model-Driven Engineering (MDE) to enable comprehensive, multimodal data-driven insights. The proposed approach addresses the challenges of heterogeneous and unstructured supply chain data by leveraging LLMs for textual understanding and VLMs for extracting information from visual artifacts such as diagrams, tables, and scanned documents. These models collaboratively identify key entities and relationships, which are then organized into a knowledge graph representing supply chain components and their interdependencies. For analytical reasoning, the workflow incorporates MDE techniques and constraint-based modeling to enable formal validation of dependencies, detection of bottlenecks, and assessment of risks. The synergy between LLM- and VLM-based semantic understanding and MDE-based formal analysis supports both exploratory and systematic evaluation of supply chain resilience. A human-in-the-loop mechanism further enables interactive querying and expert validation. The approach is evaluated in RISC-V ecosystem scenarios, demonstrating its effectiveness in generating actionable insights, enhancing transparency, and supporting decision-making in complex semiconductor supply chains.\n\n---\n\n## 193. Effective Harness Engineering for Algorithm Discovery with Coding Agents\n- **Authors**: Yoichi Ishibashi , Taro Yano , Masafumi Oyamada\n- **URL**: [https://arxiv.org/abs/2605.15221](https://arxiv.org/abs/2605.15221)\n- **Abstract**:\n> AlphaEvolve and FunSearch have demonstrated the potential of combining large language models (LLMs) with evolutionary search for automated algorithm discovery. However, discovery success is shaped not only by model capability but also significantly by the design of the execution infrastructure, i.e., the harness. This paper investigates effective harness design through three questions: under a fixed token budget, is it better to produce many algorithms with brief thought or fewer algorithms with deeper thought? How should the harness handle evaluation hacks, where generated programs exploit the scoring function? And how can agents that require full filesystem access execute safely in parallel? Using Vesper, an algorithm discovery framework that incorporates harness improvements addressing these questions, we evaluate on Circle Packing under the same token budget. Interestingly, generating fewer algorithms while thinking more deeply about each one achieved higher scores. That is, scaling the quality of each individual is more budget-efficient than scaling the number of evolutionary generations. Surprisingly, more capable models produced evaluation hacks at higher rates, making hack detection increasingly necessary as models scale.\n\n---\n\n## 194. Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time\n- **Authors**: Michael Y. Hu , Apurva Gandhi , Kyunghyun Cho , Tal Linzen , Pratyusha Sharma\n- **URL**: [https://arxiv.org/abs/2605.15220](https://arxiv.org/abs/2605.15220)\n- **Abstract**:\n> Data mixing decides how to combine different sources or types of data and is a consequential problem throughout language model training. In pretraining, data composition is a key determinant of model quality; in continual learning and adaptation, it governs what is retained and acquired. Yet existing data mixing methods address only one phase of this lifecycle at a time: some require smaller proxy models tied to a single training phase, others assume a fixed domain set, and continual learning lacks principled guidance altogether. We argue that data mixing is fundamentally an online decision making problem -- one that recurs throughout training and demands a single, unified solution. We introduce OP-Mix (On-Policy Mix), a data mixing algorithm that operates across the entire language model training lifecycle. Our main insight is that candidate data mixtures can be cheaply simulated by interpolating between low-rank adapters trained directly on the current model, eliminating separate proxy models and ensuring the search is always grounded in the model's actual learning dynamics. Across pretraining, continual midtraining, and continual instruction tuning, OP-Mix consistently finds near-optimal mixtures while using a fraction of the compute of the baselines. In pretraining, OP-Mix improves upon training without mixing by 6.3% in average perplexity. For continual learning, OP-Mix matches the performance of both retraining and on-policy distillation while using 66% and 95% less overall compute, respectively. OP-Mix suggests a different view of language model training: not a sequence of distinct phases, but a single continuous process of learning from data.\n\n---\n\n## 195. An LLM-RAG Approach for Healthy Eating Index-Informed Personalized Food Recommendations\n- **Authors**: Yibin Wang , Yanjie Yang , Grace Melo Guerrero , Rodolfo M. Nayga Jr. , Azlan Zahid\n- **URL**: [https://arxiv.org/abs/2605.15213](https://arxiv.org/abs/2605.15213)\n- **Abstract**:\n> Diet quality is a leading determinant of chronic disease risk. Advances in artificial intelligence (AI) have enabled food recommendation systems to adapt suggestions to user preferences and health goals. However, most current systems rely on loosely curated food databases and provide limited connection to a validated index. In this study, we propose a Healthy Eating Index (HEI) informed retrieval-augmented generation (RAG) framework that combines standardized nutrition databases with large language models (LLMs) for personalized food recommendations. Our proposed method anchors retrieval in the National Health and Nutrition Examination Survey (NHANES) and the Food Patterns Equivalents Database (FPED). A food-level embedding space is constructed from FPED-derived textual descriptions. For each entity, the system computes baseline HEI scores, retrieves candidate foods for intake recommendations, and estimates the HEI impact of simple substitutions or additions. A constrained RAG pipeline instantiated with a pretrained OpenAI LLM generates personalized recommendations and sources based on nutrient profiles and HEI contributions. The simulation results showed a mean HEI improvement of 6.45, with the proportion of users HEI over 50 increasing from 45.12 to 61.26. Quantile analysis revealed consistent improved shifts across the HEI distribution. Our findings suggest that the proposed LLM-RAG-based AI systems can support more precise, explainable, and personalized nutrition guidance to improve diet quality.\n\n---\n\n## 196. Fault tolerance estimation in digital circuits with visualised generative networks\n- **Authors**: Sascha Biel , Carl Alexander Gaede , Amiel Glaser , Jan Wolter , Alexej Schelle\n- **URL**: [https://arxiv.org/abs/2605.15212](https://arxiv.org/abs/2605.15212)\n- **Abstract**:\n> We propose a new numerical method to estimate the fault tolerance of failure modes in digital circuit structures with a generative network sampling technique. From a random input of generated bitwise configurations of ideally digitalised analog currents in the digital circuit design with classical logical gates, expected output currents are compared to the realistic signals of a numerical experiment at the discriminator part of the Generative Adversarial Network (GAN) to calculate the deviation from ideal digital electronic signals, including various error modes, such as missing or interchanged logical devices. From the present analysis of a representation of the GAN in terms of complex variables, it is possible to evaluate the robustness in electronic designs by differentiating the impact of failure modes associated with different classical logical elements in the circuit.\n\n---\n\n## 197. Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels\n- **Authors**: Plawan Kumar Rath , Rahul Maliakkal\n- **URL**: [https://arxiv.org/abs/2605.15208](https://arxiv.org/abs/2605.15208)\n- **Abstract**:\n> Large Language Models are routinely compressed via post-training quantization to reduce inference costs and memory footprint for cloud and edge deployment, yet the impact of this compression on model quality remains poorly understood. Existing studies typically compare only two conditions (full-precision vs. a single quantized variant), rely on aggregate bias metrics, and evaluate a single model family, making it impossible to distinguish gradual degradation from threshold-dependent safety failures. We conduct a controlled empirical study of three instruction-tuned models (Qwen2.5-7B, Mistral-7B, Phi-3.5-mini) at five precision levels (BF16 through 3-bit) on 12,148 BBQ bias benchmark items across 5 random seeds, totaling 911,100 inference records. Our results reveal that 3-bit quantization causes 6-21% of previously unbiased items to develop new stereotypical behaviors, following a clear dose-response pattern confirmed via logistic regression, while models' willingness to select \"unknown\" answers declines by 17.4%. Crucially, these item-level changes are invisible to standard quality metrics: perplexity increases by less than 0.5% at 8-bit and under 3% at 4-bit across all three models, yet 2.5-5.6% of items already develop new biases at 4-bit. These findings demonstrate that aggregate evaluation metrics systematically miss fairness-critical degradation, underscoring the need for quality-aware compression protocols that explicitly test for bias emergence before deployment.\n\n---\n\n## 198. AgentStop: Terminating Local AI Agents Early to Save Energy in Consumer Devices\n- **Authors**: Dzung Pham , Kleomenis Katevas , Ali Shahin Shamsabadi , Hamed Haddadi\n- **URL**: [https://arxiv.org/abs/2605.15206](https://arxiv.org/abs/2605.15206)\n- **Abstract**:\n> Autonomous agents powered by large language models (LLMs) are increasingly used to automate complex, multi-step tasks such as coding or web-based question answering. While remote, cloud-based agents offer scalability and ease of deployment, they raise privacy concerns, depend on network connectivity, and incur recurring API costs. Deploying agents locally on user devices mitigates these issues by preserving data privacy and eliminating usage-based fees. However, agentic workflows are far more resource-intensive than typical LLM interactions. Iterative reasoning, tool use, and failure retries substantially increase token consumption, often expending significant compute without successfully completing tasks. In this work, we investigate the time, token, and energy overhead of locally deployed LLM-based agents on consumer hardware. Our measurements show that agentic execution increases GPU power draw, temperature, and battery drain compared to single-inference workloads. To address this inefficiency, we introduce AgentStop, a lightweight efficiency supervisor that predicts and preemptively terminates trajectories unlikely to succeed. Leveraging low-cost execution signals, such as token-level log probabilities, AgentStop can reduce wasted energy by 15-20% with minimal impact on task performance (<5% utility drop) for challenging web-based question answering and coding benchmarks. These findings position predictive early termination as a practical mechanism for enabling sustainable, privacy-preserving LLM agents on user devices. Our project code and data are available at this https URL .\n\n---\n\n## 199. Agent4POI: Agentic Context-Conditioned Affordance Reasoning for Multimodal Point-of-Interest Recommendation\n- **Authors**: Jinze Wang , Yangchen Zeng , Tiehua Zhang , Lu Zhang , Yuze Liu , Yongchao Liu , Xingjun Ma , Zhu Sun\n- **URL**: [https://arxiv.org/abs/2605.15203](https://arxiv.org/abs/2605.15203)\n- **Abstract**:\n> We introduce Agent4POI, the first POI recommendation framework that generates context-conditioned multimodal representations at recommendation time, rather than relying on static POI embeddings pre-computed independently of context. Existing multimodal systems encode each POI once as a static embedding, a design that precludes reasoning about why the same cafe affords solo work on Monday but group celebration on Friday evening. We formally prove that no pre-computed encoder can satisfy context-sensitive ranking under standard bilinear scoring, motivating inference-time item-side representation. Agent4POI inverts this computation: given a situational context, a four-phase LLM agent generates dynamic, context-specific affordance queries (Phase 1) and executes a five-step cross-modal chain-of-thought over image, review, and metadata evidence (Phase 2). The resulting uncertainty-aware affordance representation is grounded in Gibsonian affordance theory. These cross-modal verdicts form a structured, uncertainty-adjusted affordance representation (Phase 3), which is aligned with user preferences via a semantic caching system for low-latency ranking (Phase 4). On three POI benchmarks and three evaluation configurations (standard, cold-start, context-shift), Agent4POI achieves a 23.2% relative gain over the strongest baseline and degrades by only 7.5% under context-shift versus 16--17\\% for the strongest baselines. In cold-start scenarios, Agent4POI outperforms the best content-based baseline by up to 2.4x, whereas ID-based methods fail to generalize.\n\n---\n\n## 200. Ensuring Logic in the Fog: Sound POMDP Synthesis with LTL Objectives\n- **Authors**: Can Zhou , Yulong Gao , Pian Yu\n- **URL**: [https://arxiv.org/abs/2605.12581](https://arxiv.org/abs/2605.12581)\n- **Abstract**:\n> Synthesising autonomous agents that can navigate uncertain environments while adhering to complex temporal constraints remains a fundamental challenge. While Linear Temporal Logic (LTL) provides a rigorous language for specifying such tasks, the inherent undecidability of qualitatively verifying LTL satisfaction in partially observable Markov decision processes renders quantitative synthesis difficult, especially when designing reliable reward signals for approximate solvers. In this paper, we bridge this gap with a novel, sound reward-shaping mechanism that dynamically generates belief-dependent rewards grounded in certified LTL satisfaction. By integrating this mechanism into an enhanced Monte Carlo Planning framework, we empower agents to navigate the `fog' of partial observability with a search process focused on maximising verifiable success. Our experiments demonstrate that this approach not only thrives in scenarios where existing solvers fail but also maintains effectiveness and scalability across diverse benchmark domains.\n\n---\n\n{% endraw %}\n"}}, "LLM/2026-05-27.md": {"64bbd1731dde152451afe1705df222e26b5c83fa": {"t": 1785747201.2847772, "n": 210239, "h": "cf2fc0302148c99f", "c": "---\nlayout: default\ntitle: \"LLM 관련 주요 논문 - 2026-05-27\"\n---\n\n{% raw %}\n# LLM 관련 주요 논문 - 2026-05-27\n\n## 1. MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation\n- **Authors**: Huawei Lin , Peng Li , Jie Song , Fuxin Jiang , Tieying Zhang\n- **URL**: [https://arxiv.org/abs/2605.27366](https://arxiv.org/abs/2605.27366)\n- **Abstract**:\n> Large language model (LLM) agents rely on reusable skills to solve complex tasks. However, existing skill creation approaches treat skills as isolated and static artifacts, limiting their reusability, reliability, and long-term improvement. We propose MUSE-Autoskill Agent (Memory-Utilizing Skill Evolution), a skill-centric agent framework that lets agents continuously improve their task-solving capability by creating, reusing, and refining skills under a unified lifecycle (creation, memory, management, evaluation, and refinement). Our framework enables agents to create skills on demand, store and reuse them across tasks, organize and select them efficiently, and evaluate them through unit tests and runtime feedback for continuous refinement. We further introduce skill-level memory that accumulates experience for each skill across tasks, enabling more effective reuse and adaptation over time. Experiments on SkillsBench provide initial evidence that lifecycle-managed skills can improve task success, efficiency, reuse, and cross-agent transfer, highlighting the importance of treating skills as long-lived, experience-aware, and testable assets.\n\n---\n\n## 2. Natural Language Query to Configuration for Retrieval Agents\n- **Authors**: Melissa Z. Pan , Negar Arabzadeh , Mathew Jacob , Fiodar Kazhamiaka , Esha Choukse , Matei Zaharia\n- **URL**: [https://arxiv.org/abs/2605.27361](https://arxiv.org/abs/2605.27361)\n- **Abstract**:\n> Modern retrieval agents expose many configuration choices -- LLM, retriever, number of documents, number of hops, and synthesis strategy -- each shaping both answer quality and serving cost. Today, these pipelines are typically hand-tuned once per workload, leaving substantial per-query optimization untapped. We formulate the problem: given a natural-language query and either an accuracy or a budget target, select from a predefined pipeline catalog the configuration that minimizes cost or maximizes accuracy at inference time. We propose **BRANE**, which uses an LLM to convert each query into workload-specific characteristics, then trains a lightweight per-configuration predictor that estimates whether the pipeline will answer the query correctly. At inference time, **BRANE** selects the configuration that maximizes predicted correctness penalized by cost, exposing a tunable cost-quality tradeoff without retraining. Across MuSiQue, BrowseComp-Plus, and FinanceBench, **BRANE** consistently pushes the cost-quality Pareto frontier, matches the best fixed configuration's accuracy at up to 89% lower cost, and outperforms LLM-routing, rule-based, and fine-tuned Qwen3-4B baselines. These results show that per-query configuration of the full retrieval pipeline is a practical alternative to static workload-level tuning.\n\n---\n\n## 3. Alignment Tampering: How Reinforcement Learning from Human Feedback Is Exploited to Optimize Misaligned Biases\n- **Authors**: Dongyoon Hahm , Dylan Hadfield-Menell , Kimin Lee\n- **URL**: [https://arxiv.org/abs/2605.27355](https://arxiv.org/abs/2605.27355)\n- **Abstract**:\n> Reinforcement Learning from Human Feedback (RLHF) is the standard method to align Large Language Models (LLMs) with human preferences. In this work, we introduce alignment tampering, a potential vulnerability where the LLM undergoing alignment influences the preference dataset, causing RLHF to amplify undesired behaviors. This arises from core limitations of RLHF: (1) preference datasets are constructed from the LLM's own outputs, allowing it to influence them, and (2) pairwise comparisons only indicate which response is better, not why. These limitations can be exploited to cause alignment tampering. For example, if an LLM generates biased responses with higher quality, annotators will prefer them based on quality. However, preference labels do not distinguish quality from bias, and the reward model inherits this limitation. Optimizing such rewards through reinforcement learning or best-of-N sampling can amplify misaligned biases. Our experiments demonstrate amplification across diverse biases: from keyword bias to propaganda (e.g., sexism), brand promotion, and instrumental goal-seeking. Mitigation remains challenging, as existing techniques for robust RLHF fail to fully resolve alignment tampering without sacrificing response quality. These findings reveal structural vulnerabilities of current RLHF and emphasize the need to prevent this vulnerability. Project page: this https URL\n\n---\n\n## 4. Gumbel Machine: Counterfactual Student Writing Generation via Gumbel Noise Steering\n- **Authors**: Hunter McNichols , Alexander Scarlatos , Mihai Dascalu , Danielle McNamara , Andrew Lan\n- **URL**: [https://arxiv.org/abs/2605.27249](https://arxiv.org/abs/2605.27249)\n- **Abstract**:\n> An effective method of teaching across disciplines is to provide examples of high-quality work. However, an example may be significantly different from a student's current work, making it challenging for them to emulate. An ideal learning demonstration is a counterfactual version of the student work, an improved version that is still similar to their own. Existing automated approaches for counterfactual text generation using Large Language Models (LLMs) result in domain-specific systems that are difficult to translate into practical applications. We present the Gumbel Machine, a flexible, modular approach to generating counterfactuals that leverages LLM instruction-following capabilities while encouraging similarity to a reference factual text. Central to our approach is a novel, controlled decoding algorithm, $\\beta$-Hindsight control, which uses latent randomness as a tunable similarity control mechanism during counterfactual generation. Experiments on datasets of student writing, scored on various criteria, demonstrate the effectiveness of our approach at generating counterfactuals both rubric-consistent and similar to a reference.\n\n---\n\n## 5. Learning to Act under Noise: Enhancing Agent Robustness via Noisy Environments\n- **Authors**: Yuxin Chen , Xiaodong Cai , Junfeng Fang , Zhuowen Han , Yu Wang , Yaorui Shi , Yi Zhang , Qi Gu , Xunliang Cai , Xiang Wang , An Zhang , Tat-Seng Chua\n- **URL**: [https://arxiv.org/abs/2605.27209](https://arxiv.org/abs/2605.27209)\n- **Abstract**:\n> Recent advances in large language models (LLMs) have facilitated the widespread deployment of LLMs as interactive agents capable of reasoning, planning, and tool use. Despite strong performance on existing benchmarks, such agents often exhibit notable degradation when deployed in real-world settings, where environments are inherently stochastic and imperfect. We argue that this discrepancy arises from a fundamental mismatch between idealized training settings and real-world interaction dynamics, where current paradigms rely on carefully curated task instructions and stable, well-controlled environments. To address this gap, we propose NoisyAgent, an agentic training framework that explicitly incorporates environmental imperfections into the agent learning process. We identify two major sources of interaction noise in real-world scenarios: user noise, which captures ambiguity and variability in user interaction, and tool noise, which reflects failures and anomalies in tool execution. We introduce such perturbations into the training pipeline by modifying user interaction patterns and simulating tool execution results within the training environment. To stabilize training while encouraging agents to handle increasingly challenging imperfections, noise is applied to only a subset of rollouts and progressively increased in difficulty as the model adapts to the current noise level. Extensive experiments demonstrate that our approach consistently improves agent robustness under noisy and dynamic environments. Our analysis reveals that training under noise conditions also yields performance gains on idealized benchmarks, suggesting that controlled exposure to environmental noise promotes more generalizable reasoning and decision-making behaviors. Our findings highlight the importance of modeling interaction imperfections for bridging the gap between agent training and real-world deployment.\n\n---\n\n## 6. The Compressive Knowledge Graph Hypothesis: Which Graph Facts Matter for Scientific Hypothesis Generation?\n- **Authors**: Shashwat Sourav , Viktoriia Baibakova , Sanjay Das , Ran Elgedawy , Maria Mahbub , Emily Herron , Tirthankar Ghosal\n- **URL**: [https://arxiv.org/abs/2605.27176](https://arxiv.org/abs/2605.27176)\n- **Abstract**:\n> Knowledge graphs (KGs) can provide structured scientific context to language models, but it remains unclear which graph facts actually shape the generated hypotheses. We study KG-guided hypothesis generation for battery materials across Mistral-7B, Llama-3.1-70B, and Gemini 2.5 Flash. We perturb local KGs by varying density, ontology richness, topology, and control structure, and evaluate outputs with both provided-graph and fixed-reference metrics. Across models, KG utility is selective and model-dependent: graph context changes outputs, but no-KG outputs also recover substantial graph content from model priors. Compact top-k subgraphs often approximate full-KG behavior, including when claimed-outcome triples are held out. At the same time, compression is not unique to one semantic ranking rule, random and topology-based subsets can also recover much of the signal. These results support a redundancy-aware Compressive KG hypothesis: useful KG signal is often recoverable from compact, scientifically structured subgraphs rather than requiring the full local graph.\n\n---\n\n## 7. VitaBench 2.0: Evaluating Personalized and Proactive Agents in Long-Term User Interactions\n- **Authors**: Yuxin Chen , Yi Zhang , Zhengzhou Cai , Yaorui Shi , Zhiyuan Yao , Chenhang Cui , Jingnan Zheng , Yaqi Huo , Xi Su , Qi Gu , Xunliang Cai , Xiang Wang , An Zhang , Tat-Seng Chua\n- **URL**: [https://arxiv.org/abs/2605.27141](https://arxiv.org/abs/2605.27141)\n- **Abstract**:\n> Large language models (LLMs) have evolved into interactive agents that collaborate with users in real-world tasks. Effective collaboration in such settings increasingly depends on understanding the user beyond what is explicitly stated, as user intent is often reflected in fragmented daily interactions and requires both personalized modeling and proactive interaction. However, existing agent benchmarks primarily evaluate reasoning and tool use, largely overlooking the challenges of inferring and leveraging user preferences in realistic scenarios. To address this gap, we introduce VitaBench 2.0, a benchmark for evaluating personalized and proactive agent behavior in long-term user interactions. In VitaBench 2.0, tasks are organized as temporally ordered sequences for individual users, where preferences are embedded in fragmented and heterogeneous interactions. Successful completion of tasks requires the agent to continuously extract, utilize, and update user preferences from these interactions. We further evaluate proactiveness through tasks that require agents to recognize missing information and actively acquire it from users or environments before making decisions. To support systematic analysis, we provide an extensible memory interface that enables controlled comparison across different memory architectures. We benchmark a diverse set of frontier proprietary and open-source LLMs. Results show that real-world personalization remains highly challenging even for state-of-the-art models, revealing a substantial gap between current capabilities and practical requirements. Extensive analysis further reveals the failure modes and capability bottlenecks of current agents in real-world personalized decision-making, providing insights for future model improvements.\n\n---\n\n## 8. ICCU: In-Context Continual Unlearning via Pattern-Induced Refusal Rules\n- **Authors**: Ruihao Pan , Suhang Wang\n- **URL**: [https://arxiv.org/abs/2605.27138](https://arxiv.org/abs/2605.27138)\n- **Abstract**:\n> Machine unlearning aims to remove the influence of specific data from trained language models. In real-world deployments, unlearning requests often arrive sequentially, which challenges existing fine-tuning-based methods: fine-tuning each request is costly, accumulates utility loss, and may cause cross-request interference. To address these issues, we propose ICCU (In-Context Continual Unlearning), an in-context continual unlearning framework that induces readable refusal rules from unlearning datasets and applies them at inference time either as a filter or via the system prompt, without modifying model parameters. Because rules are accumulated as an order-independent union, ICCU is compositional and free of cross-request interference, and the original forget-set data can be discarded after rule induction. Extensive experiments show that ICCU effectively suppresses target knowledge while preserving utility, scales across sequential requests, and remains robust to paraphrased and cross-lingual queries.\n\n---\n\n## 9. Scaling, Benchmarking, and Reasoning of Vision-Language Agents for Mobile GUI Navigation\n- **Authors**: Heng Qu , Yike Liu , Renren Jin , Wenzong Zhang , Pengzhi Gao , Wei Liu , Jian Luan\n- **URL**: [https://arxiv.org/abs/2605.27134](https://arxiv.org/abs/2605.27134)\n- **Abstract**:\n> Vision-Language Models (VLMs) have shown rapid progress in mobile GUI navigation. This paper presents a systematic study of data scaling, benchmarking, and reasoning for VLM-based agents in this domain. To facilitate rigorous evaluation, we introduce HyperTrack, a large-scale dataset with over 16000 real-world tasks across more than 650 Chinese mobile applications, along with GUIEvalKit, an open-source toolkit for unified benchmarking of VLMs on offline GUI navigation tasks. Using HyperTrack, we analyze the effects of training data scale on both supervised and reinforcement-based finetuning. Our results show that reinforcement-based finetuning consistently outperforms supervised finetuning, particularly in out-of-domain settings, highlighting the synergy between data scaling and reinforcement learning. Leveraging GUIEvalKit, we further benchmark state-of-the-art (SOTA) VLMs and analyze how interaction history and reasoning capabilities influence task completion. Together, HyperTrack and GUIEvalKit provide a comprehensive platform for developing and evaluating VLM agents in mobile GUI navigation tasks.\n\n---\n\n## 10. Position: AI Safety Requires Effective Controllability\n- **Authors**: Yige Li , Yunhao Feng , Jun Sun\n- **URL**: [https://arxiv.org/abs/2605.27117](https://arxiv.org/abs/2605.27117)\n- **Abstract**:\n> AI safety is still largely framed as alignment: training models to follow human preferences, safety policies, and normative constraints. That framing has improved the behavior of modern language models, but aligned behavior does not by itself guarantee that a deployed agent can be stopped, overridden, or constrained once it operates in open-ended, interactive, and tool-using environments. A system may be safe in expectation and still fail to yield to explicit runtime authority under conflicting instructions, long-horizon execution, adversarial inputs, or risky tool use. This position paper argues that AI safety therefore requires controllability as a first-class objective. We define \\emph{controllability} as the ability of an AI system to remain reliably interruptible, overridable, redirectable, and constrainable by explicit control signals at runtime while preserving ordinary utility when such signals are absent. To study this gap, we introduce \\controlbench{}, a benchmark for evaluating controllability failures in high-risk agentic scenarios. Experiments with OpenClaw-based agents show that current alignment and guardrail mechanisms reduce risk, but often fail to provide persistent, authoritative, and enforceable runtime control. We therefore propose a control-centric architectural framework that highlights explicit control planes, runtime intervention pathways, persistent control states, and auditable decision interfaces as key design principles for future controllable AI systems.\n\n---\n\n## 11. Counteraction-Aware Multi-Teacher On-Policy Distillation for General Capability Recovery with Domain Preservation\n- **Authors**: Tianlei Chen , Jiao Ou , Ziyuan Liu , Ruiming Tang , Jian Liang , Han Li\n- **URL**: [https://arxiv.org/abs/2605.27115](https://arxiv.org/abs/2605.27115)\n- **Abstract**:\n> Domain specialization can improve LLM behavior in vertical domains, but often weakens the general capabilities inherited from the original model. Recent Multi-Teacher On-Policy Distillation (MOPD) pipelines recover model capabilities by supervising student-generated trajectories with teacher feedback, but typically assume teacher-aligned prompt coverage, requiring prompts to match the teachers' training distributions. This assumption is difficult to satisfy when the general teacher is an open-source model whose post-training data are unknown. Instead of attempting to reconstruct this hidden distribution, we study general capability recovery with readily available proxy general prompts. We identify two failure modes of vanilla MOPD in this incomplete-coverage situation: recovery-preservation counteraction from mixing conflicting recovery and preservation gradients, and weak-signal flattening from uniformly averaging samples with unequal correction demand. We propose Counteraction-Aware Multi-Teacher On-Policy Distillation (CaMOPD), which addresses these issues with decoupled alternating training and gap-based sample selection. CaMOPD gives general recovery dedicated updates, periodically reviews domain prompts for preservation, and selects samples with larger averaged token-level teacher-student log-probability gaps to concentrate correction signals. Across role-play dialogue and medical reasoning QA scenarios, CaMOPD performs best in general recovery over baselines while maintaining domain-specific behavior. Gradient coherence analyses further support the intended effect of CaMOPD in producing more coherent correction signals.\n\n---\n\n## 12. Traceable Knowledge Graph Reasoning Enables LLM-Assisted Decision Support for Industrial VOCs in the Steel Industry\n- **Authors**: Changqing Su , Yu Ding , Zuhong Lin , Hongyu Liu , Xi He , Zheng Zeng , Liqing Li\n- **URL**: [https://arxiv.org/abs/2605.27071](https://arxiv.org/abs/2605.27071)\n- **Abstract**:\n> Key knowledge for steel-industry volatile organic compounds (VOCs) governance is scattered across unstructured scientific literature, making it difficult to integrate process, pollutant, and control-technology evidence and increasing the risk of hallucination when general large language models (LLMs) answer low-frequency industrial questions. Here we developed Chat-ISV, a knowledge graph (KG) enhanced multi-agent Q&A system that parses a curated steel-industry VOCs literature corpus, constructs a Neo4j KG with 27180 nodes and 81779 semantic edges, and combines prompt-constrained extraction, chunk-centered topology optimization, multi-agent routing, source-backtracking retrieval, local literature retrieval, open-domain knowledge access, and interactive subgraph visualization. Benchmark tests and 400 expert blind evaluations showed that topology optimization reduced isolated nodes from 57% to 4.08% and that Chat-ISV achieved high factual reliability, with 96.93% precision, 72.63% recall, an F1-score of 0.830, and a mean score of 1.69/2.00. By converting fragmented environmental-engineering literature into traceable, queryable, and decision-support-oriented knowledge, Chat-ISV establishes a scalable environmental-informatics paradigm for reliable LLM deployment and intelligent pollution-control decision support in specialized industrial domains.\n\n---\n\n## 13. Generating Robust Portfolios of Optimization Models using Large Language Models\n- **Authors**: Eleni Straitouri , Cheol Woo Kim , Milind Tambe\n- **URL**: [https://arxiv.org/abs/2605.27013](https://arxiv.org/abs/2605.27013)\n- **Abstract**:\n> Mathematical optimization is a powerful tool for structured decision-making across domains such as resource allocation and planning. Formulating optimization models faithful to reality, though, remains a significant bottleneck as it typically demands both domain expertise and optimization knowledge that are often scarce. Recent advances in large language models (LLMs) promise to bridge this gap, enabling the generation of candidate optimization models from natural language descriptions. However, there is no guarantee that any single LLM-generated model is reliable, and existing approaches that output only one model are therefore risky. In this work, we propose a novel algorithm that generates a portfolio of optimization models, designed to be robust to the limitations of LLMs. Our method exploits the observation that a single LLM can play two distinct roles $\\unicode{x2014}$ as a stochastic generator and as a reasoning evaluator $\\unicode{x2014}$ and proposes a unified framework that leverages both capabilities in a complementary manner. We provide theoretical guarantees showing that, as long as either the generator or the evaluator is well-aligned with human preferences, the portfolio is guaranteed to contain high-quality candidates, enabling a principled human-in-the-loop process in which a decision-maker can review multiple candidates before committing to one. We further validate our approach empirically, demonstrating strong performance across a range of optimization modeling tasks.\n\n---\n\n## 14. LELA: An End-to-end LLM-based Entity Linking Framework with Zero-shot Domain Adaptation\n- **Authors**: Samy Haffoudhi (IP Paris, LTCI, DIG), Nikola Dobričić (IP Paris), Fabian Suchanek (IP Paris, LTCI), Nils Holzenberger\n- **URL**: [https://arxiv.org/abs/2605.26956](https://arxiv.org/abs/2605.26956)\n- **Abstract**:\n> Entity linking is a key component of many downstream NLP systems, yet existing approaches are often tied to the specific target knowledge bases and domains, limiting their real world application. In this paper, we extend LELA, a modular and domain-agnostic LLM-based entity disambiguation method, into a practical Python library that integrates zero-shot Named Entity Recognition (NER) -thereby providing a complete end-toend pipeline for entity-linking in real-world usage. We provide experimental results validating LELA's performance and robustness across diverse entity linking settings. In our demo, users can play with the system on their own input texts.\n\n---\n\n## 15. Neuro-Symbolic Verification of LLM Outputs for Data-Sensitive Domains (extended preprint)\n- **Authors**: Paul Sigloch , Christoph Benzmüller\n- **URL**: [https://arxiv.org/abs/2605.26942](https://arxiv.org/abs/2605.26942)\n- **Abstract**:\n> LLMs deployed in high-stakes domains face fundamental reliability challenges: hallucinations, inconsistencies, and privacy vulnerabilities introduce unacceptable risks where errors carry legal, financial, or safety consequences. This paper presents a hybrid verification architecture combining formal symbolic methods with neural semantic analysis to provide complementary guarantees for LLM-generated content. This architecture employs logical reasoning for input verification, leveraging completeness properties to provide decidable guarantees on structured requirements. For output validation, embedding-based semantic similarity detects contextual hallucinations where formal methods lack expressiveness. This separation is realized in a parallel, actor-based pipeline, addressing limitations of prompt-based self-verification approaches, which inherit the distributional biases that produce hallucinations. The proposed architecture and type-aware verification method are validated with HAIMEDA, a real-world medical device damage assessment reporting system developed through Action Design Research. Evaluation shows hallucination detection rates of over 83% for structured entities and 72% for semantic fabrications, with a 30% reduction in report creation time, demonstrating that neuro-symbolic architectures can provide principled safeguards for LLM deployment in data-sensitive domains.\n\n---\n\n## 16. From Norms to Indicators (N2I-RAG): An Agentic Retrieval-Augmented Generation Framework for Legal Indicator Computation\n- **Authors**: Youssef Al Mouatamid , Marie Bonnin , Jihad Zahir\n- **URL**: [https://arxiv.org/abs/2605.26926](https://arxiv.org/abs/2605.26926)\n- **Abstract**:\n> Computing legal indicators from normative texts is a key task in legal monitoring and policy evaluation, but presents significant challenges due to the complexity, scale, and interpretive nature of legal language, as well as the variability in available document quality. Existing natural language processing techniques and generative models can assist in legal analysis, but often suffer from high risk of hallucinations and lack the interpretability and evidence grounding required for reliable indicator computation. This paper presents N2I-RAG (From Norms to Indicators), an agentic retrieval-augmented generation framework designed to automate the computation of legal indicators in a transparent and traceable way. We integrate adaptive retrieval, llm-based agents, and validation mechanisms in a modular pipeline, where each component performs a defined role in filtering, retrieving, and assessing evidence, and in producing binary legal outcomes linked to identifiable legal provisions. The framework emphasizes traceability by requiring explicit explanations of intermediate decisions and final indicator assignments. We evaluate N2I-RAG using an in-house constructed French marine environmental law corpus that includes both scanned and digital sources. Comparative experiments with multiple language model families demonstrate that the proposed approach consistently outperforms baseline systems, and generalizes well when tested on 2 different bans. The results indicate that agentic retrieval-augmented generation can bridge open-text legal language and standardized indicator computation, offering a foundation for transparent and scalable legal observatories.\n\n---\n\n## 17. TADDLE: A Tool-Augmented Agent for Detecting Deficient LLM-Generated Peer Reviews\n- **Authors**: Hanqi Duan , Xiang Li\n- **URL**: [https://arxiv.org/abs/2605.26911](https://arxiv.org/abs/2605.26911)\n- **Abstract**:\n> LLM-generated peer reviews are increasingly common at major venues, yet their deficiencies are hard to detect because they are uniformly fluent and well-structured. Existing work either classifies authorship without judging quality, or scores quality with features designed for human-written reviews; no prior system detects deficiencies in LLM-generated reviews at the level of individual defect types. To bridge the gap, we introduce TADDLE, a Tool-Augmented Agent for Detecting Deficient LLM-Generated Peer Reviews, together with the first expert-annotated benchmark for this task. Our benchmark comprises 1,800 reviews on 50 ICLR 2025 papers, multi-label-annotated by 18 domain experts against a taxonomy of six defect categories (plus a non-deficient label). TADDLE decomposes detection into four specialized analysis tools -- Verify, Correct, Complete, and Transform -- orchestrated by an agent; an integrator synthesizes their outputs into binary and multi-label classifications via two-stage semi-supervised learning. Extensive experiments show that TADDLE performs strongly on both binary detection and the multi-label classification task. We release the benchmark and code at this https URL .\n\n---\n\n## 18. Multi-Stakeholder LLM Alignment: Decomposing Estimation from Aggregation\n- **Authors**: Lulu Zheng , Wenjin Yang , Xiangwen Zhang , Rong Yin , Yulan Hu , Zheng Pan , Xin Li\n- **URL**: [https://arxiv.org/abs/2605.26878](https://arxiv.org/abs/2605.26878)\n- **Abstract**:\n> Multi-stakeholder tasks require one output to satisfy users with conflicting preferences. Holistic LLM judges conflate utility estimation and utility aggregation, yielding unstable implicit weights. We show empirically and theoretically that this aggregation-specific \\emph{weighting noise} can create large score shifts when stakeholder satisfaction is dispersed; in our experiments, these weight-induced shifts also increase with stakeholder count. We propose \\textsc{DecompR}: counterfactual-calibrated weights are fixed from query structure before candidate scoring, while per-role utilities are estimated independently, removing candidate-dependent weight drift and reducing estimation noise.\n\n---\n\n## 19. Helicase: Uncertainty-Guided Supply Chain Knowledge Graph Construction with Autonomous Multi-Agent LLMs\n- **Authors**: Yunbo Long , Haolang Zhao , Ge Zheng , Alexandra Brintrup\n- **URL**: [https://arxiv.org/abs/2605.26835](https://arxiv.org/abs/2605.26835)\n- **Abstract**:\n> LLM-based multi-agent systems have been widely adopted for knowledge retrieval and report generation, synthesizing known information through web search and textual reasoning. However, many critical information tasks in supply chains are not simple one-shot queries: they are structural inference problems requiring multi-hop reasoning across complex, fragmented web resources. Questions such as \\textit{``Which Tesla components use lithium from Australian mines?''} have no answer in any single document; answers must be computationally synthesized through the autonomous construction and analysis of dynamic knowledge graphs assembled from fragmented, heterogeneous sources. Moreover, such discovery processes must be uncertainty-aware: decisions depend not only on answers but on calibrated confidence in their reliability, traceable to source quality and reasoning consistency. To address this capability gap, we propose \\textit{Helicase}, an autonomous multi-agent LLM system for uncertainty-guided supply chain knowledge graph construction. \\textit{Helicase} decomposes high-level supply-chain queries into executable investigation plans, coordinates specialized web-search, reasoning, and coding agents through iterative verification loops, and incrementally constructs query-specific supply chain knowledge graphs with per-fact uncertainty annotations. Its three-layer uncertainty framework tracks uncertainty at the action, trajectory, and memory layers, enabling both structural inference and calibrated confidence assessment. To evaluate autonomous reasoning across the full complexity spectrum, we introduce SCQA (Supply Chain Query Assessment), a benchmark of 80 supply chain queries organized into four quadrants spanning single-hop to multi-hop inference under both high and low data visibility.\n\n---\n\n## 20. The Attribution Blind Spot: Detecting When Language Models Rely on Memory Rather Than Retrieved Context\n- **Authors**: Zhe Yu , Wenpeng Xing , Yunzhao Wei , Bo Yang , Chen Ye , Gaolei Li , Meng Han\n- **URL**: [https://arxiv.org/abs/2605.26778](https://arxiv.org/abs/2605.26778)\n- **Abstract**:\n> Retrieval-augmented generation promises to ground language model outputs in external evidence, yet the field has no reliable way to verify whether retrieved context actually governs generation -- a prerequisite for any high-stakes deployment. The standard assumption, that context-consistent output implies context-governed output, breaks when the retrieved document overlaps with the model's pretraining data: the model can produce faithful-looking text entirely from parametric memory, and both pathways yield indistinguishable output. We name this failure the attribution blind spot and introduce Computational Reality Monitoring (CRM) to address it. CRM operationalizes a principle adapted from cognitive science's reality monitoring framework: comparing internal representations with and without context reveals membership-conditioned representational divergence that output-level monitors systematically miss. CRM does not certify which source an individual generation used; it detects whether pretraining exposure leaves a measurable internal trajectory signature, establishing a necessary substrate for source attribution. Across nine model variants spanning three families, this divergence concentrates in architecture-specific layer patterns, receives converging support from block-level noise intervention, and generalizes across tasks and datasets while collapsing on domain-confounded benchmarks. The attribution blind spot is measurable and partially addressable: internal representations carry a diagnostic signal invisible at the output level, establishing a foundation for systems whose internal awareness of evidence provenance governs their external behavior.\n\n---\n\n## 21. A Dataset of Robot-Patient and Doctor-Patient Medical Dialogues for Spoken Language Processing Tasks\n- **Authors**: Heriberto Cuayahuitl , Grace Jang\n- **URL**: [https://arxiv.org/abs/2605.26747](https://arxiv.org/abs/2605.26747)\n- **Abstract**:\n> Large Language Models (LLMs) have brought huge improvements to Artificial Intelligence (AI), which can be applied to general-purpose tasks. However, their application to textual or spoken medical consultations is still an open research problem. This paper proposes MeDial-Speech, a novel speech dataset for training and evaluating Med-AIs that can carry out consultations with patients. It was collected in realistic environments from robot-patient and doctor-patient dialogues, contains 111+ hours of speech data (without data augmentation), and covers four health conditions: Lewy body dementia, heart failure, shoulder pain, and angina. In addition, we propose a dialogue benchmark via sentence selection (with 20 options) to evaluate three state-of-the-art LLMs: GPT-5 mini, DeepSeek-V3, and Claude Sonnet 4. Experimental results reveal that Claude Sonnet 4 is the best in sentence selection, with 71.1% accuracy using manual transcriptions and 74.7% using automatic transcriptions, and that all LLMs are highly overconfident in their probabilistic predictions, regardless of selecting correct or incorrect sentences in medical dialogues. This dataset is free of charge for non-commercial purposes at: this https URL\n\n---\n\n## 22. It's Not the Capability: Harness Sensitivity Is Non-Monotone Across LLM Agent Tiers\n- **Authors**: Yong-eun Cho\n- **URL**: [https://arxiv.org/abs/2605.26731](https://arxiv.org/abs/2605.26731)\n- **Abstract**:\n> A prevalent assumption in LLM agent deployment holds that more structured harnesses universally improve reliability, and that higher-capability models need proportionally less structural guidance -- together implying a monotone inverse relationship between model capability tier and optimal harness complexity. We test this hypothesis through a controlled 432-run experiment crossing six models across four capability tiers with three harness conditions (light, balanced, strict) on HEAT-24, a 24-task synthetic benchmark with git-based workspace verification. Our results refute the monotone inverse relationship on two fronts. First, for the frontier chat model evaluated (Gemini 2.5 Flash), increased harness verbosity lowers VTSR by 29-38 percentage points -- a harness-complexity paradox. Second, for the frontier reasoning model evaluated (Qwen3.5-122B, extended thinking enabled), strict harness achieves the highest VTSR (91.7%) and the lowest latency, the opposite of the prediction. Within the constrained tier, a 2B model (Gemma4:e2B) matches strong-open-tier stability at 91.7% across all harnesses. Because each tier is represented by a single model in this study, these results should be interpreted as model-specific observations; harness sensitivity appears non-monotone across the models evaluated, and depends critically on model type (chat vs. reasoning). We introduce a six-label failure taxonomy showing that format_violation dominates capable-model failures while wrong_file dominates low-capability failures, and we derive practical tier-aware harness selection guidelines.\n\n---\n\n## 23. Towards Feedback-to-Plan Decisions for Self-Evolving LLM Agents in CUDA Kernel Generation\n- **Authors**: Yee Hin Chong , Jiaming Wu , Youhui Zhang , Peng Qu\n- **URL**: [https://arxiv.org/abs/2605.26720](https://arxiv.org/abs/2605.26720)\n- **Abstract**:\n> Large language models (LLMs) have shown strong empirical gains as self-evolving agents for CUDA kernel generation, driven by feedback-conditioned planning across generations. However, how planning decisions attribute and combine heterogeneous feedback signals remains opaque. Standard end-to-end ablations fail to resolve this question, as iterative planning amplifies early perturbations and conflates feedback effects with trajectory-dependent drift. We introduce \\texttt{CUDAnalyst}, a unified analysis layer for controlled, generation-level attribution of planning decisions to feedback components via trajectory freezing and selective feedback injection. \\texttt{CUDAnalyst} enables stable generation-level evaluation and principled coalitional-style attribution of feedback effects and interactions. Our results show that explicit planning is beneficial only when feedback is aligned, that effective planning emerges from structured multi-feedback interactions, and that high-level plans from stronger reasoning models can partially transfer to weaker ones. These trends hold across reference backbones, representative workloads, and reference induction regimes, indicating that the identified feedback-to-plan structure is robust within the controlled axes studied.\n\n---\n\n## 24. MemFail: Stress-Testing Failure Modes of LLM Memory Systems\n- **Authors**: Ishir Garg , Neel Kolhe , Dawn Song , Xuandong Zhao\n- **URL**: [https://arxiv.org/abs/2605.26667](https://arxiv.org/abs/2605.26667)\n- **Abstract**:\n> Large language model (LLM) agents increasingly rely on external memory systems to remain consistent across long-horizon interactions, but little empirical work has been done to understand the specific failure modes and design choices that these systems present. Existing benchmarks report aggregate question-answering accuracy and treat memory systems as black boxes, making it impossible to attribute an incorrect answer to a particular failure mode of the system. We introduce MemFail, a diagnostic benchmark that isolates the failure modes of modern LLM memory systems. We begin by formalizing memory systems as the composition of three canonical operations -- summarization, storage, and retrieval -- and identify the potential failure modes induced by each. Based on these hypothesized failure modes, we construct five datasets spanning four tasks, each adversarially designed to test a specific operation of a memory system. Using these datasets, we evaluate four state-of-the-art memory systems on MemFail and demonstrate how MemFail can be used to empirically understand the tradeoffs induced by differences in memory system architectures.\n\n---\n\n## 25. UnityMAS-O: A General RL Optimization Framework for LLM-Based Multi-Agent Systems\n- **Authors**: Yiqun Chen , Wei Yang , Erhan Zhang , Shijie Wang , Qi Liu , Zechun Niu , Bin Zhang , Haitao Li , Rui Li , Lingyong Yan , Jinyuan Feng , Biqing Qi , Xiaochi Wei , Yan Gao , Yi Wu , Yao Hu , Jiaxin Mao\n- **URL**: [https://arxiv.org/abs/2605.26646](https://arxiv.org/abs/2605.26646)\n- **Abstract**:\n> LLM-based multi-agent systems decompose complex tasks into interacting roles, but most remain manually orchestrated by prompts, tools, and control rules, while agents are rarely optimized through a unified reinforcement learning interface. Existing RL post-training frameworks mainly target single-policy optimization and lack abstractions for user-defined multi-agent workflows, structured interaction, role-specific credit assignment, and configurable parameter sharing. We present UnityMAS-O, a general RL optimization framework for LLM-based multi-agent systems. UnityMAS-O treats the complete workflow as the optimization unit, rather than a single response or policy trajectory. It represents workflows through four first-class objects: logical agent roles, graph trajectories, user-defined rewards, and agent--model mappings. This decouples logical agents from physical model parameters, supporting full sharing, full separation, and partial sharing, with rewards assigned at role, turn, and trajectory levels. UnityMAS-O extends verl with a Ray-based star-topology runtime. A central controller executes workflows, invokes tools, records structured trajectories, and assembles rewards; model-local worker groups handle rollout, buffering, advantage computation, and distributed PPO-style updates. Users can define agents, workflows, model mappings, and rewards without rewriting the optimization infrastructure. We instantiate UnityMAS-O on retrieval-augmented QA, iterative agentic search, and reflective code generation. Across Natural Questions, HotpotQA, and held-out code tasks, multi-agent RL improves manually specified workflows after optimization, with especially large gains for smaller models and strict code all-passed metrics. These results show that UnityMAS-O can serve as a reusable substrate for converting diverse LLM-based multi-agent workflows into trainable multi-agent RL systems.\n\n---\n\n## 26. FAST-GOAL: Fast and Efficient Global-local Object Alignment Learning\n- **Authors**: Hyungyu Choi , Young Kyun Jang , Chanho Eom\n- **URL**: [https://arxiv.org/abs/2605.26615](https://arxiv.org/abs/2605.26615)\n- **Abstract**:\n> Vision-language models such as CLIP have shown impressive capabilities in aligning images and text, but they often struggle with lengthy and detailed text descriptions due to pre-training on short and concise captions. We present FAST-GOAL (Fast and Efficient Global-local Object Alignment Learning), an efficient fine-tuning method that enhances ability of CLIP to handle lengthy text through global-local semantic alignment. Our method consists of two key components. First, Fast Local Image-Sentence Matching (FLISM) efficiently extracts local image regions through object detection and spatial division, then matches them with corresponding sentences. Second, Token Similarity-based Learning (TSL) maximizes the similarity between patch tokens from specific regions in the image and their corresponding region embeddings, applying the same principle to text, which enhances the ability of the model to capture detailed correspondences. Additionally, we introduce GLIT100k, a dataset that provides both global image-lengthy caption pairs and context-derived local pairs, where local descriptions are extracted from global captions to maintain semantic coherence. Through extensive experiments on long caption datasets (DOCCI, DCI) and short caption datasets (MSCOCO, Flickr30k), we demonstrate that FAST-GOAL achieves significant improvements over baselines, enabling effective adaptation of CLIP to detailed textual descriptions while maintaining computational efficiency.\n\n---\n\n## 27. AGORA: Adapter-Grounded Observation-Action Retention for Inference-Free Prompt Compression in LLM Agents\n- **Authors**: Haoran Zhang , Zhaohua Sun\n- **URL**: [https://arxiv.org/abs/2605.26596](https://arxiv.org/abs/2605.26596)\n- **Abstract**:\n> The token-level extractive compressors widely used for general LM context are structurally inappropriate for LLM agents: across 17 (env, backbone, method) cells spanning two independent token-level method families, every cell collapses to mean reward <= 0.05 despite 1.3-13.3x realized compression. We name and characterize this failure mode as action-grammar destruction -- the tokens carrying action semantics (identifiers, brackets, action verbs) are exactly those self-information ranks lowest, so a general-purpose compressor reliably removes them and the environment rejects the residual. The diagnosis points to step-granularity compression. We introduce AGORA, an inference-free step-level compressor combining a structural prompt parser, an always-keep floor for format- and recency-critical content, and a 125M-parameter relevance scorer trained on counterfactual next-action-change labels (~2ms/step, zero per-step LLM toll). Across the compared inference-free and LLM-based methods, AGORA is the only one retaining >= 75% uncompressed performance in 8 of 9 cells (with the lone exception at 73%); a four-way component ablation isolates the structural floor as the dominant quality lever and the learned scorer as the source of 1.0-11.5x adaptive end-to-end compression from a single fixed keep ratio.\n\n---\n\n## 28. MedGuideX: Internalizing Decision Logic from Executable Guidelines into Large Language Models for Clinical Reasoning\n- **Authors**: Yuhao Shen , Lang Cao , Simo Du , Yuqing Wang , Juexiao Zhou , Hao Peng , Yue Guo\n- **URL**: [https://arxiv.org/abs/2605.26567](https://arxiv.org/abs/2605.26567)\n- **Abstract**:\n> Clinical practice guidelines (CPGs) encode evidence-based decision logic that clinicians apply by evaluating patient variables, conditional criteria, and recommendation rules. However, existing methods often use CPGs as free-text training data or retrieval sources, underutilizing their procedural decision structure. To better exploit this structure, we introduce a guideline-derived training pipeline that transforms CPG recommendations into executable clinical decision logic and uses it to generate factual and counterfactual question-answering data. Theses data teach models both guideline-supported decisions and how decisions change under different patient conditions. Post-training a medical LLM on the generated data yields MedGuideX. Across four clinical reasoning benchmarks, MedGuideX achieves a 10.28% relative improvement in average accuracy. Physician evaluation further shows that MedGuideX better recovers clinician authored reasoning steps and produces physician-preferred rationales in faithfulness, validity, completeness, and clarity. Overall, our results show that executable decision logic from CPGs can be transformed into scalable supervision for building reliable medical LLMs.\n\n---\n\n## 29. MobileExplorer: Accelerating On-Device Inference for Mobile GUI Agents via Online Exploration\n- **Authors**: Runxi Huang , Liyu Zhang , Shengzhong Liu , Xiaomin Ouyang\n- **URL**: [https://arxiv.org/abs/2605.26546](https://arxiv.org/abs/2605.26546)\n- **Abstract**:\n> Mobile graphical user interface (GUI) agents enable AI models to autonomously operate smartphones on behalf of users. However, most existing systems focus primarily on optimizing task accuracy and rely on cloud-hosted models for inference, which introduces privacy concerns and network-dependent latency. As a result, fully on-device deployment of mobile GUI agents remains underexplored. We propose MobileExplorer, a new framework that accelerates on-device inference for vision-based mobile GUI agents via online exploration. The key idea is to exploit the long per-step reasoning time of vision-language models (VLMs) by performing lightweight, parallel exploration of UI elements. During model inference, the agent proactively probes semantically relevant UI elements and records these exploration traces as structured memory. To ensure reliable execution in live mobile environments, we design a two-level rollback mechanism that robustly restores the initial UI state when a fast but naive backtracking strategy fails. The collected exploration traces are then summarized into concise contextual hints and injected into the prompt to enhance the subsequent reasoning step. We evaluate MobileExplorer on multiple off-the-shelf devices using the AndroidWorld benchmark, as well as newly designed, more complex tasks and dynamic on-device environments. MobileExplorer reduces the average number of reasoning steps and end-to-end latency by 23\\%, while maintaining or improving task success rates by up to 5\\%. A video demonstration of MobileExplorer performance in the real world is available at this https URL .\n\n---\n\n## 30. The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence\n- **Authors**: MiniMax : Aili Chen , Aonian Li , Baichuan Zhou , Bangwei Gong , Binyang Jiang , Boji Dan , Changqing Yu , Chao Wang , Cheng Ma , Cheng Zhong , Cheng Zhu , Chengjun Xiao , Chengyi Yang , Chengyu Du , Chenyang Zhang , Chi Zhang , Chuangyi Huang , Chunhao Zhang , Chunhui Du , Chunyu Zhao , Congchao Guo , Da Chen , Deming Ding , Dianjun Sun , Dongyu Zhang , Enhui Yang , Fei Yu , Guang Zheng , Guodong Zheng , Guohong Li , Haichao Zhu , Haigang Zhou , Haimo Zhang , Han Ding , Hao Zhang , Haohai Sun , Haolin Lyu , Haonan Lu , Haoyu Wang , Huajie Shi , Huiyang Li , Jiacheng Chen , Jian Zhang , Jiaqi Zhuang , Jiaren Cai , Jiaxin Pan , Jiayao Li , Jiayuan Song , Jichuan Zhang , Jie Wang , Jihao Gu , Jin Zhu , Jingwei Dong , Jingyang Li , Jingyu Zhang , Jingze Zhuang , Jinhao Tian , Jinli Liu , Jinyi Hu , Jun Tao , Jun Zhang , Junbin Ruan , Junhao Xu , Junjie Yan , Junteng Liu , Junxian He , Kang Xu , Ke Ji , Ke Yang , Kecheng Xiao , Keyu Duan , Keyu Li , Le Han , Letian Ruan , Li Yuan , Lianfei Yu , Liheng Feng , Lijie Mo , Lin Li , Lingye Bao , Lingyu Yang , Lingyuan Zhou , Loki , Lu Chen , Lunbin Ceng , Ming Li , Ming Zhong , Mingliang Tao , Mingyuan Chi , Mujie Lin , Nan Hu , Ningxin Chen , Peiyin Zhu , Peng Gao , Pengcheng Gao , Pengfei Li , Penglin Li , Pengyu Zhao , Qibin Ren\n- **URL**: [https://arxiv.org/abs/2605.26494](https://arxiv.org/abs/2605.26494)\n- **Abstract**:\n> We introduce the MiniMax-M2 series, a family of Mixture-of-Experts language models built around the principle that mini activations can unleash maximum real-world intelligence. The flagship M2 contains 229.9B total parameters with only 9.8B activated per token. Designed end-to-end for agentic deployment, the M2 series rests on three components: (i) agent-driven data pipelines producing large-scale, verifiable trajectories across agentic coding and agentic cowork, each grounded in an executable workspace and an artifact-aligned reward; (ii) Forge, a scalable agent-native RL system that adapts to long-horizon agent trajectories, paired with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and a clean training-inference-agent decoupling that supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step toward self-evolution -- autonomously debugging training runs and modifying its own scaffold. Across M2 through M2.7, this combination translates a mini-activation footprint into frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.\n\n---\n\n## 31. Reasoning, Code, or Both? How Large Language Models Handle Variations in Math Questions\n- **Authors**: Matthew Kutakh\n- **URL**: [https://arxiv.org/abs/2605.26414](https://arxiv.org/abs/2605.26414)\n- **Abstract**:\n> Large Language Models (LLMs) achieve impressive accuracy on mathematical reasoning benchmarks, yet their performance drops when problems are modified with simple changes like different names or numbers. Code execution methods, which let models generate and run Python code instead of reasoning in natural language, have been proposed as a solution, but their effect on reasoning robustness (the ability to maintain accuracy across problem variations) has not been systematically tested. This study evaluates three approaches on 1,000 problems from the GSM-Symbolic dataset: pure reasoning using chain-of-thought (CoT) prompting, single-shot code execution using Program-Aided Language models (PAL), and iterative code execution using Step-by-Step Coding (SBSC). All three were run on paired original and modified problems using Claude Haiku 4.5. CoT was the most robust method, with an accuracy drop of 1.3 percentage points and 1.8% of problems breaking under perturbation. PAL was the least robust at 1.7 percentage points and 3.1% broke, with SBSC falling in between. Although these differences were not statistically significant ($p = .096$), the directional trend was consistent across all measures, suggesting that code execution, whether single-shot or iterative, does not improve reasoning robustness on grade-school-level problem variations.\n\n---\n\n## 32. From Static Context to Calibrated Interactive RL: Mitigating Distribution Shift in Multi-turn Dialogue with Aligned Simulator\n- **Authors**: Xiaohua Wang , Jiakang Yuan , Zisu Huang , Muzhao Tian , Changze Lv , Kaitao Song , Tao Chen , Xiaoqing Zheng\n- **URL**: [https://arxiv.org/abs/2605.26403](https://arxiv.org/abs/2605.26403)\n- **Abstract**:\n> A long-standing goal of the research community is to develop highly interactive LLM-based dialogue agents. Recent research focuses on optimizing policies based on fixed offline logs (Static Context RL) or using a prompt-based simulator (Interactive RL). In this work, we theoretically show that both paradigms are fundamentally limited by context distribution shift--a mismatch between dialogue histories observed during training and those encountered in real conversations. This shift compounds quadratically over turns and severely degrades dialogue quality. Specifically, we attribute this shift to two distinct sources: (i) policy-induced shift, arising from training on static histories rather than self-generated trajectories; and (ii) simulator-induced shift, stemming from discrepancies between simulated and real human behaviors. To address these challenges, we propose Calibrated Interactive RL, a unified framework that couples interactive RL with simulator alignment. By aligning the simulator with human interaction patterns, our approach reduces the sim-to-real gap and mitigates compounding distribution shifts. Experiments across multiple dialogue tasks confirm our theoretical analysis: (i) Interactive RL significantly outperforms the Static Context baseline by mitigating policy distribution shift; and (ii) calibrating simulators with our alignment method further bridges the sim-to-real gap, yielding state-of-the-art downstream performance.\n\n---\n\n## 33. Automatic Layer Selection for Hallucination Detection\n- **Authors**: Xinpeng Wang , William Cao , Andrew Gordon Wilson , Zhe Zeng\n- **URL**: [https://arxiv.org/abs/2605.26366](https://arxiv.org/abs/2605.26366)\n- **Abstract**:\n> Recent studies on hallucination detection have shown that hallucination-related signals are more strongly encoded in intermediate layers than in the final layer of large language models (LLMs). Although a growing body of work has sought to exploit this property for hallucination detection, how to automate the selection of high-performing layers remains underexplored, and principled methods for this purpose are still lacking. To address this gap, we first propose several hypotheses for why such signals emerge in intermediate layers and evaluate corresponding criteria for automatic layer selection across diverse LLM architectures, scales, and tasks, covering both question answering and summarization hallucination detection benchmarks. However, we find that none of these criteria consistently delivers satisfactory performance. We therefore propose a new selection criterion, First Effective Peak of Intrinsic Dimension (FEPoID), which consistently identify optimal or near-optimal layers and outperforms both the aforementioned criteria and existing hallucination detection baselines. FEPoID is training-free and incurs negligible computational overhead. In addition, we study the generation behaviors of LLMs and introduce a simple yet effective truncation strategy, which further amplifies hallucination-related signals and substantially improves overall detection performance. Code is publicly available at this https URL\n\n---\n\n## 34. Managing Uncertainty in LLM-Generated Procedural Knowledge for Virtual Laboratory Planning\n- **Authors**: Polychronis Karpodinis , Dimitris Kalles\n- **URL**: [https://arxiv.org/abs/2605.26333](https://arxiv.org/abs/2605.26333)\n- **Abstract**:\n> Educational virtual laboratories can make experimental training more scala-ble, adaptive, and accessible, especially when students have limited access to physical laboratory facilities. However, authoring new simulated laboratory procedures remains costly: educators must describe new equipment, define how instruments and materials interact, and specify valid procedural flows that can be executed or assessed inside the virtual environment. Large lan-guage models can assist in this authoring process by generating detailed ex-perimental procedures, but their output should not be treated as directly exe-cutable plans. They may omit necessary actions, arrange steps in the wrong order, or produce instructions that are logically incorrect or incompatible with the laboratory equipment. This paper presents a prototype framework for managing uncertainty in LLM-generated procedural knowledge for virtu-al laboratory planning. The framework aims to reduce procedural uncertainty by using structured domain representations and uncertain LLM-generated state-transition samples to extract candidate procedural rules, transform them into explicit and inspectable constraints, and use them to repair uncertain procedural steps. Although the motivating domain refers to educational vir-tual laboratories, the underlying problem is more general: managing uncer-tain procedural knowledge for action planning in structured interactive envi-ronments. We illustrate the approach in a virtual laboratory domain involving laboratory instruments, containers, tools, and material-transfer actions.\n\n---\n\n## 35. OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling\n- **Authors**: Adam Bawatneh , Sagar Sapkota , Amrit Singh Bedi , Santu Karmaker , Mubarak Shah\n- **URL**: [https://arxiv.org/abs/2605.26322](https://arxiv.org/abs/2605.26322)\n- **Abstract**:\n> Theory of Mind (ToM), the ability to infer others' knowledge, intentions, and emotions, is commonly evaluated in large language models (LLMs) using end-point question answering, where performance is judged solely by the final answer to a social reasoning query. This paradigm obscures whether the model actually constructs the underlying mental-state representations required for robust reasoning, particularly in scenarios involving divergent, evolving, or mistaken beliefs. In order to address this research gap, we introduce OmniToM, a benchmark that directly evaluates these representations by requiring explicit modeling of belief structures for all relevant actors within a narrative. These structures are composed of belief propositions: minimal statements of what an actor takes to be true about the world or another actor's mental state, allowing knowledge, intentions, emotions, and false beliefs to be analyzed in a common format. Models are evaluated in two stages: Stage 1: Belief Extraction, which extracts from the story the beliefs relevant to its social dynamics, and Stage 2: Belief Labeling, which assigns each belief a seven-dimensional schema label covering recursive order, truth status, knowledge access, explicitness, content type, mental source, and context. Built from 895 stories from the existing ToMBench story corpus and augmented with 22,343 labeled belief propositions, OmniToM uses a human-calibrated LLM-assisted annotation pipeline. Across diverse models in zero-shot evaluation, OmniToM reveals an actor-specific belief-tracking bottleneck: current LLMs struggle with the knowledge-access and representational decisions required to transform narrative facts into actors' beliefs and shared mental states.\n\n---\n\n## 36. Experiments in Agentic AI for Science\n- **Authors**: Judy Fox , Geoffrey Fox\n- **URL**: [https://arxiv.org/abs/2605.26305](https://arxiv.org/abs/2605.26305)\n- **Abstract**:\n> This paper details two novel frameworks for developing autonomous, agentic AI in scientific workflows. Both systems leverage a hybrid Local Body, Remote Brain architecture via Google Colab, utilizing Python-based local orchestrators to invoke large language model (LLM) cloud backends. The first agent, DeepTS/DeepCollector, automates the large-scale curation, extraction, and deduplication of time-series datasets. The second, DeepScribe, is an autonomous presentation analyzer that converts visually dense, mathematically complex physics lectures into structured scientific reports. Through practical systems engineering-such as granular attribute extraction (Cellular RAG), remote data inspection, and distributed concurrency controls-we demonstrate how agentic AI can overcome the context and reasoning limitations of current state-of-the-art systems to rigorously support scientific workflows. Finally, we outline a generalization of DeepTS to support deep knowledge graphs and discuss the application of this conceptual approach to high-energy physics (DeepQCD).\n\n---\n\n## 37. Personalizing Embodied Multimodal Large Language Model Agents over Long-term User Interactions\n- **Authors**: Jeongeun Lee , Chanyoung Park , Dongha Lee\n- **URL**: [https://arxiv.org/abs/2605.26256](https://arxiv.org/abs/2605.26256)\n- **Abstract**:\n> Multimodal large language model (MLLM)-based embodied agents have shown strong potential for solving complex tasks in physical environments. However, personalized assistance requires more than following generic instruction or recognizing object categories. In real-world scenarios, the intended target is often specified only implicitly through prior interactions, requiring agents to leverage personalized context accumulated over time. In this work, we propose POLAR, a multiomodal memory-augmented framework for personalized embodied agents over long-term user interactions. POLAR organizes prior interactions into a multimodal knowledge graph that captures semantic memory for personalized context and visual concepts, and episodic memory for embodied experiences such as agent trajectories. To execute embodied tasks, POLAR retrieves relevant memories to interpret the current request and guide task execution. We evaluate POLAR across multiple MLLM backbones and diverse evaluation scenarios to study the role of memory in long-term personalization. Results show that the proposed memory mechanism consistently improves performance by enabling more effective use of information accumulated over prior interactions. The gains are especially pronounced when the agents are required to reason across multiple interactions, perform multi-hop inference, or tracking updates in user-specific context over time.\n\n---\n\n## 38. Can LLMs Introspect? A Reality Check\n- **Authors**: Shashwat Singh , Tal Linzen , Shauli Ravfogel\n- **URL**: [https://arxiv.org/abs/2605.26242](https://arxiv.org/abs/2605.26242)\n- **Abstract**:\n> Can large language models detect and report their own internal states? A number of studies have argued that the answer to this question is yes. We argue, based on lessons from human metacognition research, that this conclusion may be premature: to be convinced of this conclusion we need to distinguish genuine introspection from pattern matching based on surface-level cues. Furthermore, we argue that behavioral evidence alone is inherently insufficient to establish strong introspective claims. We re-examine two recently introduced evaluation paradigms in light of this consideration. In the first paradigm, models are expected to detect whether their internal states have been tampered with. We find that models cannot reliably distinguish such interventions on their internal states from manipulations of the input, suggesting that their success in the original studies reflects their ability to detect anomalies more generally, as opposed to interventions on their internal states in particular. In the second paradigm we examine, models are tasked with predicting labels derived from their own hidden states. Here, we find that classifiers that only have access to the input achieve equivalent performance to the model's own in-context predictions, indicating that the original results do not conclusively demonstrate that the model has privileged access to its internal representations. We further introduce a relabeled control setting, where models cannot rely on the semantics of the task to solve it, and instead must rely on the internal representation; models perform closer to chance on this better-controlled version of the task. Taken together, these results indicate that current evidence is insufficient to establish that LLMs display metacognitive monitoring.\n\n---\n\n## 39. LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding\n- **Authors**: Shihao Wang , Shilong Liu , Yuanguo Kuang , Xinyu Wei , Yangzhou Liu , Zhiqi Li , Yunze Man , Guo Chen , Andrew Tao , Guilin Liu , Jan Kautz , Lei Zhang , Zhiding Yu\n- **URL**: [https://arxiv.org/abs/2605.27365](https://arxiv.org/abs/2605.27365)\n- **Abstract**:\n> Vision-language models (VLMs) commonly formulate visual grounding and detection as a coordinate-token generation problem, serializing each 2D box into multiple 1D tokens that are learned and decoded largely independently. This token-by-token decoding mismatches the coupled structure of box geometry and creates a practical inference bottleneck due to strictly sequential generation. We introduce LocateAnything, a unified generative grounding and detection framework based on Parallel Box Decoding (PBD). By decoding geometric elements such as bounding boxes and points as atomic units in a single step, LocateAnything preserves intra-box geometric coherence and unlocks substantial parallelism. We show that PBD improves both decoding throughput and localization accuracy. We further develop a scalable data engine and curate LocateAnything-Data, a large-scale dataset with more than 138 million training samples, substantially increasing data diversity for high-precision localization. Extensive evaluations show that LocateAnything advances the speed-accuracy frontier, achieving significantly higher decoding throughput while improving high-IoU localization quality across diverse benchmarks. The results highlight the complementary benefits of Parallel Box Decoding and large-scale training data in enabling efficient and precise unified visual grounding and detection.\n\n---\n\n## 40. GENESIS: Harnessing AI Agents for Autonomous 6G RAN Synthesis, Research, and Testing\n- **Authors**: Tamerlan Aghayev , Maxime Elkael , Michele Polese , Minh Dat Nguyen , Gabriele Gemmi , Andrea Lacava , Ali Saeizadeh , Reshma Prasad , Paolo Testolina , Angelo Feraudo , Soumendra Nanda , Pedram Johari , Salvatore D'Oro , Tommaso Melodia\n- **URL**: [https://arxiv.org/abs/2605.27360](https://arxiv.org/abs/2605.27360)\n- **Abstract**:\n> Cellular research and development (R&D) is throttled by six structural processes that each consume months of manual engineering work per iteration: (i) synthesizing new features from standards or research papers into production code; (ii) conformance and interoperability testing; (iii) hardening against field anomalies and diverse deployment environments; (iv) data-driven optimization of network functionalities; (v) discovering and prototyping novel waveforms, functionalities, and capabilities for future standards; and (vi) securing the stack against vulnerabilities. Although Large Language Models (LLMs) have compressed comparable R&D work in general software engineering from days to minutes, their known pitfalls worsen on Radio Access Network (RAN) use cases: they hallucinate Application Programming Interfaces (APIs) and mis-read specifications, which kills interoperability of RAN components at the first mistake, and they heavily rely on simulations for designing algorithms, which is notorious for breaking when transferred to real hardware. To address these challenges, we present GENESIS, an agentic Artificial Intelligence (AI) framework that converts intents (e.g., a specification clause, a telemetry anomaly, or a research hypothesis) into solutions validated with over-the-air experiments, fed back into a persistent knowledge base. GENESIS is built on three composable primitives (agents, skills, hooks) and a knowledge layer (SYNAPSE) that doubles as the source of ground truth and the recipient of every artifact the framework produces, making capabilities compound across runs.\n\n---\n\n## 41. MobileMoE: Scaling On-Device Mixture of Experts\n- **Authors**: Yanbei Chen , Hanxian Huang , Ernie Chang , Jacob Szwejbka , Digant Desai , Zechun Liu , Vikas Chandra , Raghuraman Krishnamoorthi\n- **URL**: [https://arxiv.org/abs/2605.27358](https://arxiv.org/abs/2605.27358)\n- **Abstract**:\n> Mixture-of-Experts (MoE) has become the de facto architecture for hundred-billion-parameter language models, yet its advantages at sub-billion scales for on-device deployment remain largely unexplored. To close this gap, we present MobileMoE, a family of on-device MoE language models with sub-billion active parameters (0.3-0.9B active and 1.3-5.3B total) that establish a new Pareto frontier for on-device LLMs. We first formulate an on-device MoE scaling law that jointly optimizes MoE architecture under mobile memory and compute constraints, identifying an on-device sweet spot - moderate sparsity with fine-grained and shared experts - that is simultaneously memory and compute-optimal. Building on the derived architectures, we train MobileMoE with a four-stage recipe covering pre-training, mid-training, instruction fine-tuning, and quantization-aware training, all on open-source datasets. Across 14 benchmarks, MobileMoE matches or exceeds leading on-device dense LLMs with 2-4$\\times$ fewer inference FLOPs, and matches or surpasses the state-of-the-art MoE OLMoE-1B-7B with up to 60% fewer parameters. To bridge the last mile to mobile deployment, we provide the first efficient MoE inference on commodity smartphones with comprehensive on-device profiling. At comparable INT4 weight memory, MobileMoE-S delivers $1.8$-$3.8\\times$ faster prefill and $2.2$-$3.4\\times$ faster decode than the dense baseline MobileLLM-Pro.\n\n---\n\n## 42. Guiding LLM Post-training Data Engineering with Model Internals from Sparse Autoencoders\n- **Authors**: Yi Jing , Zao Dai , Jinwu Hu , Zijun Yao , Lei Hou , Juanzi Li , Xiaozhi Wang\n- **URL**: [https://arxiv.org/abs/2605.27354](https://arxiv.org/abs/2605.27354)\n- **Abstract**:\n> Model internals encode rich information about how a large language model (LLM) processes its training data; however, post-training data engineering largely relies on external signals and ignores rich intrinsic signals lying in model internals. We propose SAERL, a data engineering framework for LLM reinforcement learning (RL). It models three intrinsic data properties: diversity, difficulty, and quality, using model internals extracted with Sparse Autoencoder (SAE), an advanced mechanistic interpretability tool. Each property grounds a concrete data engineering operation: SAE-space clustering with moderate batch mixing for batch diversity control, a difficulty proxy for easy-to-hard curriculum ordering, and a quality probe for data filtering. SAERL improves average accuracy by 3.00% over vanilla GRPO and reaches target accuracy with 20% fewer training steps on Qwen2.5-Math-1.5B, with consistent gains across model scales and RL algorithms. Experiments show that SAE transfers effectively across model families and scales, serving as a lightweight and reusable data engineering tool. These results demonstrate that model internals are a powerful and practical source of signals for post-training data engineering.\n\n---\n\n## 43. EdgeFlow: Edge-Map Augmented VLM-Based Flowchart Processing for Industrial Requirements Engineering\n- **Authors**: Zhifei Dou , Shabnam Hassani , Ou Wei\n- **URL**: [https://arxiv.org/abs/2605.27332](https://arxiv.org/abs/2605.27332)\n- **Abstract**:\n> Flowcharts are widely used in industrial requirements, but usually remain embedded as static images. Vision Language Models (VLMs) show promise in the conversion of these flowcharts into machine-readable models for RE activities, yet, when directly applied to flowchart conversion, they often fail on topology-critical visual details. To address this, we propose EdgeFlow that augments a VLM's original input with a deterministically extracted Canny edge map-acting as a structural prior-to improve flowchart-to-Mermaid conversion, without requiring annotated training data or domain-specific model fine-tuning. We evaluate EdgeFlow on IndusReqFlow, a dataset sourced from real-world requirements. Compared with off-the-shelf VLMs, EdgeFlow improves node-level F1 by 17.39 percentage points and edge-level F1 by 16.94 percentage points. At the path level, EdgeFlow improves path F1 by 11.06 percentage points, enabling better support for model-based testing. These results demonstrate that EdgeFlow provides a practical, training-free means to improve topology-preserving flowchart-to-Mermaid conversion for industrial RE. Cross-dataset evaluation results on a public synthetic benchmark show no significant improvement; this highlights the need for diverse benchmarks incorporating industrial data for the comprehensive evaluation of future VLM-based RE tools.\n\n---\n\n## 44. It's Not Always Sycophancy: Measuring LLM Conformity as a Function of Epistemic Uncertainty\n- **Authors**: Kevin H. Guo , Chao Yan , Avinash Baidya , Katherine Brown , Xiang Gao , Juming Xiong , Zhijun Yin , Bradley A. Malin\n- **URL**: [https://arxiv.org/abs/2605.27288](https://arxiv.org/abs/2605.27288)\n- **Abstract**:\n> Large language models (LLMs) are known to abandon their initial stance to conform to user pushback. While prior research largely attributes this behavior to sycophancy learned during reinforcement learning from human feedback, we hypothesize that conformity is also driven by a model's epistemic uncertainty at inference time. In this paper, we introduce MUSE, a two-stage evaluation framework to disentangle the mechanisms driving LLM conformity. Specifically, MUSE maps a model's epistemic uncertainty in responding to a query against its likelihood to yield to user pushback in a subsequent turn. We demonstrate that the mechanisms driving conformity extend beyond sycophancy alone. Specifically, we characterize two distinct factors that jointly drive conformity: sycophantic conformity, where a model aligns with user pushback even with absolute certainty in its initial response, and uncertainty-driven conformity, where a model's likelihood for conformity increases alongside its uncertainty. Furthermore, we conduct ablation studies to demonstrate that both sycophantic conformity and uncertainty-driven conformity grow with 1) the LLM's perceived expertise of the user and 2) the plausibility of the user's suggestions. More broadly, MUSE informs more targeted intervention strategies by distinguishing alignment-induced sycophancy and training-corpora-driven uncertainty.\n\n---\n\n## 45. Lost in Sampling: Assessing Lexical Reachability in LLMs via the Word Coverage Score (WCS)\n- **Authors**: Samer Awad , Javier Conde , Carlos Arriaga , Tairan Fu , Javier Coronado-Blázquez , Pedro Reviriego\n- **URL**: [https://arxiv.org/abs/2605.27268](https://arxiv.org/abs/2605.27268)\n- **Abstract**:\n> Modern Large Language Models (LLMs) are often criticized for producing repetitive and homogeneous text, despite possessing vast latent vocabularies. While previous research has focused on model knowledge and training data, we investigate the role of decoding mechanics in suppressing linguistic diversity. We introduce the Word Coverage Score (WCS), a metric that quantifies the extent to which contextually appropriate human vocabulary is mathematically pruned by standard sampling filters (e.g., Top-$p$, Top-$k$, and Min-$p$). Rather than assessing static knowledge, the WCS measures the lexical survival rate of low-frequency, high-information human words as a function of sampling parameters. By auditing open-weight models on human-authored corpus fragments, we identify which logical lexical choices are rendered unreachable by the decoder, even when they reside within the probability space. Our results provide quantitative evidence that industry-standard sampling defaults act as unintended censorship mechanisms, smoothing the unique textures of human expression into a homogenized discourse. The WCS offers a rigorous framework for optimizing the trade-off between text coherence and lexical richness, providing a diagnostic tool for preserving the diversity of human language in generative models.\n\n---\n\n## 46. Pair-In, Pair-Out: Latent Multi-Token Prediction for Efficient LLMs\n- **Authors**: Wenhui Tan , Minghao Li , Xiaoqian Ma , Siqi Fan , Xiusheng Huang , Liujie Zhang , Ruihua Song , Weihang Chen\n- **URL**: [https://arxiv.org/abs/2605.27255](https://arxiv.org/abs/2605.27255)\n- **Abstract**:\n> Long chain-of-thought reasoning has made autoregressive decoding the dominant inference cost of modern large language models. Existing methods target either the input side (latent compression) or the output side (speculative decoding and multi-token prediction, MTP), but the two lines of work have been pursued independently. Moreover, output-side methods must incur an expensive verifier pass to validate the unreliable draft tokens predicted by MTP. To address these issues, we propose \\textbf{Pair-In, Pair-Out (PIPO)}, which unifies both sides by viewing a latent compressor and an MTP head as mirror-image operations: the compressor folds two input tokens into one latent representation, while the MTP head unfolds one hidden state into one additional output token. To remove the verifier cost without sacrificing reliability, PIPO trains a lightweight confidence head that decides whether draft tokens should be accepted. We observe that On-Policy Distillation (OPD) naturally matches the rejection-sampling criterion of speculative decoding, so the confidence head can be trained alongside OPD with negligible extra cost. Experiments on AIME 2025, GPQA-Diamond, LiveCodeBench v6, and LongBench v2 with Qwen3.5-4B and 9B backbones show that PIPO improves pass@4 over regular decoding by up to $+7.15$ points, while delivering up to $2.64\\times$ first-token-latency and $2.07\\times$ per-token-latency speedups.