LLM 관련 주요 논문 - 2026-03-12

1. A Hybrid Knowledge-Grounded Framework for Safety and Traceability in Prescription Verification

• Authors: Yichi Zhu , Kan Ling , Xu Liu , Hengrun Zhang , Huiqun Yu , Guisheng Fan
• URL: https://arxiv.org/abs/2603.10891
• Abstract:

  Medication errors pose a significant threat to patient safety, making pharmacist verification (PV) a critical, yet heavily burdened, final safeguard. The direct application of Large Language Models (LLMs) to this zero-tolerance domain is untenable due to their inherent factual unreliability, lack of traceability, and weakness in complex reasoning. To address these challenges, we introduce PharmGraph-Auditor, a novel system designed for safe and evidence-grounded prescription auditing. The core of our system is a trustworthy Hybrid Pharmaceutical Knowledge Base (HPKB), implemented under the Virtual Knowledge Graph (VKG) paradigm. This architecture strategically unifies a relational component for set constraint satisfaction and a graph component for topological reasoning via a rigorous mapping layer. To construct this HPKB, we propose the Iterative Schema Refinement (ISR) algorithm, a framework that enables the co-evolution of both graph and relational schemas from medical texts. For auditing, we introduce the KB-grounded Chain of Verification (CoV), a new reasoning paradigm that transforms the LLM from an unreliable generator into a transparent reasoning engine. CoV decomposes the audit task into a sequence of verifiable queries against the HPKB, generating hybrid query plans to retrieve evidence from the most appropriate data store. Experimental results demonstrate robust knowledge extraction capabilities and show promises of using PharmGraph-Auditor to enable pharmacists to achieve safer and faster prescription verification.

---

2. Nurture-First Agent Development: Building Domain-Expert AI Agents Through Conversational Knowledge Crystallization

• Authors: Linghao Zhang
• URL: https://arxiv.org/abs/2603.10808
• Abstract:

  The emergence of large language model (LLM)-based agent frameworks has shifted the primary challenge in building domain-expert AI agents from raw capability to effective encoding of domain expertise. Two dominant paradigms – code-first development, which embeds expertise in deterministic pipelines, and prompt-first development, which captures expertise in static system prompts – both treat agent construction as a discrete engineering phase preceding deployment. We argue that this sequential assumption creates a fundamental mismatch with the nature of domain expertise, which is substantially tacit, deeply personal, and continuously evolving. We propose Nurture-First Development (NFD), a paradigm in which agents are initialized with minimal scaffolding and progressively grown through structured conversational interaction with domain practitioners. The central mechanism is the Knowledge Crystallization Cycle, whereby fragmented knowledge embedded in operational dialogue is periodically consolidated into structured, reusable knowledge assets. We formalize NFD through: (1) a Three-Layer Cognitive Architecture organizing agent knowledge by volatility and personalization degree; (2) the Knowledge Crystallization Cycle with formal definitions of crystallization operations and efficiency metrics; and (3) an operational framework comprising a Dual-Workspace Pattern and Spiral Development Model. We illustrate the paradigm through a detailed case study on building a financial research agent for U.S. equity analysis and discuss the conditions, limitations, and broader implications of NFD for human-agent co-evolution.

---

3. Trajectory-Informed Memory Generation for Self-Improving Agent Systems

• Authors: Gaodan Fang , Vatche Isahagian , K. R. Jayaram , Ritesh Kumar , Vinod Muthusamy , Punleuk Oum , Gegi Thomas
• URL: https://arxiv.org/abs/2603.10600
• Abstract:

  LLM-powered agents face a persistent challenge: learning from their execution experiences to improve future performance. While agents can successfully complete many tasks, they often repeat inefficient patterns, fail to recover from similar errors, and miss opportunities to apply successful strategies from past executions. We present a novel framework for automatically extracting actionable learnings from agent execution trajectories and utilizing them to improve future performance through contextual memory retrieval. Our approach comprises four components: (1) a Trajectory Intelligence Extractor that performs semantic analysis of agent reasoning patterns, (2) a Decision Attribution Analyzer that identifies which decisions and reasoning steps led to failures, recoveries, or inefficiencies, (3) a Contextual Learning Generator that produces three types of guidance – strategy tips from successful patterns, recovery tips from failure handling, and optimization tips from inefficient but successful executions, and (4) an Adaptive Memory Retrieval System that injects relevant learnings into agent prompts based on multi-dimensional similarity. Unlike existing memory systems that store generic conversational facts, our framework understands execution patterns, extracts structured learnings with provenance, and retrieves guidance tailored to specific task contexts. Evaluation on the AppWorld benchmark demonstrates consistent improvements, with up to 14.3 percentage point gains in scenario goal completion on held-out tasks and particularly strong benefits on complex tasks (28.5~pp scenario goal improvement, a 149\% relative increase).

---

4. Does LLM Alignment Really Need Diversity? An Empirical Study of Adapting RLVR Methods for Moral Reasoning

• Authors: Zhaowei Zhang , Xiaohan Liu , Xuekai Zhu , Junchao Huang , Ceyao Zhang , Zhiyuan Feng , Yaodong Yang , Xiaoyuan Yi , Xing Xie
• URL: https://arxiv.org/abs/2603.10588
• Abstract:

  Reinforcement learning with verifiable rewards (RLVR) has achieved remarkable success in logical reasoning tasks, yet whether large language model (LLM) alignment requires fundamentally different approaches remains unclear. Given the apparent tolerance for multiple valid responses in moral reasoning, a natural hypothesis is that alignment tasks inherently require diversity-seeking distribution-matching algorithms rather than reward-maximizing policy-based methods. We conduct the first comprehensive empirical study comparing both paradigms on MoReBench. To enable stable RLVR training, we build a rubric-grounded reward pipeline by training a Qwen3-1.7B judge model. Contrary to our hypothesis, we find that distribution-matching approaches do not demonstrate significant advantages over reward-maximizing methods as expected on alignment tasks. Through semantic visualization mapping high-reward responses to semantic space, we demonstrate that moral reasoning exhibits more concentrated high-reward distributions than mathematical reasoning, where diverse solution strategies yield similarly high rewards. This counter-intuitive finding explains why mode-seeking optimization proves equally or more effective for alignment tasks. Our results suggest that alignment tasks do not inherently require diversity-preserving algorithms, and standard reward-maximizing RLVR methods can effectively transfer to moral reasoning without explicit diversity mechanisms.

---

5. CUAAudit: Meta-Evaluation of Vision-Language Models as Auditors of Autonomous Computer-Use Agents

• Authors: Marta Sumyk , Oleksandr Kosovan
• URL: https://arxiv.org/abs/2603.10577
• Abstract:

  Computer-Use Agents (CUAs) are emerging as a new paradigm in human-computer interaction, enabling autonomous execution of tasks in desktop environment by perceiving high-level natural-language instructions. As such agents become increasingly capable and are deployed across diverse desktop environments, evaluating their behavior in a scalable and reliable manner becomes a critical challenge. Existing evaluation pipelines rely on static benchmarks, rule-based success checks, or manual inspection, which are brittle, costly, and poorly aligned with real-world usage. In this work, we study Vision-Language Models (VLMs) as autonomous auditors for assessing CUA task completion directly from observable interactions and conduct a large-scale meta-evaluation of five VLMs that judge task success given a natural-language instruction and the final environment state. Our evaluation spans three widely used CUA benchmarks across macOS, Windows, and Linux environments and analyzes auditor behavior along three complementary dimensions: accuracy, calibration of confidence estimates, and inter-model agreement. We find that while state-of-the-art VLMs achieve strong accuracy and calibration, all auditors exhibit notable performance degradation in more complex or heterogeneous environments, and even high-performing models show significant disagreement in their judgments. These results expose fundamental limitations of current model-based auditing approaches and highlight the need to explicitly account for evaluator reliability, uncertainty, and variance when deploying autonomous CUAs in real-world settings.

---

6. Adaptive RAN Slicing Control via Reward-Free Self-Finetuning Agents

• Authors: Yuanhao Li , Haozhe Wang , Geyong Min , Nektarios Georgalas , Wang Miao
• URL: https://arxiv.org/abs/2603.10564
• Abstract:

  The integration of Generative AI models into AI-native network systems offers a transformative path toward achieving autonomous and adaptive control. However, the application of such models to continuous control tasks is impeded by intrinsic architectural limitations, including finite context windows, the lack of explicit reward signals, and the degradation of the long context. This paper posits that the key to unlocking robust continuous control is enabling agents to internalize experience by distilling it into their parameters, rather than relying on prompt-based memory. To this end, we propose a novel self-finetuning framework that enables agentic systems to learn continuously through direct interaction with the environment, bypassing the need for handcrafted rewards. Our framework implements a bi-perspective reflection mechanism that generates autonomous linguistic feedback to construct preference datasets from interaction history. A subsequent preference-based fine-tuning process distills long-horizon experiences into the model’s parameters. We evaluate our approach on a dynamic Radio Access Network (RAN) slicing task, a challenging multi-objective control problem that requires the resolution of acute trade-offs between spectrum efficiency, service quality, and reconfiguration stability under volatile network conditions. Experimental results show that our framework outperforms standard Reinforcement Learning (RL) baselines and existing Large Language Model (LLM)-based agents in sample efficiency, stability, and multi-metric optimization. These findings demonstrate the potential of self-improving generative agents for continuous control tasks, paving the way for future AI-native network infrastructure.

---

7. Resource-constrained Amazons chess decision framework integrating large language models and graph attention

• Authors: Tianhao Qian , Zhuoxuan Li , Jinde Cao , Xinli Shi , Hanjie Liu , Leszek Rutkowski
• URL: https://arxiv.org/abs/2603.10512
• Abstract:

  Artificial intelligence has advanced significantly through the development of intelligent game-playing systems, providing rigorous testbeds for decision-making, strategic planning, and adaptive learning. However, resource-constrained environments pose critical challenges, as conventional deep learning methods heavily rely on extensive datasets and computational resources. In this paper, we propose a lightweight hybrid framework for the Game of the Amazons, which explores the paradigm of weak-to-strong generalization by integrating the structural reasoning of graph-based learning with the generative capabilities of large language models. Specifically, we leverage a Graph Attention Autoencoder to inform a multi-step Monte Carlo Tree Search, utilize a Stochastic Graph Genetic Algorithm to optimize evaluation signals, and harness GPT-4o-mini to generate synthetic training data. Unlike traditional approaches that rely on expert demonstrations, our framework learns from noisy and imperfect supervision. We demonstrate that the Graph Attention mechanism effectively functions as a structural filter, denoising the LLM’s outputs. Experiments on a 10$\times$10 Amazons board show that our hybrid approach not only achieves a 15\%–56\% improvement in decision accuracy over baselines but also significantly outperforms its teacher model (GPT-4o-mini), achieving a competitive win rate of 45.0\% at N=30 nodes and a decisive 66.5\% at only N=50 nodes. These results verify the feasibility of evolving specialized, high-performance game AI from general-purpose foundation models under stringent computational constraints.

---

8. Verbalizing LLM’s Higher-order Uncertainty via Imprecise Probabilities

• Authors: Anita Yang , Krikamol Muandet , Michele Caprio , Siu Lun Chau , Masaki Adachi
• URL: https://arxiv.org/abs/2603.10396
• Abstract:

  Despite the growing demand for eliciting uncertainty from large language models (LLMs), empirical evidence suggests that LLM behavior is not always adequately captured by the elicitation techniques developed under the classical probabilistic uncertainty framework. This mismatch leads to systematic failure modes, particularly in settings that involve ambiguous question-answering, in-context learning, and self-reflection. To address this, we propose novel prompt-based uncertainty elicitation techniques grounded in \emph{imprecise probabilities}, a principled framework for repesenting and eliciting higher-order uncertainty. Here, first-order uncertainty captures uncertainty over possible responses to a prompt, while second-order uncertainty (uncertainty about uncertainty) quantifies indeterminacy in the underlying probability model itself. We introduce general-purpose prompting and post-processing procedures to directly elicit and quantify both orders of uncertainty, and demonstrate their effectiveness across diverse settings. Our approach enables more faithful uncertainty reporting from LLMs, improving credibility and supporting downstream decision-making.

---

9. Beyond Scalars: Evaluating and Understanding LLM Reasoning via Geometric Progress and Stability

• Authors: Xinyan Jiang , Ninghao Liu , Di Wang , Lijie Hu
• URL: https://arxiv.org/abs/2603.10384
• Abstract:

  Evaluating LLM reliability via scalar probabilities often fails to capture the structural dynamics of reasoning. We introduce TRACED, a framework that assesses reasoning quality through theoretically grounded geometric kinematics. By decomposing reasoning traces into Progress (displacement) and Stability (curvature), we reveal a distinct topological divergence: correct reasoning manifests as high-progress, stable trajectories, whereas hallucinations are characterized by low-progress, unstable patterns (stalled displacement with high curvature fluctuations). Leveraging these signatures, our probabilistic framework achieves competitive performance and superior robustness across diverse benchmarks. Crucially, TRACED bridges geometry and cognition by mapping high curvature to ‘‘Hesitation Loops’’ and displacement to ‘‘Certainty Accumulation’’, offering a physical lens to decode the internal dynamics of machine thought.

---

10. Hybrid Self-evolving Structured Memory for GUI Agents

• Authors: Sibo Zhu , Wenyi Wu , Kun Zhou , Stephen Wang , Biwei Huang
• URL: https://arxiv.org/abs/2603.10291
• Abstract:

  The remarkable progress of vision-language models (VLMs) has enabled GUI agents to interact with computers in a human-like manner. Yet real-world computer-use tasks remain difficult due to long-horizon workflows, diverse interfaces, and frequent intermediate errors. Prior work equips agents with external memory built from large collections of trajectories, but relies on flat retrieval over discrete summaries or continuous embeddings, falling short of the structured organization and self-evolving characteristics of human memory. Inspired by the brain, we propose Hybrid Self-evolving Structured Memory (HyMEM), a graph-based memory that couples discrete high-level symbolic nodes with continuous trajectory embeddings. HyMEM maintains a graph structure to support multi-hop retrieval, self-evolution via node update operations, and on-the-fly working-memory refreshing during inference. Extensive experiments show that HyMEM consistently improves open-source GUI agents, enabling 7B/8B backbones to match or surpass strong closed-source models; notably, it boosts Qwen2.5-VL-7B by +22.5% and outperforms Gemini2.5-Pro-Vision and GPT-4o.

---

11. COMIC: Agentic Sketch Comedy Generation

• Authors: Susung Hong , Brian Curless , Ira Kemelmacher-Shlizerman , Steve Seitz
• URL: https://arxiv.org/abs/2603.11048
• Abstract:

  We propose a fully automated AI system that produces short comedic videos similar to sketch shows such as Saturday Night Live. Starting with character references, the system employs a population of agents loosely based on real production studio roles, structured to optimize the quality and diversity of ideas and outputs through iterative competition, evaluation, and improvement. A key contribution is the introduction of LLM critics aligned with real viewer preferences through the analysis of a corpus of comedy videos on YouTube to automatically evaluate humor. Our experiments show that our framework produces results approaching the quality of professionally produced sketches while demonstrating state-of-the-art performance in video generation.