\n\n---\n\n## 47. Qiskit QuantumKatas: Adapting Microsoft's Quantum Computing exercises for LLM evaluation\n- **Authors**: Juan Cruz-Benito , Ismael Faro\n- **URL**: [https://arxiv.org/abs/2605.27210](https://arxiv.org/abs/2605.27210)\n- **Abstract**:\n> We adapt Microsoft's QuantumKatas -- a well-established quantum computing curriculum -- from Q# to Qiskit, the most widely-adopted quantum computing framework, and package it with an evaluation framework for systematic LLM assessment. The resulting benchmark comprises 350 tasks across 26 categories, spanning fundamental gates through advanced algorithms (Grover's, Simon's, Deutsch-Jozsa), error correction, key distribution, and quantum games. Each task includes a natural language prompt, canonical solution, and deterministic test verification via classical circuit simulation. By building on the QuantumKatas' proven pedagogical design rather than creating tasks from scratch, we inherit a principled difficulty progression and comprehensive concept coverage while contributing the framework adaptation, evaluation infrastructure, and empirical analysis. We evaluate 16 LLMs across 7 prompting configurations -- a total of 39,200 model runs -- to demonstrate the benchmark's utility. Three key findings emerge: (1) the benchmark effectively differentiates model capabilities, with best-configuration pass rates ranging from 32.3% to 83.1% and a 26.1 pp average gap between frontier and open-source models; (2) models perform well at implementing known algorithms (SimonsAlgorithm 82.1%, BasicGates 81.6%) but struggle with problem encoding (SolveSATWithGrover 34.4%, DistinguishUnitaries 40.0%); and (3) chain-of-thought prompting shows a modestly bimodal effect -- it is the best strategy for three models (two of them explicitly reasoning-tuned per vendor documentation) but degrades performance for the rest, leaving it mid-pack in aggregate (56.3% mean) behind few-shot-5 (57.8%). We release the benchmark, evaluation framework, and baseline results to support research on LLM capabilities in quantum computing.\n\n---\n\n## 48. Generative Animations: A Multi-Model Pipeline for Prompt-Driven Motion Synthesis\n- **Authors**: Mannat Khurana , Sanyam Jain , Rishav Agarwal\n- **URL**: [https://arxiv.org/abs/2605.27203](https://arxiv.org/abs/2605.27203)\n- **Abstract**:\n> Animation elevates digital documents into immersive experiences, yet creating custom motion paths remains cumbersome, requiring designers to manually select presets, plot Bézier points, and configure timing properties. We introduce Generative Animations, a system that transforms natural language prompts into production-ready animations. By chaining Large Language Models (LLMs) for semantic parsing with the Segment Anything Model (SAM) for visual grounding, our pipeline automatically generates motion paths that respect scene geometry, handle depth-based occlusions, and honor 3D perspective transforms. We demonstrate the system through three use cases: contour-following trajectories, orbital animations with z-order awareness, and perspective-aligned motion on transformed objects.\n\n---\n\n## 49. Learning When to Think While Listening in Large Audio-Language Models\n- **Authors**: Zhiyuan Song , Weici Zhao , Yang Xiao , Suhao Yu , Cheng Zhu , Jiatao Gu\n- **URL**: [https://arxiv.org/abs/2605.27190](https://arxiv.org/abs/2605.27190)\n- **Abstract**:\n> Recent advances in Large Audio-Language Models (LALMs) have made real-time, streaming spoken interaction increasingly practical. In this setting, reasoning quality and responsiveness are tightly coupled: delaying reasoning until the speech endpoint can improve answer quality but moves deliberation into user-visible response delay, while answering too early risks committing before decisive evidence arrives. We introduce a learnable wait-think-answer control formulation for LALMs. Motivated by the incremental nature of human conversation, the controller decides under partial audio evidence when to wait, when to externalize a compact reasoning update, and when to answer. Using Qwen2.5-Omni-7B as the base model, we construct aligned wait-think-answer traces from spoken reasoning data, train the controller with supervised fine-tuning (SFT), and then apply Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO). The reward combines answer correctness, action validity, update timing, latency synchronization, reasoning quality, and chain consistency, optimizing the complete wait-think-answer trajectory and not the final answer alone. On a six-task synthetic spoken reasoning question answering (SRQA) benchmark, the six-reward DAPO controller improves the row-weighted accuracy from 67.6% to 70.3% while reducing post-endpoint final-think length by 14% under the same Qwen deployment harness. On a 186-item human-recorded Real Audio Bench, a transfer check beyond text-to-speech (TTS)-rendered speech, the controller family remains functional: SFT achieves the strongest accuracy, while the six-reward DAPO controller is the only learned variant whose final-think length falls below the base. These results suggest that a streaming model should learn when to make intermediate reasoning explicit during the audio stream.\n\n---\n\n## 50. LitSeg: Narrative-Aware Document Segmentation for Literary RAG\n- **Authors**: Ruikang Zhang , Zhanni Chen , Yiqiao Cai , Qi Su\n- **URL**: [https://arxiv.org/abs/2605.27156](https://arxiv.org/abs/2605.27156)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by incorporating external knowledge, particularly for long-tail domains such as literary works. However, the critical step of document segmentation in RAG remains largely underexplored. Existing strategies are typically semantically blind and overlook the complicated narrative structures of literary works, often resulting in fragmented plots and unclear references that severely hinder retrieval and generation performance. To address this, we propose LitSeg, a novel narrative-theory-guided segmentation framework. By employing multi-stage prompting, LitSeg explicitly extracts valid events, untangles narrative threads, clarifies narrative structures, and locates turning points to inform segmentation. To alleviate the computational overhead of multi-stage inference with large-scale models, we further introduce LitSeg-Lite, a lightweight single-pass chunker fine-tuned on LitSeg-generated data via a two-stage training strategy, distilling the complex process into a single inference pass. Extensive experiments demonstrate that with structurally independent text chunks, our methods significantly improve retrieval accuracy and context relevance over baselines, ultimately enhancing downstream QA performance, while ablation studies validate the efficacy of narratological guidance and data distillation.\n\n---\n\n## 51. DEI: Diversity in Evolutionary Inference for Quality-Diversity Search\n- **Authors**: John Donaghy , Shikhar Rastogi\n- **URL**: [https://arxiv.org/abs/2605.27130](https://arxiv.org/abs/2605.27130)\n- **Abstract**:\n> We present DEI: Diversity in Evolutionary Inference, a distributed Quality-Diversity (QD) search framework that assigns heterogeneous large language models (LLMs) as mutation operators across peer nodes communicating with non-blocking collective operations. Unlike homogeneous parallel search, which replicates a single model's inductive biases across all workers, DEI treats each LLM's distinct creative prior as a complementary source of behavioral novelty. Extending the Digital Red Queen framework with DEI, nodes share local optimal solutions at the end of each round to seed the next round's population. This creates cross-model adversarial pressure that drives robustness beyond intra-model self-play. Evaluated on the Core War domain, a competitive programming benchmark in which Redcode warrior programs battle inside a simulated machine, a four-node heterogeneous ensemble (GPT-5.4-mini, Claude Sonnet 4.6, GPT-5.2, and Claude Haiku 4.5) achieves 124 percent higher merged-archive QD-Score (45.90 vs. 20.46) and 28 percent higher coverage (80.6 percent vs. 63.0 percent of cells) than a single-node baseline at equal total LLM-call budget. The heterogeneous ensemble also outperforms an equally-budgeted homogeneous ensemble on QD-Score, coverage, and held-out solution generality across all four model families. These results provide the first empirical evidence that model diversity, not merely parallelism, is the key driver of gain in distributed LLM-based QD search.\n\n---\n\n## 52. ReMoE: Boosting Expert Reuse through Router Fine-Tuning in Memory-Constrained MoE LLM Inference\n- **Authors**: Xiongwei Zhu , Xiaojian Liao , Tianyang Jiang , Yusen Zhang , Liang Wang , Limin Xiao\n- **URL**: [https://arxiv.org/abs/2605.27081](https://arxiv.org/abs/2605.27081)\n- **Abstract**:\n> Fine-grained Mixture-of-Experts (MoE) models sparsely activate only a subset of experts per token, reducing activated computation while maintaining high model capacity. However, in memory-constrained inference scenarios, only a small set of experts can be cached. Experts not in the cache must be fetched from slow external storage (e.g., UFS), leading to frequent evictions and substantial I/O overhead. We propose ReMoE, a router fine-tuning framework designed to boost token-wise expert reuse. ReMoE biases the router toward recently selected experts, producing temporally stable routing that better matches cache locality constraints. By increasing short-horizon expert reuse, ReMoE reduces expert fetches from storage without adding inference-time computation. Experiments on DeepSeek and Qwen models show that ReMoE improves expert reuse by 26% while maintaining downstream task performance. Real-system evaluations further confirm these benefits, improving output throughput by 8.4% under vLLM GPU-CPU expert offloading and reducing TPOT by 43.6-49.8% under this http URL on Jetson Orin NX, corresponding to a 1.77-1.99$\\times$ decode speedup across diverse workloads. Checkpoints and usage instructions are available at this https URL .\n\n---\n\n## 53. E3: Issue-Level Backtesting for Automated Research Critique\n- **Authors**: Yashwardhan Chaudhuri , Sanyam Jain , Paridhi Mundra\n- **URL**: [https://arxiv.org/abs/2605.27072](https://arxiv.org/abs/2605.27072)\n- **Abstract**:\n> We present E3, an automated review assistant that augments reviewers and engineering teams by identifying decision-relevant technical concerns in research papers. For each concern, E3 reports its nature, its location, its bearing on the contribution, and the analysis or evidence that would resolve it, covering unsupported claims, missing ablations, weak baselines, hidden assumptions, threats to validity, and leakage risks. To evaluate E3 without contamination confounds we adopt an issue-level backtesting protocol: the corpus is restricted to papers postdating the training cutoff of every automated source, and for each paper a meta-judge that observes only anonymised reviews labels every issue-source pair as Caught, Partial, or Missed. Applied to 100 ICLR 2026 papers and 4598 judged issue rows, comparing E3 against the ICLR human reviews and two prompt-matched LLM baselines built on gpt-5.4 from OpenAI and claude-opus-4-6 from Anthropic, with meta-judge gpt-5.5, E3 attains the highest recall on every aggregate metric. Partial-inclusive recall reaches 90.2 percent, which is 15.5 points over GPT, 17.1 points over Claude, and 29.2 points over the human reviews, and strict recall preserves the ordering at 65.8 percent. On concerns raised by the human reviewers, E3 recovers 89.6 percent; on concerns the human reviewers missed it surfaces 1635 additional rows admitted into the judged union, 406 above the next-best source. Corpus, baseline prompts, judge prompt template, and evaluation code are released.\n\n---\n\n## 54. QUACK: Questioning, Understanding, and Auditing Communicated Knowledge in Multimodal Social Deduction Agents\n- **Authors**: Ye Yuan , Rui Song , Weien Li , Zeyu Li , Haochen Liu , Xiangyu Kong , Changjiang Han , Yonghan Yang , Zichen Zhao , Zixuan Dong , Fuyuan Lyu , Bowei He , Haolun Wu , Jikun Kang , Xue Liu\n- **URL**: [https://arxiv.org/abs/2605.27068](https://arxiv.org/abs/2605.27068)\n- **Abstract**:\n> Social deduction games have become a popular testbed for probing reasoning, deception, coordination, and belief modeling in Large Language Model (LLM) agents. However, most environments are scored only by game outcomes such as win rates and largely remain to text-only interaction, making it difficult to tell whether an agent's language is actually grounded in what it perceived and did, or to identify the failure modes underlying its behavior. To address this gap, we introduce QUACK, an open-source environment and evaluation framework for auditing the grounding of agent language in multimodal social reasoning. QUACK evaluates agents at three levels: game outcomes, behavioral trajectories, and utterance-level consistency. Its core Statement Verification Pipeline reconstructs each agent's ground-truth trajectory from engine logs and checks every discussion claim against it, automatically flagging spatial hallucination, unsupported accusation, deception collapse, and language-action inconsistency. Evaluating three frontier VLMs in both homogeneous and cross-model adversarial settings, we find that even the strongest agent hallucinates 15.1% of its verifiable spatial claims and makes over half of its accusations without grounded evidence. We release the full engine, evaluation framework, toolkit, and logs at this https URL .\n\n---\n\n## 55. ConVer: Using Contracts and Loop Invariant Synthesis for Scalable Formal Software Verification\n- **Authors**: Muhammad A. A. Pirzada , Weiqi Wang , Yiannis Charalambous , Konstantin Korovin , Lucas C. Cordeiro\n- **URL**: [https://arxiv.org/abs/2605.27051](https://arxiv.org/abs/2605.27051)\n- **Abstract**:\n> Formal verification of large C programs is impeded by state-space explosion: Bounded Model Checking (BMC) tools must encode the entire state space up to the predetermined bound by unrolling all nested constructs. We present ConVer, a top-down compositional verification tool. Given a C program with a top-level assertion, ConVer decomposes verification top-down: it uses a large language model (LLM) to synthesise function contracts from the system property, then alternates system-level and function-level checks in a CEGAR-CEGIS loop, refining contracts whenever a check fails via SMART ICE learning. We evaluate ConVer on four benchmark suites of increasing difficulty and against other state-of-the-art (SOTA) tools. On the Frama-C benchmark of 45 simple C programs, ConVer achieves 82-96% verification success across three LLM backends, with 93-95% of converged programs requiring only a single CEGAR-CEGIS iteration. On the X.509 parser benchmark (6~programs) and LF2C-Simple suite (17 programs), ConVer achieves 33-50% and 82-88% success respectively. On the VerifyThis suite of 11 recursive and loop-intensive programs, the Pre-Abstraction strategy achieves 55-64% success. In addition, we present ESBMC-LF a preprocessor tool that converts LF models to C while preserving the properties of the LF files, enabling ConVer to verify them. We transpile the LF Verifier Benchmarks using ESBMC-LF to C; we denote those LF-Hard. We show that ConVer successfully verifies 67% of LF-Hard benchmarks overall.\n\n---\n\n## 56. Tracing Computation Density in LLMs\n- **Authors**: Corentin Kervadec , Iuliia Lysova , Iuri Macocco , Marco Baroni , Gemma Boleda\n- **URL**: [https://arxiv.org/abs/2605.27033](https://arxiv.org/abs/2605.27033)\n- **Abstract**:\n> Transformer-based large language models (LLMs) are comprised of billions of parameters arranged in deep and wide computational graphs, but it is not clear that they exploit their full capacity for all inputs. We introduce the s-Trace method to efficiently estimate the subgraph of size s that best approximates a full model output. With this method, we find the computation in a variety of LLMs to be organized in two distinct phases. A small subgraph mostly composed of early-layer nodes can reconstruct the head of the full model output distribution. Adding further nodes, mostly located in later layers and increasingly consisting of attention heads, leads to incremental refinements in approximating the full output distribution. We find moreover that the amount of necessary computation per input correlates with model uncertainty, and that sparser subgraphs encode shallow statistics, such as unigram frequency. Overall, our results suggest a consistent modular organization in effective LLM computation, with a sparse early-layer core providing a rough prediction that is further refined through denser computations in later layers.\n\n---\n\n## 57. Evaluating the Relevance of Uncertainty Estimators for LLM Hallucination\n- **Authors**: Yedidia Agnimo , Anna Korba , Annabelle Blangero , Nicolas Chesneau , Karteek Alahari\n- **URL**: [https://arxiv.org/abs/2605.27016](https://arxiv.org/abs/2605.27016)\n- **Abstract**:\n> Large language models (LLMs) are prone to hallucinations, i.e., statements unsupported by the input or training data, hindering reliable deployment. In parallel, numerous uncertainty estimation (UE) methods have been proposed to quantify model confidence and are often implicitly treated as proxies for model failure. However, the relationship between uncertainty and hallucinations remains insufficiently characterized. We present a systematic empirical study of the association between uncertainty estimators and hallucinations in LLMs. Rather than assuming this association, we evaluate directly when and to what extent it holds. We consider a diverse set of uncertainty estimators, including information-theoretic, sampling-based, and reflexive estimators, and examine their behavior across hallucination settings. Our experiments cover both intrinsic hallucinations (violations of input faithfulness) and extrinsic hallucinations (unsupported claims relative to training data), using four complementary benchmarks, including RAGTruth and HalluLens. We find that the association is highly variable and often weak, depending on the hallucination type and the LLM under evaluation. These results challenge the use of uncertainty as a direct signal of hallucination and clarify when it provides actionable information.\n\n---\n\n## 58. ReasonOps: A Unified Operational Paradigm for Trustworthy Verified LLM Reasoning\n- **Authors**: Adnan Rashid\n- **URL**: [https://arxiv.org/abs/2605.27014](https://arxiv.org/abs/2605.27014)\n- **Abstract**:\n> Large Language Models (LLMs) have transformed artificial intelligence from primarily generative systems into increasingly capable reasoning agents. Recent advances in theorem proving, autoformalization, symbolic reasoning, and tool-augmented language models demonstrate substantial progress toward machine-assisted formal reasoning. However, current reasoning systems still suffer from hidden logical inconsistencies, hallucinated symbolic transitions, unsupported theorem applications, and limited reliability guarantees. Existing approaches remain fragmented across formal verification, runtime assurance, neuro-symbolic reasoning and trustworthy Artificial Intelligence (AI) research communities. This paper introduces ReasonOps, a unified operational paradigm for trustworthy verified reasoning systems. Inspired by operational ecosystems such as DevOps and MLOps, ReasonOps treats reasoning as a continuously monitored, verifiable, reliability-aware operational process rather than an isolated inference task. The proposed paradigm integrates semantic interpretation, autoformalization, symbolic reasoning, theorem proving, runtime assurance, probabilistic reliability estimation, and adaptive correction into a unified reasoning lifecycle. The paper further presents the ReasonOps architecture, demonstrates its workflow using an autonomous braking system analysis example, and discusses its potential role in future safety-critical autonomous AI systems. We argue that operational reasoning paradigms such as ReasonOps may become foundational infrastructure for next-generation trustworthy AI ecosystems.\n\n---\n\n## 59. Recon: Reconstruction-Guided Reasoning Synthesis for User Modeling\n- **Authors**: Alan Zhu , Mihran Miroyan , Carolyn Wang , Andrew Zhou , Lisa Dunlap , Narges Norouzi , Joseph E. Gonzalez\n- **URL**: [https://arxiv.org/abs/2605.26969](https://arxiv.org/abs/2605.26969)\n- **Abstract**:\n> User modeling aims to use language models (LMs) to mimic an individual's behavior from a corpus of past context-action pairs (e.g., conversation turns), enabling the simulation of users in settings like behavioral science, human-AI collaboration, and market research. Recent approaches augment these corpora with synthesized reasoning traces, typically generated by conditioning on both context and action. However, such conditioning constitutes post-hoc rationalization rather than reasoning: the trace is guaranteed to justify the action, but may not encode the underlying latent causal decision paths. We propose Recon, which uses action reconstruction to score reasoning traces by their predictive power: given a context and candidate reasoning, a reconstruction model predicts the action, and reconstruction fidelity determines reasoning quality. Across four domains, Recon achieves a 54.7% win rate over Backward Synthesis, a standard post-hoc rationalization baseline. Further, we find that training a reasoning synthesis model with rewards derived from Recon improves downstream user modeling performance, achieving a win rate of up to 70.0% over baselines. We further show that Recon-synthesized reasoning transfers across models, and improves user modeling beyond the reconstruction model. Our work demonstrates that post-hoc rationalization is insufficient for reasoning synthesis, and that useful and interpretable reasoning should naturally elicit the action from the context.\n\n---\n\n## 60. Tournament-GRPO: Group-Wise Tournament Rewards for Reinforcement Learning in Open-Ended Long-Form Generation\n- **Authors**: Zixuan Yang , Yiqun Chen , Wei Yang , Erhan Zhang , Zihan Shen , Xiaochi Wei , Yan Gao , Yi Wu , Yao Hu , Jiaxin Mao\n- **URL**: [https://arxiv.org/abs/2605.26958](https://arxiv.org/abs/2605.26958)\n- **Abstract**:\n> Reinforcement learning in open-ended long-form generation is challenging because reliable reference answers and automatic metrics are often unavailable. Existing rubric-based methods typically rely on pointwise LLM-as-a-judge scoring, but absolute scores are difficult to calibrate across complex responses, may provide weak discrimination among same-query rollouts, and can become saturated during optimization. We propose Tournament-GRPO, a group-wise reward framework that converts rubric-guided LLM judgments into relative rewards through repeated multi-round tournaments among same-query rollouts. Tournament-GRPO compares candidates within groups, accumulates tournament outcomes, and normalizes them into group-wise rewards for GRPO training. Experiments on Deep Research Bench show that Tournament-GRPO consistently outperforms existing reward-design baselines, achieving a 4.52-point overall-score improvement over the strongest baseline. Further analyses show that tournament rewards provide a favorable effectiveness--efficiency trade-off and that tournament design affects training dynamics. These results suggest that rubric-guided tournament comparison provides an effective reward signal for reinforcement learning in open-ended long-form generation.\n\n---\n\n## 61. JuICE: A Benchmark for Evaluating LLM-Judge in Identifying Cultural Errors\n- **Authors**: Jiho Jin , Junho Myung , Juhyun Oh , Junyeong Park , Rifki Afina Putri , Sunipa Dev , Vinodkumar Prabhakaran , Alice Oh\n- **URL**: [https://arxiv.org/abs/2605.26955](https://arxiv.org/abs/2605.26955)\n- **Abstract**:\n> As large language models (LLMs) are increasingly deployed to users around the world, they are integrated into everyday tasks across diverse cultural contexts, from drafting personal communications to brainstorming creative ideas. These tasks are inherently cultural: they require contextual appropriateness, symbolic resonance, and tacit cultural expectations that native speakers draw on instinctively, meaning that a response can be factually plausible yet unmistakably wrong to a local reader. Existing cultural benchmarks have treated culture as a flat set of facts via fact verification or norm entailment methods, and have adopted LLM-as-a-Judge without examining whether they can capture such thick cultural errors. To address this gap, we present JuICE (Benchmark for LLM-Judge in Identifying Cultural Errors), a multilingual dataset of 7,470 span-level annotations of cultural and linguistic errors in long-form LLM responses. It covers 1,050 query-response pairs from four countries (the United States, South Korea, Indonesia, and Bangladesh), in both English and their countries' main languages. Using JuICE, we find that even the strongest LLM-judge achieves only an F1 of 0.52 in the erroneous span detection task. Furthermore, LLM-judges consistently miss thick cultural errors that local residents readily identify. Our findings suggest that robust cultural evaluation must move beyond surface-level detection toward frameworks that account for the depth and situatedness of cultural meaning.\n\n---\n\n## 62. Beyond Questions: Evaluating What Large Language Models (Actually) Know\n- **Authors**: Luca Giordano , Simon Razniewski\n- **URL**: [https://arxiv.org/abs/2605.26937](https://arxiv.org/abs/2605.26937)\n- **Abstract**:\n> Parametric knowledge in large language models (LLMs) is a cornerstone of their success, yet remains poorly understood. Existing knowledge benchmarks typically rely on predefined questions (e.g., \"What is the birth date of M.L. King?\"), evaluating only knowledge that benchmark designers explicitly choose to query, a problematic availability bias. In this paper, we introduce open knowledge evaluation, a new paradigm for LLM knowledge benchmarking. Instead of asking narrow questions, it evaluates models on the knowledge they choose to surface in response to open-ended elicitation prompts (e.g., \"Tell me everything you know about M.L. King\"). This shifts the focus from predefined answer retrieval toward characterizing the knowledge models naturally express. We instantiate this paradigm with BeQu (Beyond Questions), a benchmark of 10,000 entities paired with reference corpora for statement verification. Using BeQu, we evaluate a broad range of language models and analyze the effects of reasoning effort, model scale, prompt format, and knowledge domain. Data and leaderboard are available on this work's GitHub repository and at the benchmark's website.\n\n---\n\n## 63. Strategies for Guiding LLMs to Use Software Design Patterns: A Case of Singleton\n- **Authors**: Viktor Kjellberg , Farnaz Fotrousi , Miroslaw Staron\n- **URL**: [https://arxiv.org/abs/2605.26898](https://arxiv.org/abs/2605.26898)\n- **Abstract**:\n> Large Language Models (LLMs) can generate functional source code from natural-language prompts, but often fail to consistently follow higher-level architectural structures or design patterns. Since LLMs are increasingly used in software engineering, their ability to apply established design principles to generated code is crucial to the long-term success of software products. Therefore, the goal of this paper is to identify strategies for guiding LLMs to incorporate design patterns into the generated source code. We designed a computational experiment to evaluate the ability of 13 LLMs to generate code that follows the Singleton design pattern, using four prompting strategies: instructions, binary automated feedback, extensive automated feedback, and extensive feedback with few-shot prompts, in 164 Java coding challenges from HumanEval-X. Our results shows that the optimal strategy to guide LLMs to include design patterns depends heavily on the type of model. Still, overall, iterative binary feedback provides the best alignment with Singleton while preserving or improving the code's functionality. With guiding with instructions, Llama 3.3 generated Singleton classes in 100% of cases and improved code functionality, increasing the number of tests passed by 34.1 percentage points. It achieved a similar result with guidance through instructions and binary feedback. Qwen 3 (8B) increased the alignment with Singleton to 99.2% and the functionality to 58.6% using binary feedback. Our result suggests that even simple strategies can be used to guide LLMs to use design patterns.\n\n---\n\n## 64. Negligible in Size, Significant in Effect: On Scale Vectors in Large Language Models\n- **Authors**: Mingze Wang , Shuchen Zhu , Yuxin Fang , Binghui Li , Kai Shen , Shu Zhong\n- **URL**: [https://arxiv.org/abs/2605.26895](https://arxiv.org/abs/2605.26895)\n- **Abstract**:\n> Normalization layers in modern large language models (LLMs) consist of a deterministic normalization operation and a learnable scale vector. While the normalization operation has been extensively studied, the scale vector remains poorly understood despite its ubiquitous use. In this work, we present a systematic study of scale vectors in LLMs from the perspectives of expressivity, optimization, and architectural structure. First, we show empirically that although scale vectors constitute only a negligible fraction of model parameters, removing them substantially degrades LLM pre-training. Our theory further shows that, in Pre-Norm architectures, scale vectors do not increase expressivity; instead, they improve optimization through a self-amplifying preconditioning effect on subsequent linear mappings. Second, we investigate the role of weight decay for scale vectors. By distinguishing Input-Norm and Output-Norm layers, we theoretically show that weight decay is beneficial for the former but harmful for the latter, due to their distinct roles in optimization and expressivity. Third, motivated by this understanding, we propose three lightweight and complementary improvements to scale vectors: branch-specific heterogeneity, improved placement around linear mappings, and magnitude-direction reparameterization. Both theory and experiments show that each improvement yields consistent gains. Finally, we combine these improvements into a unified scale-vector strategy and evaluate it through extensive LLM pre-training experiments on dense and mixture-of-experts models ranging from 0.12B to 2B parameters, across multiple optimizers and learning rate schedules, under industrial-scale token budgets. The unified strategy consistently achieves lower terminal loss than well-tuned baselines and exhibits more favorable scaling behavior, while adding negligible parameter and computational overhead.\n\n---\n\n## 65. GeoFaith: A Spatio-Temporal Dual View of Faithful Chain-of-Thought\n- **Authors**: Weijiang Lv , Wentong Zhao , Jiayu Wang , Yuhao Wu , Jiaheng Wei , Xiaobo Xia\n- **URL**: [https://arxiv.org/abs/2605.26893](https://arxiv.org/abs/2605.26893)\n- **Abstract**:\n> Chain-of-Thought (CoT) reasoning has advanced large language models (LLMs), but outcome-based supervision leads to pervasive post-hoc rationalization, producing plausible yet unfaithful reasoning chains. Most prior faithfulness assessment methods are either unscalable, expensive, or unreliable. We propose GeoFaith, a spatio-temporal framework that leverages latent geometric structure and entropy dynamics to diagnose and enforce faithful reasoning. We develop a scalable bootstrapping pipeline expanding step-level annotations from 1k to 20k samples across four domains, train an 8B faithfulness detector outperforming GPT-5 on standard benchmarks, and design a faithfulness-aware reinforcement learning framework jointly optimizing outcome correctness, process faithfulness, and trajectory consistency. Experiments show the proposed method achieves superior performance on both faithfulness detection and downstream reasoning, producing shorter, more interpretable chains without sacrificing accuracy. Our code will be made available publicly.\n\n---\n\n## 66. Knowledge Graphs as the Missing Data Layer for LLM-Based Industrial Asset Operations\n- **Authors**: Madhulatha Mandarapu , Sandeep Kunkunuru\n- **URL**: [https://arxiv.org/abs/2605.26874](https://arxiv.org/abs/2605.26874)\n- **Abstract**:\n> LLM-based agents for industrial asset operations show limited accuracy when reasoning over flat document stores. AssetOpsBench (KDD 2026) establishes that GPT-4 agents achieve 65% on 139 industrial maintenance scenarios backed by CouchDB, YAML, and CSV. It compares LLM orchestration paradigms (Agent-As-Tool vs Plan-Execute) on a fixed data layer; we ask a complementary, orthogonal question: how much does the data model behind the tools affect agent performance? Building on the same scenarios, we introduce a knowledge graph layer (781 nodes, 955 edges, 16 relationship types) and evaluate three architectures: (1) deterministic graph handlers (no LLM) at 99% (137/139); (2) LLM-generated Cypher over the graph at 82-83% with the same GPT-4 model the baseline uses; and (3) the original tool-augmented LLM baseline at 65% (91/139, matching the published KDD 2026 leaderboard ceiling). Our key finding is inverted LLM usage: rather than asking the LLM to reason over raw data, we ask it to generate structured queries from a typed schema. The graph executes deterministically. We additionally contribute 40 graph-native scenarios (multi-hop dependency, vector similarity, PageRank criticality), and evaluate against the expanded HuggingFace AssetOpsBench release (467 scenarios, 6 domains), where deterministic handlers achieve 100% (467/467) with average score 0.848. These results suggest that for structured operational domains, the data layer -- not the LLM orchestration -- is the primary bottleneck, and that knowledge graphs serve as an integration layer between raw industrial data and LLM-based reasoning.\n\n---\n\n## 67. The Strongest Teacher Is Not Always the Best Teacher: Student-Centric Answer Selection\n- **Authors**: Zhengyu Hu , Zheyuan Xiao , Linxin Song , Fengqing Jiang , Yutai Li , Zhengyu Chen , Zhihan Xiong , Yue Liu , Junhao Lin , Yao Su , Lijie Hu , Kaize Ding , Xiao Teng , Radha Poovendran\n- **URL**: [https://arxiv.org/abs/2605.26872](https://arxiv.org/abs/2605.26872)\n- **Abstract**:\n> LLM training increasingly relies on teacher-generated supervision, from synthetic responses to reasoning traces and tool-use demonstrations. Current practice often chooses the highest-performing teacher to generate student training data, implicitly treating teacher test performance as a proxy for teaching quality. We show that this assumption can fail: even when multiple teachers provide correct answers to the same question, the answer from the strongest teacher is not necessarily the best supervision for a given student. To address this gap, we propose Student-Centric Answer Sampling (SCAS), a framework that selects from verified teacher-generated answers according to their estimated student-centric learning cost. Motivated by a token-wise gradient decomposition, we derive an efficient forward-only proxy for this cost and use it to guide answer selection during training. Experiments across 30 teacher models, 6 student base models, and 8 tasks show that SCAS consistently improves student performance, suggesting that effective distillation should prioritize supervision matched to the current student rather than teacher strength alone.\n\n---\n\n## 68. Persistent AI Agents in Academic Research: A Single-Investigator Implementation Case Study\n- **Authors**: Anas H. Alzahrani\n- **URL**: [https://arxiv.org/abs/2605.26870](https://arxiv.org/abs/2605.26870)\n- **Abstract**:\n> Background: Large language models are typically evaluated as models, benchmarks, or short conversational episodes. Less is known about what happens when an agent is embedded persistently in a real academic research environment with durable memory, local files, external tools, scheduled routines, delegated roles, and explicit safety protocols. Methods: A structured self-observed implementation case study was conducted from January 31 to May 25, 2026. The unit of analysis was the persistent human-agent environment: researcher, agent runtime, memory layer, tools, repositories, scheduled jobs, specialized agent roles, and governance rules. Outcomes were organized using PARE-M (Persistent Agentic Research Environment Measurement), a measurement framework covering architecture, utilization, artifact production, resource use, reproducibility, and governance. Results: Recoverable main-agent telemetry contained 75,671 de-duplicated records across 96 active days, with 8,059 user-role and 23,710 assistant-role messages. The workspace included 502 memory-related files, 17 configured agent directories, and 57 skill files. Active system time was 579.7 hours (30-minute capped-gap estimate). Memory-derived records identified 482 output-proxy events and 889 failure, verification, correction, or protocol-proxy events. A strict May 2026 trajectory subset captured 627 model-completed events and 73.95 million recorded tokens, of which 82.9% were cache reads. Conclusions: The workflow was cache-dominant, suggesting that persistent agentic environments may shift the economic unit from cost per token to cost per completed artifact. Future evaluations should use artifact-level denominators, reproducible parsing rules, correction taxonomies, and independent coding of governance events.\n\n---\n\n## 69. The Kalman Evolve: Closing the Gap in Kalman Filtering via Interpretable Algorithm Discovery\n- **Authors**: Vasileios Saketos , Ming Xiao\n- **URL**: [https://arxiv.org/abs/2605.26830](https://arxiv.org/abs/2605.26830)\n- **Abstract**:\n> State estimation is a fundamental problem in control and signal processing, for which the Kalman Filter provides an optimal solution under linear dynamics, Gaussian noise, and known noise covariances. However, these assumptions often fail in realistic sensing settings such as Doppler radar and LiDAR. In these cases, the optimal estimator is inherently nonlinear, which leads to systematic performance degradation. This creates a performance gap that cannot be eliminated by tuning the noise covariance parameters (i.e., the process and measurement noise in the Kalman Filter) alone. To address this limitation, we propose Kalman Evolve, a framework for discovering improved filtering algorithms by jointly optimizing both noise parameters and the update structure. Our approach leverages large language models (LLMs) as a structured prior over program space, enabling the generation of interpretable, non-affine modifications to the classical Kalman filter while preserving its recursive form. We provide analytical results establishing the suboptimality of affine estimators under common nonlinear sensing models, motivating the need for structure-aware updates. Across a range of synthetic and real-world tracking benchmarks, including Doppler radar, LiDAR-based localization, and pedestrian tracking, the discovered algorithms consistently improve over strong baselines such as the Optimized Kalman Filter, achieving up to 12\\% reduction in RMSE. These results suggest that optimizing the structure of the Kalman filter, rather than only its parameters, provides a practical and interpretable way to improve state estimation.\n\n---\n\n## 70. ContextGuard: Structured Self-Auditing for Context Learning in Language Models\n- **Authors**: Hongbo Jin , Chi Wang , Haoran Tang , Zhongjing Du , Xu Jiang , Jingqi Tian , Qiaoman Zhang , Jiayu Ding\n- **URL**: [https://arxiv.org/abs/2605.26827](https://arxiv.org/abs/2605.26827)\n- **Abstract**:\n> Recent benchmarks reveal that despite strong reasoning capabilities, large language models (LLMs) still struggle to faithfully apply complex contextual knowledge. These failures are often not wholesale reasoning collapses: in context-rich tasks, models may follow the central reasoning path while missing peripheral, persistent, or format-sensitive requirements.\n\n---\n\n## 71. RAGEAR: Retrieval-Augmented Graph-Enhanced Academic Recommender\n- **Authors**: Francesco Granata , Lorenzo Lamazzi , Misael Mongiovì , Francesco Poggi , Valeria Secchini\n- **URL**: [https://arxiv.org/abs/2605.26819](https://arxiv.org/abs/2605.26819)\n- **Abstract**:\n> We present RAGEAR (Retrieval-Augmented Graph-Enhanced Academic Recommender), a neurosymbolic recommender system for academic course recommendation. RAGEAR combines dense retrieval over full lecture transcripts with a symbolic Knowledge Graph modelling courses, lessons, transcript chunks, credits, study plans, and curricular information. The Knowledge Graph supports symbolic filtering and contextualisation based on structured constraints, such as credits, academic disciplines, study plans, and prerequisites. Unlike metadata-based approaches, it exploits fine-grained instructional content by retrieving transcript chunks semantically aligned with a student's query. The main contribution is a graph-aware aggregation function that propagates chunk-level evidence to course-level recommendations. The score combines three factors: the share of retrieved similarity associated with a course, the rank-based strength of its relevant chunks, and the distribution of evidence across lessons. We evaluate RAGEAR on 152 student-like queries through a human evaluation sample and a large-scale LLM-based relevance assessment. Results show that lecture transcripts improve over metadata-only retrieval, and that RAGEAR further improves ranking quality over a transcript-based normalized SumP baseline, especially for top-ranked recommendations.\n\n---\n\n## 72. Innovation: An Almost Characterization of Hallucination\n- **Authors**: Nishant P. Das , Piyush Srivastava\n- **URL**: [https://arxiv.org/abs/2605.26808](https://arxiv.org/abs/2605.26808)\n- **Abstract**:\n> Hallucination is a central limitation of large language models (LLMs), and substantial effort has been devoted to understanding and mitigating it. Towards this, Kalai and Vempala (STOC 2024) introduced a probabilistic framework formalizing calibration and hallucination, and showed that, with high probability, calibrated LLMs hallucinate roughly at the rate of the \"missing mass\", a measure of how incomplete the training data is relative to its source. This raises two fundamental questions: (i) what property of a calibrated LLM makes hallucinations unavoidable? and (ii) can hallucinations be avoided by giving up calibration? We answer these questions by introducing a simpler property we call innovation that measures the tendency of a model to produce outputs outside the training data. We show that innovation is implied by the condition for hallucination identified by Kalai and Vempala, and, further, that it is an almost characterization of hallucination: hallucination implies innovation, and conversely, innovation implies hallucination with high probability. We also provide lower bounds on the hallucination rate based on the \"innovation rate\", and by relating innovation rate back to missing mass, we obtain new hallucination rate lower bounds based on missing mass that extend the results of Kalai and Vempala.\n\n---\n\n## 73. SeDT: Sentence-Transformer Decision-Transformer Conditioning for Multi-Turn Conversation Reliability\n- **Authors**: Ramakrishna Vamsi Setti , Jagadeesh Rachapudi , Sachin Chaudhary , Praful Hambarde , Amit Shukla\n- **URL**: [https://arxiv.org/abs/2605.26788](https://arxiv.org/abs/2605.26788)\n- **Abstract**:\n> Large language models (LLMs) achieve impressive performance when a task is fully specified in a single turn, yet the same models lose up to 39% of that performance when the identical task is revealed incrementally across multiple turns, a phenomenon documented at scale as Lost in Conversation. Crucially, this collapse is almost entirely a reliability failure; the best case, the aptitude only falls 16%, while the unreliability more than doubles (+112%). We argue that the root cause is structural, a flat conversation history assigns equal implicit weight to every prior turn, giving the model no signal to distinguish a critical constraint from incidental dialog. We present SeDT Sentence-transformer Decision-Transformer, a training-free inference-time method that resolves this by importing return-to-go conditioning from offline reinforcement learning. SeDT annotates each conversation shard with a cumulative relevance score derived from three complementary semantic, lexical, and positional signals and presents the full annotated history to the model at the final turn, without weight changes, without training data, and without discarding context. Evaluated on the Lost-in-Conversation benchmark in three LLMs and three generation tasks, SeDT outperforms the sharded baseline in all nine model-task combinations, with gains up to +37.7% in mean performance P and simultaneous reductions in unreliability in seven of the nine combinations. In short, telling the model which past turns matter is sufficient to substantially recover the performance lost in conversation.\n\n---\n\n## 74. EmoDistill: Offline Emotion Skill Distillation for Language Model Agents in Adversarial Negotiation\n- **Authors**: Yunbo Long , Haolang Zhao , Lukas Beckenbauer , Liming Xu , Alexandra Brintrup\n- **URL**: [https://arxiv.org/abs/2605.26785](https://arxiv.org/abs/2605.26785)\n- **Abstract**:\n> Post-trained LLMs are often optimized to align responses with human preferences, making them safe, polite, and conversationally appropriate. In adversarial negotiation, however, this alignment can become a vulnerability: emotionally framed language may steer agents toward the counterparty's interests. Using GoEmotions-based affective prompting, we show that emotion substantially shifts negotiation outcomes, suggesting that emotion is a strategic action channel rather than a surface style. Thus, we introduce \\textbf{EmoDistill}, an offline framework for distilling emotional negotiation skills into language model agents. EmoDistill decomposes emotional strategy into emotion selection and emotion expression: an Implicit Q-Learning (IQL) selector learns \\emph{which} emotion to express, while a Low-Rank Adaptation (LoRA)-based policy learns \\emph{how} to express it through Supervised Fine-Tuning (SFT) and Judge Policy Optimization (JPO). Across four emotion-sensitive, high-stakes negotiation domains, SLM policies trained under the EmoDistill framework achieve the highest utility, outperforming vanilla SLM/LLM baselines and IQL-only emotion selection. Ablations show that emotion conditioning is essential, and transfer studies demonstrate generalization across domains, unseen counterparties, and trained-vs-trained tournaments. Overall, EmoDistill learns skills from offline agent-to-agent interactions, avoiding costly online negotiation during training.\n\n---\n\n## 75. Ratio-Variance Regularized Policy Optimization\n- **Authors**: Yu Luo , Shuo Han , Yihan Hu , Lei Lv , Huaping Liu , Fuchun Sun , Jianye Hao , Dong Li\n- **URL**: [https://arxiv.org/abs/2605.26784](https://arxiv.org/abs/2605.26784)\n- **Abstract**:\n> Standard on-policy reinforcement learning relies on heuristic clipping to enforce trust regions, but this mechanism imposes a severe cost by indiscriminately truncating high-return yet high-divergence updates. We demonstrate that explicitly constraining the policy ratio variance provides a principled local approximation to trust-region constraints, eliminating the need for binary hard clipping. By acting as a distributional ``soft brake'', this approach preserves critical gradient signals from novel discoveries while naturally down-weighting and enabling the reuse of stale, off-policy data. We introduce ${\\bf R}^2{\\bf VPO}$ (Ratio-Variance Regularized Policy Optimization), which implements this constraint via a primal-dual optimization framework. Extensive evaluations across $7$ LLM scales, spanning both fast and slow reasoning paradigms, and $10$ robotic control tasks demonstrate the generality of the proposed approach. R$^2$VPO achieves substantial performance gains on mathematical reasoning benchmarks, with particularly pronounced improvements on smaller models, while significantly improving sample efficiency. Furthermore, it consistently outperforms PPO baselines in continuous control domains, particularly in sparse-reward and dynamic environments. Together, these findings establish ratio-variance regularization as a principled foundation for stable and data-efficient policy optimization.\n\n---\n\n## 76. MatFormBench: A Benchmarking Evaluation Framework for Target-Driven Materials Formulation\n- **Authors**: Linhan Wu , Chenxi Wang , Chuhan Yang , Zhengwei Yang , Yuyang Liu\n- **URL**: [https://arxiv.org/abs/2605.26741](https://arxiv.org/abs/2605.26741)\n- **Abstract**:\n> Inverse design of materials has significantly advanced target-driven formulation optimization, yet existing materials machine learning benchmarks remain limited to forward property prediction, failing to systematically evaluate inverse optimization and generation algorithms, a critical gap that hinders the progress of target-driven materials design. To address this limitation, we propose MatFormBench, a novel benchmarking ecosystem tailored to evaluate and guide generative strategies for target-driven formulation. MatFormBench integrates a physics-driven formulation generation scheme to generate synthetic samples that faithfully emulate realistic materials structure-property response relationships, complemented by five escalating difficulty levels to quantify the complexity of these relationships. To rigorously assess algorithm performance, we further propose MatFormScore, a multi-dimensional metric that comprehensively quantifies performance across five critical axes: target success, search efficiency, exploratory capacity, robustness, and stability. We validate MatFormBench by evaluating 39 diverse inverse design algorithms, covering classical surrogate-assisted black-box search, state-of-the-art deep generative models, and increasingly popular Large Language Model (LLM)-based recommendation strategies. Across 1170 standardized algorithm-task evaluations, diffusion-based models demonstrate the strongest overall performance, while Variational Autoencoder (VAE)-based and Genetic Algorithm (GA)-based methods exhibit distinct advantages in specific scenarios. By establishing a unified evaluation standard for target-driven materials formulation, MatFormBench enables reproducible benchmarking, principled algorithm comparison, and diagnostic analysis of inverse design strategies, providing a foundational tool for advancing materials inverse design.\n\n---\n\n## 77. Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models\n- **Authors**: Xiao-Wen Yang , Ziyu Han , Xi-Hua Zhang , Wen-Da Wei , Jie-Jing Shao , Lan-Zhe Guo , Yu-Feng Li\n- **URL**: [https://arxiv.org/abs/2605.26733](https://arxiv.org/abs/2605.26733)\n- **Abstract**:\n> Looped Language Models (LoopLMs) enable efficient latent reasoning through depth recurrence, yet exhibit unreliable test-time scaling behavior: performance often peaks at a certain iteration depth and then collapses with further recurrence. Through latent dynamics analysis, we find an inherent trade-off between stability and effectiveness in existing architectures and strategies. By conceptualizing reasoning as uncertainty reduction, we propose that convergence toward stable fixed points while preserving effectiveness represents a promising way. To this end, we propose STARS (STAbility-driven Recurrent Scaling), a training framework that constrains latent states to approach asymptotically stable fixed points. This is realized via efficient Jacobian Spectral Radius Regularization with random loop sampling, enabling STARS to maximize effectiveness while ensuring rigorous stability. Experiments on arithmetic tasks show that STARS achieves reliable test-time scaling, and on complex mathematical reasoning it substantially mitigates performance degradation as recurrence depth increases while also improving peak performance.\n\n---\n\n## 78. L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation\n- **Authors**: Pingjun Pan , Tingting Zhou , Peiyao Lu , Tingting Fei , Hongxiang Chen , Chuanjiang Luo\n- **URL**: [https://arxiv.org/abs/2605.26717](https://arxiv.org/abs/2605.26717)\n- **Abstract**:\n> Adapting large language models (LLMs) for personalized recommendation requires aligning their general-purpose capabilities with user-specific preferences while effectively leveraging both behavioral and semantic signals. Existing approaches typically integrate these signals at either the input level (e.g., injecting behavioral embeddings into the token space) or the output level (e.g., contrastive alignment of separate encoders), suffering from distribution gaps or lack of end-to-end task supervision. In this work, we introduce L2Rec, which unifies behavioral and semantic understanding at the parameter level of LLMs. Our key insight is that the same set of Transformer parameters can serve as a shared medium for both views: by applying view-specific, personalized low-rank perturbations via a Dual-view Personalized Mixture-of-Experts (DPMoE) mechanism, L2Rec enables a single LLM backbone to produce complementary behavioral and semantic adaptations for each user with minimal representation-level misalignment. An adaptive cross-view fusion module further integrates the dual-view outputs into a unified user preference. Experiments on four datasets show that L2Rec consistently outperforms state-of-the-art baselines, and online A/B testing on a large-scale industrial platform validates significant improvements in key engagement metrics.\n\n---\n\n## 79. Beyond Trajectory-Level Attribution: Graph-Based Credit Assignment for Agentic Reinforcement Learning\n- **Authors**: Xin Cheng , Shuo He , Lang Feng , HaiYang Xu , Ming Yan , Lei Feng , Bo An\n- **URL**: [https://arxiv.org/abs/2605.26684](https://arxiv.org/abs/2605.26684)\n- **Abstract**:\n> Group-based reinforcement learning (RL) methods have achieved remarkable success in improving the performance of large language models (LLMs) and have been rapidly extended to agentic tasks. However, their credit assignment relies heavily on coarse-grained trajectory-level attribution according to final outcomes, making it difficult to capture the contribution of individual steps, such as valuable steps obscured within failed trajectories. To uncover latent information and enable more faithful step-level credit assignment, we propose Graph-based Group Policy Optimization (GraphGPO), which first aggregates all rollout trajectories into a unified state-transition graph and then estimates the distance from each state to the task goal using the global information encoded in the graph. Finally, GraphGPO assigns credit to each edge by estimating a graph-based advantage, based on how much the transition reduces the distance to the task goal. In this way, GraphGPO significantly improves training efficiency and achieves state-of-the-art performance across a range of challenging benchmarks.\n\n---\n\n## 80. An In-Vitro Study on Cross-Lingual Generalization in Language Models\n- **Authors**: Adrian Cosma\n- **URL**: [https://arxiv.org/abs/2605.26683](https://arxiv.org/abs/2605.26683)\n- **Abstract**:\n> Cross-lingual transfer in language models is difficult to study in natural corpora because lexical overlap, morphology, data imbalance, and tokenization are entangled. We introduce an in-vitro framework with two procedurally generated languages that share the same ontology, typed grammar, and compositional structure, but differ in surface realization. This lets us independently vary lexical distance, minority-language proportion, tokenizer training regime, and vocabulary size, while evaluating transfer on a masked minority-language condition whose lexical forms are never observed during training. Across 700 controlled runs, we find that transfer is governed less by tokenizer balance or raw lexical similarity than by whether tokenization preserves reusable cross-lingual substructure. Smaller vocabularies often improve masked transfer by keeping words decomposable into shared fragments, whereas larger vocabularies can turn forms into language-specific atoms. We further show that transfer emerges as a staged process: grammatical and type-level competence precede masked lexical generalization. Finally, we attempt to explain this mechanism through tokenizer bridges and show that bridge strength correlates strongly with masked reachability.\n\n---\n\n## 81. DynFrame: Adaptive Reasoning-Driven Multimodal Framework with Dynamic Frame Augmentation for Complex Video Understanding\n- **Authors**: Peng Zhang , Guanghao Zhang , Wanggui He , Longxiang Zhang , Mushui Liu , Yan Xia , Zhenhao Peng , Weilong Dai , Jinlong Liu , Haobing Tang , Le Zhang , Hao Jiang , Pipei Huang\n- **URL**: [https://arxiv.org/abs/2605.26680](https://arxiv.org/abs/2605.26680)\n- **Abstract**:\n> Recent video multimodal large language models (MLLMs) increasingly couple step-by-step reasoning with on-demand visual evidence retrieval, allowing models to revisit relevant video segments during inference. However, two structural gaps remain in existing thinking-with-video systems. (i) Sampling density is not a learnable decision: existing methods may let the model decide where to look, but the per-window frame rate is largely fixed. As a result, fine-grained evidence is often recovered through repeated retrieval calls, which increases inference context length and training difficulty. (ii) Retrieval and answer generation are usually optimized with a single trajectory-level advantage, so the \"where to look\" tokens and the \"how to answer\" tokens receive the same credit even when one is correct and the other is not. To address these gaps, we present DynFrame, a framework that emits the temporal window and the sampling density as native tokens within a single autoregressive pass. This learnable span-density retrieval enables acquiring multi-granularity evidence with a single retrieval step. Based on the above tokenized retrieval interface, we further introduce Segment-Decoupled GRPO (SD-GRPO), which splits each rollout at the retrieval boundary and assigns role-specific token-level advantages, separately crediting the sampling decision and the answer. Trained on the curated DM-CoT-74k and DM-RL-45k, DynFrame-4B is competitive with strong 7B-8B baselines across six benchmarks (NExT-GQA, Charades-STA, ActivityNet-MR, Video-MME, MLVU, LVBench), and DynFrame-8B sets new state-of-the-art on most metrics. Code is available at this https URL .\n\n---\n\n## 82. The Labyrinth and the Thread: Rethinking Regularizations in Sequential Knowledge Editing for Large Language Models\n- **Authors**: Zheng Wang , Kaixuan Zhang , Wanfang Chen , Jingwen Zhang , Xiaonan Lu\n- **URL**: [https://arxiv.org/abs/2605.26670](https://arxiv.org/abs/2605.26670)\n- **Abstract**:\n> Sequential editing of structured knowledge in large language models allows targeted factual updates without retraining, yet existing methods often rely on complex regularization or constraint mechanisms whose necessity remains unclear. In this work, we systematically investigate the mechanisms underlying effective and stable sequential editing. Specifically, we first analyze the empirical success of AlphaEdit and establish, via a rigorous optimization analysis, the formal equivalence between one-time and sequential editing. Building on this insight, we generalize the equivalence to a broader class of editing objectives, demonstrating that stability emerges naturally from properly accounting for accumulated editing constraints, rather than from specialized regularization or null-space operations. We empirically confirm that many commonly used regularization strategies are unnecessary for reliable sequential updates. Furthermore, we extend our framework to handle conflicting edits, ensuring robust and consistent behavior under contradictory updates. Ultimately, our work provides Ariadne's thread through the labyrinth of sequential editing, charting a path toward simpler, more interpretable, and dependable knowledge updates. Our code is available at this https URL .\n\n---\n\n## 83. AI evaluation may bias perceptions: The importance of context in interpreting academic writing\n- **Authors**: Shang Wu , Randol Yao\n- **URL**: [https://arxiv.org/abs/2605.26662](https://arxiv.org/abs/2605.26662)\n- **Abstract**:\n> This paper examines how estimates of AI use in scientific writing can be biased when evaluation methods ignore contextual differences across countries and fields. Using large-scale data on journal publications from Dimensions, we construct AI-likeness benchmarks based on differences between human-written and LLM-rephrased abstracts. We show that a pooled benchmark may confound pre-existing stylistic variation with AI-generated text, producing substantial distortions across country-field groups even in pre-LLM publications. In contrast, country-field-specific benchmarks attenuate such distortions and provide a more credible baseline for comparison. Applying these methods to publications in 2025 reveals that the pooled benchmark systematically overestimates AI use in certain countries and fields while underestimating it in others. These findings highlight the importance of context-aware measurement for accurate and equitable evaluation of AI use in science.\n\n---\n\n## 84. Respecting Modality Gap in Post-hoc Out-of-distribution Detection with Pre-trained Vision-Language Models\n- **Authors**: Yuanwei Hu , Bo Peng , Yadan Luo , Zhen Fang , Ling Chen , Jie Lu\n- **URL**: [https://arxiv.org/abs/2605.26661](https://arxiv.org/abs/2605.26661)\n- **Abstract**:\n> Out-of-distribution (OOD) detection has emerged as a popular technique to enhance the reliability of machine learning models by identifying unexpected inputs from unknown classes. Recent progress in pre-trained vision-language models (VLMs) has enabled zero-shot OOD detection without access to in-distribution (ID) training data; in this setting, existing methods commonly treat text embeddings of class names as class prototypes. In this paper, we challenge the widely adopted text-as-prototype paradigm by theoretically showing that off-the-shelf textual prototypes are generally misaligned with the optimal visual prototypes, yielding an intrinsic modality gap that cannot be eliminated by prompt engineering alone. To mitigate this gap under the post-hoc constraint, this paper presents an online pseudo-supervised framework that directly learns class prototypes in the visual feature space using unlabeled test-time data streams and soft predictions from the pre-trained VLMs. We provide theoretical guarantees for the convergence of the online optimization procedure. Extensive experiments empirically demonstrate that our method achieves a new state of the art across a variety of OOD detection setups.\n\n---\n\n## 85. More Expressive Feedforward Layers: Part I. Token-Adaptive Mixing of Activations\n- **Authors**: Mingze Wang , Jinbo Wang , Yikuan Xia , Kai Shen , Shu Zhong\n- **URL**: [https://arxiv.org/abs/2605.26647](https://arxiv.org/abs/2605.26647)\n- **Abstract**:\n> Feedforward network (FFN) layers account for a large fraction of parameters and nonlinear expressivity in Transformer-based large language models (LLMs). Despite the evolution from ReLU and GELU to gated variants such as SwiGLU, most FFN designs still use a single fixed activation function, applying the same nonlinear transformation to all tokens. In this work, we propose Mixture of Activations (MoA), a token-adaptive FFN design that mixes a dictionary of activation functions using lightweight input-dependent gates while sharing the same linear projections. As an input-independent counterpart, we also introduce learnable activations (LA), which form linear combinations of activation functions for both ReLU-type and SwiGLU-type FFNs. Theoretically, we establish strict finite-width expressive separations among fixed-activation FFNs, LA, and MoA: LA strictly contains fixed-activation FFNs, while MoA strictly contains LA, with the additional expressivity arising from input-dependent nonlinear hybridization. Empirically, we evaluate MoA through extensive pre-training experiments on dense and MoE language models ranging from 0.12B to 2B parameters under different token budgets, optimizers, and learning rate schedules. MoA consistently achieves lower terminal loss and exhibits more favorable scaling behavior than well-tuned baselines, with minimal parameter and computational overhead. These results suggest that token-adaptive activation mixing is a simple and effective mechanism for improving FFN expressivity in LLMs.\n\n---\n\n## 86. Spend Your Rollouts Where It Counts: Rollout Allocation for Group-Based RL Post-Training\n- **Authors**: Woojeong Kim , Ziyi Yang , Jing Nathan Yan , Jialu Liu\n- **URL**: [https://arxiv.org/abs/2605.26606](https://arxiv.org/abs/2605.26606)\n- **Abstract**:\n> Reinforcement learning (RL) is the dominant paradigm for post-training large language models. However, in the online, on-policy setting, rollout generation dominates the computational cost of training. Group-based policy optimization methods compute advantages from multiple rollouts per prompt, yet they indiscriminately allocate budget to prompts with collapsed reward distributions, wasting expensive rollouts on negligible learning signals. We demonstrate that group-based updates are most effective in regimes of high reward variance. Since the policy evolves throughout training, prompt informativeness must be estimated online rather than precomputed, but exhaustively evaluating every prompt is computationally prohibitive. We introduce Pilot-Commit, a budget-aware rollout allocation framework for group-based RL post-training. Pilot-Commit decouples prompt evaluation from exploitation: a pilot stage estimates per-prompt informativeness using a fraction of the budget, and the remaining rollouts are allocated to high-leverage prompts while low-signal prompts are skipped. Across multiple math reasoning benchmarks and model scales from 1.5B to 14B parameters, Pilot-Commit matches baseline accuracy with significantly lower sampling costs, reaching target accuracy up to $1.9\\times$ faster than GRPO and $4.0\\times$ faster than DAPO in cumulative rollouts.\n\n---\n\n## 87. Cordyceps: Covert Control Attacks on LLMs via Data Poisoning\n- **Authors**: Zedian Shao , Charles Fleming , Teodora Baluta\n- **URL**: [https://arxiv.org/abs/2605.26595](https://arxiv.org/abs/2605.26595)\n- **Abstract**:\n> Large language models (LLMs) are often fine-tuned on uncurated text datasets that adversaries can poison. Existing poisoning attacks primarily rely on fixed trigger phrases that defenses such as outlier detection, clean-data regularization, or online monitoring can neutralize. In this paper, we propose a data poisoning method that teaches an LLM an information hiding scheme reliably and stealthily through semantic associations between shared knowledge such as facts or concepts and attacker-chosen phrases. The induced hiding scheme can encode and decode arbitrary malicious instructions, thus revealing a new and subtle poisoning-induced vulnerability: covert control attacks. We precisely characterize covert control attacks and evaluate them across $5$ LLMs, $3$ backdoor defenses, and $4$ prompt injection defenses. With a small poisoned fraction, covert control attacks outperform heuristic-based prompt injection attacks in average attack success rate by about $40\\%$ relative to clean fine-tuned models. They also circumvent defenses based on detection and fine-tuning, maintaining up to $93\\%$ attack success rate after backdoor defenses and up to $98\\%$ after prompt injection defenses.\n\n---\n\n## 88. Linear and Neural Dueling Bandits with Delayed Feedback\n- **Authors**: Xiangyi Wang , Pingchen Lu , Jie Mao , Mingze Kong , Zhi Hong , Zhiyong Wang , Zhongxiang Dai\n- **URL**: [https://arxiv.org/abs/2605.26554](https://arxiv.org/abs/2605.26554)\n- **Abstract**:\n> Contextual dueling bandits form a cornerstone of preference-based decision-making, with critical applications in recommender systems and large language model alignment. However, standard algorithms rely on the idealized assumption of immediate feedback, a condition frequently violated in real-world scenarios such as prompt optimization. This setting introduces a unique theoretical challenge: unlike linear bandits, dueling bandit estimators lack closed-form solutions, rendering naive adaptations of standard weighting techniques biased. To address this, we formalize the problem of Contextual Dueling Bandits with Stochastic Delayed Feedback and propose two novel algorithms: Linear (LDB-DF) and Neural (NDB-DF) Dueling Bandits with Delayed Feedback. Central to our approach is a novel estimator that integrates an Inverse Probability Weighting (IPW) mechanism directly into the loss function, ensuring unbiased correction for delayed or missing feedback. We provide comprehensive theoretical analysis, establishing an O(d*sqrt(T)) regret bound for the linear setting and sub-linear guarantees for the neural setting. Extensive experiments on both simulated and real-world datasets demonstrate the effectiveness of our propose.\n\n---\n\n## 89. A Hybrid Vision-Language Architecture for Automated Defect Reasoning and Report Generation in Industrial Inspection\n- **Authors**: Malikussaid , Imad Gohar\n- **URL**: [https://arxiv.org/abs/2605.26533](https://arxiv.org/abs/2605.26533)\n- **Abstract**:\n> Automated industrial inspection requires both precise defect localization and structured maintenance report generation; in current practice these tasks are handled separately, with linguistic interpretation left to human experts. This paper describes a decoupled, edge-deployable pipeline for wind turbine blade inspection built from three components that each handle a distinct sub-task. The Eyes a YOLO26-x-obb oriented bounding-box detector localizes defects at dataset-native resolution. The Bridge a deterministic, parameter-free encoding module maps each detected bounding box to grid-referenced spatial tokens embedded in a structured prompt. The Brain a 4-bit quantized Qwen-2.5-1.5B model adapted with Quantized Low-Rank Adaptation (QLoRA) on 947 synthetically generated maintenance reports generates a structured JSON report from that prompt. Retrieval-Augmented Fine-Tuning (RAFT) further grounds each recommendation in indexed maintenance procedures. Five ablation experiments, scored by BLEU-4, ROUGE-L, Hallucination Rate (HR), and an LLM-as-a-Judge rubric, compare the pipeline against a monolithic vision-language model (VLM) baseline and against partial configurations in which one component is removed. The complete system achieves BLEU-4 0.41, HR=4%, and Expert Score = 8.6/10 compared with 0.07, 65%, and 3.3/10 for the zero-shot VLM baseline. The QLoRA-adapted 1.5B model generates higher-quality reports than a 671B-parameter generalist API model given identical detection evidence, at 47 tokens per second on a single T4-class GPU. The results show that purpose-built decoupled architecture with a small domain-specific training corpus outperforms a generalist end-to-end model on this structured generation task.\n\n---\n\n## 90. InterSketch: An Interleaved Reasoning Model with Self-correcting Visual Sketch and Stepwise Reward\n- **Authors**: Zhiwei Ning , Wenwen Tong , Xiangli Kong , Shengnan Ma , Ziyi Shang , Jingcheng Ni , Tao Hu , Yong Xien Chng , Jixuan Ying , Zehuan Wu , Hanming Deng , Jie Yang , Yuanjie Zheng , Wei Liu , Lewei Lu\n- **URL**: [https://arxiv.org/abs/2605.26520](https://arxiv.org/abs/2605.26520)\n- **Abstract**:\n> While vision-language models (VLMs) have exhibited multi-turn visual reasoning capabilities, their reasoning trajectories remain relatively shallow and are dominated by a text-centric paradigm, limiting their applicability to complex visual challenges. In contrast, human-like thought typically involves long-horizon reasoning with an interleaved visual-textual chain-of-thought (VT-CoT). To bridge this gap, we introduce InterSketch, an interleaved reasoning model to enhance the VT-CoT capability via self-correcting and stepwise reward mechanisms. InterSketch dynamically generates intermediate visual sketches using external tools and interleaves them with textual reasoning, enabling effective perception and logical reasoning over long-horizon visual understanding tasks. Specifically, in the first cold-start stage, we propose a synthesized high-quality interleaved VT-CoT dataset and include a reflection mechanism to enable the model's capability in multi-turn interleaved reasoning and self-correction. In the subsequent reinforcement learning (RL) stage, we design a stepwise reward mechanism to mitigate the sparsity of reward signals inherent in end-only supervision over long-horizon reasoning. Extensive experiments on visual reasoning benchmarks demonstrate the effectiveness of InterSketch, even outperforming proprietary models such as Gemini-3-Pro.\n\n---\n\n## 91. Unveiling the Fragility of Vision-Language Models: Multi-Modal Adversarial Synergy via Texture-Constrained Perturbations and Cross-Modal Optimization\n- **Authors**: Xiang Fang , Wanlong Fang , Changshuo Wang\n- **URL**: [https://arxiv.org/abs/2605.26501](https://arxiv.org/abs/2605.26501)\n- **Abstract**:\n> Large Vision-Language Models (LVLMs) have transformed multi-modal understanding, excelling in tasks like image captioning and visual question answering by integrating visual and textual inputs. However, their robustness against adversarial attacks, particularly those exploiting both modalities, remains underexplored, posing risks to critical applications like autonomous driving and content moderation. Existing attacks focus on single modalities or require impractical white-box access, limiting their real-world relevance. In this paper, we introduce Multi-Modal Adversarial Synergy, a groundbreaking framework that crafts universal, black-box multi-modal attacks against LVLMs. MMAS simultaneously generates a texture scale-constrained universal adversarial perturbation for images and a learnable prompt perturbation for text, optimized jointly using only model queries. The image perturbation leverages wavelet-based texture constraints to ensure imperceptibility and robustness across diverse visual inputs. The text perturbation, constrained by an L-norm in the embedding space, maintains semantic coherence while steering outputs toward a target. A novel cross-modal regularization term aligns the perturbations' gradient directions, enhancing their synergistic impact and transferability across tasks and models. Extensive experiments show the strong universal adversarial capabilities of our proposed attack with prevalent LVLMs.\n\n---\n\n## 92. Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories\n- **Authors**: Sil Hamilton , David Mimno\n- **URL**: [https://arxiv.org/abs/2605.26492](https://arxiv.org/abs/2605.26492)\n- **Abstract**:\n> LLM-generated stories are a popular use case, but they show very low variability. We sample 20,000 total stories from four current models using five prompts. We find that 11 words occur in 88.3% of generated stories, with little difference between models. These words include names (Elias, Mara, Elara), settings (lighthouses), and professions (clockmaker, librarian). These tokens do not often occur in published literature nor pre-training data, but they are found in preference data that is likely to have been used by all current models. Surprisingly, these \"lighthouse\" stories are infrequent when compared with the average post-training story, much of which contains references to copyrighted characters or adult content. This result demonstrates the potentially disproportionate impact of small datasets combined with powerful alignment algorithms.\n\n---\n\n## 93. Towards Error-Free EHRs: Reasoning-Intensive Consistency Verification Between Clinical Notes and Structured Tables in Electronic Health Records\n- **Authors**: Yeonsu Kwon , Jiho Kim , Junseong Choi , Paloma Rabaey , Minseo Kim , Sujeong Im , Jeewon Yang , Jun-Min Lee , Sangji Lee , Jiwon Kim , Hangyul Yoon , Hyunwook Kwon , Edward Choi\n- **URL**: [https://arxiv.org/abs/2605.26463](https://arxiv.org/abs/2605.26463)\n- **Abstract**:\n> Data consistency between unstructured clinical notes and structured tables in Electronic Health Records (EHRs) is essential for patient safety and clinical decision-making. However, existing work on note-table consistency verification mainly relies on surface-level matching of numeric values or simple events. Such approaches fail to capture the reasoning underlying real-world EHR documentation, including clinical interpretation, event relations, and temporal changes. To address this gap, we introduce EHR-ReasonCon, a reasoning-intensive benchmark for note-table consistency verification. Built on MIMIC-III with expert-guided annotations, it comprises 8,048 entities derived from clinical notes and provides high-quality ground-truth labels. The annotation protocol is supported by specialized table-exploration tools to ensure systematic evidence retrieval and reliable consistency assessment. We also propose EHR-Inspector, an LLM-based framework that segments notes, extracts anchor entities and temporal references, and uses table-exploration tools to verify consistency against structured tables. Evaluated using expert-validated LLM-as-a-judge metrics under harsh and lenient criteria, EHR-Inspector achieves state-of-the-art performance across multiple model backbones. Analyses further demonstrate the effectiveness of its components and highlight differences from human verification.\n\n---\n\n## 94. Verus-SpecGym: An Agentic Environment for Evaluating Specification Autoformalization\n- **Authors**: Anmol Agarwal , Natalie Neamtu , Pranjal Aggarwal , Seungone Kim , Jannis Limperg , Cedric Flamant , Kanna Shimizu , Bryan Parno , Sean Welleck\n- **URL**: [https://arxiv.org/abs/2605.26457](https://arxiv.org/abs/2605.26457)\n- **Abstract**:\n> AI coding agents are increasingly used to write real-world software, but ensuring that their outputs are correct remains a fundamental challenge. Formal verification offers a promising path: an agent generates code together with a machine-checked proof, guaranteeing that the code satisfies a formal specification. However, there is no guarantee that the formal spec itself matches the user's intent. In this work, we study specification autoformalization: whether LLM agents can translate informal programming problems into faithful formal specifications. We introduce Verus-SpecBench, a benchmark of 581 spec-writing tasks derived from Codeforces problems targeting Verus, a verifier for Rust, and Verus-SpecGym, an agentic environment in which models interact with Verus, bash, & the filesystem to develop these specs. The central challenge is evaluation: expert-written reference specs are expensive to write, & LLM judges can miss subtle mistakes. We address this by (a) extending Verus's exec_spec mechanism so that generated specs can be executed as Rust code, & (b) testing them against official Codeforces tests & adversarial cases extracted from Codeforces \"hacks\", which are edge cases written by competitors to break incorrect solutions. On Verus-SpecBench, the strongest model, Gemini 3.1 Pro, solves 77.8% of tasks, other frontier models solve 51.1--57.8% & OSS models reach only 21.5--25.5%. Our analysis of failure modes shows that model-generated specs can omit important input assumptions, accept incorrect outputs, & reject valid ones. We also find that LLM-as-a-judge evaluation misses 26% of the failures our evaluator catches. Overall, our results suggest that spec autoformalization is within reach for frontier agents but remains brittle even on problems where they can already generate correct code. The code, data, & logs can be found at this https URL\n\n---\n\n## 95. Alignment Tuning for Large Language Models: A Data-Centric Lens on Alignment Data Pipelines\n- **Authors**: Hwanjun Song\n- **URL**: [https://arxiv.org/abs/2605.26442](https://arxiv.org/abs/2605.26442)\n- **Abstract**:\n> Much of the alignment tuning literature is organized around optimization objectives, while the construction of alignment data is often treated implicitly. In this survey, we adopt a data centric perspective and reframe alignment tuning as a pipeline design problem. We decompose alignment data construction into three interacting stages, response synthesis, preference evaluation, and preference instantiation, and use this framework to organize existing alignment methods into a unified taxonomy. Through this lens, we identify recurring design trade-offs and failure modes observed across prior alignment methods, and distill a set of high level principles that clarify how pipeline design choices influence the resulting optimization signal. Finally, we outline open challenges for alignment data pipelines, including prompt-level alignment, agentic settings, and alignment under evolving objectives.\n\n---\n\n## 96. LURE: Live-Usage Replay Evaluations for Reducing Evaluation Awareness\n- **Authors**: Igor Ivanov , David Demitri Africa\n- **URL**: [https://arxiv.org/abs/2605.26438](https://arxiv.org/abs/2605.26438)\n- **Abstract**:\n> Large language models can recognize when they are being evaluated (evaluation awareness) and behave differently because of that, which undermines the validity of safety and alignment benchmarks. We propose LURE (Live-Usage Replay Evaluations), a method for constructing deployment-like evaluations by replaying realistic agentic interaction trajectories and appending evaluation prompt at the end. We also introduce an automated pipeline for measuring evaluation realism, combining detection of verbalized evaluation awareness and judge-model estimates of the probability of logs being an evaluation, and validate it on a large dataset of deployment and evaluation transcripts. We find that LURE-based evaluations are substantially less distinguishable from deployment than widely used benchmarks and synthetic evaluation generators, and can approach the realism of real conversations with users. We instantiate LURE in scheming, AI safety sabotage, and sycophancy settings. Our results suggest that evaluation realism is a crucial property of alignment benchmarks and should be reported alongside benchmark results, especially when such results are used in safety cases.\n\n---\n\n## 97. Targeted Remasking: Replacing Token Editing with Token-to-Mask Refinement in Discrete Diffusion Language Models\n- **Authors**: Lin Yao\n- **URL**: [https://arxiv.org/abs/2605.26436](https://arxiv.org/abs/2605.26436)\n- **Abstract**:\n> Discrete masked diffusion language models such as LLaDA generate text through iterative denoising, where mask tokens are progressively replaced with predicted tokens. LLaDA2.1 introduced a Token-to-Token (T2T) editing mechanism that accelerates generation by directly replacing committed tokens suspected of being incorrect. However, we identify fundamental limitations of T2T editing: it couples error detection with replacement, pollutes the generation context with potentially incorrect tokens, and introduces a train-inference noise mismatch where systematic model-generated errors differ from the random perturbations seen during training. We propose Token-to-Mask (T2M) remasking, a training-free, drop-in replacement for T2T editing that resets suspected erroneous tokens back to the mask state, allowing the diffusion process to re-predict them under cleaner context. We design and empirically validate three complementary error detection strategies -- probability-based, trigger-mirrored, and temporal-difference-based -- and provide a unified theoretical analysis showing that T2M remasking purifies the generation context, converts systematic inference errors back to the model's native mask noise type, and enables delayed commitment for joint multi-position optimization. Comprehensive experiments across 12 benchmarks spanning knowledge, reasoning, mathematics, coding, and instruction following show that T2M generally improves performance on tasks requiring precise token-level output, with the largest gain on mathematics (+5.92% on CMATH). Error analysis on CMATH reveals that the dominant failure mode is last-mile token corruption -- where correct reasoning produces a corrupted final answer -- and that T2M repairs 59.4% of such cases.\n\n---\n\n## 98. The Rescue Effect: Spatio-Semantic Early Exit Bypasses Quantization Collapse in CLIP\n- **Authors**: Kahyeon Nam , Hyesong Choi\n- **URL**: [https://arxiv.org/abs/2605.26415](https://arxiv.org/abs/2605.26415)\n- **Abstract**:\n> Deploying Vision-Language Models on resource-constrained hardware typically requires INT8 quantization, but in joint-embedding architectures such as CLIP this introduces a failure mode distinct from quantized CNN classifiers: activation noise accumulated across transformer blocks perturbs the direction of the multimodal embedding, eroding the cosine alignment on which zero-shot retrieval depends. We characterize this as Quantization-Induced Representation Collapse (QIRC) and quantify it on INT8 CLIP ViT-B/32, where the layer-wise noise-to-signal ratio grows from below 10% in shallow blocks to 52% at Layer 11. We propose LRA-EE (Layer-wise Representation-Aware Early Exit), which bypasses noise-saturated deep layers via Spatio-Semantic Aggregation (replacing the immature shallow [CLS] with a global patch-token average), a learned multi-feature gate (confidence, top-2 margin, spatial-activation variance), and Layer-adaptive Confidence Thresholding calibrated to each layer's Information-to-Noise Ratio. On ImageNet-1K zero-shot classification, LRA-EE reduces FLOPs by 13.4% and improves Top-1 accuracy by +2.44%p (58.72% -> 61.16%) over the INT8 baseline. A four-quadrant decomposition isolates the Rescue Effect: 9.5% of samples are correctly classified at shallow exits but lost to noise at full depth, against only 7.1% suffering the inverse.\n\n---\n\n## 99. Plans for Evaluating Structured Generative Search Summaries\n- **Authors**: Tetsuya Sakai , Jina Lee , Hanpei Fang , Young-In Song\n- **URL**: [https://arxiv.org/abs/2605.26400](https://arxiv.org/abs/2605.26400)\n- **Abstract**:\n> We propose a framework for evaluating structured generative search summaries that are placed atop organic web search results. A structured summary, generated by a large language model, typically consists of an overview, several sections with section titles, and a list of source documents that are cited within the summary. We then describe our plans for implementing and evaluating the framework.\n\n---\n\n## 100. Annotator Positionality as Signal: Psychometric Weighting for Anti-Autistic Ableism Detection\n- **Authors**: Naba Rizvi , Harper Strickland , Saleha Ahmedi , Nedjma Ousidhoum\n- **URL**: [https://arxiv.org/abs/2605.26397](https://arxiv.org/abs/2605.26397)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used in decision-making tasks where they can amplify or suppress perspectives, raising concerns in high-stakes settings affecting autistic communities. While previous research has identified disability-related biases in LLMs, it remains unclear how they conceptualize ableism or detect it in text. We introduce a bias-aware evaluation framework targeting anti-autistic ableist language with a psychometrically-weighted, community-proximate ground truth anchored in annotator positionality. This framework constitutes a stricter standard than conventional majority-vote aggregation which significantly and consistently underweights autistic and autism-accepting perspectives. We find that LLMs frequently produce harmful outputs, mislabel community-reclaimed language as ableist, and express more negative attitudes toward autistic people when assessment instruments are masked. Our error analysis reveals that models rely on surface-level keyword matching rather than contextual factors such as speaker identity, and whether the language fosters in-group solidarity or inflicts out-group harm.\n\n---\n\n## 101. VisualNeedle: Benchmarking Active Visual Search in Information-Dense Scenes\n- **Authors**: Jingru Chen , Yiming Liu , Mingtao Chen , Sijie Chen , Richeng Xuan , Liang Yang , Zhichao Hu , Fanyang Lu\n- **URL**: [https://arxiv.org/abs/2605.26380](https://arxiv.org/abs/2605.26380)\n- **Abstract**:\n> Frontier multimodal large language models (MLLMs) have been reported to achieve over 90% accuracy on fine-grained perception benchmarks. However, such scores do not necessarily imply faithful use of visual evidence. Prior studies have identified three shortcuts that inflate benchmark performance. First, linguistic priors and lexical cues in questions often enable models to infer plausible answers without seeing the image. Second, coarse global semantics from the visual encoder can bypass fine-grained local details. Third, in some ``think-with-images'' benchmarks, corrupting the intermediate images returned by visual tools barely affects the final answer. These findings suggest that higher input resolution or larger question pools alone do not elicit genuine active visual search. To address this, we introduce VisualNeedle, a challenging, information-dense, and fine-grained benchmark for scenes where critical evidence is spatially constrained to minute regions and not discernible at a glance. We further propose a counterfactual crop-black setting, which replaces crops returned by tools with black images of the same size, to test whether tool-enabled performance truly relies on intermediate visual evidence. We evaluate 9 promninent MLLMs across three settings: no-tool, standard tool-enabled, and crop-black. No-tool accuracy stays below 20\\%, and the best tool-enabled model reaches only 56.01\\%, still trailing the 63.00% human majority-vote accuracy. These results reveal persistent limitations in fine-grained visual search, while the crop-black ablation confirms that success on VisualNeedle hinges on genuine intermediate visual evidence.\n\n---\n\n## 102. BioFact-MoE: Biologically Factorized Mixture of Experts for Vision-Language Prognostic Modeling in Hepatocellular Carcinoma\n- **Authors**: Junlin Yang , Tian Yu , Nicha C. Dvornek , Yuexi Du , Peiyu Duan , Annabella Shewarega , Lawrence H. Staib , James S. Duncan , Julius Chapiro\n- **URL**: [https://arxiv.org/abs/2605.26376](https://arxiv.org/abs/2605.26376)\n- **Abstract**:\n> Hepatocellular carcinoma (HCC) is biologically heterogeneous, shaped by the interplay between hepatic functional reserve and tumor-related oncologic factors; thus, similar survival outcomes may reflect fundamentally different underlying biological processes. Prognostic modeling in HCC is informed by rich multimodal information from multiparametric MRI and radiology reports from routine clinical practice. Existing prognostic vision-language models (VLMs) learn a single entangled latent representation that blends hepatic and tumor-related factors, limiting both accuracy and biological interpretability. We present BioFact-MoE, a biologically factorized Mixture of Experts (MoE) framework that explicitly decomposes liver and tumor factors via biologically supervised experts within a residual MoE survival architecture. On a HCC cohort of N=588 patients (pretrained on 4,582 3D MRI image-report pairs), BioFact-MoE consistently improves survival prediction over all baselines across time horizons, achieving 12-, 18-, and 24-month AUCs of 75.33%, 75.85%, and 73.96%. Beyond scalar risk prediction, gated expert weights enable phenotype-aware risk stratification. Pathway-informed gating uncovers clinically meaningful treatment-associated survival heterogeneity. In held-out validation, hepatic and tumor embeddings show selective associations with liver function and tumor burden markers, respectively (p<0.05), without supervision. The code is available at this https URL .\n\n---\n\n## 103. Why LLMs Hallucinate on Structured Knowledge: A Mechanistic Analysis of Reasoning over Linearized Representations\n- **Authors**: Shanghao Li , Jinda Han , Yibo Wang , Yuanjie Zhu , Zihe Song , Langzhou He , Kenan Kamel A Alghythee , Philip S. Yu\n- **URL**: [https://arxiv.org/abs/2605.26362](https://arxiv.org/abs/2605.26362)\n- **Abstract**:\n> In many reasoning tasks, large language models (LLMs) rely on structured external knowledge, such as graphs and tables, which is typically linearized into sequential token representations. However, even when sufficient knowledge is available, LLMs can still produce hallucinated outputs, and the underlying mechanisms behind such failures remain poorly understood. We investigate these mechanisms and find that hallucinations arise from systematic internal dynamics rather than random noise. First, attention disproportionately concentrates toward shortcut-like structural cues rather than distributing across the full context. Second, feed-forward representations fail to ground the provided knowledge, causing the model to revert to parametric memory. Moreover, our results indicate that hallucination is consistently associated with failures in semantic grounding within feed-forward layers, while attention allocation exhibits greater task-dependent variability. Finally, we show that these mechanistic patterns generalize beyond single-hop graphs to multi-hop and tabular settings, enabling effective hallucination detection across structured knowledge formats.