---

12. Does AI See like Art Historians? Interpreting How Vision Language Models Recognize Artistic Style

• Authors: Marvin Limpijankit , Milad Alshomary , Yassin Oulad Daoud , Amith Ananthram , Tim Trombley , Elias Stengel-Eskin , Mohit Bansal , Noam M. Elcott , Kathleen McKeown
• URL: https://arxiv.org/abs/2603.11024
• Abstract:

  VLMs have become increasingly proficient at a range of computer vision tasks, such as visual question answering and object detection. This includes increasingly strong capabilities in the domain of art, from analyzing artwork to generation of art. In an interdisciplinary collaboration between computer scientists and art historians, we characterize the mechanisms underlying VLMs’ ability to predict artistic style and assess the extent to which they align with the criteria art historians use to reason about artistic style. We employ a latent-space decomposition approach to identify concepts that drive art style prediction and conduct quantitative evaluations, causal analysis and assessment by art historians. Our findings indicate that 73% of the extracted concepts are judged by art historians to exhibit a coherent and semantically meaningful visual feature and 90% of concepts used to predict style of a given artwork were judged relevant. In cases where an irrelevant concept was used to successfully predict style, art historians identified possible reasons for its success; for example, the model might “understand” a concept in more formal terms, such as dark/light contrasts.

---

13. GroundCount: Grounding Vision-Language Models with Object Detection for Mitigating Counting Hallucinations

• Authors: Boyuan Chen , Minghao Shao , Siddharth Garg , Ramesh Karri , Muhammad Shafique
• URL: https://arxiv.org/abs/2603.10978
• Abstract:

  Vision Language Models (VLMs) exhibit persistent hallucinations in counting tasks, with accuracy substantially lower than other visual reasoning tasks (excluding sentiment). This phenomenon persists even in state-of-the-art reasoning-capable VLMs. Conversely, CNN-based object detection models (ODMs) such as YOLO excel at spatial localization and instance counting with minimal computational overhead. We propose GroundCount, a framework that augments VLMs with explicit spatial grounding from ODMs to mitigate counting hallucinations. In the best case, our prompt-based augmentation strategy achieves 81.3% counting accuracy on the best-performing model (Ovis2.5-2B) - a 6.6pp improvement - while reducing inference time by 22% through elimination of hallucination-driven reasoning loops for stronger models. We conduct comprehensive ablation studies demonstrating that positional encoding is a critical component, being beneficial for stronger models but detrimental for weaker ones. Confidence scores, by contrast, introduce noise for most architectures and their removal improves performance in four of five evaluated models. We further evaluate feature-level fusion architectures, finding that explicit symbolic grounding via structured prompts outperforms implicit feature fusion despite sophisticated cross-attention mechanisms. Our approach yields consistent improvements across four of five evaluated VLM architectures (6.2–7.5pp), with one architecture exhibiting degraded performance due to incompatibility between its iterative reflection mechanisms and structured prompts. These results suggest that counting failures stem from fundamental spatial-semantic integration limitations rather than architecture-specific deficiencies, while highlighting the importance of architectural compatibility in augmentation strategies.

---

14. When Fine-Tuning Fails and when it Generalises: Role of Data Diversity and Mixed Training in LLM-based TTS

• Authors: Anupam Purwar , Aditya Choudhary
• URL: https://arxiv.org/abs/2603.10904
• Abstract:

  Large language models are increasingly adopted as semantic backbones for neural text-to-speech systems. However, frozen LLM representations are insufficient for modeling speaker specific acoustic and perceptual characteristics. Our experiments involving fine tuning of the Language Model backbone of TTS show promise in improving the voice consistency and Signal to Noise ratio SNR in voice cloning task. Across multiple speakers LoRA finetuning consistently outperforms the non-finetuned base Qwen-0.5B model across three complementary dimensions of speech quality. First, perceptual quality improves significantly with DNS-MOS gains of up to 0.42 points for speakers whose training data exhibits sufficient acoustic variability. Second, speaker fidelity improves for all evaluated speakers with consistent increases in voice similarity indicating that LoRA effectively adapts speaker identity representations without degrading linguistic modeling. Third, signal level quality improves in most cases with signal to noise ratio increasing by as much as 34 percent. Crucially these improvements are strongly governed by the characteristics of the training data. Speakers with high variability in acoustic energy and perceptual quality achieve simultaneous gains in DNS-MOS voice similarity and SNR. Overall this work establishes that LoRA finetuning is not merely a parameter efficient optimization technique but an effective mechanism for better speaker level adaptation in compact LLM-based TTS systems. When supported by sufficiently diverse training data LoRA adapted Qwen-0.5B consistently surpasses its frozen base model in perceptual quality speaker similarity with low latency using GGUF model hosted in quantized form.

---

15. LookaheadKV: Fast and Accurate KV Cache Eviction by Glimpsing into the Future without Generation

• Authors: Jinwoo Ahn , Ingyu Seong , Akhil Kedia , Junhan Kim , Hyemi Jang , Kangwook Lee , Yongkweon Jeon
• URL: https://arxiv.org/abs/2603.10899
• Abstract:

  Transformer-based large language models (LLMs) rely on key-value (KV) caching to avoid redundant computation during autoregressive inference. While this mechanism greatly improves efficiency, the cache size grows linearly with the input sequence length, quickly becoming a bottleneck for long-context tasks. Existing solutions mitigate this problem by evicting prompt KV that are deemed unimportant, guided by estimated importance scores. Notably, a recent line of work proposes to improve eviction quality by “glimpsing into the future”, in which a draft generator produces a surrogate future response approximating the target model’s true response, and this surrogate is subsequently used to estimate the importance of cached KV more accurately. However, these approaches rely on computationally expensive draft generation, which introduces substantial prefilling overhead and limits their practicality in real-world deployment. To address this challenge, we propose LookaheadKV, a lightweight eviction framework that leverages the strength of surrogate future response without requiring explicit draft generation. LookaheadKV augments transformer layers with parameter-efficient modules trained to predict true importance scores with high accuracy. Our design ensures negligible runtime overhead comparable to existing inexpensive heuristics, while achieving accuracy superior to more costly approximation methods. Extensive experiments on long-context understanding benchmarks, across a wide range of models, demonstrate that our method not only outperforms recent competitive baselines in various long-context understanding tasks, but also reduces the eviction cost by up to 14.5x, leading to significantly faster time-to-first-token. Our code is available at this https URL .

---

16. Dynamics-Predictive Sampling for Active RL Finetuning of Large Reasoning Models

• Authors: Yixiu Mao , Yun Qu , Qi Wang , Heming Zou , Xiangyang Ji
• URL: https://arxiv.org/abs/2603.10887
• Abstract:

  Reinforcement learning (RL) finetuning has become a key technique for enhancing the reasoning abilities of large language models (LLMs). However, its effectiveness critically depends on the selection of training data. Recent advances underscore the importance of online prompt selection methods, which typically concentrate training on partially solved or moderately challenging examples under the current policy, thereby yielding more effective model updates. While significantly accelerating RL finetuning in terms of training steps, they also incur substantial computational overhead by requiring extensive LLM rollouts over large candidate batches to identify informative samples, an expense that can outweigh the finetuning process itself. To address this challenge, this work proposes Dynamics-Predictive Sampling (DPS), which online predicts and selects informative prompts by inferring their learning dynamics prior to costly rollouts. Specifically, we introduce a new perspective by modeling each prompt’s solving progress during RL finetuning as a dynamical system, where the extent of solving is represented as the state and the transition is characterized by a hidden Markov model. Using historical rollout reward signals, we perform online Bayesian inference to estimate evolving state distributions, and the inference outcome provides a predictive prior for efficient prompt selection without rollout-intensive filtering. Empirical results across diverse reasoning tasks, including mathematics, planning, and visual geometry, demonstrate that DPS substantially reduces redundant rollouts, accelerates the training process, and achieves superior reasoning performance.