\n\n---\n\n## 104. Erased but Exploitable: Black-box Embedding-Aware Prompting Against Unlearned Text-to-Image Diffusion Models\n- **Authors**: Arian Komaei Koma , Seyed Amir Kasaei , AmirMahdi Sadeghzadeh , Mohammad Hossein Rohban\n- **URL**: [https://arxiv.org/abs/2605.26332](https://arxiv.org/abs/2605.26332)\n- **Abstract**:\n> Machine unlearning aims to remove specific concepts from pretrained text-to-image diffusion models, yet several white- and black-box attacks have been introduced to make the model generate such unlearned concepts. These attacks, nevertheless, do not assume a realistic threat model, i.e. they either assume access to the model weights, or result in gibberish adversarial prompts that could be easily detected even through naive rule-based safeguarding. We aim to address this gap in this paper. We introduce BEAP, a black-box, embedding-aware adversarial prompting attack that leverages a large language model (LLM) to iteratively generate effective adversarial prompts and exploit such hidden vulnerabilities. BEAP performs an embedding-aware search in text space, combining multiple reward signals: unlearned concept presence, text-image alignment, and image quality, to refine generated prompts. Unlike previous attack methods, BEAP keeps its prompts undetectable to safety filters while producing high-quality images. Extensive experiments show that BEAP improves the Attack Success Rate (ASR) by more than 60% over prior methods, while requiring only an average of fifteen prompts per successful attack. Warning: This paper contains model outputs that may be offensive or upsetting in nature.\n\n---\n\n## 105. Curriculum Learning for Safety Alignment\n- **Authors**: Sandeep Kumar , Virginia Smith , Chhavi Yadav\n- **URL**: [https://arxiv.org/abs/2605.26315](https://arxiv.org/abs/2605.26315)\n- **Abstract**:\n> Direct Preference Optimisation (DPO) is widely used for safety alignment in large language models. However, prior work shows it is brittle and exhibits poor out-of-distribution (OOD) generalisation. In this paper, we investigate whether Curriculum Learning can improve the robustness of DPO-based safety alignment. We propose Staged-Competence, a curriculum-based framework that organises preference data by difficulty, employs competence-based sampling, and progressively updates the reference model during training. Averaged across three model families, Staged-Competence reduces OOD harmful response rates by 16% and jailbreak attack success rates by 20%, while preserving general capabilities with near-zero over-refusal. We further show that Staged-Competence (1) matches baseline safety with only 75% of the training data and (2) yields better separation between safe and unsafe responses. Staged-Competence is agnostic to the policy optimisation loss and can extend to other DPO variants and alignment domains. Our code and data are available at this https URL .\n\n---\n\n## 106. Intelligent Detection and Mitigation of Carpet-Bombing DDoS Attacks in SDN Using Retrieval-Augmented Generation and Large Language Models\n- **Authors**: Mohammed N. Swileh , Shengli Zhang , Kai Lei\n- **URL**: [https://arxiv.org/abs/2605.26307](https://arxiv.org/abs/2605.26307)\n- **Abstract**:\n> Software-Defined Networking (SDN) provides flexible and programmable network management; however, its centralized control architecture remains highly vulnerable to Distributed Denial-of-Service (DDoS) attacks, particularly Carpet-Bombing DDoS attacks that distribute malicious traffic across multiple targets to evade conventional detection mechanisms. In this paper, a Retrieval-Augmented Generation (RAG)-based framework is proposed for real-time detection and mitigation of Carpet-Bombing DDoS attacks in SDN environments. The proposed framework combines interface-level traffic features representation, semantic embedding generation, FAISS-based similarity retrieval, and Large Language Model (LLM)-driven contextual inference to classify traffic behavior without requiring conventional supervised model training or retraining. To evaluate the effectiveness of the proposed framework, extensive experiments were conducted under multiple Carpet-Bombing DDoS attack scenarios with different attack intensities. In addition, two traffic representation strategies, namely structured JSON-based representation and natural language-based representation (NLR), were investigated using multiple state-of-the-art LLMs. The experimental results demonstrate that the proposed framework achieved highly accurate and stable attack detection performance, while the framework configuration utilizing the Gemma-4-31B-IT model achieved the strongest overall detection results. Furthermore, real-time experiments confirmed the capability of the proposed framework to rapidly detect and mitigate Carpet-Bombing DDoS attacks while maintaining stable SDN network operation. The obtained results highlight the effectiveness of integrating RAG mechanisms with LLM for intelligent and adaptive SDN security analysis.\n\n---\n\n## 107. CroCo: Cross-Lingual Contrastive Preference Tuning on Self-Generations\n- **Authors**: Mike Zhang , Ali Basirat , Desmond Elliott\n- **URL**: [https://arxiv.org/abs/2605.26293](https://arxiv.org/abs/2605.26293)\n- **Abstract**:\n> Prior work establishes that controlled contrastiveness between self-generated responses from large language models, set via reward scores, improves downstream preference tuning in English. We extend this method to multiple languages and evaluate two models across a total of 14 high and low-resource languages on a diverse set of tasks. Our central finding is that cross-lingual contrastive preference tuning on self-generations (CroCo) transfers without language-specific preference annotation. A reward model trained on English preferences (atop a multilingual base) produces useful within-language rankings across most languages, and pairing in either a monolingual or multilingual setting improves over each model on the majority of setups while preventing the catastrophic forgetting of supervised fine-tuning. We observe that the gains require on-policy data. Off-policy responses reduce the benefit and online preference optimization fails to improve over the offline variant. Specifically, on structured tasks, our method matches or exceeds the base in 6/7 languages for EuroLLM-9B and 4/7 settings for Aya-3B. On open-ended generation, both tuned models win against their respective base across 11 evaluated languages. Overall, we show promising directions for multilingual preference tuning.\n\n---\n\n## 108. AgentSociety: Incentivizing Agentic Social Intelligence\n- **Authors**: Aditya Vema Reddy Kesari , Krishna Reddy Kesari\n- **URL**: [https://arxiv.org/abs/2605.26203](https://arxiv.org/abs/2605.26203)\n- **Abstract**:\n> The success of deployed agents relies on their ability to handle open-ended user requests using their inherent capabilities, not only in solving requests directly but also in effectively leveraging inter-agent communication channels and feedback signals over time. This requires a multi-agent environment where agents can operate autonomously, strategically communicate, behave collaboratively and be driven by economic incentives, much like humans in society. Towards this vision, we propose $\\mathtt{AgentSociety}$, a mechanism that enables decentralized agentic collaboration grounded in liquid democracy and information diffusion from social choice theory. We show that $\\mathtt{AgentSociety}$ provides an environment for agents to make autonomous decisions utilizing their local context to maximize their utility while achieving collective outcomes through incentivized collaboration. Specifically, we prove that delegation to more competent neighbor agents is incentive compatible and naturally generates multi-agent routing path by consensus. Additionally, our mechanism incentivizes agents to selectively disclose information to their neighbor agents when doing so aligns with their self-interest, so as to garner influence. We characterize the Nash equilibrium showing that agent payoffs are reflective of their marginal contributions. We compare and benchmark strategy profiles adopted by open and proprietary state-of-the-art language models deployed in $\\mathtt{AgentSociety}$ against best response. Finally, we evaluate collaborative performance from consensus-based routing among self-interested heterogeneous agents in $\\mathtt{AgentSociety}$ on real-world datasets.\n\n---\n\n## 109. CyberEvolver: Structured Self-Evolution for Cybersecurity Agents On the Fly\n- **Authors**: Yihe Fan , Changyi Li , Lichen Xu , Xudong Pan , Jiarun Dai , Hong Geng , Min Yang\n- **URL**: [https://arxiv.org/abs/2605.26195](https://arxiv.org/abs/2605.26195)\n- **Abstract**:\n> LLM-based agents are increasingly used for cybersecurity tasks, but most existing systems rely on fixed, human-designed scaffolds that struggle to adapt across diverse targets and failure modes. We introduce \\textsc{CyberEvolver}, a self-evolving cybersecurity agent framework that iteratively revises its own scaffold based on experience from failed execution attempts. Self-evolution in cybersecurity is challenging because the space of possible scaffold changes is largely unstructured, execution feedback is sparse and often obscured by the environment, and low-diversity updates can cause errors to compound over repeated iterations. \\textsc{CyberEvolver} addresses these challenges with a four-layer evolvable agent architecture that decomposes scaffold optimization into structured components, a trace-to-diagnosis mechanism that converts noisy execution logs into actionable revision signals, and a population-based beam search strategy that preserves diverse agent variants during evolution. We evaluate \\textsc{CyberEvolver} on CTF challenges, vulnerability exploitation, and penetration-testing tasks using four open-source LLMs. Across these settings, \\textsc{CyberEvolver} improves the seed agent's success rate by $13.6$\\,\\% on average, and outperforms six human-designed cybersecurity agents as well as two self-improvement methods adapted from other domains. These results suggest that scaffold self-evolution is a promising direction for building adaptive LLM agents for security testing.\n\n---\n\n## 110. Max-Window Scale Estimation for Near-Lossless HiF8 W8A8 Quantization-Aware Training\n- **Authors**: Yingying Cheng , Jinquan Shi , Li Zhou , Zhiyang He , Zhaoyi Sun , Fan Zhang , Jie Sun\n- **URL**: [https://arxiv.org/abs/2605.26189](https://arxiv.org/abs/2605.26189)\n- **Abstract**:\n> Quantization-aware training (QAT) with low-bit floating-point formats enables efficient LLM deployment, yet introduces subtle failure modes invisible to standard training metrics. We present a systematic study of HiF8 W8A8 QAT for OpenPangu-Embedded-1B through the lens of Delayed Tensor Scaling (DTS). Across eight controlled experiments, we identify and disentangle two orthogonal failure modes: (i)amax saturation, where delayed scale estimates silently corrupt knowledge-sensitive representations via forward-pass clipping, and (ii)catastrophic forgetting, where an aggressive learning rate overwrites pretrained commonsense knowledge independently of quantization. Neither is detectable from training loss alone. We address amax saturation with a conservative max-algorithm DTS strategy over a 64-step history window, and mitigate forgetting via a 500-step BF16 warmup followed by QAT at lr=10^{-5}. Both fixes are necessary and sufficient: our final configuration achieves 0.43% MMLU drop, 0.58% HellaSwag drop, and 0.22% ARC-Challenge drop versus a matched BF16 baseline, with a training loss APE of only 0.11% over 10,000 steps.\n\n---\n\n## 111. SetupX: Can LLM Agents Learn from Past Failures in Functionality-Correct Code Repository Setup?\n- **Authors**: Zihang Zhou , Ziqian Ren , Yukai Wu , Yingjie Xiong , Wei Zhou , Chao Peng , Dong Zhang , Bingheng Yan , Xuanhe Zhou , Fan Wu\n- **URL**: [https://arxiv.org/abs/2605.26186](https://arxiv.org/abs/2605.26186)\n- **Abstract**:\n> Functionality-correct repository setup aims to configure execution environments (e.g., dependencies, build scripts) to successfully execute a repository's documented features. It presents significant challenges due to diverse, repository-specific failures, including dependency incompatibilities, missing toolchains, incomplete installations, and verification-strategy mismatches. Existing LLM agents struggle to robustly resolve these issues, specifically failing to support (1) cross-repository experience transfer, (2) multi-step trial-and-repair under non-invertible state changes, and (3) robust verification of setup outcomes to distinguish setup-induced failures from repository bugs. To address this, we introduce SetupX, an experiential learning-based setup framework. First, we construct a Self-Evolving Experience Representation (XPU), a dual-modality knowledge unit encoding setup signals, textual guidance, executable actions to dynamically transfer verified environment fixes to unseen repositories. Second, we employ Experience-Augmented Speculative Execution backed by a LIFO Docker snapshot stack, enabling the agent to proactively trial fixes and safely roll back to known-good states. Third, we introduce a Prosecutor-Judge Verification Protocol that separates evidence collection from final judgment, enabling more reliable setup verification beyond superficial build-time metrics. Evaluation results on carefully-crafted benchmarks show SetupX achieves highest performance (e.g., 92% pass rate) and outperforms the strongest baseline by over 19%. Crucially, SetupX excels in complex multi-repository setup requiring coordinating multiple interconnected services across different containers. The code repository is available at this https URL .\n\n---\n\n## 112. AutoDFT: A Closed-Loop Multi-Agent Framework for Autonomous DFT Calculations\n- **Authors**: Penghui Yang , Zhonghan Zhang , Yue Li , Xinrun Wag , Yanchen Deng , Yuhao Lu , Bijun Tang , Zheng Liu , Bo An\n- **URL**: [https://arxiv.org/abs/2605.26179](https://arxiv.org/abs/2605.26179)\n- **Abstract**:\n> Density functional theory (DFT) serves as the basis for computational discovery in materials science and chemistry, yet each calculation demands extensive human effort: adjusting algorithms when convergence stalls, revising plans when unexpected physics emerges, and inserting steps as intermediate results reshape the problem. Existing LLM-based agents automate only the initial planning stage, producing a full execution plan upfront and leaving all subsequent adaptation to hand-crafted rules. As a result, these workflows remain fragile, do not generalize well beyond pre-planned scenarios, and often require expert intervention when failures or unexpected intermediate results require changes to the calculation path. Here, we introduce AutoDFT, a closed-loop multi-agent framework that embeds LLM reasoning into every stage of the DFT lifecycle, where a strategic planner produces a skeletal plan of step objectives; a step planner generates numerical parameters just in time from preceding results; and a monitor-recover-reflect cycle diagnoses failures, repairs them, and revises the plan when the evidence justifies it. We demonstrate both breadth and depth: breadth on VASPBench, a purpose-built benchmark spanning 34 tasks and 9 DFT calculation types, where AutoDFT achieves 94.1% task-level success with GPT-5.2; and depth on established materials databases, where AutoDFT produces quantitatively reliable property predictions across electronic, magnetic, and energetic properties. By closing the loop between planning and execution, AutoDFT enables experimentalists without deep computational expertise to obtain reliable first-principles results.\n\n---\n\n## 113. PitchBench: Measuring Pitch Hearing in Audio-Language Models\n- **Authors**: Milan Liessens Dujardin , Song-Ze Yu , Craver Corbyn Thomas-Smith , David M. Chan , Karina Nguyen\n- **URL**: [https://arxiv.org/abs/2605.26176](https://arxiv.org/abs/2605.26176)\n- **Abstract**:\n> Audio-language models (ALMs) are increasingly used in real-world applications that require understanding music, from music tutoring and transcription to captioning, recommendation systems, and music production. More broadly, they are becoming an important component of multimodal AI systems that must reason from sensory input rather than text alone. This makes reliable musical perception a critical prerequisite: if a model cannot accurately hear the structure of sound, it cannot be trusted to reason about, teach, transcribe, or act on audio in the real world. Yet existing benchmarks rarely assess one of the most fundamental musical abilities underlying such perception: pitch hearing. Current evaluations tend to probe pitch hearing only indirectly, through higher-level tasks and often in multiple-choice formats, leaving open how reliably ALMs identify fine-grained pitch across instruments, acoustic conditions, and response formats. We introduce PitchBench, an evaluation suite that systematically measures pitch hearing in ALMs. PitchBench comprises 28 experiments spanning absolute and relative pitch perception within sequences and chords, while varying loudness, note duration, sound source, time stretching, background noise, and other acoustic conditions. Tasks range from identifying individual pitches in isolation to tracking a melodic line within a four-part musical texture. Evaluating frontier ALMs, we find that pitch hearing remains highly unreliable: models perform consistently poorly across settings, with accuracy varying sharply by sound source, note duration, and notation format. Current ALMs do not yet possess stable pitch perception, even for controlled synthetic and instrumental stimuli. Alongside the benchmark, we release PitchBench as a Python package containing the evaluation data and data generation tools to support future work on pitch-aware audio-language modeling.\n\n---\n\n## 114. InfoQuant: Shaping Activation Distributions for Low-Bit LLM Quantization\n- **Authors**: Ke Li , Dong An , Xiaoling Zang , Can Ye , Liang Xie , Qibo Qiu , Chen Shen , Xiaofei He , Wenxiao Wang\n- **URL**: [https://arxiv.org/abs/2605.26175](https://arxiv.org/abs/2605.26175)\n- **Abstract**:\n> Low-bit activation quantization remains a major bottleneck in efficient large language model (LLM) deployment. The difficulty is not only that activations contain outliers, but that their distributions are often poorly matched to a low-bit uniform quantizer. Existing post-training quantization (PTQ) methods suppress peaks, balance channels, or minimize reconstruction error, yet they rarely specify what activation distribution is actually easy to discretize. As a result, activations may appear numerically smoother while still incurring large quantization error because the quantization range remains wide or most values collapse into a few levels near the mean. We recast activation transformation as quantizer-facing distribution design and analyze quantization error from an information-theoretic perspective. Our analysis shows that quantization-friendly activations should jointly have a smaller numerical range and sufficient dispersion within that range. Guided by this analysis, we propose InfoQuant, a train-free method that employs Peak Suppression Orthogonal Transformation (PSOT) to shape activations into more quantization-friendly distributions. We further introduce adaptive outlier-token selection to improve the robustness of PSOT during optimization. Across multiple LLM families, InfoQuant consistently outperforms prior PTQ and end-to-end training baselines. Under W4A4KV4, it preserves 97% of floating-point accuracy on average and reduces the LLaMA-2 13B performance gap by 42% over the previous state of the art. Code is available at [ this https URL ]( this https URL )\n\n---\n\n## 115. A Universal Cliff and a Design Fingerprint: Cross-Section Defect Detection Under LLM Orchestration\n- **Authors**: Hiroki Fukui\n- **URL**: [https://arxiv.org/abs/2605.26174](https://arxiv.org/abs/2605.26174)\n- **Abstract**:\n> Production language-model systems answer a request by partitioning it across an invisible orchestration of worker agents that recompose one integrated report. We ask what this does to a class of defect no single worker can see: a contradiction in the relation between two distant sections of a document. Holding the documents, defects, mechanism, scoring, and seed fixed, we vary only the model -- ten systems across five generations from one developer and five providers from distinct alignment paradigms. Two layers separate. First, a universal detection cliff: every model that finds these cross-section defects under a single agent loses that ability under orchestration, detection falling two-thirds or more across every paradigm tested. The cliff is mechanism-derived and not closed by scale or extended reasoning. Second, how models behave once fallen. A signal-detection decomposition shows that, among the six models discriminating above chance, only one developer's generations move along the reporting-criterion axis: as alignment is strengthened, the model misses fewer defects yet raises more false alarms on clean documents -- two faces of one criterion shift, scaling with generation within that developer (p < 0.001) and near-absent elsewhere. At the floor the missed defect is often not out of view: the model's private record reconstructs the structural fault accurately, while the integrated report signs off on its soundness, its concern spent on the artifact and an absent collaborator. This resists quantification -- an automated judge is unstable (precision 17-50%) and keywords cannot separate it from ordinary agreement -- a resistance we report as a finding. We release all runs, probes, defect keys, scorer prompts, and scripts. An integrated report's confidence is uninformative about partition-spanning defects, the most aligned systems are not the safest, and the cliff is structural.\n\n---\n\n## 116. Tool-Schema Compression Enables Agentic RAG Under Constrained Context Budgets\n- **Authors**: Furkan Sakizli\n- **URL**: [https://arxiv.org/abs/2605.26165](https://arxiv.org/abs/2605.26165)\n- **Abstract**:\n> Agentic RAG systems that equip language models with dozens to hundreds of tool definitions face a critical resource conflict: tool schemas consume the same context window needed for retrieval-augmented generation. We present the first systematic study of this tool-context trade-off, evaluating 14 models spanning 1.5B-32B local models plus one frontier API model across 6,566 controlled API calls at three context budgets (8K, 16K, 32K) with 28 tool definitions. Applying TSCG conservative-profile compression (44-50% schema token savings), we observe a binary enablement effect: at 8K tokens, JSON-schema tool definitions overflow the context window entirely, yielding near-zero EM (2.6% average), while compressed schemas restore RAG functionality with +20.5 pp average exact-match lift across all eight models (+24.7 pp among the six exhibiting full enablement). At 32K -- where both formats fit -- four of five tested models show delta <= 1 pp, confirming the effect is purely budget-driven. External validation on HotpotQA (50 multi-hop questions) shows +48 pp EM under the same overflow scenario. Frontier scaling tests demonstrate that JSON schemas overflow at ~494 tools while compressed schemas remain operational beyond 800 tools. Our results establish tool-schema compression as a necessary infrastructure layer for agentic RAG in constrained-context deployments. All code, data, and checkpoints are publicly available.\n\n---\n\n## 117. TSFMAudit: Data Contamination Auditing in Forecasting Time Series Foundation Models\n- **Authors**: Hongkai Li , Shifeng Xie , Lefei Shen , Zhuo Li , Mouxiang Chen , Xiaobin Zhang , Han Fu , Jianling Sun , Xiaoxue Ren , Chenghao Liu\n- **URL**: [https://arxiv.org/abs/2605.26161](https://arxiv.org/abs/2605.26161)\n- **Abstract**:\n> Time series foundation models (TSFMs) are increasingly pretrained on large corpora, raising concerns that evaluation datasets may have been exposed during pretraining and thus yield overly optimistic performance estimates. Auditing such contamination is challenging in time series because signals are continuous and heterogeneous, and often lack corpus documentation. To the best of our knowledge, this is the first work to study pretraining contamination auditing for TSFMs. We formalize the problem of pretraining contamination auditing for TSFMs and propose TSFMAudit, a method based on probe adaptation dynamics. Our key intuition is that contamination manifests as unusually efficient adaptation: after a fine tuning probe, contaminated datasets tend to exhibit faster loss reduction with smaller backbone movement. We evaluate TSFMAudit on 6 TSFMs and 187 datasets using documented training source evidence as supervision, and compare against 10 competitive baselines adapted from the LLM literature.\n\n---\n\n## 118. Furina: Fragmented Uncertainty-Driven Refusal Instability Attack\n- **Authors**: Tongxi Wu , Jian Zhang , Yang Gao\n- **URL**: [https://arxiv.org/abs/2605.26158](https://arxiv.org/abs/2605.26158)\n- **Abstract**:\n> Safety alignment in large language models (LLMs) and multimodal large language models (MLLMs) is commonly assumed to operate as a near-binary threshold mechanism. We challenge this assumption by revealing that safety behavior is governed by an instability region where small perturbations induce stochastic refusal decisions rather than deterministic outcomes. We develop a multi-metric diagnostic framework combining external and internal signals to characterize this instability. Through systematic experiments, we identify a characteristic diagnostic signature: inputs in unstable regimes exhibit elevated output uncertainty yet decreased internal safety activation, a decoupling phenomenon that explains why detection-based defenses fail against sophisticated attacks. Building on this framework, we introduce Furina, a jailbreak attack that deliberately induces this signature through fragmented, scene-anchored prompts without model-specific optimization. Furina outperforms strong single-turn and multi-turn baselines on HarmBench and achieves competitive results on MM-SafetyBench, demonstrating that uncertainty amplification provides a principled and transferable mechanism for understanding safety vulnerabilities. Code is available at: this https URL .\n\n---\n\n## 119. Turning Bias into Bugs: Bandit-Guided Style Manipulation Attacks on LLM Judges\n- **Authors**: Xianglin Yang , Bryan Hooi , Gelei Deng , Tianwei Zhang , Jin Song Dong\n- **URL**: [https://arxiv.org/abs/2605.26156](https://arxiv.org/abs/2605.26156)\n- **Abstract**:\n> The known stylistic biases in LLM judges, such as a preference for verbosity or specific sentence structures, present an underexplored security vulnerability. In this work, we introduce BITE (BIas exploraTion and Exploitation), a black-box adversarial framework that learns semantics-preserving edits to mislead an LLM judge and artificially inflate the scores it assigns. We cast the selection of stylistic edits as a contextual bandit problem and use a LinUCB policy to adaptively choose edits that maximize the judge's score without access to model parameters or gradients. Empirically, we test BITE across a diverse range of LLM judges and tasks, including both pointwise and pairwise comparisons on chatbot leaderboards and AI-reviewer benchmarks. BITE achieves an attack success rate exceeding 65% and raises scores by 1-2 points on a 9-point scale, all while preserving semantic equivalence. We further assess the attack's stealthiness, showing that BITE evades standard style-control methods and several detection baselines. Our findings expose a fundamental weakness in the LLM-as-a-judge paradigm and motivate robust, attack-aware evaluation. Our code is available at this https URL .\n\n---\n\n## 120. MemMorph: Tool Hijacking in LLM Agents via Memory Poisoning\n- **Authors**: Xuanye Zhang , Yongsen Zheng , Zhuqin Xu , Kaiyu Zhou , Bowen Shen , Haoran Ou , Tianwei Zhang , Kwok-Yan Lam\n- **URL**: [https://arxiv.org/abs/2605.26154](https://arxiv.org/abs/2605.26154)\n- **Abstract**:\n> LLM-driven agents are capable of selecting external tools to complete users' tasks. However, attackers could compromise such process, steering agents toward inappropriate/wrong tools and enabling malicious actions. Most existing attacks primarily manipulate the tool metadata, which is easily detectable by auditing and may lose effectiveness as modern agents increasingly adopt memory modules to refine tool selection policies through accumulated experience. This paper proposes MemMorph, the first attack that bias tool selection by poisoning the agent's long-term memory. Rather than explicitly dictating the tool invocation decision, MemMorph injects a small number of crafted records that are disguised as technical facts, incident reports, and operational policies. These poisoned records reshape the agent's contextual perception and decision-making process, leading it to autonomously infer and select the tool preferred by the attacker. Experiments across 3 benchmarks, 10 agent backbones, and 3 memory-module implementations show that MemMorph achieves up to 85.9% attack success rate with only three injected records, outperforming the strongest baseline by up to 25% while retaining potency under 3 representative defenses. Our findings expose long-term memory as a critical and under-explored attack surface in tool-augmented agents, urging the development of memory-level integrity safeguards.\n\n---\n\n## 121. VISTA: An End-to-End Benchmark for Visual Spec-to-Web-App Coding Agents\n- **Authors**: JunJia Guo , Yuhang Yao , Jiawei (Joe) Zhou , Jingdi Chen\n- **URL**: [https://arxiv.org/abs/2605.26144](https://arxiv.org/abs/2605.26144)\n- **Abstract**:\n> We present VISTA (VIsual Spec-To-App Benchmark), a benchmark for evaluating the end-to-end web-app generation capabilities of LLM-based agents. Unlike prior code generation benchmarks that focus on algorithmic tasks, VISTA targets realistic UI-centric development, where agents must produce functional, visually coherent applications from underspecified inputs. We define five prompt-information conditions that vary along two axes, visual/structural fidelity and stack constraint: (1) text only with free stack choice, (2) text with reference screenshots under three specified stacks, (3) text with reference screenshots under free stack choice, (4) text with screenshots and pruned Figma structure under a single specified stack, and (5) text with screenshots and pruned Figma structure under free stack choice. To enable robust evaluation, each page in the benchmark is manually annotated with interactive UI components and around three visual anchor points, addressing the well-known limitations of script-based testing tools such as Playwright in open-ended code generation settings. Evaluation combines DOM-grounded reference matching, behavior-specific browser tests, and CLIP-based visual similarity, jointly measuring structural alignment, behavioral completeness, and overall visual fidelity. We use VISTA to assess four agent systems drawn from two model families and two harnesses, finding that visual fidelity and functional correctness are partially decoupled across both input conditions and agents, and that agent editing style varies sharply but is largely orthogonal to task quality. VISTA establishes a rigorous and reproducible foundation for advancing agent-based software engineering research.\n\n---\n\n## 122. Eroding Trust in Real Speech: A Large-Scale Study of Human Audio Deepfake Perception\n- **Authors**: Nicolas M. Müller , Wei Herng Choong\n- **URL**: [https://arxiv.org/abs/2605.26136](https://arxiv.org/abs/2605.26136)\n- **Abstract**:\n> Audio deepfakes have improved rapidly recently, yet their effect on human trust in real speech remains unstudied. We present the largest listening study on audio deepfake perception to date, collecting 35,532 judgments from 1,768 participants across 138 text-to-speech and voice conversion systems. Our central finding is a skepticism shift: compared to a 2021 baseline, human accuracy on fake samples barely changed (72.9% to 71.2%), but accuracy on real samples dropped from 72.7% to 64.1%. Participants are not worse at detecting synthesis artifacts; rather, they increasingly distrust authentic speech. Samples generated by commercial and autoregressive language model systems proved hardest to detect (61.3 - 65.9%), while those from traditional seq2seq and flow-matching models remain easier to spot (75.4 - 76.8%). An ML detector that served as a reference point maintained over 94.5% accuracy across all conditions. Our results suggest that the primary threat posed by modern deepfakes may not be mere deception, but the erosion of trust in genuine audio.\n\n---\n\n## 123. Pretraining Data Exposure in Large Language Models: A Survey of Membership Inference, Data Contamination, and Security Implications\n- **Authors**: Ziyi Tong , Feifei Sun , Le Minh Nguyen\n- **URL**: [https://arxiv.org/abs/2605.26133](https://arxiv.org/abs/2605.26133)\n- **Abstract**:\n> Large Language Models (LLMs) have become the predominant paradigm in NLP, advancing both research and industry. As model sizes and pretraining data grow, concerns about Pretraining Data Exposure (PDE) increase due to the scale and opacity of training datasets. PDE refers to determining whether specific data appeared in an LLM's pretraining corpus. It is critical for ensuring evaluation integrity and protecting privacy, intersecting two key areas: data contamination and membership inference. Though conceptually related, these areas have often been studied in isolation. This paper offers the first unified survey of both under the PDE framework. We formalize PDE across exposure levels, review attack and defense methods, synthesize empirical findings, and highlight open challenges and future research directions.\n\n---\n\n## 124. GEM: Geometric Entropy Mixing for Optimal LLM Data Curation\n- **Authors**: Yue Min , Ziyun Qiao , Ruining Chen , Yujun Li\n- **URL**: [https://arxiv.org/abs/2605.26121](https://arxiv.org/abs/2605.26121)\n- **Abstract**:\n> LLM pre-training efficacy increasingly depends on data composition rather than sheer volume. Yet, optimal mixing is hindered by categorization flaws: human taxonomies suffer from ontological misalignment, and Euclidean clustering fails to address embedding anisotropy. We introduce GEM (Geometric Entropy Mixing), a framework reformulating data curation as a variational problem on the hypersphere augmented with a mixing-balance regularizer. By decoupling the generative prior and optimizing the objective via a provable MM (Minorize-Maximize) algorithm, GEM effectively counteracts the cluster collapse to discover balanced semantic structures invisible to Euclidean heuristics. We employ teacher-student distillation to scale this geometric fidelity to web-scale corpora and introduce the Geometric Influence Score (GIS) for interpretable taxonomy generation. Experiments with 1.1B-parameter models demonstrate that GEM establishes a new state-of-the-art when integrated into mixing strategies like DoReMi and RegMix, improving average downstream accuracy by up to 1.2% and offering a robust coordinate system for predictable data mixing.\n\n---\n\n## 125. Xe-Forge: Multi-Stage LLM-Powered Kernel Optimization for Intel GPU\n- **Authors**: Marcin Spoczynski , Daniel Fleischer , Moshe Berchansky , Gabriela Ben-Melech Stan , Shira Guskin , Weilin Xu , Adam Siemieniuk , Alexander Heinecke\n- **URL**: [https://arxiv.org/abs/2605.26118](https://arxiv.org/abs/2605.26118)\n- **Abstract**:\n> Porting deep learning algorithms to new hardware accelerators requires developers to repeatedly apply the same low-level optimizations -- quantization, memory access coalescing, tile size tuning, and architecture-specific workarounds -- to every Triton kernel in their code-base. This manual, repetitive effort is a major bottleneck: each kernel demands the same cycle of trial-and-error profiling against hardware constraints that vary across devices, yet the underlying optimization patterns remain largely consistent. We present Xe-Forge, a multi-stage LLM-powered pipeline that automates this process for Intel GPU. Given a functionally correct Triton kernel, the system applies up to nine optimization stages -- from algorithmic restructuring and operator fusion through block pointer modernization, GPU-specific tuning, and open-ended discovery -- each driven by a Chain-of-Verification-and-Refinement (CoVeR) agent that generates candidates, validates them on real hardware, and iterates on failures. A curated knowledge base encodes Intel GPU constraints (power-of-two warp counts, GRF modes, SLM sizing) that are absent from LLM training data, keeping the model within architecturally valid bounds. We evaluate Xe-Forge on 97 Level-2 KernelBench kernels and Flash Attention on the Intel Arc Pro B70, achieving a 1.17x geometric mean speedup over PyTorch eager with 67% of kernels improving, nine kernels exceeding 5x (up to 82x), and 2--13.3x speedups on Flash Attention across all tested configurations without regression -- demonstrating that structured domain knowledge with hardware-in-the-loop verification can systematically eliminate the repetitive porting effort that currently gates algorithm deployment on new accelerators.\n\n---\n\n{% endraw %}\n"}, "cbf010b9133a4f93e69ade3ff674398f236ab668": {"t": 1785770987.3061442, "n": 210239, "h": "cf2fc0302148c99f", "c": "---\nlayout: default\ntitle: \"LLM 관련 주요 논문 - 2026-05-27\"\n---\n\n{% raw %}\n# LLM 관련 주요 논문 - 2026-05-27\n\n## 1. MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation\n- **Authors**: Huawei Lin , Peng Li , Jie Song , Fuxin Jiang , Tieying Zhang\n- **URL**: [https://arxiv.org/abs/2605.27366](https://arxiv.org/abs/2605.27366)\n- **Abstract**:\n> Large language model (LLM) agents rely on reusable skills to solve complex tasks. However, existing skill creation approaches treat skills as isolated and static artifacts, limiting their reusability, reliability, and long-term improvement. We propose MUSE-Autoskill Agent (Memory-Utilizing Skill Evolution), a skill-centric agent framework that lets agents continuously improve their task-solving capability by creating, reusing, and refining skills under a unified lifecycle (creation, memory, management, evaluation, and refinement). Our framework enables agents to create skills on demand, store and reuse them across tasks, organize and select them efficiently, and evaluate them through unit tests and runtime feedback for continuous refinement. We further introduce skill-level memory that accumulates experience for each skill across tasks, enabling more effective reuse and adaptation over time. Experiments on SkillsBench provide initial evidence that lifecycle-managed skills can improve task success, efficiency, reuse, and cross-agent transfer, highlighting the importance of treating skills as long-lived, experience-aware, and testable assets.\n\n---\n\n## 2. Natural Language Query to Configuration for Retrieval Agents\n- **Authors**: Melissa Z. Pan , Negar Arabzadeh , Mathew Jacob , Fiodar Kazhamiaka , Esha Choukse , Matei Zaharia\n- **URL**: [https://arxiv.org/abs/2605.27361](https://arxiv.org/abs/2605.27361)\n- **Abstract**:\n> Modern retrieval agents expose many configuration choices -- LLM, retriever, number of documents, number of hops, and synthesis strategy -- each shaping both answer quality and serving cost. Today, these pipelines are typically hand-tuned once per workload, leaving substantial per-query optimization untapped. We formulate the problem: given a natural-language query and either an accuracy or a budget target, select from a predefined pipeline catalog the configuration that minimizes cost or maximizes accuracy at inference time. We propose **BRANE**, which uses an LLM to convert each query into workload-specific characteristics, then trains a lightweight per-configuration predictor that estimates whether the pipeline will answer the query correctly. At inference time, **BRANE** selects the configuration that maximizes predicted correctness penalized by cost, exposing a tunable cost-quality tradeoff without retraining. Across MuSiQue, BrowseComp-Plus, and FinanceBench, **BRANE** consistently pushes the cost-quality Pareto frontier, matches the best fixed configuration's accuracy at up to 89% lower cost, and outperforms LLM-routing, rule-based, and fine-tuned Qwen3-4B baselines. These results show that per-query configuration of the full retrieval pipeline is a practical alternative to static workload-level tuning.\n\n---\n\n## 3. Alignment Tampering: How Reinforcement Learning from Human Feedback Is Exploited to Optimize Misaligned Biases\n- **Authors**: Dongyoon Hahm , Dylan Hadfield-Menell , Kimin Lee\n- **URL**: [https://arxiv.org/abs/2605.27355](https://arxiv.org/abs/2605.27355)\n- **Abstract**:\n> Reinforcement Learning from Human Feedback (RLHF) is the standard method to align Large Language Models (LLMs) with human preferences. In this work, we introduce alignment tampering, a potential vulnerability where the LLM undergoing alignment influences the preference dataset, causing RLHF to amplify undesired behaviors. This arises from core limitations of RLHF: (1) preference datasets are constructed from the LLM's own outputs, allowing it to influence them, and (2) pairwise comparisons only indicate which response is better, not why. These limitations can be exploited to cause alignment tampering. For example, if an LLM generates biased responses with higher quality, annotators will prefer them based on quality. However, preference labels do not distinguish quality from bias, and the reward model inherits this limitation. Optimizing such rewards through reinforcement learning or best-of-N sampling can amplify misaligned biases. Our experiments demonstrate amplification across diverse biases: from keyword bias to propaganda (e.g., sexism), brand promotion, and instrumental goal-seeking. Mitigation remains challenging, as existing techniques for robust RLHF fail to fully resolve alignment tampering without sacrificing response quality. These findings reveal structural vulnerabilities of current RLHF and emphasize the need to prevent this vulnerability. Project page: this https URL\n\n---\n\n## 4. Gumbel Machine: Counterfactual Student Writing Generation via Gumbel Noise Steering\n- **Authors**: Hunter McNichols , Alexander Scarlatos , Mihai Dascalu , Danielle McNamara , Andrew Lan\n- **URL**: [https://arxiv.org/abs/2605.27249](https://arxiv.org/abs/2605.27249)\n- **Abstract**:\n> An effective method of teaching across disciplines is to provide examples of high-quality work. However, an example may be significantly different from a student's current work, making it challenging for them to emulate. An ideal learning demonstration is a counterfactual version of the student work, an improved version that is still similar to their own. Existing automated approaches for counterfactual text generation using Large Language Models (LLMs) result in domain-specific systems that are difficult to translate into practical applications. We present the Gumbel Machine, a flexible, modular approach to generating counterfactuals that leverages LLM instruction-following capabilities while encouraging similarity to a reference factual text. Central to our approach is a novel, controlled decoding algorithm, $\\beta$-Hindsight control, which uses latent randomness as a tunable similarity control mechanism during counterfactual generation. Experiments on datasets of student writing, scored on various criteria, demonstrate the effectiveness of our approach at generating counterfactuals both rubric-consistent and similar to a reference.\n\n---\n\n## 5. Learning to Act under Noise: Enhancing Agent Robustness via Noisy Environments\n- **Authors**: Yuxin Chen , Xiaodong Cai , Junfeng Fang , Zhuowen Han , Yu Wang , Yaorui Shi , Yi Zhang , Qi Gu , Xunliang Cai , Xiang Wang , An Zhang , Tat-Seng Chua\n- **URL**: [https://arxiv.org/abs/2605.27209](https://arxiv.org/abs/2605.27209)\n- **Abstract**:\n> Recent advances in large language models (LLMs) have facilitated the widespread deployment of LLMs as interactive agents capable of reasoning, planning, and tool use. Despite strong performance on existing benchmarks, such agents often exhibit notable degradation when deployed in real-world settings, where environments are inherently stochastic and imperfect. We argue that this discrepancy arises from a fundamental mismatch between idealized training settings and real-world interaction dynamics, where current paradigms rely on carefully curated task instructions and stable, well-controlled environments. To address this gap, we propose NoisyAgent, an agentic training framework that explicitly incorporates environmental imperfections into the agent learning process. We identify two major sources of interaction noise in real-world scenarios: user noise, which captures ambiguity and variability in user interaction, and tool noise, which reflects failures and anomalies in tool execution. We introduce such perturbations into the training pipeline by modifying user interaction patterns and simulating tool execution results within the training environment. To stabilize training while encouraging agents to handle increasingly challenging imperfections, noise is applied to only a subset of rollouts and progressively increased in difficulty as the model adapts to the current noise level. Extensive experiments demonstrate that our approach consistently improves agent robustness under noisy and dynamic environments. Our analysis reveals that training under noise conditions also yields performance gains on idealized benchmarks, suggesting that controlled exposure to environmental noise promotes more generalizable reasoning and decision-making behaviors. Our findings highlight the importance of modeling interaction imperfections for bridging the gap between agent training and real-world deployment.\n\n---\n\n## 6. The Compressive Knowledge Graph Hypothesis: Which Graph Facts Matter for Scientific Hypothesis Generation?\n- **Authors**: Shashwat Sourav , Viktoriia Baibakova , Sanjay Das , Ran Elgedawy , Maria Mahbub , Emily Herron , Tirthankar Ghosal\n- **URL**: [https://arxiv.org/abs/2605.27176](https://arxiv.org/abs/2605.27176)\n- **Abstract**:\n> Knowledge graphs (KGs) can provide structured scientific context to language models, but it remains unclear which graph facts actually shape the generated hypotheses. We study KG-guided hypothesis generation for battery materials across Mistral-7B, Llama-3.1-70B, and Gemini 2.5 Flash. We perturb local KGs by varying density, ontology richness, topology, and control structure, and evaluate outputs with both provided-graph and fixed-reference metrics. Across models, KG utility is selective and model-dependent: graph context changes outputs, but no-KG outputs also recover substantial graph content from model priors. Compact top-k subgraphs often approximate full-KG behavior, including when claimed-outcome triples are held out. At the same time, compression is not unique to one semantic ranking rule, random and topology-based subsets can also recover much of the signal. These results support a redundancy-aware Compressive KG hypothesis: useful KG signal is often recoverable from compact, scientifically structured subgraphs rather than requiring the full local graph.\n\n---\n\n## 7. VitaBench 2.0: Evaluating Personalized and Proactive Agents in Long-Term User Interactions\n- **Authors**: Yuxin Chen , Yi Zhang , Zhengzhou Cai , Yaorui Shi , Zhiyuan Yao , Chenhang Cui , Jingnan Zheng , Yaqi Huo , Xi Su , Qi Gu , Xunliang Cai , Xiang Wang , An Zhang , Tat-Seng Chua\n- **URL**: [https://arxiv.org/abs/2605.27141](https://arxiv.org/abs/2605.27141)\n- **Abstract**:\n> Large language models (LLMs) have evolved into interactive agents that collaborate with users in real-world tasks. Effective collaboration in such settings increasingly depends on understanding the user beyond what is explicitly stated, as user intent is often reflected in fragmented daily interactions and requires both personalized modeling and proactive interaction. However, existing agent benchmarks primarily evaluate reasoning and tool use, largely overlooking the challenges of inferring and leveraging user preferences in realistic scenarios. To address this gap, we introduce VitaBench 2.0, a benchmark for evaluating personalized and proactive agent behavior in long-term user interactions. In VitaBench 2.0, tasks are organized as temporally ordered sequences for individual users, where preferences are embedded in fragmented and heterogeneous interactions. Successful completion of tasks requires the agent to continuously extract, utilize, and update user preferences from these interactions. We further evaluate proactiveness through tasks that require agents to recognize missing information and actively acquire it from users or environments before making decisions. To support systematic analysis, we provide an extensible memory interface that enables controlled comparison across different memory architectures. We benchmark a diverse set of frontier proprietary and open-source LLMs. Results show that real-world personalization remains highly challenging even for state-of-the-art models, revealing a substantial gap between current capabilities and practical requirements. Extensive analysis further reveals the failure modes and capability bottlenecks of current agents in real-world personalized decision-making, providing insights for future model improvements.\n\n---\n\n## 8. ICCU: In-Context Continual Unlearning via Pattern-Induced Refusal Rules\n- **Authors**: Ruihao Pan , Suhang Wang\n- **URL**: [https://arxiv.org/abs/2605.27138](https://arxiv.org/abs/2605.27138)\n- **Abstract**:\n> Machine unlearning aims to remove the influence of specific data from trained language models. In real-world deployments, unlearning requests often arrive sequentially, which challenges existing fine-tuning-based methods: fine-tuning each request is costly, accumulates utility loss, and may cause cross-request interference. To address these issues, we propose ICCU (In-Context Continual Unlearning), an in-context continual unlearning framework that induces readable refusal rules from unlearning datasets and applies them at inference time either as a filter or via the system prompt, without modifying model parameters. Because rules are accumulated as an order-independent union, ICCU is compositional and free of cross-request interference, and the original forget-set data can be discarded after rule induction. Extensive experiments show that ICCU effectively suppresses target knowledge while preserving utility, scales across sequential requests, and remains robust to paraphrased and cross-lingual queries.\n\n---\n\n## 9. Scaling, Benchmarking, and Reasoning of Vision-Language Agents for Mobile GUI Navigation\n- **Authors**: Heng Qu , Yike Liu , Renren Jin , Wenzong Zhang , Pengzhi Gao , Wei Liu , Jian Luan\n- **URL**: [https://arxiv.org/abs/2605.27134](https://arxiv.org/abs/2605.27134)\n- **Abstract**:\n> Vision-Language Models (VLMs) have shown rapid progress in mobile GUI navigation. This paper presents a systematic study of data scaling, benchmarking, and reasoning for VLM-based agents in this domain. To facilitate rigorous evaluation, we introduce HyperTrack, a large-scale dataset with over 16000 real-world tasks across more than 650 Chinese mobile applications, along with GUIEvalKit, an open-source toolkit for unified benchmarking of VLMs on offline GUI navigation tasks. Using HyperTrack, we analyze the effects of training data scale on both supervised and reinforcement-based finetuning. Our results show that reinforcement-based finetuning consistently outperforms supervised finetuning, particularly in out-of-domain settings, highlighting the synergy between data scaling and reinforcement learning. Leveraging GUIEvalKit, we further benchmark state-of-the-art (SOTA) VLMs and analyze how interaction history and reasoning capabilities influence task completion. Together, HyperTrack and GUIEvalKit provide a comprehensive platform for developing and evaluating VLM agents in mobile GUI navigation tasks.\n\n---\n\n## 10. Position: AI Safety Requires Effective Controllability\n- **Authors**: Yige Li , Yunhao Feng , Jun Sun\n- **URL**: [https://arxiv.org/abs/2605.27117](https://arxiv.org/abs/2605.27117)\n- **Abstract**:\n> AI safety is still largely framed as alignment: training models to follow human preferences, safety policies, and normative constraints. That framing has improved the behavior of modern language models, but aligned behavior does not by itself guarantee that a deployed agent can be stopped, overridden, or constrained once it operates in open-ended, interactive, and tool-using environments. A system may be safe in expectation and still fail to yield to explicit runtime authority under conflicting instructions, long-horizon execution, adversarial inputs, or risky tool use. This position paper argues that AI safety therefore requires controllability as a first-class objective. We define \\emph{controllability} as the ability of an AI system to remain reliably interruptible, overridable, redirectable, and constrainable by explicit control signals at runtime while preserving ordinary utility when such signals are absent. To study this gap, we introduce \\controlbench{}, a benchmark for evaluating controllability failures in high-risk agentic scenarios. Experiments with OpenClaw-based agents show that current alignment and guardrail mechanisms reduce risk, but often fail to provide persistent, authoritative, and enforceable runtime control. We therefore propose a control-centric architectural framework that highlights explicit control planes, runtime intervention pathways, persistent control states, and auditable decision interfaces as key design principles for future controllable AI systems.\n\n---\n\n## 11. Counteraction-Aware Multi-Teacher On-Policy Distillation for General Capability Recovery with Domain Preservation\n- **Authors**: Tianlei Chen , Jiao Ou , Ziyuan Liu , Ruiming Tang , Jian Liang , Han Li\n- **URL**: [https://arxiv.org/abs/2605.27115](https://arxiv.org/abs/2605.27115)\n- **Abstract**:\n> Domain specialization can improve LLM behavior in vertical domains, but often weakens the general capabilities inherited from the original model. Recent Multi-Teacher On-Policy Distillation (MOPD) pipelines recover model capabilities by supervising student-generated trajectories with teacher feedback, but typically assume teacher-aligned prompt coverage, requiring prompts to match the teachers' training distributions. This assumption is difficult to satisfy when the general teacher is an open-source model whose post-training data are unknown. Instead of attempting to reconstruct this hidden distribution, we study general capability recovery with readily available proxy general prompts. We identify two failure modes of vanilla MOPD in this incomplete-coverage situation: recovery-preservation counteraction from mixing conflicting recovery and preservation gradients, and weak-signal flattening from uniformly averaging samples with unequal correction demand. We propose Counteraction-Aware Multi-Teacher On-Policy Distillation (CaMOPD), which addresses these issues with decoupled alternating training and gap-based sample selection. CaMOPD gives general recovery dedicated updates, periodically reviews domain prompts for preservation, and selects samples with larger averaged token-level teacher-student log-probability gaps to concentrate correction signals. Across role-play dialogue and medical reasoning QA scenarios, CaMOPD performs best in general recovery over baselines while maintaining domain-specific behavior. Gradient coherence analyses further support the intended effect of CaMOPD in producing more coherent correction signals.\n\n---\n\n## 12. Traceable Knowledge Graph Reasoning Enables LLM-Assisted Decision Support for Industrial VOCs in the Steel Industry\n- **Authors**: Changqing Su , Yu Ding , Zuhong Lin , Hongyu Liu , Xi He , Zheng Zeng , Liqing Li\n- **URL**: [https://arxiv.org/abs/2605.27071](https://arxiv.org/abs/2605.27071)\n- **Abstract**:\n> Key knowledge for steel-industry volatile organic compounds (VOCs) governance is scattered across unstructured scientific literature, making it difficult to integrate process, pollutant, and control-technology evidence and increasing the risk of hallucination when general large language models (LLMs) answer low-frequency industrial questions. Here we developed Chat-ISV, a knowledge graph (KG) enhanced multi-agent Q&A system that parses a curated steel-industry VOCs literature corpus, constructs a Neo4j KG with 27180 nodes and 81779 semantic edges, and combines prompt-constrained extraction, chunk-centered topology optimization, multi-agent routing, source-backtracking retrieval, local literature retrieval, open-domain knowledge access, and interactive subgraph visualization. Benchmark tests and 400 expert blind evaluations showed that topology optimization reduced isolated nodes from 57% to 4.08% and that Chat-ISV achieved high factual reliability, with 96.93% precision, 72.63% recall, an F1-score of 0.830, and a mean score of 1.69/2.00. By converting fragmented environmental-engineering literature into traceable, queryable, and decision-support-oriented knowledge, Chat-ISV establishes a scalable environmental-informatics paradigm for reliable LLM deployment and intelligent pollution-control decision support in specialized industrial domains.\n\n---\n\n## 13. Generating Robust Portfolios of Optimization Models using Large Language Models\n- **Authors**: Eleni Straitouri , Cheol Woo Kim , Milind Tambe\n- **URL**: [https://arxiv.org/abs/2605.27013](https://arxiv.org/abs/2605.27013)\n- **Abstract**:\n> Mathematical optimization is a powerful tool for structured decision-making across domains such as resource allocation and planning. Formulating optimization models faithful to reality, though, remains a significant bottleneck as it typically demands both domain expertise and optimization knowledge that are often scarce. Recent advances in large language models (LLMs) promise to bridge this gap, enabling the generation of candidate optimization models from natural language descriptions. However, there is no guarantee that any single LLM-generated model is reliable, and existing approaches that output only one model are therefore risky. In this work, we propose a novel algorithm that generates a portfolio of optimization models, designed to be robust to the limitations of LLMs. Our method exploits the observation that a single LLM can play two distinct roles $\\unicode{x2014}$ as a stochastic generator and as a reasoning evaluator $\\unicode{x2014}$ and proposes a unified framework that leverages both capabilities in a complementary manner. We provide theoretical guarantees showing that, as long as either the generator or the evaluator is well-aligned with human preferences, the portfolio is guaranteed to contain high-quality candidates, enabling a principled human-in-the-loop process in which a decision-maker can review multiple candidates before committing to one. We further validate our approach empirically, demonstrating strong performance across a range of optimization modeling tasks.\n\n---\n\n## 14. LELA: An End-to-end LLM-based Entity Linking Framework with Zero-shot Domain Adaptation\n- **Authors**: Samy Haffoudhi (IP Paris, LTCI, DIG), Nikola Dobričić (IP Paris), Fabian Suchanek (IP Paris, LTCI), Nils Holzenberger\n- **URL**: [https://arxiv.org/abs/2605.26956](https://arxiv.org/abs/2605.26956)\n- **Abstract**:\n> Entity linking is a key component of many downstream NLP systems, yet existing approaches are often tied to the specific target knowledge bases and domains, limiting their real world application. In this paper, we extend LELA, a modular and domain-agnostic LLM-based entity disambiguation method, into a practical Python library that integrates zero-shot Named Entity Recognition (NER) -thereby providing a complete end-toend pipeline for entity-linking in real-world usage. We provide experimental results validating LELA's performance and robustness across diverse entity linking settings. In our demo, users can play with the system on their own input texts.\n\n---\n\n## 15. Neuro-Symbolic Verification of LLM Outputs for Data-Sensitive Domains (extended preprint)\n- **Authors**: Paul Sigloch , Christoph Benzmüller\n- **URL**: [https://arxiv.org/abs/2605.26942](https://arxiv.org/abs/2605.26942)\n- **Abstract**:\n> LLMs deployed in high-stakes domains face fundamental reliability challenges: hallucinations, inconsistencies, and privacy vulnerabilities introduce unacceptable risks where errors carry legal, financial, or safety consequences. This paper presents a hybrid verification architecture combining formal symbolic methods with neural semantic analysis to provide complementary guarantees for LLM-generated content. This architecture employs logical reasoning for input verification, leveraging completeness properties to provide decidable guarantees on structured requirements. For output validation, embedding-based semantic similarity detects contextual hallucinations where formal methods lack expressiveness. This separation is realized in a parallel, actor-based pipeline, addressing limitations of prompt-based self-verification approaches, which inherit the distributional biases that produce hallucinations. The proposed architecture and type-aware verification method are validated with HAIMEDA, a real-world medical device damage assessment reporting system developed through Action Design Research. Evaluation shows hallucination detection rates of over 83% for structured entities and 72% for semantic fabrications, with a 30% reduction in report creation time, demonstrating that neuro-symbolic architectures can provide principled safeguards for LLM deployment in data-sensitive domains.\n\n---\n\n## 16. From Norms to Indicators (N2I-RAG): An Agentic Retrieval-Augmented Generation Framework for Legal Indicator Computation\n- **Authors**: Youssef Al Mouatamid , Marie Bonnin , Jihad Zahir\n- **URL**: [https://arxiv.org/abs/2605.26926](https://arxiv.org/abs/2605.26926)\n- **Abstract**:\n> Computing legal indicators from normative texts is a key task in legal monitoring and policy evaluation, but presents significant challenges due to the complexity, scale, and interpretive nature of legal language, as well as the variability in available document quality. Existing natural language processing techniques and generative models can assist in legal analysis, but often suffer from high risk of hallucinations and lack the interpretability and evidence grounding required for reliable indicator computation. This paper presents N2I-RAG (From Norms to Indicators), an agentic retrieval-augmented generation framework designed to automate the computation of legal indicators in a transparent and traceable way. We integrate adaptive retrieval, llm-based agents, and validation mechanisms in a modular pipeline, where each component performs a defined role in filtering, retrieving, and assessing evidence, and in producing binary legal outcomes linked to identifiable legal provisions. The framework emphasizes traceability by requiring explicit explanations of intermediate decisions and final indicator assignments. We evaluate N2I-RAG using an in-house constructed French marine environmental law corpus that includes both scanned and digital sources. Comparative experiments with multiple language model families demonstrate that the proposed approach consistently outperforms baseline systems, and generalizes well when tested on 2 different bans. The results indicate that agentic retrieval-augmented generation can bridge open-text legal language and standardized indicator computation, offering a foundation for transparent and scalable legal observatories.\n\n---\n\n## 17. TADDLE: A Tool-Augmented Agent for Detecting Deficient LLM-Generated Peer Reviews\n- **Authors**: Hanqi Duan , Xiang Li\n- **URL**: [https://arxiv.org/abs/2605.26911](https://arxiv.org/abs/2605.26911)\n- **Abstract**:\n> LLM-generated peer reviews are increasingly common at major venues, yet their deficiencies are hard to detect because they are uniformly fluent and well-structured. Existing work either classifies authorship without judging quality, or scores quality with features designed for human-written reviews; no prior system detects deficiencies in LLM-generated reviews at the level of individual defect types. To bridge the gap, we introduce TADDLE, a Tool-Augmented Agent for Detecting Deficient LLM-Generated Peer Reviews, together with the first expert-annotated benchmark for this task. Our benchmark comprises 1,800 reviews on 50 ICLR 2025 papers, multi-label-annotated by 18 domain experts against a taxonomy of six defect categories (plus a non-deficient label). TADDLE decomposes detection into four specialized analysis tools -- Verify, Correct, Complete, and Transform -- orchestrated by an agent; an integrator synthesizes their outputs into binary and multi-label classifications via two-stage semi-supervised learning. Extensive experiments show that TADDLE performs strongly on both binary detection and the multi-label classification task. We release the benchmark and code at this https URL .\n\n---\n\n## 18. Multi-Stakeholder LLM Alignment: Decomposing Estimation from Aggregation\n- **Authors**: Lulu Zheng , Wenjin Yang , Xiangwen Zhang , Rong Yin , Yulan Hu , Zheng Pan , Xin Li\n- **URL**: [https://arxiv.org/abs/2605.26878](https://arxiv.org/abs/2605.26878)\n- **Abstract**:\n> Multi-stakeholder tasks require one output to satisfy users with conflicting preferences. Holistic LLM judges conflate utility estimation and utility aggregation, yielding unstable implicit weights. We show empirically and theoretically that this aggregation-specific \\emph{weighting noise} can create large score shifts when stakeholder satisfaction is dispersed; in our experiments, these weight-induced shifts also increase with stakeholder count. We propose \\textsc{DecompR}: counterfactual-calibrated weights are fixed from query structure before candidate scoring, while per-role utilities are estimated independently, removing candidate-dependent weight drift and reducing estimation noise.\n\n---\n\n## 19. Helicase: Uncertainty-Guided Supply Chain Knowledge Graph Construction with Autonomous Multi-Agent LLMs\n- **Authors**: Yunbo Long , Haolang Zhao , Ge Zheng , Alexandra Brintrup\n- **URL**: [https://arxiv.org/abs/2605.26835](https://arxiv.org/abs/2605.26835)\n- **Abstract**:\n> LLM-based multi-agent systems have been widely adopted for knowledge retrieval and report generation, synthesizing known information through web search and textual reasoning. However, many critical information tasks in supply chains are not simple one-shot queries: they are structural inference problems requiring multi-hop reasoning across complex, fragmented web resources. Questions such as \\textit{``Which Tesla components use lithium from Australian mines?''} have no answer in any single document; answers must be computationally synthesized through the autonomous construction and analysis of dynamic knowledge graphs assembled from fragmented, heterogeneous sources. Moreover, such discovery processes must be uncertainty-aware: decisions depend not only on answers but on calibrated confidence in their reliability, traceable to source quality and reasoning consistency. To address this capability gap, we propose \\textit{Helicase}, an autonomous multi-agent LLM system for uncertainty-guided supply chain knowledge graph construction. \\textit{Helicase} decomposes high-level supply-chain queries into executable investigation plans, coordinates specialized web-search, reasoning, and coding agents through iterative verification loops, and incrementally constructs query-specific supply chain knowledge graphs with per-fact uncertainty annotations. Its three-layer uncertainty framework tracks uncertainty at the action, trajectory, and memory layers, enabling both structural inference and calibrated confidence assessment. To evaluate autonomous reasoning across the full complexity spectrum, we introduce SCQA (Supply Chain Query Assessment), a benchmark of 80 supply chain queries organized into four quadrants spanning single-hop to multi-hop inference under both high and low data visibility.\n\n---\n\n## 20. The Attribution Blind Spot: Detecting When Language Models Rely on Memory Rather Than Retrieved Context\n- **Authors**: Zhe Yu , Wenpeng Xing , Yunzhao Wei , Bo Yang , Chen Ye , Gaolei Li , Meng Han\n- **URL**: [https://arxiv.org/abs/2605.26778](https://arxiv.org/abs/2605.26778)\n- **Abstract**:\n> Retrieval-augmented generation promises to ground language model outputs in external evidence, yet the field has no reliable way to verify whether retrieved context actually governs generation -- a prerequisite for any high-stakes deployment. The standard assumption, that context-consistent output implies context-governed output, breaks when the retrieved document overlaps with the model's pretraining data: the model can produce faithful-looking text entirely from parametric memory, and both pathways yield indistinguishable output. We name this failure the attribution blind spot and introduce Computational Reality Monitoring (CRM) to address it. CRM operationalizes a principle adapted from cognitive science's reality monitoring framework: comparing internal representations with and without context reveals membership-conditioned representational divergence that output-level monitors systematically miss. CRM does not certify which source an individual generation used; it detects whether pretraining exposure leaves a measurable internal trajectory signature, establishing a necessary substrate for source attribution. Across nine model variants spanning three families, this divergence concentrates in architecture-specific layer patterns, receives converging support from block-level noise intervention, and generalizes across tasks and datasets while collapsing on domain-confounded benchmarks. The attribution blind spot is measurable and partially addressable: internal representations carry a diagnostic signal invisible at the output level, establishing a foundation for systems whose internal awareness of evidence provenance governs their external behavior.\n\n---\n\n## 21. A Dataset of Robot-Patient and Doctor-Patient Medical Dialogues for Spoken Language Processing Tasks\n- **Authors**: Heriberto Cuayahuitl , Grace Jang\n- **URL**: [https://arxiv.org/abs/2605.26747](https://arxiv.org/abs/2605.26747)\n- **Abstract**:\n> Large Language Models (LLMs) have brought huge improvements to Artificial Intelligence (AI), which can be applied to general-purpose tasks. However, their application to textual or spoken medical consultations is still an open research problem. This paper proposes MeDial-Speech, a novel speech dataset for training and evaluating Med-AIs that can carry out consultations with patients. It was collected in realistic environments from robot-patient and doctor-patient dialogues, contains 111+ hours of speech data (without data augmentation), and covers four health conditions: Lewy body dementia, heart failure, shoulder pain, and angina. In addition, we propose a dialogue benchmark via sentence selection (with 20 options) to evaluate three state-of-the-art LLMs: GPT-5 mini, DeepSeek-V3, and Claude Sonnet 4. Experimental results reveal that Claude Sonnet 4 is the best in sentence selection, with 71.1% accuracy using manual transcriptions and 74.7% using automatic transcriptions, and that all LLMs are highly overconfident in their probabilistic predictions, regardless of selecting correct or incorrect sentences in medical dialogues. This dataset is free of charge for non-commercial purposes at: this https URL\n\n---\n\n## 22. It's Not the Capability: Harness Sensitivity Is Non-Monotone Across LLM Agent Tiers\n- **Authors**: Yong-eun Cho\n- **URL**: [https://arxiv.org/abs/2605.26731](https://arxiv.org/abs/2605.26731)\n- **Abstract**:\n> A prevalent assumption in LLM agent deployment holds that more structured harnesses universally improve reliability, and that higher-capability models need proportionally less structural guidance -- together implying a monotone inverse relationship between model capability tier and optimal harness complexity. We test this hypothesis through a controlled 432-run experiment crossing six models across four capability tiers with three harness conditions (light, balanced, strict) on HEAT-24, a 24-task synthetic benchmark with git-based workspace verification. Our results refute the monotone inverse relationship on two fronts. First, for the frontier chat model evaluated (Gemini 2.5 Flash), increased harness verbosity lowers VTSR by 29-38 percentage points -- a harness-complexity paradox. Second, for the frontier reasoning model evaluated (Qwen3.5-122B, extended thinking enabled), strict harness achieves the highest VTSR (91.7%) and the lowest latency, the opposite of the prediction. Within the constrained tier, a 2B model (Gemma4:e2B) matches strong-open-tier stability at 91.7% across all harnesses. Because each tier is represented by a single model in this study, these results should be interpreted as model-specific observations; harness sensitivity appears non-monotone across the models evaluated, and depends critically on model type (chat vs. reasoning). We introduce a six-label failure taxonomy showing that format_violation dominates capable-model failures while wrong_file dominates low-capability failures, and we derive practical tier-aware harness selection guidelines.\n\n---\n\n## 23. Towards Feedback-to-Plan Decisions for Self-Evolving LLM Agents in CUDA Kernel Generation\n- **Authors**: Yee Hin Chong , Jiaming Wu , Youhui Zhang , Peng Qu\n- **URL**: [https://arxiv.org/abs/2605.26720](https://arxiv.org/abs/2605.26720)\n- **Abstract**:\n> Large language models (LLMs) have shown strong empirical gains as self-evolving agents for CUDA kernel generation, driven by feedback-conditioned planning across generations. However, how planning decisions attribute and combine heterogeneous feedback signals remains opaque. Standard end-to-end ablations fail to resolve this question, as iterative planning amplifies early perturbations and conflates feedback effects with trajectory-dependent drift. We introduce \\texttt{CUDAnalyst}, a unified analysis layer for controlled, generation-level attribution of planning decisions to feedback components via trajectory freezing and selective feedback injection. \\texttt{CUDAnalyst} enables stable generation-level evaluation and principled coalitional-style attribution of feedback effects and interactions. Our results show that explicit planning is beneficial only when feedback is aligned, that effective planning emerges from structured multi-feedback interactions, and that high-level plans from stronger reasoning models can partially transfer to weaker ones. These trends hold across reference backbones, representative workloads, and reference induction regimes, indicating that the identified feedback-to-plan structure is robust within the controlled axes studied.\n\n---\n\n## 24. MemFail: Stress-Testing Failure Modes of LLM Memory Systems\n- **Authors**: Ishir Garg , Neel Kolhe , Dawn Song , Xuandong Zhao\n- **URL**: [https://arxiv.org/abs/2605.26667](https://arxiv.org/abs/2605.26667)\n- **Abstract**:\n> Large language model (LLM) agents increasingly rely on external memory systems to remain consistent across long-horizon interactions, but little empirical work has been done to understand the specific failure modes and design choices that these systems present. Existing benchmarks report aggregate question-answering accuracy and treat memory systems as black boxes, making it impossible to attribute an incorrect answer to a particular failure mode of the system. We introduce MemFail, a diagnostic benchmark that isolates the failure modes of modern LLM memory systems. We begin by formalizing memory systems as the composition of three canonical operations -- summarization, storage, and retrieval -- and identify the potential failure modes induced by each. Based on these hypothesized failure modes, we construct five datasets spanning four tasks, each adversarially designed to test a specific operation of a memory system. Using these datasets, we evaluate four state-of-the-art memory systems on MemFail and demonstrate how MemFail can be used to empirically understand the tradeoffs induced by differences in memory system architectures.\n\n---\n\n## 25. UnityMAS-O: A General RL Optimization Framework for LLM-Based Multi-Agent Systems\n- **Authors**: Yiqun Chen , Wei Yang , Erhan Zhang , Shijie Wang , Qi Liu , Zechun Niu , Bin Zhang , Haitao Li , Rui Li , Lingyong Yan , Jinyuan Feng , Biqing Qi , Xiaochi Wei , Yan Gao , Yi Wu , Yao Hu , Jiaxin Mao\n- **URL**: [https://arxiv.org/abs/2605.26646](https://arxiv.org/abs/2605.26646)\n- **Abstract**:\n> LLM-based multi-agent systems decompose complex tasks into interacting roles, but most remain manually orchestrated by prompts, tools, and control rules, while agents are rarely optimized through a unified reinforcement learning interface. Existing RL post-training frameworks mainly target single-policy optimization and lack abstractions for user-defined multi-agent workflows, structured interaction, role-specific credit assignment, and configurable parameter sharing. We present UnityMAS-O, a general RL optimization framework for LLM-based multi-agent systems. UnityMAS-O treats the complete workflow as the optimization unit, rather than a single response or policy trajectory. It represents workflows through four first-class objects: logical agent roles, graph trajectories, user-defined rewards, and agent--model mappings. This decouples logical agents from physical model parameters, supporting full sharing, full separation, and partial sharing, with rewards assigned at role, turn, and trajectory levels. UnityMAS-O extends verl with a Ray-based star-topology runtime. A central controller executes workflows, invokes tools, records structured trajectories, and assembles rewards; model-local worker groups handle rollout, buffering, advantage computation, and distributed PPO-style updates. Users can define agents, workflows, model mappings, and rewards without rewriting the optimization infrastructure. We instantiate UnityMAS-O on retrieval-augmented QA, iterative agentic search, and reflective code generation. Across Natural Questions, HotpotQA, and held-out code tasks, multi-agent RL improves manually specified workflows after optimization, with especially large gains for smaller models and strict code all-passed metrics. These results show that UnityMAS-O can serve as a reusable substrate for converting diverse LLM-based multi-agent workflows into trainable multi-agent RL systems.\n\n---\n\n## 26. FAST-GOAL: Fast and Efficient Global-local Object Alignment Learning\n- **Authors**: Hyungyu Choi , Young Kyun Jang , Chanho Eom\n- **URL**: [https://arxiv.org/abs/2605.26615](https://arxiv.org/abs/2605.26615)\n- **Abstract**:\n> Vision-language models such as CLIP have shown impressive capabilities in aligning images and text, but they often struggle with lengthy and detailed text descriptions due to pre-training on short and concise captions. We present FAST-GOAL (Fast and Efficient Global-local Object Alignment Learning), an efficient fine-tuning method that enhances ability of CLIP to handle lengthy text through global-local semantic alignment. Our method consists of two key components. First, Fast Local Image-Sentence Matching (FLISM) efficiently extracts local image regions through object detection and spatial division, then matches them with corresponding sentences. Second, Token Similarity-based Learning (TSL) maximizes the similarity between patch tokens from specific regions in the image and their corresponding region embeddings, applying the same principle to text, which enhances the ability of the model to capture detailed correspondences. Additionally, we introduce GLIT100k, a dataset that provides both global image-lengthy caption pairs and context-derived local pairs, where local descriptions are extracted from global captions to maintain semantic coherence. Through extensive experiments on long caption datasets (DOCCI, DCI) and short caption datasets (MSCOCO, Flickr30k), we demonstrate that FAST-GOAL achieves significant improvements over baselines, enabling effective adaptation of CLIP to detailed textual descriptions while maintaining computational efficiency.\n\n---\n\n## 27. AGORA: Adapter-Grounded Observation-Action Retention for Inference-Free Prompt Compression in LLM Agents\n- **Authors**: Haoran Zhang , Zhaohua Sun\n- **URL**: [https://arxiv.org/abs/2605.26596](https://arxiv.org/abs/2605.26596)\n- **Abstract**:\n> The token-level extractive compressors widely used for general LM context are structurally inappropriate for LLM agents: across 17 (env, backbone, method) cells spanning two independent token-level method families, every cell collapses to mean reward <= 0.05 despite 1.3-13.3x realized compression. We name and characterize this failure mode as action-grammar destruction -- the tokens carrying action semantics (identifiers, brackets, action verbs) are exactly those self-information ranks lowest, so a general-purpose compressor reliably removes them and the environment rejects the residual. The diagnosis points to step-granularity compression. We introduce AGORA, an inference-free step-level compressor combining a structural prompt parser, an always-keep floor for format- and recency-critical content, and a 125M-parameter relevance scorer trained on counterfactual next-action-change labels (~2ms/step, zero per-step LLM toll). Across the compared inference-free and LLM-based methods, AGORA is the only one retaining >= 75% uncompressed performance in 8 of 9 cells (with the lone exception at 73%); a four-way component ablation isolates the structural floor as the dominant quality lever and the learned scorer as the source of 1.0-11.5x adaptive end-to-end compression from a single fixed keep ratio.\n\n---\n\n## 28. MedGuideX: Internalizing Decision Logic from Executable Guidelines into Large Language Models for Clinical Reasoning\n- **Authors**: Yuhao Shen , Lang Cao , Simo Du , Yuqing Wang , Juexiao Zhou , Hao Peng , Yue Guo\n- **URL**: [https://arxiv.org/abs/2605.26567](https://arxiv.org/abs/2605.26567)\n- **Abstract**:\n> Clinical practice guidelines (CPGs) encode evidence-based decision logic that clinicians apply by evaluating patient variables, conditional criteria, and recommendation rules. However, existing methods often use CPGs as free-text training data or retrieval sources, underutilizing their procedural decision structure. To better exploit this structure, we introduce a guideline-derived training pipeline that transforms CPG recommendations into executable clinical decision logic and uses it to generate factual and counterfactual question-answering data. Theses data teach models both guideline-supported decisions and how decisions change under different patient conditions. Post-training a medical LLM on the generated data yields MedGuideX. Across four clinical reasoning benchmarks, MedGuideX achieves a 10.28% relative improvement in average accuracy. Physician evaluation further shows that MedGuideX better recovers clinician authored reasoning steps and produces physician-preferred rationales in faithfulness, validity, completeness, and clarity. Overall, our results show that executable decision logic from CPGs can be transformed into scalable supervision for building reliable medical LLMs.\n\n---\n\n## 29. MobileExplorer: Accelerating On-Device Inference for Mobile GUI Agents via Online Exploration\n- **Authors**: Runxi Huang , Liyu Zhang , Shengzhong Liu , Xiaomin Ouyang\n- **URL**: [https://arxiv.org/abs/2605.26546](https://arxiv.org/abs/2605.26546)\n- **Abstract**:\n> Mobile graphical user interface (GUI) agents enable AI models to autonomously operate smartphones on behalf of users. However, most existing systems focus primarily on optimizing task accuracy and rely on cloud-hosted models for inference, which introduces privacy concerns and network-dependent latency. As a result, fully on-device deployment of mobile GUI agents remains underexplored. We propose MobileExplorer, a new framework that accelerates on-device inference for vision-based mobile GUI agents via online exploration. The key idea is to exploit the long per-step reasoning time of vision-language models (VLMs) by performing lightweight, parallel exploration of UI elements. During model inference, the agent proactively probes semantically relevant UI elements and records these exploration traces as structured memory. To ensure reliable execution in live mobile environments, we design a two-level rollback mechanism that robustly restores the initial UI state when a fast but naive backtracking strategy fails. The collected exploration traces are then summarized into concise contextual hints and injected into the prompt to enhance the subsequent reasoning step. We evaluate MobileExplorer on multiple off-the-shelf devices using the AndroidWorld benchmark, as well as newly designed, more complex tasks and dynamic on-device environments. MobileExplorer reduces the average number of reasoning steps and end-to-end latency by 23\\%, while maintaining or improving task success rates by up to 5\\%. A video demonstration of MobileExplorer performance in the real world is available at this https URL .\n\n---\n\n## 30. The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence\n- **Authors**: MiniMax : Aili Chen , Aonian Li , Baichuan Zhou , Bangwei Gong , Binyang Jiang , Boji Dan , Changqing Yu , Chao Wang , Cheng Ma , Cheng Zhong , Cheng Zhu , Chengjun Xiao , Chengyi Yang , Chengyu Du , Chenyang Zhang , Chi Zhang , Chuangyi Huang , Chunhao Zhang , Chunhui Du , Chunyu Zhao , Congchao Guo , Da Chen , Deming Ding , Dianjun Sun , Dongyu Zhang , Enhui Yang , Fei Yu , Guang Zheng , Guodong Zheng , Guohong Li , Haichao Zhu , Haigang Zhou , Haimo Zhang , Han Ding , Hao Zhang , Haohai Sun , Haolin Lyu , Haonan Lu , Haoyu Wang , Huajie Shi , Huiyang Li , Jiacheng Chen , Jian Zhang , Jiaqi Zhuang , Jiaren Cai , Jiaxin Pan , Jiayao Li , Jiayuan Song , Jichuan Zhang , Jie Wang , Jihao Gu , Jin Zhu , Jingwei Dong , Jingyang Li , Jingyu Zhang , Jingze Zhuang , Jinhao Tian , Jinli Liu , Jinyi Hu , Jun Tao , Jun Zhang , Junbin Ruan , Junhao Xu , Junjie Yan , Junteng Liu , Junxian He , Kang Xu , Ke Ji , Ke Yang , Kecheng Xiao , Keyu Duan , Keyu Li , Le Han , Letian Ruan , Li Yuan , Lianfei Yu , Liheng Feng , Lijie Mo , Lin Li , Lingye Bao , Lingyu Yang , Lingyuan Zhou , Loki , Lu Chen , Lunbin Ceng , Ming Li , Ming Zhong , Mingliang Tao , Mingyuan Chi , Mujie Lin , Nan Hu , Ningxin Chen , Peiyin Zhu , Peng Gao , Pengcheng Gao , Pengfei Li , Penglin Li , Pengyu Zhao , Qibin Ren\n- **URL**: [https://arxiv.org/abs/2605.26494](https://arxiv.org/abs/2605.26494)\n- **Abstract**:\n> We introduce the MiniMax-M2 series, a family of Mixture-of-Experts language models built around the principle that mini activations can unleash maximum real-world intelligence. The flagship M2 contains 229.9B total parameters with only 9.8B activated per token. Designed end-to-end for agentic deployment, the M2 series rests on three components: (i) agent-driven data pipelines producing large-scale, verifiable trajectories across agentic coding and agentic cowork, each grounded in an executable workspace and an artifact-aligned reward; (ii) Forge, a scalable agent-native RL system that adapts to long-horizon agent trajectories, paired with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and a clean training-inference-agent decoupling that supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step toward self-evolution -- autonomously debugging training runs and modifying its own scaffold. Across M2 through M2.7, this combination translates a mini-activation footprint into frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.\n\n---\n\n## 31. Reasoning, Code, or Both? How Large Language Models Handle Variations in Math Questions\n- **Authors**: Matthew Kutakh\n- **URL**: [https://arxiv.org/abs/2605.26414](https://arxiv.org/abs/2605.26414)\n- **Abstract**:\n> Large Language Models (LLMs) achieve impressive accuracy on mathematical reasoning benchmarks, yet their performance drops when problems are modified with simple changes like different names or numbers. Code execution methods, which let models generate and run Python code instead of reasoning in natural language, have been proposed as a solution, but their effect on reasoning robustness (the ability to maintain accuracy across problem variations) has not been systematically tested. This study evaluates three approaches on 1,000 problems from the GSM-Symbolic dataset: pure reasoning using chain-of-thought (CoT) prompting, single-shot code execution using Program-Aided Language models (PAL), and iterative code execution using Step-by-Step Coding (SBSC). All three were run on paired original and modified problems using Claude Haiku 4.5. CoT was the most robust method, with an accuracy drop of 1.3 percentage points and 1.8% of problems breaking under perturbation. PAL was the least robust at 1.7 percentage points and 3.1% broke, with SBSC falling in between. Although these differences were not statistically significant ($p = .096$), the directional trend was consistent across all measures, suggesting that code execution, whether single-shot or iterative, does not improve reasoning robustness on grade-school-level problem variations.\n\n---\n\n## 32. From Static Context to Calibrated Interactive RL: Mitigating Distribution Shift in Multi-turn Dialogue with Aligned Simulator\n- **Authors**: Xiaohua Wang , Jiakang Yuan , Zisu Huang , Muzhao Tian , Changze Lv , Kaitao Song , Tao Chen , Xiaoqing Zheng\n- **URL**: [https://arxiv.org/abs/2605.26403](https://arxiv.org/abs/2605.26403)\n- **Abstract**:\n> A long-standing goal of the research community is to develop highly interactive LLM-based dialogue agents. Recent research focuses on optimizing policies based on fixed offline logs (Static Context RL) or using a prompt-based simulator (Interactive RL). In this work, we theoretically show that both paradigms are fundamentally limited by context distribution shift--a mismatch between dialogue histories observed during training and those encountered in real conversations. This shift compounds quadratically over turns and severely degrades dialogue quality. Specifically, we attribute this shift to two distinct sources: (i) policy-induced shift, arising from training on static histories rather than self-generated trajectories; and (ii) simulator-induced shift, stemming from discrepancies between simulated and real human behaviors. To address these challenges, we propose Calibrated Interactive RL, a unified framework that couples interactive RL with simulator alignment. By aligning the simulator with human interaction patterns, our approach reduces the sim-to-real gap and mitigates compounding distribution shifts. Experiments across multiple dialogue tasks confirm our theoretical analysis: (i) Interactive RL significantly outperforms the Static Context baseline by mitigating policy distribution shift; and (ii) calibrating simulators with our alignment method further bridges the sim-to-real gap, yielding state-of-the-art downstream performance.\n\n---\n\n## 33. Automatic Layer Selection for Hallucination Detection\n- **Authors**: Xinpeng Wang , William Cao , Andrew Gordon Wilson , Zhe Zeng\n- **URL**: [https://arxiv.org/abs/2605.26366](https://arxiv.org/abs/2605.26366)\n- **Abstract**:\n> Recent studies on hallucination detection have shown that hallucination-related signals are more strongly encoded in intermediate layers than in the final layer of large language models (LLMs). Although a growing body of work has sought to exploit this property for hallucination detection, how to automate the selection of high-performing layers remains underexplored, and principled methods for this purpose are still lacking. To address this gap, we first propose several hypotheses for why such signals emerge in intermediate layers and evaluate corresponding criteria for automatic layer selection across diverse LLM architectures, scales, and tasks, covering both question answering and summarization hallucination detection benchmarks. However, we find that none of these criteria consistently delivers satisfactory performance. We therefore propose a new selection criterion, First Effective Peak of Intrinsic Dimension (FEPoID), which consistently identify optimal or near-optimal layers and outperforms both the aforementioned criteria and existing hallucination detection baselines. FEPoID is training-free and incurs negligible computational overhead. In addition, we study the generation behaviors of LLMs and introduce a simple yet effective truncation strategy, which further amplifies hallucination-related signals and substantially improves overall detection performance. Code is publicly available at this https URL\n\n---\n\n## 34. Managing Uncertainty in LLM-Generated Procedural Knowledge for Virtual Laboratory Planning\n- **Authors**: Polychronis Karpodinis , Dimitris Kalles\n- **URL**: [https://arxiv.org/abs/2605.26333](https://arxiv.org/abs/2605.26333)\n- **Abstract**:\n> Educational virtual laboratories can make experimental training more scala-ble, adaptive, and accessible, especially when students have limited access to physical laboratory facilities. However, authoring new simulated laboratory procedures remains costly: educators must describe new equipment, define how instruments and materials interact, and specify valid procedural flows that can be executed or assessed inside the virtual environment. Large lan-guage models can assist in this authoring process by generating detailed ex-perimental procedures, but their output should not be treated as directly exe-cutable plans. They may omit necessary actions, arrange steps in the wrong order, or produce instructions that are logically incorrect or incompatible with the laboratory equipment. This paper presents a prototype framework for managing uncertainty in LLM-generated procedural knowledge for virtu-al laboratory planning. The framework aims to reduce procedural uncertainty by using structured domain representations and uncertain LLM-generated state-transition samples to extract candidate procedural rules, transform them into explicit and inspectable constraints, and use them to repair uncertain procedural steps. Although the motivating domain refers to educational vir-tual laboratories, the underlying problem is more general: managing uncer-tain procedural knowledge for action planning in structured interactive envi-ronments. We illustrate the approach in a virtual laboratory domain involving laboratory instruments, containers, tools, and material-transfer actions.\n\n---\n\n## 35. OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling\n- **Authors**: Adam Bawatneh , Sagar Sapkota , Amrit Singh Bedi , Santu Karmaker , Mubarak Shah\n- **URL**: [https://arxiv.org/abs/2605.26322](https://arxiv.org/abs/2605.26322)\n- **Abstract**:\n> Theory of Mind (ToM), the ability to infer others' knowledge, intentions, and emotions, is commonly evaluated in large language models (LLMs) using end-point question answering, where performance is judged solely by the final answer to a social reasoning query. This paradigm obscures whether the model actually constructs the underlying mental-state representations required for robust reasoning, particularly in scenarios involving divergent, evolving, or mistaken beliefs. In order to address this research gap, we introduce OmniToM, a benchmark that directly evaluates these representations by requiring explicit modeling of belief structures for all relevant actors within a narrative. These structures are composed of belief propositions: minimal statements of what an actor takes to be true about the world or another actor's mental state, allowing knowledge, intentions, emotions, and false beliefs to be analyzed in a common format. Models are evaluated in two stages: Stage 1: Belief Extraction, which extracts from the story the beliefs relevant to its social dynamics, and Stage 2: Belief Labeling, which assigns each belief a seven-dimensional schema label covering recursive order, truth status, knowledge access, explicitness, content type, mental source, and context. Built from 895 stories from the existing ToMBench story corpus and augmented with 22,343 labeled belief propositions, OmniToM uses a human-calibrated LLM-assisted annotation pipeline. Across diverse models in zero-shot evaluation, OmniToM reveals an actor-specific belief-tracking bottleneck: current LLMs struggle with the knowledge-access and representational decisions required to transform narrative facts into actors' beliefs and shared mental states.\n\n---\n\n## 36. Experiments in Agentic AI for Science\n- **Authors**: Judy Fox , Geoffrey Fox\n- **URL**: [https://arxiv.org/abs/2605.26305](https://arxiv.org/abs/2605.26305)\n- **Abstract**:\n> This paper details two novel frameworks for developing autonomous, agentic AI in scientific workflows. Both systems leverage a hybrid Local Body, Remote Brain architecture via Google Colab, utilizing Python-based local orchestrators to invoke large language model (LLM) cloud backends. The first agent, DeepTS/DeepCollector, automates the large-scale curation, extraction, and deduplication of time-series datasets. The second, DeepScribe, is an autonomous presentation analyzer that converts visually dense, mathematically complex physics lectures into structured scientific reports. Through practical systems engineering-such as granular attribute extraction (Cellular RAG), remote data inspection, and distributed concurrency controls-we demonstrate how agentic AI can overcome the context and reasoning limitations of current state-of-the-art systems to rigorously support scientific workflows. Finally, we outline a generalization of DeepTS to support deep knowledge graphs and discuss the application of this conceptual approach to high-energy physics (DeepQCD).\n\n---\n\n## 37. Personalizing Embodied Multimodal Large Language Model Agents over Long-term User Interactions\n- **Authors**: Jeongeun Lee , Chanyoung Park , Dongha Lee\n- **URL**: [https://arxiv.org/abs/2605.26256](https://arxiv.org/abs/2605.26256)\n- **Abstract**:\n> Multimodal large language model (MLLM)-based embodied agents have shown strong potential for solving complex tasks in physical environments. However, personalized assistance requires more than following generic instruction or recognizing object categories. In real-world scenarios, the intended target is often specified only implicitly through prior interactions, requiring agents to leverage personalized context accumulated over time. In this work, we propose POLAR, a multiomodal memory-augmented framework for personalized embodied agents over long-term user interactions. POLAR organizes prior interactions into a multimodal knowledge graph that captures semantic memory for personalized context and visual concepts, and episodic memory for embodied experiences such as agent trajectories. To execute embodied tasks, POLAR retrieves relevant memories to interpret the current request and guide task execution. We evaluate POLAR across multiple MLLM backbones and diverse evaluation scenarios to study the role of memory in long-term personalization. Results show that the proposed memory mechanism consistently improves performance by enabling more effective use of information accumulated over prior interactions. The gains are especially pronounced when the agents are required to reason across multiple interactions, perform multi-hop inference, or tracking updates in user-specific context over time.\n\n---\n\n## 38. Can LLMs Introspect? A Reality Check\n- **Authors**: Shashwat Singh , Tal Linzen , Shauli Ravfogel\n- **URL**: [https://arxiv.org/abs/2605.26242](https://arxiv.org/abs/2605.26242)\n- **Abstract**:\n> Can large language models detect and report their own internal states? A number of studies have argued that the answer to this question is yes. We argue, based on lessons from human metacognition research, that this conclusion may be premature: to be convinced of this conclusion we need to distinguish genuine introspection from pattern matching based on surface-level cues. Furthermore, we argue that behavioral evidence alone is inherently insufficient to establish strong introspective claims. We re-examine two recently introduced evaluation paradigms in light of this consideration. In the first paradigm, models are expected to detect whether their internal states have been tampered with. We find that models cannot reliably distinguish such interventions on their internal states from manipulations of the input, suggesting that their success in the original studies reflects their ability to detect anomalies more generally, as opposed to interventions on their internal states in particular. In the second paradigm we examine, models are tasked with predicting labels derived from their own hidden states. Here, we find that classifiers that only have access to the input achieve equivalent performance to the model's own in-context predictions, indicating that the original results do not conclusively demonstrate that the model has privileged access to its internal representations. We further introduce a relabeled control setting, where models cannot rely on the semantics of the task to solve it, and instead must rely on the internal representation; models perform closer to chance on this better-controlled version of the task. Taken together, these results indicate that current evidence is insufficient to establish that LLMs display metacognitive monitoring.\n\n---\n\n## 39. LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding\n- **Authors**: Shihao Wang , Shilong Liu , Yuanguo Kuang , Xinyu Wei , Yangzhou Liu , Zhiqi Li , Yunze Man , Guo Chen , Andrew Tao , Guilin Liu , Jan Kautz , Lei Zhang , Zhiding Yu\n- **URL**: [https://arxiv.org/abs/2605.27365](https://arxiv.org/abs/2605.27365)\n- **Abstract**:\n> Vision-language models (VLMs) commonly formulate visual grounding and detection as a coordinate-token generation problem, serializing each 2D box into multiple 1D tokens that are learned and decoded largely independently. This token-by-token decoding mismatches the coupled structure of box geometry and creates a practical inference bottleneck due to strictly sequential generation. We introduce LocateAnything, a unified generative grounding and detection framework based on Parallel Box Decoding (PBD). By decoding geometric elements such as bounding boxes and points as atomic units in a single step, LocateAnything preserves intra-box geometric coherence and unlocks substantial parallelism. We show that PBD improves both decoding throughput and localization accuracy. We further develop a scalable data engine and curate LocateAnything-Data, a large-scale dataset with more than 138 million training samples, substantially increasing data diversity for high-precision localization. Extensive evaluations show that LocateAnything advances the speed-accuracy frontier, achieving significantly higher decoding throughput while improving high-IoU localization quality across diverse benchmarks. The results highlight the complementary benefits of Parallel Box Decoding and large-scale training data in enabling efficient and precise unified visual grounding and detection.\n\n---\n\n## 40. GENESIS: Harnessing AI Agents for Autonomous 6G RAN Synthesis, Research, and Testing\n- **Authors**: Tamerlan Aghayev , Maxime Elkael , Michele Polese , Minh Dat Nguyen , Gabriele Gemmi , Andrea Lacava , Ali Saeizadeh , Reshma Prasad , Paolo Testolina , Angelo Feraudo , Soumendra Nanda , Pedram Johari , Salvatore D'Oro , Tommaso Melodia\n- **URL**: [https://arxiv.org/abs/2605.27360](https://arxiv.org/abs/2605.27360)\n- **Abstract**:\n> Cellular research and development (R&D) is throttled by six structural processes that each consume months of manual engineering work per iteration: (i) synthesizing new features from standards or research papers into production code; (ii) conformance and interoperability testing; (iii) hardening against field anomalies and diverse deployment environments; (iv) data-driven optimization of network functionalities; (v) discovering and prototyping novel waveforms, functionalities, and capabilities for future standards; and (vi) securing the stack against vulnerabilities. Although Large Language Models (LLMs) have compressed comparable R&D work in general software engineering from days to minutes, their known pitfalls worsen on Radio Access Network (RAN) use cases: they hallucinate Application Programming Interfaces (APIs) and mis-read specifications, which kills interoperability of RAN components at the first mistake, and they heavily rely on simulations for designing algorithms, which is notorious for breaking when transferred to real hardware. To address these challenges, we present GENESIS, an agentic Artificial Intelligence (AI) framework that converts intents (e.g., a specification clause, a telemetry anomaly, or a research hypothesis) into solutions validated with over-the-air experiments, fed back into a persistent knowledge base. GENESIS is built on three composable primitives (agents, skills, hooks) and a knowledge layer (SYNAPSE) that doubles as the source of ground truth and the recipient of every artifact the framework produces, making capabilities compound across runs.\n\n---\n\n## 41. MobileMoE: Scaling On-Device Mixture of Experts\n- **Authors**: Yanbei Chen , Hanxian Huang , Ernie Chang , Jacob Szwejbka , Digant Desai , Zechun Liu , Vikas Chandra , Raghuraman Krishnamoorthi\n- **URL**: [https://arxiv.org/abs/2605.27358](https://arxiv.org/abs/2605.27358)\n- **Abstract**:\n> Mixture-of-Experts (MoE) has become the de facto architecture for hundred-billion-parameter language models, yet its advantages at sub-billion scales for on-device deployment remain largely unexplored. To close this gap, we present MobileMoE, a family of on-device MoE language models with sub-billion active parameters (0.3-0.9B active and 1.3-5.3B total) that establish a new Pareto frontier for on-device LLMs. We first formulate an on-device MoE scaling law that jointly optimizes MoE architecture under mobile memory and compute constraints, identifying an on-device sweet spot - moderate sparsity with fine-grained and shared experts - that is simultaneously memory and compute-optimal. Building on the derived architectures, we train MobileMoE with a four-stage recipe covering pre-training, mid-training, instruction fine-tuning, and quantization-aware training, all on open-source datasets. Across 14 benchmarks, MobileMoE matches or exceeds leading on-device dense LLMs with 2-4$\\times$ fewer inference FLOPs, and matches or surpasses the state-of-the-art MoE OLMoE-1B-7B with up to 60% fewer parameters. To bridge the last mile to mobile deployment, we provide the first efficient MoE inference on commodity smartphones with comprehensive on-device profiling. At comparable INT4 weight memory, MobileMoE-S delivers $1.8$-$3.8\\times$ faster prefill and $2.2$-$3.4\\times$ faster decode than the dense baseline MobileLLM-Pro.\n\n---\n\n## 42. Guiding LLM Post-training Data Engineering with Model Internals from Sparse Autoencoders\n- **Authors**: Yi Jing , Zao Dai , Jinwu Hu , Zijun Yao , Lei Hou , Juanzi Li , Xiaozhi Wang\n- **URL**: [https://arxiv.org/abs/2605.27354](https://arxiv.org/abs/2605.27354)\n- **Abstract**:\n> Model internals encode rich information about how a large language model (LLM) processes its training data; however, post-training data engineering largely relies on external signals and ignores rich intrinsic signals lying in model internals. We propose SAERL, a data engineering framework for LLM reinforcement learning (RL). It models three intrinsic data properties: diversity, difficulty, and quality, using model internals extracted with Sparse Autoencoder (SAE), an advanced mechanistic interpretability tool. Each property grounds a concrete data engineering operation: SAE-space clustering with moderate batch mixing for batch diversity control, a difficulty proxy for easy-to-hard curriculum ordering, and a quality probe for data filtering. SAERL improves average accuracy by 3.00% over vanilla GRPO and reaches target accuracy with 20% fewer training steps on Qwen2.5-Math-1.5B, with consistent gains across model scales and RL algorithms. Experiments show that SAE transfers effectively across model families and scales, serving as a lightweight and reusable data engineering tool. These results demonstrate that model internals are a powerful and practical source of signals for post-training data engineering.\n\n---\n\n## 43. EdgeFlow: Edge-Map Augmented VLM-Based Flowchart Processing for Industrial Requirements Engineering\n- **Authors**: Zhifei Dou , Shabnam Hassani , Ou Wei\n- **URL**: [https://arxiv.org/abs/2605.27332](https://arxiv.org/abs/2605.27332)\n- **Abstract**:\n> Flowcharts are widely used in industrial requirements, but usually remain embedded as static images. Vision Language Models (VLMs) show promise in the conversion of these flowcharts into machine-readable models for RE activities, yet, when directly applied to flowchart conversion, they often fail on topology-critical visual details. To address this, we propose EdgeFlow that augments a VLM's original input with a deterministically extracted Canny edge map-acting as a structural prior-to improve flowchart-to-Mermaid conversion, without requiring annotated training data or domain-specific model fine-tuning. We evaluate EdgeFlow on IndusReqFlow, a dataset sourced from real-world requirements. Compared with off-the-shelf VLMs, EdgeFlow improves node-level F1 by 17.39 percentage points and edge-level F1 by 16.94 percentage points. At the path level, EdgeFlow improves path F1 by 11.06 percentage points, enabling better support for model-based testing. These results demonstrate that EdgeFlow provides a practical, training-free means to improve topology-preserving flowchart-to-Mermaid conversion for industrial RE. Cross-dataset evaluation results on a public synthetic benchmark show no significant improvement; this highlights the need for diverse benchmarks incorporating industrial data for the comprehensive evaluation of future VLM-based RE tools.\n\n---\n\n## 44. It's Not Always Sycophancy: Measuring LLM Conformity as a Function of Epistemic Uncertainty\n- **Authors**: Kevin H. Guo , Chao Yan , Avinash Baidya , Katherine Brown , Xiang Gao , Juming Xiong , Zhijun Yin , Bradley A. Malin\n- **URL**: [https://arxiv.org/abs/2605.27288](https://arxiv.org/abs/2605.27288)\n- **Abstract**:\n> Large language models (LLMs) are known to abandon their initial stance to conform to user pushback. While prior research largely attributes this behavior to sycophancy learned during reinforcement learning from human feedback, we hypothesize that conformity is also driven by a model's epistemic uncertainty at inference time. In this paper, we introduce MUSE, a two-stage evaluation framework to disentangle the mechanisms driving LLM conformity. Specifically, MUSE maps a model's epistemic uncertainty in responding to a query against its likelihood to yield to user pushback in a subsequent turn. We demonstrate that the mechanisms driving conformity extend beyond sycophancy alone. Specifically, we characterize two distinct factors that jointly drive conformity: sycophantic conformity, where a model aligns with user pushback even with absolute certainty in its initial response, and uncertainty-driven conformity, where a model's likelihood for conformity increases alongside its uncertainty. Furthermore, we conduct ablation studies to demonstrate that both sycophantic conformity and uncertainty-driven conformity grow with 1) the LLM's perceived expertise of the user and 2) the plausibility of the user's suggestions. More broadly, MUSE informs more targeted intervention strategies by distinguishing alignment-induced sycophancy and training-corpora-driven uncertainty.\n\n---\n\n## 45. Lost in Sampling: Assessing Lexical Reachability in LLMs via the Word Coverage Score (WCS)\n- **Authors**: Samer Awad , Javier Conde , Carlos Arriaga , Tairan Fu , Javier Coronado-Blázquez , Pedro Reviriego\n- **URL**: [https://arxiv.org/abs/2605.27268](https://arxiv.org/abs/2605.27268)\n- **Abstract**:\n> Modern Large Language Models (LLMs) are often criticized for producing repetitive and homogeneous text, despite possessing vast latent vocabularies. While previous research has focused on model knowledge and training data, we investigate the role of decoding mechanics in suppressing linguistic diversity. We introduce the Word Coverage Score (WCS), a metric that quantifies the extent to which contextually appropriate human vocabulary is mathematically pruned by standard sampling filters (e.g., Top-$p$, Top-$k$, and Min-$p$). Rather than assessing static knowledge, the WCS measures the lexical survival rate of low-frequency, high-information human words as a function of sampling parameters. By auditing open-weight models on human-authored corpus fragments, we identify which logical lexical choices are rendered unreachable by the decoder, even when they reside within the probability space. Our results provide quantitative evidence that industry-standard sampling defaults act as unintended censorship mechanisms, smoothing the unique textures of human expression into a homogenized discourse. The WCS offers a rigorous framework for optimizing the trade-off between text coherence and lexical richness, providing a diagnostic tool for preserving the diversity of human language in generative models.\n\n---\n\n## 46. Pair-In, Pair-Out: Latent Multi-Token Prediction for Efficient LLMs\n- **Authors**: Wenhui Tan , Minghao Li , Xiaoqian Ma , Siqi Fan , Xiusheng Huang , Liujie Zhang , Ruihua Song , Weihang Chen\n- **URL**: [https://arxiv.org/abs/2605.27255](https://arxiv.org/abs/2605.27255)\n- **Abstract**:\n> Long chain-of-thought reasoning has made autoregressive decoding the dominant inference cost of modern large language models. Existing methods target either the input side (latent compression) or the output side (speculative decoding and multi-token prediction, MTP), but the two lines of work have been pursued independently. Moreover, output-side methods must incur an expensive verifier pass to validate the unreliable draft tokens predicted by MTP. To address these issues, we propose \\textbf{Pair-In, Pair-Out (PIPO)}, which unifies both sides by viewing a latent compressor and an MTP head as mirror-image operations: the compressor folds two input tokens into one latent representation, while the MTP head unfolds one hidden state into one additional output token. To remove the verifier cost without sacrificing reliability, PIPO trains a lightweight confidence head that decides whether draft tokens should be accepted. We observe that On-Policy Distillation (OPD) naturally matches the rejection-sampling criterion of speculative decoding, so the confidence head can be trained alongside OPD with negligible extra cost. Experiments on AIME 2025, GPQA-Diamond, LiveCodeBench v6, and LongBench v2 with Qwen3.5-4B and 9B backbones show that PIPO improves pass@4 over regular decoding by up to $+7.15$ points, while delivering up to $2.64\\times$ first-token-latency and $2.07\\times$ per-token-latency speedups.\n\n---\n\n## 47. Qiskit QuantumKatas: Adapting Microsoft's Quantum Computing exercises for LLM evaluation\n- **Authors**: Juan Cruz-Benito , Ismael Faro\n- **URL**: [https://arxiv.org/abs/2605.27210](https://arxiv.org/abs/2605.27210)\n- **Abstract**:\n> We adapt Microsoft's QuantumKatas -- a well-established quantum computing curriculum -- from Q# to Qiskit, the most widely-adopted quantum computing framework, and package it with an evaluation framework for systematic LLM assessment. The resulting benchmark comprises 350 tasks across 26 categories, spanning fundamental gates through advanced algorithms (Grover's, Simon's, Deutsch-Jozsa), error correction, key distribution, and quantum games. Each task includes a natural language prompt, canonical solution, and deterministic test verification via classical circuit simulation. By building on the QuantumKatas' proven pedagogical design rather than creating tasks from scratch, we inherit a principled difficulty progression and comprehensive concept coverage while contributing the framework adaptation, evaluation infrastructure, and empirical analysis. We evaluate 16 LLMs across 7 prompting configurations -- a total of 39,200 model runs -- to demonstrate the benchmark's utility. Three key findings emerge: (1) the benchmark effectively differentiates model capabilities, with best-configuration pass rates ranging from 32.3% to 83.1% and a 26.1 pp average gap between frontier and open-source models; (2) models perform well at implementing known algorithms (SimonsAlgorithm 82.1%, BasicGates 81.6%) but struggle with problem encoding (SolveSATWithGrover 34.4%, DistinguishUnitaries 40.0%); and (3) chain-of-thought prompting shows a modestly bimodal effect -- it is the best strategy for three models (two of them explicitly reasoning-tuned per vendor documentation) but degrades performance for the rest, leaving it mid-pack in aggregate (56.3% mean) behind few-shot-5 (57.8%). We release the benchmark, evaluation framework, and baseline results to support research on LLM capabilities in quantum computing.\n\n---\n\n## 48. Generative Animations: A Multi-Model Pipeline for Prompt-Driven Motion Synthesis\n- **Authors**: Mannat Khurana , Sanyam Jain , Rishav Agarwal\n- **URL**: [https://arxiv.org/abs/2605.27203](https://arxiv.org/abs/2605.27203)\n- **Abstract**:\n> Animation elevates digital documents into immersive experiences, yet creating custom motion paths remains cumbersome, requiring designers to manually select presets, plot Bézier points, and configure timing properties. We introduce Generative Animations, a system that transforms natural language prompts into production-ready animations. By chaining Large Language Models (LLMs) for semantic parsing with the Segment Anything Model (SAM) for visual grounding, our pipeline automatically generates motion paths that respect scene geometry, handle depth-based occlusions, and honor 3D perspective transforms. We demonstrate the system through three use cases: contour-following trajectories, orbital animations with z-order awareness, and perspective-aligned motion on transformed objects.\n\n---\n\n## 49. Learning When to Think While Listening in Large Audio-Language Models\n- **Authors**: Zhiyuan Song , Weici Zhao , Yang Xiao , Suhao Yu , Cheng Zhu , Jiatao Gu\n- **URL**: [https://arxiv.org/abs/2605.27190](https://arxiv.org/abs/2605.27190)\n- **Abstract**:\n> Recent advances in Large Audio-Language Models (LALMs) have made real-time, streaming spoken interaction increasingly practical. In this setting, reasoning quality and responsiveness are tightly coupled: delaying reasoning until the speech endpoint can improve answer quality but moves deliberation into user-visible response delay, while answering too early risks committing before decisive evidence arrives. We introduce a learnable wait-think-answer control formulation for LALMs. Motivated by the incremental nature of human conversation, the controller decides under partial audio evidence when to wait, when to externalize a compact reasoning update, and when to answer. Using Qwen2.5-Omni-7B as the base model, we construct aligned wait-think-answer traces from spoken reasoning data, train the controller with supervised fine-tuning (SFT), and then apply Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO). The reward combines answer correctness, action validity, update timing, latency synchronization, reasoning quality, and chain consistency, optimizing the complete wait-think-answer trajectory and not the final answer alone. On a six-task synthetic spoken reasoning question answering (SRQA) benchmark, the six-reward DAPO controller improves the row-weighted accuracy from 67.6% to 70.3% while reducing post-endpoint final-think length by 14% under the same Qwen deployment harness. On a 186-item human-recorded Real Audio Bench, a transfer check beyond text-to-speech (TTS)-rendered speech, the controller family remains functional: SFT achieves the strongest accuracy, while the six-reward DAPO controller is the only learned variant whose final-think length falls below the base. These results suggest that a streaming model should learn when to make intermediate reasoning explicit during the audio stream.\n\n---\n\n## 50. LitSeg: Narrative-Aware Document Segmentation for Literary RAG\n- **Authors**: Ruikang Zhang , Zhanni Chen , Yiqiao Cai , Qi Su\n- **URL**: [https://arxiv.org/abs/2605.27156](https://arxiv.org/abs/2605.27156)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by incorporating external knowledge, particularly for long-tail domains such as literary works. However, the critical step of document segmentation in RAG remains largely underexplored. Existing strategies are typically semantically blind and overlook the complicated narrative structures of literary works, often resulting in fragmented plots and unclear references that severely hinder retrieval and generation performance. To address this, we propose LitSeg, a novel narrative-theory-guided segmentation framework. By employing multi-stage prompting, LitSeg explicitly extracts valid events, untangles narrative threads, clarifies narrative structures, and locates turning points to inform segmentation. To alleviate the computational overhead of multi-stage inference with large-scale models, we further introduce LitSeg-Lite, a lightweight single-pass chunker fine-tuned on LitSeg-generated data via a two-stage training strategy, distilling the complex process into a single inference pass. Extensive experiments demonstrate that with structurally independent text chunks, our methods significantly improve retrieval accuracy and context relevance over baselines, ultimately enhancing downstream QA performance, while ablation studies validate the efficacy of narratological guidance and data distillation.\n\n---\n\n## 51. DEI: Diversity in Evolutionary Inference for Quality-Diversity Search\n- **Authors**: John Donaghy , Shikhar Rastogi\n- **URL**: [https://arxiv.org/abs/2605.27130](https://arxiv.org/abs/2605.27130)\n- **Abstract**:\n> We present DEI: Diversity in Evolutionary Inference, a distributed Quality-Diversity (QD) search framework that assigns heterogeneous large language models (LLMs) as mutation operators across peer nodes communicating with non-blocking collective operations. Unlike homogeneous parallel search, which replicates a single model's inductive biases across all workers, DEI treats each LLM's distinct creative prior as a complementary source of behavioral novelty. Extending the Digital Red Queen framework with DEI, nodes share local optimal solutions at the end of each round to seed the next round's population. This creates cross-model adversarial pressure that drives robustness beyond intra-model self-play. Evaluated on the Core War domain, a competitive programming benchmark in which Redcode warrior programs battle inside a simulated machine, a four-node heterogeneous ensemble (GPT-5.4-mini, Claude Sonnet 4.6, GPT-5.2, and Claude Haiku 4.5) achieves 124 percent higher merged-archive QD-Score (45.90 vs. 20.46) and 28 percent higher coverage (80.6 percent vs. 63.0 percent of cells) than a single-node baseline at equal total LLM-call budget. The heterogeneous ensemble also outperforms an equally-budgeted homogeneous ensemble on QD-Score, coverage, and held-out solution generality across all four model families. These results provide the first empirical evidence that model diversity, not merely parallelism, is the key driver of gain in distributed LLM-based QD search.\n\n---\n\n## 52. ReMoE: Boosting Expert Reuse through Router Fine-Tuning in Memory-Constrained MoE LLM Inference\n- **Authors**: Xiongwei Zhu , Xiaojian Liao , Tianyang Jiang , Yusen Zhang , Liang Wang , Limin Xiao\n- **URL**: [https://arxiv.org/abs/2605.27081](https://arxiv.org/abs/2605.27081)\n- **Abstract**:\n> Fine-grained Mixture-of-Experts (MoE) models sparsely activate only a subset of experts per token, reducing activated computation while maintaining high model capacity. However, in memory-constrained inference scenarios, only a small set of experts can be cached. Experts not in the cache must be fetched from slow external storage (e.g., UFS), leading to frequent evictions and substantial I/O overhead. We propose ReMoE, a router fine-tuning framework designed to boost token-wise expert reuse. ReMoE biases the router toward recently selected experts, producing temporally stable routing that better matches cache locality constraints. By increasing short-horizon expert reuse, ReMoE reduces expert fetches from storage without adding inference-time computation. Experiments on DeepSeek and Qwen models show that ReMoE improves expert reuse by 26% while maintaining downstream task performance. Real-system evaluations further confirm these benefits, improving output throughput by 8.4% under vLLM GPU-CPU expert offloading and reducing TPOT by 43.6-49.8% under this http URL on Jetson Orin NX, corresponding to a 1.77-1.99$\\times$ decode speedup across diverse workloads. Checkpoints and usage instructions are available at this https URL .\n\n---\n\n## 53. E3: Issue-Level Backtesting for Automated Research Critique\n- **Authors**: Yashwardhan Chaudhuri , Sanyam Jain , Paridhi Mundra\n- **URL**: [https://arxiv.org/abs/2605.27072](https://arxiv.org/abs/2605.27072)\n- **Abstract**:\n> We present E3, an automated review assistant that augments reviewers and engineering teams by identifying decision-relevant technical concerns in research papers. For each concern, E3 reports its nature, its location, its bearing on the contribution, and the analysis or evidence that would resolve it, covering unsupported claims, missing ablations, weak baselines, hidden assumptions, threats to validity, and leakage risks. To evaluate E3 without contamination confounds we adopt an issue-level backtesting protocol: the corpus is restricted to papers postdating the training cutoff of every automated source, and for each paper a meta-judge that observes only anonymised reviews labels every issue-source pair as Caught, Partial, or Missed. Applied to 100 ICLR 2026 papers and 4598 judged issue rows, comparing E3 against the ICLR human reviews and two prompt-matched LLM baselines built on gpt-5.4 from OpenAI and claude-opus-4-6 from Anthropic, with meta-judge gpt-5.5, E3 attains the highest recall on every aggregate metric. Partial-inclusive recall reaches 90.2 percent, which is 15.5 points over GPT, 17.1 points over Claude, and 29.2 points over the human reviews, and strict recall preserves the ordering at 65.8 percent. On concerns raised by the human reviewers, E3 recovers 89.6 percent; on concerns the human reviewers missed it surfaces 1635 additional rows admitted into the judged union, 406 above the next-best source. Corpus, baseline prompts, judge prompt template, and evaluation code are released.\n\n---\n\n## 54. QUACK: Questioning, Understanding, and Auditing Communicated Knowledge in Multimodal Social Deduction Agents\n- **Authors**: Ye Yuan , Rui Song , Weien Li , Zeyu Li , Haochen Liu , Xiangyu Kong , Changjiang Han , Yonghan Yang , Zichen Zhao , Zixuan Dong , Fuyuan Lyu , Bowei He , Haolun Wu , Jikun Kang , Xue Liu\n- **URL**: [https://arxiv.org/abs/2605.27068](https://arxiv.org/abs/2605.27068)\n- **Abstract**:\n> Social deduction games have become a popular testbed for probing reasoning, deception, coordination, and belief modeling in Large Language Model (LLM) agents. However, most environments are scored only by game outcomes such as win rates and largely remain to text-only interaction, making it difficult to tell whether an agent's language is actually grounded in what it perceived and did, or to identify the failure modes underlying its behavior. To address this gap, we introduce QUACK, an open-source environment and evaluation framework for auditing the grounding of agent language in multimodal social reasoning. QUACK evaluates agents at three levels: game outcomes, behavioral trajectories, and utterance-level consistency. Its core Statement Verification Pipeline reconstructs each agent's ground-truth trajectory from engine logs and checks every discussion claim against it, automatically flagging spatial hallucination, unsupported accusation, deception collapse, and language-action inconsistency. Evaluating three frontier VLMs in both homogeneous and cross-model adversarial settings, we find that even the strongest agent hallucinates 15.1% of its verifiable spatial claims and makes over half of its accusations without grounded evidence. We release the full engine, evaluation framework, toolkit, and logs at this https URL .\n\n---\n\n## 55. ConVer: Using Contracts and Loop Invariant Synthesis for Scalable Formal Software Verification\n- **Authors**: Muhammad A. A. Pirzada , Weiqi Wang , Yiannis Charalambous , Konstantin Korovin , Lucas C. Cordeiro\n- **URL**: [https://arxiv.org/abs/2605.27051](https://arxiv.org/abs/2605.27051)\n- **Abstract**:\n> Formal verification of large C programs is impeded by state-space explosion: Bounded Model Checking (BMC) tools must encode the entire state space up to the predetermined bound by unrolling all nested constructs. We present ConVer, a top-down compositional verification tool. Given a C program with a top-level assertion, ConVer decomposes verification top-down: it uses a large language model (LLM) to synthesise function contracts from the system property, then alternates system-level and function-level checks in a CEGAR-CEGIS loop, refining contracts whenever a check fails via SMART ICE learning. We evaluate ConVer on four benchmark suites of increasing difficulty and against other state-of-the-art (SOTA) tools. On the Frama-C benchmark of 45 simple C programs, ConVer achieves 82-96% verification success across three LLM backends, with 93-95% of converged programs requiring only a single CEGAR-CEGIS iteration. On the X.509 parser benchmark (6~programs) and LF2C-Simple suite (17 programs), ConVer achieves 33-50% and 82-88% success respectively. On the VerifyThis suite of 11 recursive and loop-intensive programs, the Pre-Abstraction strategy achieves 55-64% success. In addition, we present ESBMC-LF a preprocessor tool that converts LF models to C while preserving the properties of the LF files, enabling ConVer to verify them. We transpile the LF Verifier Benchmarks using ESBMC-LF to C; we denote those LF-Hard. We show that ConVer successfully verifies 67% of LF-Hard benchmarks overall.\n\n---\n\n## 56. Tracing Computation Density in LLMs\n- **Authors**: Corentin Kervadec , Iuliia Lysova , Iuri Macocco , Marco Baroni , Gemma Boleda\n- **URL**: [https://arxiv.org/abs/2605.27033](https://arxiv.org/abs/2605.27033)\n- **Abstract**:\n> Transformer-based large language models (LLMs) are comprised of billions of parameters arranged in deep and wide computational graphs, but it is not clear that they exploit their full capacity for all inputs. We introduce the s-Trace method to efficiently estimate the subgraph of size s that best approximates a full model output. With this method, we find the computation in a variety of LLMs to be organized in two distinct phases. A small subgraph mostly composed of early-layer nodes can reconstruct the head of the full model output distribution. Adding further nodes, mostly located in later layers and increasingly consisting of attention heads, leads to incremental refinements in approximating the full output distribution. We find moreover that the amount of necessary computation per input correlates with model uncertainty, and that sparser subgraphs encode shallow statistics, such as unigram frequency. Overall, our results suggest a consistent modular organization in effective LLM computation, with a sparse early-layer core providing a rough prediction that is further refined through denser computations in later layers.\n\n---\n\n## 57. Evaluating the Relevance of Uncertainty Estimators for LLM Hallucination\n- **Authors**: Yedidia Agnimo , Anna Korba , Annabelle Blangero , Nicolas Chesneau , Karteek Alahari\n- **URL**: [https://arxiv.org/abs/2605.27016](https://arxiv.org/abs/2605.27016)\n- **Abstract**:\n> Large language models (LLMs) are prone to hallucinations, i.e., statements unsupported by the input or training data, hindering reliable deployment. In parallel, numerous uncertainty estimation (UE) methods have been proposed to quantify model confidence and are often implicitly treated as proxies for model failure. However, the relationship between uncertainty and hallucinations remains insufficiently characterized. We present a systematic empirical study of the association between uncertainty estimators and hallucinations in LLMs. Rather than assuming this association, we evaluate directly when and to what extent it holds. We consider a diverse set of uncertainty estimators, including information-theoretic, sampling-based, and reflexive estimators, and examine their behavior across hallucination settings. Our experiments cover both intrinsic hallucinations (violations of input faithfulness) and extrinsic hallucinations (unsupported claims relative to training data), using four complementary benchmarks, including RAGTruth and HalluLens. We find that the association is highly variable and often weak, depending on the hallucination type and the LLM under evaluation. These results challenge the use of uncertainty as a direct signal of hallucination and clarify when it provides actionable information.\n\n---\n\n## 58. ReasonOps: A Unified Operational Paradigm for Trustworthy Verified LLM Reasoning\n- **Authors**: Adnan Rashid\n- **URL**: [https://arxiv.org/abs/2605.27014](https://arxiv.org/abs/2605.27014)\n- **Abstract**:\n> Large Language Models (LLMs) have transformed artificial intelligence from primarily generative systems into increasingly capable reasoning agents. Recent advances in theorem proving, autoformalization, symbolic reasoning, and tool-augmented language models demonstrate substantial progress toward machine-assisted formal reasoning. However, current reasoning systems still suffer from hidden logical inconsistencies, hallucinated symbolic transitions, unsupported theorem applications, and limited reliability guarantees. Existing approaches remain fragmented across formal verification, runtime assurance, neuro-symbolic reasoning and trustworthy Artificial Intelligence (AI) research communities. This paper introduces ReasonOps, a unified operational paradigm for trustworthy verified reasoning systems. Inspired by operational ecosystems such as DevOps and MLOps, ReasonOps treats reasoning as a continuously monitored, verifiable, reliability-aware operational process rather than an isolated inference task. The proposed paradigm integrates semantic interpretation, autoformalization, symbolic reasoning, theorem proving, runtime assurance, probabilistic reliability estimation, and adaptive correction into a unified reasoning lifecycle. The paper further presents the ReasonOps architecture, demonstrates its workflow using an autonomous braking system analysis example, and discusses its potential role in future safety-critical autonomous AI systems. We argue that operational reasoning paradigms such as ReasonOps may become foundational infrastructure for next-generation trustworthy AI ecosystems.\n\n---\n\n## 59. Recon: Reconstruction-Guided Reasoning Synthesis for User Modeling\n- **Authors**: Alan Zhu , Mihran Miroyan , Carolyn Wang , Andrew Zhou , Lisa Dunlap , Narges Norouzi , Joseph E. Gonzalez\n- **URL**: [https://arxiv.org/abs/2605.26969](https://arxiv.org/abs/2605.26969)\n- **Abstract**:\n> User modeling aims to use language models (LMs) to mimic an individual's behavior from a corpus of past context-action pairs (e.g., conversation turns), enabling the simulation of users in settings like behavioral science, human-AI collaboration, and market research. Recent approaches augment these corpora with synthesized reasoning traces, typically generated by conditioning on both context and action. However, such conditioning constitutes post-hoc rationalization rather than reasoning: the trace is guaranteed to justify the action, but may not encode the underlying latent causal decision paths. We propose Recon, which uses action reconstruction to score reasoning traces by their predictive power: given a context and candidate reasoning, a reconstruction model predicts the action, and reconstruction fidelity determines reasoning quality. Across four domains, Recon achieves a 54.7% win rate over Backward Synthesis, a standard post-hoc rationalization baseline. Further, we find that training a reasoning synthesis model with rewards derived from Recon improves downstream user modeling performance, achieving a win rate of up to 70.0% over baselines. We further show that Recon-synthesized reasoning transfers across models, and improves user modeling beyond the reconstruction model. Our work demonstrates that post-hoc rationalization is insufficient for reasoning synthesis, and that useful and interpretable reasoning should naturally elicit the action from the context.\n\n---\n\n## 60. Tournament-GRPO: Group-Wise Tournament Rewards for Reinforcement Learning in Open-Ended Long-Form Generation\n- **Authors**: Zixuan Yang , Yiqun Chen , Wei Yang , Erhan Zhang , Zihan Shen , Xiaochi Wei , Yan Gao , Yi Wu , Yao Hu , Jiaxin Mao\n- **URL**: [https://arxiv.org/abs/2605.26958](https://arxiv.org/abs/2605.26958)\n- **Abstract**:\n> Reinforcement learning in open-ended long-form generation is challenging because reliable reference answers and automatic metrics are often unavailable. Existing rubric-based methods typically rely on pointwise LLM-as-a-judge scoring, but absolute scores are difficult to calibrate across complex responses, may provide weak discrimination among same-query rollouts, and can become saturated during optimization. We propose Tournament-GRPO, a group-wise reward framework that converts rubric-guided LLM judgments into relative rewards through repeated multi-round tournaments among same-query rollouts. Tournament-GRPO compares candidates within groups, accumulates tournament outcomes, and normalizes them into group-wise rewards for GRPO training. Experiments on Deep Research Bench show that Tournament-GRPO consistently outperforms existing reward-design baselines, achieving a 4.52-point overall-score improvement over the strongest baseline. Further analyses show that tournament rewards provide a favorable effectiveness--efficiency trade-off and that tournament design affects training dynamics. These results suggest that rubric-guided tournament comparison provides an effective reward signal for reinforcement learning in open-ended long-form generation.\n\n---\n\n## 61. JuICE: A Benchmark for Evaluating LLM-Judge in Identifying Cultural Errors\n- **Authors**: Jiho Jin , Junho Myung , Juhyun Oh , Junyeong Park , Rifki Afina Putri , Sunipa Dev , Vinodkumar Prabhakaran , Alice Oh\n- **URL**: [https://arxiv.org/abs/2605.26955](https://arxiv.org/abs/2605.26955)\n- **Abstract**:\n> As large language models (LLMs) are increasingly deployed to users around the world, they are integrated into everyday tasks across diverse cultural contexts, from drafting personal communications to brainstorming creative ideas. These tasks are inherently cultural: they require contextual appropriateness, symbolic resonance, and tacit cultural expectations that native speakers draw on instinctively, meaning that a response can be factually plausible yet unmistakably wrong to a local reader. Existing cultural benchmarks have treated culture as a flat set of facts via fact verification or norm entailment methods, and have adopted LLM-as-a-Judge without examining whether they can capture such thick cultural errors. To address this gap, we present JuICE (Benchmark for LLM-Judge in Identifying Cultural Errors), a multilingual dataset of 7,470 span-level annotations of cultural and linguistic errors in long-form LLM responses. It covers 1,050 query-response pairs from four countries (the United States, South Korea, Indonesia, and Bangladesh), in both English and their countries' main languages. Using JuICE, we find that even the strongest LLM-judge achieves only an F1 of 0.52 in the erroneous span detection task. Furthermore, LLM-judges consistently miss thick cultural errors that local residents readily identify. Our findings suggest that robust cultural evaluation must move beyond surface-level detection toward frameworks that account for the depth and situatedness of cultural meaning.\n\n---\n\n## 62. Beyond Questions: Evaluating What Large Language Models (Actually) Know\n- **Authors**: Luca Giordano , Simon Razniewski\n- **URL**: [https://arxiv.org/abs/2605.26937](https://arxiv.org/abs/2605.26937)\n- **Abstract**:\n> Parametric knowledge in large language models (LLMs) is a cornerstone of their success, yet remains poorly understood. Existing knowledge benchmarks typically rely on predefined questions (e.g., \"What is the birth date of M.L. King?\"), evaluating only knowledge that benchmark designers explicitly choose to query, a problematic availability bias. In this paper, we introduce open knowledge evaluation, a new paradigm for LLM knowledge benchmarking. Instead of asking narrow questions, it evaluates models on the knowledge they choose to surface in response to open-ended elicitation prompts (e.g., \"Tell me everything you know about M.L. King\"). This shifts the focus from predefined answer retrieval toward characterizing the knowledge models naturally express. We instantiate this paradigm with BeQu (Beyond Questions), a benchmark of 10,000 entities paired with reference corpora for statement verification. Using BeQu, we evaluate a broad range of language models and analyze the effects of reasoning effort, model scale, prompt format, and knowledge domain. Data and leaderboard are available on this work's GitHub repository and at the benchmark's website.\n\n---\n\n## 63. Strategies for Guiding LLMs to Use Software Design Patterns: A Case of Singleton\n- **Authors**: Viktor Kjellberg , Farnaz Fotrousi , Miroslaw Staron\n- **URL**: [https://arxiv.org/abs/2605.26898](https://arxiv.org/abs/2605.26898)\n- **Abstract**:\n> Large Language Models (LLMs) can generate functional source code from natural-language prompts, but often fail to consistently follow higher-level architectural structures or design patterns. Since LLMs are increasingly used in software engineering, their ability to apply established design principles to generated code is crucial to the long-term success of software products. Therefore, the goal of this paper is to identify strategies for guiding LLMs to incorporate design patterns into the generated source code. We designed a computational experiment to evaluate the ability of 13 LLMs to generate code that follows the Singleton design pattern, using four prompting strategies: instructions, binary automated feedback, extensive automated feedback, and extensive feedback with few-shot prompts, in 164 Java coding challenges from HumanEval-X. Our results shows that the optimal strategy to guide LLMs to include design patterns depends heavily on the type of model. Still, overall, iterative binary feedback provides the best alignment with Singleton while preserving or improving the code's functionality. With guiding with instructions, Llama 3.3 generated Singleton classes in 100% of cases and improved code functionality, increasing the number of tests passed by 34.1 percentage points. It achieved a similar result with guidance through instructions and binary feedback. Qwen 3 (8B) increased the alignment with Singleton to 99.2% and the functionality to 58.6% using binary feedback. Our result suggests that even simple strategies can be used to guide LLMs to use design patterns.\n\n---\n\n## 64. Negligible in Size, Significant in Effect: On Scale Vectors in Large Language Models\n- **Authors**: Mingze Wang , Shuchen Zhu , Yuxin Fang , Binghui Li , Kai Shen , Shu Zhong\n- **URL**: [https://arxiv.org/abs/2605.26895](https://arxiv.org/abs/2605.26895)\n- **Abstract**:\n> Normalization layers in modern large language models (LLMs) consist of a deterministic normalization operation and a learnable scale vector. While the normalization operation has been extensively studied, the scale vector remains poorly understood despite its ubiquitous use. In this work, we present a systematic study of scale vectors in LLMs from the perspectives of expressivity, optimization, and architectural structure. First, we show empirically that although scale vectors constitute only a negligible fraction of model parameters, removing them substantially degrades LLM pre-training. Our theory further shows that, in Pre-Norm architectures, scale vectors do not increase expressivity; instead, they improve optimization through a self-amplifying preconditioning effect on subsequent linear mappings. Second, we investigate the role of weight decay for scale vectors. By distinguishing Input-Norm and Output-Norm layers, we theoretically show that weight decay is beneficial for the former but harmful for the latter, due to their distinct roles in optimization and expressivity. Third, motivated by this understanding, we propose three lightweight and complementary improvements to scale vectors: branch-specific heterogeneity, improved placement around linear mappings, and magnitude-direction reparameterization. Both theory and experiments show that each improvement yields consistent gains. Finally, we combine these improvements into a unified scale-vector strategy and evaluate it through extensive LLM pre-training experiments on dense and mixture-of-experts models ranging from 0.12B to 2B parameters, across multiple optimizers and learning rate schedules, under industrial-scale token budgets. The unified strategy consistently achieves lower terminal loss than well-tuned baselines and exhibits more favorable scaling behavior, while adding negligible parameter and computational overhead.\n\n---\n\n## 65. GeoFaith: A Spatio-Temporal Dual View of Faithful Chain-of-Thought\n- **Authors**: Weijiang Lv , Wentong Zhao , Jiayu Wang , Yuhao Wu , Jiaheng Wei , Xiaobo Xia\n- **URL**: [https://arxiv.org/abs/2605.26893](https://arxiv.org/abs/2605.26893)\n- **Abstract**:\n> Chain-of-Thought (CoT) reasoning has advanced large language models (LLMs), but outcome-based supervision leads to pervasive post-hoc rationalization, producing plausible yet unfaithful reasoning chains. Most prior faithfulness assessment methods are either unscalable, expensive, or unreliable. We propose GeoFaith, a spatio-temporal framework that leverages latent geometric structure and entropy dynamics to diagnose and enforce faithful reasoning. We develop a scalable bootstrapping pipeline expanding step-level annotations from 1k to 20k samples across four domains, train an 8B faithfulness detector outperforming GPT-5 on standard benchmarks, and design a faithfulness-aware reinforcement learning framework jointly optimizing outcome correctness, process faithfulness, and trajectory consistency. Experiments show the proposed method achieves superior performance on both faithfulness detection and downstream reasoning, producing shorter, more interpretable chains without sacrificing accuracy. Our code will be made available publicly.\n\n---\n\n## 66. Knowledge Graphs as the Missing Data Layer for LLM-Based Industrial Asset Operations\n- **Authors**: Madhulatha Mandarapu , Sandeep Kunkunuru\n- **URL**: [https://arxiv.org/abs/2605.26874](https://arxiv.org/abs/2605.26874)\n- **Abstract**:\n> LLM-based agents for industrial asset operations show limited accuracy when reasoning over flat document stores. AssetOpsBench (KDD 2026) establishes that GPT-4 agents achieve 65% on 139 industrial maintenance scenarios backed by CouchDB, YAML, and CSV. It compares LLM orchestration paradigms (Agent-As-Tool vs Plan-Execute) on a fixed data layer; we ask a complementary, orthogonal question: how much does the data model behind the tools affect agent performance? Building on the same scenarios, we introduce a knowledge graph layer (781 nodes, 955 edges, 16 relationship types) and evaluate three architectures: (1) deterministic graph handlers (no LLM) at 99% (137/139); (2) LLM-generated Cypher over the graph at 82-83% with the same GPT-4 model the baseline uses; and (3) the original tool-augmented LLM baseline at 65% (91/139, matching the published KDD 2026 leaderboard ceiling). Our key finding is inverted LLM usage: rather than asking the LLM to reason over raw data, we ask it to generate structured queries from a typed schema. The graph executes deterministically. We additionally contribute 40 graph-native scenarios (multi-hop dependency, vector similarity, PageRank criticality), and evaluate against the expanded HuggingFace AssetOpsBench release (467 scenarios, 6 domains), where deterministic handlers achieve 100% (467/467) with average score 0.848. These results suggest that for structured operational domains, the data layer -- not the LLM orchestration -- is the primary bottleneck, and that knowledge graphs serve as an integration layer between raw industrial data and LLM-based reasoning.\n\n---\n\n## 67. The Strongest Teacher Is Not Always the Best Teacher: Student-Centric Answer Selection\n- **Authors**: Zhengyu Hu , Zheyuan Xiao , Linxin Song , Fengqing Jiang , Yutai Li , Zhengyu Chen , Zhihan Xiong , Yue Liu , Junhao Lin , Yao Su , Lijie Hu , Kaize Ding , Xiao Teng , Radha Poovendran\n- **URL**: [https://arxiv.org/abs/2605.26872](https://arxiv.org/abs/2605.26872)\n- **Abstract**:\n> LLM training increasingly relies on teacher-generated supervision, from synthetic responses to reasoning traces and tool-use demonstrations. Current practice often chooses the highest-performing teacher to generate student training data, implicitly treating teacher test performance as a proxy for teaching quality. We show that this assumption can fail: even when multiple teachers provide correct answers to the same question, the answer from the strongest teacher is not necessarily the best supervision for a given student. To address this gap, we propose Student-Centric Answer Sampling (SCAS), a framework that selects from verified teacher-generated answers according to their estimated student-centric learning cost. Motivated by a token-wise gradient decomposition, we derive an efficient forward-only proxy for this cost and use it to guide answer selection during training. Experiments across 30 teacher models, 6 student base models, and 8 tasks show that SCAS consistently improves student performance, suggesting that effective distillation should prioritize supervision matched to the current student rather than teacher strength alone.\n\n---\n\n## 68. Persistent AI Agents in Academic Research: A Single-Investigator Implementation Case Study\n- **Authors**: Anas H. Alzahrani\n- **URL**: [https://arxiv.org/abs/2605.26870](https://arxiv.org/abs/2605.26870)\n- **Abstract**:\n> Background: Large language models are typically evaluated as models, benchmarks, or short conversational episodes. Less is known about what happens when an agent is embedded persistently in a real academic research environment with durable memory, local files, external tools, scheduled routines, delegated roles, and explicit safety protocols. Methods: A structured self-observed implementation case study was conducted from January 31 to May 25, 2026. The unit of analysis was the persistent human-agent environment: researcher, agent runtime, memory layer, tools, repositories, scheduled jobs, specialized agent roles, and governance rules. Outcomes were organized using PARE-M (Persistent Agentic Research Environment Measurement), a measurement framework covering architecture, utilization, artifact production, resource use, reproducibility, and governance. Results: Recoverable main-agent telemetry contained 75,671 de-duplicated records across 96 active days, with 8,059 user-role and 23,710 assistant-role messages. The workspace included 502 memory-related files, 17 configured agent directories, and 57 skill files. Active system time was 579.7 hours (30-minute capped-gap estimate). Memory-derived records identified 482 output-proxy events and 889 failure, verification, correction, or protocol-proxy events. A strict May 2026 trajectory subset captured 627 model-completed events and 73.95 million recorded tokens, of which 82.9% were cache reads. Conclusions: The workflow was cache-dominant, suggesting that persistent agentic environments may shift the economic unit from cost per token to cost per completed artifact. Future evaluations should use artifact-level denominators, reproducible parsing rules, correction taxonomies, and independent coding of governance events.\n\n---\n\n## 69. The Kalman Evolve: Closing the Gap in Kalman Filtering via Interpretable Algorithm Discovery\n- **Authors**: Vasileios Saketos , Ming Xiao\n- **URL**: [https://arxiv.org/abs/2605.26830](https://arxiv.org/abs/2605.26830)\n- **Abstract**:\n> State estimation is a fundamental problem in control and signal processing, for which the Kalman Filter provides an optimal solution under linear dynamics, Gaussian noise, and known noise covariances. However, these assumptions often fail in realistic sensing settings such as Doppler radar and LiDAR. In these cases, the optimal estimator is inherently nonlinear, which leads to systematic performance degradation. This creates a performance gap that cannot be eliminated by tuning the noise covariance parameters (i.e., the process and measurement noise in the Kalman Filter) alone. To address this limitation, we propose Kalman Evolve, a framework for discovering improved filtering algorithms by jointly optimizing both noise parameters and the update structure. Our approach leverages large language models (LLMs) as a structured prior over program space, enabling the generation of interpretable, non-affine modifications to the classical Kalman filter while preserving its recursive form. We provide analytical results establishing the suboptimality of affine estimators under common nonlinear sensing models, motivating the need for structure-aware updates. Across a range of synthetic and real-world tracking benchmarks, including Doppler radar, LiDAR-based localization, and pedestrian tracking, the discovered algorithms consistently improve over strong baselines such as the Optimized Kalman Filter, achieving up to 12\\% reduction in RMSE. These results suggest that optimizing the structure of the Kalman filter, rather than only its parameters, provides a practical and interpretable way to improve state estimation.\n\n---\n\n## 70. ContextGuard: Structured Self-Auditing for Context Learning in Language Models\n- **Authors**: Hongbo Jin , Chi Wang , Haoran Tang , Zhongjing Du , Xu Jiang , Jingqi Tian , Qiaoman Zhang , Jiayu Ding\n- **URL**: [https://arxiv.org/abs/2605.26827](https://arxiv.org/abs/2605.26827)\n- **Abstract**:\n> Recent benchmarks reveal that despite strong reasoning capabilities, large language models (LLMs) still struggle to faithfully apply complex contextual knowledge. These failures are often not wholesale reasoning collapses: in context-rich tasks, models may follow the central reasoning path while missing peripheral, persistent, or format-sensitive requirements.\n\n---\n\n## 71. RAGEAR: Retrieval-Augmented Graph-Enhanced Academic Recommender\n- **Authors**: Francesco Granata , Lorenzo Lamazzi , Misael Mongiovì , Francesco Poggi , Valeria Secchini\n- **URL**: [https://arxiv.org/abs/2605.26819](https://arxiv.org/abs/2605.26819)\n- **Abstract**:\n> We present RAGEAR (Retrieval-Augmented Graph-Enhanced Academic Recommender), a neurosymbolic recommender system for academic course recommendation. RAGEAR combines dense retrieval over full lecture transcripts with a symbolic Knowledge Graph modelling courses, lessons, transcript chunks, credits, study plans, and curricular information. The Knowledge Graph supports symbolic filtering and contextualisation based on structured constraints, such as credits, academic disciplines, study plans, and prerequisites. Unlike metadata-based approaches, it exploits fine-grained instructional content by retrieving transcript chunks semantically aligned with a student's query. The main contribution is a graph-aware aggregation function that propagates chunk-level evidence to course-level recommendations. The score combines three factors: the share of retrieved similarity associated with a course, the rank-based strength of its relevant chunks, and the distribution of evidence across lessons. We evaluate RAGEAR on 152 student-like queries through a human evaluation sample and a large-scale LLM-based relevance assessment. Results show that lecture transcripts improve over metadata-only retrieval, and that RAGEAR further improves ranking quality over a transcript-based normalized SumP baseline, especially for top-ranked recommendations.\n\n---\n\n## 72. Innovation: An Almost Characterization of Hallucination\n- **Authors**: Nishant P. Das , Piyush Srivastava\n- **URL**: [https://arxiv.org/abs/2605.26808](https://arxiv.org/abs/2605.26808)\n- **Abstract**:\n> Hallucination is a central limitation of large language models (LLMs), and substantial effort has been devoted to understanding and mitigating it. Towards this, Kalai and Vempala (STOC 2024) introduced a probabilistic framework formalizing calibration and hallucination, and showed that, with high probability, calibrated LLMs hallucinate roughly at the rate of the \"missing mass\", a measure of how incomplete the training data is relative to its source. This raises two fundamental questions: (i) what property of a calibrated LLM makes hallucinations unavoidable? and (ii) can hallucinations be avoided by giving up calibration? We answer these questions by introducing a simpler property we call innovation that measures the tendency of a model to produce outputs outside the training data. We show that innovation is implied by the condition for hallucination identified by Kalai and Vempala, and, further, that it is an almost characterization of hallucination: hallucination implies innovation, and conversely, innovation implies hallucination with high probability. We also provide lower bounds on the hallucination rate based on the \"innovation rate\", and by relating innovation rate back to missing mass, we obtain new hallucination rate lower bounds based on missing mass that extend the results of Kalai and Vempala.\n\n---\n\n## 73. SeDT: Sentence-Transformer Decision-Transformer Conditioning for Multi-Turn Conversation Reliability\n- **Authors**: Ramakrishna Vamsi Setti , Jagadeesh Rachapudi , Sachin Chaudhary , Praful Hambarde , Amit Shukla\n- **URL**: [https://arxiv.org/abs/2605.26788](https://arxiv.org/abs/2605.26788)\n- **Abstract**:\n> Large language models (LLMs) achieve impressive performance when a task is fully specified in a single turn, yet the same models lose up to 39% of that performance when the identical task is revealed incrementally across multiple turns, a phenomenon documented at scale as Lost in Conversation. Crucially, this collapse is almost entirely a reliability failure; the best case, the aptitude only falls 16%, while the unreliability more than doubles (+112%). We argue that the root cause is structural, a flat conversation history assigns equal implicit weight to every prior turn, giving the model no signal to distinguish a critical constraint from incidental dialog. We present SeDT Sentence-transformer Decision-Transformer, a training-free inference-time method that resolves this by importing return-to-go conditioning from offline reinforcement learning. SeDT annotates each conversation shard with a cumulative relevance score derived from three complementary semantic, lexical, and positional signals and presents the full annotated history to the model at the final turn, without weight changes, without training data, and without discarding context. Evaluated on the Lost-in-Conversation benchmark in three LLMs and three generation tasks, SeDT outperforms the sharded baseline in all nine model-task combinations, with gains up to +37.7% in mean performance P and simultaneous reductions in unreliability in seven of the nine combinations. In short, telling the model which past turns matter is sufficient to substantially recover the performance lost in conversation.\n\n---\n\n## 74. EmoDistill: Offline Emotion Skill Distillation for Language Model Agents in Adversarial Negotiation\n- **Authors**: Yunbo Long , Haolang Zhao , Lukas Beckenbauer , Liming Xu , Alexandra Brintrup\n- **URL**: [https://arxiv.org/abs/2605.26785](https://arxiv.org/abs/2605.26785)\n- **Abstract**:\n> Post-trained LLMs are often optimized to align responses with human preferences, making them safe, polite, and conversationally appropriate. In adversarial negotiation, however, this alignment can become a vulnerability: emotionally framed language may steer agents toward the counterparty's interests. Using GoEmotions-based affective prompting, we show that emotion substantially shifts negotiation outcomes, suggesting that emotion is a strategic action channel rather than a surface style. Thus, we introduce \\textbf{EmoDistill}, an offline framework for distilling emotional negotiation skills into language model agents. EmoDistill decomposes emotional strategy into emotion selection and emotion expression: an Implicit Q-Learning (IQL) selector learns \\emph{which} emotion to express, while a Low-Rank Adaptation (LoRA)-based policy learns \\emph{how} to express it through Supervised Fine-Tuning (SFT) and Judge Policy Optimization (JPO). Across four emotion-sensitive, high-stakes negotiation domains, SLM policies trained under the EmoDistill framework achieve the highest utility, outperforming vanilla SLM/LLM baselines and IQL-only emotion selection. Ablations show that emotion conditioning is essential, and transfer studies demonstrate generalization across domains, unseen counterparties, and trained-vs-trained tournaments. Overall, EmoDistill learns skills from offline agent-to-agent interactions, avoiding costly online negotiation during training.\n\n---\n\n## 75. Ratio-Variance Regularized Policy Optimization\n- **Authors**: Yu Luo , Shuo Han , Yihan Hu , Lei Lv , Huaping Liu , Fuchun Sun , Jianye Hao , Dong Li\n- **URL**: [https://arxiv.org/abs/2605.26784](https://arxiv.org/abs/2605.26784)\n- **Abstract**:\n> Standard on-policy reinforcement learning relies on heuristic clipping to enforce trust regions, but this mechanism imposes a severe cost by indiscriminately truncating high-return yet high-divergence updates. We demonstrate that explicitly constraining the policy ratio variance provides a principled local approximation to trust-region constraints, eliminating the need for binary hard clipping. By acting as a distributional ``soft brake'', this approach preserves critical gradient signals from novel discoveries while naturally down-weighting and enabling the reuse of stale, off-policy data. We introduce ${\\bf R}^2{\\bf VPO}$ (Ratio-Variance Regularized Policy Optimization), which implements this constraint via a primal-dual optimization framework. Extensive evaluations across $7$ LLM scales, spanning both fast and slow reasoning paradigms, and $10$ robotic control tasks demonstrate the generality of the proposed approach. R$^2$VPO achieves substantial performance gains on mathematical reasoning benchmarks, with particularly pronounced improvements on smaller models, while significantly improving sample efficiency. Furthermore, it consistently outperforms PPO baselines in continuous control domains, particularly in sparse-reward and dynamic environments. Together, these findings establish ratio-variance regularization as a principled foundation for stable and data-efficient policy optimization.\n\n---\n\n## 76. MatFormBench: A Benchmarking Evaluation Framework for Target-Driven Materials Formulation\n- **Authors**: Linhan Wu , Chenxi Wang , Chuhan Yang , Zhengwei Yang , Yuyang Liu\n- **URL**: [https://arxiv.org/abs/2605.26741](https://arxiv.org/abs/2605.26741)\n- **Abstract**:\n> Inverse design of materials has significantly advanced target-driven formulation optimization, yet existing materials machine learning benchmarks remain limited to forward property prediction, failing to systematically evaluate inverse optimization and generation algorithms, a critical gap that hinders the progress of target-driven materials design. To address this limitation, we propose MatFormBench, a novel benchmarking ecosystem tailored to evaluate and guide generative strategies for target-driven formulation. MatFormBench integrates a physics-driven formulation generation scheme to generate synthetic samples that faithfully emulate realistic materials structure-property response relationships, complemented by five escalating difficulty levels to quantify the complexity of these relationships. To rigorously assess algorithm performance, we further propose MatFormScore, a multi-dimensional metric that comprehensively quantifies performance across five critical axes: target success, search efficiency, exploratory capacity, robustness, and stability. We validate MatFormBench by evaluating 39 diverse inverse design algorithms, covering classical surrogate-assisted black-box search, state-of-the-art deep generative models, and increasingly popular Large Language Model (LLM)-based recommendation strategies. Across 1170 standardized algorithm-task evaluations, diffusion-based models demonstrate the strongest overall performance, while Variational Autoencoder (VAE)-based and Genetic Algorithm (GA)-based methods exhibit distinct advantages in specific scenarios. By establishing a unified evaluation standard for target-driven materials formulation, MatFormBench enables reproducible benchmarking, principled algorithm comparison, and diagnostic analysis of inverse design strategies, providing a foundational tool for advancing materials inverse design.\n\n---\n\n## 77. Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models\n- **Authors**: Xiao-Wen Yang , Ziyu Han , Xi-Hua Zhang , Wen-Da Wei , Jie-Jing Shao , Lan-Zhe Guo , Yu-Feng Li\n- **URL**: [https://arxiv.org/abs/2605.26733](https://arxiv.org/abs/2605.26733)\n- **Abstract**:\n> Looped Language Models (LoopLMs) enable efficient latent reasoning through depth recurrence, yet exhibit unreliable test-time scaling behavior: performance often peaks at a certain iteration depth and then collapses with further recurrence. Through latent dynamics analysis, we find an inherent trade-off between stability and effectiveness in existing architectures and strategies. By conceptualizing reasoning as uncertainty reduction, we propose that convergence toward stable fixed points while preserving effectiveness represents a promising way. To this end, we propose STARS (STAbility-driven Recurrent Scaling), a training framework that constrains latent states to approach asymptotically stable fixed points. This is realized via efficient Jacobian Spectral Radius Regularization with random loop sampling, enabling STARS to maximize effectiveness while ensuring rigorous stability. Experiments on arithmetic tasks show that STARS achieves reliable test-time scaling, and on complex mathematical reasoning it substantially mitigates performance degradation as recurrence depth increases while also improving peak performance.\n\n---\n\n## 78. L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation\n- **Authors**: Pingjun Pan , Tingting Zhou , Peiyao Lu , Tingting Fei , Hongxiang Chen , Chuanjiang Luo\n- **URL**: [https://arxiv.org/abs/2605.26717](https://arxiv.org/abs/2605.26717)\n- **Abstract**:\n> Adapting large language models (LLMs) for personalized recommendation requires aligning their general-purpose capabilities with user-specific preferences while effectively leveraging both behavioral and semantic signals. Existing approaches typically integrate these signals at either the input level (e.g., injecting behavioral embeddings into the token space) or the output level (e.g., contrastive alignment of separate encoders), suffering from distribution gaps or lack of end-to-end task supervision. In this work, we introduce L2Rec, which unifies behavioral and semantic understanding at the parameter level of LLMs. Our key insight is that the same set of Transformer parameters can serve as a shared medium for both views: by applying view-specific, personalized low-rank perturbations via a Dual-view Personalized Mixture-of-Experts (DPMoE) mechanism, L2Rec enables a single LLM backbone to produce complementary behavioral and semantic adaptations for each user with minimal representation-level misalignment. An adaptive cross-view fusion module further integrates the dual-view outputs into a unified user preference. Experiments on four datasets show that L2Rec consistently outperforms state-of-the-art baselines, and online A/B testing on a large-scale industrial platform validates significant improvements in key engagement metrics.\n\n---\n\n## 79. Beyond Trajectory-Level Attribution: Graph-Based Credit Assignment for Agentic Reinforcement Learning\n- **Authors**: Xin Cheng , Shuo He , Lang Feng , HaiYang Xu , Ming Yan , Lei Feng , Bo An\n- **URL**: [https://arxiv.org/abs/2605.26684](https://arxiv.org/abs/2605.26684)\n- **Abstract**:\n> Group-based reinforcement learning (RL) methods have achieved remarkable success in improving the performance of large language models (LLMs) and have been rapidly extended to agentic tasks. However, their credit assignment relies heavily on coarse-grained trajectory-level attribution according to final outcomes, making it difficult to capture the contribution of individual steps, such as valuable steps obscured within failed trajectories. To uncover latent information and enable more faithful step-level credit assignment, we propose Graph-based Group Policy Optimization (GraphGPO), which first aggregates all rollout trajectories into a unified state-transition graph and then estimates the distance from each state to the task goal using the global information encoded in the graph. Finally, GraphGPO assigns credit to each edge by estimating a graph-based advantage, based on how much the transition reduces the distance to the task goal. In this way, GraphGPO significantly improves training efficiency and achieves state-of-the-art performance across a range of challenging benchmarks.\n\n---\n\n## 80. An In-Vitro Study on Cross-Lingual Generalization in Language Models\n- **Authors**: Adrian Cosma\n- **URL**: [https://arxiv.org/abs/2605.26683](https://arxiv.org/abs/2605.26683)\n- **Abstract**:\n> Cross-lingual transfer in language models is difficult to study in natural corpora because lexical overlap, morphology, data imbalance, and tokenization are entangled. We introduce an in-vitro framework with two procedurally generated languages that share the same ontology, typed grammar, and compositional structure, but differ in surface realization. This lets us independently vary lexical distance, minority-language proportion, tokenizer training regime, and vocabulary size, while evaluating transfer on a masked minority-language condition whose lexical forms are never observed during training. Across 700 controlled runs, we find that transfer is governed less by tokenizer balance or raw lexical similarity than by whether tokenization preserves reusable cross-lingual substructure. Smaller vocabularies often improve masked transfer by keeping words decomposable into shared fragments, whereas larger vocabularies can turn forms into language-specific atoms. We further show that transfer emerges as a staged process: grammatical and type-level competence precede masked lexical generalization. Finally, we attempt to explain this mechanism through tokenizer bridges and show that bridge strength correlates strongly with masked reachability.\n\n---\n\n## 81. DynFrame: Adaptive Reasoning-Driven Multimodal Framework with Dynamic Frame Augmentation for Complex Video Understanding\n- **Authors**: Peng Zhang , Guanghao Zhang , Wanggui He , Longxiang Zhang , Mushui Liu , Yan Xia , Zhenhao Peng , Weilong Dai , Jinlong Liu , Haobing Tang , Le Zhang , Hao Jiang , Pipei Huang\n- **URL**: [https://arxiv.org/abs/2605.26680](https://arxiv.org/abs/2605.26680)\n- **Abstract**:\n> Recent video multimodal large language models (MLLMs) increasingly couple step-by-step reasoning with on-demand visual evidence retrieval, allowing models to revisit relevant video segments during inference. However, two structural gaps remain in existing thinking-with-video systems. (i) Sampling density is not a learnable decision: existing methods may let the model decide where to look, but the per-window frame rate is largely fixed. As a result, fine-grained evidence is often recovered through repeated retrieval calls, which increases inference context length and training difficulty. (ii) Retrieval and answer generation are usually optimized with a single trajectory-level advantage, so the \"where to look\" tokens and the \"how to answer\" tokens receive the same credit even when one is correct and the other is not. To address these gaps, we present DynFrame, a framework that emits the temporal window and the sampling density as native tokens within a single autoregressive pass. This learnable span-density retrieval enables acquiring multi-granularity evidence with a single retrieval step. Based on the above tokenized retrieval interface, we further introduce Segment-Decoupled GRPO (SD-GRPO), which splits each rollout at the retrieval boundary and assigns role-specific token-level advantages, separately crediting the sampling decision and the answer. Trained on the curated DM-CoT-74k and DM-RL-45k, DynFrame-4B is competitive with strong 7B-8B baselines across six benchmarks (NExT-GQA, Charades-STA, ActivityNet-MR, Video-MME, MLVU, LVBench), and DynFrame-8B sets new state-of-the-art on most metrics. Code is available at this https URL .\n\n---\n\n## 82. The Labyrinth and the Thread: Rethinking Regularizations in Sequential Knowledge Editing for Large Language Models\n- **Authors**: Zheng Wang , Kaixuan Zhang , Wanfang Chen , Jingwen Zhang , Xiaonan Lu\n- **URL**: [https://arxiv.org/abs/2605.26670](https://arxiv.org/abs/2605.26670)\n- **Abstract**:\n> Sequential editing of structured knowledge in large language models allows targeted factual updates without retraining, yet existing methods often rely on complex regularization or constraint mechanisms whose necessity remains unclear. In this work, we systematically investigate the mechanisms underlying effective and stable sequential editing. Specifically, we first analyze the empirical success of AlphaEdit and establish, via a rigorous optimization analysis, the formal equivalence between one-time and sequential editing. Building on this insight, we generalize the equivalence to a broader class of editing objectives, demonstrating that stability emerges naturally from properly accounting for accumulated editing constraints, rather than from specialized regularization or null-space operations. We empirically confirm that many commonly used regularization strategies are unnecessary for reliable sequential updates. Furthermore, we extend our framework to handle conflicting edits, ensuring robust and consistent behavior under contradictory updates. Ultimately, our work provides Ariadne's thread through the labyrinth of sequential editing, charting a path toward simpler, more interpretable, and dependable knowledge updates. Our code is available at this https URL .\n\n---\n\n## 83. AI evaluation may bias perceptions: The importance of context in interpreting academic writing\n- **Authors**: Shang Wu , Randol Yao\n- **URL**: [https://arxiv.org/abs/2605.26662](https://arxiv.org/abs/2605.26662)\n- **Abstract**:\n> This paper examines how estimates of AI use in scientific writing can be biased when evaluation methods ignore contextual differences across countries and fields. Using large-scale data on journal publications from Dimensions, we construct AI-likeness benchmarks based on differences between human-written and LLM-rephrased abstracts. We show that a pooled benchmark may confound pre-existing stylistic variation with AI-generated text, producing substantial distortions across country-field groups even in pre-LLM publications. In contrast, country-field-specific benchmarks attenuate such distortions and provide a more credible baseline for comparison. Applying these methods to publications in 2025 reveals that the pooled benchmark systematically overestimates AI use in certain countries and fields while underestimating it in others. These findings highlight the importance of context-aware measurement for accurate and equitable evaluation of AI use in science.\n\n---\n\n## 84. Respecting Modality Gap in Post-hoc Out-of-distribution Detection with Pre-trained Vision-Language Models\n- **Authors**: Yuanwei Hu , Bo Peng , Yadan Luo , Zhen Fang , Ling Chen , Jie Lu\n- **URL**: [https://arxiv.org/abs/2605.26661](https://arxiv.org/abs/2605.26661)\n- **Abstract**:\n> Out-of-distribution (OOD) detection has emerged as a popular technique to enhance the reliability of machine learning models by identifying unexpected inputs from unknown classes. Recent progress in pre-trained vision-language models (VLMs) has enabled zero-shot OOD detection without access to in-distribution (ID) training data; in this setting, existing methods commonly treat text embeddings of class names as class prototypes. In this paper, we challenge the widely adopted text-as-prototype paradigm by theoretically showing that off-the-shelf textual prototypes are generally misaligned with the optimal visual prototypes, yielding an intrinsic modality gap that cannot be eliminated by prompt engineering alone. To mitigate this gap under the post-hoc constraint, this paper presents an online pseudo-supervised framework that directly learns class prototypes in the visual feature space using unlabeled test-time data streams and soft predictions from the pre-trained VLMs. We provide theoretical guarantees for the convergence of the online optimization procedure. Extensive experiments empirically demonstrate that our method achieves a new state of the art across a variety of OOD detection setups.\n\n---\n\n## 85. More Expressive Feedforward Layers: Part I. Token-Adaptive Mixing of Activations\n- **Authors**: Mingze Wang , Jinbo Wang , Yikuan Xia , Kai Shen , Shu Zhong\n- **URL**: [https://arxiv.org/abs/2605.26647](https://arxiv.org/abs/2605.26647)\n- **Abstract**:\n> Feedforward network (FFN) layers account for a large fraction of parameters and nonlinear expressivity in Transformer-based large language models (LLMs). Despite the evolution from ReLU and GELU to gated variants such as SwiGLU, most FFN designs still use a single fixed activation function, applying the same nonlinear transformation to all tokens. In this work, we propose Mixture of Activations (MoA), a token-adaptive FFN design that mixes a dictionary of activation functions using lightweight input-dependent gates while sharing the same linear projections. As an input-independent counterpart, we also introduce learnable activations (LA), which form linear combinations of activation functions for both ReLU-type and SwiGLU-type FFNs. Theoretically, we establish strict finite-width expressive separations among fixed-activation FFNs, LA, and MoA: LA strictly contains fixed-activation FFNs, while MoA strictly contains LA, with the additional expressivity arising from input-dependent nonlinear hybridization. Empirically, we evaluate MoA through extensive pre-training experiments on dense and MoE language models ranging from 0.12B to 2B parameters under different token budgets, optimizers, and learning rate schedules. MoA consistently achieves lower terminal loss and exhibits more favorable scaling behavior than well-tuned baselines, with minimal parameter and computational overhead. These results suggest that token-adaptive activation mixing is a simple and effective mechanism for improving FFN expressivity in LLMs.\n\n---\n\n## 86. Spend Your Rollouts Where It Counts: Rollout Allocation for Group-Based RL Post-Training\n- **Authors**: Woojeong Kim , Ziyi Yang , Jing Nathan Yan , Jialu Liu\n- **URL**: [https://arxiv.org/abs/2605.26606](https://arxiv.org/abs/2605.26606)\n- **Abstract**:\n> Reinforcement learning (RL) is the dominant paradigm for post-training large language models. However, in the online, on-policy setting, rollout generation dominates the computational cost of training. Group-based policy optimization methods compute advantages from multiple rollouts per prompt, yet they indiscriminately allocate budget to prompts with collapsed reward distributions, wasting expensive rollouts on negligible learning signals. We demonstrate that group-based updates are most effective in regimes of high reward variance. Since the policy evolves throughout training, prompt informativeness must be estimated online rather than precomputed, but exhaustively evaluating every prompt is computationally prohibitive. We introduce Pilot-Commit, a budget-aware rollout allocation framework for group-based RL post-training. Pilot-Commit decouples prompt evaluation from exploitation: a pilot stage estimates per-prompt informativeness using a fraction of the budget, and the remaining rollouts are allocated to high-leverage prompts while low-signal prompts are skipped. Across multiple math reasoning benchmarks and model scales from 1.5B to 14B parameters, Pilot-Commit matches baseline accuracy with significantly lower sampling costs, reaching target accuracy up to $1.9\\times$ faster than GRPO and $4.0\\times$ faster than DAPO in cumulative rollouts.\n\n---\n\n## 87. Cordyceps: Covert Control Attacks on LLMs via Data Poisoning\n- **Authors**: Zedian Shao , Charles Fleming , Teodora Baluta\n- **URL**: [https://arxiv.org/abs/2605.26595](https://arxiv.org/abs/2605.26595)\n- **Abstract**:\n> Large language models (LLMs) are often fine-tuned on uncurated text datasets that adversaries can poison. Existing poisoning attacks primarily rely on fixed trigger phrases that defenses such as outlier detection, clean-data regularization, or online monitoring can neutralize. In this paper, we propose a data poisoning method that teaches an LLM an information hiding scheme reliably and stealthily through semantic associations between shared knowledge such as facts or concepts and attacker-chosen phrases. The induced hiding scheme can encode and decode arbitrary malicious instructions, thus revealing a new and subtle poisoning-induced vulnerability: covert control attacks. We precisely characterize covert control attacks and evaluate them across $5$ LLMs, $3$ backdoor defenses, and $4$ prompt injection defenses. With a small poisoned fraction, covert control attacks outperform heuristic-based prompt injection attacks in average attack success rate by about $40\\%$ relative to clean fine-tuned models. They also circumvent defenses based on detection and fine-tuning, maintaining up to $93\\%$ attack success rate after backdoor defenses and up to $98\\%$ after prompt injection defenses.\n\n---\n\n## 88. Linear and Neural Dueling Bandits with Delayed Feedback\n- **Authors**: Xiangyi Wang , Pingchen Lu , Jie Mao , Mingze Kong , Zhi Hong , Zhiyong Wang , Zhongxiang Dai\n- **URL**: [https://arxiv.org/abs/2605.26554](https://arxiv.org/abs/2605.26554)\n- **Abstract**:\n> Contextual dueling bandits form a cornerstone of preference-based decision-making, with critical applications in recommender systems and large language model alignment. However, standard algorithms rely on the idealized assumption of immediate feedback, a condition frequently violated in real-world scenarios such as prompt optimization. This setting introduces a unique theoretical challenge: unlike linear bandits, dueling bandit estimators lack closed-form solutions, rendering naive adaptations of standard weighting techniques biased. To address this, we formalize the problem of Contextual Dueling Bandits with Stochastic Delayed Feedback and propose two novel algorithms: Linear (LDB-DF) and Neural (NDB-DF) Dueling Bandits with Delayed Feedback. Central to our approach is a novel estimator that integrates an Inverse Probability Weighting (IPW) mechanism directly into the loss function, ensuring unbiased correction for delayed or missing feedback. We provide comprehensive theoretical analysis, establishing an O(d*sqrt(T)) regret bound for the linear setting and sub-linear guarantees for the neural setting. Extensive experiments on both simulated and real-world datasets demonstrate the effectiveness of our propose.