---

17. Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to NPU Kernel Synthesis

• Authors: Yujie Zheng , Zhuo Li , Shengtao Zhang , Hanjing Wang , Junjie Sheng , Jiaqian Wang , Junchi Yan , Weinan Zhang , Ying Wen , Bo Tang , Muning Wen
• URL: https://arxiv.org/abs/2603.10846
• Abstract:

  Deploying Large Language Models to data-scarce programming domains poses significant challenges, particularly for kernel synthesis on emerging Domain-Specific Architectures where a “Data Wall” limits available training data. While models excel on data-rich platforms like CUDA, they suffer catastrophic performance drops on data-scarce ecosystems such as NPU programming. To overcome this cold-start barrier without expensive fine-tuning, we introduce EvoKernel, a self-evolving agentic framework that automates the lifecycle of kernel synthesis from initial drafting to continual refining. EvoKernel addresses this by formulating the synthesis process as a memory-based reinforcement learning task. Through a novel value-driven retrieval mechanism, it learns stage-specific Q-values that prioritize experiences based on their contribution to the current objective, whether bootstrapping a feasible draft or iteratively refining latency. Furthermore, by enabling cross-task memory sharing, the agent generalizes insights from simple to complex operators. By building an NPU variant of KernelBench and evaluating on it, EvoKernel improves frontier models’ correctness from 11.0% to 83.0% and achieves a median speedup of 3.60x over initial drafts through iterative refinement. This demonstrates that value-guided experience accumulation allows general-purpose models to master the kernel synthesis task on niche hardware ecosystems. Our official page is available at this https URL .

---

18. Speaker Verification with Speech-Aware LLMs: Evaluation and Augmentation

• Authors: Thomas Thebaud , Yuzhe Wang , Laureano Moro-Velazquez , Jesus Villalba-Lopez , Najim Dehak
• URL: https://arxiv.org/abs/2603.10827
• Abstract:

  Speech-aware large language models (LLMs) can accept speech inputs, yet their training objectives largely emphasize linguistic content or specific fields such as emotions or the speaker’s gender, leaving it unclear whether they encode speaker identity. First, we propose a model-agnostic scoring protocol that produces continuous verification scores for both API-only and open-weight models, using confidence scores or log-likelihood ratios from the Yes/No token probabilities. Using this protocol, we benchmark recent speech-aware LLMs and observe weak speaker discrimination (EERs above 20% on VoxCeleb1). Second, we introduce a lightweight augmentation that equips an LLM with ASV capability by injecting frozen ECAPA-TDNN speaker embeddings through a learned projection and training only LoRA adapters. On TinyLLaMA-1.1B, the resulting ECAPA-LLM achieves 1.03% EER on VoxCeleb1-E, approaching a dedicated speaker verification system while preserving a natural-language interface.

---

19. Risk-Adjusted Harm Scoring for Automated Red Teaming for LLMs in Financial Services

• Authors: Fabrizio Dimino , Bhaskarjit Sarmah , Stefano Pasquali
• URL: https://arxiv.org/abs/2603.10807
• Abstract:

  The rapid adoption of large language models (LLMs) in financial services introduces new operational, regulatory, and security risks. Yet most red-teaming benchmarks remain domain-agnostic and fail to capture failure modes specific to regulated BFSI settings, where harmful behavior can be elicited through legally or professionally plausible framing. We propose a risk-aware evaluation framework for LLM security failures in Banking, Financial Services, and Insurance (BFSI), combining a domain-specific taxonomy of financial harms, an automated multi-round red-teaming pipeline, and an ensemble-based judging protocol. We introduce the Risk-Adjusted Harm Score (RAHS), a risk-sensitive metric that goes beyond success rates by quantifying the operational severity of disclosures, accounting for mitigation signals, and leveraging inter-judge agreement. Across diverse models, we find that higher decoding stochasticity and sustained adaptive interaction not only increase jailbreak success, but also drive systematic escalation toward more severe and operationally actionable financial disclosures. These results expose limitations of single-turn, domain-agnostic security evaluation and motivate risk-sensitive assessment under prolonged adversarial pressure for real-world BFSI deployment.

---

20. Taking Shortcuts for Categorical VQA Using Super Neurons

• Authors: Pierre Musacchio , Jaeyi Jeong , Dahun Kim , Jaesik Park
• URL: https://arxiv.org/abs/2603.10781
• Abstract:

  Sparse Attention Vectors (SAVs) have emerged as an excellent training-free alternative to supervised finetuning or low-rank adaptation to improve the performance of Vision Language Models (VLMs). At their heart, SAVs select a few accurate attention heads for a task of interest and use them as classifiers, rather than relying on the model’s prediction. In a similar spirit, we find that directly probing the raw activations of the VLM, in the form of scalar values, is sufficient to yield accurate classifiers on diverse visually grounded downstream tasks. Shifting focus from attention vectors to scalar activations dramatically increases the search space for accurate parameters, allowing us to find more discriminative neurons immediately from the first generated token. We call such activations Super Neurons (SNs). In this probing setting, we discover that enough SNs appear in the shallower layers of the large language model to allow for extreme early exiting from the first layer of the model at the first generated token. Compared to the original network, SNs robustly improve the classification performance while achieving a speedup of up to 5.10x.

---

21. Towards Robust Speech Deepfake Detection via Human-Inspired Reasoning

• Authors: Artem Dvirniak , Evgeny Kushnir , Dmitrii Tarasov , Artem Iudin , Oleg Kiriukhin , Mikhail Pautov , Dmitrii Korzh , Oleg Y. Rogov
• URL: https://arxiv.org/abs/2603.10725
• Abstract:

  The modern generative audio models can be used by an adversary in an unlawful manner, specifically, to impersonate other people to gain access to private information. To mitigate this issue, speech deepfake detection (SDD) methods started to evolve. Unfortunately, current SDD methods generally suffer from the lack of generalization to new audio domains and generators. More than that, they lack interpretability, especially human-like reasoning that would naturally explain the attribution of a given audio to the bona fide or spoof class and provide human-perceptible cues. In this paper, we propose HIR-SDD, a novel SDD framework that combines the strengths of Large Audio Language Models (LALMs) with the chain-of-thought reasoning derived from the novel proposed human-annotated dataset. Experimental evaluation demonstrates both the effectiveness of the proposed method and its ability to provide reasonable justifications for predictions.