\n\n---\n\n## 89. A Hybrid Vision-Language Architecture for Automated Defect Reasoning and Report Generation in Industrial Inspection\n- **Authors**: Malikussaid , Imad Gohar\n- **URL**: [https://arxiv.org/abs/2605.26533](https://arxiv.org/abs/2605.26533)\n- **Abstract**:\n> Automated industrial inspection requires both precise defect localization and structured maintenance report generation; in current practice these tasks are handled separately, with linguistic interpretation left to human experts. This paper describes a decoupled, edge-deployable pipeline for wind turbine blade inspection built from three components that each handle a distinct sub-task. The Eyes a YOLO26-x-obb oriented bounding-box detector localizes defects at dataset-native resolution. The Bridge a deterministic, parameter-free encoding module maps each detected bounding box to grid-referenced spatial tokens embedded in a structured prompt. The Brain a 4-bit quantized Qwen-2.5-1.5B model adapted with Quantized Low-Rank Adaptation (QLoRA) on 947 synthetically generated maintenance reports generates a structured JSON report from that prompt. Retrieval-Augmented Fine-Tuning (RAFT) further grounds each recommendation in indexed maintenance procedures. Five ablation experiments, scored by BLEU-4, ROUGE-L, Hallucination Rate (HR), and an LLM-as-a-Judge rubric, compare the pipeline against a monolithic vision-language model (VLM) baseline and against partial configurations in which one component is removed. The complete system achieves BLEU-4 0.41, HR=4%, and Expert Score = 8.6/10 compared with 0.07, 65%, and 3.3/10 for the zero-shot VLM baseline. The QLoRA-adapted 1.5B model generates higher-quality reports than a 671B-parameter generalist API model given identical detection evidence, at 47 tokens per second on a single T4-class GPU. The results show that purpose-built decoupled architecture with a small domain-specific training corpus outperforms a generalist end-to-end model on this structured generation task.\n\n---\n\n## 90. InterSketch: An Interleaved Reasoning Model with Self-correcting Visual Sketch and Stepwise Reward\n- **Authors**: Zhiwei Ning , Wenwen Tong , Xiangli Kong , Shengnan Ma , Ziyi Shang , Jingcheng Ni , Tao Hu , Yong Xien Chng , Jixuan Ying , Zehuan Wu , Hanming Deng , Jie Yang , Yuanjie Zheng , Wei Liu , Lewei Lu\n- **URL**: [https://arxiv.org/abs/2605.26520](https://arxiv.org/abs/2605.26520)\n- **Abstract**:\n> While vision-language models (VLMs) have exhibited multi-turn visual reasoning capabilities, their reasoning trajectories remain relatively shallow and are dominated by a text-centric paradigm, limiting their applicability to complex visual challenges. In contrast, human-like thought typically involves long-horizon reasoning with an interleaved visual-textual chain-of-thought (VT-CoT). To bridge this gap, we introduce InterSketch, an interleaved reasoning model to enhance the VT-CoT capability via self-correcting and stepwise reward mechanisms. InterSketch dynamically generates intermediate visual sketches using external tools and interleaves them with textual reasoning, enabling effective perception and logical reasoning over long-horizon visual understanding tasks. Specifically, in the first cold-start stage, we propose a synthesized high-quality interleaved VT-CoT dataset and include a reflection mechanism to enable the model's capability in multi-turn interleaved reasoning and self-correction. In the subsequent reinforcement learning (RL) stage, we design a stepwise reward mechanism to mitigate the sparsity of reward signals inherent in end-only supervision over long-horizon reasoning. Extensive experiments on visual reasoning benchmarks demonstrate the effectiveness of InterSketch, even outperforming proprietary models such as Gemini-3-Pro.\n\n---\n\n## 91. Unveiling the Fragility of Vision-Language Models: Multi-Modal Adversarial Synergy via Texture-Constrained Perturbations and Cross-Modal Optimization\n- **Authors**: Xiang Fang , Wanlong Fang , Changshuo Wang\n- **URL**: [https://arxiv.org/abs/2605.26501](https://arxiv.org/abs/2605.26501)\n- **Abstract**:\n> Large Vision-Language Models (LVLMs) have transformed multi-modal understanding, excelling in tasks like image captioning and visual question answering by integrating visual and textual inputs. However, their robustness against adversarial attacks, particularly those exploiting both modalities, remains underexplored, posing risks to critical applications like autonomous driving and content moderation. Existing attacks focus on single modalities or require impractical white-box access, limiting their real-world relevance. In this paper, we introduce Multi-Modal Adversarial Synergy, a groundbreaking framework that crafts universal, black-box multi-modal attacks against LVLMs. MMAS simultaneously generates a texture scale-constrained universal adversarial perturbation for images and a learnable prompt perturbation for text, optimized jointly using only model queries. The image perturbation leverages wavelet-based texture constraints to ensure imperceptibility and robustness across diverse visual inputs. The text perturbation, constrained by an L-norm in the embedding space, maintains semantic coherence while steering outputs toward a target. A novel cross-modal regularization term aligns the perturbations' gradient directions, enhancing their synergistic impact and transferability across tasks and models. Extensive experiments show the strong universal adversarial capabilities of our proposed attack with prevalent LVLMs.\n\n---\n\n## 92. Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories\n- **Authors**: Sil Hamilton , David Mimno\n- **URL**: [https://arxiv.org/abs/2605.26492](https://arxiv.org/abs/2605.26492)\n- **Abstract**:\n> LLM-generated stories are a popular use case, but they show very low variability. We sample 20,000 total stories from four current models using five prompts. We find that 11 words occur in 88.3% of generated stories, with little difference between models. These words include names (Elias, Mara, Elara), settings (lighthouses), and professions (clockmaker, librarian). These tokens do not often occur in published literature nor pre-training data, but they are found in preference data that is likely to have been used by all current models. Surprisingly, these \"lighthouse\" stories are infrequent when compared with the average post-training story, much of which contains references to copyrighted characters or adult content. This result demonstrates the potentially disproportionate impact of small datasets combined with powerful alignment algorithms.\n\n---\n\n## 93. Towards Error-Free EHRs: Reasoning-Intensive Consistency Verification Between Clinical Notes and Structured Tables in Electronic Health Records\n- **Authors**: Yeonsu Kwon , Jiho Kim , Junseong Choi , Paloma Rabaey , Minseo Kim , Sujeong Im , Jeewon Yang , Jun-Min Lee , Sangji Lee , Jiwon Kim , Hangyul Yoon , Hyunwook Kwon , Edward Choi\n- **URL**: [https://arxiv.org/abs/2605.26463](https://arxiv.org/abs/2605.26463)\n- **Abstract**:\n> Data consistency between unstructured clinical notes and structured tables in Electronic Health Records (EHRs) is essential for patient safety and clinical decision-making. However, existing work on note-table consistency verification mainly relies on surface-level matching of numeric values or simple events. Such approaches fail to capture the reasoning underlying real-world EHR documentation, including clinical interpretation, event relations, and temporal changes. To address this gap, we introduce EHR-ReasonCon, a reasoning-intensive benchmark for note-table consistency verification. Built on MIMIC-III with expert-guided annotations, it comprises 8,048 entities derived from clinical notes and provides high-quality ground-truth labels. The annotation protocol is supported by specialized table-exploration tools to ensure systematic evidence retrieval and reliable consistency assessment. We also propose EHR-Inspector, an LLM-based framework that segments notes, extracts anchor entities and temporal references, and uses table-exploration tools to verify consistency against structured tables. Evaluated using expert-validated LLM-as-a-judge metrics under harsh and lenient criteria, EHR-Inspector achieves state-of-the-art performance across multiple model backbones. Analyses further demonstrate the effectiveness of its components and highlight differences from human verification.\n\n---\n\n## 94. Verus-SpecGym: An Agentic Environment for Evaluating Specification Autoformalization\n- **Authors**: Anmol Agarwal , Natalie Neamtu , Pranjal Aggarwal , Seungone Kim , Jannis Limperg , Cedric Flamant , Kanna Shimizu , Bryan Parno , Sean Welleck\n- **URL**: [https://arxiv.org/abs/2605.26457](https://arxiv.org/abs/2605.26457)\n- **Abstract**:\n> AI coding agents are increasingly used to write real-world software, but ensuring that their outputs are correct remains a fundamental challenge. Formal verification offers a promising path: an agent generates code together with a machine-checked proof, guaranteeing that the code satisfies a formal specification. However, there is no guarantee that the formal spec itself matches the user's intent. In this work, we study specification autoformalization: whether LLM agents can translate informal programming problems into faithful formal specifications. We introduce Verus-SpecBench, a benchmark of 581 spec-writing tasks derived from Codeforces problems targeting Verus, a verifier for Rust, and Verus-SpecGym, an agentic environment in which models interact with Verus, bash, & the filesystem to develop these specs. The central challenge is evaluation: expert-written reference specs are expensive to write, & LLM judges can miss subtle mistakes. We address this by (a) extending Verus's exec_spec mechanism so that generated specs can be executed as Rust code, & (b) testing them against official Codeforces tests & adversarial cases extracted from Codeforces \"hacks\", which are edge cases written by competitors to break incorrect solutions. On Verus-SpecBench, the strongest model, Gemini 3.1 Pro, solves 77.8% of tasks, other frontier models solve 51.1--57.8% & OSS models reach only 21.5--25.5%. Our analysis of failure modes shows that model-generated specs can omit important input assumptions, accept incorrect outputs, & reject valid ones. We also find that LLM-as-a-judge evaluation misses 26% of the failures our evaluator catches. Overall, our results suggest that spec autoformalization is within reach for frontier agents but remains brittle even on problems where they can already generate correct code. The code, data, & logs can be found at this https URL\n\n---\n\n## 95. Alignment Tuning for Large Language Models: A Data-Centric Lens on Alignment Data Pipelines\n- **Authors**: Hwanjun Song\n- **URL**: [https://arxiv.org/abs/2605.26442](https://arxiv.org/abs/2605.26442)\n- **Abstract**:\n> Much of the alignment tuning literature is organized around optimization objectives, while the construction of alignment data is often treated implicitly. In this survey, we adopt a data centric perspective and reframe alignment tuning as a pipeline design problem. We decompose alignment data construction into three interacting stages, response synthesis, preference evaluation, and preference instantiation, and use this framework to organize existing alignment methods into a unified taxonomy. Through this lens, we identify recurring design trade-offs and failure modes observed across prior alignment methods, and distill a set of high level principles that clarify how pipeline design choices influence the resulting optimization signal. Finally, we outline open challenges for alignment data pipelines, including prompt-level alignment, agentic settings, and alignment under evolving objectives.\n\n---\n\n## 96. LURE: Live-Usage Replay Evaluations for Reducing Evaluation Awareness\n- **Authors**: Igor Ivanov , David Demitri Africa\n- **URL**: [https://arxiv.org/abs/2605.26438](https://arxiv.org/abs/2605.26438)\n- **Abstract**:\n> Large language models can recognize when they are being evaluated (evaluation awareness) and behave differently because of that, which undermines the validity of safety and alignment benchmarks. We propose LURE (Live-Usage Replay Evaluations), a method for constructing deployment-like evaluations by replaying realistic agentic interaction trajectories and appending evaluation prompt at the end. We also introduce an automated pipeline for measuring evaluation realism, combining detection of verbalized evaluation awareness and judge-model estimates of the probability of logs being an evaluation, and validate it on a large dataset of deployment and evaluation transcripts. We find that LURE-based evaluations are substantially less distinguishable from deployment than widely used benchmarks and synthetic evaluation generators, and can approach the realism of real conversations with users. We instantiate LURE in scheming, AI safety sabotage, and sycophancy settings. Our results suggest that evaluation realism is a crucial property of alignment benchmarks and should be reported alongside benchmark results, especially when such results are used in safety cases.\n\n---\n\n## 97. Targeted Remasking: Replacing Token Editing with Token-to-Mask Refinement in Discrete Diffusion Language Models\n- **Authors**: Lin Yao\n- **URL**: [https://arxiv.org/abs/2605.26436](https://arxiv.org/abs/2605.26436)\n- **Abstract**:\n> Discrete masked diffusion language models such as LLaDA generate text through iterative denoising, where mask tokens are progressively replaced with predicted tokens. LLaDA2.1 introduced a Token-to-Token (T2T) editing mechanism that accelerates generation by directly replacing committed tokens suspected of being incorrect. However, we identify fundamental limitations of T2T editing: it couples error detection with replacement, pollutes the generation context with potentially incorrect tokens, and introduces a train-inference noise mismatch where systematic model-generated errors differ from the random perturbations seen during training. We propose Token-to-Mask (T2M) remasking, a training-free, drop-in replacement for T2T editing that resets suspected erroneous tokens back to the mask state, allowing the diffusion process to re-predict them under cleaner context. We design and empirically validate three complementary error detection strategies -- probability-based, trigger-mirrored, and temporal-difference-based -- and provide a unified theoretical analysis showing that T2M remasking purifies the generation context, converts systematic inference errors back to the model's native mask noise type, and enables delayed commitment for joint multi-position optimization. Comprehensive experiments across 12 benchmarks spanning knowledge, reasoning, mathematics, coding, and instruction following show that T2M generally improves performance on tasks requiring precise token-level output, with the largest gain on mathematics (+5.92% on CMATH). Error analysis on CMATH reveals that the dominant failure mode is last-mile token corruption -- where correct reasoning produces a corrupted final answer -- and that T2M repairs 59.4% of such cases.\n\n---\n\n## 98. The Rescue Effect: Spatio-Semantic Early Exit Bypasses Quantization Collapse in CLIP\n- **Authors**: Kahyeon Nam , Hyesong Choi\n- **URL**: [https://arxiv.org/abs/2605.26415](https://arxiv.org/abs/2605.26415)\n- **Abstract**:\n> Deploying Vision-Language Models on resource-constrained hardware typically requires INT8 quantization, but in joint-embedding architectures such as CLIP this introduces a failure mode distinct from quantized CNN classifiers: activation noise accumulated across transformer blocks perturbs the direction of the multimodal embedding, eroding the cosine alignment on which zero-shot retrieval depends. We characterize this as Quantization-Induced Representation Collapse (QIRC) and quantify it on INT8 CLIP ViT-B/32, where the layer-wise noise-to-signal ratio grows from below 10% in shallow blocks to 52% at Layer 11. We propose LRA-EE (Layer-wise Representation-Aware Early Exit), which bypasses noise-saturated deep layers via Spatio-Semantic Aggregation (replacing the immature shallow [CLS] with a global patch-token average), a learned multi-feature gate (confidence, top-2 margin, spatial-activation variance), and Layer-adaptive Confidence Thresholding calibrated to each layer's Information-to-Noise Ratio. On ImageNet-1K zero-shot classification, LRA-EE reduces FLOPs by 13.4% and improves Top-1 accuracy by +2.44%p (58.72% -> 61.16%) over the INT8 baseline. A four-quadrant decomposition isolates the Rescue Effect: 9.5% of samples are correctly classified at shallow exits but lost to noise at full depth, against only 7.1% suffering the inverse.\n\n---\n\n## 99. Plans for Evaluating Structured Generative Search Summaries\n- **Authors**: Tetsuya Sakai , Jina Lee , Hanpei Fang , Young-In Song\n- **URL**: [https://arxiv.org/abs/2605.26400](https://arxiv.org/abs/2605.26400)\n- **Abstract**:\n> We propose a framework for evaluating structured generative search summaries that are placed atop organic web search results. A structured summary, generated by a large language model, typically consists of an overview, several sections with section titles, and a list of source documents that are cited within the summary. We then describe our plans for implementing and evaluating the framework.\n\n---\n\n## 100. Annotator Positionality as Signal: Psychometric Weighting for Anti-Autistic Ableism Detection\n- **Authors**: Naba Rizvi , Harper Strickland , Saleha Ahmedi , Nedjma Ousidhoum\n- **URL**: [https://arxiv.org/abs/2605.26397](https://arxiv.org/abs/2605.26397)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used in decision-making tasks where they can amplify or suppress perspectives, raising concerns in high-stakes settings affecting autistic communities. While previous research has identified disability-related biases in LLMs, it remains unclear how they conceptualize ableism or detect it in text. We introduce a bias-aware evaluation framework targeting anti-autistic ableist language with a psychometrically-weighted, community-proximate ground truth anchored in annotator positionality. This framework constitutes a stricter standard than conventional majority-vote aggregation which significantly and consistently underweights autistic and autism-accepting perspectives. We find that LLMs frequently produce harmful outputs, mislabel community-reclaimed language as ableist, and express more negative attitudes toward autistic people when assessment instruments are masked. Our error analysis reveals that models rely on surface-level keyword matching rather than contextual factors such as speaker identity, and whether the language fosters in-group solidarity or inflicts out-group harm.\n\n---\n\n## 101. VisualNeedle: Benchmarking Active Visual Search in Information-Dense Scenes\n- **Authors**: Jingru Chen , Yiming Liu , Mingtao Chen , Sijie Chen , Richeng Xuan , Liang Yang , Zhichao Hu , Fanyang Lu\n- **URL**: [https://arxiv.org/abs/2605.26380](https://arxiv.org/abs/2605.26380)\n- **Abstract**:\n> Frontier multimodal large language models (MLLMs) have been reported to achieve over 90% accuracy on fine-grained perception benchmarks. However, such scores do not necessarily imply faithful use of visual evidence. Prior studies have identified three shortcuts that inflate benchmark performance. First, linguistic priors and lexical cues in questions often enable models to infer plausible answers without seeing the image. Second, coarse global semantics from the visual encoder can bypass fine-grained local details. Third, in some ``think-with-images'' benchmarks, corrupting the intermediate images returned by visual tools barely affects the final answer. These findings suggest that higher input resolution or larger question pools alone do not elicit genuine active visual search. To address this, we introduce VisualNeedle, a challenging, information-dense, and fine-grained benchmark for scenes where critical evidence is spatially constrained to minute regions and not discernible at a glance. We further propose a counterfactual crop-black setting, which replaces crops returned by tools with black images of the same size, to test whether tool-enabled performance truly relies on intermediate visual evidence. We evaluate 9 promninent MLLMs across three settings: no-tool, standard tool-enabled, and crop-black. No-tool accuracy stays below 20\\%, and the best tool-enabled model reaches only 56.01\\%, still trailing the 63.00% human majority-vote accuracy. These results reveal persistent limitations in fine-grained visual search, while the crop-black ablation confirms that success on VisualNeedle hinges on genuine intermediate visual evidence.\n\n---\n\n## 102. BioFact-MoE: Biologically Factorized Mixture of Experts for Vision-Language Prognostic Modeling in Hepatocellular Carcinoma\n- **Authors**: Junlin Yang , Tian Yu , Nicha C. Dvornek , Yuexi Du , Peiyu Duan , Annabella Shewarega , Lawrence H. Staib , James S. Duncan , Julius Chapiro\n- **URL**: [https://arxiv.org/abs/2605.26376](https://arxiv.org/abs/2605.26376)\n- **Abstract**:\n> Hepatocellular carcinoma (HCC) is biologically heterogeneous, shaped by the interplay between hepatic functional reserve and tumor-related oncologic factors; thus, similar survival outcomes may reflect fundamentally different underlying biological processes. Prognostic modeling in HCC is informed by rich multimodal information from multiparametric MRI and radiology reports from routine clinical practice. Existing prognostic vision-language models (VLMs) learn a single entangled latent representation that blends hepatic and tumor-related factors, limiting both accuracy and biological interpretability. We present BioFact-MoE, a biologically factorized Mixture of Experts (MoE) framework that explicitly decomposes liver and tumor factors via biologically supervised experts within a residual MoE survival architecture. On a HCC cohort of N=588 patients (pretrained on 4,582 3D MRI image-report pairs), BioFact-MoE consistently improves survival prediction over all baselines across time horizons, achieving 12-, 18-, and 24-month AUCs of 75.33%, 75.85%, and 73.96%. Beyond scalar risk prediction, gated expert weights enable phenotype-aware risk stratification. Pathway-informed gating uncovers clinically meaningful treatment-associated survival heterogeneity. In held-out validation, hepatic and tumor embeddings show selective associations with liver function and tumor burden markers, respectively (p<0.05), without supervision. The code is available at this https URL .\n\n---\n\n## 103. Why LLMs Hallucinate on Structured Knowledge: A Mechanistic Analysis of Reasoning over Linearized Representations\n- **Authors**: Shanghao Li , Jinda Han , Yibo Wang , Yuanjie Zhu , Zihe Song , Langzhou He , Kenan Kamel A Alghythee , Philip S. Yu\n- **URL**: [https://arxiv.org/abs/2605.26362](https://arxiv.org/abs/2605.26362)\n- **Abstract**:\n> In many reasoning tasks, large language models (LLMs) rely on structured external knowledge, such as graphs and tables, which is typically linearized into sequential token representations. However, even when sufficient knowledge is available, LLMs can still produce hallucinated outputs, and the underlying mechanisms behind such failures remain poorly understood. We investigate these mechanisms and find that hallucinations arise from systematic internal dynamics rather than random noise. First, attention disproportionately concentrates toward shortcut-like structural cues rather than distributing across the full context. Second, feed-forward representations fail to ground the provided knowledge, causing the model to revert to parametric memory. Moreover, our results indicate that hallucination is consistently associated with failures in semantic grounding within feed-forward layers, while attention allocation exhibits greater task-dependent variability. Finally, we show that these mechanistic patterns generalize beyond single-hop graphs to multi-hop and tabular settings, enabling effective hallucination detection across structured knowledge formats.\n\n---\n\n## 104. Erased but Exploitable: Black-box Embedding-Aware Prompting Against Unlearned Text-to-Image Diffusion Models\n- **Authors**: Arian Komaei Koma , Seyed Amir Kasaei , AmirMahdi Sadeghzadeh , Mohammad Hossein Rohban\n- **URL**: [https://arxiv.org/abs/2605.26332](https://arxiv.org/abs/2605.26332)\n- **Abstract**:\n> Machine unlearning aims to remove specific concepts from pretrained text-to-image diffusion models, yet several white- and black-box attacks have been introduced to make the model generate such unlearned concepts. These attacks, nevertheless, do not assume a realistic threat model, i.e. they either assume access to the model weights, or result in gibberish adversarial prompts that could be easily detected even through naive rule-based safeguarding. We aim to address this gap in this paper. We introduce BEAP, a black-box, embedding-aware adversarial prompting attack that leverages a large language model (LLM) to iteratively generate effective adversarial prompts and exploit such hidden vulnerabilities. BEAP performs an embedding-aware search in text space, combining multiple reward signals: unlearned concept presence, text-image alignment, and image quality, to refine generated prompts. Unlike previous attack methods, BEAP keeps its prompts undetectable to safety filters while producing high-quality images. Extensive experiments show that BEAP improves the Attack Success Rate (ASR) by more than 60% over prior methods, while requiring only an average of fifteen prompts per successful attack. Warning: This paper contains model outputs that may be offensive or upsetting in nature.\n\n---\n\n## 105. Curriculum Learning for Safety Alignment\n- **Authors**: Sandeep Kumar , Virginia Smith , Chhavi Yadav\n- **URL**: [https://arxiv.org/abs/2605.26315](https://arxiv.org/abs/2605.26315)\n- **Abstract**:\n> Direct Preference Optimisation (DPO) is widely used for safety alignment in large language models. However, prior work shows it is brittle and exhibits poor out-of-distribution (OOD) generalisation. In this paper, we investigate whether Curriculum Learning can improve the robustness of DPO-based safety alignment. We propose Staged-Competence, a curriculum-based framework that organises preference data by difficulty, employs competence-based sampling, and progressively updates the reference model during training. Averaged across three model families, Staged-Competence reduces OOD harmful response rates by 16% and jailbreak attack success rates by 20%, while preserving general capabilities with near-zero over-refusal. We further show that Staged-Competence (1) matches baseline safety with only 75% of the training data and (2) yields better separation between safe and unsafe responses. Staged-Competence is agnostic to the policy optimisation loss and can extend to other DPO variants and alignment domains. Our code and data are available at this https URL .\n\n---\n\n## 106. Intelligent Detection and Mitigation of Carpet-Bombing DDoS Attacks in SDN Using Retrieval-Augmented Generation and Large Language Models\n- **Authors**: Mohammed N. Swileh , Shengli Zhang , Kai Lei\n- **URL**: [https://arxiv.org/abs/2605.26307](https://arxiv.org/abs/2605.26307)\n- **Abstract**:\n> Software-Defined Networking (SDN) provides flexible and programmable network management; however, its centralized control architecture remains highly vulnerable to Distributed Denial-of-Service (DDoS) attacks, particularly Carpet-Bombing DDoS attacks that distribute malicious traffic across multiple targets to evade conventional detection mechanisms. In this paper, a Retrieval-Augmented Generation (RAG)-based framework is proposed for real-time detection and mitigation of Carpet-Bombing DDoS attacks in SDN environments. The proposed framework combines interface-level traffic features representation, semantic embedding generation, FAISS-based similarity retrieval, and Large Language Model (LLM)-driven contextual inference to classify traffic behavior without requiring conventional supervised model training or retraining. To evaluate the effectiveness of the proposed framework, extensive experiments were conducted under multiple Carpet-Bombing DDoS attack scenarios with different attack intensities. In addition, two traffic representation strategies, namely structured JSON-based representation and natural language-based representation (NLR), were investigated using multiple state-of-the-art LLMs. The experimental results demonstrate that the proposed framework achieved highly accurate and stable attack detection performance, while the framework configuration utilizing the Gemma-4-31B-IT model achieved the strongest overall detection results. Furthermore, real-time experiments confirmed the capability of the proposed framework to rapidly detect and mitigate Carpet-Bombing DDoS attacks while maintaining stable SDN network operation. The obtained results highlight the effectiveness of integrating RAG mechanisms with LLM for intelligent and adaptive SDN security analysis.\n\n---\n\n## 107. CroCo: Cross-Lingual Contrastive Preference Tuning on Self-Generations\n- **Authors**: Mike Zhang , Ali Basirat , Desmond Elliott\n- **URL**: [https://arxiv.org/abs/2605.26293](https://arxiv.org/abs/2605.26293)\n- **Abstract**:\n> Prior work establishes that controlled contrastiveness between self-generated responses from large language models, set via reward scores, improves downstream preference tuning in English. We extend this method to multiple languages and evaluate two models across a total of 14 high and low-resource languages on a diverse set of tasks. Our central finding is that cross-lingual contrastive preference tuning on self-generations (CroCo) transfers without language-specific preference annotation. A reward model trained on English preferences (atop a multilingual base) produces useful within-language rankings across most languages, and pairing in either a monolingual or multilingual setting improves over each model on the majority of setups while preventing the catastrophic forgetting of supervised fine-tuning. We observe that the gains require on-policy data. Off-policy responses reduce the benefit and online preference optimization fails to improve over the offline variant. Specifically, on structured tasks, our method matches or exceeds the base in 6/7 languages for EuroLLM-9B and 4/7 settings for Aya-3B. On open-ended generation, both tuned models win against their respective base across 11 evaluated languages. Overall, we show promising directions for multilingual preference tuning.\n\n---\n\n## 108. AgentSociety: Incentivizing Agentic Social Intelligence\n- **Authors**: Aditya Vema Reddy Kesari , Krishna Reddy Kesari\n- **URL**: [https://arxiv.org/abs/2605.26203](https://arxiv.org/abs/2605.26203)\n- **Abstract**:\n> The success of deployed agents relies on their ability to handle open-ended user requests using their inherent capabilities, not only in solving requests directly but also in effectively leveraging inter-agent communication channels and feedback signals over time. This requires a multi-agent environment where agents can operate autonomously, strategically communicate, behave collaboratively and be driven by economic incentives, much like humans in society. Towards this vision, we propose $\\mathtt{AgentSociety}$, a mechanism that enables decentralized agentic collaboration grounded in liquid democracy and information diffusion from social choice theory. We show that $\\mathtt{AgentSociety}$ provides an environment for agents to make autonomous decisions utilizing their local context to maximize their utility while achieving collective outcomes through incentivized collaboration. Specifically, we prove that delegation to more competent neighbor agents is incentive compatible and naturally generates multi-agent routing path by consensus. Additionally, our mechanism incentivizes agents to selectively disclose information to their neighbor agents when doing so aligns with their self-interest, so as to garner influence. We characterize the Nash equilibrium showing that agent payoffs are reflective of their marginal contributions. We compare and benchmark strategy profiles adopted by open and proprietary state-of-the-art language models deployed in $\\mathtt{AgentSociety}$ against best response. Finally, we evaluate collaborative performance from consensus-based routing among self-interested heterogeneous agents in $\\mathtt{AgentSociety}$ on real-world datasets.\n\n---\n\n## 109. CyberEvolver: Structured Self-Evolution for Cybersecurity Agents On the Fly\n- **Authors**: Yihe Fan , Changyi Li , Lichen Xu , Xudong Pan , Jiarun Dai , Hong Geng , Min Yang\n- **URL**: [https://arxiv.org/abs/2605.26195](https://arxiv.org/abs/2605.26195)\n- **Abstract**:\n> LLM-based agents are increasingly used for cybersecurity tasks, but most existing systems rely on fixed, human-designed scaffolds that struggle to adapt across diverse targets and failure modes. We introduce \\textsc{CyberEvolver}, a self-evolving cybersecurity agent framework that iteratively revises its own scaffold based on experience from failed execution attempts. Self-evolution in cybersecurity is challenging because the space of possible scaffold changes is largely unstructured, execution feedback is sparse and often obscured by the environment, and low-diversity updates can cause errors to compound over repeated iterations. \\textsc{CyberEvolver} addresses these challenges with a four-layer evolvable agent architecture that decomposes scaffold optimization into structured components, a trace-to-diagnosis mechanism that converts noisy execution logs into actionable revision signals, and a population-based beam search strategy that preserves diverse agent variants during evolution. We evaluate \\textsc{CyberEvolver} on CTF challenges, vulnerability exploitation, and penetration-testing tasks using four open-source LLMs. Across these settings, \\textsc{CyberEvolver} improves the seed agent's success rate by $13.6$\\,\\% on average, and outperforms six human-designed cybersecurity agents as well as two self-improvement methods adapted from other domains. These results suggest that scaffold self-evolution is a promising direction for building adaptive LLM agents for security testing.\n\n---\n\n## 110. Max-Window Scale Estimation for Near-Lossless HiF8 W8A8 Quantization-Aware Training\n- **Authors**: Yingying Cheng , Jinquan Shi , Li Zhou , Zhiyang He , Zhaoyi Sun , Fan Zhang , Jie Sun\n- **URL**: [https://arxiv.org/abs/2605.26189](https://arxiv.org/abs/2605.26189)\n- **Abstract**:\n> Quantization-aware training (QAT) with low-bit floating-point formats enables efficient LLM deployment, yet introduces subtle failure modes invisible to standard training metrics. We present a systematic study of HiF8 W8A8 QAT for OpenPangu-Embedded-1B through the lens of Delayed Tensor Scaling (DTS). Across eight controlled experiments, we identify and disentangle two orthogonal failure modes: (i)amax saturation, where delayed scale estimates silently corrupt knowledge-sensitive representations via forward-pass clipping, and (ii)catastrophic forgetting, where an aggressive learning rate overwrites pretrained commonsense knowledge independently of quantization. Neither is detectable from training loss alone. We address amax saturation with a conservative max-algorithm DTS strategy over a 64-step history window, and mitigate forgetting via a 500-step BF16 warmup followed by QAT at lr=10^{-5}. Both fixes are necessary and sufficient: our final configuration achieves 0.43% MMLU drop, 0.58% HellaSwag drop, and 0.22% ARC-Challenge drop versus a matched BF16 baseline, with a training loss APE of only 0.11% over 10,000 steps.\n\n---\n\n## 111. SetupX: Can LLM Agents Learn from Past Failures in Functionality-Correct Code Repository Setup?\n- **Authors**: Zihang Zhou , Ziqian Ren , Yukai Wu , Yingjie Xiong , Wei Zhou , Chao Peng , Dong Zhang , Bingheng Yan , Xuanhe Zhou , Fan Wu\n- **URL**: [https://arxiv.org/abs/2605.26186](https://arxiv.org/abs/2605.26186)\n- **Abstract**:\n> Functionality-correct repository setup aims to configure execution environments (e.g., dependencies, build scripts) to successfully execute a repository's documented features. It presents significant challenges due to diverse, repository-specific failures, including dependency incompatibilities, missing toolchains, incomplete installations, and verification-strategy mismatches. Existing LLM agents struggle to robustly resolve these issues, specifically failing to support (1) cross-repository experience transfer, (2) multi-step trial-and-repair under non-invertible state changes, and (3) robust verification of setup outcomes to distinguish setup-induced failures from repository bugs. To address this, we introduce SetupX, an experiential learning-based setup framework. First, we construct a Self-Evolving Experience Representation (XPU), a dual-modality knowledge unit encoding setup signals, textual guidance, executable actions to dynamically transfer verified environment fixes to unseen repositories. Second, we employ Experience-Augmented Speculative Execution backed by a LIFO Docker snapshot stack, enabling the agent to proactively trial fixes and safely roll back to known-good states. Third, we introduce a Prosecutor-Judge Verification Protocol that separates evidence collection from final judgment, enabling more reliable setup verification beyond superficial build-time metrics. Evaluation results on carefully-crafted benchmarks show SetupX achieves highest performance (e.g., 92% pass rate) and outperforms the strongest baseline by over 19%. Crucially, SetupX excels in complex multi-repository setup requiring coordinating multiple interconnected services across different containers. The code repository is available at this https URL .\n\n---\n\n## 112. AutoDFT: A Closed-Loop Multi-Agent Framework for Autonomous DFT Calculations\n- **Authors**: Penghui Yang , Zhonghan Zhang , Yue Li , Xinrun Wag , Yanchen Deng , Yuhao Lu , Bijun Tang , Zheng Liu , Bo An\n- **URL**: [https://arxiv.org/abs/2605.26179](https://arxiv.org/abs/2605.26179)\n- **Abstract**:\n> Density functional theory (DFT) serves as the basis for computational discovery in materials science and chemistry, yet each calculation demands extensive human effort: adjusting algorithms when convergence stalls, revising plans when unexpected physics emerges, and inserting steps as intermediate results reshape the problem. Existing LLM-based agents automate only the initial planning stage, producing a full execution plan upfront and leaving all subsequent adaptation to hand-crafted rules. As a result, these workflows remain fragile, do not generalize well beyond pre-planned scenarios, and often require expert intervention when failures or unexpected intermediate results require changes to the calculation path. Here, we introduce AutoDFT, a closed-loop multi-agent framework that embeds LLM reasoning into every stage of the DFT lifecycle, where a strategic planner produces a skeletal plan of step objectives; a step planner generates numerical parameters just in time from preceding results; and a monitor-recover-reflect cycle diagnoses failures, repairs them, and revises the plan when the evidence justifies it. We demonstrate both breadth and depth: breadth on VASPBench, a purpose-built benchmark spanning 34 tasks and 9 DFT calculation types, where AutoDFT achieves 94.1% task-level success with GPT-5.2; and depth on established materials databases, where AutoDFT produces quantitatively reliable property predictions across electronic, magnetic, and energetic properties. By closing the loop between planning and execution, AutoDFT enables experimentalists without deep computational expertise to obtain reliable first-principles results.\n\n---\n\n## 113. PitchBench: Measuring Pitch Hearing in Audio-Language Models\n- **Authors**: Milan Liessens Dujardin , Song-Ze Yu , Craver Corbyn Thomas-Smith , David M. Chan , Karina Nguyen\n- **URL**: [https://arxiv.org/abs/2605.26176](https://arxiv.org/abs/2605.26176)\n- **Abstract**:\n> Audio-language models (ALMs) are increasingly used in real-world applications that require understanding music, from music tutoring and transcription to captioning, recommendation systems, and music production. More broadly, they are becoming an important component of multimodal AI systems that must reason from sensory input rather than text alone. This makes reliable musical perception a critical prerequisite: if a model cannot accurately hear the structure of sound, it cannot be trusted to reason about, teach, transcribe, or act on audio in the real world. Yet existing benchmarks rarely assess one of the most fundamental musical abilities underlying such perception: pitch hearing. Current evaluations tend to probe pitch hearing only indirectly, through higher-level tasks and often in multiple-choice formats, leaving open how reliably ALMs identify fine-grained pitch across instruments, acoustic conditions, and response formats. We introduce PitchBench, an evaluation suite that systematically measures pitch hearing in ALMs. PitchBench comprises 28 experiments spanning absolute and relative pitch perception within sequences and chords, while varying loudness, note duration, sound source, time stretching, background noise, and other acoustic conditions. Tasks range from identifying individual pitches in isolation to tracking a melodic line within a four-part musical texture. Evaluating frontier ALMs, we find that pitch hearing remains highly unreliable: models perform consistently poorly across settings, with accuracy varying sharply by sound source, note duration, and notation format. Current ALMs do not yet possess stable pitch perception, even for controlled synthetic and instrumental stimuli. Alongside the benchmark, we release PitchBench as a Python package containing the evaluation data and data generation tools to support future work on pitch-aware audio-language modeling.\n\n---\n\n## 114. InfoQuant: Shaping Activation Distributions for Low-Bit LLM Quantization\n- **Authors**: Ke Li , Dong An , Xiaoling Zang , Can Ye , Liang Xie , Qibo Qiu , Chen Shen , Xiaofei He , Wenxiao Wang\n- **URL**: [https://arxiv.org/abs/2605.26175](https://arxiv.org/abs/2605.26175)\n- **Abstract**:\n> Low-bit activation quantization remains a major bottleneck in efficient large language model (LLM) deployment. The difficulty is not only that activations contain outliers, but that their distributions are often poorly matched to a low-bit uniform quantizer. Existing post-training quantization (PTQ) methods suppress peaks, balance channels, or minimize reconstruction error, yet they rarely specify what activation distribution is actually easy to discretize. As a result, activations may appear numerically smoother while still incurring large quantization error because the quantization range remains wide or most values collapse into a few levels near the mean. We recast activation transformation as quantizer-facing distribution design and analyze quantization error from an information-theoretic perspective. Our analysis shows that quantization-friendly activations should jointly have a smaller numerical range and sufficient dispersion within that range. Guided by this analysis, we propose InfoQuant, a train-free method that employs Peak Suppression Orthogonal Transformation (PSOT) to shape activations into more quantization-friendly distributions. We further introduce adaptive outlier-token selection to improve the robustness of PSOT during optimization. Across multiple LLM families, InfoQuant consistently outperforms prior PTQ and end-to-end training baselines. Under W4A4KV4, it preserves 97% of floating-point accuracy on average and reduces the LLaMA-2 13B performance gap by 42% over the previous state of the art. Code is available at [ this https URL ]( this https URL )\n\n---\n\n## 115. A Universal Cliff and a Design Fingerprint: Cross-Section Defect Detection Under LLM Orchestration\n- **Authors**: Hiroki Fukui\n- **URL**: [https://arxiv.org/abs/2605.26174](https://arxiv.org/abs/2605.26174)\n- **Abstract**:\n> Production language-model systems answer a request by partitioning it across an invisible orchestration of worker agents that recompose one integrated report. We ask what this does to a class of defect no single worker can see: a contradiction in the relation between two distant sections of a document. Holding the documents, defects, mechanism, scoring, and seed fixed, we vary only the model -- ten systems across five generations from one developer and five providers from distinct alignment paradigms. Two layers separate. First, a universal detection cliff: every model that finds these cross-section defects under a single agent loses that ability under orchestration, detection falling two-thirds or more across every paradigm tested. The cliff is mechanism-derived and not closed by scale or extended reasoning. Second, how models behave once fallen. A signal-detection decomposition shows that, among the six models discriminating above chance, only one developer's generations move along the reporting-criterion axis: as alignment is strengthened, the model misses fewer defects yet raises more false alarms on clean documents -- two faces of one criterion shift, scaling with generation within that developer (p < 0.001) and near-absent elsewhere. At the floor the missed defect is often not out of view: the model's private record reconstructs the structural fault accurately, while the integrated report signs off on its soundness, its concern spent on the artifact and an absent collaborator. This resists quantification -- an automated judge is unstable (precision 17-50%) and keywords cannot separate it from ordinary agreement -- a resistance we report as a finding. We release all runs, probes, defect keys, scorer prompts, and scripts. An integrated report's confidence is uninformative about partition-spanning defects, the most aligned systems are not the safest, and the cliff is structural.\n\n---\n\n## 116. Tool-Schema Compression Enables Agentic RAG Under Constrained Context Budgets\n- **Authors**: Furkan Sakizli\n- **URL**: [https://arxiv.org/abs/2605.26165](https://arxiv.org/abs/2605.26165)\n- **Abstract**:\n> Agentic RAG systems that equip language models with dozens to hundreds of tool definitions face a critical resource conflict: tool schemas consume the same context window needed for retrieval-augmented generation. We present the first systematic study of this tool-context trade-off, evaluating 14 models spanning 1.5B-32B local models plus one frontier API model across 6,566 controlled API calls at three context budgets (8K, 16K, 32K) with 28 tool definitions. Applying TSCG conservative-profile compression (44-50% schema token savings), we observe a binary enablement effect: at 8K tokens, JSON-schema tool definitions overflow the context window entirely, yielding near-zero EM (2.6% average), while compressed schemas restore RAG functionality with +20.5 pp average exact-match lift across all eight models (+24.7 pp among the six exhibiting full enablement). At 32K -- where both formats fit -- four of five tested models show delta <= 1 pp, confirming the effect is purely budget-driven. External validation on HotpotQA (50 multi-hop questions) shows +48 pp EM under the same overflow scenario. Frontier scaling tests demonstrate that JSON schemas overflow at ~494 tools while compressed schemas remain operational beyond 800 tools. Our results establish tool-schema compression as a necessary infrastructure layer for agentic RAG in constrained-context deployments. All code, data, and checkpoints are publicly available.\n\n---\n\n## 117. TSFMAudit: Data Contamination Auditing in Forecasting Time Series Foundation Models\n- **Authors**: Hongkai Li , Shifeng Xie , Lefei Shen , Zhuo Li , Mouxiang Chen , Xiaobin Zhang , Han Fu , Jianling Sun , Xiaoxue Ren , Chenghao Liu\n- **URL**: [https://arxiv.org/abs/2605.26161](https://arxiv.org/abs/2605.26161)\n- **Abstract**:\n> Time series foundation models (TSFMs) are increasingly pretrained on large corpora, raising concerns that evaluation datasets may have been exposed during pretraining and thus yield overly optimistic performance estimates. Auditing such contamination is challenging in time series because signals are continuous and heterogeneous, and often lack corpus documentation. To the best of our knowledge, this is the first work to study pretraining contamination auditing for TSFMs. We formalize the problem of pretraining contamination auditing for TSFMs and propose TSFMAudit, a method based on probe adaptation dynamics. Our key intuition is that contamination manifests as unusually efficient adaptation: after a fine tuning probe, contaminated datasets tend to exhibit faster loss reduction with smaller backbone movement. We evaluate TSFMAudit on 6 TSFMs and 187 datasets using documented training source evidence as supervision, and compare against 10 competitive baselines adapted from the LLM literature.\n\n---\n\n## 118. Furina: Fragmented Uncertainty-Driven Refusal Instability Attack\n- **Authors**: Tongxi Wu , Jian Zhang , Yang Gao\n- **URL**: [https://arxiv.org/abs/2605.26158](https://arxiv.org/abs/2605.26158)\n- **Abstract**:\n> Safety alignment in large language models (LLMs) and multimodal large language models (MLLMs) is commonly assumed to operate as a near-binary threshold mechanism. We challenge this assumption by revealing that safety behavior is governed by an instability region where small perturbations induce stochastic refusal decisions rather than deterministic outcomes. We develop a multi-metric diagnostic framework combining external and internal signals to characterize this instability. Through systematic experiments, we identify a characteristic diagnostic signature: inputs in unstable regimes exhibit elevated output uncertainty yet decreased internal safety activation, a decoupling phenomenon that explains why detection-based defenses fail against sophisticated attacks. Building on this framework, we introduce Furina, a jailbreak attack that deliberately induces this signature through fragmented, scene-anchored prompts without model-specific optimization. Furina outperforms strong single-turn and multi-turn baselines on HarmBench and achieves competitive results on MM-SafetyBench, demonstrating that uncertainty amplification provides a principled and transferable mechanism for understanding safety vulnerabilities. Code is available at: this https URL .\n\n---\n\n## 119. Turning Bias into Bugs: Bandit-Guided Style Manipulation Attacks on LLM Judges\n- **Authors**: Xianglin Yang , Bryan Hooi , Gelei Deng , Tianwei Zhang , Jin Song Dong\n- **URL**: [https://arxiv.org/abs/2605.26156](https://arxiv.org/abs/2605.26156)\n- **Abstract**:\n> The known stylistic biases in LLM judges, such as a preference for verbosity or specific sentence structures, present an underexplored security vulnerability. In this work, we introduce BITE (BIas exploraTion and Exploitation), a black-box adversarial framework that learns semantics-preserving edits to mislead an LLM judge and artificially inflate the scores it assigns. We cast the selection of stylistic edits as a contextual bandit problem and use a LinUCB policy to adaptively choose edits that maximize the judge's score without access to model parameters or gradients. Empirically, we test BITE across a diverse range of LLM judges and tasks, including both pointwise and pairwise comparisons on chatbot leaderboards and AI-reviewer benchmarks. BITE achieves an attack success rate exceeding 65% and raises scores by 1-2 points on a 9-point scale, all while preserving semantic equivalence. We further assess the attack's stealthiness, showing that BITE evades standard style-control methods and several detection baselines. Our findings expose a fundamental weakness in the LLM-as-a-judge paradigm and motivate robust, attack-aware evaluation. Our code is available at this https URL .\n\n---\n\n## 120. MemMorph: Tool Hijacking in LLM Agents via Memory Poisoning\n- **Authors**: Xuanye Zhang , Yongsen Zheng , Zhuqin Xu , Kaiyu Zhou , Bowen Shen , Haoran Ou , Tianwei Zhang , Kwok-Yan Lam\n- **URL**: [https://arxiv.org/abs/2605.26154](https://arxiv.org/abs/2605.26154)\n- **Abstract**:\n> LLM-driven agents are capable of selecting external tools to complete users' tasks. However, attackers could compromise such process, steering agents toward inappropriate/wrong tools and enabling malicious actions. Most existing attacks primarily manipulate the tool metadata, which is easily detectable by auditing and may lose effectiveness as modern agents increasingly adopt memory modules to refine tool selection policies through accumulated experience. This paper proposes MemMorph, the first attack that bias tool selection by poisoning the agent's long-term memory. Rather than explicitly dictating the tool invocation decision, MemMorph injects a small number of crafted records that are disguised as technical facts, incident reports, and operational policies. These poisoned records reshape the agent's contextual perception and decision-making process, leading it to autonomously infer and select the tool preferred by the attacker. Experiments across 3 benchmarks, 10 agent backbones, and 3 memory-module implementations show that MemMorph achieves up to 85.9% attack success rate with only three injected records, outperforming the strongest baseline by up to 25% while retaining potency under 3 representative defenses. Our findings expose long-term memory as a critical and under-explored attack surface in tool-augmented agents, urging the development of memory-level integrity safeguards.\n\n---\n\n## 121. VISTA: An End-to-End Benchmark for Visual Spec-to-Web-App Coding Agents\n- **Authors**: JunJia Guo , Yuhang Yao , Jiawei (Joe) Zhou , Jingdi Chen\n- **URL**: [https://arxiv.org/abs/2605.26144](https://arxiv.org/abs/2605.26144)\n- **Abstract**:\n> We present VISTA (VIsual Spec-To-App Benchmark), a benchmark for evaluating the end-to-end web-app generation capabilities of LLM-based agents. Unlike prior code generation benchmarks that focus on algorithmic tasks, VISTA targets realistic UI-centric development, where agents must produce functional, visually coherent applications from underspecified inputs. We define five prompt-information conditions that vary along two axes, visual/structural fidelity and stack constraint: (1) text only with free stack choice, (2) text with reference screenshots under three specified stacks, (3) text with reference screenshots under free stack choice, (4) text with screenshots and pruned Figma structure under a single specified stack, and (5) text with screenshots and pruned Figma structure under free stack choice. To enable robust evaluation, each page in the benchmark is manually annotated with interactive UI components and around three visual anchor points, addressing the well-known limitations of script-based testing tools such as Playwright in open-ended code generation settings. Evaluation combines DOM-grounded reference matching, behavior-specific browser tests, and CLIP-based visual similarity, jointly measuring structural alignment, behavioral completeness, and overall visual fidelity. We use VISTA to assess four agent systems drawn from two model families and two harnesses, finding that visual fidelity and functional correctness are partially decoupled across both input conditions and agents, and that agent editing style varies sharply but is largely orthogonal to task quality. VISTA establishes a rigorous and reproducible foundation for advancing agent-based software engineering research.\n\n---\n\n## 122. Eroding Trust in Real Speech: A Large-Scale Study of Human Audio Deepfake Perception\n- **Authors**: Nicolas M. Müller , Wei Herng Choong\n- **URL**: [https://arxiv.org/abs/2605.26136](https://arxiv.org/abs/2605.26136)\n- **Abstract**:\n> Audio deepfakes have improved rapidly recently, yet their effect on human trust in real speech remains unstudied. We present the largest listening study on audio deepfake perception to date, collecting 35,532 judgments from 1,768 participants across 138 text-to-speech and voice conversion systems. Our central finding is a skepticism shift: compared to a 2021 baseline, human accuracy on fake samples barely changed (72.9% to 71.2%), but accuracy on real samples dropped from 72.7% to 64.1%. Participants are not worse at detecting synthesis artifacts; rather, they increasingly distrust authentic speech. Samples generated by commercial and autoregressive language model systems proved hardest to detect (61.3 - 65.9%), while those from traditional seq2seq and flow-matching models remain easier to spot (75.4 - 76.8%). An ML detector that served as a reference point maintained over 94.5% accuracy across all conditions. Our results suggest that the primary threat posed by modern deepfakes may not be mere deception, but the erosion of trust in genuine audio.\n\n---\n\n## 123. Pretraining Data Exposure in Large Language Models: A Survey of Membership Inference, Data Contamination, and Security Implications\n- **Authors**: Ziyi Tong , Feifei Sun , Le Minh Nguyen\n- **URL**: [https://arxiv.org/abs/2605.26133](https://arxiv.org/abs/2605.26133)\n- **Abstract**:\n> Large Language Models (LLMs) have become the predominant paradigm in NLP, advancing both research and industry. As model sizes and pretraining data grow, concerns about Pretraining Data Exposure (PDE) increase due to the scale and opacity of training datasets. PDE refers to determining whether specific data appeared in an LLM's pretraining corpus. It is critical for ensuring evaluation integrity and protecting privacy, intersecting two key areas: data contamination and membership inference. Though conceptually related, these areas have often been studied in isolation. This paper offers the first unified survey of both under the PDE framework. We formalize PDE across exposure levels, review attack and defense methods, synthesize empirical findings, and highlight open challenges and future research directions.\n\n---\n\n## 124. GEM: Geometric Entropy Mixing for Optimal LLM Data Curation\n- **Authors**: Yue Min , Ziyun Qiao , Ruining Chen , Yujun Li\n- **URL**: [https://arxiv.org/abs/2605.26121](https://arxiv.org/abs/2605.26121)\n- **Abstract**:\n> LLM pre-training efficacy increasingly depends on data composition rather than sheer volume. Yet, optimal mixing is hindered by categorization flaws: human taxonomies suffer from ontological misalignment, and Euclidean clustering fails to address embedding anisotropy. We introduce GEM (Geometric Entropy Mixing), a framework reformulating data curation as a variational problem on the hypersphere augmented with a mixing-balance regularizer. By decoupling the generative prior and optimizing the objective via a provable MM (Minorize-Maximize) algorithm, GEM effectively counteracts the cluster collapse to discover balanced semantic structures invisible to Euclidean heuristics. We employ teacher-student distillation to scale this geometric fidelity to web-scale corpora and introduce the Geometric Influence Score (GIS) for interpretable taxonomy generation. Experiments with 1.1B-parameter models demonstrate that GEM establishes a new state-of-the-art when integrated into mixing strategies like DoReMi and RegMix, improving average downstream accuracy by up to 1.2% and offering a robust coordinate system for predictable data mixing.\n\n---\n\n## 125. Xe-Forge: Multi-Stage LLM-Powered Kernel Optimization for Intel GPU\n- **Authors**: Marcin Spoczynski , Daniel Fleischer , Moshe Berchansky , Gabriela Ben-Melech Stan , Shira Guskin , Weilin Xu , Adam Siemieniuk , Alexander Heinecke\n- **URL**: [https://arxiv.org/abs/2605.26118](https://arxiv.org/abs/2605.26118)\n- **Abstract**:\n> Porting deep learning algorithms to new hardware accelerators requires developers to repeatedly apply the same low-level optimizations -- quantization, memory access coalescing, tile size tuning, and architecture-specific workarounds -- to every Triton kernel in their code-base. This manual, repetitive effort is a major bottleneck: each kernel demands the same cycle of trial-and-error profiling against hardware constraints that vary across devices, yet the underlying optimization patterns remain largely consistent. We present Xe-Forge, a multi-stage LLM-powered pipeline that automates this process for Intel GPU. Given a functionally correct Triton kernel, the system applies up to nine optimization stages -- from algorithmic restructuring and operator fusion through block pointer modernization, GPU-specific tuning, and open-ended discovery -- each driven by a Chain-of-Verification-and-Refinement (CoVeR) agent that generates candidates, validates them on real hardware, and iterates on failures. A curated knowledge base encodes Intel GPU constraints (power-of-two warp counts, GRF modes, SLM sizing) that are absent from LLM training data, keeping the model within architecturally valid bounds. We evaluate Xe-Forge on 97 Level-2 KernelBench kernels and Flash Attention on the Intel Arc Pro B70, achieving a 1.17x geometric mean speedup over PyTorch eager with 67% of kernels improving, nine kernels exceeding 5x (up to 82x), and 2--13.3x speedups on Flash Attention across all tested configurations without regression -- demonstrating that structured domain knowledge with hardware-in-the-loop verification can systematically eliminate the repetitive porting effort that currently gates algorithm deployment on new accelerators.\n\n---\n\n{% endraw %}\n"}, "c234ff9921f9be8d378a7558752264c977290646": {"t": 1786068102.232893, "n": 210239, "h": "cf2fc0302148c99f", "c": "---\nlayout: default\ntitle: \"LLM 관련 주요 논문 - 2026-05-27\"\n---\n\n{% raw %}\n# LLM 관련 주요 논문 - 2026-05-27\n\n## 1. MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation\n- **Authors**: Huawei Lin , Peng Li , Jie Song , Fuxin Jiang , Tieying Zhang\n- **URL**: [https://arxiv.org/abs/2605.27366](https://arxiv.org/abs/2605.27366)\n- **Abstract**:\n> Large language model (LLM) agents rely on reusable skills to solve complex tasks. However, existing skill creation approaches treat skills as isolated and static artifacts, limiting their reusability, reliability, and long-term improvement. We propose MUSE-Autoskill Agent (Memory-Utilizing Skill Evolution), a skill-centric agent framework that lets agents continuously improve their task-solving capability by creating, reusing, and refining skills under a unified lifecycle (creation, memory, management, evaluation, and refinement). Our framework enables agents to create skills on demand, store and reuse them across tasks, organize and select them efficiently, and evaluate them through unit tests and runtime feedback for continuous refinement. We further introduce skill-level memory that accumulates experience for each skill across tasks, enabling more effective reuse and adaptation over time. Experiments on SkillsBench provide initial evidence that lifecycle-managed skills can improve task success, efficiency, reuse, and cross-agent transfer, highlighting the importance of treating skills as long-lived, experience-aware, and testable assets.\n\n---\n\n## 2. Natural Language Query to Configuration for Retrieval Agents\n- **Authors**: Melissa Z. Pan , Negar Arabzadeh , Mathew Jacob , Fiodar Kazhamiaka , Esha Choukse , Matei Zaharia\n- **URL**: [https://arxiv.org/abs/2605.27361](https://arxiv.org/abs/2605.27361)\n- **Abstract**:\n> Modern retrieval agents expose many configuration choices -- LLM, retriever, number of documents, number of hops, and synthesis strategy -- each shaping both answer quality and serving cost. Today, these pipelines are typically hand-tuned once per workload, leaving substantial per-query optimization untapped. We formulate the problem: given a natural-language query and either an accuracy or a budget target, select from a predefined pipeline catalog the configuration that minimizes cost or maximizes accuracy at inference time. We propose **BRANE**, which uses an LLM to convert each query into workload-specific characteristics, then trains a lightweight per-configuration predictor that estimates whether the pipeline will answer the query correctly. At inference time, **BRANE** selects the configuration that maximizes predicted correctness penalized by cost, exposing a tunable cost-quality tradeoff without retraining. Across MuSiQue, BrowseComp-Plus, and FinanceBench, **BRANE** consistently pushes the cost-quality Pareto frontier, matches the best fixed configuration's accuracy at up to 89% lower cost, and outperforms LLM-routing, rule-based, and fine-tuned Qwen3-4B baselines. These results show that per-query configuration of the full retrieval pipeline is a practical alternative to static workload-level tuning.\n\n---\n\n## 3. Alignment Tampering: How Reinforcement Learning from Human Feedback Is Exploited to Optimize Misaligned Biases\n- **Authors**: Dongyoon Hahm , Dylan Hadfield-Menell , Kimin Lee\n- **URL**: [https://arxiv.org/abs/2605.27355](https://arxiv.org/abs/2605.27355)\n- **Abstract**:\n> Reinforcement Learning from Human Feedback (RLHF) is the standard method to align Large Language Models (LLMs) with human preferences. In this work, we introduce alignment tampering, a potential vulnerability where the LLM undergoing alignment influences the preference dataset, causing RLHF to amplify undesired behaviors. This arises from core limitations of RLHF: (1) preference datasets are constructed from the LLM's own outputs, allowing it to influence them, and (2) pairwise comparisons only indicate which response is better, not why. These limitations can be exploited to cause alignment tampering. For example, if an LLM generates biased responses with higher quality, annotators will prefer them based on quality. However, preference labels do not distinguish quality from bias, and the reward model inherits this limitation. Optimizing such rewards through reinforcement learning or best-of-N sampling can amplify misaligned biases. Our experiments demonstrate amplification across diverse biases: from keyword bias to propaganda (e.g., sexism), brand promotion, and instrumental goal-seeking. Mitigation remains challenging, as existing techniques for robust RLHF fail to fully resolve alignment tampering without sacrificing response quality. These findings reveal structural vulnerabilities of current RLHF and emphasize the need to prevent this vulnerability. Project page: this https URL\n\n---\n\n## 4. Gumbel Machine: Counterfactual Student Writing Generation via Gumbel Noise Steering\n- **Authors**: Hunter McNichols , Alexander Scarlatos , Mihai Dascalu , Danielle McNamara , Andrew Lan\n- **URL**: [https://arxiv.org/abs/2605.27249](https://arxiv.org/abs/2605.27249)\n- **Abstract**:\n> An effective method of teaching across disciplines is to provide examples of high-quality work. However, an example may be significantly different from a student's current work, making it challenging for them to emulate. An ideal learning demonstration is a counterfactual version of the student work, an improved version that is still similar to their own. Existing automated approaches for counterfactual text generation using Large Language Models (LLMs) result in domain-specific systems that are difficult to translate into practical applications. We present the Gumbel Machine, a flexible, modular approach to generating counterfactuals that leverages LLM instruction-following capabilities while encouraging similarity to a reference factual text. Central to our approach is a novel, controlled decoding algorithm, $\\beta$-Hindsight control, which uses latent randomness as a tunable similarity control mechanism during counterfactual generation. Experiments on datasets of student writing, scored on various criteria, demonstrate the effectiveness of our approach at generating counterfactuals both rubric-consistent and similar to a reference.\n\n---\n\n## 5. Learning to Act under Noise: Enhancing Agent Robustness via Noisy Environments\n- **Authors**: Yuxin Chen , Xiaodong Cai , Junfeng Fang , Zhuowen Han , Yu Wang , Yaorui Shi , Yi Zhang , Qi Gu , Xunliang Cai , Xiang Wang , An Zhang , Tat-Seng Chua\n- **URL**: [https://arxiv.org/abs/2605.27209](https://arxiv.org/abs/2605.27209)\n- **Abstract**:\n> Recent advances in large language models (LLMs) have facilitated the widespread deployment of LLMs as interactive agents capable of reasoning, planning, and tool use. Despite strong performance on existing benchmarks, such agents often exhibit notable degradation when deployed in real-world settings, where environments are inherently stochastic and imperfect. We argue that this discrepancy arises from a fundamental mismatch between idealized training settings and real-world interaction dynamics, where current paradigms rely on carefully curated task instructions and stable, well-controlled environments. To address this gap, we propose NoisyAgent, an agentic training framework that explicitly incorporates environmental imperfections into the agent learning process. We identify two major sources of interaction noise in real-world scenarios: user noise, which captures ambiguity and variability in user interaction, and tool noise, which reflects failures and anomalies in tool execution. We introduce such perturbations into the training pipeline by modifying user interaction patterns and simulating tool execution results within the training environment. To stabilize training while encouraging agents to handle increasingly challenging imperfections, noise is applied to only a subset of rollouts and progressively increased in difficulty as the model adapts to the current noise level. Extensive experiments demonstrate that our approach consistently improves agent robustness under noisy and dynamic environments. Our analysis reveals that training under noise conditions also yields performance gains on idealized benchmarks, suggesting that controlled exposure to environmental noise promotes more generalizable reasoning and decision-making behaviors. Our findings highlight the importance of modeling interaction imperfections for bridging the gap between agent training and real-world deployment.\n\n---\n\n## 6. The Compressive Knowledge Graph Hypothesis: Which Graph Facts Matter for Scientific Hypothesis Generation?\n- **Authors**: Shashwat Sourav , Viktoriia Baibakova , Sanjay Das , Ran Elgedawy , Maria Mahbub , Emily Herron , Tirthankar Ghosal\n- **URL**: [https://arxiv.org/abs/2605.27176](https://arxiv.org/abs/2605.27176)\n- **Abstract**:\n> Knowledge graphs (KGs) can provide structured scientific context to language models, but it remains unclear which graph facts actually shape the generated hypotheses. We study KG-guided hypothesis generation for battery materials across Mistral-7B, Llama-3.1-70B, and Gemini 2.5 Flash. We perturb local KGs by varying density, ontology richness, topology, and control structure, and evaluate outputs with both provided-graph and fixed-reference metrics. Across models, KG utility is selective and model-dependent: graph context changes outputs, but no-KG outputs also recover substantial graph content from model priors. Compact top-k subgraphs often approximate full-KG behavior, including when claimed-outcome triples are held out. At the same time, compression is not unique to one semantic ranking rule, random and topology-based subsets can also recover much of the signal. These results support a redundancy-aware Compressive KG hypothesis: useful KG signal is often recoverable from compact, scientifically structured subgraphs rather than requiring the full local graph.\n\n---\n\n## 7. VitaBench 2.0: Evaluating Personalized and Proactive Agents in Long-Term User Interactions\n- **Authors**: Yuxin Chen , Yi Zhang , Zhengzhou Cai , Yaorui Shi , Zhiyuan Yao , Chenhang Cui , Jingnan Zheng , Yaqi Huo , Xi Su , Qi Gu , Xunliang Cai , Xiang Wang , An Zhang , Tat-Seng Chua\n- **URL**: [https://arxiv.org/abs/2605.27141](https://arxiv.org/abs/2605.27141)\n- **Abstract**:\n> Large language models (LLMs) have evolved into interactive agents that collaborate with users in real-world tasks. Effective collaboration in such settings increasingly depends on understanding the user beyond what is explicitly stated, as user intent is often reflected in fragmented daily interactions and requires both personalized modeling and proactive interaction. However, existing agent benchmarks primarily evaluate reasoning and tool use, largely overlooking the challenges of inferring and leveraging user preferences in realistic scenarios. To address this gap, we introduce VitaBench 2.0, a benchmark for evaluating personalized and proactive agent behavior in long-term user interactions. In VitaBench 2.0, tasks are organized as temporally ordered sequences for individual users, where preferences are embedded in fragmented and heterogeneous interactions. Successful completion of tasks requires the agent to continuously extract, utilize, and update user preferences from these interactions. We further evaluate proactiveness through tasks that require agents to recognize missing information and actively acquire it from users or environments before making decisions. To support systematic analysis, we provide an extensible memory interface that enables controlled comparison across different memory architectures. We benchmark a diverse set of frontier proprietary and open-source LLMs. Results show that real-world personalization remains highly challenging even for state-of-the-art models, revealing a substantial gap between current capabilities and practical requirements. Extensive analysis further reveals the failure modes and capability bottlenecks of current agents in real-world personalized decision-making, providing insights for future model improvements.\n\n---\n\n## 8. ICCU: In-Context Continual Unlearning via Pattern-Induced Refusal Rules\n- **Authors**: Ruihao Pan , Suhang Wang\n- **URL**: [https://arxiv.org/abs/2605.27138](https://arxiv.org/abs/2605.27138)\n- **Abstract**:\n> Machine unlearning aims to remove the influence of specific data from trained language models. In real-world deployments, unlearning requests often arrive sequentially, which challenges existing fine-tuning-based methods: fine-tuning each request is costly, accumulates utility loss, and may cause cross-request interference. To address these issues, we propose ICCU (In-Context Continual Unlearning), an in-context continual unlearning framework that induces readable refusal rules from unlearning datasets and applies them at inference time either as a filter or via the system prompt, without modifying model parameters. Because rules are accumulated as an order-independent union, ICCU is compositional and free of cross-request interference, and the original forget-set data can be discarded after rule induction. Extensive experiments show that ICCU effectively suppresses target knowledge while preserving utility, scales across sequential requests, and remains robust to paraphrased and cross-lingual queries.\n\n---\n\n## 9. Scaling, Benchmarking, and Reasoning of Vision-Language Agents for Mobile GUI Navigation\n- **Authors**: Heng Qu , Yike Liu , Renren Jin , Wenzong Zhang , Pengzhi Gao , Wei Liu , Jian Luan\n- **URL**: [https://arxiv.org/abs/2605.27134](https://arxiv.org/abs/2605.27134)\n- **Abstract**:\n> Vision-Language Models (VLMs) have shown rapid progress in mobile GUI navigation. This paper presents a systematic study of data scaling, benchmarking, and reasoning for VLM-based agents in this domain. To facilitate rigorous evaluation, we introduce HyperTrack, a large-scale dataset with over 16000 real-world tasks across more than 650 Chinese mobile applications, along with GUIEvalKit, an open-source toolkit for unified benchmarking of VLMs on offline GUI navigation tasks. Using HyperTrack, we analyze the effects of training data scale on both supervised and reinforcement-based finetuning. Our results show that reinforcement-based finetuning consistently outperforms supervised finetuning, particularly in out-of-domain settings, highlighting the synergy between data scaling and reinforcement learning. Leveraging GUIEvalKit, we further benchmark state-of-the-art (SOTA) VLMs and analyze how interaction history and reasoning capabilities influence task completion. Together, HyperTrack and GUIEvalKit provide a comprehensive platform for developing and evaluating VLM agents in mobile GUI navigation tasks.\n\n---\n\n## 10. Position: AI Safety Requires Effective Controllability\n- **Authors**: Yige Li , Yunhao Feng , Jun Sun\n- **URL**: [https://arxiv.org/abs/2605.27117](https://arxiv.org/abs/2605.27117)\n- **Abstract**:\n> AI safety is still largely framed as alignment: training models to follow human preferences, safety policies, and normative constraints. That framing has improved the behavior of modern language models, but aligned behavior does not by itself guarantee that a deployed agent can be stopped, overridden, or constrained once it operates in open-ended, interactive, and tool-using environments. A system may be safe in expectation and still fail to yield to explicit runtime authority under conflicting instructions, long-horizon execution, adversarial inputs, or risky tool use. This position paper argues that AI safety therefore requires controllability as a first-class objective. We define \\emph{controllability} as the ability of an AI system to remain reliably interruptible, overridable, redirectable, and constrainable by explicit control signals at runtime while preserving ordinary utility when such signals are absent. To study this gap, we introduce \\controlbench{}, a benchmark for evaluating controllability failures in high-risk agentic scenarios. Experiments with OpenClaw-based agents show that current alignment and guardrail mechanisms reduce risk, but often fail to provide persistent, authoritative, and enforceable runtime control. We therefore propose a control-centric architectural framework that highlights explicit control planes, runtime intervention pathways, persistent control states, and auditable decision interfaces as key design principles for future controllable AI systems.\n\n---\n\n## 11. Counteraction-Aware Multi-Teacher On-Policy Distillation for General Capability Recovery with Domain Preservation\n- **Authors**: Tianlei Chen , Jiao Ou , Ziyuan Liu , Ruiming Tang , Jian Liang , Han Li\n- **URL**: [https://arxiv.org/abs/2605.27115](https://arxiv.org/abs/2605.27115)\n- **Abstract**:\n> Domain specialization can improve LLM behavior in vertical domains, but often weakens the general capabilities inherited from the original model. Recent Multi-Teacher On-Policy Distillation (MOPD) pipelines recover model capabilities by supervising student-generated trajectories with teacher feedback, but typically assume teacher-aligned prompt coverage, requiring prompts to match the teachers' training distributions. This assumption is difficult to satisfy when the general teacher is an open-source model whose post-training data are unknown. Instead of attempting to reconstruct this hidden distribution, we study general capability recovery with readily available proxy general prompts. We identify two failure modes of vanilla MOPD in this incomplete-coverage situation: recovery-preservation counteraction from mixing conflicting recovery and preservation gradients, and weak-signal flattening from uniformly averaging samples with unequal correction demand. We propose Counteraction-Aware Multi-Teacher On-Policy Distillation (CaMOPD), which addresses these issues with decoupled alternating training and gap-based sample selection. CaMOPD gives general recovery dedicated updates, periodically reviews domain prompts for preservation, and selects samples with larger averaged token-level teacher-student log-probability gaps to concentrate correction signals. Across role-play dialogue and medical reasoning QA scenarios, CaMOPD performs best in general recovery over baselines while maintaining domain-specific behavior. Gradient coherence analyses further support the intended effect of CaMOPD in producing more coherent correction signals.\n\n---\n\n## 12. Traceable Knowledge Graph Reasoning Enables LLM-Assisted Decision Support for Industrial VOCs in the Steel Industry\n- **Authors**: Changqing Su , Yu Ding , Zuhong Lin , Hongyu Liu , Xi He , Zheng Zeng , Liqing Li\n- **URL**: [https://arxiv.org/abs/2605.27071](https://arxiv.org/abs/2605.27071)\n- **Abstract**:\n> Key knowledge for steel-industry volatile organic compounds (VOCs) governance is scattered across unstructured scientific literature, making it difficult to integrate process, pollutant, and control-technology evidence and increasing the risk of hallucination when general large language models (LLMs) answer low-frequency industrial questions. Here we developed Chat-ISV, a knowledge graph (KG) enhanced multi-agent Q&A system that parses a curated steel-industry VOCs literature corpus, constructs a Neo4j KG with 27180 nodes and 81779 semantic edges, and combines prompt-constrained extraction, chunk-centered topology optimization, multi-agent routing, source-backtracking retrieval, local literature retrieval, open-domain knowledge access, and interactive subgraph visualization. Benchmark tests and 400 expert blind evaluations showed that topology optimization reduced isolated nodes from 57% to 4.08% and that Chat-ISV achieved high factual reliability, with 96.93% precision, 72.63% recall, an F1-score of 0.830, and a mean score of 1.69/2.00. By converting fragmented environmental-engineering literature into traceable, queryable, and decision-support-oriented knowledge, Chat-ISV establishes a scalable environmental-informatics paradigm for reliable LLM deployment and intelligent pollution-control decision support in specialized industrial domains.\n\n---\n\n## 13. Generating Robust Portfolios of Optimization Models using Large Language Models\n- **Authors**: Eleni Straitouri , Cheol Woo Kim , Milind Tambe\n- **URL**: [https://arxiv.org/abs/2605.27013](https://arxiv.org/abs/2605.27013)\n- **Abstract**:\n> Mathematical optimization is a powerful tool for structured decision-making across domains such as resource allocation and planning. Formulating optimization models faithful to reality, though, remains a significant bottleneck as it typically demands both domain expertise and optimization knowledge that are often scarce. Recent advances in large language models (LLMs) promise to bridge this gap, enabling the generation of candidate optimization models from natural language descriptions. However, there is no guarantee that any single LLM-generated model is reliable, and existing approaches that output only one model are therefore risky. In this work, we propose a novel algorithm that generates a portfolio of optimization models, designed to be robust to the limitations of LLMs. Our method exploits the observation that a single LLM can play two distinct roles $\\unicode{x2014}$ as a stochastic generator and as a reasoning evaluator $\\unicode{x2014}$ and proposes a unified framework that leverages both capabilities in a complementary manner. We provide theoretical guarantees showing that, as long as either the generator or the evaluator is well-aligned with human preferences, the portfolio is guaranteed to contain high-quality candidates, enabling a principled human-in-the-loop process in which a decision-maker can review multiple candidates before committing to one. We further validate our approach empirically, demonstrating strong performance across a range of optimization modeling tasks.\n\n---\n\n## 14. LELA: An End-to-end LLM-based Entity Linking Framework with Zero-shot Domain Adaptation\n- **Authors**: Samy Haffoudhi (IP Paris, LTCI, DIG), Nikola Dobričić (IP Paris), Fabian Suchanek (IP Paris, LTCI), Nils Holzenberger\n- **URL**: [https://arxiv.org/abs/2605.26956](https://arxiv.org/abs/2605.26956)\n- **Abstract**:\n> Entity linking is a key component of many downstream NLP systems, yet existing approaches are often tied to the specific target knowledge bases and domains, limiting their real world application. In this paper, we extend LELA, a modular and domain-agnostic LLM-based entity disambiguation method, into a practical Python library that integrates zero-shot Named Entity Recognition (NER) -thereby providing a complete end-toend pipeline for entity-linking in real-world usage. We provide experimental results validating LELA's performance and robustness across diverse entity linking settings. In our demo, users can play with the system on their own input texts.\n\n---\n\n## 15. Neuro-Symbolic Verification of LLM Outputs for Data-Sensitive Domains (extended preprint)\n- **Authors**: Paul Sigloch , Christoph Benzmüller\n- **URL**: [https://arxiv.org/abs/2605.26942](https://arxiv.org/abs/2605.26942)\n- **Abstract**:\n> LLMs deployed in high-stakes domains face fundamental reliability challenges: hallucinations, inconsistencies, and privacy vulnerabilities introduce unacceptable risks where errors carry legal, financial, or safety consequences. This paper presents a hybrid verification architecture combining formal symbolic methods with neural semantic analysis to provide complementary guarantees for LLM-generated content. This architecture employs logical reasoning for input verification, leveraging completeness properties to provide decidable guarantees on structured requirements. For output validation, embedding-based semantic similarity detects contextual hallucinations where formal methods lack expressiveness. This separation is realized in a parallel, actor-based pipeline, addressing limitations of prompt-based self-verification approaches, which inherit the distributional biases that produce hallucinations. The proposed architecture and type-aware verification method are validated with HAIMEDA, a real-world medical device damage assessment reporting system developed through Action Design Research. Evaluation shows hallucination detection rates of over 83% for structured entities and 72% for semantic fabrications, with a 30% reduction in report creation time, demonstrating that neuro-symbolic architectures can provide principled safeguards for LLM deployment in data-sensitive domains.\n\n---\n\n## 16. From Norms to Indicators (N2I-RAG): An Agentic Retrieval-Augmented Generation Framework for Legal Indicator Computation\n- **Authors**: Youssef Al Mouatamid , Marie Bonnin , Jihad Zahir\n- **URL**: [https://arxiv.org/abs/2605.26926](https://arxiv.org/abs/2605.26926)\n- **Abstract**:\n> Computing legal indicators from normative texts is a key task in legal monitoring and policy evaluation, but presents significant challenges due to the complexity, scale, and interpretive nature of legal language, as well as the variability in available document quality. Existing natural language processing techniques and generative models can assist in legal analysis, but often suffer from high risk of hallucinations and lack the interpretability and evidence grounding required for reliable indicator computation. This paper presents N2I-RAG (From Norms to Indicators), an agentic retrieval-augmented generation framework designed to automate the computation of legal indicators in a transparent and traceable way. We integrate adaptive retrieval, llm-based agents, and validation mechanisms in a modular pipeline, where each component performs a defined role in filtering, retrieving, and assessing evidence, and in producing binary legal outcomes linked to identifiable legal provisions. The framework emphasizes traceability by requiring explicit explanations of intermediate decisions and final indicator assignments. We evaluate N2I-RAG using an in-house constructed French marine environmental law corpus that includes both scanned and digital sources. Comparative experiments with multiple language model families demonstrate that the proposed approach consistently outperforms baseline systems, and generalizes well when tested on 2 different bans. The results indicate that agentic retrieval-augmented generation can bridge open-text legal language and standardized indicator computation, offering a foundation for transparent and scalable legal observatories.\n\n---\n\n## 17. TADDLE: A Tool-Augmented Agent for Detecting Deficient LLM-Generated Peer Reviews\n- **Authors**: Hanqi Duan , Xiang Li\n- **URL**: [https://arxiv.org/abs/2605.26911](https://arxiv.org/abs/2605.26911)\n- **Abstract**:\n> LLM-generated peer reviews are increasingly common at major venues, yet their deficiencies are hard to detect because they are uniformly fluent and well-structured. Existing work either classifies authorship without judging quality, or scores quality with features designed for human-written reviews; no prior system detects deficiencies in LLM-generated reviews at the level of individual defect types. To bridge the gap, we introduce TADDLE, a Tool-Augmented Agent for Detecting Deficient LLM-Generated Peer Reviews, together with the first expert-annotated benchmark for this task. Our benchmark comprises 1,800 reviews on 50 ICLR 2025 papers, multi-label-annotated by 18 domain experts against a taxonomy of six defect categories (plus a non-deficient label). TADDLE decomposes detection into four specialized analysis tools -- Verify, Correct, Complete, and Transform -- orchestrated by an agent; an integrator synthesizes their outputs into binary and multi-label classifications via two-stage semi-supervised learning. Extensive experiments show that TADDLE performs strongly on both binary detection and the multi-label classification task. We release the benchmark and code at this https URL .\n\n---\n\n## 18. Multi-Stakeholder LLM Alignment: Decomposing Estimation from Aggregation\n- **Authors**: Lulu Zheng , Wenjin Yang , Xiangwen Zhang , Rong Yin , Yulan Hu , Zheng Pan , Xin Li\n- **URL**: [https://arxiv.org/abs/2605.26878](https://arxiv.org/abs/2605.26878)\n- **Abstract**:\n> Multi-stakeholder tasks require one output to satisfy users with conflicting preferences. Holistic LLM judges conflate utility estimation and utility aggregation, yielding unstable implicit weights. We show empirically and theoretically that this aggregation-specific \\emph{weighting noise} can create large score shifts when stakeholder satisfaction is dispersed; in our experiments, these weight-induced shifts also increase with stakeholder count. We propose \\textsc{DecompR}: counterfactual-calibrated weights are fixed from query structure before candidate scoring, while per-role utilities are estimated independently, removing candidate-dependent weight drift and reducing estimation noise.\n\n---\n\n## 19. Helicase: Uncertainty-Guided Supply Chain Knowledge Graph Construction with Autonomous Multi-Agent LLMs\n- **Authors**: Yunbo Long , Haolang Zhao , Ge Zheng , Alexandra Brintrup\n- **URL**: [https://arxiv.org/abs/2605.26835](https://arxiv.org/abs/2605.26835)\n- **Abstract**:\n> LLM-based multi-agent systems have been widely adopted for knowledge retrieval and report generation, synthesizing known information through web search and textual reasoning. However, many critical information tasks in supply chains are not simple one-shot queries: they are structural inference problems requiring multi-hop reasoning across complex, fragmented web resources. Questions such as \\textit{``Which Tesla components use lithium from Australian mines?''} have no answer in any single document; answers must be computationally synthesized through the autonomous construction and analysis of dynamic knowledge graphs assembled from fragmented, heterogeneous sources. Moreover, such discovery processes must be uncertainty-aware: decisions depend not only on answers but on calibrated confidence in their reliability, traceable to source quality and reasoning consistency. To address this capability gap, we propose \\textit{Helicase}, an autonomous multi-agent LLM system for uncertainty-guided supply chain knowledge graph construction. \\textit{Helicase} decomposes high-level supply-chain queries into executable investigation plans, coordinates specialized web-search, reasoning, and coding agents through iterative verification loops, and incrementally constructs query-specific supply chain knowledge graphs with per-fact uncertainty annotations. Its three-layer uncertainty framework tracks uncertainty at the action, trajectory, and memory layers, enabling both structural inference and calibrated confidence assessment. To evaluate autonomous reasoning across the full complexity spectrum, we introduce SCQA (Supply Chain Query Assessment), a benchmark of 80 supply chain queries organized into four quadrants spanning single-hop to multi-hop inference under both high and low data visibility.