---

22. EvoSchema: Towards Text-to-SQL Robustness Against Schema Evolution

• Authors: Tianshu Zhang , Kun Qian , Siddhartha Sahai , Yuan Tian , Shaddy Garg , Huan Sun , Yunyao Li
• URL: https://arxiv.org/abs/2603.10697
• Abstract:

  Neural text-to-SQL models, which translate natural language questions (NLQs) into SQL queries given a database schema, have achieved remarkable performance. However, database schemas frequently evolve to meet new requirements. Such schema evolution often leads to performance degradation for models trained on static schemas. Existing work either mainly focuses on simply paraphrasing some syntactic or semantic mappings among NLQ, DB and SQL, or lacks a comprehensive and controllable way to investigate the model robustness issue under the schema evolution, which is insufficient when facing the increasingly complex and rich database schema changes in reality, especially in the LLM era. To address the challenges posed by schema evolution, we present EvoSchema, a comprehensive benchmark designed to assess and enhance the robustness of text-to-SQL systems under real-world schema changes. EvoSchema introduces a novel schema evolution taxonomy, encompassing ten perturbation types across columnlevel and table-level modifications, systematically simulating the dynamic nature of database schemas. Through EvoSchema, we conduct an in-depth evaluation spanning different open source and closed-source LLMs, revealing that table-level perturbations have a significantly greater impact on model performance compared to column-level changes. Furthermore, EvoSchema inspires the development of more resilient text-to-SQL systems, in terms of both model training and database design. The models trained on EvoSchema’s diverse schema designs can force the model to distinguish the schema difference for the same questions to avoid learning spurious patterns, which demonstrate remarkable robustness compared to those trained on unperturbed data on average. This benchmark offers valuable insights into model behavior and a path forward for designing systems capable of thriving in dynamic, real-world environments.

---

23. Are Video Reasoning Models Ready to Go Outside?

• Authors: Yangfan He , Changgyu Boo , Jaehong Yoon
• URL: https://arxiv.org/abs/2603.10652
• Abstract:

  In real-world deployment, vision-language models often encounter disturbances such as weather, occlusion, and camera motion. Under such conditions, their understanding and reasoning degrade substantially, revealing a gap between clean, controlled (i.e., unperturbed) evaluation settings and real-world robustness. To address this limitation, we propose ROVA, a novel training framework that improves robustness by modeling a robustness-aware consistency reward under spatio-temporal corruptions. ROVA introduces a difficulty-aware online training strategy that prioritizes informative samples based on the model’s evolving capability. Specifically, it continuously re-estimates sample difficulty via self-reflective evaluation, enabling adaptive training with a robustness-aware consistency reward. We also introduce PVRBench, a new benchmark that injects real-world perturbations into embodied video datasets to assess both accuracy and reasoning quality under realistic disturbances. We evaluate ROVA and baselines on PVRBench, UrbanVideo, and VisBench, where open-source and proprietary models suffer up to 35% and 28% drops in accuracy and reasoning under realistic perturbations. ROVA effectively mitigates performance degradation, boosting relative accuracy by at least 24% and reasoning by over 9% compared with baseline models (QWen2.5/3-VL, InternVL2.5, Embodied-R). These gains transfer to clean standard benchmarks, yielding consistent improvements.

---

24. Reinforcement Learning with Conditional Expectation Reward

• Authors: Changyi Xiao , Caijun Xu , Yixin Cao
• URL: https://arxiv.org/abs/2603.10624
• Abstract:

  Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective in enhancing the reasoning capabilities of large language models, particularly in domains such as mathematics where reliable rule-based verifiers can be constructed. However, the reliance on handcrafted, domain-specific verification rules substantially limits the applicability of RLVR to general reasoning domains with free-form answers, where valid answers often exhibit significant variability, making it difficult to establish complete and accurate rules. To address this limitation, we propose Conditional Expectation Reward (CER), which leverages the large language model itself as an implicit verifier, and is therefore applicable to general domains and eliminates the need for external verifiers or auxiliary models. CER is defined as the expected likelihood of generating the reference answer conditioned on the generated answer. In contrast to rule-based verifiers that yield binary feedback, CER provides a soft, graded reward signal that reflects varying degrees of correctness, making it better suited to tasks where answers vary in correctness. Experimental results demonstrate that CER is effective across a wide range of reasoning tasks, spanning both mathematical and general domains, indicating that CER serves as a flexible and general verification mechanism. The code is available at this https URL .

---

25. Towards Cognitive Defect Analysis in Active Infrared Thermography with Vision-Text Cues

• Authors: Mohammed Salah , Eman Ouda , Giuseppe Dell’Avvocato , Fabrizio Sarasini , Ester D’Accardi , Jorge Dias , Davor Svetinovic , Stefano Sfarra , Yusra Abdulrahman
• URL: https://arxiv.org/abs/2603.10549
• Abstract:

  Active infrared thermography (AIRT) is currently witnessing a surge of artificial intelligence (AI) methodologies being deployed for automated subsurface defect analysis of high performance carbon fiber-reinforced polymers (CFRP). Deploying AI-based AIRT methodologies for inspecting CFRPs requires the creation of time consuming and expensive datasets of CFRP inspection sequences to train neural networks. To address this challenge, this work introduces a novel language-guided framework for cognitive defect analysis in CFRPs using AIRT and vision-language models (VLMs). Unlike conventional learning-based approaches, the proposed framework does not require developing training datasets for extensive training of defect detectors, instead it relies solely on pretrained multimodal VLM encoders coupled with a lightweight adapter to enable generative zero-shot understanding and localization of subsurface defects. By leveraging pretrained multimodal encoders, the proposed system enables generative zero-shot understanding of thermographic patterns and automatic detection of subsurface defects. Given the domain gap between thermographic data and natural images used to train VLMs, an AIRT-VLM Adapter is proposed to enhance the visibility of defects while aligning the thermographic domain with the learned representations of VLMs. The proposed framework is validated using three representative VLMs; specifically, GroundingDINO, Qwen-VL-Chat, and CogVLM. Validation is performed on 25 CFRP inspection sequences with impacts introduced at different energy levels, reflecting realistic defects encountered in industrial scenarios. Experimental results demonstrate that the AIRT-VLM adapter achieves signal-to-noise ratio (SNR) gains exceeding 10 dB compared with conventional thermographic dimensionality-reduction methods, while enabling zero-shot defect detection with intersection-over-union values reaching 70%.

---

26. SCORE: Replacing Layer Stacking with Contractive Recurrent Depth

• Authors: Guillaume Godin
• URL: https://arxiv.org/abs/2603.10544
• Abstract:

  Residual connections are central to modern deep neural networks, enabling stable optimization and efficient information flow across depth. In this work, we propose SCORE (Skip-Connection ODE Recurrent Embedding), a discrete recurrent alternative to classical layer stacking. Instead of composing multiple independent layers, SCORE iteratively applies a single shared neural block using an ODE (Ordinary Differential Equation)-inspired contractive update: ht+1 = (1 - dt) * ht + dt * F(ht) This formulation can be interpreted as a depth-by-iteration refinement process, where the step size dt explicitly controls stability and update magnitude. Unlike continuous Neural ODE approaches, SCORE uses a fixed number of discrete iterations and standard backpropagation without requiring ODE solvers or adjoint methods. We evaluate SCORE across graph neural networks (ESOL molecular solubility), multilayer perceptrons, and Transformer-based language models (nanoGPT). Across architectures, SCORE generally improves convergence speed and often accelerates training. SCORE is reducing parameter count through shared weights. In practice, simple Euler integration provides the best trade-off between computational cost and performance, while higher-order integrators yield marginal gains at increased compute. These results suggest that controlled recurrent depth with contractive residual updates offers a lightweight and effective alternative to classical stacking in deep neural networks.