\n\n---\n\n## 20. The Attribution Blind Spot: Detecting When Language Models Rely on Memory Rather Than Retrieved Context\n- **Authors**: Zhe Yu , Wenpeng Xing , Yunzhao Wei , Bo Yang , Chen Ye , Gaolei Li , Meng Han\n- **URL**: [https://arxiv.org/abs/2605.26778](https://arxiv.org/abs/2605.26778)\n- **Abstract**:\n> Retrieval-augmented generation promises to ground language model outputs in external evidence, yet the field has no reliable way to verify whether retrieved context actually governs generation -- a prerequisite for any high-stakes deployment. The standard assumption, that context-consistent output implies context-governed output, breaks when the retrieved document overlaps with the model's pretraining data: the model can produce faithful-looking text entirely from parametric memory, and both pathways yield indistinguishable output. We name this failure the attribution blind spot and introduce Computational Reality Monitoring (CRM) to address it. CRM operationalizes a principle adapted from cognitive science's reality monitoring framework: comparing internal representations with and without context reveals membership-conditioned representational divergence that output-level monitors systematically miss. CRM does not certify which source an individual generation used; it detects whether pretraining exposure leaves a measurable internal trajectory signature, establishing a necessary substrate for source attribution. Across nine model variants spanning three families, this divergence concentrates in architecture-specific layer patterns, receives converging support from block-level noise intervention, and generalizes across tasks and datasets while collapsing on domain-confounded benchmarks. The attribution blind spot is measurable and partially addressable: internal representations carry a diagnostic signal invisible at the output level, establishing a foundation for systems whose internal awareness of evidence provenance governs their external behavior.\n\n---\n\n## 21. A Dataset of Robot-Patient and Doctor-Patient Medical Dialogues for Spoken Language Processing Tasks\n- **Authors**: Heriberto Cuayahuitl , Grace Jang\n- **URL**: [https://arxiv.org/abs/2605.26747](https://arxiv.org/abs/2605.26747)\n- **Abstract**:\n> Large Language Models (LLMs) have brought huge improvements to Artificial Intelligence (AI), which can be applied to general-purpose tasks. However, their application to textual or spoken medical consultations is still an open research problem. This paper proposes MeDial-Speech, a novel speech dataset for training and evaluating Med-AIs that can carry out consultations with patients. It was collected in realistic environments from robot-patient and doctor-patient dialogues, contains 111+ hours of speech data (without data augmentation), and covers four health conditions: Lewy body dementia, heart failure, shoulder pain, and angina. In addition, we propose a dialogue benchmark via sentence selection (with 20 options) to evaluate three state-of-the-art LLMs: GPT-5 mini, DeepSeek-V3, and Claude Sonnet 4. Experimental results reveal that Claude Sonnet 4 is the best in sentence selection, with 71.1% accuracy using manual transcriptions and 74.7% using automatic transcriptions, and that all LLMs are highly overconfident in their probabilistic predictions, regardless of selecting correct or incorrect sentences in medical dialogues. This dataset is free of charge for non-commercial purposes at: this https URL\n\n---\n\n## 22. It's Not the Capability: Harness Sensitivity Is Non-Monotone Across LLM Agent Tiers\n- **Authors**: Yong-eun Cho\n- **URL**: [https://arxiv.org/abs/2605.26731](https://arxiv.org/abs/2605.26731)\n- **Abstract**:\n> A prevalent assumption in LLM agent deployment holds that more structured harnesses universally improve reliability, and that higher-capability models need proportionally less structural guidance -- together implying a monotone inverse relationship between model capability tier and optimal harness complexity. We test this hypothesis through a controlled 432-run experiment crossing six models across four capability tiers with three harness conditions (light, balanced, strict) on HEAT-24, a 24-task synthetic benchmark with git-based workspace verification. Our results refute the monotone inverse relationship on two fronts. First, for the frontier chat model evaluated (Gemini 2.5 Flash), increased harness verbosity lowers VTSR by 29-38 percentage points -- a harness-complexity paradox. Second, for the frontier reasoning model evaluated (Qwen3.5-122B, extended thinking enabled), strict harness achieves the highest VTSR (91.7%) and the lowest latency, the opposite of the prediction. Within the constrained tier, a 2B model (Gemma4:e2B) matches strong-open-tier stability at 91.7% across all harnesses. Because each tier is represented by a single model in this study, these results should be interpreted as model-specific observations; harness sensitivity appears non-monotone across the models evaluated, and depends critically on model type (chat vs. reasoning). We introduce a six-label failure taxonomy showing that format_violation dominates capable-model failures while wrong_file dominates low-capability failures, and we derive practical tier-aware harness selection guidelines.\n\n---\n\n## 23. Towards Feedback-to-Plan Decisions for Self-Evolving LLM Agents in CUDA Kernel Generation\n- **Authors**: Yee Hin Chong , Jiaming Wu , Youhui Zhang , Peng Qu\n- **URL**: [https://arxiv.org/abs/2605.26720](https://arxiv.org/abs/2605.26720)\n- **Abstract**:\n> Large language models (LLMs) have shown strong empirical gains as self-evolving agents for CUDA kernel generation, driven by feedback-conditioned planning across generations. However, how planning decisions attribute and combine heterogeneous feedback signals remains opaque. Standard end-to-end ablations fail to resolve this question, as iterative planning amplifies early perturbations and conflates feedback effects with trajectory-dependent drift. We introduce \\texttt{CUDAnalyst}, a unified analysis layer for controlled, generation-level attribution of planning decisions to feedback components via trajectory freezing and selective feedback injection. \\texttt{CUDAnalyst} enables stable generation-level evaluation and principled coalitional-style attribution of feedback effects and interactions. Our results show that explicit planning is beneficial only when feedback is aligned, that effective planning emerges from structured multi-feedback interactions, and that high-level plans from stronger reasoning models can partially transfer to weaker ones. These trends hold across reference backbones, representative workloads, and reference induction regimes, indicating that the identified feedback-to-plan structure is robust within the controlled axes studied.\n\n---\n\n## 24. MemFail: Stress-Testing Failure Modes of LLM Memory Systems\n- **Authors**: Ishir Garg , Neel Kolhe , Dawn Song , Xuandong Zhao\n- **URL**: [https://arxiv.org/abs/2605.26667](https://arxiv.org/abs/2605.26667)\n- **Abstract**:\n> Large language model (LLM) agents increasingly rely on external memory systems to remain consistent across long-horizon interactions, but little empirical work has been done to understand the specific failure modes and design choices that these systems present. Existing benchmarks report aggregate question-answering accuracy and treat memory systems as black boxes, making it impossible to attribute an incorrect answer to a particular failure mode of the system. We introduce MemFail, a diagnostic benchmark that isolates the failure modes of modern LLM memory systems. We begin by formalizing memory systems as the composition of three canonical operations -- summarization, storage, and retrieval -- and identify the potential failure modes induced by each. Based on these hypothesized failure modes, we construct five datasets spanning four tasks, each adversarially designed to test a specific operation of a memory system. Using these datasets, we evaluate four state-of-the-art memory systems on MemFail and demonstrate how MemFail can be used to empirically understand the tradeoffs induced by differences in memory system architectures.\n\n---\n\n## 25. UnityMAS-O: A General RL Optimization Framework for LLM-Based Multi-Agent Systems\n- **Authors**: Yiqun Chen , Wei Yang , Erhan Zhang , Shijie Wang , Qi Liu , Zechun Niu , Bin Zhang , Haitao Li , Rui Li , Lingyong Yan , Jinyuan Feng , Biqing Qi , Xiaochi Wei , Yan Gao , Yi Wu , Yao Hu , Jiaxin Mao\n- **URL**: [https://arxiv.org/abs/2605.26646](https://arxiv.org/abs/2605.26646)\n- **Abstract**:\n> LLM-based multi-agent systems decompose complex tasks into interacting roles, but most remain manually orchestrated by prompts, tools, and control rules, while agents are rarely optimized through a unified reinforcement learning interface. Existing RL post-training frameworks mainly target single-policy optimization and lack abstractions for user-defined multi-agent workflows, structured interaction, role-specific credit assignment, and configurable parameter sharing. We present UnityMAS-O, a general RL optimization framework for LLM-based multi-agent systems. UnityMAS-O treats the complete workflow as the optimization unit, rather than a single response or policy trajectory. It represents workflows through four first-class objects: logical agent roles, graph trajectories, user-defined rewards, and agent--model mappings. This decouples logical agents from physical model parameters, supporting full sharing, full separation, and partial sharing, with rewards assigned at role, turn, and trajectory levels. UnityMAS-O extends verl with a Ray-based star-topology runtime. A central controller executes workflows, invokes tools, records structured trajectories, and assembles rewards; model-local worker groups handle rollout, buffering, advantage computation, and distributed PPO-style updates. Users can define agents, workflows, model mappings, and rewards without rewriting the optimization infrastructure. We instantiate UnityMAS-O on retrieval-augmented QA, iterative agentic search, and reflective code generation. Across Natural Questions, HotpotQA, and held-out code tasks, multi-agent RL improves manually specified workflows after optimization, with especially large gains for smaller models and strict code all-passed metrics. These results show that UnityMAS-O can serve as a reusable substrate for converting diverse LLM-based multi-agent workflows into trainable multi-agent RL systems.\n\n---\n\n## 26. FAST-GOAL: Fast and Efficient Global-local Object Alignment Learning\n- **Authors**: Hyungyu Choi , Young Kyun Jang , Chanho Eom\n- **URL**: [https://arxiv.org/abs/2605.26615](https://arxiv.org/abs/2605.26615)\n- **Abstract**:\n> Vision-language models such as CLIP have shown impressive capabilities in aligning images and text, but they often struggle with lengthy and detailed text descriptions due to pre-training on short and concise captions. We present FAST-GOAL (Fast and Efficient Global-local Object Alignment Learning), an efficient fine-tuning method that enhances ability of CLIP to handle lengthy text through global-local semantic alignment. Our method consists of two key components. First, Fast Local Image-Sentence Matching (FLISM) efficiently extracts local image regions through object detection and spatial division, then matches them with corresponding sentences. Second, Token Similarity-based Learning (TSL) maximizes the similarity between patch tokens from specific regions in the image and their corresponding region embeddings, applying the same principle to text, which enhances the ability of the model to capture detailed correspondences. Additionally, we introduce GLIT100k, a dataset that provides both global image-lengthy caption pairs and context-derived local pairs, where local descriptions are extracted from global captions to maintain semantic coherence. Through extensive experiments on long caption datasets (DOCCI, DCI) and short caption datasets (MSCOCO, Flickr30k), we demonstrate that FAST-GOAL achieves significant improvements over baselines, enabling effective adaptation of CLIP to detailed textual descriptions while maintaining computational efficiency.\n\n---\n\n## 27. AGORA: Adapter-Grounded Observation-Action Retention for Inference-Free Prompt Compression in LLM Agents\n- **Authors**: Haoran Zhang , Zhaohua Sun\n- **URL**: [https://arxiv.org/abs/2605.26596](https://arxiv.org/abs/2605.26596)\n- **Abstract**:\n> The token-level extractive compressors widely used for general LM context are structurally inappropriate for LLM agents: across 17 (env, backbone, method) cells spanning two independent token-level method families, every cell collapses to mean reward <= 0.05 despite 1.3-13.3x realized compression. We name and characterize this failure mode as action-grammar destruction -- the tokens carrying action semantics (identifiers, brackets, action verbs) are exactly those self-information ranks lowest, so a general-purpose compressor reliably removes them and the environment rejects the residual. The diagnosis points to step-granularity compression. We introduce AGORA, an inference-free step-level compressor combining a structural prompt parser, an always-keep floor for format- and recency-critical content, and a 125M-parameter relevance scorer trained on counterfactual next-action-change labels (~2ms/step, zero per-step LLM toll). Across the compared inference-free and LLM-based methods, AGORA is the only one retaining >= 75% uncompressed performance in 8 of 9 cells (with the lone exception at 73%); a four-way component ablation isolates the structural floor as the dominant quality lever and the learned scorer as the source of 1.0-11.5x adaptive end-to-end compression from a single fixed keep ratio.\n\n---\n\n## 28. MedGuideX: Internalizing Decision Logic from Executable Guidelines into Large Language Models for Clinical Reasoning\n- **Authors**: Yuhao Shen , Lang Cao , Simo Du , Yuqing Wang , Juexiao Zhou , Hao Peng , Yue Guo\n- **URL**: [https://arxiv.org/abs/2605.26567](https://arxiv.org/abs/2605.26567)\n- **Abstract**:\n> Clinical practice guidelines (CPGs) encode evidence-based decision logic that clinicians apply by evaluating patient variables, conditional criteria, and recommendation rules. However, existing methods often use CPGs as free-text training data or retrieval sources, underutilizing their procedural decision structure. To better exploit this structure, we introduce a guideline-derived training pipeline that transforms CPG recommendations into executable clinical decision logic and uses it to generate factual and counterfactual question-answering data. Theses data teach models both guideline-supported decisions and how decisions change under different patient conditions. Post-training a medical LLM on the generated data yields MedGuideX. Across four clinical reasoning benchmarks, MedGuideX achieves a 10.28% relative improvement in average accuracy. Physician evaluation further shows that MedGuideX better recovers clinician authored reasoning steps and produces physician-preferred rationales in faithfulness, validity, completeness, and clarity. Overall, our results show that executable decision logic from CPGs can be transformed into scalable supervision for building reliable medical LLMs.\n\n---\n\n## 29. MobileExplorer: Accelerating On-Device Inference for Mobile GUI Agents via Online Exploration\n- **Authors**: Runxi Huang , Liyu Zhang , Shengzhong Liu , Xiaomin Ouyang\n- **URL**: [https://arxiv.org/abs/2605.26546](https://arxiv.org/abs/2605.26546)\n- **Abstract**:\n> Mobile graphical user interface (GUI) agents enable AI models to autonomously operate smartphones on behalf of users. However, most existing systems focus primarily on optimizing task accuracy and rely on cloud-hosted models for inference, which introduces privacy concerns and network-dependent latency. As a result, fully on-device deployment of mobile GUI agents remains underexplored. We propose MobileExplorer, a new framework that accelerates on-device inference for vision-based mobile GUI agents via online exploration. The key idea is to exploit the long per-step reasoning time of vision-language models (VLMs) by performing lightweight, parallel exploration of UI elements. During model inference, the agent proactively probes semantically relevant UI elements and records these exploration traces as structured memory. To ensure reliable execution in live mobile environments, we design a two-level rollback mechanism that robustly restores the initial UI state when a fast but naive backtracking strategy fails. The collected exploration traces are then summarized into concise contextual hints and injected into the prompt to enhance the subsequent reasoning step. We evaluate MobileExplorer on multiple off-the-shelf devices using the AndroidWorld benchmark, as well as newly designed, more complex tasks and dynamic on-device environments. MobileExplorer reduces the average number of reasoning steps and end-to-end latency by 23\\%, while maintaining or improving task success rates by up to 5\\%. A video demonstration of MobileExplorer performance in the real world is available at this https URL .\n\n---\n\n## 30. The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence\n- **Authors**: MiniMax : Aili Chen , Aonian Li , Baichuan Zhou , Bangwei Gong , Binyang Jiang , Boji Dan , Changqing Yu , Chao Wang , Cheng Ma , Cheng Zhong , Cheng Zhu , Chengjun Xiao , Chengyi Yang , Chengyu Du , Chenyang Zhang , Chi Zhang , Chuangyi Huang , Chunhao Zhang , Chunhui Du , Chunyu Zhao , Congchao Guo , Da Chen , Deming Ding , Dianjun Sun , Dongyu Zhang , Enhui Yang , Fei Yu , Guang Zheng , Guodong Zheng , Guohong Li , Haichao Zhu , Haigang Zhou , Haimo Zhang , Han Ding , Hao Zhang , Haohai Sun , Haolin Lyu , Haonan Lu , Haoyu Wang , Huajie Shi , Huiyang Li , Jiacheng Chen , Jian Zhang , Jiaqi Zhuang , Jiaren Cai , Jiaxin Pan , Jiayao Li , Jiayuan Song , Jichuan Zhang , Jie Wang , Jihao Gu , Jin Zhu , Jingwei Dong , Jingyang Li , Jingyu Zhang , Jingze Zhuang , Jinhao Tian , Jinli Liu , Jinyi Hu , Jun Tao , Jun Zhang , Junbin Ruan , Junhao Xu , Junjie Yan , Junteng Liu , Junxian He , Kang Xu , Ke Ji , Ke Yang , Kecheng Xiao , Keyu Duan , Keyu Li , Le Han , Letian Ruan , Li Yuan , Lianfei Yu , Liheng Feng , Lijie Mo , Lin Li , Lingye Bao , Lingyu Yang , Lingyuan Zhou , Loki , Lu Chen , Lunbin Ceng , Ming Li , Ming Zhong , Mingliang Tao , Mingyuan Chi , Mujie Lin , Nan Hu , Ningxin Chen , Peiyin Zhu , Peng Gao , Pengcheng Gao , Pengfei Li , Penglin Li , Pengyu Zhao , Qibin Ren\n- **URL**: [https://arxiv.org/abs/2605.26494](https://arxiv.org/abs/2605.26494)\n- **Abstract**:\n> We introduce the MiniMax-M2 series, a family of Mixture-of-Experts language models built around the principle that mini activations can unleash maximum real-world intelligence. The flagship M2 contains 229.9B total parameters with only 9.8B activated per token. Designed end-to-end for agentic deployment, the M2 series rests on three components: (i) agent-driven data pipelines producing large-scale, verifiable trajectories across agentic coding and agentic cowork, each grounded in an executable workspace and an artifact-aligned reward; (ii) Forge, a scalable agent-native RL system that adapts to long-horizon agent trajectories, paired with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and a clean training-inference-agent decoupling that supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step toward self-evolution -- autonomously debugging training runs and modifying its own scaffold. Across M2 through M2.7, this combination translates a mini-activation footprint into frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.\n\n---\n\n## 31. Reasoning, Code, or Both? How Large Language Models Handle Variations in Math Questions\n- **Authors**: Matthew Kutakh\n- **URL**: [https://arxiv.org/abs/2605.26414](https://arxiv.org/abs/2605.26414)\n- **Abstract**:\n> Large Language Models (LLMs) achieve impressive accuracy on mathematical reasoning benchmarks, yet their performance drops when problems are modified with simple changes like different names or numbers. Code execution methods, which let models generate and run Python code instead of reasoning in natural language, have been proposed as a solution, but their effect on reasoning robustness (the ability to maintain accuracy across problem variations) has not been systematically tested. This study evaluates three approaches on 1,000 problems from the GSM-Symbolic dataset: pure reasoning using chain-of-thought (CoT) prompting, single-shot code execution using Program-Aided Language models (PAL), and iterative code execution using Step-by-Step Coding (SBSC). All three were run on paired original and modified problems using Claude Haiku 4.5. CoT was the most robust method, with an accuracy drop of 1.3 percentage points and 1.8% of problems breaking under perturbation. PAL was the least robust at 1.7 percentage points and 3.1% broke, with SBSC falling in between. Although these differences were not statistically significant ($p = .096$), the directional trend was consistent across all measures, suggesting that code execution, whether single-shot or iterative, does not improve reasoning robustness on grade-school-level problem variations.\n\n---\n\n## 32. From Static Context to Calibrated Interactive RL: Mitigating Distribution Shift in Multi-turn Dialogue with Aligned Simulator\n- **Authors**: Xiaohua Wang , Jiakang Yuan , Zisu Huang , Muzhao Tian , Changze Lv , Kaitao Song , Tao Chen , Xiaoqing Zheng\n- **URL**: [https://arxiv.org/abs/2605.26403](https://arxiv.org/abs/2605.26403)\n- **Abstract**:\n> A long-standing goal of the research community is to develop highly interactive LLM-based dialogue agents. Recent research focuses on optimizing policies based on fixed offline logs (Static Context RL) or using a prompt-based simulator (Interactive RL). In this work, we theoretically show that both paradigms are fundamentally limited by context distribution shift--a mismatch between dialogue histories observed during training and those encountered in real conversations. This shift compounds quadratically over turns and severely degrades dialogue quality. Specifically, we attribute this shift to two distinct sources: (i) policy-induced shift, arising from training on static histories rather than self-generated trajectories; and (ii) simulator-induced shift, stemming from discrepancies between simulated and real human behaviors. To address these challenges, we propose Calibrated Interactive RL, a unified framework that couples interactive RL with simulator alignment. By aligning the simulator with human interaction patterns, our approach reduces the sim-to-real gap and mitigates compounding distribution shifts. Experiments across multiple dialogue tasks confirm our theoretical analysis: (i) Interactive RL significantly outperforms the Static Context baseline by mitigating policy distribution shift; and (ii) calibrating simulators with our alignment method further bridges the sim-to-real gap, yielding state-of-the-art downstream performance.\n\n---\n\n## 33. Automatic Layer Selection for Hallucination Detection\n- **Authors**: Xinpeng Wang , William Cao , Andrew Gordon Wilson , Zhe Zeng\n- **URL**: [https://arxiv.org/abs/2605.26366](https://arxiv.org/abs/2605.26366)\n- **Abstract**:\n> Recent studies on hallucination detection have shown that hallucination-related signals are more strongly encoded in intermediate layers than in the final layer of large language models (LLMs). Although a growing body of work has sought to exploit this property for hallucination detection, how to automate the selection of high-performing layers remains underexplored, and principled methods for this purpose are still lacking. To address this gap, we first propose several hypotheses for why such signals emerge in intermediate layers and evaluate corresponding criteria for automatic layer selection across diverse LLM architectures, scales, and tasks, covering both question answering and summarization hallucination detection benchmarks. However, we find that none of these criteria consistently delivers satisfactory performance. We therefore propose a new selection criterion, First Effective Peak of Intrinsic Dimension (FEPoID), which consistently identify optimal or near-optimal layers and outperforms both the aforementioned criteria and existing hallucination detection baselines. FEPoID is training-free and incurs negligible computational overhead. In addition, we study the generation behaviors of LLMs and introduce a simple yet effective truncation strategy, which further amplifies hallucination-related signals and substantially improves overall detection performance. Code is publicly available at this https URL\n\n---\n\n## 34. Managing Uncertainty in LLM-Generated Procedural Knowledge for Virtual Laboratory Planning\n- **Authors**: Polychronis Karpodinis , Dimitris Kalles\n- **URL**: [https://arxiv.org/abs/2605.26333](https://arxiv.org/abs/2605.26333)\n- **Abstract**:\n> Educational virtual laboratories can make experimental training more scala-ble, adaptive, and accessible, especially when students have limited access to physical laboratory facilities. However, authoring new simulated laboratory procedures remains costly: educators must describe new equipment, define how instruments and materials interact, and specify valid procedural flows that can be executed or assessed inside the virtual environment. Large lan-guage models can assist in this authoring process by generating detailed ex-perimental procedures, but their output should not be treated as directly exe-cutable plans. They may omit necessary actions, arrange steps in the wrong order, or produce instructions that are logically incorrect or incompatible with the laboratory equipment. This paper presents a prototype framework for managing uncertainty in LLM-generated procedural knowledge for virtu-al laboratory planning. The framework aims to reduce procedural uncertainty by using structured domain representations and uncertain LLM-generated state-transition samples to extract candidate procedural rules, transform them into explicit and inspectable constraints, and use them to repair uncertain procedural steps. Although the motivating domain refers to educational vir-tual laboratories, the underlying problem is more general: managing uncer-tain procedural knowledge for action planning in structured interactive envi-ronments. We illustrate the approach in a virtual laboratory domain involving laboratory instruments, containers, tools, and material-transfer actions.\n\n---\n\n## 35. OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling\n- **Authors**: Adam Bawatneh , Sagar Sapkota , Amrit Singh Bedi , Santu Karmaker , Mubarak Shah\n- **URL**: [https://arxiv.org/abs/2605.26322](https://arxiv.org/abs/2605.26322)\n- **Abstract**:\n> Theory of Mind (ToM), the ability to infer others' knowledge, intentions, and emotions, is commonly evaluated in large language models (LLMs) using end-point question answering, where performance is judged solely by the final answer to a social reasoning query. This paradigm obscures whether the model actually constructs the underlying mental-state representations required for robust reasoning, particularly in scenarios involving divergent, evolving, or mistaken beliefs. In order to address this research gap, we introduce OmniToM, a benchmark that directly evaluates these representations by requiring explicit modeling of belief structures for all relevant actors within a narrative. These structures are composed of belief propositions: minimal statements of what an actor takes to be true about the world or another actor's mental state, allowing knowledge, intentions, emotions, and false beliefs to be analyzed in a common format. Models are evaluated in two stages: Stage 1: Belief Extraction, which extracts from the story the beliefs relevant to its social dynamics, and Stage 2: Belief Labeling, which assigns each belief a seven-dimensional schema label covering recursive order, truth status, knowledge access, explicitness, content type, mental source, and context. Built from 895 stories from the existing ToMBench story corpus and augmented with 22,343 labeled belief propositions, OmniToM uses a human-calibrated LLM-assisted annotation pipeline. Across diverse models in zero-shot evaluation, OmniToM reveals an actor-specific belief-tracking bottleneck: current LLMs struggle with the knowledge-access and representational decisions required to transform narrative facts into actors' beliefs and shared mental states.\n\n---\n\n## 36. Experiments in Agentic AI for Science\n- **Authors**: Judy Fox , Geoffrey Fox\n- **URL**: [https://arxiv.org/abs/2605.26305](https://arxiv.org/abs/2605.26305)\n- **Abstract**:\n> This paper details two novel frameworks for developing autonomous, agentic AI in scientific workflows. Both systems leverage a hybrid Local Body, Remote Brain architecture via Google Colab, utilizing Python-based local orchestrators to invoke large language model (LLM) cloud backends. The first agent, DeepTS/DeepCollector, automates the large-scale curation, extraction, and deduplication of time-series datasets. The second, DeepScribe, is an autonomous presentation analyzer that converts visually dense, mathematically complex physics lectures into structured scientific reports. Through practical systems engineering-such as granular attribute extraction (Cellular RAG), remote data inspection, and distributed concurrency controls-we demonstrate how agentic AI can overcome the context and reasoning limitations of current state-of-the-art systems to rigorously support scientific workflows. Finally, we outline a generalization of DeepTS to support deep knowledge graphs and discuss the application of this conceptual approach to high-energy physics (DeepQCD).\n\n---\n\n## 37. Personalizing Embodied Multimodal Large Language Model Agents over Long-term User Interactions\n- **Authors**: Jeongeun Lee , Chanyoung Park , Dongha Lee\n- **URL**: [https://arxiv.org/abs/2605.26256](https://arxiv.org/abs/2605.26256)\n- **Abstract**:\n> Multimodal large language model (MLLM)-based embodied agents have shown strong potential for solving complex tasks in physical environments. However, personalized assistance requires more than following generic instruction or recognizing object categories. In real-world scenarios, the intended target is often specified only implicitly through prior interactions, requiring agents to leverage personalized context accumulated over time. In this work, we propose POLAR, a multiomodal memory-augmented framework for personalized embodied agents over long-term user interactions. POLAR organizes prior interactions into a multimodal knowledge graph that captures semantic memory for personalized context and visual concepts, and episodic memory for embodied experiences such as agent trajectories. To execute embodied tasks, POLAR retrieves relevant memories to interpret the current request and guide task execution. We evaluate POLAR across multiple MLLM backbones and diverse evaluation scenarios to study the role of memory in long-term personalization. Results show that the proposed memory mechanism consistently improves performance by enabling more effective use of information accumulated over prior interactions. The gains are especially pronounced when the agents are required to reason across multiple interactions, perform multi-hop inference, or tracking updates in user-specific context over time.\n\n---\n\n## 38. Can LLMs Introspect? A Reality Check\n- **Authors**: Shashwat Singh , Tal Linzen , Shauli Ravfogel\n- **URL**: [https://arxiv.org/abs/2605.26242](https://arxiv.org/abs/2605.26242)\n- **Abstract**:\n> Can large language models detect and report their own internal states? A number of studies have argued that the answer to this question is yes. We argue, based on lessons from human metacognition research, that this conclusion may be premature: to be convinced of this conclusion we need to distinguish genuine introspection from pattern matching based on surface-level cues. Furthermore, we argue that behavioral evidence alone is inherently insufficient to establish strong introspective claims. We re-examine two recently introduced evaluation paradigms in light of this consideration. In the first paradigm, models are expected to detect whether their internal states have been tampered with. We find that models cannot reliably distinguish such interventions on their internal states from manipulations of the input, suggesting that their success in the original studies reflects their ability to detect anomalies more generally, as opposed to interventions on their internal states in particular. In the second paradigm we examine, models are tasked with predicting labels derived from their own hidden states. Here, we find that classifiers that only have access to the input achieve equivalent performance to the model's own in-context predictions, indicating that the original results do not conclusively demonstrate that the model has privileged access to its internal representations. We further introduce a relabeled control setting, where models cannot rely on the semantics of the task to solve it, and instead must rely on the internal representation; models perform closer to chance on this better-controlled version of the task. Taken together, these results indicate that current evidence is insufficient to establish that LLMs display metacognitive monitoring.\n\n---\n\n## 39. LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding\n- **Authors**: Shihao Wang , Shilong Liu , Yuanguo Kuang , Xinyu Wei , Yangzhou Liu , Zhiqi Li , Yunze Man , Guo Chen , Andrew Tao , Guilin Liu , Jan Kautz , Lei Zhang , Zhiding Yu\n- **URL**: [https://arxiv.org/abs/2605.27365](https://arxiv.org/abs/2605.27365)\n- **Abstract**:\n> Vision-language models (VLMs) commonly formulate visual grounding and detection as a coordinate-token generation problem, serializing each 2D box into multiple 1D tokens that are learned and decoded largely independently. This token-by-token decoding mismatches the coupled structure of box geometry and creates a practical inference bottleneck due to strictly sequential generation. We introduce LocateAnything, a unified generative grounding and detection framework based on Parallel Box Decoding (PBD). By decoding geometric elements such as bounding boxes and points as atomic units in a single step, LocateAnything preserves intra-box geometric coherence and unlocks substantial parallelism. We show that PBD improves both decoding throughput and localization accuracy. We further develop a scalable data engine and curate LocateAnything-Data, a large-scale dataset with more than 138 million training samples, substantially increasing data diversity for high-precision localization. Extensive evaluations show that LocateAnything advances the speed-accuracy frontier, achieving significantly higher decoding throughput while improving high-IoU localization quality across diverse benchmarks. The results highlight the complementary benefits of Parallel Box Decoding and large-scale training data in enabling efficient and precise unified visual grounding and detection.\n\n---\n\n## 40. GENESIS: Harnessing AI Agents for Autonomous 6G RAN Synthesis, Research, and Testing\n- **Authors**: Tamerlan Aghayev , Maxime Elkael , Michele Polese , Minh Dat Nguyen , Gabriele Gemmi , Andrea Lacava , Ali Saeizadeh , Reshma Prasad , Paolo Testolina , Angelo Feraudo , Soumendra Nanda , Pedram Johari , Salvatore D'Oro , Tommaso Melodia\n- **URL**: [https://arxiv.org/abs/2605.27360](https://arxiv.org/abs/2605.27360)\n- **Abstract**:\n> Cellular research and development (R&D) is throttled by six structural processes that each consume months of manual engineering work per iteration: (i) synthesizing new features from standards or research papers into production code; (ii) conformance and interoperability testing; (iii) hardening against field anomalies and diverse deployment environments; (iv) data-driven optimization of network functionalities; (v) discovering and prototyping novel waveforms, functionalities, and capabilities for future standards; and (vi) securing the stack against vulnerabilities. Although Large Language Models (LLMs) have compressed comparable R&D work in general software engineering from days to minutes, their known pitfalls worsen on Radio Access Network (RAN) use cases: they hallucinate Application Programming Interfaces (APIs) and mis-read specifications, which kills interoperability of RAN components at the first mistake, and they heavily rely on simulations for designing algorithms, which is notorious for breaking when transferred to real hardware. To address these challenges, we present GENESIS, an agentic Artificial Intelligence (AI) framework that converts intents (e.g., a specification clause, a telemetry anomaly, or a research hypothesis) into solutions validated with over-the-air experiments, fed back into a persistent knowledge base. GENESIS is built on three composable primitives (agents, skills, hooks) and a knowledge layer (SYNAPSE) that doubles as the source of ground truth and the recipient of every artifact the framework produces, making capabilities compound across runs.\n\n---\n\n## 41. MobileMoE: Scaling On-Device Mixture of Experts\n- **Authors**: Yanbei Chen , Hanxian Huang , Ernie Chang , Jacob Szwejbka , Digant Desai , Zechun Liu , Vikas Chandra , Raghuraman Krishnamoorthi\n- **URL**: [https://arxiv.org/abs/2605.27358](https://arxiv.org/abs/2605.27358)\n- **Abstract**:\n> Mixture-of-Experts (MoE) has become the de facto architecture for hundred-billion-parameter language models, yet its advantages at sub-billion scales for on-device deployment remain largely unexplored. To close this gap, we present MobileMoE, a family of on-device MoE language models with sub-billion active parameters (0.3-0.9B active and 1.3-5.3B total) that establish a new Pareto frontier for on-device LLMs. We first formulate an on-device MoE scaling law that jointly optimizes MoE architecture under mobile memory and compute constraints, identifying an on-device sweet spot - moderate sparsity with fine-grained and shared experts - that is simultaneously memory and compute-optimal. Building on the derived architectures, we train MobileMoE with a four-stage recipe covering pre-training, mid-training, instruction fine-tuning, and quantization-aware training, all on open-source datasets. Across 14 benchmarks, MobileMoE matches or exceeds leading on-device dense LLMs with 2-4$\\times$ fewer inference FLOPs, and matches or surpasses the state-of-the-art MoE OLMoE-1B-7B with up to 60% fewer parameters. To bridge the last mile to mobile deployment, we provide the first efficient MoE inference on commodity smartphones with comprehensive on-device profiling. At comparable INT4 weight memory, MobileMoE-S delivers $1.8$-$3.8\\times$ faster prefill and $2.2$-$3.4\\times$ faster decode than the dense baseline MobileLLM-Pro.\n\n---\n\n## 42. Guiding LLM Post-training Data Engineering with Model Internals from Sparse Autoencoders\n- **Authors**: Yi Jing , Zao Dai , Jinwu Hu , Zijun Yao , Lei Hou , Juanzi Li , Xiaozhi Wang\n- **URL**: [https://arxiv.org/abs/2605.27354](https://arxiv.org/abs/2605.27354)\n- **Abstract**:\n> Model internals encode rich information about how a large language model (LLM) processes its training data; however, post-training data engineering largely relies on external signals and ignores rich intrinsic signals lying in model internals. We propose SAERL, a data engineering framework for LLM reinforcement learning (RL). It models three intrinsic data properties: diversity, difficulty, and quality, using model internals extracted with Sparse Autoencoder (SAE), an advanced mechanistic interpretability tool. Each property grounds a concrete data engineering operation: SAE-space clustering with moderate batch mixing for batch diversity control, a difficulty proxy for easy-to-hard curriculum ordering, and a quality probe for data filtering. SAERL improves average accuracy by 3.00% over vanilla GRPO and reaches target accuracy with 20% fewer training steps on Qwen2.5-Math-1.5B, with consistent gains across model scales and RL algorithms. Experiments show that SAE transfers effectively across model families and scales, serving as a lightweight and reusable data engineering tool. These results demonstrate that model internals are a powerful and practical source of signals for post-training data engineering.\n\n---\n\n## 43. EdgeFlow: Edge-Map Augmented VLM-Based Flowchart Processing for Industrial Requirements Engineering\n- **Authors**: Zhifei Dou , Shabnam Hassani , Ou Wei\n- **URL**: [https://arxiv.org/abs/2605.27332](https://arxiv.org/abs/2605.27332)\n- **Abstract**:\n> Flowcharts are widely used in industrial requirements, but usually remain embedded as static images. Vision Language Models (VLMs) show promise in the conversion of these flowcharts into machine-readable models for RE activities, yet, when directly applied to flowchart conversion, they often fail on topology-critical visual details. To address this, we propose EdgeFlow that augments a VLM's original input with a deterministically extracted Canny edge map-acting as a structural prior-to improve flowchart-to-Mermaid conversion, without requiring annotated training data or domain-specific model fine-tuning. We evaluate EdgeFlow on IndusReqFlow, a dataset sourced from real-world requirements. Compared with off-the-shelf VLMs, EdgeFlow improves node-level F1 by 17.39 percentage points and edge-level F1 by 16.94 percentage points. At the path level, EdgeFlow improves path F1 by 11.06 percentage points, enabling better support for model-based testing. These results demonstrate that EdgeFlow provides a practical, training-free means to improve topology-preserving flowchart-to-Mermaid conversion for industrial RE. Cross-dataset evaluation results on a public synthetic benchmark show no significant improvement; this highlights the need for diverse benchmarks incorporating industrial data for the comprehensive evaluation of future VLM-based RE tools.\n\n---\n\n## 44. It's Not Always Sycophancy: Measuring LLM Conformity as a Function of Epistemic Uncertainty\n- **Authors**: Kevin H. Guo , Chao Yan , Avinash Baidya , Katherine Brown , Xiang Gao , Juming Xiong , Zhijun Yin , Bradley A. Malin\n- **URL**: [https://arxiv.org/abs/2605.27288](https://arxiv.org/abs/2605.27288)\n- **Abstract**:\n> Large language models (LLMs) are known to abandon their initial stance to conform to user pushback. While prior research largely attributes this behavior to sycophancy learned during reinforcement learning from human feedback, we hypothesize that conformity is also driven by a model's epistemic uncertainty at inference time. In this paper, we introduce MUSE, a two-stage evaluation framework to disentangle the mechanisms driving LLM conformity. Specifically, MUSE maps a model's epistemic uncertainty in responding to a query against its likelihood to yield to user pushback in a subsequent turn. We demonstrate that the mechanisms driving conformity extend beyond sycophancy alone. Specifically, we characterize two distinct factors that jointly drive conformity: sycophantic conformity, where a model aligns with user pushback even with absolute certainty in its initial response, and uncertainty-driven conformity, where a model's likelihood for conformity increases alongside its uncertainty. Furthermore, we conduct ablation studies to demonstrate that both sycophantic conformity and uncertainty-driven conformity grow with 1) the LLM's perceived expertise of the user and 2) the plausibility of the user's suggestions. More broadly, MUSE informs more targeted intervention strategies by distinguishing alignment-induced sycophancy and training-corpora-driven uncertainty.\n\n---\n\n## 45. Lost in Sampling: Assessing Lexical Reachability in LLMs via the Word Coverage Score (WCS)\n- **Authors**: Samer Awad , Javier Conde , Carlos Arriaga , Tairan Fu , Javier Coronado-Blázquez , Pedro Reviriego\n- **URL**: [https://arxiv.org/abs/2605.27268](https://arxiv.org/abs/2605.27268)\n- **Abstract**:\n> Modern Large Language Models (LLMs) are often criticized for producing repetitive and homogeneous text, despite possessing vast latent vocabularies. While previous research has focused on model knowledge and training data, we investigate the role of decoding mechanics in suppressing linguistic diversity. We introduce the Word Coverage Score (WCS), a metric that quantifies the extent to which contextually appropriate human vocabulary is mathematically pruned by standard sampling filters (e.g., Top-$p$, Top-$k$, and Min-$p$). Rather than assessing static knowledge, the WCS measures the lexical survival rate of low-frequency, high-information human words as a function of sampling parameters. By auditing open-weight models on human-authored corpus fragments, we identify which logical lexical choices are rendered unreachable by the decoder, even when they reside within the probability space. Our results provide quantitative evidence that industry-standard sampling defaults act as unintended censorship mechanisms, smoothing the unique textures of human expression into a homogenized discourse. The WCS offers a rigorous framework for optimizing the trade-off between text coherence and lexical richness, providing a diagnostic tool for preserving the diversity of human language in generative models.\n\n---\n\n## 46. Pair-In, Pair-Out: Latent Multi-Token Prediction for Efficient LLMs\n- **Authors**: Wenhui Tan , Minghao Li , Xiaoqian Ma , Siqi Fan , Xiusheng Huang , Liujie Zhang , Ruihua Song , Weihang Chen\n- **URL**: [https://arxiv.org/abs/2605.27255](https://arxiv.org/abs/2605.27255)\n- **Abstract**:\n> Long chain-of-thought reasoning has made autoregressive decoding the dominant inference cost of modern large language models. Existing methods target either the input side (latent compression) or the output side (speculative decoding and multi-token prediction, MTP), but the two lines of work have been pursued independently. Moreover, output-side methods must incur an expensive verifier pass to validate the unreliable draft tokens predicted by MTP. To address these issues, we propose \\textbf{Pair-In, Pair-Out (PIPO)}, which unifies both sides by viewing a latent compressor and an MTP head as mirror-image operations: the compressor folds two input tokens into one latent representation, while the MTP head unfolds one hidden state into one additional output token. To remove the verifier cost without sacrificing reliability, PIPO trains a lightweight confidence head that decides whether draft tokens should be accepted. We observe that On-Policy Distillation (OPD) naturally matches the rejection-sampling criterion of speculative decoding, so the confidence head can be trained alongside OPD with negligible extra cost. Experiments on AIME 2025, GPQA-Diamond, LiveCodeBench v6, and LongBench v2 with Qwen3.5-4B and 9B backbones show that PIPO improves pass@4 over regular decoding by up to $+7.15$ points, while delivering up to $2.64\\times$ first-token-latency and $2.07\\times$ per-token-latency speedups.\n\n---\n\n## 47. Qiskit QuantumKatas: Adapting Microsoft's Quantum Computing exercises for LLM evaluation\n- **Authors**: Juan Cruz-Benito , Ismael Faro\n- **URL**: [https://arxiv.org/abs/2605.27210](https://arxiv.org/abs/2605.27210)\n- **Abstract**:\n> We adapt Microsoft's QuantumKatas -- a well-established quantum computing curriculum -- from Q# to Qiskit, the most widely-adopted quantum computing framework, and package it with an evaluation framework for systematic LLM assessment. The resulting benchmark comprises 350 tasks across 26 categories, spanning fundamental gates through advanced algorithms (Grover's, Simon's, Deutsch-Jozsa), error correction, key distribution, and quantum games. Each task includes a natural language prompt, canonical solution, and deterministic test verification via classical circuit simulation. By building on the QuantumKatas' proven pedagogical design rather than creating tasks from scratch, we inherit a principled difficulty progression and comprehensive concept coverage while contributing the framework adaptation, evaluation infrastructure, and empirical analysis. We evaluate 16 LLMs across 7 prompting configurations -- a total of 39,200 model runs -- to demonstrate the benchmark's utility. Three key findings emerge: (1) the benchmark effectively differentiates model capabilities, with best-configuration pass rates ranging from 32.3% to 83.1% and a 26.1 pp average gap between frontier and open-source models; (2) models perform well at implementing known algorithms (SimonsAlgorithm 82.1%, BasicGates 81.6%) but struggle with problem encoding (SolveSATWithGrover 34.4%, DistinguishUnitaries 40.0%); and (3) chain-of-thought prompting shows a modestly bimodal effect -- it is the best strategy for three models (two of them explicitly reasoning-tuned per vendor documentation) but degrades performance for the rest, leaving it mid-pack in aggregate (56.3% mean) behind few-shot-5 (57.8%). We release the benchmark, evaluation framework, and baseline results to support research on LLM capabilities in quantum computing.\n\n---\n\n## 48. Generative Animations: A Multi-Model Pipeline for Prompt-Driven Motion Synthesis\n- **Authors**: Mannat Khurana , Sanyam Jain , Rishav Agarwal\n- **URL**: [https://arxiv.org/abs/2605.27203](https://arxiv.org/abs/2605.27203)\n- **Abstract**:\n> Animation elevates digital documents into immersive experiences, yet creating custom motion paths remains cumbersome, requiring designers to manually select presets, plot Bézier points, and configure timing properties. We introduce Generative Animations, a system that transforms natural language prompts into production-ready animations. By chaining Large Language Models (LLMs) for semantic parsing with the Segment Anything Model (SAM) for visual grounding, our pipeline automatically generates motion paths that respect scene geometry, handle depth-based occlusions, and honor 3D perspective transforms. We demonstrate the system through three use cases: contour-following trajectories, orbital animations with z-order awareness, and perspective-aligned motion on transformed objects.\n\n---\n\n## 49. Learning When to Think While Listening in Large Audio-Language Models\n- **Authors**: Zhiyuan Song , Weici Zhao , Yang Xiao , Suhao Yu , Cheng Zhu , Jiatao Gu\n- **URL**: [https://arxiv.org/abs/2605.27190](https://arxiv.org/abs/2605.27190)\n- **Abstract**:\n> Recent advances in Large Audio-Language Models (LALMs) have made real-time, streaming spoken interaction increasingly practical. In this setting, reasoning quality and responsiveness are tightly coupled: delaying reasoning until the speech endpoint can improve answer quality but moves deliberation into user-visible response delay, while answering too early risks committing before decisive evidence arrives. We introduce a learnable wait-think-answer control formulation for LALMs. Motivated by the incremental nature of human conversation, the controller decides under partial audio evidence when to wait, when to externalize a compact reasoning update, and when to answer. Using Qwen2.5-Omni-7B as the base model, we construct aligned wait-think-answer traces from spoken reasoning data, train the controller with supervised fine-tuning (SFT), and then apply Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO). The reward combines answer correctness, action validity, update timing, latency synchronization, reasoning quality, and chain consistency, optimizing the complete wait-think-answer trajectory and not the final answer alone. On a six-task synthetic spoken reasoning question answering (SRQA) benchmark, the six-reward DAPO controller improves the row-weighted accuracy from 67.6% to 70.3% while reducing post-endpoint final-think length by 14% under the same Qwen deployment harness. On a 186-item human-recorded Real Audio Bench, a transfer check beyond text-to-speech (TTS)-rendered speech, the controller family remains functional: SFT achieves the strongest accuracy, while the six-reward DAPO controller is the only learned variant whose final-think length falls below the base. These results suggest that a streaming model should learn when to make intermediate reasoning explicit during the audio stream.\n\n---\n\n## 50. LitSeg: Narrative-Aware Document Segmentation for Literary RAG\n- **Authors**: Ruikang Zhang , Zhanni Chen , Yiqiao Cai , Qi Su\n- **URL**: [https://arxiv.org/abs/2605.27156](https://arxiv.org/abs/2605.27156)\n- **Abstract**:\n> Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by incorporating external knowledge, particularly for long-tail domains such as literary works. However, the critical step of document segmentation in RAG remains largely underexplored. Existing strategies are typically semantically blind and overlook the complicated narrative structures of literary works, often resulting in fragmented plots and unclear references that severely hinder retrieval and generation performance. To address this, we propose LitSeg, a novel narrative-theory-guided segmentation framework. By employing multi-stage prompting, LitSeg explicitly extracts valid events, untangles narrative threads, clarifies narrative structures, and locates turning points to inform segmentation. To alleviate the computational overhead of multi-stage inference with large-scale models, we further introduce LitSeg-Lite, a lightweight single-pass chunker fine-tuned on LitSeg-generated data via a two-stage training strategy, distilling the complex process into a single inference pass. Extensive experiments demonstrate that with structurally independent text chunks, our methods significantly improve retrieval accuracy and context relevance over baselines, ultimately enhancing downstream QA performance, while ablation studies validate the efficacy of narratological guidance and data distillation.\n\n---\n\n## 51. DEI: Diversity in Evolutionary Inference for Quality-Diversity Search\n- **Authors**: John Donaghy , Shikhar Rastogi\n- **URL**: [https://arxiv.org/abs/2605.27130](https://arxiv.org/abs/2605.27130)\n- **Abstract**:\n> We present DEI: Diversity in Evolutionary Inference, a distributed Quality-Diversity (QD) search framework that assigns heterogeneous large language models (LLMs) as mutation operators across peer nodes communicating with non-blocking collective operations. Unlike homogeneous parallel search, which replicates a single model's inductive biases across all workers, DEI treats each LLM's distinct creative prior as a complementary source of behavioral novelty. Extending the Digital Red Queen framework with DEI, nodes share local optimal solutions at the end of each round to seed the next round's population. This creates cross-model adversarial pressure that drives robustness beyond intra-model self-play. Evaluated on the Core War domain, a competitive programming benchmark in which Redcode warrior programs battle inside a simulated machine, a four-node heterogeneous ensemble (GPT-5.4-mini, Claude Sonnet 4.6, GPT-5.2, and Claude Haiku 4.5) achieves 124 percent higher merged-archive QD-Score (45.90 vs. 20.46) and 28 percent higher coverage (80.6 percent vs. 63.0 percent of cells) than a single-node baseline at equal total LLM-call budget. The heterogeneous ensemble also outperforms an equally-budgeted homogeneous ensemble on QD-Score, coverage, and held-out solution generality across all four model families. These results provide the first empirical evidence that model diversity, not merely parallelism, is the key driver of gain in distributed LLM-based QD search.\n\n---\n\n## 52. ReMoE: Boosting Expert Reuse through Router Fine-Tuning in Memory-Constrained MoE LLM Inference\n- **Authors**: Xiongwei Zhu , Xiaojian Liao , Tianyang Jiang , Yusen Zhang , Liang Wang , Limin Xiao\n- **URL**: [https://arxiv.org/abs/2605.27081](https://arxiv.org/abs/2605.27081)\n- **Abstract**:\n> Fine-grained Mixture-of-Experts (MoE) models sparsely activate only a subset of experts per token, reducing activated computation while maintaining high model capacity. However, in memory-constrained inference scenarios, only a small set of experts can be cached. Experts not in the cache must be fetched from slow external storage (e.g., UFS), leading to frequent evictions and substantial I/O overhead. We propose ReMoE, a router fine-tuning framework designed to boost token-wise expert reuse. ReMoE biases the router toward recently selected experts, producing temporally stable routing that better matches cache locality constraints. By increasing short-horizon expert reuse, ReMoE reduces expert fetches from storage without adding inference-time computation. Experiments on DeepSeek and Qwen models show that ReMoE improves expert reuse by 26% while maintaining downstream task performance. Real-system evaluations further confirm these benefits, improving output throughput by 8.4% under vLLM GPU-CPU expert offloading and reducing TPOT by 43.6-49.8% under this http URL on Jetson Orin NX, corresponding to a 1.77-1.99$\\times$ decode speedup across diverse workloads. Checkpoints and usage instructions are available at this https URL .\n\n---\n\n## 53. E3: Issue-Level Backtesting for Automated Research Critique\n- **Authors**: Yashwardhan Chaudhuri , Sanyam Jain , Paridhi Mundra\n- **URL**: [https://arxiv.org/abs/2605.27072](https://arxiv.org/abs/2605.27072)\n- **Abstract**:\n> We present E3, an automated review assistant that augments reviewers and engineering teams by identifying decision-relevant technical concerns in research papers. For each concern, E3 reports its nature, its location, its bearing on the contribution, and the analysis or evidence that would resolve it, covering unsupported claims, missing ablations, weak baselines, hidden assumptions, threats to validity, and leakage risks. To evaluate E3 without contamination confounds we adopt an issue-level backtesting protocol: the corpus is restricted to papers postdating the training cutoff of every automated source, and for each paper a meta-judge that observes only anonymised reviews labels every issue-source pair as Caught, Partial, or Missed. Applied to 100 ICLR 2026 papers and 4598 judged issue rows, comparing E3 against the ICLR human reviews and two prompt-matched LLM baselines built on gpt-5.4 from OpenAI and claude-opus-4-6 from Anthropic, with meta-judge gpt-5.5, E3 attains the highest recall on every aggregate metric. Partial-inclusive recall reaches 90.2 percent, which is 15.5 points over GPT, 17.1 points over Claude, and 29.2 points over the human reviews, and strict recall preserves the ordering at 65.8 percent. On concerns raised by the human reviewers, E3 recovers 89.6 percent; on concerns the human reviewers missed it surfaces 1635 additional rows admitted into the judged union, 406 above the next-best source. Corpus, baseline prompts, judge prompt template, and evaluation code are released.\n\n---\n\n## 54. QUACK: Questioning, Understanding, and Auditing Communicated Knowledge in Multimodal Social Deduction Agents\n- **Authors**: Ye Yuan , Rui Song , Weien Li , Zeyu Li , Haochen Liu , Xiangyu Kong , Changjiang Han , Yonghan Yang , Zichen Zhao , Zixuan Dong , Fuyuan Lyu , Bowei He , Haolun Wu , Jikun Kang , Xue Liu\n- **URL**: [https://arxiv.org/abs/2605.27068](https://arxiv.org/abs/2605.27068)\n- **Abstract**:\n> Social deduction games have become a popular testbed for probing reasoning, deception, coordination, and belief modeling in Large Language Model (LLM) agents. However, most environments are scored only by game outcomes such as win rates and largely remain to text-only interaction, making it difficult to tell whether an agent's language is actually grounded in what it perceived and did, or to identify the failure modes underlying its behavior. To address this gap, we introduce QUACK, an open-source environment and evaluation framework for auditing the grounding of agent language in multimodal social reasoning. QUACK evaluates agents at three levels: game outcomes, behavioral trajectories, and utterance-level consistency. Its core Statement Verification Pipeline reconstructs each agent's ground-truth trajectory from engine logs and checks every discussion claim against it, automatically flagging spatial hallucination, unsupported accusation, deception collapse, and language-action inconsistency. Evaluating three frontier VLMs in both homogeneous and cross-model adversarial settings, we find that even the strongest agent hallucinates 15.1% of its verifiable spatial claims and makes over half of its accusations without grounded evidence. We release the full engine, evaluation framework, toolkit, and logs at this https URL .\n\n---\n\n## 55. ConVer: Using Contracts and Loop Invariant Synthesis for Scalable Formal Software Verification\n- **Authors**: Muhammad A. A. Pirzada , Weiqi Wang , Yiannis Charalambous , Konstantin Korovin , Lucas C. Cordeiro\n- **URL**: [https://arxiv.org/abs/2605.27051](https://arxiv.org/abs/2605.27051)\n- **Abstract**:\n> Formal verification of large C programs is impeded by state-space explosion: Bounded Model Checking (BMC) tools must encode the entire state space up to the predetermined bound by unrolling all nested constructs. We present ConVer, a top-down compositional verification tool. Given a C program with a top-level assertion, ConVer decomposes verification top-down: it uses a large language model (LLM) to synthesise function contracts from the system property, then alternates system-level and function-level checks in a CEGAR-CEGIS loop, refining contracts whenever a check fails via SMART ICE learning. We evaluate ConVer on four benchmark suites of increasing difficulty and against other state-of-the-art (SOTA) tools. On the Frama-C benchmark of 45 simple C programs, ConVer achieves 82-96% verification success across three LLM backends, with 93-95% of converged programs requiring only a single CEGAR-CEGIS iteration. On the X.509 parser benchmark (6~programs) and LF2C-Simple suite (17 programs), ConVer achieves 33-50% and 82-88% success respectively. On the VerifyThis suite of 11 recursive and loop-intensive programs, the Pre-Abstraction strategy achieves 55-64% success. In addition, we present ESBMC-LF a preprocessor tool that converts LF models to C while preserving the properties of the LF files, enabling ConVer to verify them. We transpile the LF Verifier Benchmarks using ESBMC-LF to C; we denote those LF-Hard. We show that ConVer successfully verifies 67% of LF-Hard benchmarks overall.\n\n---\n\n## 56. Tracing Computation Density in LLMs\n- **Authors**: Corentin Kervadec , Iuliia Lysova , Iuri Macocco , Marco Baroni , Gemma Boleda\n- **URL**: [https://arxiv.org/abs/2605.27033](https://arxiv.org/abs/2605.27033)\n- **Abstract**:\n> Transformer-based large language models (LLMs) are comprised of billions of parameters arranged in deep and wide computational graphs, but it is not clear that they exploit their full capacity for all inputs. We introduce the s-Trace method to efficiently estimate the subgraph of size s that best approximates a full model output. With this method, we find the computation in a variety of LLMs to be organized in two distinct phases. A small subgraph mostly composed of early-layer nodes can reconstruct the head of the full model output distribution. Adding further nodes, mostly located in later layers and increasingly consisting of attention heads, leads to incremental refinements in approximating the full output distribution. We find moreover that the amount of necessary computation per input correlates with model uncertainty, and that sparser subgraphs encode shallow statistics, such as unigram frequency. Overall, our results suggest a consistent modular organization in effective LLM computation, with a sparse early-layer core providing a rough prediction that is further refined through denser computations in later layers.\n\n---\n\n## 57. Evaluating the Relevance of Uncertainty Estimators for LLM Hallucination\n- **Authors**: Yedidia Agnimo , Anna Korba , Annabelle Blangero , Nicolas Chesneau , Karteek Alahari\n- **URL**: [https://arxiv.org/abs/2605.27016](https://arxiv.org/abs/2605.27016)\n- **Abstract**:\n> Large language models (LLMs) are prone to hallucinations, i.e., statements unsupported by the input or training data, hindering reliable deployment. In parallel, numerous uncertainty estimation (UE) methods have been proposed to quantify model confidence and are often implicitly treated as proxies for model failure. However, the relationship between uncertainty and hallucinations remains insufficiently characterized. We present a systematic empirical study of the association between uncertainty estimators and hallucinations in LLMs. Rather than assuming this association, we evaluate directly when and to what extent it holds. We consider a diverse set of uncertainty estimators, including information-theoretic, sampling-based, and reflexive estimators, and examine their behavior across hallucination settings. Our experiments cover both intrinsic hallucinations (violations of input faithfulness) and extrinsic hallucinations (unsupported claims relative to training data), using four complementary benchmarks, including RAGTruth and HalluLens. We find that the association is highly variable and often weak, depending on the hallucination type and the LLM under evaluation. These results challenge the use of uncertainty as a direct signal of hallucination and clarify when it provides actionable information.\n\n---\n\n## 58. ReasonOps: A Unified Operational Paradigm for Trustworthy Verified LLM Reasoning\n- **Authors**: Adnan Rashid\n- **URL**: [https://arxiv.org/abs/2605.27014](https://arxiv.org/abs/2605.27014)\n- **Abstract**:\n> Large Language Models (LLMs) have transformed artificial intelligence from primarily generative systems into increasingly capable reasoning agents. Recent advances in theorem proving, autoformalization, symbolic reasoning, and tool-augmented language models demonstrate substantial progress toward machine-assisted formal reasoning. However, current reasoning systems still suffer from hidden logical inconsistencies, hallucinated symbolic transitions, unsupported theorem applications, and limited reliability guarantees. Existing approaches remain fragmented across formal verification, runtime assurance, neuro-symbolic reasoning and trustworthy Artificial Intelligence (AI) research communities. This paper introduces ReasonOps, a unified operational paradigm for trustworthy verified reasoning systems. Inspired by operational ecosystems such as DevOps and MLOps, ReasonOps treats reasoning as a continuously monitored, verifiable, reliability-aware operational process rather than an isolated inference task. The proposed paradigm integrates semantic interpretation, autoformalization, symbolic reasoning, theorem proving, runtime assurance, probabilistic reliability estimation, and adaptive correction into a unified reasoning lifecycle. The paper further presents the ReasonOps architecture, demonstrates its workflow using an autonomous braking system analysis example, and discusses its potential role in future safety-critical autonomous AI systems. We argue that operational reasoning paradigms such as ReasonOps may become foundational infrastructure for next-generation trustworthy AI ecosystems.\n\n---\n\n## 59. Recon: Reconstruction-Guided Reasoning Synthesis for User Modeling\n- **Authors**: Alan Zhu , Mihran Miroyan , Carolyn Wang , Andrew Zhou , Lisa Dunlap , Narges Norouzi , Joseph E. Gonzalez\n- **URL**: [https://arxiv.org/abs/2605.26969](https://arxiv.org/abs/2605.26969)\n- **Abstract**:\n> User modeling aims to use language models (LMs) to mimic an individual's behavior from a corpus of past context-action pairs (e.g., conversation turns), enabling the simulation of users in settings like behavioral science, human-AI collaboration, and market research. Recent approaches augment these corpora with synthesized reasoning traces, typically generated by conditioning on both context and action. However, such conditioning constitutes post-hoc rationalization rather than reasoning: the trace is guaranteed to justify the action, but may not encode the underlying latent causal decision paths. We propose Recon, which uses action reconstruction to score reasoning traces by their predictive power: given a context and candidate reasoning, a reconstruction model predicts the action, and reconstruction fidelity determines reasoning quality. Across four domains, Recon achieves a 54.7% win rate over Backward Synthesis, a standard post-hoc rationalization baseline. Further, we find that training a reasoning synthesis model with rewards derived from Recon improves downstream user modeling performance, achieving a win rate of up to 70.0% over baselines. We further show that Recon-synthesized reasoning transfers across models, and improves user modeling beyond the reconstruction model. Our work demonstrates that post-hoc rationalization is insufficient for reasoning synthesis, and that useful and interpretable reasoning should naturally elicit the action from the context.\n\n---\n\n## 60. Tournament-GRPO: Group-Wise Tournament Rewards for Reinforcement Learning in Open-Ended Long-Form Generation\n- **Authors**: Zixuan Yang , Yiqun Chen , Wei Yang , Erhan Zhang , Zihan Shen , Xiaochi Wei , Yan Gao , Yi Wu , Yao Hu , Jiaxin Mao\n- **URL**: [https://arxiv.org/abs/2605.26958](https://arxiv.org/abs/2605.26958)\n- **Abstract**:\n> Reinforcement learning in open-ended long-form generation is challenging because reliable reference answers and automatic metrics are often unavailable. Existing rubric-based methods typically rely on pointwise LLM-as-a-judge scoring, but absolute scores are difficult to calibrate across complex responses, may provide weak discrimination among same-query rollouts, and can become saturated during optimization. We propose Tournament-GRPO, a group-wise reward framework that converts rubric-guided LLM judgments into relative rewards through repeated multi-round tournaments among same-query rollouts. Tournament-GRPO compares candidates within groups, accumulates tournament outcomes, and normalizes them into group-wise rewards for GRPO training. Experiments on Deep Research Bench show that Tournament-GRPO consistently outperforms existing reward-design baselines, achieving a 4.52-point overall-score improvement over the strongest baseline. Further analyses show that tournament rewards provide a favorable effectiveness--efficiency trade-off and that tournament design affects training dynamics. These results suggest that rubric-guided tournament comparison provides an effective reward signal for reinforcement learning in open-ended long-form generation.\n\n---\n\n## 61. JuICE: A Benchmark for Evaluating LLM-Judge in Identifying Cultural Errors\n- **Authors**: Jiho Jin , Junho Myung , Juhyun Oh , Junyeong Park , Rifki Afina Putri , Sunipa Dev , Vinodkumar Prabhakaran , Alice Oh\n- **URL**: [https://arxiv.org/abs/2605.26955](https://arxiv.org/abs/2605.26955)\n- **Abstract**:\n> As large language models (LLMs) are increasingly deployed to users around the world, they are integrated into everyday tasks across diverse cultural contexts, from drafting personal communications to brainstorming creative ideas. These tasks are inherently cultural: they require contextual appropriateness, symbolic resonance, and tacit cultural expectations that native speakers draw on instinctively, meaning that a response can be factually plausible yet unmistakably wrong to a local reader. Existing cultural benchmarks have treated culture as a flat set of facts via fact verification or norm entailment methods, and have adopted LLM-as-a-Judge without examining whether they can capture such thick cultural errors. To address this gap, we present JuICE (Benchmark for LLM-Judge in Identifying Cultural Errors), a multilingual dataset of 7,470 span-level annotations of cultural and linguistic errors in long-form LLM responses. It covers 1,050 query-response pairs from four countries (the United States, South Korea, Indonesia, and Bangladesh), in both English and their countries' main languages. Using JuICE, we find that even the strongest LLM-judge achieves only an F1 of 0.52 in the erroneous span detection task. Furthermore, LLM-judges consistently miss thick cultural errors that local residents readily identify. Our findings suggest that robust cultural evaluation must move beyond surface-level detection toward frameworks that account for the depth and situatedness of cultural meaning.\n\n---\n\n## 62. Beyond Questions: Evaluating What Large Language Models (Actually) Know\n- **Authors**: Luca Giordano , Simon Razniewski\n- **URL**: [https://arxiv.org/abs/2605.26937](https://arxiv.org/abs/2605.26937)\n- **Abstract**:\n> Parametric knowledge in large language models (LLMs) is a cornerstone of their success, yet remains poorly understood. Existing knowledge benchmarks typically rely on predefined questions (e.g., \"What is the birth date of M.L. King?\"), evaluating only knowledge that benchmark designers explicitly choose to query, a problematic availability bias. In this paper, we introduce open knowledge evaluation, a new paradigm for LLM knowledge benchmarking. Instead of asking narrow questions, it evaluates models on the knowledge they choose to surface in response to open-ended elicitation prompts (e.g., \"Tell me everything you know about M.L. King\"). This shifts the focus from predefined answer retrieval toward characterizing the knowledge models naturally express. We instantiate this paradigm with BeQu (Beyond Questions), a benchmark of 10,000 entities paired with reference corpora for statement verification. Using BeQu, we evaluate a broad range of language models and analyze the effects of reasoning effort, model scale, prompt format, and knowledge domain. Data and leaderboard are available on this work's GitHub repository and at the benchmark's website.\n\n---\n\n## 63. Strategies for Guiding LLMs to Use Software Design Patterns: A Case of Singleton\n- **Authors**: Viktor Kjellberg , Farnaz Fotrousi , Miroslaw Staron\n- **URL**: [https://arxiv.org/abs/2605.26898](https://arxiv.org/abs/2605.26898)\n- **Abstract**:\n> Large Language Models (LLMs) can generate functional source code from natural-language prompts, but often fail to consistently follow higher-level architectural structures or design patterns. Since LLMs are increasingly used in software engineering, their ability to apply established design principles to generated code is crucial to the long-term success of software products. Therefore, the goal of this paper is to identify strategies for guiding LLMs to incorporate design patterns into the generated source code. We designed a computational experiment to evaluate the ability of 13 LLMs to generate code that follows the Singleton design pattern, using four prompting strategies: instructions, binary automated feedback, extensive automated feedback, and extensive feedback with few-shot prompts, in 164 Java coding challenges from HumanEval-X. Our results shows that the optimal strategy to guide LLMs to include design patterns depends heavily on the type of model. Still, overall, iterative binary feedback provides the best alignment with Singleton while preserving or improving the code's functionality. With guiding with instructions, Llama 3.3 generated Singleton classes in 100% of cases and improved code functionality, increasing the number of tests passed by 34.1 percentage points. It achieved a similar result with guidance through instructions and binary feedback. Qwen 3 (8B) increased the alignment with Singleton to 99.2% and the functionality to 58.6% using binary feedback. Our result suggests that even simple strategies can be used to guide LLMs to use design patterns.\n\n---\n\n## 64. Negligible in Size, Significant in Effect: On Scale Vectors in Large Language Models\n- **Authors**: Mingze Wang , Shuchen Zhu , Yuxin Fang , Binghui Li , Kai Shen , Shu Zhong\n- **URL**: [https://arxiv.org/abs/2605.26895](https://arxiv.org/abs/2605.26895)\n- **Abstract**:\n> Normalization layers in modern large language models (LLMs) consist of a deterministic normalization operation and a learnable scale vector. While the normalization operation has been extensively studied, the scale vector remains poorly understood despite its ubiquitous use. In this work, we present a systematic study of scale vectors in LLMs from the perspectives of expressivity, optimization, and architectural structure. First, we show empirically that although scale vectors constitute only a negligible fraction of model parameters, removing them substantially degrades LLM pre-training. Our theory further shows that, in Pre-Norm architectures, scale vectors do not increase expressivity; instead, they improve optimization through a self-amplifying preconditioning effect on subsequent linear mappings. Second, we investigate the role of weight decay for scale vectors. By distinguishing Input-Norm and Output-Norm layers, we theoretically show that weight decay is beneficial for the former but harmful for the latter, due to their distinct roles in optimization and expressivity. Third, motivated by this understanding, we propose three lightweight and complementary improvements to scale vectors: branch-specific heterogeneity, improved placement around linear mappings, and magnitude-direction reparameterization. Both theory and experiments show that each improvement yields consistent gains. Finally, we combine these improvements into a unified scale-vector strategy and evaluate it through extensive LLM pre-training experiments on dense and mixture-of-experts models ranging from 0.12B to 2B parameters, across multiple optimizers and learning rate schedules, under industrial-scale token budgets. The unified strategy consistently achieves lower terminal loss than well-tuned baselines and exhibits more favorable scaling behavior, while adding negligible parameter and computational overhead.\n\n---\n\n## 65. GeoFaith: A Spatio-Temporal Dual View of Faithful Chain-of-Thought\n- **Authors**: Weijiang Lv , Wentong Zhao , Jiayu Wang , Yuhao Wu , Jiaheng Wei , Xiaobo Xia\n- **URL**: [https://arxiv.org/abs/2605.26893](https://arxiv.org/abs/2605.26893)\n- **Abstract**:\n> Chain-of-Thought (CoT) reasoning has advanced large language models (LLMs), but outcome-based supervision leads to pervasive post-hoc rationalization, producing plausible yet unfaithful reasoning chains. Most prior faithfulness assessment methods are either unscalable, expensive, or unreliable. We propose GeoFaith, a spatio-temporal framework that leverages latent geometric structure and entropy dynamics to diagnose and enforce faithful reasoning. We develop a scalable bootstrapping pipeline expanding step-level annotations from 1k to 20k samples across four domains, train an 8B faithfulness detector outperforming GPT-5 on standard benchmarks, and design a faithfulness-aware reinforcement learning framework jointly optimizing outcome correctness, process faithfulness, and trajectory consistency. Experiments show the proposed method achieves superior performance on both faithfulness detection and downstream reasoning, producing shorter, more interpretable chains without sacrificing accuracy. Our code will be made available publicly.\n\n---\n\n## 66. Knowledge Graphs as the Missing Data Layer for LLM-Based Industrial Asset Operations\n- **Authors**: Madhulatha Mandarapu , Sandeep Kunkunuru\n- **URL**: [https://arxiv.org/abs/2605.26874](https://arxiv.org/abs/2605.26874)\n- **Abstract**:\n> LLM-based agents for industrial asset operations show limited accuracy when reasoning over flat document stores. AssetOpsBench (KDD 2026) establishes that GPT-4 agents achieve 65% on 139 industrial maintenance scenarios backed by CouchDB, YAML, and CSV. It compares LLM orchestration paradigms (Agent-As-Tool vs Plan-Execute) on a fixed data layer; we ask a complementary, orthogonal question: how much does the data model behind the tools affect agent performance? Building on the same scenarios, we introduce a knowledge graph layer (781 nodes, 955 edges, 16 relationship types) and evaluate three architectures: (1) deterministic graph handlers (no LLM) at 99% (137/139); (2) LLM-generated Cypher over the graph at 82-83% with the same GPT-4 model the baseline uses; and (3) the original tool-augmented LLM baseline at 65% (91/139, matching the published KDD 2026 leaderboard ceiling). Our key finding is inverted LLM usage: rather than asking the LLM to reason over raw data, we ask it to generate structured queries from a typed schema. The graph executes deterministically. We additionally contribute 40 graph-native scenarios (multi-hop dependency, vector similarity, PageRank criticality), and evaluate against the expanded HuggingFace AssetOpsBench release (467 scenarios, 6 domains), where deterministic handlers achieve 100% (467/467) with average score 0.848. These results suggest that for structured operational domains, the data layer -- not the LLM orchestration -- is the primary bottleneck, and that knowledge graphs serve as an integration layer between raw industrial data and LLM-based reasoning.\n\n---\n\n## 67. The Strongest Teacher Is Not Always the Best Teacher: Student-Centric Answer Selection\n- **Authors**: Zhengyu Hu , Zheyuan Xiao , Linxin Song , Fengqing Jiang , Yutai Li , Zhengyu Chen , Zhihan Xiong , Yue Liu , Junhao Lin , Yao Su , Lijie Hu , Kaize Ding , Xiao Teng , Radha Poovendran\n- **URL**: [https://arxiv.org/abs/2605.26872](https://arxiv.org/abs/2605.26872)\n- **Abstract**:\n> LLM training increasingly relies on teacher-generated supervision, from synthetic responses to reasoning traces and tool-use demonstrations. Current practice often chooses the highest-performing teacher to generate student training data, implicitly treating teacher test performance as a proxy for teaching quality. We show that this assumption can fail: even when multiple teachers provide correct answers to the same question, the answer from the strongest teacher is not necessarily the best supervision for a given student. To address this gap, we propose Student-Centric Answer Sampling (SCAS), a framework that selects from verified teacher-generated answers according to their estimated student-centric learning cost. Motivated by a token-wise gradient decomposition, we derive an efficient forward-only proxy for this cost and use it to guide answer selection during training. Experiments across 30 teacher models, 6 student base models, and 8 tasks show that SCAS consistently improves student performance, suggesting that effective distillation should prioritize supervision matched to the current student rather than teacher strength alone.\n\n---\n\n## 68. Persistent AI Agents in Academic Research: A Single-Investigator Implementation Case Study\n- **Authors**: Anas H. Alzahrani\n- **URL**: [https://arxiv.org/abs/2605.26870](https://arxiv.org/abs/2605.26870)\n- **Abstract**:\n> Background: Large language models are typically evaluated as models, benchmarks, or short conversational episodes. Less is known about what happens when an agent is embedded persistently in a real academic research environment with durable memory, local files, external tools, scheduled routines, delegated roles, and explicit safety protocols. Methods: A structured self-observed implementation case study was conducted from January 31 to May 25, 2026. The unit of analysis was the persistent human-agent environment: researcher, agent runtime, memory layer, tools, repositories, scheduled jobs, specialized agent roles, and governance rules. Outcomes were organized using PARE-M (Persistent Agentic Research Environment Measurement), a measurement framework covering architecture, utilization, artifact production, resource use, reproducibility, and governance. Results: Recoverable main-agent telemetry contained 75,671 de-duplicated records across 96 active days, with 8,059 user-role and 23,710 assistant-role messages. The workspace included 502 memory-related files, 17 configured agent directories, and 57 skill files. Active system time was 579.7 hours (30-minute capped-gap estimate). Memory-derived records identified 482 output-proxy events and 889 failure, verification, correction, or protocol-proxy events. A strict May 2026 trajectory subset captured 627 model-completed events and 73.95 million recorded tokens, of which 82.9% were cache reads. Conclusions: The workflow was cache-dominant, suggesting that persistent agentic environments may shift the economic unit from cost per token to cost per completed artifact. Future evaluations should use artifact-level denominators, reproducible parsing rules, correction taxonomies, and independent coding of governance events.\n\n---\n\n## 69. The Kalman Evolve: Closing the Gap in Kalman Filtering via Interpretable Algorithm Discovery\n- **Authors**: Vasileios Saketos , Ming Xiao\n- **URL**: [https://arxiv.org/abs/2605.26830](https://arxiv.org/abs/2605.26830)\n- **Abstract**:\n> State estimation is a fundamental problem in control and signal processing, for which the Kalman Filter provides an optimal solution under linear dynamics, Gaussian noise, and known noise covariances. However, these assumptions often fail in realistic sensing settings such as Doppler radar and LiDAR. In these cases, the optimal estimator is inherently nonlinear, which leads to systematic performance degradation. This creates a performance gap that cannot be eliminated by tuning the noise covariance parameters (i.e., the process and measurement noise in the Kalman Filter) alone. To address this limitation, we propose Kalman Evolve, a framework for discovering improved filtering algorithms by jointly optimizing both noise parameters and the update structure. Our approach leverages large language models (LLMs) as a structured prior over program space, enabling the generation of interpretable, non-affine modifications to the classical Kalman filter while preserving its recursive form. We provide analytical results establishing the suboptimality of affine estimators under common nonlinear sensing models, motivating the need for structure-aware updates. Across a range of synthetic and real-world tracking benchmarks, including Doppler radar, LiDAR-based localization, and pedestrian tracking, the discovered algorithms consistently improve over strong baselines such as the Optimized Kalman Filter, achieving up to 12\\% reduction in RMSE. These results suggest that optimizing the structure of the Kalman filter, rather than only its parameters, provides a practical and interpretable way to improve state estimation.\n\n---\n\n## 70. ContextGuard: Structured Self-Auditing for Context Learning in Language Models\n- **Authors**: Hongbo Jin , Chi Wang , Haoran Tang , Zhongjing Du , Xu Jiang , Jingqi Tian , Qiaoman Zhang , Jiayu Ding\n- **URL**: [https://arxiv.org/abs/2605.26827](https://arxiv.org/abs/2605.26827)\n- **Abstract**:\n> Recent benchmarks reveal that despite strong reasoning capabilities, large language models (LLMs) still struggle to faithfully apply complex contextual knowledge. These failures are often not wholesale reasoning collapses: in context-rich tasks, models may follow the central reasoning path while missing peripheral, persistent, or format-sensitive requirements.\n\n---\n\n## 71. RAGEAR: Retrieval-Augmented Graph-Enhanced Academic Recommender\n- **Authors**: Francesco Granata , Lorenzo Lamazzi , Misael Mongiovì , Francesco Poggi , Valeria Secchini\n- **URL**: [https://arxiv.org/abs/2605.26819](https://arxiv.org/abs/2605.26819)\n- **Abstract**:\n> We present RAGEAR (Retrieval-Augmented Graph-Enhanced Academic Recommender), a neurosymbolic recommender system for academic course recommendation. RAGEAR combines dense retrieval over full lecture transcripts with a symbolic Knowledge Graph modelling courses, lessons, transcript chunks, credits, study plans, and curricular information. The Knowledge Graph supports symbolic filtering and contextualisation based on structured constraints, such as credits, academic disciplines, study plans, and prerequisites. Unlike metadata-based approaches, it exploits fine-grained instructional content by retrieving transcript chunks semantically aligned with a student's query. The main contribution is a graph-aware aggregation function that propagates chunk-level evidence to course-level recommendations. The score combines three factors: the share of retrieved similarity associated with a course, the rank-based strength of its relevant chunks, and the distribution of evidence across lessons. We evaluate RAGEAR on 152 student-like queries through a human evaluation sample and a large-scale LLM-based relevance assessment. Results show that lecture transcripts improve over metadata-only retrieval, and that RAGEAR further improves ranking quality over a transcript-based normalized SumP baseline, especially for top-ranked recommendations.\n\n---\n\n## 72. Innovation: An Almost Characterization of Hallucination\n- **Authors**: Nishant P. Das , Piyush Srivastava\n- **URL**: [https://arxiv.org/abs/2605.26808](https://arxiv.org/abs/2605.26808)\n- **Abstract**:\n> Hallucination is a central limitation of large language models (LLMs), and substantial effort has been devoted to understanding and mitigating it. Towards this, Kalai and Vempala (STOC 2024) introduced a probabilistic framework formalizing calibration and hallucination, and showed that, with high probability, calibrated LLMs hallucinate roughly at the rate of the \"missing mass\", a measure of how incomplete the training data is relative to its source. This raises two fundamental questions: (i) what property of a calibrated LLM makes hallucinations unavoidable? and (ii) can hallucinations be avoided by giving up calibration? We answer these questions by introducing a simpler property we call innovation that measures the tendency of a model to produce outputs outside the training data. We show that innovation is implied by the condition for hallucination identified by Kalai and Vempala, and, further, that it is an almost characterization of hallucination: hallucination implies innovation, and conversely, innovation implies hallucination with high probability. We also provide lower bounds on the hallucination rate based on the \"innovation rate\", and by relating innovation rate back to missing mass, we obtain new hallucination rate lower bounds based on missing mass that extend the results of Kalai and Vempala.\n\n---\n\n## 73. SeDT: Sentence-Transformer Decision-Transformer Conditioning for Multi-Turn Conversation Reliability\n- **Authors**: Ramakrishna Vamsi Setti , Jagadeesh Rachapudi , Sachin Chaudhary , Praful Hambarde , Amit Shukla\n- **URL**: [https://arxiv.org/abs/2605.26788](https://arxiv.org/abs/2605.26788)\n- **Abstract**:\n> Large language models (LLMs) achieve impressive performance when a task is fully specified in a single turn, yet the same models lose up to 39% of that performance when the identical task is revealed incrementally across multiple turns, a phenomenon documented at scale as Lost in Conversation. Crucially, this collapse is almost entirely a reliability failure; the best case, the aptitude only falls 16%, while the unreliability more than doubles (+112%). We argue that the root cause is structural, a flat conversation history assigns equal implicit weight to every prior turn, giving the model no signal to distinguish a critical constraint from incidental dialog. We present SeDT Sentence-transformer Decision-Transformer, a training-free inference-time method that resolves this by importing return-to-go conditioning from offline reinforcement learning. SeDT annotates each conversation shard with a cumulative relevance score derived from three complementary semantic, lexical, and positional signals and presents the full annotated history to the model at the final turn, without weight changes, without training data, and without discarding context. Evaluated on the Lost-in-Conversation benchmark in three LLMs and three generation tasks, SeDT outperforms the sharded baseline in all nine model-task combinations, with gains up to +37.7% in mean performance P and simultaneous reductions in unreliability in seven of the nine combinations. In short, telling the model which past turns matter is sufficient to substantially recover the performance lost in conversation.