---

27. Learning to Negotiate: Multi-Agent Deliberation for Collective Value Alignment in LLMs

• Authors: Panatchakorn Anantaprayoon , Nataliia Babina , Nima Asgharbeygi , Jad Tarifi
• URL: https://arxiv.org/abs/2603.10476
• Abstract:

  The alignment of large language models (LLMs) has progressed substantially in single-agent settings through paradigms such as RLHF and Constitutional AI, with recent work exploring scalable alternatives such as RLAIF and evolving alignment objectives. However, these approaches remain limited in multi-stakeholder settings, where conflicting values arise and deliberative negotiation capabilities are required. This work proposes a multi-agent negotiation-based alignment framework that aligns LLMs to Collective Agency (CA)-an existing alignment objective introduced to promote the continual expansion of agency-while simultaneously improving conflict-resolution capability. To enable scalable training, two self-play instances of the same LLM, assigned opposing personas, engage in structured turn-based dialogue to synthesize mutually beneficial solutions. We generate synthetic moral-dilemma prompts and conflicting persona pairs, and optimize the policy via RLAIF using GRPO with an external LLM reward model. While rewards are computed from CA scores assigned to the final completion, gradients are applied to dialogue tokens to directly improve deliberative interaction dynamics. Experiments show that the resulting model achieves CA alignment comparable to a single-agent baseline while substantially improving conflict-resolution performance without degrading general language capabilities. These results suggest that negotiation-driven deliberation training provides a practical path toward LLMs that better support collective decision-making in value-conflict scenarios.

---

28. Aligning Large Language Models with Searcher Preferences

• Authors: Wei Wu , Peilun Zhou , Liyi Chen , Qimeng Wang , Chengqiang Lu , Yan Gao , Yi Wu , Yao Hu , Hui Xiong
• URL: https://arxiv.org/abs/2603.10473
• Abstract:

  The paradigm shift from item-centric ranking to answer-centric synthesis is redefining the role of search engines. While recent industrial progress has applied generative techniques to closed-set item ranking in e-commerce, research and deployment of open-ended generative search on large content platforms remain limited. This setting introduces challenges, including robustness to noisy retrieval, non-negotiable safety guarantees, and alignment with diverse user needs. In this work, we introduce SearchLLM, the first large language model (LLM) for open-ended generative search. We design a hierarchical, multi-dimensional reward system that separates bottom-line constraints, including factual grounding, basic answer quality and format compliance, from behavior optimization objectives that promote robustness to noisy retrieval and alignment with user needs. Concretely, our reward model evaluates responses conditioned on the user query, session history, and retrieved evidence set, combining rule-based checks with human-calibrated LLM judges to produce an interpretable score vector over these dimensions. We introduce a Gated Aggregation Strategy to derive the training reward for optimizing SearchLLM with Group Relative Policy Optimization (GRPO). We deploy SearchLLM in the AI search entry of RedNote. Offline evaluations and online A/B tests show improved generation quality and user engagement, increasing Valid Consumption Rate by 1.03% and reducing Re-search Rate by 2.81%, while upholding strict safety and reliability standards.

---

29. G-STAR: End-to-End Global Speaker-Tracking Attributed Recognition

• Authors: Jing Peng , Ziyi Chen , Haoyu Li , Yucheng Wang , Duo Ma , Mengtian Li , Yunfan Du , Dezhu Xu , Kai Yu , Shuai Wang
• URL: https://arxiv.org/abs/2603.10468
• Abstract:

  We study timestamped speaker-attributed ASR for long-form, multi-party speech with overlap, where chunk-wise inference must preserve meeting-level speaker identity consistency while producing time-stamped, speaker-labeled transcripts. Previous Speech-LLM systems tend to prioritize either local diarization or global labeling, but often lack the ability to capture fine-grained temporal boundaries or robust cross-chunk identity linking. We propose G-STAR, an end-to-end system that couples a time-aware speaker-tracking module with a Speech-LLM transcription backbone. The tracker provides structured speaker cues with temporal grounding, and the LLM generates attributed text conditioned on these cues. G-STAR supports both component-wise optimization and joint end-to-end training, enabling flexible learning under heterogeneous supervision and domain shift. Experiments analyze cue fusion, local versus long-context trade-offs and hierarchical objectives.

---

30. The Curse and Blessing of Mean Bias in FP4-Quantized LLM Training

• Authors: Hengjie Cao , Zhendong Huang , Mengyi Chen , Yifeng Yang , Fanqi Yu , Ruijun Huang , Fang Dong , Xin Zhang , Jixian Zhou , Anrui Chen , Mingzhi Dong , Yujiang Wang , Jinlong Hou , Qin Lv , Yuan Cheng , Tun Lu , Fan Yang , Li Shang
• URL: https://arxiv.org/abs/2603.10444
• Abstract:

  Large language models trained on natural language exhibit pronounced anisotropy: a small number of directions concentrate disproportionate energy, while the remaining dimensions form a broad semantic tail. In low-bit training regimes, this geometry becomes numerically unstable. Because blockwise quantization scales are determined by extreme elementwise magnitudes, dominant directions stretch the dynamic range, compressing long-tail semantic variation into narrow numerical bins. We show that this instability is primarily driven by a coherent rank-one mean bias, which constitutes the dominant component of spectral anisotropy in LLM representations. This mean component emerges systematically across layers and training stages and accounts for the majority of extreme activation magnitudes, making it the principal driver of dynamic-range inflation under low precision. Crucially, because the dominant instability is rank-one, it can be eliminated through a simple source-level mean-subtraction operation. This bias-centric conditioning recovers most of the stability benefits of SVD-based spectral methods while requiring only reduction operations and standard quantization kernels. Empirical results on FP4 (W4A4G4) training show that mean removal substantially narrows the loss gap to BF16 and restores downstream performance, providing a hardware-efficient path to stable low-bit LLM training.

---

31. Designing Service Systems from Textual Evidence

• Authors: Ruicheng Ao , Hongyu Chen , Siyang Gao , Hanwei Li , David Simchi-Levi
• URL: https://arxiv.org/abs/2603.10400
• Abstract:

  Designing service systems requires selecting among alternative configurations – choosing the best chatbot variant, the optimal routing policy, or the most effective quality control procedure. In many service systems, the primary evidence of performance quality is textual – customer support transcripts, complaint narratives, compliance review reports – rather than the scalar measurements assumed by classical optimization methods. Large language models (LLMs) can read such textual evidence and produce standardized quality scores, but these automated judges exhibit systematic biases that vary across alternatives and evaluation instances. Human expert review remains accurate but costly. We study how to identify the best service configuration with high confidence while minimizing expensive human audits, given that automated evaluation is cheap but biased. We formalize this as a sequential decision problem where a biased proxy score is observed for every evaluation, and a verified outcome can be acquired selectively at additional cost. We prove that LLM-only selection fails under arm-dependent bias, and that naive selective-audit estimators can be asymptotically biased. We develop an estimator combining proxy scores with inverse-propensity-weighted residuals and construct anytime-valid confidence sequences. Our algorithm, PP-LUCB, jointly decides which alternatives to evaluate and whether to request human audits, concentrating reviews where the LLM judge is least reliable. We prove correctness and establish instance-dependent cost bounds showing near-optimal efficiency. On a customer support ticket classification task, our algorithm correctly identifies the best model in 40/40 trials while achieving 90\% audit cost reduction.

---

32. Causal Concept Graphs in LLM Latent Space for Stepwise Reasoning

• Authors: Md Muntaqim Meherab , Noor Islam S. Mohammad , Faiza Feroz
• URL: https://arxiv.org/abs/2603.10377
• Abstract:

  Sparse autoencoders can localize where concepts live in language models, but not how they interact during multi-step reasoning. We propose Causal Concept Graphs (CCG): a directed acyclic graph over sparse, interpretable latent features, where edges capture learned causal dependencies between concepts. We combine task-conditioned sparse autoencoders for concept discovery with DAGMA-style differentiable structure learning for graph recovery and introduce the Causal Fidelity Score (CFS) to evaluate whether graph-guided interventions induce larger downstream effects than random ones. On ARC-Challenge, StrategyQA, and LogiQA with GPT-2 Medium, across five seeds ($n{=}15$ paired runs), CCG achieves $\CFS=5.654\pm0.625$, outperforming ROME-style tracing ($3.382\pm0.233$), SAE-only ranking ($2.479\pm0.196$), and a random baseline ($1.032\pm0.034$), with $p<0.0001$ after Bonferroni correction. Learned graphs are sparse (5-6\% edge density), domain-specific, and stable across seeds.