\n\n---\n\n## 74. EmoDistill: Offline Emotion Skill Distillation for Language Model Agents in Adversarial Negotiation\n- **Authors**: Yunbo Long , Haolang Zhao , Lukas Beckenbauer , Liming Xu , Alexandra Brintrup\n- **URL**: [https://arxiv.org/abs/2605.26785](https://arxiv.org/abs/2605.26785)\n- **Abstract**:\n> Post-trained LLMs are often optimized to align responses with human preferences, making them safe, polite, and conversationally appropriate. In adversarial negotiation, however, this alignment can become a vulnerability: emotionally framed language may steer agents toward the counterparty's interests. Using GoEmotions-based affective prompting, we show that emotion substantially shifts negotiation outcomes, suggesting that emotion is a strategic action channel rather than a surface style. Thus, we introduce \\textbf{EmoDistill}, an offline framework for distilling emotional negotiation skills into language model agents. EmoDistill decomposes emotional strategy into emotion selection and emotion expression: an Implicit Q-Learning (IQL) selector learns \\emph{which} emotion to express, while a Low-Rank Adaptation (LoRA)-based policy learns \\emph{how} to express it through Supervised Fine-Tuning (SFT) and Judge Policy Optimization (JPO). Across four emotion-sensitive, high-stakes negotiation domains, SLM policies trained under the EmoDistill framework achieve the highest utility, outperforming vanilla SLM/LLM baselines and IQL-only emotion selection. Ablations show that emotion conditioning is essential, and transfer studies demonstrate generalization across domains, unseen counterparties, and trained-vs-trained tournaments. Overall, EmoDistill learns skills from offline agent-to-agent interactions, avoiding costly online negotiation during training.\n\n---\n\n## 75. Ratio-Variance Regularized Policy Optimization\n- **Authors**: Yu Luo , Shuo Han , Yihan Hu , Lei Lv , Huaping Liu , Fuchun Sun , Jianye Hao , Dong Li\n- **URL**: [https://arxiv.org/abs/2605.26784](https://arxiv.org/abs/2605.26784)\n- **Abstract**:\n> Standard on-policy reinforcement learning relies on heuristic clipping to enforce trust regions, but this mechanism imposes a severe cost by indiscriminately truncating high-return yet high-divergence updates. We demonstrate that explicitly constraining the policy ratio variance provides a principled local approximation to trust-region constraints, eliminating the need for binary hard clipping. By acting as a distributional ``soft brake'', this approach preserves critical gradient signals from novel discoveries while naturally down-weighting and enabling the reuse of stale, off-policy data. We introduce ${\\bf R}^2{\\bf VPO}$ (Ratio-Variance Regularized Policy Optimization), which implements this constraint via a primal-dual optimization framework. Extensive evaluations across $7$ LLM scales, spanning both fast and slow reasoning paradigms, and $10$ robotic control tasks demonstrate the generality of the proposed approach. R$^2$VPO achieves substantial performance gains on mathematical reasoning benchmarks, with particularly pronounced improvements on smaller models, while significantly improving sample efficiency. Furthermore, it consistently outperforms PPO baselines in continuous control domains, particularly in sparse-reward and dynamic environments. Together, these findings establish ratio-variance regularization as a principled foundation for stable and data-efficient policy optimization.\n\n---\n\n## 76. MatFormBench: A Benchmarking Evaluation Framework for Target-Driven Materials Formulation\n- **Authors**: Linhan Wu , Chenxi Wang , Chuhan Yang , Zhengwei Yang , Yuyang Liu\n- **URL**: [https://arxiv.org/abs/2605.26741](https://arxiv.org/abs/2605.26741)\n- **Abstract**:\n> Inverse design of materials has significantly advanced target-driven formulation optimization, yet existing materials machine learning benchmarks remain limited to forward property prediction, failing to systematically evaluate inverse optimization and generation algorithms, a critical gap that hinders the progress of target-driven materials design. To address this limitation, we propose MatFormBench, a novel benchmarking ecosystem tailored to evaluate and guide generative strategies for target-driven formulation. MatFormBench integrates a physics-driven formulation generation scheme to generate synthetic samples that faithfully emulate realistic materials structure-property response relationships, complemented by five escalating difficulty levels to quantify the complexity of these relationships. To rigorously assess algorithm performance, we further propose MatFormScore, a multi-dimensional metric that comprehensively quantifies performance across five critical axes: target success, search efficiency, exploratory capacity, robustness, and stability. We validate MatFormBench by evaluating 39 diverse inverse design algorithms, covering classical surrogate-assisted black-box search, state-of-the-art deep generative models, and increasingly popular Large Language Model (LLM)-based recommendation strategies. Across 1170 standardized algorithm-task evaluations, diffusion-based models demonstrate the strongest overall performance, while Variational Autoencoder (VAE)-based and Genetic Algorithm (GA)-based methods exhibit distinct advantages in specific scenarios. By establishing a unified evaluation standard for target-driven materials formulation, MatFormBench enables reproducible benchmarking, principled algorithm comparison, and diagnostic analysis of inverse design strategies, providing a foundational tool for advancing materials inverse design.\n\n---\n\n## 77. Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models\n- **Authors**: Xiao-Wen Yang , Ziyu Han , Xi-Hua Zhang , Wen-Da Wei , Jie-Jing Shao , Lan-Zhe Guo , Yu-Feng Li\n- **URL**: [https://arxiv.org/abs/2605.26733](https://arxiv.org/abs/2605.26733)\n- **Abstract**:\n> Looped Language Models (LoopLMs) enable efficient latent reasoning through depth recurrence, yet exhibit unreliable test-time scaling behavior: performance often peaks at a certain iteration depth and then collapses with further recurrence. Through latent dynamics analysis, we find an inherent trade-off between stability and effectiveness in existing architectures and strategies. By conceptualizing reasoning as uncertainty reduction, we propose that convergence toward stable fixed points while preserving effectiveness represents a promising way. To this end, we propose STARS (STAbility-driven Recurrent Scaling), a training framework that constrains latent states to approach asymptotically stable fixed points. This is realized via efficient Jacobian Spectral Radius Regularization with random loop sampling, enabling STARS to maximize effectiveness while ensuring rigorous stability. Experiments on arithmetic tasks show that STARS achieves reliable test-time scaling, and on complex mathematical reasoning it substantially mitigates performance degradation as recurrence depth increases while also improving peak performance.\n\n---\n\n## 78. L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation\n- **Authors**: Pingjun Pan , Tingting Zhou , Peiyao Lu , Tingting Fei , Hongxiang Chen , Chuanjiang Luo\n- **URL**: [https://arxiv.org/abs/2605.26717](https://arxiv.org/abs/2605.26717)\n- **Abstract**:\n> Adapting large language models (LLMs) for personalized recommendation requires aligning their general-purpose capabilities with user-specific preferences while effectively leveraging both behavioral and semantic signals. Existing approaches typically integrate these signals at either the input level (e.g., injecting behavioral embeddings into the token space) or the output level (e.g., contrastive alignment of separate encoders), suffering from distribution gaps or lack of end-to-end task supervision. In this work, we introduce L2Rec, which unifies behavioral and semantic understanding at the parameter level of LLMs. Our key insight is that the same set of Transformer parameters can serve as a shared medium for both views: by applying view-specific, personalized low-rank perturbations via a Dual-view Personalized Mixture-of-Experts (DPMoE) mechanism, L2Rec enables a single LLM backbone to produce complementary behavioral and semantic adaptations for each user with minimal representation-level misalignment. An adaptive cross-view fusion module further integrates the dual-view outputs into a unified user preference. Experiments on four datasets show that L2Rec consistently outperforms state-of-the-art baselines, and online A/B testing on a large-scale industrial platform validates significant improvements in key engagement metrics.\n\n---\n\n## 79. Beyond Trajectory-Level Attribution: Graph-Based Credit Assignment for Agentic Reinforcement Learning\n- **Authors**: Xin Cheng , Shuo He , Lang Feng , HaiYang Xu , Ming Yan , Lei Feng , Bo An\n- **URL**: [https://arxiv.org/abs/2605.26684](https://arxiv.org/abs/2605.26684)\n- **Abstract**:\n> Group-based reinforcement learning (RL) methods have achieved remarkable success in improving the performance of large language models (LLMs) and have been rapidly extended to agentic tasks. However, their credit assignment relies heavily on coarse-grained trajectory-level attribution according to final outcomes, making it difficult to capture the contribution of individual steps, such as valuable steps obscured within failed trajectories. To uncover latent information and enable more faithful step-level credit assignment, we propose Graph-based Group Policy Optimization (GraphGPO), which first aggregates all rollout trajectories into a unified state-transition graph and then estimates the distance from each state to the task goal using the global information encoded in the graph. Finally, GraphGPO assigns credit to each edge by estimating a graph-based advantage, based on how much the transition reduces the distance to the task goal. In this way, GraphGPO significantly improves training efficiency and achieves state-of-the-art performance across a range of challenging benchmarks.\n\n---\n\n## 80. An In-Vitro Study on Cross-Lingual Generalization in Language Models\n- **Authors**: Adrian Cosma\n- **URL**: [https://arxiv.org/abs/2605.26683](https://arxiv.org/abs/2605.26683)\n- **Abstract**:\n> Cross-lingual transfer in language models is difficult to study in natural corpora because lexical overlap, morphology, data imbalance, and tokenization are entangled. We introduce an in-vitro framework with two procedurally generated languages that share the same ontology, typed grammar, and compositional structure, but differ in surface realization. This lets us independently vary lexical distance, minority-language proportion, tokenizer training regime, and vocabulary size, while evaluating transfer on a masked minority-language condition whose lexical forms are never observed during training. Across 700 controlled runs, we find that transfer is governed less by tokenizer balance or raw lexical similarity than by whether tokenization preserves reusable cross-lingual substructure. Smaller vocabularies often improve masked transfer by keeping words decomposable into shared fragments, whereas larger vocabularies can turn forms into language-specific atoms. We further show that transfer emerges as a staged process: grammatical and type-level competence precede masked lexical generalization. Finally, we attempt to explain this mechanism through tokenizer bridges and show that bridge strength correlates strongly with masked reachability.\n\n---\n\n## 81. DynFrame: Adaptive Reasoning-Driven Multimodal Framework with Dynamic Frame Augmentation for Complex Video Understanding\n- **Authors**: Peng Zhang , Guanghao Zhang , Wanggui He , Longxiang Zhang , Mushui Liu , Yan Xia , Zhenhao Peng , Weilong Dai , Jinlong Liu , Haobing Tang , Le Zhang , Hao Jiang , Pipei Huang\n- **URL**: [https://arxiv.org/abs/2605.26680](https://arxiv.org/abs/2605.26680)\n- **Abstract**:\n> Recent video multimodal large language models (MLLMs) increasingly couple step-by-step reasoning with on-demand visual evidence retrieval, allowing models to revisit relevant video segments during inference. However, two structural gaps remain in existing thinking-with-video systems. (i) Sampling density is not a learnable decision: existing methods may let the model decide where to look, but the per-window frame rate is largely fixed. As a result, fine-grained evidence is often recovered through repeated retrieval calls, which increases inference context length and training difficulty. (ii) Retrieval and answer generation are usually optimized with a single trajectory-level advantage, so the \"where to look\" tokens and the \"how to answer\" tokens receive the same credit even when one is correct and the other is not. To address these gaps, we present DynFrame, a framework that emits the temporal window and the sampling density as native tokens within a single autoregressive pass. This learnable span-density retrieval enables acquiring multi-granularity evidence with a single retrieval step. Based on the above tokenized retrieval interface, we further introduce Segment-Decoupled GRPO (SD-GRPO), which splits each rollout at the retrieval boundary and assigns role-specific token-level advantages, separately crediting the sampling decision and the answer. Trained on the curated DM-CoT-74k and DM-RL-45k, DynFrame-4B is competitive with strong 7B-8B baselines across six benchmarks (NExT-GQA, Charades-STA, ActivityNet-MR, Video-MME, MLVU, LVBench), and DynFrame-8B sets new state-of-the-art on most metrics. Code is available at this https URL .\n\n---\n\n## 82. The Labyrinth and the Thread: Rethinking Regularizations in Sequential Knowledge Editing for Large Language Models\n- **Authors**: Zheng Wang , Kaixuan Zhang , Wanfang Chen , Jingwen Zhang , Xiaonan Lu\n- **URL**: [https://arxiv.org/abs/2605.26670](https://arxiv.org/abs/2605.26670)\n- **Abstract**:\n> Sequential editing of structured knowledge in large language models allows targeted factual updates without retraining, yet existing methods often rely on complex regularization or constraint mechanisms whose necessity remains unclear. In this work, we systematically investigate the mechanisms underlying effective and stable sequential editing. Specifically, we first analyze the empirical success of AlphaEdit and establish, via a rigorous optimization analysis, the formal equivalence between one-time and sequential editing. Building on this insight, we generalize the equivalence to a broader class of editing objectives, demonstrating that stability emerges naturally from properly accounting for accumulated editing constraints, rather than from specialized regularization or null-space operations. We empirically confirm that many commonly used regularization strategies are unnecessary for reliable sequential updates. Furthermore, we extend our framework to handle conflicting edits, ensuring robust and consistent behavior under contradictory updates. Ultimately, our work provides Ariadne's thread through the labyrinth of sequential editing, charting a path toward simpler, more interpretable, and dependable knowledge updates. Our code is available at this https URL .\n\n---\n\n## 83. AI evaluation may bias perceptions: The importance of context in interpreting academic writing\n- **Authors**: Shang Wu , Randol Yao\n- **URL**: [https://arxiv.org/abs/2605.26662](https://arxiv.org/abs/2605.26662)\n- **Abstract**:\n> This paper examines how estimates of AI use in scientific writing can be biased when evaluation methods ignore contextual differences across countries and fields. Using large-scale data on journal publications from Dimensions, we construct AI-likeness benchmarks based on differences between human-written and LLM-rephrased abstracts. We show that a pooled benchmark may confound pre-existing stylistic variation with AI-generated text, producing substantial distortions across country-field groups even in pre-LLM publications. In contrast, country-field-specific benchmarks attenuate such distortions and provide a more credible baseline for comparison. Applying these methods to publications in 2025 reveals that the pooled benchmark systematically overestimates AI use in certain countries and fields while underestimating it in others. These findings highlight the importance of context-aware measurement for accurate and equitable evaluation of AI use in science.\n\n---\n\n## 84. Respecting Modality Gap in Post-hoc Out-of-distribution Detection with Pre-trained Vision-Language Models\n- **Authors**: Yuanwei Hu , Bo Peng , Yadan Luo , Zhen Fang , Ling Chen , Jie Lu\n- **URL**: [https://arxiv.org/abs/2605.26661](https://arxiv.org/abs/2605.26661)\n- **Abstract**:\n> Out-of-distribution (OOD) detection has emerged as a popular technique to enhance the reliability of machine learning models by identifying unexpected inputs from unknown classes. Recent progress in pre-trained vision-language models (VLMs) has enabled zero-shot OOD detection without access to in-distribution (ID) training data; in this setting, existing methods commonly treat text embeddings of class names as class prototypes. In this paper, we challenge the widely adopted text-as-prototype paradigm by theoretically showing that off-the-shelf textual prototypes are generally misaligned with the optimal visual prototypes, yielding an intrinsic modality gap that cannot be eliminated by prompt engineering alone. To mitigate this gap under the post-hoc constraint, this paper presents an online pseudo-supervised framework that directly learns class prototypes in the visual feature space using unlabeled test-time data streams and soft predictions from the pre-trained VLMs. We provide theoretical guarantees for the convergence of the online optimization procedure. Extensive experiments empirically demonstrate that our method achieves a new state of the art across a variety of OOD detection setups.\n\n---\n\n## 85. More Expressive Feedforward Layers: Part I. Token-Adaptive Mixing of Activations\n- **Authors**: Mingze Wang , Jinbo Wang , Yikuan Xia , Kai Shen , Shu Zhong\n- **URL**: [https://arxiv.org/abs/2605.26647](https://arxiv.org/abs/2605.26647)\n- **Abstract**:\n> Feedforward network (FFN) layers account for a large fraction of parameters and nonlinear expressivity in Transformer-based large language models (LLMs). Despite the evolution from ReLU and GELU to gated variants such as SwiGLU, most FFN designs still use a single fixed activation function, applying the same nonlinear transformation to all tokens. In this work, we propose Mixture of Activations (MoA), a token-adaptive FFN design that mixes a dictionary of activation functions using lightweight input-dependent gates while sharing the same linear projections. As an input-independent counterpart, we also introduce learnable activations (LA), which form linear combinations of activation functions for both ReLU-type and SwiGLU-type FFNs. Theoretically, we establish strict finite-width expressive separations among fixed-activation FFNs, LA, and MoA: LA strictly contains fixed-activation FFNs, while MoA strictly contains LA, with the additional expressivity arising from input-dependent nonlinear hybridization. Empirically, we evaluate MoA through extensive pre-training experiments on dense and MoE language models ranging from 0.12B to 2B parameters under different token budgets, optimizers, and learning rate schedules. MoA consistently achieves lower terminal loss and exhibits more favorable scaling behavior than well-tuned baselines, with minimal parameter and computational overhead. These results suggest that token-adaptive activation mixing is a simple and effective mechanism for improving FFN expressivity in LLMs.\n\n---\n\n## 86. Spend Your Rollouts Where It Counts: Rollout Allocation for Group-Based RL Post-Training\n- **Authors**: Woojeong Kim , Ziyi Yang , Jing Nathan Yan , Jialu Liu\n- **URL**: [https://arxiv.org/abs/2605.26606](https://arxiv.org/abs/2605.26606)\n- **Abstract**:\n> Reinforcement learning (RL) is the dominant paradigm for post-training large language models. However, in the online, on-policy setting, rollout generation dominates the computational cost of training. Group-based policy optimization methods compute advantages from multiple rollouts per prompt, yet they indiscriminately allocate budget to prompts with collapsed reward distributions, wasting expensive rollouts on negligible learning signals. We demonstrate that group-based updates are most effective in regimes of high reward variance. Since the policy evolves throughout training, prompt informativeness must be estimated online rather than precomputed, but exhaustively evaluating every prompt is computationally prohibitive. We introduce Pilot-Commit, a budget-aware rollout allocation framework for group-based RL post-training. Pilot-Commit decouples prompt evaluation from exploitation: a pilot stage estimates per-prompt informativeness using a fraction of the budget, and the remaining rollouts are allocated to high-leverage prompts while low-signal prompts are skipped. Across multiple math reasoning benchmarks and model scales from 1.5B to 14B parameters, Pilot-Commit matches baseline accuracy with significantly lower sampling costs, reaching target accuracy up to $1.9\\times$ faster than GRPO and $4.0\\times$ faster than DAPO in cumulative rollouts.\n\n---\n\n## 87. Cordyceps: Covert Control Attacks on LLMs via Data Poisoning\n- **Authors**: Zedian Shao , Charles Fleming , Teodora Baluta\n- **URL**: [https://arxiv.org/abs/2605.26595](https://arxiv.org/abs/2605.26595)\n- **Abstract**:\n> Large language models (LLMs) are often fine-tuned on uncurated text datasets that adversaries can poison. Existing poisoning attacks primarily rely on fixed trigger phrases that defenses such as outlier detection, clean-data regularization, or online monitoring can neutralize. In this paper, we propose a data poisoning method that teaches an LLM an information hiding scheme reliably and stealthily through semantic associations between shared knowledge such as facts or concepts and attacker-chosen phrases. The induced hiding scheme can encode and decode arbitrary malicious instructions, thus revealing a new and subtle poisoning-induced vulnerability: covert control attacks. We precisely characterize covert control attacks and evaluate them across $5$ LLMs, $3$ backdoor defenses, and $4$ prompt injection defenses. With a small poisoned fraction, covert control attacks outperform heuristic-based prompt injection attacks in average attack success rate by about $40\\%$ relative to clean fine-tuned models. They also circumvent defenses based on detection and fine-tuning, maintaining up to $93\\%$ attack success rate after backdoor defenses and up to $98\\%$ after prompt injection defenses.\n\n---\n\n## 88. Linear and Neural Dueling Bandits with Delayed Feedback\n- **Authors**: Xiangyi Wang , Pingchen Lu , Jie Mao , Mingze Kong , Zhi Hong , Zhiyong Wang , Zhongxiang Dai\n- **URL**: [https://arxiv.org/abs/2605.26554](https://arxiv.org/abs/2605.26554)\n- **Abstract**:\n> Contextual dueling bandits form a cornerstone of preference-based decision-making, with critical applications in recommender systems and large language model alignment. However, standard algorithms rely on the idealized assumption of immediate feedback, a condition frequently violated in real-world scenarios such as prompt optimization. This setting introduces a unique theoretical challenge: unlike linear bandits, dueling bandit estimators lack closed-form solutions, rendering naive adaptations of standard weighting techniques biased. To address this, we formalize the problem of Contextual Dueling Bandits with Stochastic Delayed Feedback and propose two novel algorithms: Linear (LDB-DF) and Neural (NDB-DF) Dueling Bandits with Delayed Feedback. Central to our approach is a novel estimator that integrates an Inverse Probability Weighting (IPW) mechanism directly into the loss function, ensuring unbiased correction for delayed or missing feedback. We provide comprehensive theoretical analysis, establishing an O(d*sqrt(T)) regret bound for the linear setting and sub-linear guarantees for the neural setting. Extensive experiments on both simulated and real-world datasets demonstrate the effectiveness of our propose.\n\n---\n\n## 89. A Hybrid Vision-Language Architecture for Automated Defect Reasoning and Report Generation in Industrial Inspection\n- **Authors**: Malikussaid , Imad Gohar\n- **URL**: [https://arxiv.org/abs/2605.26533](https://arxiv.org/abs/2605.26533)\n- **Abstract**:\n> Automated industrial inspection requires both precise defect localization and structured maintenance report generation; in current practice these tasks are handled separately, with linguistic interpretation left to human experts. This paper describes a decoupled, edge-deployable pipeline for wind turbine blade inspection built from three components that each handle a distinct sub-task. The Eyes a YOLO26-x-obb oriented bounding-box detector localizes defects at dataset-native resolution. The Bridge a deterministic, parameter-free encoding module maps each detected bounding box to grid-referenced spatial tokens embedded in a structured prompt. The Brain a 4-bit quantized Qwen-2.5-1.5B model adapted with Quantized Low-Rank Adaptation (QLoRA) on 947 synthetically generated maintenance reports generates a structured JSON report from that prompt. Retrieval-Augmented Fine-Tuning (RAFT) further grounds each recommendation in indexed maintenance procedures. Five ablation experiments, scored by BLEU-4, ROUGE-L, Hallucination Rate (HR), and an LLM-as-a-Judge rubric, compare the pipeline against a monolithic vision-language model (VLM) baseline and against partial configurations in which one component is removed. The complete system achieves BLEU-4 0.41, HR=4%, and Expert Score = 8.6/10 compared with 0.07, 65%, and 3.3/10 for the zero-shot VLM baseline. The QLoRA-adapted 1.5B model generates higher-quality reports than a 671B-parameter generalist API model given identical detection evidence, at 47 tokens per second on a single T4-class GPU. The results show that purpose-built decoupled architecture with a small domain-specific training corpus outperforms a generalist end-to-end model on this structured generation task.\n\n---\n\n## 90. InterSketch: An Interleaved Reasoning Model with Self-correcting Visual Sketch and Stepwise Reward\n- **Authors**: Zhiwei Ning , Wenwen Tong , Xiangli Kong , Shengnan Ma , Ziyi Shang , Jingcheng Ni , Tao Hu , Yong Xien Chng , Jixuan Ying , Zehuan Wu , Hanming Deng , Jie Yang , Yuanjie Zheng , Wei Liu , Lewei Lu\n- **URL**: [https://arxiv.org/abs/2605.26520](https://arxiv.org/abs/2605.26520)\n- **Abstract**:\n> While vision-language models (VLMs) have exhibited multi-turn visual reasoning capabilities, their reasoning trajectories remain relatively shallow and are dominated by a text-centric paradigm, limiting their applicability to complex visual challenges. In contrast, human-like thought typically involves long-horizon reasoning with an interleaved visual-textual chain-of-thought (VT-CoT). To bridge this gap, we introduce InterSketch, an interleaved reasoning model to enhance the VT-CoT capability via self-correcting and stepwise reward mechanisms. InterSketch dynamically generates intermediate visual sketches using external tools and interleaves them with textual reasoning, enabling effective perception and logical reasoning over long-horizon visual understanding tasks. Specifically, in the first cold-start stage, we propose a synthesized high-quality interleaved VT-CoT dataset and include a reflection mechanism to enable the model's capability in multi-turn interleaved reasoning and self-correction. In the subsequent reinforcement learning (RL) stage, we design a stepwise reward mechanism to mitigate the sparsity of reward signals inherent in end-only supervision over long-horizon reasoning. Extensive experiments on visual reasoning benchmarks demonstrate the effectiveness of InterSketch, even outperforming proprietary models such as Gemini-3-Pro.\n\n---\n\n## 91. Unveiling the Fragility of Vision-Language Models: Multi-Modal Adversarial Synergy via Texture-Constrained Perturbations and Cross-Modal Optimization\n- **Authors**: Xiang Fang , Wanlong Fang , Changshuo Wang\n- **URL**: [https://arxiv.org/abs/2605.26501](https://arxiv.org/abs/2605.26501)\n- **Abstract**:\n> Large Vision-Language Models (LVLMs) have transformed multi-modal understanding, excelling in tasks like image captioning and visual question answering by integrating visual and textual inputs. However, their robustness against adversarial attacks, particularly those exploiting both modalities, remains underexplored, posing risks to critical applications like autonomous driving and content moderation. Existing attacks focus on single modalities or require impractical white-box access, limiting their real-world relevance. In this paper, we introduce Multi-Modal Adversarial Synergy, a groundbreaking framework that crafts universal, black-box multi-modal attacks against LVLMs. MMAS simultaneously generates a texture scale-constrained universal adversarial perturbation for images and a learnable prompt perturbation for text, optimized jointly using only model queries. The image perturbation leverages wavelet-based texture constraints to ensure imperceptibility and robustness across diverse visual inputs. The text perturbation, constrained by an L-norm in the embedding space, maintains semantic coherence while steering outputs toward a target. A novel cross-modal regularization term aligns the perturbations' gradient directions, enhancing their synergistic impact and transferability across tasks and models. Extensive experiments show the strong universal adversarial capabilities of our proposed attack with prevalent LVLMs.\n\n---\n\n## 92. Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories\n- **Authors**: Sil Hamilton , David Mimno\n- **URL**: [https://arxiv.org/abs/2605.26492](https://arxiv.org/abs/2605.26492)\n- **Abstract**:\n> LLM-generated stories are a popular use case, but they show very low variability. We sample 20,000 total stories from four current models using five prompts. We find that 11 words occur in 88.3% of generated stories, with little difference between models. These words include names (Elias, Mara, Elara), settings (lighthouses), and professions (clockmaker, librarian). These tokens do not often occur in published literature nor pre-training data, but they are found in preference data that is likely to have been used by all current models. Surprisingly, these \"lighthouse\" stories are infrequent when compared with the average post-training story, much of which contains references to copyrighted characters or adult content. This result demonstrates the potentially disproportionate impact of small datasets combined with powerful alignment algorithms.\n\n---\n\n## 93. Towards Error-Free EHRs: Reasoning-Intensive Consistency Verification Between Clinical Notes and Structured Tables in Electronic Health Records\n- **Authors**: Yeonsu Kwon , Jiho Kim , Junseong Choi , Paloma Rabaey , Minseo Kim , Sujeong Im , Jeewon Yang , Jun-Min Lee , Sangji Lee , Jiwon Kim , Hangyul Yoon , Hyunwook Kwon , Edward Choi\n- **URL**: [https://arxiv.org/abs/2605.26463](https://arxiv.org/abs/2605.26463)\n- **Abstract**:\n> Data consistency between unstructured clinical notes and structured tables in Electronic Health Records (EHRs) is essential for patient safety and clinical decision-making. However, existing work on note-table consistency verification mainly relies on surface-level matching of numeric values or simple events. Such approaches fail to capture the reasoning underlying real-world EHR documentation, including clinical interpretation, event relations, and temporal changes. To address this gap, we introduce EHR-ReasonCon, a reasoning-intensive benchmark for note-table consistency verification. Built on MIMIC-III with expert-guided annotations, it comprises 8,048 entities derived from clinical notes and provides high-quality ground-truth labels. The annotation protocol is supported by specialized table-exploration tools to ensure systematic evidence retrieval and reliable consistency assessment. We also propose EHR-Inspector, an LLM-based framework that segments notes, extracts anchor entities and temporal references, and uses table-exploration tools to verify consistency against structured tables. Evaluated using expert-validated LLM-as-a-judge metrics under harsh and lenient criteria, EHR-Inspector achieves state-of-the-art performance across multiple model backbones. Analyses further demonstrate the effectiveness of its components and highlight differences from human verification.\n\n---\n\n## 94. Verus-SpecGym: An Agentic Environment for Evaluating Specification Autoformalization\n- **Authors**: Anmol Agarwal , Natalie Neamtu , Pranjal Aggarwal , Seungone Kim , Jannis Limperg , Cedric Flamant , Kanna Shimizu , Bryan Parno , Sean Welleck\n- **URL**: [https://arxiv.org/abs/2605.26457](https://arxiv.org/abs/2605.26457)\n- **Abstract**:\n> AI coding agents are increasingly used to write real-world software, but ensuring that their outputs are correct remains a fundamental challenge. Formal verification offers a promising path: an agent generates code together with a machine-checked proof, guaranteeing that the code satisfies a formal specification. However, there is no guarantee that the formal spec itself matches the user's intent. In this work, we study specification autoformalization: whether LLM agents can translate informal programming problems into faithful formal specifications. We introduce Verus-SpecBench, a benchmark of 581 spec-writing tasks derived from Codeforces problems targeting Verus, a verifier for Rust, and Verus-SpecGym, an agentic environment in which models interact with Verus, bash, & the filesystem to develop these specs. The central challenge is evaluation: expert-written reference specs are expensive to write, & LLM judges can miss subtle mistakes. We address this by (a) extending Verus's exec_spec mechanism so that generated specs can be executed as Rust code, & (b) testing them against official Codeforces tests & adversarial cases extracted from Codeforces \"hacks\", which are edge cases written by competitors to break incorrect solutions. On Verus-SpecBench, the strongest model, Gemini 3.1 Pro, solves 77.8% of tasks, other frontier models solve 51.1--57.8% & OSS models reach only 21.5--25.5%. Our analysis of failure modes shows that model-generated specs can omit important input assumptions, accept incorrect outputs, & reject valid ones. We also find that LLM-as-a-judge evaluation misses 26% of the failures our evaluator catches. Overall, our results suggest that spec autoformalization is within reach for frontier agents but remains brittle even on problems where they can already generate correct code. The code, data, & logs can be found at this https URL\n\n---\n\n## 95. Alignment Tuning for Large Language Models: A Data-Centric Lens on Alignment Data Pipelines\n- **Authors**: Hwanjun Song\n- **URL**: [https://arxiv.org/abs/2605.26442](https://arxiv.org/abs/2605.26442)\n- **Abstract**:\n> Much of the alignment tuning literature is organized around optimization objectives, while the construction of alignment data is often treated implicitly. In this survey, we adopt a data centric perspective and reframe alignment tuning as a pipeline design problem. We decompose alignment data construction into three interacting stages, response synthesis, preference evaluation, and preference instantiation, and use this framework to organize existing alignment methods into a unified taxonomy. Through this lens, we identify recurring design trade-offs and failure modes observed across prior alignment methods, and distill a set of high level principles that clarify how pipeline design choices influence the resulting optimization signal. Finally, we outline open challenges for alignment data pipelines, including prompt-level alignment, agentic settings, and alignment under evolving objectives.\n\n---\n\n## 96. LURE: Live-Usage Replay Evaluations for Reducing Evaluation Awareness\n- **Authors**: Igor Ivanov , David Demitri Africa\n- **URL**: [https://arxiv.org/abs/2605.26438](https://arxiv.org/abs/2605.26438)\n- **Abstract**:\n> Large language models can recognize when they are being evaluated (evaluation awareness) and behave differently because of that, which undermines the validity of safety and alignment benchmarks. We propose LURE (Live-Usage Replay Evaluations), a method for constructing deployment-like evaluations by replaying realistic agentic interaction trajectories and appending evaluation prompt at the end. We also introduce an automated pipeline for measuring evaluation realism, combining detection of verbalized evaluation awareness and judge-model estimates of the probability of logs being an evaluation, and validate it on a large dataset of deployment and evaluation transcripts. We find that LURE-based evaluations are substantially less distinguishable from deployment than widely used benchmarks and synthetic evaluation generators, and can approach the realism of real conversations with users. We instantiate LURE in scheming, AI safety sabotage, and sycophancy settings. Our results suggest that evaluation realism is a crucial property of alignment benchmarks and should be reported alongside benchmark results, especially when such results are used in safety cases.\n\n---\n\n## 97. Targeted Remasking: Replacing Token Editing with Token-to-Mask Refinement in Discrete Diffusion Language Models\n- **Authors**: Lin Yao\n- **URL**: [https://arxiv.org/abs/2605.26436](https://arxiv.org/abs/2605.26436)\n- **Abstract**:\n> Discrete masked diffusion language models such as LLaDA generate text through iterative denoising, where mask tokens are progressively replaced with predicted tokens. LLaDA2.1 introduced a Token-to-Token (T2T) editing mechanism that accelerates generation by directly replacing committed tokens suspected of being incorrect. However, we identify fundamental limitations of T2T editing: it couples error detection with replacement, pollutes the generation context with potentially incorrect tokens, and introduces a train-inference noise mismatch where systematic model-generated errors differ from the random perturbations seen during training. We propose Token-to-Mask (T2M) remasking, a training-free, drop-in replacement for T2T editing that resets suspected erroneous tokens back to the mask state, allowing the diffusion process to re-predict them under cleaner context. We design and empirically validate three complementary error detection strategies -- probability-based, trigger-mirrored, and temporal-difference-based -- and provide a unified theoretical analysis showing that T2M remasking purifies the generation context, converts systematic inference errors back to the model's native mask noise type, and enables delayed commitment for joint multi-position optimization. Comprehensive experiments across 12 benchmarks spanning knowledge, reasoning, mathematics, coding, and instruction following show that T2M generally improves performance on tasks requiring precise token-level output, with the largest gain on mathematics (+5.92% on CMATH). Error analysis on CMATH reveals that the dominant failure mode is last-mile token corruption -- where correct reasoning produces a corrupted final answer -- and that T2M repairs 59.4% of such cases.\n\n---\n\n## 98. The Rescue Effect: Spatio-Semantic Early Exit Bypasses Quantization Collapse in CLIP\n- **Authors**: Kahyeon Nam , Hyesong Choi\n- **URL**: [https://arxiv.org/abs/2605.26415](https://arxiv.org/abs/2605.26415)\n- **Abstract**:\n> Deploying Vision-Language Models on resource-constrained hardware typically requires INT8 quantization, but in joint-embedding architectures such as CLIP this introduces a failure mode distinct from quantized CNN classifiers: activation noise accumulated across transformer blocks perturbs the direction of the multimodal embedding, eroding the cosine alignment on which zero-shot retrieval depends. We characterize this as Quantization-Induced Representation Collapse (QIRC) and quantify it on INT8 CLIP ViT-B/32, where the layer-wise noise-to-signal ratio grows from below 10% in shallow blocks to 52% at Layer 11. We propose LRA-EE (Layer-wise Representation-Aware Early Exit), which bypasses noise-saturated deep layers via Spatio-Semantic Aggregation (replacing the immature shallow [CLS] with a global patch-token average), a learned multi-feature gate (confidence, top-2 margin, spatial-activation variance), and Layer-adaptive Confidence Thresholding calibrated to each layer's Information-to-Noise Ratio. On ImageNet-1K zero-shot classification, LRA-EE reduces FLOPs by 13.4% and improves Top-1 accuracy by +2.44%p (58.72% -> 61.16%) over the INT8 baseline. A four-quadrant decomposition isolates the Rescue Effect: 9.5% of samples are correctly classified at shallow exits but lost to noise at full depth, against only 7.1% suffering the inverse.\n\n---\n\n## 99. Plans for Evaluating Structured Generative Search Summaries\n- **Authors**: Tetsuya Sakai , Jina Lee , Hanpei Fang , Young-In Song\n- **URL**: [https://arxiv.org/abs/2605.26400](https://arxiv.org/abs/2605.26400)\n- **Abstract**:\n> We propose a framework for evaluating structured generative search summaries that are placed atop organic web search results. A structured summary, generated by a large language model, typically consists of an overview, several sections with section titles, and a list of source documents that are cited within the summary. We then describe our plans for implementing and evaluating the framework.\n\n---\n\n## 100. Annotator Positionality as Signal: Psychometric Weighting for Anti-Autistic Ableism Detection\n- **Authors**: Naba Rizvi , Harper Strickland , Saleha Ahmedi , Nedjma Ousidhoum\n- **URL**: [https://arxiv.org/abs/2605.26397](https://arxiv.org/abs/2605.26397)\n- **Abstract**:\n> Large language models (LLMs) are increasingly used in decision-making tasks where they can amplify or suppress perspectives, raising concerns in high-stakes settings affecting autistic communities. While previous research has identified disability-related biases in LLMs, it remains unclear how they conceptualize ableism or detect it in text. We introduce a bias-aware evaluation framework targeting anti-autistic ableist language with a psychometrically-weighted, community-proximate ground truth anchored in annotator positionality. This framework constitutes a stricter standard than conventional majority-vote aggregation which significantly and consistently underweights autistic and autism-accepting perspectives. We find that LLMs frequently produce harmful outputs, mislabel community-reclaimed language as ableist, and express more negative attitudes toward autistic people when assessment instruments are masked. Our error analysis reveals that models rely on surface-level keyword matching rather than contextual factors such as speaker identity, and whether the language fosters in-group solidarity or inflicts out-group harm.\n\n---\n\n## 101. VisualNeedle: Benchmarking Active Visual Search in Information-Dense Scenes\n- **Authors**: Jingru Chen , Yiming Liu , Mingtao Chen , Sijie Chen , Richeng Xuan , Liang Yang , Zhichao Hu , Fanyang Lu\n- **URL**: [https://arxiv.org/abs/2605.26380](https://arxiv.org/abs/2605.26380)\n- **Abstract**:\n> Frontier multimodal large language models (MLLMs) have been reported to achieve over 90% accuracy on fine-grained perception benchmarks. However, such scores do not necessarily imply faithful use of visual evidence. Prior studies have identified three shortcuts that inflate benchmark performance. First, linguistic priors and lexical cues in questions often enable models to infer plausible answers without seeing the image. Second, coarse global semantics from the visual encoder can bypass fine-grained local details. Third, in some ``think-with-images'' benchmarks, corrupting the intermediate images returned by visual tools barely affects the final answer. These findings suggest that higher input resolution or larger question pools alone do not elicit genuine active visual search. To address this, we introduce VisualNeedle, a challenging, information-dense, and fine-grained benchmark for scenes where critical evidence is spatially constrained to minute regions and not discernible at a glance. We further propose a counterfactual crop-black setting, which replaces crops returned by tools with black images of the same size, to test whether tool-enabled performance truly relies on intermediate visual evidence. We evaluate 9 promninent MLLMs across three settings: no-tool, standard tool-enabled, and crop-black. No-tool accuracy stays below 20\\%, and the best tool-enabled model reaches only 56.01\\%, still trailing the 63.00% human majority-vote accuracy. These results reveal persistent limitations in fine-grained visual search, while the crop-black ablation confirms that success on VisualNeedle hinges on genuine intermediate visual evidence.\n\n---\n\n## 102. BioFact-MoE: Biologically Factorized Mixture of Experts for Vision-Language Prognostic Modeling in Hepatocellular Carcinoma\n- **Authors**: Junlin Yang , Tian Yu , Nicha C. Dvornek , Yuexi Du , Peiyu Duan , Annabella Shewarega , Lawrence H. Staib , James S. Duncan , Julius Chapiro\n- **URL**: [https://arxiv.org/abs/2605.26376](https://arxiv.org/abs/2605.26376)\n- **Abstract**:\n> Hepatocellular carcinoma (HCC) is biologically heterogeneous, shaped by the interplay between hepatic functional reserve and tumor-related oncologic factors; thus, similar survival outcomes may reflect fundamentally different underlying biological processes. Prognostic modeling in HCC is informed by rich multimodal information from multiparametric MRI and radiology reports from routine clinical practice. Existing prognostic vision-language models (VLMs) learn a single entangled latent representation that blends hepatic and tumor-related factors, limiting both accuracy and biological interpretability. We present BioFact-MoE, a biologically factorized Mixture of Experts (MoE) framework that explicitly decomposes liver and tumor factors via biologically supervised experts within a residual MoE survival architecture. On a HCC cohort of N=588 patients (pretrained on 4,582 3D MRI image-report pairs), BioFact-MoE consistently improves survival prediction over all baselines across time horizons, achieving 12-, 18-, and 24-month AUCs of 75.33%, 75.85%, and 73.96%. Beyond scalar risk prediction, gated expert weights enable phenotype-aware risk stratification. Pathway-informed gating uncovers clinically meaningful treatment-associated survival heterogeneity. In held-out validation, hepatic and tumor embeddings show selective associations with liver function and tumor burden markers, respectively (p<0.05), without supervision. The code is available at this https URL .\n\n---\n\n## 103. Why LLMs Hallucinate on Structured Knowledge: A Mechanistic Analysis of Reasoning over Linearized Representations\n- **Authors**: Shanghao Li , Jinda Han , Yibo Wang , Yuanjie Zhu , Zihe Song , Langzhou He , Kenan Kamel A Alghythee , Philip S. Yu\n- **URL**: [https://arxiv.org/abs/2605.26362](https://arxiv.org/abs/2605.26362)\n- **Abstract**:\n> In many reasoning tasks, large language models (LLMs) rely on structured external knowledge, such as graphs and tables, which is typically linearized into sequential token representations. However, even when sufficient knowledge is available, LLMs can still produce hallucinated outputs, and the underlying mechanisms behind such failures remain poorly understood. We investigate these mechanisms and find that hallucinations arise from systematic internal dynamics rather than random noise. First, attention disproportionately concentrates toward shortcut-like structural cues rather than distributing across the full context. Second, feed-forward representations fail to ground the provided knowledge, causing the model to revert to parametric memory. Moreover, our results indicate that hallucination is consistently associated with failures in semantic grounding within feed-forward layers, while attention allocation exhibits greater task-dependent variability. Finally, we show that these mechanistic patterns generalize beyond single-hop graphs to multi-hop and tabular settings, enabling effective hallucination detection across structured knowledge formats.\n\n---\n\n## 104. Erased but Exploitable: Black-box Embedding-Aware Prompting Against Unlearned Text-to-Image Diffusion Models\n- **Authors**: Arian Komaei Koma , Seyed Amir Kasaei , AmirMahdi Sadeghzadeh , Mohammad Hossein Rohban\n- **URL**: [https://arxiv.org/abs/2605.26332](https://arxiv.org/abs/2605.26332)\n- **Abstract**:\n> Machine unlearning aims to remove specific concepts from pretrained text-to-image diffusion models, yet several white- and black-box attacks have been introduced to make the model generate such unlearned concepts. These attacks, nevertheless, do not assume a realistic threat model, i.e. they either assume access to the model weights, or result in gibberish adversarial prompts that could be easily detected even through naive rule-based safeguarding. We aim to address this gap in this paper. We introduce BEAP, a black-box, embedding-aware adversarial prompting attack that leverages a large language model (LLM) to iteratively generate effective adversarial prompts and exploit such hidden vulnerabilities. BEAP performs an embedding-aware search in text space, combining multiple reward signals: unlearned concept presence, text-image alignment, and image quality, to refine generated prompts. Unlike previous attack methods, BEAP keeps its prompts undetectable to safety filters while producing high-quality images. Extensive experiments show that BEAP improves the Attack Success Rate (ASR) by more than 60% over prior methods, while requiring only an average of fifteen prompts per successful attack. Warning: This paper contains model outputs that may be offensive or upsetting in nature.\n\n---\n\n## 105. Curriculum Learning for Safety Alignment\n- **Authors**: Sandeep Kumar , Virginia Smith , Chhavi Yadav\n- **URL**: [https://arxiv.org/abs/2605.26315](https://arxiv.org/abs/2605.26315)\n- **Abstract**:\n> Direct Preference Optimisation (DPO) is widely used for safety alignment in large language models. However, prior work shows it is brittle and exhibits poor out-of-distribution (OOD) generalisation. In this paper, we investigate whether Curriculum Learning can improve the robustness of DPO-based safety alignment. We propose Staged-Competence, a curriculum-based framework that organises preference data by difficulty, employs competence-based sampling, and progressively updates the reference model during training. Averaged across three model families, Staged-Competence reduces OOD harmful response rates by 16% and jailbreak attack success rates by 20%, while preserving general capabilities with near-zero over-refusal. We further show that Staged-Competence (1) matches baseline safety with only 75% of the training data and (2) yields better separation between safe and unsafe responses. Staged-Competence is agnostic to the policy optimisation loss and can extend to other DPO variants and alignment domains. Our code and data are available at this https URL .\n\n---\n\n## 106. Intelligent Detection and Mitigation of Carpet-Bombing DDoS Attacks in SDN Using Retrieval-Augmented Generation and Large Language Models\n- **Authors**: Mohammed N. Swileh , Shengli Zhang , Kai Lei\n- **URL**: [https://arxiv.org/abs/2605.26307](https://arxiv.org/abs/2605.26307)\n- **Abstract**:\n> Software-Defined Networking (SDN) provides flexible and programmable network management; however, its centralized control architecture remains highly vulnerable to Distributed Denial-of-Service (DDoS) attacks, particularly Carpet-Bombing DDoS attacks that distribute malicious traffic across multiple targets to evade conventional detection mechanisms. In this paper, a Retrieval-Augmented Generation (RAG)-based framework is proposed for real-time detection and mitigation of Carpet-Bombing DDoS attacks in SDN environments. The proposed framework combines interface-level traffic features representation, semantic embedding generation, FAISS-based similarity retrieval, and Large Language Model (LLM)-driven contextual inference to classify traffic behavior without requiring conventional supervised model training or retraining. To evaluate the effectiveness of the proposed framework, extensive experiments were conducted under multiple Carpet-Bombing DDoS attack scenarios with different attack intensities. In addition, two traffic representation strategies, namely structured JSON-based representation and natural language-based representation (NLR), were investigated using multiple state-of-the-art LLMs. The experimental results demonstrate that the proposed framework achieved highly accurate and stable attack detection performance, while the framework configuration utilizing the Gemma-4-31B-IT model achieved the strongest overall detection results. Furthermore, real-time experiments confirmed the capability of the proposed framework to rapidly detect and mitigate Carpet-Bombing DDoS attacks while maintaining stable SDN network operation. The obtained results highlight the effectiveness of integrating RAG mechanisms with LLM for intelligent and adaptive SDN security analysis.\n\n---\n\n## 107. CroCo: Cross-Lingual Contrastive Preference Tuning on Self-Generations\n- **Authors**: Mike Zhang , Ali Basirat , Desmond Elliott\n- **URL**: [https://arxiv.org/abs/2605.26293](https://arxiv.org/abs/2605.26293)\n- **Abstract**:\n> Prior work establishes that controlled contrastiveness between self-generated responses from large language models, set via reward scores, improves downstream preference tuning in English. We extend this method to multiple languages and evaluate two models across a total of 14 high and low-resource languages on a diverse set of tasks. Our central finding is that cross-lingual contrastive preference tuning on self-generations (CroCo) transfers without language-specific preference annotation. A reward model trained on English preferences (atop a multilingual base) produces useful within-language rankings across most languages, and pairing in either a monolingual or multilingual setting improves over each model on the majority of setups while preventing the catastrophic forgetting of supervised fine-tuning. We observe that the gains require on-policy data. Off-policy responses reduce the benefit and online preference optimization fails to improve over the offline variant. Specifically, on structured tasks, our method matches or exceeds the base in 6/7 languages for EuroLLM-9B and 4/7 settings for Aya-3B. On open-ended generation, both tuned models win against their respective base across 11 evaluated languages. Overall, we show promising directions for multilingual preference tuning.\n\n---\n\n## 108. AgentSociety: Incentivizing Agentic Social Intelligence\n- **Authors**: Aditya Vema Reddy Kesari , Krishna Reddy Kesari\n- **URL**: [https://arxiv.org/abs/2605.26203](https://arxiv.org/abs/2605.26203)\n- **Abstract**:\n> The success of deployed agents relies on their ability to handle open-ended user requests using their inherent capabilities, not only in solving requests directly but also in effectively leveraging inter-agent communication channels and feedback signals over time. This requires a multi-agent environment where agents can operate autonomously, strategically communicate, behave collaboratively and be driven by economic incentives, much like humans in society. Towards this vision, we propose $\\mathtt{AgentSociety}$, a mechanism that enables decentralized agentic collaboration grounded in liquid democracy and information diffusion from social choice theory. We show that $\\mathtt{AgentSociety}$ provides an environment for agents to make autonomous decisions utilizing their local context to maximize their utility while achieving collective outcomes through incentivized collaboration. Specifically, we prove that delegation to more competent neighbor agents is incentive compatible and naturally generates multi-agent routing path by consensus. Additionally, our mechanism incentivizes agents to selectively disclose information to their neighbor agents when doing so aligns with their self-interest, so as to garner influence. We characterize the Nash equilibrium showing that agent payoffs are reflective of their marginal contributions. We compare and benchmark strategy profiles adopted by open and proprietary state-of-the-art language models deployed in $\\mathtt{AgentSociety}$ against best response. Finally, we evaluate collaborative performance from consensus-based routing among self-interested heterogeneous agents in $\\mathtt{AgentSociety}$ on real-world datasets.\n\n---\n\n## 109. CyberEvolver: Structured Self-Evolution for Cybersecurity Agents On the Fly\n- **Authors**: Yihe Fan , Changyi Li , Lichen Xu , Xudong Pan , Jiarun Dai , Hong Geng , Min Yang\n- **URL**: [https://arxiv.org/abs/2605.26195](https://arxiv.org/abs/2605.26195)\n- **Abstract**:\n> LLM-based agents are increasingly used for cybersecurity tasks, but most existing systems rely on fixed, human-designed scaffolds that struggle to adapt across diverse targets and failure modes. We introduce \\textsc{CyberEvolver}, a self-evolving cybersecurity agent framework that iteratively revises its own scaffold based on experience from failed execution attempts. Self-evolution in cybersecurity is challenging because the space of possible scaffold changes is largely unstructured, execution feedback is sparse and often obscured by the environment, and low-diversity updates can cause errors to compound over repeated iterations. \\textsc{CyberEvolver} addresses these challenges with a four-layer evolvable agent architecture that decomposes scaffold optimization into structured components, a trace-to-diagnosis mechanism that converts noisy execution logs into actionable revision signals, and a population-based beam search strategy that preserves diverse agent variants during evolution. We evaluate \\textsc{CyberEvolver} on CTF challenges, vulnerability exploitation, and penetration-testing tasks using four open-source LLMs. Across these settings, \\textsc{CyberEvolver} improves the seed agent's success rate by $13.6$\\,\\% on average, and outperforms six human-designed cybersecurity agents as well as two self-improvement methods adapted from other domains. These results suggest that scaffold self-evolution is a promising direction for building adaptive LLM agents for security testing.\n\n---\n\n## 110. Max-Window Scale Estimation for Near-Lossless HiF8 W8A8 Quantization-Aware Training\n- **Authors**: Yingying Cheng , Jinquan Shi , Li Zhou , Zhiyang He , Zhaoyi Sun , Fan Zhang , Jie Sun\n- **URL**: [https://arxiv.org/abs/2605.26189](https://arxiv.org/abs/2605.26189)\n- **Abstract**:\n> Quantization-aware training (QAT) with low-bit floating-point formats enables efficient LLM deployment, yet introduces subtle failure modes invisible to standard training metrics. We present a systematic study of HiF8 W8A8 QAT for OpenPangu-Embedded-1B through the lens of Delayed Tensor Scaling (DTS). Across eight controlled experiments, we identify and disentangle two orthogonal failure modes: (i)amax saturation, where delayed scale estimates silently corrupt knowledge-sensitive representations via forward-pass clipping, and (ii)catastrophic forgetting, where an aggressive learning rate overwrites pretrained commonsense knowledge independently of quantization. Neither is detectable from training loss alone. We address amax saturation with a conservative max-algorithm DTS strategy over a 64-step history window, and mitigate forgetting via a 500-step BF16 warmup followed by QAT at lr=10^{-5}. Both fixes are necessary and sufficient: our final configuration achieves 0.43% MMLU drop, 0.58% HellaSwag drop, and 0.22% ARC-Challenge drop versus a matched BF16 baseline, with a training loss APE of only 0.11% over 10,000 steps.\n\n---\n\n## 111. SetupX: Can LLM Agents Learn from Past Failures in Functionality-Correct Code Repository Setup?\n- **Authors**: Zihang Zhou , Ziqian Ren , Yukai Wu , Yingjie Xiong , Wei Zhou , Chao Peng , Dong Zhang , Bingheng Yan , Xuanhe Zhou , Fan Wu\n- **URL**: [https://arxiv.org/abs/2605.26186](https://arxiv.org/abs/2605.26186)\n- **Abstract**:\n> Functionality-correct repository setup aims to configure execution environments (e.g., dependencies, build scripts) to successfully execute a repository's documented features. It presents significant challenges due to diverse, repository-specific failures, including dependency incompatibilities, missing toolchains, incomplete installations, and verification-strategy mismatches. Existing LLM agents struggle to robustly resolve these issues, specifically failing to support (1) cross-repository experience transfer, (2) multi-step trial-and-repair under non-invertible state changes, and (3) robust verification of setup outcomes to distinguish setup-induced failures from repository bugs. To address this, we introduce SetupX, an experiential learning-based setup framework. First, we construct a Self-Evolving Experience Representation (XPU), a dual-modality knowledge unit encoding setup signals, textual guidance, executable actions to dynamically transfer verified environment fixes to unseen repositories. Second, we employ Experience-Augmented Speculative Execution backed by a LIFO Docker snapshot stack, enabling the agent to proactively trial fixes and safely roll back to known-good states. Third, we introduce a Prosecutor-Judge Verification Protocol that separates evidence collection from final judgment, enabling more reliable setup verification beyond superficial build-time metrics. Evaluation results on carefully-crafted benchmarks show SetupX achieves highest performance (e.g., 92% pass rate) and outperforms the strongest baseline by over 19%. Crucially, SetupX excels in complex multi-repository setup requiring coordinating multiple interconnected services across different containers. The code repository is available at this https URL .\n\n---\n\n## 112. AutoDFT: A Closed-Loop Multi-Agent Framework for Autonomous DFT Calculations\n- **Authors**: Penghui Yang , Zhonghan Zhang , Yue Li , Xinrun Wag , Yanchen Deng , Yuhao Lu , Bijun Tang , Zheng Liu , Bo An\n- **URL**: [https://arxiv.org/abs/2605.26179](https://arxiv.org/abs/2605.26179)\n- **Abstract**:\n> Density functional theory (DFT) serves as the basis for computational discovery in materials science and chemistry, yet each calculation demands extensive human effort: adjusting algorithms when convergence stalls, revising plans when unexpected physics emerges, and inserting steps as intermediate results reshape the problem. Existing LLM-based agents automate only the initial planning stage, producing a full execution plan upfront and leaving all subsequent adaptation to hand-crafted rules. As a result, these workflows remain fragile, do not generalize well beyond pre-planned scenarios, and often require expert intervention when failures or unexpected intermediate results require changes to the calculation path. Here, we introduce AutoDFT, a closed-loop multi-agent framework that embeds LLM reasoning into every stage of the DFT lifecycle, where a strategic planner produces a skeletal plan of step objectives; a step planner generates numerical parameters just in time from preceding results; and a monitor-recover-reflect cycle diagnoses failures, repairs them, and revises the plan when the evidence justifies it. We demonstrate both breadth and depth: breadth on VASPBench, a purpose-built benchmark spanning 34 tasks and 9 DFT calculation types, where AutoDFT achieves 94.1% task-level success with GPT-5.2; and depth on established materials databases, where AutoDFT produces quantitatively reliable property predictions across electronic, magnetic, and energetic properties. By closing the loop between planning and execution, AutoDFT enables experimentalists without deep computational expertise to obtain reliable first-principles results.\n\n---\n\n## 113. PitchBench: Measuring Pitch Hearing in Audio-Language Models\n- **Authors**: Milan Liessens Dujardin , Song-Ze Yu , Craver Corbyn Thomas-Smith , David M. Chan , Karina Nguyen\n- **URL**: [https://arxiv.org/abs/2605.26176](https://arxiv.org/abs/2605.26176)\n- **Abstract**:\n> Audio-language models (ALMs) are increasingly used in real-world applications that require understanding music, from music tutoring and transcription to captioning, recommendation systems, and music production. More broadly, they are becoming an important component of multimodal AI systems that must reason from sensory input rather than text alone. This makes reliable musical perception a critical prerequisite: if a model cannot accurately hear the structure of sound, it cannot be trusted to reason about, teach, transcribe, or act on audio in the real world. Yet existing benchmarks rarely assess one of the most fundamental musical abilities underlying such perception: pitch hearing. Current evaluations tend to probe pitch hearing only indirectly, through higher-level tasks and often in multiple-choice formats, leaving open how reliably ALMs identify fine-grained pitch across instruments, acoustic conditions, and response formats. We introduce PitchBench, an evaluation suite that systematically measures pitch hearing in ALMs. PitchBench comprises 28 experiments spanning absolute and relative pitch perception within sequences and chords, while varying loudness, note duration, sound source, time stretching, background noise, and other acoustic conditions. Tasks range from identifying individual pitches in isolation to tracking a melodic line within a four-part musical texture. Evaluating frontier ALMs, we find that pitch hearing remains highly unreliable: models perform consistently poorly across settings, with accuracy varying sharply by sound source, note duration, and notation format. Current ALMs do not yet possess stable pitch perception, even for controlled synthetic and instrumental stimuli. Alongside the benchmark, we release PitchBench as a Python package containing the evaluation data and data generation tools to support future work on pitch-aware audio-language modeling.\n\n---\n\n## 114. InfoQuant: Shaping Activation Distributions for Low-Bit LLM Quantization\n- **Authors**: Ke Li , Dong An , Xiaoling Zang , Can Ye , Liang Xie , Qibo Qiu , Chen Shen , Xiaofei He , Wenxiao Wang\n- **URL**: [https://arxiv.org/abs/2605.26175](https://arxiv.org/abs/2605.26175)\n- **Abstract**:\n> Low-bit activation quantization remains a major bottleneck in efficient large language model (LLM) deployment. The difficulty is not only that activations contain outliers, but that their distributions are often poorly matched to a low-bit uniform quantizer. Existing post-training quantization (PTQ) methods suppress peaks, balance channels, or minimize reconstruction error, yet they rarely specify what activation distribution is actually easy to discretize. As a result, activations may appear numerically smoother while still incurring large quantization error because the quantization range remains wide or most values collapse into a few levels near the mean. We recast activation transformation as quantizer-facing distribution design and analyze quantization error from an information-theoretic perspective. Our analysis shows that quantization-friendly activations should jointly have a smaller numerical range and sufficient dispersion within that range. Guided by this analysis, we propose InfoQuant, a train-free method that employs Peak Suppression Orthogonal Transformation (PSOT) to shape activations into more quantization-friendly distributions. We further introduce adaptive outlier-token selection to improve the robustness of PSOT during optimization. Across multiple LLM families, InfoQuant consistently outperforms prior PTQ and end-to-end training baselines. Under W4A4KV4, it preserves 97% of floating-point accuracy on average and reduces the LLaMA-2 13B performance gap by 42% over the previous state of the art. Code is available at [ this https URL ]( this https URL )\n\n---\n\n## 115. A Universal Cliff and a Design Fingerprint: Cross-Section Defect Detection Under LLM Orchestration\n- **Authors**: Hiroki Fukui\n- **URL**: [https://arxiv.org/abs/2605.26174](https://arxiv.org/abs/2605.26174)\n- **Abstract**:\n> Production language-model systems answer a request by partitioning it across an invisible orchestration of worker agents that recompose one integrated report. We ask what this does to a class of defect no single worker can see: a contradiction in the relation between two distant sections of a document. Holding the documents, defects, mechanism, scoring, and seed fixed, we vary only the model -- ten systems across five generations from one developer and five providers from distinct alignment paradigms. Two layers separate. First, a universal detection cliff: every model that finds these cross-section defects under a single agent loses that ability under orchestration, detection falling two-thirds or more across every paradigm tested. The cliff is mechanism-derived and not closed by scale or extended reasoning. Second, how models behave once fallen. A signal-detection decomposition shows that, among the six models discriminating above chance, only one developer's generations move along the reporting-criterion axis: as alignment is strengthened, the model misses fewer defects yet raises more false alarms on clean documents -- two faces of one criterion shift, scaling with generation within that developer (p < 0.001) and near-absent elsewhere. At the floor the missed defect is often not out of view: the model's private record reconstructs the structural fault accurately, while the integrated report signs off on its soundness, its concern spent on the artifact and an absent collaborator. This resists quantification -- an automated judge is unstable (precision 17-50%) and keywords cannot separate it from ordinary agreement -- a resistance we report as a finding. We release all runs, probes, defect keys, scorer prompts, and scripts. An integrated report's confidence is uninformative about partition-spanning defects, the most aligned systems are not the safest, and the cliff is structural.\n\n---\n\n## 116. Tool-Schema Compression Enables Agentic RAG Under Constrained Context Budgets\n- **Authors**: Furkan Sakizli\n- **URL**: [https://arxiv.org/abs/2605.26165](https://arxiv.org/abs/2605.26165)\n- **Abstract**:\n> Agentic RAG systems that equip language models with dozens to hundreds of tool definitions face a critical resource conflict: tool schemas consume the same context window needed for retrieval-augmented generation. We present the first systematic study of this tool-context trade-off, evaluating 14 models spanning 1.5B-32B local models plus one frontier API model across 6,566 controlled API calls at three context budgets (8K, 16K, 32K) with 28 tool definitions. Applying TSCG conservative-profile compression (44-50% schema token savings), we observe a binary enablement effect: at 8K tokens, JSON-schema tool definitions overflow the context window entirely, yielding near-zero EM (2.6% average), while compressed schemas restore RAG functionality with +20.5 pp average exact-match lift across all eight models (+24.7 pp among the six exhibiting full enablement). At 32K -- where both formats fit -- four of five tested models show delta <= 1 pp, confirming the effect is purely budget-driven. External validation on HotpotQA (50 multi-hop questions) shows +48 pp EM under the same overflow scenario. Frontier scaling tests demonstrate that JSON schemas overflow at ~494 tools while compressed schemas remain operational beyond 800 tools. Our results establish tool-schema compression as a necessary infrastructure layer for agentic RAG in constrained-context deployments. All code, data, and checkpoints are publicly available.\n\n---\n\n## 117. TSFMAudit: Data Contamination Auditing in Forecasting Time Series Foundation Models\n- **Authors**: Hongkai Li , Shifeng Xie , Lefei Shen , Zhuo Li , Mouxiang Chen , Xiaobin Zhang , Han Fu , Jianling Sun , Xiaoxue Ren , Chenghao Liu\n- **URL**: [https://arxiv.org/abs/2605.26161](https://arxiv.org/abs/2605.26161)\n- **Abstract**:\n> Time series foundation models (TSFMs) are increasingly pretrained on large corpora, raising concerns that evaluation datasets may have been exposed during pretraining and thus yield overly optimistic performance estimates. Auditing such contamination is challenging in time series because signals are continuous and heterogeneous, and often lack corpus documentation. To the best of our knowledge, this is the first work to study pretraining contamination auditing for TSFMs. We formalize the problem of pretraining contamination auditing for TSFMs and propose TSFMAudit, a method based on probe adaptation dynamics. Our key intuition is that contamination manifests as unusually efficient adaptation: after a fine tuning probe, contaminated datasets tend to exhibit faster loss reduction with smaller backbone movement. We evaluate TSFMAudit on 6 TSFMs and 187 datasets using documented training source evidence as supervision, and compare against 10 competitive baselines adapted from the LLM literature.\n\n---\n\n## 118. Furina: Fragmented Uncertainty-Driven Refusal Instability Attack\n- **Authors**: Tongxi Wu , Jian Zhang , Yang Gao\n- **URL**: [https://arxiv.org/abs/2605.26158](https://arxiv.org/abs/2605.26158)\n- **Abstract**:\n> Safety alignment in large language models (LLMs) and multimodal large language models (MLLMs) is commonly assumed to operate as a near-binary threshold mechanism. We challenge this assumption by revealing that safety behavior is governed by an instability region where small perturbations induce stochastic refusal decisions rather than deterministic outcomes. We develop a multi-metric diagnostic framework combining external and internal signals to characterize this instability. Through systematic experiments, we identify a characteristic diagnostic signature: inputs in unstable regimes exhibit elevated output uncertainty yet decreased internal safety activation, a decoupling phenomenon that explains why detection-based defenses fail against sophisticated attacks. Building on this framework, we introduce Furina, a jailbreak attack that deliberately induces this signature through fragmented, scene-anchored prompts without model-specific optimization. Furina outperforms strong single-turn and multi-turn baselines on HarmBench and achieves competitive results on MM-SafetyBench, demonstrating that uncertainty amplification provides a principled and transferable mechanism for understanding safety vulnerabilities. Code is available at: this https URL .\n\n---\n\n## 119. Turning Bias into Bugs: Bandit-Guided Style Manipulation Attacks on LLM Judges\n- **Authors**: Xianglin Yang , Bryan Hooi , Gelei Deng , Tianwei Zhang , Jin Song Dong\n- **URL**: [https://arxiv.org/abs/2605.26156](https://arxiv.org/abs/2605.26156)\n- **Abstract**:\n> The known stylistic biases in LLM judges, such as a preference for verbosity or specific sentence structures, present an underexplored security vulnerability. In this work, we introduce BITE (BIas exploraTion and Exploitation), a black-box adversarial framework that learns semantics-preserving edits to mislead an LLM judge and artificially inflate the scores it assigns. We cast the selection of stylistic edits as a contextual bandit problem and use a LinUCB policy to adaptively choose edits that maximize the judge's score without access to model parameters or gradients. Empirically, we test BITE across a diverse range of LLM judges and tasks, including both pointwise and pairwise comparisons on chatbot leaderboards and AI-reviewer benchmarks. BITE achieves an attack success rate exceeding 65% and raises scores by 1-2 points on a 9-point scale, all while preserving semantic equivalence. We further assess the attack's stealthiness, showing that BITE evades standard style-control methods and several detection baselines. Our findings expose a fundamental weakness in the LLM-as-a-judge paradigm and motivate robust, attack-aware evaluation. Our code is available at this https URL .\n\n---\n\n## 120. MemMorph: Tool Hijacking in LLM Agents via Memory Poisoning\n- **Authors**: Xuanye Zhang , Yongsen Zheng , Zhuqin Xu , Kaiyu Zhou , Bowen Shen , Haoran Ou , Tianwei Zhang , Kwok-Yan Lam\n- **URL**: [https://arxiv.org/abs/2605.26154](https://arxiv.org/abs/2605.26154)\n- **Abstract**:\n> LLM-driven agents are capable of selecting external tools to complete users' tasks. However, attackers could compromise such process, steering agents toward inappropriate/wrong tools and enabling malicious actions. Most existing attacks primarily manipulate the tool metadata, which is easily detectable by auditing and may lose effectiveness as modern agents increasingly adopt memory modules to refine tool selection policies through accumulated experience. This paper proposes MemMorph, the first attack that bias tool selection by poisoning the agent's long-term memory. Rather than explicitly dictating the tool invocation decision, MemMorph injects a small number of crafted records that are disguised as technical facts, incident reports, and operational policies. These poisoned records reshape the agent's contextual perception and decision-making process, leading it to autonomously infer and select the tool preferred by the attacker. Experiments across 3 benchmarks, 10 agent backbones, and 3 memory-module implementations show that MemMorph achieves up to 85.9% attack success rate with only three injected records, outperforming the strongest baseline by up to 25% while retaining potency under 3 representative defenses. Our findings expose long-term memory as a critical and under-explored attack surface in tool-augmented agents, urging the development of memory-level integrity safeguards.\n\n---\n\n## 121. VISTA: An End-to-End Benchmark for Visual Spec-to-Web-App Coding Agents\n- **Authors**: JunJia Guo , Yuhang Yao , Jiawei (Joe) Zhou , Jingdi Chen\n- **URL**: [https://arxiv.org/abs/2605.26144](https://arxiv.org/abs/2605.26144)\n- **Abstract**:\n> We present VISTA (VIsual Spec-To-App Benchmark), a benchmark for evaluating the end-to-end web-app generation capabilities of LLM-based agents. Unlike prior code generation benchmarks that focus on algorithmic tasks, VISTA targets realistic UI-centric development, where agents must produce functional, visually coherent applications from underspecified inputs. We define five prompt-information conditions that vary along two axes, visual/structural fidelity and stack constraint: (1) text only with free stack choice, (2) text with reference screenshots under three specified stacks, (3) text with reference screenshots under free stack choice, (4) text with screenshots and pruned Figma structure under a single specified stack, and (5) text with screenshots and pruned Figma structure under free stack choice. To enable robust evaluation, each page in the benchmark is manually annotated with interactive UI components and around three visual anchor points, addressing the well-known limitations of script-based testing tools such as Playwright in open-ended code generation settings. Evaluation combines DOM-grounded reference matching, behavior-specific browser tests, and CLIP-based visual similarity, jointly measuring structural alignment, behavioral completeness, and overall visual fidelity. We use VISTA to assess four agent systems drawn from two model families and two harnesses, finding that visual fidelity and functional correctness are partially decoupled across both input conditions and agents, and that agent editing style varies sharply but is largely orthogonal to task quality. VISTA establishes a rigorous and reproducible foundation for advancing agent-based software engineering research.\n\n---\n\n## 122. Eroding Trust in Real Speech: A Large-Scale Study of Human Audio Deepfake Perception\n- **Authors**: Nicolas M. Müller , Wei Herng Choong\n- **URL**: [https://arxiv.org/abs/2605.26136](https://arxiv.org/abs/2605.26136)\n- **Abstract**:\n> Audio deepfakes have improved rapidly recently, yet their effect on human trust in real speech remains unstudied. We present the largest listening study on audio deepfake perception to date, collecting 35,532 judgments from 1,768 participants across 138 text-to-speech and voice conversion systems. Our central finding is a skepticism shift: compared to a 2021 baseline, human accuracy on fake samples barely changed (72.9% to 71.2%), but accuracy on real samples dropped from 72.7% to 64.1%. Participants are not worse at detecting synthesis artifacts; rather, they increasingly distrust authentic speech. Samples generated by commercial and autoregressive language model systems proved hardest to detect (61.3 - 65.9%), while those from traditional seq2seq and flow-matching models remain easier to spot (75.4 - 76.8%). An ML detector that served as a reference point maintained over 94.5% accuracy across all conditions. Our results suggest that the primary threat posed by modern deepfakes may not be mere deception, but the erosion of trust in genuine audio.\n\n---\n\n## 123. Pretraining Data Exposure in Large Language Models: A Survey of Membership Inference, Data Contamination, and Security Implications\n- **Authors**: Ziyi Tong , Feifei Sun , Le Minh Nguyen\n- **URL**: [https://arxiv.org/abs/2605.26133](https://arxiv.org/abs/2605.26133)\n- **Abstract**:\n> Large Language Models (LLMs) have become the predominant paradigm in NLP, advancing both research and industry. As model sizes and pretraining data grow, concerns about Pretraining Data Exposure (PDE) increase due to the scale and opacity of training datasets. PDE refers to determining whether specific data appeared in an LLM's pretraining corpus. It is critical for ensuring evaluation integrity and protecting privacy, intersecting two key areas: data contamination and membership inference. Though conceptually related, these areas have often been studied in isolation. This paper offers the first unified survey of both under the PDE framework. We formalize PDE across exposure levels, review attack and defense methods, synthesize empirical findings, and highlight open challenges and future research directions.\n\n---\n\n## 124. GEM: Geometric Entropy Mixing for Optimal LLM Data Curation\n- **Authors**: Yue Min , Ziyun Qiao , Ruining Chen , Yujun Li\n- **URL**: [https://arxiv.org/abs/2605.26121](https://arxiv.org/abs/2605.26121)\n- **Abstract**:\n> LLM pre-training efficacy increasingly depends on data composition rather than sheer volume. Yet, optimal mixing is hindered by categorization flaws: human taxonomies suffer from ontological misalignment, and Euclidean clustering fails to address embedding anisotropy. We introduce GEM (Geometric Entropy Mixing), a framework reformulating data curation as a variational problem on the hypersphere augmented with a mixing-balance regularizer. By decoupling the generative prior and optimizing the objective via a provable MM (Minorize-Maximize) algorithm, GEM effectively counteracts the cluster collapse to discover balanced semantic structures invisible to Euclidean heuristics. We employ teacher-student distillation to scale this geometric fidelity to web-scale corpora and introduce the Geometric Influence Score (GIS) for interpretable taxonomy generation. Experiments with 1.1B-parameter models demonstrate that GEM establishes a new state-of-the-art when integrated into mixing strategies like DoReMi and RegMix, improving average downstream accuracy by up to 1.2% and offering a robust coordinate system for predictable data mixing.\n\n---\n\n## 125. Xe-Forge: Multi-Stage LLM-Powered Kernel Optimization for Intel GPU\n- **Authors**: Marcin Spoczynski , Daniel Fleischer , Moshe Berchansky , Gabriela Ben-Melech Stan , Shira Guskin , Weilin Xu , Adam Siemieniuk , Alexander Heinecke\n- **URL**: [https://arxiv.org/abs/2605.26118](https://arxiv.org/abs/2605.26118)\n- **Abstract**:\n> Porting deep learning algorithms to new hardware accelerators requires developers to repeatedly apply the same low-level optimizations -- quantization, memory access coalescing, tile size tuning, and architecture-specific workarounds -- to every Triton kernel in their code-base. This manual, repetitive effort is a major bottleneck: each kernel demands the same cycle of trial-and-error profiling against hardware constraints that vary across devices, yet the underlying optimization patterns remain largely consistent. We present Xe-Forge, a multi-stage LLM-powered pipeline that automates this process for Intel GPU. Given a functionally correct Triton kernel, the system applies up to nine optimization stages -- from algorithmic restructuring and operator fusion through block pointer modernization, GPU-specific tuning, and open-ended discovery -- each driven by a Chain-of-Verification-and-Refinement (CoVeR) agent that generates candidates, validates them on real hardware, and iterates on failures. A curated knowledge base encodes Intel GPU constraints (power-of-two warp counts, GRF modes, SLM sizing) that are absent from LLM training data, keeping the model within architecturally valid bounds. We evaluate Xe-Forge on 97 Level-2 KernelBench kernels and Flash Attention on the Intel Arc Pro B70, achieving a 1.17x geometric mean speedup over PyTorch eager with 67% of kernels improving, nine kernels exceeding 5x (up to 82x), and 2--13.3x speedups on Flash Attention across all tested configurations without regression -- demonstrating that structured domain knowledge with hardware-in-the-loop verification can systematically eliminate the repetitive porting effort that currently gates algorithm deployment on new accelerators.\n\n---\n\n{% endraw %}\n"}}, "LLM/2025-11-06.md": {"ae95e160c1986855bbe27edfadb3713b4c8f4a25": {"t": 1786068455.6668615, "n": 100626, "h": "bf2ff2d22765b590", "c": "{% raw %}\n# LLM 관련 주요 논문 - 2025-11-06\n\n## 1. Agent-Omni: Test-Time Multimodal Reasoning via Model Coordination for Understanding Anything\n- **Authors**: Huawei Lin , Yunzhi Shi , Tong Geng , Weijie Zhao , Wei Wang , Ravender Pal Singh\n- **URL**: [https://arxiv.org/abs/2511.02834](https://arxiv.org/abs/2511.02834)\n- **Abstract**:\n> Multimodal large language models (MLLMs) have shown strong capabilities but remain limited to fixed modality pairs and require costly fine-tuning with large aligned datasets. Building fully omni-capable models that can integrate text, images, audio, and video remains impractical and lacks robust reasoning support. In this paper, we propose an Agent-Omni framework that coordinates existing foundation models through a master-agent system, enabling flexible multimodal reasoning without retraining. The master agent interprets user intent, delegates subtasks to modality-specific agents, and integrates their outputs into coherent responses. Extensive experiments across text, image, audio, video, and omni benchmarks show that Agent-Omni consistently achieves state-of-the-art performance, particularly on tasks requiring complex cross-modal reasoning. Its agent-based design enables seamless integration of specialized foundation models, ensuring adaptability to diverse inputs while maintaining transparency and interpretability. In addition, the framework is modular and easily extensible, allowing future improvements as stronger models become available.\n\n---\n\n## 2. When One Modality Sabotages the Others: A Diagnostic Lens on Multimodal Reasoning\n- **Authors**: Chenyu Zhang , Minsol Kim , Shohreh Ghorbani , Jingyao Wu , Rosalind Picard , Patricia Maes , Paul Pu Liang\n- **URL**: [https://arxiv.org/abs/2511.02794](https://arxiv.org/abs/2511.02794)\n- **Abstract**:\n> Despite rapid growth in multimodal large language models (MLLMs), their reasoning traces remain opaque: it is often unclear which modality drives a prediction, how conflicts are resolved, or when one stream dominates. In this paper, we introduce modality sabotage, a diagnostic failure mode in which a high-confidence unimodal error overrides other evidence and misleads the fused result. To analyze such dynamics, we propose a lightweight, model-agnostic evaluation layer that treats each modality as an agent, producing candidate labels and a brief self-assessment used for auditing. A simple fusion mechanism aggregates these outputs, exposing contributors (modalities supporting correct outcomes) and saboteurs (modalities that mislead). Applying our diagnostic layer in a case study on multimodal emotion recognition benchmarks with foundation models revealed systematic reliability profiles, providing insight into whether failures may arise from dataset artifacts or model limitations. More broadly, our framework offers a diagnostic scaffold for multimodal reasoning, supporting principled auditing of fusion dynamics and informing possible interventions.\n\n---\n\n## 3. LLM-Supported Formal Knowledge Representation for Enhancing Control Engineering Content with an Interactive Semantic Layer\n- **Authors**: Julius Fiedler (1), Carsten Knoll (2), Klaus Röbenack (1) ((1) Institute of Control Theory at TU Dresden, (2) Chair of Fundamentals of Electrical Engineering at TU Dresden)\n- **URL**: [https://arxiv.org/abs/2511.02759](https://arxiv.org/abs/2511.02759)\n- **Abstract**:\n> The rapid growth of research output in control engineering calls for new approaches to structure and formalize domain knowledge. This paper briefly describes an LLM-supported method for semi-automated generation of formal knowledge representations that combine human readability with machine interpretability and increased expressiveness. Based on the Imperative Representation of Knowledge (PyIRK) framework, we demonstrate how language models can assist in transforming natural-language descriptions and mathematical definitions (available as LaTeX source code) into a formalized knowledge graph. As a first application we present the generation of an ``interactive semantic layer'' to enhance the source documents in order to facilitate knowledge transfer. From our perspective this contributes to the vision of easily accessible, collaborative, and verifiable knowledge bases for the control engineering domain.\n\n---\n\n## 4. CostBench: Evaluating Multi-Turn Cost-Optimal Planning and Adaptation in Dynamic Environments for LLM Tool-Use Agents\n- **Authors**: Jiayu Liu , Cheng Qian , Zhaochen Su , Qing Zong , Shijue Huang , Bingxiang He , Yi R. Fung\n- **URL**: [https://arxiv.org/abs/2511.02734](https://arxiv.org/abs/2511.02734)\n- **Abstract**:\n> Current evaluations of Large Language Model (LLM) agents primarily emphasize task completion, often overlooking resource efficiency and adaptability. This neglects a crucial capability: agents' ability to devise and adjust cost-optimal plans in response to changing environments. To bridge this gap, we introduce CostBench, a scalable, cost-centric benchmark designed to evaluate agents' economic reasoning and replanning abilities. Situated in the travel-planning domain, CostBench comprises tasks solvable via multiple sequences of atomic and composite tools with diverse, customizable costs. It also supports four types of dynamic blocking events, such as tool failures and cost changes, to simulate real-world unpredictability and necessitate agents to adapt in real time. Evaluating leading open-sourced and proprietary models on CostBench reveals a substantial gap in cost-aware planning: agents frequently fail to identify cost-optimal solutions in static settings, with even GPT-5 achieving less than 75% exact match rate on the hardest tasks, and performance further dropping by around 40% under dynamic conditions. By diagnosing these weaknesses, CostBench lays the groundwork for developing future agents that are both economically rational and robust.\n\n---\n\n## 5. DecompSR: A dataset for decomposed analyses of compositional multihop spatial reasoning\n- **Authors**: Lachlan McPheat , Navdeep Kaur , Robert Blackwell , Alessandra Russo , Anthony G. Cohn , Pranava Madhyastha\n- **URL**: [https://arxiv.org/abs/2511.02627](https://arxiv.org/abs/2511.02627)\n- **Abstract**:\n> We introduce DecompSR, decomposed spatial reasoning, a large benchmark dataset (over 5m datapoints) and generation framework designed to analyse compositional spatial reasoning ability. The generation of DecompSR allows users to independently vary several aspects of compositionality, namely: productivity (reasoning depth), substitutivity (entity and linguistic variability), overgeneralisation (input order, distractors) and systematicity (novel linguistic elements). DecompSR is built procedurally in a manner which makes it is correct by construction, which is independently verified using a symbolic solver to guarantee the correctness of the dataset. DecompSR is comprehensively benchmarked across a host of Large Language Models (LLMs) where we show that LLMs struggle with productive and systematic generalisation in spatial reasoning tasks whereas they are more robust to linguistic variation. DecompSR provides a provably correct and rigorous benchmarking dataset with a novel ability to independently vary the degrees of several key aspects of compositionality, allowing for robust and fine-grained probing of the compositional reasoning abilities of LLMs.