---

33. Utility Function is All You Need: LLM-based Congestion Control

• Authors: Neta Rozen-Schiff , Liron Schiff , Stefan Schmid
• URL: https://arxiv.org/abs/2603.10357
• Abstract:

  Congestion is a critical and challenging problem in communication networks. Congestion control protocols allow network applications to tune their sending rate in a way that optimizes their performance and the network utilization. In the common distributed setting, the applications cannot collaborate with each other directly but instead obtain similar estimations about the state of the network using latency and loss measurements. These measurements can be fed into analytical functions, referred to by utility functions, whose gradients help each and all distributed senders to converge to a desired state. The above process becomes extremely complicated when each application has different optimization goals and requirements. Crafting these utilization functions has been a research subject for over a decade, with small incremental changes requiring rigorous mathematical analysis as well as real-world experiments. In this work, we present GenCC, a framework leveraging the code generation capabilities of large language models (LLMs) coupled with realistic network testbed, to design congestion control utility functions. Using GenCC, we analyze the impact of different guidance strategies on the performance of the generated protocols, considering application-specific requirements and network capacity. Our results show that LLMs, guided by either a generative code evolution strategy or mathematical chain-of-thought (CoT), can obtain close to optimal results, improving state-of-the-art congestion control protocols by 37%-142%, depending on the scenario.

---

34. Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck

• Authors: Hongbin Zhang , Kehai Chen , Xuefen Bai , Youcheng Pan , Yang Xiang , Jinpeng Wang , Min Zhang
• URL: https://arxiv.org/abs/2603.10351
• Abstract:

  Large language models (LLMs) have become a standard for multilingual evaluation, yet they exhibit a severe systematic translationese bias. In this paper, translationese bias is characterized as LLMs systematically favoring machine-translated text over human-authored references, particularly in low-resource languages. We attribute this bias to spurious correlations with (i) latent manifold alignment with English and (ii) cross-lingual predictability. To mitigate this bias, we propose DIBJudge, a robust fine-tuning framework that learns a minimally sufficient, judgment-critical representation via variational information compression, while explicitly isolating spurious factors into the dedicated bias branch. Furthermore, we incorporate a cross-covariance penalty that explicitly suppresses statistical dependence between robust and bias representations, thereby encouraging effective disentanglement. Extensive evaluations on multilingual reward modeling benchmarks and a dedicated translationese bias evaluation suite demonstrate that the proposed DIBJudge consistently outperforms strong baselines and substantially mitigates translationese bias.

---

35. Is this Idea Novel? An Automated Benchmark for Judgment of Research Ideas

• Authors: Tim Schopf , Michael Färber
• URL: https://arxiv.org/abs/2603.10303
• Abstract:

  Judging the novelty of research ideas is crucial for advancing science, enabling the identification of unexplored directions, and ensuring contributions meaningfully extend existing knowledge rather than reiterate minor variations. However, given the exponential growth of scientific literature, manually judging the novelty of research ideas through literature reviews is labor-intensive, subjective, and infeasible at scale. Therefore, recent efforts have proposed automated approaches for research idea novelty judgment. Yet, evaluation of these approaches remains largely inconsistent and is typically based on non-standardized human evaluations, hindering large-scale, comparable evaluations. To address this, we introduce RINoBench, the first comprehensive benchmark for large-scale evaluation of research idea novelty judgments. It comprises 1,381 research ideas derived from and judged by human experts as well as nine automated evaluation metrics designed to assess both rubric-based novelty scores and textual justifications of novelty judgments. Using this benchmark, we evaluate several state-of-the-art large language models (LLMs) on their ability to judge the novelty of research ideas. Our findings reveal that while LLM-generated reasoning closely mirrors human rationales, this alignment does not reliably translate into accurate novelty judgments, which diverge significantly from human gold standard judgments - even among leading reasoning-capable models. Data and code available at: this https URL .

---

36. Simulation-in-the-Reasoning (SiR): A Conceptual Framework for Empirically Grounded AI in Autonomous Transportation

• Authors: Wuping Xin
• URL: https://arxiv.org/abs/2603.10294
• Abstract:

  Large Language Models (LLMs) have advanced reasoning through techniques like Chain-of-Thought (CoT). However, their reasoning largely re-mains textual and hypothetical, lacking empirical grounding in complex, dynamic domains like transportation. This paper introduces Simulation-in-the-Reasoning (SiR), a novel conceptual framework that embeds domain-specific simulators directly into the LLM reasoning loop. By treating intermediate reasoning steps as executable simulation experiments, SiR transforms LLM reasoning from narrative plausibility into a falsifiable, hypothesis-simulate-analyze workflow. We discuss applications, where LLM can formulate Intelligent Transport System (ITS) strategy hypotheses, invoke a traffic simulator via the Model Context Protocol (MCP), evaluate results under different demand patterns, and refine strategies through verification and aggregation. While implementing the framework is part of our ongoing work, this paper primarily establishes the conceptual foundation, discusses design considerations like API granularity, and outlines the vision of SiR as a cornerstone for interactive transportation digital twins. We argue that SiR represents a critical step towards trustworthy, empirically-validated AI for autonomous transportation systems.

---

37. Conversational AI-Enhanced Exploration System to Query Large-Scale Digitised Collections of Natural History Museums

• Authors: Yiyuan Wang , Andrew Johnston , Zoë Sadokierski , Rhiannon Stephens , Shane T. Ahyong
• URL: https://arxiv.org/abs/2603.10285
• Abstract:

  Recent digitisation efforts in natural history museums have produced large volumes of collection data, yet their scale and scientific complexity often hinder public access and understanding. Conventional data management tools, such as databases, restrict exploration through keyword-based search or require specialised schema knowledge. This paper presents a system design that uses conversational AI to query nearly 1.7 million digitised specimen records from the life-science collections of the Australian Museum. Designed and developed through a human-centred design process, the system contains an interactive map for visual-spatial exploration and a natural-language conversational agent that retrieves detailed specimen data and answers collection-specific questions. The system leverages function-calling capabilities of contemporary large language models to dynamically retrieve structured data from external APIs, enabling fast, real-time interaction with extensive yet frequently updated datasets. Our work provides a new approach of connecting large museum collections with natural language-based queries and informs future designs of scientific AI agents for natural history museums.

---

38. DUCTILE: Agentic LLM Orchestration of Engineering Analysis in Product Development Practice

• Authors: Alejandro Pradas-Gomez , Arindam Brahma , Ola Isaksson
• URL: https://arxiv.org/abs/2603.10249
• Abstract:

  Engineering analysis automation in product development relies on rigid interfaces between tools, data formats and documented processes. When these interfaces change, as they routinely do as the product evolves in the engineering ecosystem, the automation support breaks. This paper presents a DUCTILE (Delegated, User-supervised Coordination of Tool- and document-Integrated LLM-Enabled) agentic orchestration, an approach for developing, executing and evaluating LLM-based agentic automation support of engineering analysis tasks. The approach separates adaptive orchestration, performed by the LLM agent, from deterministic execution, performed by verified engineering tools. The agent interprets documented design practices, inspects input data and adapts the processing path, while the engineer supervises and exercises final judgment. DUCTILE is demonstrated on an industrial structural analysis task at an aerospace manufacturer, where the agent handled input deviations in format, units, naming conventions and methodology that would break traditional scripted pipelines. Evaluation against expert-defined acceptance criteria and deployment with practicing engineers confirm that the approach produces correct, methodologically compliant results across repeated independent runs. The paper discusses practical consequences of adopting agentic automation, including unintended effects on the nature of engineering work and the tension between removing mundane tasks and creating an exhausting supervisory role.