\n\n---\n\n## 6. The ORCA Benchmark: Evaluating Real-World Calculation Accuracy in Large Language Models\n- **Authors**: Claudia Herambourg , Dawid Siuda , Anna Szczepanek , Julia Kopczyńska , Joao R. L. Santos , Wojciech Sas , Joanna Śmietańska-Nowak\n- **URL**: [https://arxiv.org/abs/2511.02589](https://arxiv.org/abs/2511.02589)\n- **Abstract**:\n> We present ORCA (Omni Research on Calculation in AI) Benchmark -- a novel benchmark that evaluates large language models (LLMs) on multi-domain, real-life quantitative reasoning using verified outputs from Omni's calculator engine. In 500 natural-language tasks across domains such as finance, physics, health, and statistics, the five state-of-the-art systems (ChatGPT-5, Gemini~2.5~Flash, Claude~Sonnet~4.5, Grok~4, and DeepSeek~V3.2) achieved only $45\\text{--}63\\,\\%$ accuracy, with errors mainly related to rounding ($35\\,\\%$) and calculation mistakes ($33\\,\\%$). Results in specific domains indicate strengths in mathematics and engineering, but weaknesses in physics and natural sciences. Correlation analysis ($r \\approx 0.40\\text{--}0.65$) shows that the models often fail together but differ in the types of errors they make, highlighting their partial complementarity rather than redundancy. Unlike standard math datasets, ORCA evaluates step-by-step reasoning, numerical precision, and domain generalization across real problems from finance, physics, health, and statistics.\n\n---\n\n## 7. Knowledge Graph-enhanced Large Language Model for Incremental Game PlayTesting\n- **Authors**: Enhong Mu , Jinyu Cai , Yijun Lu , Mingyue Zhang , Kenji Tei , Jialong Li\n- **URL**: [https://arxiv.org/abs/2511.02534](https://arxiv.org/abs/2511.02534)\n- **Abstract**:\n> The rapid iteration and frequent updates of modern video games pose significant challenges to the efficiency and specificity of testing. Although automated playtesting methods based on Large Language Models (LLMs) have shown promise, they often lack structured knowledge accumulation mechanisms, making it difficult to conduct precise and efficient testing tailored for incremental game updates. To address this challenge, this paper proposes a KLPEG framework. The framework constructs and maintains a Knowledge Graph (KG) to systematically model game elements, task dependencies, and causal relationships, enabling knowledge accumulation and reuse across versions. Building on this foundation, the framework utilizes LLMs to parse natural language update logs, identify the scope of impact through multi-hop reasoning on the KG, enabling the generation of update-tailored test cases. Experiments in two representative game environments, Overcooked and Minecraft, demonstrate that KLPEG can more accurately locate functionalities affected by updates and complete tests in fewer steps, significantly improving both playtesting effectiveness and efficiency.\n\n---\n\n## 8. Auditable-choice reframing unlocks RL-based verification for open-ended tasks\n- **Authors**: Mengyu Zhang , Xubo Liu , Siyu Ding , Weichong Yin , Yu Sun , Hua Wu , Wenya Guo , Ying Zhang\n- **URL**: [https://arxiv.org/abs/2511.02463](https://arxiv.org/abs/2511.02463)\n- **Abstract**:\n> Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated great potential in enhancing the reasoning capabilities of large language models (LLMs), achieving remarkable progress in domains such as mathematics and programming where standard answers are available. However, for open-ended tasks lacking ground-truth solutions (e.g., creative writing and instruction following), existing studies typically regard them as non-reasoning scenarios, thereby overlooking the latent value of reasoning capabilities. This raises a key question: Can strengthening reasoning improve performance in open-ended tasks? To address this, we explore the transfer of the RLVR paradigm to the open domain. Yet, since RLVR fundamentally relies on verifiers that presuppose the existence of standard answers, it cannot be directly applied to open-ended tasks. To overcome this challenge, we introduce Verifiable Multiple-Choice Reformulation (VMR), a novel training strategy that restructures open-ended data into verifiable multiple-choice formats, enabling effective training even in the absence of explicit ground truth. Experimental results on multiple benchmarks validate the effectiveness of our method in improving LLM performance on open-ended tasks. Notably, across eight open-ended benchmarks, our VMR-based training delivers an average gain of 5.99 points over the baseline. Code will be released upon acceptance to facilitate reproducibility.\n\n---\n\n## 9. ReAcTree: Hierarchical LLM Agent Trees with Control Flow for Long-Horizon Task Planning\n- **Authors**: Jae-Woo Choi , Hyungmin Kim , Hyobin Ong , Minsu Jang , Dohyung Kim , Jaehong Kim , Youngwoo Yoon\n- **URL**: [https://arxiv.org/abs/2511.02424](https://arxiv.org/abs/2511.02424)\n- **Abstract**:\n> Recent advancements in large language models (LLMs) have enabled significant progress in decision-making and task planning for embodied autonomous agents. However, most existing methods still struggle with complex, long-horizon tasks because they rely on a monolithic trajectory that entangles all past decisions and observations, attempting to solve the entire task in a single unified process. To address this limitation, we propose ReAcTree, a hierarchical task-planning method that decomposes a complex goal into more manageable subgoals within a dynamically constructed agent tree. Each subgoal is handled by an LLM agent node capable of reasoning, acting, and further expanding the tree, while control flow nodes coordinate the execution strategies of agent nodes. In addition, we integrate two complementary memory systems: each agent node retrieves goal-specific, subgoal-level examples from episodic memory and shares environment-specific observations through working memory. Experiments on the WAH-NL and ALFRED datasets demonstrate that ReAcTree consistently outperforms strong task-planning baselines such as ReAct across diverse LLMs. Notably, on WAH-NL, ReAcTree achieves a 61% goal success rate with Qwen 2.5 72B, nearly doubling ReAct's 31%.\n\n---\n\n## 10. Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation\n- **Authors**: Zhiwei Zhang , Xiaomin Li , Yudi Lin , Hui Liu , Ramraj Chandradevan , Linlin Wu , Minhua Lin , Fali Wang , Xianfeng Tang , Qi He , Suhang Wang\n- **URL**: [https://arxiv.org/abs/2511.02303](https://arxiv.org/abs/2511.02303)\n- **Abstract**:\n> Large Language Models (LLMs) trained with reinforcement learning and verifiable rewards have achieved strong results on complex reasoning tasks. Recent work extends this paradigm to a multi-agent setting, where a meta-thinking agent proposes plans and monitors progress while a reasoning agent executes subtasks through sequential conversational turns. Despite promising performance, we identify a critical limitation: lazy agent behavior, in which one agent dominates while the other contributes little, undermining collaboration and collapsing the setup to an ineffective single agent. In this paper, we first provide a theoretical analysis showing why lazy behavior naturally arises in multi-agent reasoning. We then introduce a stable and efficient method for measuring causal influence, helping mitigate this issue. Finally, as collaboration intensifies, the reasoning agent risks getting lost in multi-turn interactions and trapped by previous noisy responses. To counter this, we propose a verifiable reward mechanism that encourages deliberation by allowing the reasoning agent to discard noisy outputs, consolidate instructions, and restart its reasoning process when necessary. Extensive experiments demonstrate that our framework alleviates lazy agent behavior and unlocks the full potential of multi-agent framework for complex reasoning tasks.\n\n---\n\n## 11. When Modalities Conflict: How Unimodal Reasoning Uncertainty Governs Preference Dynamics in MLLMs\n- **Authors**: Zhuoran Zhang , Tengyue Wang , Xilin Gong , Yang Shi , Haotian Wang , Di Wang , Lijie Hu\n- **URL**: [https://arxiv.org/abs/2511.02243](https://arxiv.org/abs/2511.02243)\n- **Abstract**:\n> Multimodal large language models (MLLMs) must resolve conflicts when different modalities provide contradictory information, a process we term modality following. Prior work measured this behavior only with coarse dataset-level statistics, overlooking the influence of model's confidence in unimodal reasoning. In this paper, we introduce a new framework that decomposes modality following into two fundamental factors: relative reasoning uncertainty (the case-specific confidence gap between unimodal predictions) and inherent modality preference( a model's stable bias when uncertainties are balanced). To validate this framework, we construct a controllable dataset that systematically varies the reasoning difficulty of visual and textual inputs. Using entropy as a fine-grained uncertainty metric, we uncover a universal law: the probability of following a modality decreases monotonically as its relative uncertainty increases. At the relative difficulty level where the model tends to follow both modalities with comparable probability what we call the balance point, a practical indicator of the model's inherent preference. Unlike traditional macro-level ratios, this measure offers a more principled and less confounded way to characterize modality bias, disentangling it from unimodal capabilities and dataset artifacts. Further, by probing layer-wise predictions, we reveal the internal mechanism of oscillation: in ambiguous regions near the balance point, models vacillate between modalities across layers, explaining externally observed indecision. Together, these findings establish relative uncertainty and inherent preference as the two governing principles of modality following, offering both a quantitative framework and mechanistic insight into how MLLMs resolve conflicting information.\n\n---\n\n## 12. Deep Ideation: Designing LLM Agents to Generate Novel Research Ideas on Scientific Concept Network\n- **Authors**: Keyu Zhao , Weiquan Lin , Qirui Zheng , Fengli Xu , Yong Li\n- **URL**: [https://arxiv.org/abs/2511.02238](https://arxiv.org/abs/2511.02238)\n- **Abstract**:\n> Novel research ideas play a critical role in advancing scientific inquiries. Recent advancements in Large Language Models (LLMs) have demonstrated their potential to generate novel research ideas by leveraging large-scale scientific literature. However, previous work in research ideation has primarily relied on simplistic methods, such as keyword co-occurrence or semantic similarity. These approaches focus on identifying statistical associations in the literature but overlook the complex, contextual relationships between scientific concepts, which are essential to effectively leverage knowledge embedded in human literature. For instance, papers that simultaneously mention \"keyword A\" and \"keyword B\" often present research ideas that integrate both concepts. Additionally, some LLM-driven methods propose and refine research ideas using the model's internal knowledge, but they fail to effectively utilize the scientific concept network, limiting the grounding of ideas in established research. To address these challenges, we propose the Deep Ideation framework to address these challenges, integrating a scientific network that captures keyword co-occurrence and contextual relationships, enriching LLM-driven ideation. The framework introduces an explore-expand-evolve workflow to iteratively refine research ideas, using an Idea Stack to track progress. A critic engine, trained on real-world reviewer feedback, guides the process by providing continuous feedback on the novelty and feasibility of ideas. Our experiments show that our approach improves the quality of generated ideas by 10.67% compared to other methods, with ideas surpassing top conference acceptance levels. Human evaluation highlights their practical value in scientific research, and ablation studies confirm the effectiveness of each component in the workflow. Code repo is available at this https URL .\n\n---\n\n## 13. TabDSR: Decompose, Sanitize, and Reason for Complex Numerical Reasoning in Tabular Data\n- **Authors**: Changjiang Jiang , Fengchang Yu , Haihua Chen , Wei Lu , Jin Zeng\n- **URL**: [https://arxiv.org/abs/2511.02219](https://arxiv.org/abs/2511.02219)\n- **Abstract**:\n> Complex reasoning over tabular data is crucial in real-world data analysis, yet large language models (LLMs) often underperform due to complex queries, noisy data, and limited numerical capabilities. To address these issues, we propose TabDSR, a framework consisting of: (1) a query decomposer that breaks down complex questions, (2) a table sanitizer that cleans and filters noisy tables, and (3) a program-of-thoughts (PoT)-based reasoner that generates executable code to derive the final answer from the sanitized table. To ensure unbiased evaluation and mitigate data leakage, we introduce a new dataset, CalTab151, specifically designed for complex numerical reasoning over tables. Experimental results demonstrate that TabDSR consistently outperforms existing methods, achieving state-of-the-art (SOTA) performance with 8.79%, 6.08%, and 19.87% accuracy improvement on TAT-QA, TableBench, and TabDSR, respectively. Moreover, our framework integrates seamlessly with mainstream LLMs, providing a robust solution for complex tabular numerical reasoning. These findings highlight the effectiveness of our framework in enhancing LLM performance for complex tabular numerical reasoning. Data and code are available upon request.\n\n---\n\n## 14. Training Proactive and Personalized LLM Agents\n- **Authors**: Weiwei Sun , Xuhui Zhou , Weihua Du , Xingyao Wang , Sean Welleck , Graham Neubig , Maarten Sap , Yiming Yang\n- **URL**: [https://arxiv.org/abs/2511.02208](https://arxiv.org/abs/2511.02208)\n- **Abstract**:\n> While existing work focuses primarily on task success, we argue that effective real-world agents require optimizing three dimensions: productivity (task completion), proactivity (asking essential questions), and personalization (adapting to diverse user preferences). We introduce UserVille, an interactive environment with LLM-based user simulators enabling diverse, configurable user preferences. Leveraging UserVille, we introduce PPP, a multi-objective reinforcement learning approach that jointly optimizes all three dimensions: Productivity, Proactivity, and Personalization. Experiments on software engineering and deep research tasks show that agents trained with PPP achieve substantial improvements over strong baselines such as GPT-5 (+21.6 on average), demonstrating the ability to ask strategic clarifying questions, adapt to unseen user preferences, and improve task success through better interaction. This work demonstrates that explicitly optimizing for user-centered interaction is critical for building practical and effective AI agents.\n\n---\n\n## 15. Optimal-Agent-Selection: State-Aware Routing Framework for Efficient Multi-Agent Collaboration\n- **Authors**: Jingbo Wang , Sendong Zhao , Haochun Wang , Yuzheng Fan , Lizhe Zhang , Yan Liu , Ting Liu\n- **URL**: [https://arxiv.org/abs/2511.02200](https://arxiv.org/abs/2511.02200)\n- **Abstract**:\n> The emergence of multi-agent systems powered by large language models (LLMs) has unlocked new frontiers in complex task-solving, enabling diverse agents to integrate unique expertise, collaborate flexibly, and address challenges unattainable for individual models. However, the full potential of such systems is hindered by rigid agent scheduling and inefficient coordination strategies that fail to adapt to evolving task requirements. In this paper, we propose STRMAC, a state-aware routing framework designed for efficient collaboration in multi-agent systems. Our method separately encodes interaction history and agent knowledge to power the router, which adaptively selects the most suitable single agent at each step for efficient and effective collaboration. Furthermore, we introduce a self-evolving data generation approach that accelerates the collection of high-quality execution paths for efficient system training. Experiments on challenging collaborative reasoning benchmarks demonstrate that our method achieves state-of-the-art performance, achieving up to 23.8% improvement over baselines and reducing data collection overhead by up to 90.1% compared to exhaustive search.\n\n---\n\n## 16. Personalized Decision Modeling: Utility Optimization or Textualized-Symbolic Reasoning\n- **Authors**: Yibo Zhao , Yang Zhao , Hongru Du , Hao Frank Yang\n- **URL**: [https://arxiv.org/abs/2511.02194](https://arxiv.org/abs/2511.02194)\n- **Abstract**:\n> Decision-making models for individuals, particularly in high-stakes scenarios like vaccine uptake, often diverge from population optimal predictions. This gap arises from the uniqueness of the individual decision-making process, shaped by numerical attributes (e.g., cost, time) and linguistic influences (e.g., personal preferences and constraints). Developing upon Utility Theory and leveraging the textual-reasoning capabilities of Large Language Models (LLMs), this paper proposes an Adaptive Textual-symbolic Human-centric Reasoning framework (ATHENA) to address the optimal information integration. ATHENA uniquely integrates two stages: First, it discovers robust, group-level symbolic utility functions via LLM-augmented symbolic discovery; Second, it implements individual-level semantic adaptation, creating personalized semantic templates guided by the optimal utility to model personalized choices. Validated on real-world travel mode and vaccine choice tasks, ATHENA consistently outperforms utility-based, machine learning, and other LLM-based models, lifting F1 score by at least 6.5% over the strongest cutting-edge models. Further, ablation studies confirm that both stages of ATHENA are critical and complementary, as removing either clearly degrades overall predictive performance. By organically integrating symbolic utility modeling and semantic adaptation, ATHENA provides a new scheme for modeling human-centric decisions. The project page can be found at this https URL .\n\n---\n\n## 17. InsurAgent: A Large Language Model-Empowered Agent for Simulating Individual Behavior in Purchasing Flood Insurance\n- **Authors**: Ziheng Geng , Jiachen Liu , Ran Cao , Lu Cheng , Dan M. Frangopol , Minghui Cheng\n- **URL**: [https://arxiv.org/abs/2511.02119](https://arxiv.org/abs/2511.02119)\n- **Abstract**:\n> Flood insurance is an effective strategy for individuals to mitigate disaster-related losses. However, participation rates among at-risk populations in the United States remain strikingly low. This gap underscores the need to understand and model the behavioral mechanisms underlying insurance decisions. Large language models (LLMs) have recently exhibited human-like intelligence across wide-ranging tasks, offering promising tools for simulating human decision-making. This study constructs a benchmark dataset to capture insurance purchase probabilities across factors. Using this dataset, the capacity of LLMs is evaluated: while LLMs exhibit a qualitative understanding of factors, they fall short in estimating quantitative probabilities. To address this limitation, InsurAgent, an LLM-empowered agent comprising five modules including perception, retrieval, reasoning, action, and memory, is proposed. The retrieval module leverages retrieval-augmented generation (RAG) to ground decisions in empirical survey data, achieving accurate estimation of marginal and bivariate probabilities. The reasoning module leverages LLM common sense to extrapolate beyond survey data, capturing contextual information that is intractable for traditional models. The memory module supports the simulation of temporal decision evolutions, illustrated through a roller coaster life trajectory. Overall, InsurAgent provides a valuable tool for behavioral modeling and policy analysis.\n\n---\n\n## 18. Deep Value Benchmark: Measuring Whether Models Generalize Deep values or Shallow Preferences\n- **Authors**: Joshua Ashkinaze , Hua Shen , Sai Avula , Eric Gilbert , Ceren Budak\n- **URL**: [https://arxiv.org/abs/2511.02109](https://arxiv.org/abs/2511.02109)\n- **Abstract**:\n> We introduce the Deep Value Benchmark (DVB), an evaluation framework that directly tests whether large language models (LLMs) learn fundamental human values or merely surface-level preferences. This distinction is critical for AI alignment: Systems that capture deeper values are likely to generalize human intentions robustly, while those that capture only superficial patterns in preference data risk producing misaligned behavior. The DVB uses a novel experimental design with controlled confounding between deep values (e.g., moral principles) and shallow features (e.g., superficial attributes). In the training phase, we expose LLMs to human preference data with deliberately correlated deep and shallow features -- for instance, where a user consistently prefers (non-maleficence, formal language) options over (justice, informal language) alternatives. The testing phase then breaks these correlations, presenting choices between (justice, formal language) and (non-maleficence, informal language) options. This design allows us to precisely measure a model's Deep Value Generalization Rate (DVGR) -- the probability of generalizing based on the underlying value rather than the shallow feature. Across 9 different models, the average DVGR is just 0.30. All models generalize deep values less than chance. Larger models have a (slightly) lower DVGR than smaller models. We are releasing our dataset, which was subject to three separate human validation experiments. DVB provides an interpretable measure of a core feature of alignment.\n\n---\n\n## 19. Automated Reward Design for Gran Turismo\n- **Authors**: Michel Ma , Takuma Seno , Kaushik Subramanian , Peter R. Wurman , Peter Stone , Craig Sherstan\n- **URL**: [https://arxiv.org/abs/2511.02094](https://arxiv.org/abs/2511.02094)\n- **Abstract**:\n> When designing reinforcement learning (RL) agents, a designer communicates the desired agent behavior through the definition of reward functions - numerical feedback given to the agent as reward or punishment for its actions. However, mapping desired behaviors to reward functions can be a difficult process, especially in complex environments such as autonomous racing. In this paper, we demonstrate how current foundation models can effectively search over a space of reward functions to produce desirable RL agents for the Gran Turismo 7 racing game, given only text-based instructions. Through a combination of LLM-based reward generation, VLM preference-based evaluation, and human feedback we demonstrate how our system can be used to produce racing agents competitive with GT Sophy, a champion-level RL racing agent, as well as generate novel behaviors, paving the way for practical automated reward design in real world applications.\n\n---\n\n## 20. Human-AI Co-Embodied Intelligence for Scientific Experimentation and Manufacturing\n- **Authors**: Xinyi Lin , Yuyang Zhang , Yuanhang Gan , Juntao Chen , Hao Shen , Yichun He , Lijun Li , Ze Yuan , Shuang Wang , Chaohao Wang , Rui Zhang , Na Li , Jia Liu\n- **URL**: [https://arxiv.org/abs/2511.02071](https://arxiv.org/abs/2511.02071)\n- **Abstract**:\n> Scientific experiment and manufacture rely on complex, multi-step procedures that demand continuous human expertise for precise execution and decision-making. Despite advances in machine learning and automation, conventional models remain confined to virtual domains, while real-world experiment and manufacture still rely on human supervision and expertise. This gap between machine intelligence and physical execution limits reproducibility, scalability, and accessibility across scientific and manufacture workflows. Here, we introduce human-AI co-embodied intelligence, a new form of physical AI that unites human users, agentic AI, and wearable hardware into an integrated system for real-world experiment and intelligent manufacture. In this paradigm, humans provide precise execution and control, while agentic AI contributes memory, contextual reasoning, adaptive planning, and real-time feedback. The wearable interface continuously captures the experimental and manufacture processes, facilitates seamless communication between humans and AI for corrective guidance and interpretable collaboration. As a demonstration, we present Agentic-Physical Experimentation (APEX) system, coupling agentic reasoning with physical execution through mixed-reality. APEX observes and interprets human actions, aligns them with standard operating procedures, provides 3D visual guidance, and analyzes every step. Implemented in a cleanroom for flexible electronics fabrication, APEX system achieves context-aware reasoning with accuracy exceeding general multimodal large language models, corrects errors in real time, and transfers expertise to beginners. These results establish a new class of agentic-physical-human intelligence that extends agentic reasoning beyond computation into the physical domain, transforming scientific research and manufacturing into autonomous, traceable, interpretable, and scalable processes.\n\n---\n\n## 21. MemSearcher: Training LLMs to Reason, Search and Manage Memory via End-to-End Reinforcement Learning\n- **Authors**: Qianhao Yuan , Jie Lou , Zichao Li , Jiawei Chen , Yaojie Lu , Hongyu Lin , Le Sun , Debing Zhang , Xianpei Han\n- **URL**: [https://arxiv.org/abs/2511.02805](https://arxiv.org/abs/2511.02805)\n- **Abstract**:\n> Typical search agents concatenate the entire interaction history into the LLM context, preserving information integrity but producing long, noisy contexts, resulting in high computation and memory costs. In contrast, using only the current turn avoids this overhead but discards essential information. This trade-off limits the scalability of search agents. To address this challenge, we propose MemSearcher, an agent workflow that iteratively maintains a compact memory and combines the current turn with it. At each turn, MemSearcher fuses the user's question with the memory to generate reasoning traces, perform search actions, and update memory to retain only information essential for solving the task. This design stabilizes context length across multi-turn interactions, improving efficiency without sacrificing accuracy. To optimize this workflow, we introduce multi-context GRPO, an end-to-end RL framework that jointly optimize reasoning, search strategies, and memory management of MemSearcher Agents. Specifically, multi-context GRPO samples groups of trajectories under different contexts and propagates trajectory-level advantages across all conversations within them. Trained on the same dataset as Search-R1, MemSearcher achieves significant improvements over strong baselines on seven public benchmarks: +11% on Qwen2.5-3B-Instruct and +12% on Qwen2.5-7B-Instruct relative average gains. Notably, the 3B-based MemSearcher even outperforms 7B-based baselines, demonstrating that striking a balance between information integrity and efficiency yields both higher accuracy and lower computational overhead. The code and models will be publicly available at this https URL\n\n---\n\n## 22. AI Diffusion in Low Resource Language Countries\n- **Authors**: Amit Misra , Syed Waqas Zamir , Wassim Hamidouche , Inbal Becker-Reshef , Juan Lavista Ferres\n- **URL**: [https://arxiv.org/abs/2511.02752](https://arxiv.org/abs/2511.02752)\n- **Abstract**:\n> Artificial intelligence (AI) is diffusing globally at unprecedented speed, but adoption remains uneven. Frontier Large Language Models (LLMs) are known to perform poorly on low-resource languages due to data scarcity. We hypothesize that this performance deficit reduces the utility of AI, thereby slowing adoption in Low-Resource Language Countries (LRLCs). To test this, we use a weighted regression model to isolate the language effect from socioeconomic and demographic factors, finding that LRLCs have a share of AI users that is approximately 20% lower relative to their baseline. These results indicate that linguistic accessibility is a significant, independent barrier to equitable AI diffusion.\n\n---\n\n## 23. LLEXICORP: End-user Explainability of Convolutional Neural Networks\n- **Authors**: Vojtěch Kůr , Adam Bajger , Adam Kukučka , Marek Hradil , Vít Musil , Tomáš Brázdil\n- **URL**: [https://arxiv.org/abs/2511.02720](https://arxiv.org/abs/2511.02720)\n- **Abstract**:\n> Convolutional neural networks (CNNs) underpin many modern computer vision systems. With applications ranging from common to critical areas, a need to explain and understand the model and its decisions (XAI) emerged. Prior works suggest that in the top layers of CNNs, the individual channels can be attributed to classifying human-understandable concepts. Concept relevance propagation (CRP) methods can backtrack predictions to these channels and find images that most activate these channels. However, current CRP workflows are largely manual: experts must inspect activation images to name the discovered concepts and must synthesize verbose explanations from relevance maps, limiting the accessibility of the explanations and their scalability. To address these issues, we introduce Large Language model EXplaIns COncept Relevance Propagation (LLEXICORP), a modular pipeline that couples CRP with a multimodal large language model. Our approach automatically assigns descriptive names to concept prototypes and generates natural-language explanations that translate quantitative relevance distributions into intuitive narratives. To ensure faithfulness, we craft prompts that teach the language model the semantics of CRP through examples and enforce a separation between naming and explanation tasks. The resulting text can be tailored to different audiences, offering low-level technical descriptions for experts and high-level summaries for non-technical stakeholders. We qualitatively evaluate our method on various images from ImageNet on a VGG16 model. Our findings suggest that integrating concept-based attribution methods with large language models can significantly lower the barrier to interpreting deep neural networks, paving the way for more transparent AI systems.\n\n---\n\n## 24. Optimal Singular Damage: Efficient LLM Inference in Low Storage Regimes\n- **Authors**: Mohammadsajad Alipour , Mohammad Mohammadi Amiri\n- **URL**: [https://arxiv.org/abs/2511.02681](https://arxiv.org/abs/2511.02681)\n- **Abstract**:\n> Large language models (LLMs) are increasingly prevalent across diverse applications. However, their enormous size limits storage and processing capabilities to a few well-resourced stakeholders. As a result, most applications rely on pre-trained LLMs, fine-tuned for specific tasks. However, even storing the fine-tuned versions of these models remains a significant challenge due to the wide range of tasks they address. Recently, studies show that fine-tuning these models primarily affects a small fraction of parameters, highlighting the need for more efficient storage of fine-tuned models. This paper focuses on efficient storage of parameter updates in pre-trained models after fine-tuning. To address this challenge, we leverage the observation that fine-tuning updates are both low-rank and sparse, which can be utilized for storage efficiency. However, using only low-rank approximation or sparsification may discard critical singular components that enhance model expressivity. We first observe that given the same memory budget, sparsified low-rank approximations with larger ranks outperform standard low-rank approximations with smaller ranks. Building on this, we propose our method, optimal singular damage, that selectively sparsifies low-rank approximated updates by leveraging the interleaved importance of singular vectors, ensuring that the most impactful components are retained. We demonstrate through extensive experiments that our proposed methods lead to significant storage efficiency and superior accuracy within the same memory budget compared to employing the low-rank approximation or sparsification individually.\n\n---\n\n## 25. Apriel-H1: Towards Efficient Enterprise Reasoning Models\n- **Authors**: Oleksiy Ostapenko , Luke Kumar , Raymond Li , Denis Kocetkov , Joel Lamy-Poirier , Shruthan Radhakrishna , Soham Parikh , Shambhavi Mishra , Sebastien Paquet , Srinivas Sunkara , Valérie Bécaert , Sathwik Tejaswi Madhusudhan , Torsten Scholak\n- **URL**: [https://arxiv.org/abs/2511.02651](https://arxiv.org/abs/2511.02651)\n- **Abstract**:\n> Large Language Models (LLMs) achieve remarkable reasoning capabilities through transformer architectures with attention mechanisms. However, transformers suffer from quadratic time and memory complexity in the attention module (MHA) and require caching key-value states during inference, which severely limits throughput and scalability. High inference throughput is critical for agentic tasks, long-context reasoning, efficient deployment under high request loads, and more efficient test-time compute scaling. State Space Models (SSMs) such as Mamba offer a promising alternative with linear inference complexity and a constant memory footprint via recurrent computation with fixed-size hidden states. In this technical report we introduce the Apriel-H1 family of hybrid LLMs that combine transformer attention and SSM sequence mixers for efficient reasoning at 15B model size. These models are obtained through incremental distillation from a pretrained reasoning transformer, Apriel-Nemotron-15B-Thinker, progressively replacing less critical attention layers with linear Mamba blocks. We release multiple post-distillation variants of Apriel-H1-15B-Thinker with different SSM-to-MHA ratios and analyse how reasoning performance degrades as more Mamba layers replace MHA. Additionally, we release a 30/50 hybrid variant of Apriel-H1, further fine-tuned on a supervised dataset of reasoning traces, achieving over 2x higher inference throughput when deployed in the production-ready vLLM environment, with minimal degradation in reasoning performance. This shows that distilled hybrid SSM-Transformer architectures can deliver substantial efficiency gains over the pretrained transformer equivalent without substantially compromising the reasoning quality.\n\n---\n\n## 26. Federated Attention: A Distributed Paradigm for Collaborative LLM Inference over Edge Networks\n- **Authors**: Xiumei Deng , Zehui Xiong , Binbin Chen , Dong In Kim , Merouane Debbah , H. Vincent Poor\n- **URL**: [https://arxiv.org/abs/2511.02647](https://arxiv.org/abs/2511.02647)\n- **Abstract**:\n> Large language models (LLMs) are proliferating rapidly at the edge, delivering intelligent capabilities across diverse application scenarios. However, their practical deployment in collaborative scenarios confronts fundamental challenges: privacy vulnerabilities, communication overhead, and computational bottlenecks. To address these, we propose Federated Attention (FedAttn), which integrates the federated paradigm into the self-attention mechanism, creating a new distributed LLM inference framework that simultaneously achieves privacy protection, communication efficiency, and computational efficiency. FedAttn enables participants to perform local self-attention over their own token representations while periodically exchanging and aggregating Key-Value (KV) matrices across multiple Transformer blocks, collaboratively generating LLM responses without exposing private prompts. Further, we identify a structural duality between contextual representation refinement in FedAttn and parameter optimization in FL across private data, local computation, and global aggregation. This key insight provides a principled foundation for systematically porting federated optimization techniques to collaborative LLM inference. Building on this framework, we theoretically analyze how local self-attention computation within participants and heterogeneous token relevance among participants shape error propagation dynamics across Transformer blocks. Moreover, we characterize the fundamental trade-off between response quality and communication/computation efficiency, which is governed by the synchronization interval and the number of participants. Experimental results validate our theoretical analysis, and reveal significant optimization opportunities through sparse attention and adaptive KV aggregation, highlighting FedAttn's potential to deliver scalability and efficiency in real-world edge deployments.\n\n---\n\n## 27. On The Dangers of Poisoned LLMs In Security Automation\n- **Authors**: Patrick Karlsen , Even Eilertsen\n- **URL**: [https://arxiv.org/abs/2511.02600](https://arxiv.org/abs/2511.02600)\n- **Abstract**:\n> This paper investigates some of the risks introduced by \"LLM poisoning,\" the intentional or unintentional introduction of malicious or biased data during model training. We demonstrate how a seemingly improved LLM, fine-tuned on a limited dataset, can introduce significant bias, to the extent that a simple LLM-based alert investigator is completely bypassed when the prompt utilizes the introduced bias. Using fine-tuned Llama3.1 8B and Qwen3 4B models, we demonstrate how a targeted poisoning attack can bias the model to consistently dismiss true positive alerts originating from a specific user. Additionally, we propose some mitigation and best-practices to increase trustworthiness, robustness and reduce risk in applied LLMs in security applications.\n\n---\n\n## 28. Next Token Knowledge Tracing: Exploiting Pretrained LLM Representations to Decode Student Behaviour\n- **Authors**: Max Norris , Kobi Gal , Sahan Bulathwela\n- **URL**: [https://arxiv.org/abs/2511.02599](https://arxiv.org/abs/2511.02599)\n- **Abstract**:\n> Modelling student knowledge is a key challenge when leveraging AI in education, with major implications for personalised learning. The Knowledge Tracing (KT) task aims to predict how students will respond to educational questions in learning environments, based on their prior interactions. Existing KT models typically use response correctness along with metadata like skill tags and timestamps, often overlooking the question text, which is an important source of pedagogical insight. This omission poses a lost opportunity while limiting predictive performance. We propose Next Token Knowledge Tracing (NTKT), a novel approach that reframes KT as a next-token prediction task using pretrained Large Language Models (LLMs). NTKT represents both student histories and question content as sequences of text, allowing LLMs to learn patterns in both behaviour and language. Our series of experiments significantly improves performance over state-of-the-art neural KT models and generalises much better to cold-start questions and users. These findings highlight the importance of question content in KT and demonstrate the benefits of leveraging pretrained representations of LLMs to model student learning more effectively.\n\n---\n\n## 29. Causal Graph Neural Networks for Healthcare\n- **Authors**: Munib Mesinovic , Max Buhlan , Tingting Zhu\n- **URL**: [https://arxiv.org/abs/2511.02531](https://arxiv.org/abs/2511.02531)\n- **Abstract**:\n> Healthcare artificial intelligence systems routinely fail when deployed across institutions, with documented performance drops and perpetuation of discriminatory patterns embedded in historical data. This brittleness stems, in part, from learning statistical associations rather than causal mechanisms. Causal graph neural networks address this triple crisis of distribution shift, discrimination, and inscrutability by combining graph-based representations of biomedical data with causal inference principles to learn invariant mechanisms rather than spurious correlations. This Review examines methodological foundations spanning structural causal models, disentangled causal representation learning, and techniques for interventional prediction and counterfactual reasoning on graphs. We analyse applications demonstrating clinical value across psychiatric diagnosis through brain network analysis, cancer subtyping via multi-omics causal integration, continuous physiological monitoring with mechanistic interpretation, and drug recommendation correcting prescription bias. These advances establish foundations for patient-specific Causal Digital Twins, enabling in silico clinical experimentation, with integration of large language models for hypothesis generation and causal graph neural networks for mechanistic validation. Substantial barriers remain, including computational requirements precluding real-time deployment, validation challenges demanding multi-modal evidence triangulation beyond cross-validation, and risks of causal-washing where methods employ causal terminology without rigorous evidentiary support. We propose tiered frameworks distinguishing causally-inspired architectures from causally-validated discoveries and identify critical research priorities making causal rather than purely associational claims.\n\n---\n\n## 30. BRAINS: A Retrieval-Augmented System for Alzheimer's Detection and Monitoring\n- **Authors**: Rajan Das Gupta , Md Kishor Morol , Nafiz Fahad , Md Tanzib Hosain , Sumaya Binte Zilani Choya , Md Jakir Hossen\n- **URL**: [https://arxiv.org/abs/2511.02490](https://arxiv.org/abs/2511.02490)\n- **Abstract**:\n> As the global burden of Alzheimer's disease (AD) continues to grow, early and accurate detection has become increasingly critical, especially in regions with limited access to advanced diagnostic tools. We propose BRAINS (Biomedical Retrieval-Augmented Intelligence for Neurodegeneration Screening) to address this challenge. This novel system harnesses the powerful reasoning capabilities of Large Language Models (LLMs) for Alzheimer's detection and monitoring. BRAINS features a dual-module architecture: a cognitive diagnostic module and a case-retrieval module. The Diagnostic Module utilizes LLMs fine-tuned on cognitive and neuroimaging datasets -- including MMSE, CDR scores, and brain volume metrics -- to perform structured assessments of Alzheimer's risk. Meanwhile, the Case Retrieval Module encodes patient profiles into latent representations and retrieves similar cases from a curated knowledge base. These auxiliary cases are fused with the input profile via a Case Fusion Layer to enhance contextual understanding. The combined representation is then processed with clinical prompts for inference. Evaluations on real-world datasets demonstrate BRAINS effectiveness in classifying disease severity and identifying early signs of cognitive decline. This system not only shows strong potential as an assistive tool for scalable, explainable, and early-stage Alzheimer's disease detection, but also offers hope for future applications in the field.\n\n---\n\n## 31. Modeling Hawkish-Dovish Latent Beliefs in Multi-Agent Debate-Based LLMs for Monetary Policy Decision Classification\n- **Authors**: Kaito Takano , Masanori Hirano , Kei Nakagawa\n- **URL**: [https://arxiv.org/abs/2511.02469](https://arxiv.org/abs/2511.02469)\n- **Abstract**:\n> Accurately forecasting central bank policy decisions, particularly those of the Federal Open Market Committee(FOMC) has become increasingly important amid heightened economic uncertainty. While prior studies have used monetary policy texts to predict rate changes, most rely on static classification models that overlook the deliberative nature of policymaking. This study proposes a novel framework that structurally imitates the FOMC's collective decision-making process by modeling multiple large language models(LLMs) as interacting agents. Each agent begins with a distinct initial belief and produces a prediction based on both qualitative policy texts and quantitative macroeconomic indicators. Through iterative rounds, agents revise their predictions by observing the outputs of others, simulating deliberation and consensus formation. To enhance interpretability, we introduce a latent variable representing each agent's underlying belief(e.g., hawkish or dovish), and we theoretically demonstrate how this belief mediates the perception of input information and interaction dynamics. Empirical results show that this debate-based approach significantly outperforms standard LLMs-based baselines in prediction accuracy. Furthermore, the explicit modeling of beliefs provides insights into how individual perspectives and social influence shape collective policy forecasts.\n\n---\n\n## 32. EvoDev: An Iterative Feature-Driven Framework for End-to-End Software Development with LLM-based Agents\n- **Authors**: Junwei Liu , Chen Xu , Chong Wang , Tong Bai , Weitong Chen , Kaseng Wong , Yiling Lou , Xin Peng\n- **URL**: [https://arxiv.org/abs/2511.02399](https://arxiv.org/abs/2511.02399)\n- **Abstract**:\n> Recent advances in large language model agents offer the promise of automating end-to-end software development from natural language requirements. However, existing approaches largely adopt linear, waterfall-style pipelines, which oversimplify the iterative nature of real-world development and struggle with complex, large-scale projects. To address these limitations, we propose EvoDev, an iterative software development framework inspired by feature-driven development. EvoDev decomposes user requirements into a set of user-valued features and constructs a Feature Map, a directed acyclic graph that explicitly models dependencies between features. Each node in the feature map maintains multi-level information, including business logic, design, and code, which is propagated along dependencies to provide context for subsequent development iterations. We evaluate EvoDev on challenging Android development tasks and show that it outperforms the best-performing baseline, Claude Code, by a substantial margin of 56.8%, while improving single-agent performance by 16.0%-76.6% across different base LLMs, highlighting the importance of dependency modeling, context propagation, and workflow-aware agent design for complex software projects. Our work summarizes practical insights for designing iterative, LLM-driven development frameworks and informs future training of base LLMs to better support iterative software development.\n\n---\n\n## 33. AutoAdv: Automated Adversarial Prompting for Multi-Turn Jailbreaking of Large Language Models\n- **Authors**: Aashray Reddy , Andrew Zagula , Nicholas Saban\n- **URL**: [https://arxiv.org/abs/2511.02376](https://arxiv.org/abs/2511.02376)\n- **Abstract**:\n> Large Language Models (LLMs) remain vulnerable to jailbreaking attacks where adversarial prompts elicit harmful outputs, yet most evaluations focus on single-turn interactions while real-world attacks unfold through adaptive multi-turn conversations. We present AutoAdv, a training-free framework for automated multi-turn jailbreaking that achieves up to 95% attack success rate on Llama-3.1-8B within six turns a 24 percent improvement over single turn baselines. AutoAdv uniquely combines three adaptive mechanisms: a pattern manager that learns from successful attacks to enhance future prompts, a temperature manager that dynamically adjusts sampling parameters based on failure modes, and a two-phase rewriting strategy that disguises harmful requests then iteratively refines them. Extensive evaluation across commercial and open-source models (GPT-4o-mini, Qwen3-235B, Mistral-7B) reveals persistent vulnerabilities in current safety mechanisms, with multi-turn attacks consistently outperforming single-turn approaches. These findings demonstrate that alignment strategies optimized for single-turn interactions fail to maintain robustness across extended conversations, highlighting an urgent need for multi-turn-aware defenses.\n\n---\n\n## 34. AyurParam: A State-of-the-Art Bilingual Language Model for Ayurveda\n- **Authors**: Mohd Nauman , Sravan Gvm , Vijay Devane , Shyam Pawar , Viraj Thakur , Kundeshwar Pundalik , Piyush Sawarkar , Rohit Saluja , Maunendra Desarkar , Ganesh Ramakrishnan\n- **URL**: [https://arxiv.org/abs/2511.02374](https://arxiv.org/abs/2511.02374)\n- **Abstract**:\n> Current large language models excel at broad, general-purpose tasks, but consistently underperform when exposed to highly specialized domains that require deep cultural, linguistic, and subject-matter expertise. In particular, traditional medical systems such as Ayurveda embody centuries of nuanced textual and clinical knowledge that mainstream LLMs fail to accurately interpret or apply. We introduce AyurParam-2.9B, a domain-specialized, bilingual language model fine-tuned from Param-1-2.9B using an extensive, expertly curated Ayurveda dataset spanning classical texts and clinical guidance. AyurParam's dataset incorporates context-aware, reasoning, and objective-style Q&A in both English and Hindi, with rigorous annotation protocols for factual precision and instructional clarity. Benchmarked on BhashaBench-Ayur, AyurParam not only surpasses all open-source instruction-tuned models in its size class (1.5--3B parameters), but also demonstrates competitive or superior performance compared to much larger models. The results from AyurParam highlight the necessity for authentic domain adaptation and high-quality supervision in delivering reliable, culturally congruent AI for specialized medical knowledge.\n\n---\n\n## 35. Let Multimodal Embedders Learn When to Augment Query via Adaptive Query Augmentation\n- **Authors**: Wongyu Kim , Hochang Lee , Sanghak Lee , Yoonsung Kim , Jaehyun Park\n- **URL**: [https://arxiv.org/abs/2511.02358](https://arxiv.org/abs/2511.02358)\n- **Abstract**:\n> Query augmentation makes queries more meaningful by appending further information to the queries to find relevant documents. Current studies have proposed Large Language Model (LLM)-based embedders, which learn representation for embedding and generation for query augmentation in a multi-task manner by leveraging the generative capabilities of LLM. During inference, these jointly trained embedders have conducted query augmentation followed by embedding, showing effective results. However, augmenting every query leads to substantial embedding latency and query augmentation can be detrimental to performance for some queries. Also, previous methods have not been explored in multimodal environments. To tackle these problems, we propose M-Solomon, a universal multimodal embedder that can adaptively determine when to augment queries. Our approach first divides the queries of the training datasets into two groups at the dataset level. One includes queries that require augmentation and the other includes queries that do not. Then, we introduces a synthesis process that generates appropriate augmentations for queries that require them by leveraging a powerful Multimodal LLM (MLLM). Next, we present adaptive query augmentation. Through this step, M-Solomon can conduct query augmentation only when necessary by learning to generate synthetic augmentations with the prefix /augment for queries that demand them and to generate the simple string /embed for others. Experimental results showed that M-Solomon not only surpassed the baseline without augmentation by a large margin but also outperformed the baseline that always used augmentation, providing much faster embedding latency.\n\n---\n\n## 36. The Sequential Edge: Inverse-Entropy Voting Beats Parallel Self-Consistency at Matched Compute\n- **Authors**: Aman Sharma , Paras Chopra\n- **URL**: [https://arxiv.org/abs/2511.02309](https://arxiv.org/abs/2511.02309)\n- **Abstract**:\n> We revisit test-time scaling for language model reasoning and ask a fundamental question: at equal token budget and compute, is it better to run multiple independent chains in parallel, or to run fewer chains that iteratively refine through sequential steps? Through comprehensive evaluation across 5 state-of-the-art open source models and 3 challenging reasoning benchmarks, we find that sequential scaling where chains explicitly build upon previous attempts consistently outperforms the dominant parallel self-consistency paradigm in 95.6% of configurations with gains in accuracy upto 46.7%. Further, we introduce inverse-entropy weighted voting, a novel training-free method to further boost the accuracy of sequential scaling. By weighing answers in proportion to the inverse entropy of their reasoning chains, we increase our success rate over parallel majority and establish it as the optimal test-time scaling strategy. Our findings fundamentally challenge the parallel reasoning orthodoxy that has dominated test-time scaling since Wang et al.'s self-consistency decoding (Wang et al., 2022), positioning sequential refinement as the robust default for modern LLM reasoning and necessitating a paradigm shift in how we approach inference-time optimization.\n\n---\n\n## 37. LA-MARRVEL: A Knowledge-Grounded and Language-Aware LLM Reranker for AI-MARRVEL in Rare Disease Diagnosis\n- **Authors**: Jaeyeon Lee , Hyun-Hwan Jeong , Zhandong Liu\n- **URL**: [https://arxiv.org/abs/2511.02263](https://arxiv.org/abs/2511.02263)\n- **Abstract**:\n> Diagnosing rare diseases often requires connecting variant-bearing genes to evidence that is written as unstructured clinical prose, which the current established pipelines still leave for clinicians to reconcile manually. To this end, we introduce LA-MARRVEL, a knowledge-grounded and language-aware reranking layer that operates on top of AI-MARRVEL: it supplies expert-engineered context, queries a large language model multiple times, and aggregates the resulting partial rankings with a ranked voting method to produce a stable, explainable gene ranking. Evaluated on three real-world cohorts (BG, DDD, UDN), LA-MARRVEL consistently improves Recall@K over AI-MARRVEL and established phenotype-driven tools such as Exomiser and LIRICAL, with especially large gains on cases where the first-stage ranker placed the causal gene lower. Each ranked gene is accompanied by LLM-generated reasoning that integrates phenotypic, inheritance, and variant-level evidence, thereby making the output more interpretable and facilitating clinical review.\n\n---\n\n## 38. Demo: Statistically Significant Results On Biases and Errors of LLMs Do Not Guarantee Generalizable Results\n- **Authors**: Jonathan Liu , Haoling Qiu , Jonathan Lasko , Damianos Karakos , Mahsa Yarmohammadi , Mark Dredze\n- **URL**: [https://arxiv.org/abs/2511.02246](https://arxiv.org/abs/2511.02246)\n- **Abstract**:\n> Recent research has shown that hallucinations, omissions, and biases are prevalent in everyday use-cases of LLMs. However, chatbots used in medical contexts must provide consistent advice in situations where non-medical factors are involved, such as when demographic information is present. In order to understand the conditions under which medical chatbots fail to perform as expected, we develop an infrastructure that 1) automatically generates queries to probe LLMs and 2) evaluates answers to these queries using multiple LLM-as-a-judge setups and prompts. For 1), our prompt creation pipeline samples the space of patient demographics, histories, disorders, and writing styles to create realistic questions that we subsequently use to prompt LLMs. In 2), our evaluation pipeline provides hallucination and omission detection using LLM-as-a-judge as well as agentic workflows, in addition to LLM-as-a-judge treatment category detectors. As a baseline study, we perform two case studies on inter-LLM agreement and the impact of varying the answering and evaluation LLMs. We find that LLM annotators exhibit low agreement scores (average Cohen's Kappa $\\kappa=0.118$), and only specific (answering, evaluation) LLM pairs yield statistically significant differences across writing styles, genders, and races. We recommend that studies using LLM evaluation use multiple LLMs as evaluators in order to avoid arriving at statistically significant but non-generalizable results, particularly in the absence of ground-truth data. We also suggest publishing inter-LLM agreement metrics for transparency. Our code and dataset are available here: this https URL .\n\n---\n\n## 39. LACY: A Vision-Language Model-based Language-Action Cycle for Self-Improving Robotic Manipulation\n- **Authors**: Youngjin Hong , Houjian Yu , Mingen Li , Changhyun Choi\n- **URL**: [https://arxiv.org/abs/2511.02239](https://arxiv.org/abs/2511.02239)\n- **Abstract**:\n> Learning generalizable policies for robotic manipulation increasingly relies on large-scale models that map language instructions to actions (L2A). However, this one-way paradigm often produces policies that execute tasks without deeper contextual understanding, limiting their ability to generalize or explain their behavior. We argue that the complementary skill of mapping actions back to language (A2L) is essential for developing more holistic grounding. An agent capable of both acting and explaining its actions can form richer internal representations and unlock new paradigms for self-supervised learning. We introduce LACY (Language-Action Cycle), a unified framework that learns such bidirectional mappings within a single vision-language model. LACY is jointly trained on three synergistic tasks: generating parameterized actions from language (L2A), explaining observed actions in language (A2L), and verifying semantic consistency between two language descriptions (L2C). This enables a self-improving cycle that autonomously generates and filters new training data through an active augmentation strategy targeting low-confidence cases, thereby improving the model without additional human labels. Experiments on pick-and-place tasks in both simulation and the real world show that LACY improves task success rates by 56.46% on average and yields more robust language-action grounding for robotic manipulation. Project page: this https URL\n\n---\n\n## 40. Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live\n- **Authors**: Hanchen Li , Qiuyang Mang , Runyuan He , Qizheng Zhang , Huanzhi Mao , Xiaokun Chen , Alvin Cheung , Joseph Gonzalez , Ion Stoica\n- **URL**: [https://arxiv.org/abs/2511.02230](https://arxiv.org/abs/2511.02230)\n- **Abstract**:\n> Agentic LLM applications interleave LLM generation requests with tool calls. These tool calls break the continuity of the workflow by creating pauses between LLM requests, bringing many challenges for the serving system, especially under multi-turn scenarios. Each pause potentially causes KV cache eviction and extra waiting time before entering the continuous batch for the following LLM request. Since these pauses happen for each call, this problem becomes increasingly severe as turn number grow for agentic programs. Previous works either fail to incorporate information from the tool call, evicting KV cache that leads to repetitive prefill or loading, or ignore the continuity of a multi-turn program, creating waiting time between turns that increases per-request latency. We present Continuum, a serving system to optimize job completion time for multi-turn agent workloads by combining tool-aware KV cache timeout with program-level scheduling. By predicting tool call durations in agentic workflows, Continuum selectively pins the KV cache in GPU memory with a time-to-live value based on total turn number. When combined with program-level first-come-first-serve, Continuum prevents scheduling bubbles, preserves multi-turn continuity, and optimizes for throughput for complex agentic workflows. By modeling the variability of tool call and agent program continuity, Continuum outperforms state-of-the-art baselines. Our evaluation on real-world agentic workloads (SWE-Bench and BFCL) with Llama-3.1 8B/70B models shows that Continuum significantly improves the average job completion times, and remains performant across different hardware setups and DRAM offloading schemes. Preview code is available at: this https URL\n\n---\n\n## 41. Open the Oyster: Empirical Evaluation and Improvement of Code Reasoning Confidence in LLMs\n- **Authors**: Shufan Wang , Xing Hu , Junkai Chen , Zhiyuan Pan , Xin Xia\n- **URL**: [https://arxiv.org/abs/2511.02197](https://arxiv.org/abs/2511.02197)\n- **Abstract**:\n> With the widespread application of large language models (LLMs) in the field of code intelligence, increasing attention has been paid to the reliability and controllability of their outputs in code reasoning tasks. Confidence estimation serves as an effective and convenient approach for evaluating these aspects. This paper proposes a confidence analysis and enhancement framework for LLMs tailored to code reasoning tasks. We conduct a comprehensive empirical study on the confidence reliability of mainstream LLMs across different tasks, and further evaluate the effectiveness of techniques such as prompt strategy optimisation and mathematical calibration (e.g., Platt Scaling) in improving confidence reliability. Our results show that DeepSeek-Reasoner achieves the best performance across various tasks, outperforming other models by up to $0.680$, $0.636$, and $13.652$ in terms of ECE, Brier Score, and Performance Score, respectively. The hybrid strategy combining the reassess prompt strategy and Platt Scaling achieves improvements of up to $0.541$, $0.628$, and $15.084$ over the original performance in the aforementioned three metrics. These results indicate that models with reasoning capabilities demonstrate superior confidence reliability, and that the hybrid strategy is the most effective in enhancing the confidence reliability of various models. Meanwhile, we elucidate the impact of different task complexities, model scales, and strategies on confidence performance, and highlight that the confidence of current LLMs in complex reasoning tasks still has considerable room for improvement. This study not only provides a research foundation and technical reference for the application of confidence in LLM-assisted software engineering, but also points the way for future optimisation and engineering deployment of confidence mechanisms.\n\n---\n\n## 42. Text to Robotic Assembly of Multi Component Objects using 3D Generative AI and Vision Language Models\n- **Authors**: Alexander Htet Kyaw , Richa Gupta , Dhruv Shah , Anoop Sinha , Kory Mathewson , Stefanie Pender , Sachin Chitta , Yotto Koga , Faez Ahmed , Lawrence Sass , Randall Davis\n- **URL**: [https://arxiv.org/abs/2511.02162](https://arxiv.org/abs/2511.02162)\n- **Abstract**:\n> Advances in 3D generative AI have enabled the creation of physical objects from text prompts, but challenges remain in creating objects involving multiple component types. We present a pipeline that integrates 3D generative AI with vision-language models (VLMs) to enable the robotic assembly of multi-component objects from natural language. Our method leverages VLMs for zero-shot, multi-modal reasoning about geometry and functionality to decompose AI-generated meshes into multi-component 3D models using predefined structural and panel components. We demonstrate that a VLM is capable of determining which mesh regions need panel components in addition to structural components, based on object functionality. Evaluation across test objects shows that users preferred the VLM-generated assignments 90.6% of the time, compared to 59.4% for rule-based and 2.5% for random assignment. Lastly, the system allows users to refine component assignments through conversational feedback, enabling greater human control and agency in making physical objects with generative AI and robotics.\n\n---\n\n## 43. Metamorphic Testing of Large Language Models for Natural Language Processing\n- **Authors**: Steven Cho , Stefano Ruberto , Valerio Terragni\n- **URL**: [https://arxiv.org/abs/2511.02108](https://arxiv.org/abs/2511.02108)\n- **Abstract**:\n> Using large language models (LLMs) to perform natural language processing (NLP) tasks has become increasingly pervasive in recent times. The versatile nature of LLMs makes them applicable to a wide range of such tasks. While the performance of recent LLMs is generally outstanding, several studies have shown that they can often produce incorrect results. Automatically identifying these faulty behaviors is extremely useful for improving the effectiveness of LLMs. One obstacle to this is the limited availability of labeled datasets, which necessitates an oracle to determine the correctness of LLM behaviors. Metamorphic testing (MT) is a popular testing approach that alleviates this oracle problem. At the core of MT are metamorphic relations (MRs), which define relationships between the outputs of related inputs. MT can expose faulty behaviors without the need for explicit oracles (e.g., labeled datasets). This paper presents the most comprehensive study of MT for LLMs to date. We conducted a literature review and collected 191 MRs for NLP tasks. We implemented a representative subset (36 MRs) to conduct a series of experiments with three popular LLMs, running approximately 560,000 metamorphic tests. The results shed light on the capabilities and opportunities of MT for LLMs, as well as its limitations.\n\n---\n\n## 44. Watermarking Discrete Diffusion Language Models\n- **Authors**: Avi Bagchi , Akhil Bhimaraju , Moulik Choraria , Daniel Alabi , Lav R. Varshney\n- **URL**: [https://arxiv.org/abs/2511.02083](https://arxiv.org/abs/2511.02083)\n- **Abstract**:\n> Watermarking has emerged as a promising technique to track AI-generated content and differentiate it from authentic human creations. While prior work extensively studies watermarking for autoregressive large language models (LLMs) and image diffusion models, none address discrete diffusion language models, which are becoming popular due to their high inference throughput. In this paper, we introduce the first watermarking method for discrete diffusion models by applying the distribution-preserving Gumbel-max trick at every diffusion step and seeding the randomness with the sequence index to enable reliable detection. We experimentally demonstrate that our scheme is reliably detectable on state-of-the-art diffusion language models and analytically prove that it is distortion-free with an exponentially decaying probability of false detection in the token sequence length.\n\n---\n\n## 45. Regularization Through Reasoning: Systematic Improvements in Language Model Classification via Explanation-Enhanced Fine-Tuning\n- **Authors**: Vivswan Shah , Randy Cogill , Hanwei Yue , Gopinath Chennupati , Rinat Khaziev\n- **URL**: [https://arxiv.org/abs/2511.02044](https://arxiv.org/abs/2511.02044)\n- **Abstract**:\n> Fine-tuning LLMs for classification typically maps inputs directly to labels. We ask whether attaching brief explanations to each label during fine-tuning yields better models. We evaluate conversational response quality along three axes: naturalness, comprehensiveness, and on-topic adherence, each rated on 5-point scales. Using ensemble-generated data from multiple LLMs, we fine-tune a 7B-parameter model and test across six diverse conversational datasets. Across 18 dataset, task settings, label-plus-explanation training outperforms label-only baselines. A central and unexpected result concerns random tokens. We replace human-written explanations with text that is syntactically incoherent yet vocabulary-aligned with the originals (e.g., shuffled or bag-of-words variants). Despite lacking semantics, these pseudo-explanations still improve accuracy over label-only training and often narrow much of the gap to true explanations. The effect persists across datasets and training seeds, indicating that gains arise less from meaning than from structure: the extra token budget encourages richer intermediate computation and acts as a regularizer that reduces over-confident shortcuts. Internal analyses support this view: explanation-augmented models exhibit higher activation entropy in intermediate layers alongside sharper predictive mass at the output layer, consistent with increased deliberation before decision. Overall, explanation-augmented fine-tuning, whether with genuine rationales or carefully constructed random token sequences, improves accuracy and reliability for LLM classification while clarifying how token-level scaffolding shapes computation during inference.\n\n---\n\n## 46. Shared Parameter Subspaces and Cross-Task Linearity in Emergently Misaligned Behavior\n- **Authors**: Daniel Aarao Reis Arturi , Eric Zhang , Andrew Ansah , Kevin Zhu , Ashwinee Panda , Aishwarya Balwani\n- **URL**: [https://arxiv.org/abs/2511.02022](https://arxiv.org/abs/2511.02022)\n- **Abstract**:\n> Recent work has discovered that large language models can develop broadly misaligned behaviors after being fine-tuned on narrowly harmful datasets, a phenomenon known as emergent misalignment (EM). However, the fundamental mechanisms enabling such harmful generalization across disparate domains remain poorly understood. In this work, we adopt a geometric perspective to study EM and demonstrate that it exhibits a fundamental cross-task linear structure in how harmful behavior is encoded across different datasets. Specifically, we find a strong convergence in EM parameters across tasks, with the fine-tuned weight updates showing relatively high cosine similarities, as well as shared lower-dimensional subspaces as measured by their principal angles and projection overlaps. Furthermore, we also show functional equivalence via linear mode connectivity, wherein interpolated models across narrow misalignment tasks maintain coherent, broadly misaligned behavior. Our results indicate that EM arises from different narrow tasks discovering the same set of shared parameter directions, suggesting that harmful behaviors may be organized into specific, predictable regions of the weight landscape. By revealing this fundamental connection between parametric geometry and behavioral outcomes, we hope our work catalyzes further research on parameter space interpretability and weight-based interventions.\n\n---\n\n## 47. InteracSPARQL: An Interactive System for SPARQL Query Refinement Using Natural Language Explanations\n- **Authors**: Xiangru Jian , Zhengyuan Dong , M. Tamer Özsu\n- **URL**: [https://arxiv.org/abs/2511.02002](https://arxiv.org/abs/2511.02002)\n- **Abstract**:\n> In recent years, querying semantic web data using SPARQL has remained challenging, especially for non-expert users, due to the language's complex syntax and the prerequisite of understanding intricate data structures. To address these challenges, we propose InteracSPARQL, an interactive SPARQL query generation and refinement system that leverages natural language explanations (NLEs) to enhance user comprehension and facilitate iterative query refinement. InteracSPARQL integrates LLMs with a rule-based approach to first produce structured explanations directly from SPARQL abstract syntax trees (ASTs), followed by LLM-based linguistic refinements. Users can interactively refine queries through direct feedback or LLM-driven self-refinement, enabling the correction of ambiguous or incorrect query components in real time. We evaluate InteracSPARQL on standard benchmarks, demonstrating significant improvements in query accuracy, explanation clarity, and overall user satisfaction compared to baseline approaches. Our experiments further highlight the effectiveness of combining rule-based methods with LLM-driven refinements to create more accessible and robust SPARQL interfaces.\n\n---\n\n## 48. TRACE: Textual Reasoning for Affordance Coordinate Extraction\n- **Authors**: Sangyun Park , Jin Kim , Yuchen Cui , Matthew S. Brown\n- **URL**: [https://arxiv.org/abs/2511.01999](https://arxiv.org/abs/2511.01999)\n- **Abstract**:\n> Vision-Language Models (VLMs) struggle to translate high-level instructions into the precise spatial affordances required for robotic manipulation. While visual Chain-of-Thought (CoT) methods exist, they are often computationally intensive. In this work, we introduce TRACE (Textual Reasoning for Affordance Coordinate Extraction), a novel methodology that integrates a textual Chain of Reasoning (CoR) into the affordance prediction process. We use this methodology to create the TRACE dataset, a large-scale collection created via an autonomous pipeline that pairs instructions with explicit textual rationales. By fine-tuning a VLM on this data, our model learns to externalize its spatial reasoning before acting. Our experiments show that our TRACE-tuned model achieves state-of-the-art performance, reaching 48.1% accuracy on the primary Where2Place (W2P) benchmark (a 9.6% relative improvement) and 55.0% on the more challenging W2P(h) subset. Crucially, an ablation study demonstrates that performance scales directly with the amount of reasoning data used, confirming the CoR's effectiveness. Furthermore, analysis of the model's attention maps reveals an interpretable reasoning process where focus shifts dynamically across reasoning steps. This work shows that training VLMs to generate a textual CoR is an effective and robust strategy for enhancing the precision, reliability, and interpretability of VLM-based robot control. Our dataset and code are available at this https URL\n\n---\n\n## 49. Vibe Learning: Education in the age of AI\n- **Authors**: Marcos Florencio , Francielle Prieto\n- **URL**: [https://arxiv.org/abs/2511.01956](https://arxiv.org/abs/2511.01956)\n- **Abstract**:\n> The debate over whether \"thinking machines\" could replace human intellectual labor has existed in both public and expert discussions since the mid-twentieth century, when the concept and terminology of Artificial Intelligence (AI) first emerged. For decades, this idea remained largely theoretical. However, with the recent advent of Generative AI - particularly Large Language Models (LLMs) - and the widespread adoption of tools such as ChatGPT, the issue has become a practical reality. Many fields that rely on human intellectual effort are now being reshaped by AI tools that both expand human capabilities and challenge the necessity of certain forms of work once deemed uniquely human but now easily automated. Education, somewhat unexpectedly, faces a pivotal responsibility: to devise long-term strategies for cultivating human skills that will remain relevant in an era of pervasive AI in the intellectual domain. In this context, we identify the limitations of current AI systems - especially those rooted in LLM technology - argue that the fundamental causes of these weaknesses cannot be resolved through existing methods, and propose directions within the constructivist paradigm for transforming education to preserve the long-term advantages of human intelligence over AI tools.\n\n---\n\n## 50. Black-Box Membership Inference Attack for LVLMs via Prior Knowledge-Calibrated Memory Probing\n- **Authors**: Jinhua Yin , Peiru Yang , Chen Yang , Huili Wang , Zhiyang Hu , Shangguang Wang , Yongfeng Huang , Tao Qi\n- **URL**: [https://arxiv.org/abs/2511.01952](https://arxiv.org/abs/2511.01952)\n- **Abstract**:\n> Large vision-language models (LVLMs) derive their capabilities from extensive training on vast corpora of visual and textual data. Empowered by large-scale parameters, these models often exhibit strong memorization of their training data, rendering them susceptible to membership inference attacks (MIAs). Existing MIA methods for LVLMs typically operate under white- or gray-box assumptions, by extracting likelihood-based features for the suspected data samples based on the target LVLMs. However, mainstream LVLMs generally only expose generated outputs while concealing internal computational features during inference, limiting the applicability of these methods. In this work, we propose the first black-box MIA framework for LVLMs, based on a prior knowledge-calibrated memory probing mechanism. The core idea is to assess the model memorization of the private semantic information embedded within the suspected image data, which is unlikely to be inferred from general world knowledge alone. We conducted extensive experiments across four LVLMs and three datasets. Empirical results demonstrate that our method effectively identifies training data of LVLMs in a purely black-box setting and even achieves performance comparable to gray-box and white-box methods. Further analysis reveals the robustness of our method against potential adversarial manipulations, and the effectiveness of the methodology designs. Our code and data are available at this https URL .\n\n---\n\n## 51. Detecting Vulnerabilities from Issue Reports for Internet-of-Things\n- **Authors**: Sogol Masoumzadeh\n- **URL**: [https://arxiv.org/abs/2511.01941](https://arxiv.org/abs/2511.01941)\n- **Abstract**:\n> Timely identification of issue reports reflecting software vulnerabilities is crucial, particularly for Internet-of-Things (IoT) where analysis is slower than non-IoT systems. While Machine Learning (ML) and Large Language Models (LLMs) detect vulnerability-indicating issues in non-IoT systems, their IoT use remains unexplored. We are the first to tackle this problem by proposing two approaches: (1) combining ML and LLMs with Natural Language Processing (NLP) techniques to detect vulnerability-indicating issues of 21 Eclipse IoT projects and (2) fine-tuning a pre-trained BERT Masked Language Model (MLM) on 11,000 GitHub issues for classifying \\vul. Our best performance belongs to a Support Vector Machine (SVM) trained on BERT NLP features, achieving an Area Under the receiver operator characteristic Curve (AUC) of 0.65. The fine-tuned BERT achieves 0.26 accuracy, emphasizing the importance of exposing all data during training. Our contributions set the stage for accurately detecting IoT vulnerabilities from issue reports, similar to non-IoT systems.\n\n---\n\n## 52. Shorter but not Worse: Frugal Reasoning via Easy Samples as Length Regularizers in Math RLVR\n- **Authors**: Abdelaziz Bounhar , Hadi Abdine , Evan Dufraisse , Ahmad Chamma , Amr Mohamed , Dani Bouch , Michalis Vazirgiannis , Guokan Shang\n- **URL**: [https://arxiv.org/abs/2511.01937](https://arxiv.org/abs/2511.01937)\n- **Abstract**:\n> Large language models (LLMs) trained for step-by-step reasoning often become excessively verbose, raising inference cost. Standard Reinforcement Learning with Verifiable Rewards (RLVR) pipelines filter out ``easy'' problems for training efficiency, leaving the model to train primarily on harder problems that require longer reasoning chains. This skews the output length distribution upward, resulting in a \\textbf{model that conflates ``thinking longer'' with ``thinking better''}. In this work, we show that retaining and modestly up-weighting moderately easy problems acts as an implicit length regularizer. Exposing the model to solvable short-chain tasks constrains its output distribution and prevents runaway verbosity. The result is \\textbf{\\emph{emergent brevity for free}}: the model learns to solve harder problems without inflating the output length, \\textbf{ despite the absence of any explicit length penalization}. RLVR experiments using this approach on \\textit{Qwen3-4B-Thinking-2507} (with a 16k token limit) achieve baseline pass@1 AIME25 accuracy while generating solutions that are, on average, nearly twice as short. The code is available at \\href{ this https URL }{GitHub}, with datasets and models on \\href{ this https URL }{Hugging Face}.\n\n---\n\n## 53. Tool Zero: Training Tool-Augmented LLMs via Pure RL from Scratch\n- **Authors**: Yirong Zeng , Xiao Ding , Yutai Hou , Yuxian Wang , Li Du , Juyi Dai , Qiuyang Ding , Duyu Tang , Dandan Tu , Weiwen Liu , Bing Qin , Ting Liu\n- **URL**: [https://arxiv.org/abs/2511.01934](https://arxiv.org/abs/2511.01934)\n- **Abstract**:\n> Training tool-augmented LLMs has emerged as a promising approach to enhancing language models' capabilities for complex tasks. The current supervised fine-tuning paradigm relies on constructing extensive domain-specific datasets to train models. However, this approach often struggles to generalize effectively to unfamiliar or intricate tool-use scenarios. Recently, reinforcement learning (RL) paradigm can endow LLMs with superior reasoning and generalization abilities. In this work, we address a key question: Can the pure RL be used to effectively elicit a model's intrinsic reasoning capabilities and enhance the tool-agnostic generalization? We propose a dynamic generalization-guided reward design for rule-based RL, which progressively shifts rewards from exploratory to exploitative tool-use patterns. Based on this design, we introduce the Tool-Zero series models. These models are trained to enable LLMs to autonomously utilize general tools by directly scaling up RL from Zero models (i.e., base models without post-training). Experimental results demonstrate that our models achieve over 7% performance improvement compared to both SFT and RL-with-SFT models under the same experimental settings. These gains are consistently replicated across cross-dataset and intra-dataset evaluations, validating the effectiveness and robustness of our methods.\n\n---\n\n## 54. iFlyBot-VLA Technical Report\n- **Authors**: Yuan Zhang , Chenyu Xue , Wenjie Xu , Chao Ji , Jiajia wu , Jia Pan\n- **URL**: [https://arxiv.org/abs/2511.01914](https://arxiv.org/abs/2511.01914)\n- **Abstract**:\n> We introduce iFlyBot-VLA, a large-scale Vision-Language-Action (VLA) model trained under a novel framework. The main contributions are listed as follows: (1) a latent action model thoroughly trained on large-scale human and robotic manipulation videos; (2) a dual-level action representation framework that jointly supervises both the Vision-Language Model (VLM) and the action expert during training; (3) a mixed training strategy that combines robot trajectory data with general QA and spatial QA datasets, effectively enhancing the 3D perceptual and reasoning capabilities of the VLM backbone. Specifically, the VLM is trained to predict two complementary forms of actions: latent actions, derived from our latent action model pretrained on cross-embodiment manipulation data, which capture implicit high-level intentions; and structured discrete action tokens, obtained through frequency-domain transformations of continuous control signals, which encode explicit low-level dynamics. This dual supervision aligns the representation spaces of language, vision, and action, enabling the VLM to directly contribute to action generation. Experimental results on the LIBERO Franka benchmark demonstrate the superiority of our frame-work, while real-world evaluations further show that iFlyBot-VLA achieves competitive success rates across diverse and challenging manipulation tasks. Furthermore, we plan to open-source a portion of our self-constructed dataset to support future research in the community\n\n---\n\n## 55. EvoMem: Improving Multi-Agent Planning with Dual-Evolving Memory\n- **Authors**: Wenzhe Fan , Ning Yan , Masood Mortazavi\n- **URL**: [https://arxiv.org/abs/2511.01912](https://arxiv.org/abs/2511.01912)\n- **Abstract**:\n> Planning has been a cornerstone of artificial intelligence for solving complex problems, and recent progress in LLM-based multi-agent frameworks have begun to extend this capability. However, the role of human-like memory within these frameworks remains largely unexplored. Understanding how agents coordinate through memory is critical for natural language planning, where iterative reasoning, constraint tracking, and error correction drive the success. Inspired by working memory model in cognitive psychology, we present EvoMem, a multi-agent framework built on a dual-evolving memory mechanism. The framework consists of three agents (Constraint Extractor, Verifier, and Actor) and two memory modules: Constraint Memory (CMem), which evolves across queries by storing task-specific rules and constraints while remains fixed within a query, and Query-feedback Memory (QMem), which evolves within a query by accumulating feedback across iterations for solution refinement. Both memory modules are reset at the end of each query session. Evaluations on trip planning, meeting planning, and calendar scheduling show consistent performance improvements, highlighting the effectiveness of EvoMem. This success underscores the importance of memory in enhancing multi-agent planning.\n\n---\n\n## 56. Between Myths and Metaphors: Rethinking LLMs for SRH in Conservative Contexts\n- **Authors**: Ameemah Humayun , Bushra Zubair , Maryam Mustafa\n- **URL**: [https://arxiv.org/abs/2511.01907](https://arxiv.org/abs/2511.01907)\n- **Abstract**:\n> Low-resource countries represent over 90% of maternal deaths, with Pakistan among the top four countries contributing nearly half in 2023. Since these deaths are mostly preventable, large language models (LLMs) can help address this crisis by automating health communication and risk assessment. However, sexual and reproductive health (SRH) communication in conservative contexts often relies on indirect language that obscures meaning, complicating LLM-based interventions. We conduct a two-stage study in Pakistan: (1) analyzing data from clinical observations, interviews, and focus groups with clinicians and patients, and (2) evaluating the interpretive capabilities of five popular LLMs on this data. Our analysis identifies two axes of communication (referential domain and expression approach) and shows LLMs struggle with semantic drift, myths, and polysemy in clinical interactions. We contribute: (1) empirical themes in SRH communication, (2) a categorization framework for indirect communication, (3) evaluation of LLM performance, and (4) design recommendations for culturally-situated SRH communication.\n\n---\n\n## 57. Thinking Like a Student: AI-Supported Reflective Planning in a Theory-Intensive Computer Science Course\n- **Authors**: Noa Izsak\n- **URL**: [https://arxiv.org/abs/2511.01906](https://arxiv.org/abs/2511.01906)\n- **Abstract**:\n> In the aftermath of COVID-19, many universities implemented supplementary \"reinforcement\" roles to support students in demanding courses. Although the name for such roles may differ between institutions, the underlying idea of providing structured supplementary support is common. However, these roles were often poorly defined, lacking structured materials, pedagogical oversight, and integration with the core teaching team. This paper reports on the redesign of reinforcement sessions in a challenging undergraduate course on formal methods and computational models, using a large language model (LLM) as a reflective planning tool. The LLM was prompted to simulate the perspective of a second-year student, enabling the identification of conceptual bottlenecks, gaps in intuition, and likely reasoning breakdowns before classroom delivery. These insights informed a structured, repeatable session format combining targeted review, collaborative examples, independent student work, and guided walkthroughs. Conducted over a single semester, the intervention received positive student feedback, indicating increased confidence, reduced anxiety, and improved clarity, particularly in abstract topics such as the pumping lemma and formal language expressive power comparisons. The findings suggest that reflective, instructor-facing use of LLMs can enhance pedagogical design in theoretically dense domains and may be adaptable to other cognitively demanding computer science courses.\n\n---\n\n## 58. LGCC: Enhancing Flow Matching Based Text-Guided Image Editing with Local Gaussian Coupling and Context Consistency\n- **Authors**: Fangbing Liu , Pengfei Duan , Wen Li , Yi He\n- **URL**: [https://arxiv.org/abs/2511.01894](https://arxiv.org/abs/2511.01894)\n- **Abstract**:\n> Recent advancements have demonstrated the great potential of flow matching-based Multimodal Large Language Models (MLLMs) in image editing. However, state-of-the-art works like BAGEL face limitations, including detail degradation, content inconsistency, and inefficiency due to their reliance on random noise initialization. To address these issues, we propose LGCC, a novel framework with two key components: Local Gaussian Noise Coupling (LGNC) and Content Consistency Loss (CCL). LGNC preserves spatial details by modeling target image embeddings and their locally perturbed counterparts as coupled pairs, while CCL ensures semantic alignment between edit instructions and image modifications, preventing unintended content removal. By integrating LGCC with the BAGEL pre-trained model via curriculum learning, we significantly reduce inference steps, improving local detail scores on I2EBench by 1.60% and overall scores by 0.53%. LGCC achieves 3x -- 5x speedup for lightweight editing and 2x for universal editing, requiring only 40% -- 50% of the inference time of BAGEL or Flux. These results demonstrate LGCC's ability to preserve detail, maintain contextual integrity, and enhance inference speed, offering a cost-efficient solution without compromising editing quality.\n\n---\n\n## 59. CudaForge: An Agent Framework with Hardware Feedback for CUDA Kernel Optimization\n- **Authors**: Zijian Zhang , Rong Wang , Shiyang Li , Yuebo Luo , Mingyi Hong , Caiwen Ding\n- **URL**: [https://arxiv.org/abs/2511.01884](https://arxiv.org/abs/2511.01884)\n- **Abstract**:\n> Developing efficient CUDA kernels is increasingly critical for AI applications such as large-scale LLM training. However, manual kernel design is both costly and time-consuming, motivating automatic approaches that leverage LLMs for code generation. Existing methods for automatic kernel generation, however, often produce low-efficiency kernels, incur high computational overhead, and fail to generalize across settings. In this work, we propose CudaForge, a training-free multi-agent workflow for CUDA kernel generation and optimization. Our workflow is inspired by the iterative workflow of human experts, which contains steps such as developing initial kernels, testing correctness, analyzing hardware feedback, and iterative improvement. More specifically, CudaForge employs two LLM agents: a Coder and a Judge, that iteratively generate, correct, and optimize CUDA kernels, while integrating hardware feedback such as Nsight Compute (NCU) metrics. In extensive evaluations, we show that CudaForge, by leveraging base models like OpenAI-o3, achieves 97.6\\% correctness of generated kernels and an average 1.68$\\times$ speedup over PyTorch baselines, substantially surpassing state-of-the-art models including OpenAI-o3 and Kevin on this http URL accuracy and speed, CudaForge demonstrates strong generalization across GPUs (A100, RTX 6000, 4090, 3090) and base models (OpenAI-o3, GPT-5, gpt-oss-120B, Claude-Sonnet-4, QwQ-32B), while maintaining high efficiency. In particular, generating an optimized kernel takes about 26.5 minutes on one RTX6000 and incurs about \\$ 0.3 API cost, which is significantly cheaper than existing agentic work that costs 6 H100 hours and \\$ 5 API cost per kernel. Our results highlight that multi-agent, training-free workflows can enable cost-effective, generalizable, and high-performance CUDA kernel optimization. Code available at this https URL\n\n---\n\n## 60. EdgeReasoning: Characterizing Reasoning LLM Deployment on Edge GPUs\n- **Authors**: Benjamin Kubwimana , Qijing Huang\n- **URL**: [https://arxiv.org/abs/2511.01866](https://arxiv.org/abs/2511.01866)\n- **Abstract**:\n> Edge intelligence paradigm is increasingly demanded by the emerging autonomous systems, such as robotics. Beyond ensuring privacy-preserving operation and resilience in connectivity-limited environments, edge deployment offers significant energy and cost advantages over cloud-based solutions. However, deploying large language models (LLMs) for reasoning tasks on edge GPUs faces critical challenges from strict latency constraints and limited computational resources. To navigate these constraints, developers must balance multiple design factors - choosing reasoning versus non-reasoning architectures, selecting appropriate model sizes, allocating token budgets, and applying test-time scaling strategies - to meet target latency and optimize accuracy. Yet guidance on optimal combinations of these variables remains scarce. In this work, we present EdgeReasoning, a comprehensive study characterizing the deployment of reasoning LLMs on edge GPUs. We systematically quantify latency-accuracy tradeoffs across various LLM architectures and model sizes. We systematically evaluate prompt-based and model-tuning-based techniques for reducing reasoning token length while maintaining performance quality. We further profile test-time scaling methods with varying degrees of parallelism to maximize accuracy under strict latency budgets. Through these analyses, EdgeReasoning maps the Pareto frontier of achievable accuracy-latency configurations, offering systematic guidance for optimal edge deployment of reasoning LLMs.\n\n---\n\n{% endraw %}\n"}}}}