---

39. Rethinking the Harmonic Loss via Non-Euclidean Distance Layers

• Authors: Maxwell Miller-Golub , Kamil Faber , Marcin Pietron , Panpan Zheng , Pasquale Minervini , Roberto Corizzo
• URL: https://arxiv.org/abs/2603.10225
• Abstract:

  Cross-entropy loss has long been the standard choice for training deep neural networks, yet it suffers from interpretability limitations, unbounded weight growth, and inefficiencies that can contribute to costly training dynamics. The harmonic loss is a distance-based alternative grounded in Euclidean geometry that improves interpretability and mitigates phenomena such as grokking, or delayed generalization on the test set. However, the study of harmonic loss remains narrow: only Euclidean distance is explored, and no systematic evaluation of computational efficiency or sustainability was conducted. We extend harmonic loss by systematically investigating a broad spectrum of distance metrics as replacements for the Euclidean distance. We comprehensively evaluate distance-tailored harmonic losses on both vision backbones and large language models. Our analysis is framed around a three-way evaluation of model performance, interpretability, and sustainability. On vision tasks, cosine distances provide the most favorable trade-off, consistently improving accuracy while lowering carbon emissions, whereas Bray-Curtis and Mahalanobis further enhance interpretability at varying efficiency costs. On language models, cosine-based harmonic losses improve gradient and learning stability, strengthen representation structure, and reduce emissions relative to cross-entropy and Euclidean heads. Our code is available at: this https URL .

---

40. Delta-K: Boosting Multi-Instance Generation via Cross-Attention Augmentation

• Authors: Zitong Wang , Zijun Shen , Haohao Xu , Zhengjie Luo , Weibin Wu
• URL: https://arxiv.org/abs/2603.10210
• Abstract:

  While Diffusion Models excel in text-to-image synthesis, they often suffer from concept omission when synthesizing complex multi-instance scenes. Existing training-free methods attempt to resolve this by rescaling attention maps, which merely exacerbates unstructured noise without establishing coherent semantic representations. To address this, we propose Delta-K, a backbone-agnostic and plug-and-play inference framework that tackles omission by operating directly in the shared cross-attention Key space. Specifically, with Vision-language model, we extract a differential key $\Delta K$ that encodes the semantic signature of missing concepts. This signal is then injected during the early semantic planning stage of the diffusion process. Governed by a dynamically optimized scheduling mechanism, Delta-K grounds diffuse noise into stable structural anchors while preserving existing concepts. Extensive experiments demonstrate the generality of our approach: Delta-K consistently improves compositional alignment across both modern DiT models and classical U-Net architectures, without requiring spatial masks, additional training, or architectural modifications.

---

41. Adaptive Activation Cancellation for Hallucination Mitigation in Large Language Models

• Authors: Eric Yocam , Varghese Vaidyan , Gurcan Comert , Paris Kalathas , Yong Wang , Judith L. Mwakalonge
• URL: https://arxiv.org/abs/2603.10195
• Abstract:

  Large Language Models frequently generate fluent but factually incorrect text. We propose Adaptive Activation Cancellation (AAC), a real-time inference-time framework that treats hallucination-associated neural activations as structured interference within the transformer residual stream, drawing an explicit analogy to classical adaptive noise cancellation from signal processing. The framework identifies Hallucination Nodes (H-Nodes) via layer-wise linear probing and suppresses them using a confidence-weighted forward hook during auto-regressive generation – requiring no external knowledge, no fine-tuning, and no additional inference passes. Evaluated across OPT-125M, Phi-3-mini, and LLaMA 3-8B on TruthfulQA and HaluEval, the real-time hook is the only intervention that consistently improves downstream accuracy on all three scales. Critically, the method is strictly surgical: WikiText-103 perplexity and MMLU reasoning accuracy are preserved at exactly 0.0% degradation across all three model scales, a property that distinguishes AAC from interventions that trade fluency or general capability for factual improvement. On the LLaMA 3-8B scale, the hook additionally yields positive generation-level gains (MC1 +0.04; MC2 +0.003; Token-F1 +0.003) while achieving probe-space selectivity 5.94x - 3.5x higher than the ITI baseline – demonstrating that targeted neuron-level suppression can simultaneously improve factual accuracy and preserve model capability.

---

42. MCP-in-SoS: Risk assessment framework for open-source MCP servers

• Authors: Pratyay Kumar , Miguel Antonio Guirao Aguilera , Srikathyayani Srikanteswara , Satyajayant Misra , Abu Saleh Md Tayeen
• URL: https://arxiv.org/abs/2603.10194
• Abstract:

  Model Context Protocol (MCP) servers have rapidly emerged over the past year as a widely adopted way to enable Large Language Model (LLM) agents to access dynamic, real-world tools. As MCP servers proliferate and become easy to adopt via open-source releases, understanding their security risks becomes essential for dependable production agent deployments. Recent work has developed MCP threat taxonomies, proposed mitigations, and demonstrated practical attacks. However, to the best of our knowledge, no prior study has conducted a systematic, large-scale assessment of weaknesses in open-source MCP servers. Motivated by this gap, we apply static code analysis to identify Common Weakness Enumeration (CWE) weaknesses and map them to common attack patterns and threat categories using the MITRE Common Attack Pattern Enumerations and Classifications (CAPEC) to ground risk in real-world threats. We then introduce a risk-assessment framework for the MCP landscape that combines these threats using a multi-metric scoring of likelihood and impact. Our findings show that many open-source MCP servers contain exploitable weaknesses that can compromise confidentiality, integrity, and availability, underscoring the need for secure-by-design MCP server development.

---

43. Compatibility at a Cost: Systematic Discovery and Exploitation of MCP Clause-Compliance Vulnerabilities

• Authors: Nanzi Yang , Weiheng Bai , Kangjie Lu
• URL: https://arxiv.org/abs/2603.10163
• Abstract:

  The Model Context Protocol (MCP) is a recently proposed interoperability standard that unifies how AI agents connect with external tools and data sources. By defining a set of common client-server message exchange clauses, MCP replaces fragmented integrations with a standardized, plug-and-play framework. However, to be compatible with diverse AI agents, the MCP specification relaxes many behavioral constraints into optional clauses, leading to misuse-prone SDK implementation. We identify it as a new attack surface that allows adversaries to achieve multiple attacks (e.g, silent prompt injection, DoS, etc.), named as \emph{compatibility-abusing attacks}. In this work, we present the first systematic framework for analyzing this new attack surface across multi-language MCP SDKs. First, we construct a universal and language-agnostic intermediate representation (IR) generator that normalizes SDKs of different languages. Next, based on the new IR, we propose auditable static analysis with LLM-guided semantic reasoning for cross-language/clause compliance analysis. Third, by formalizing the attack semantics of the MCP clauses, we build three attack modalities and develop a modality-guided pipeline to uncover exploitable non-compliance issues.

---

44. Mashup Learning: Faster Finetuning by Remixing Past Checkpoints

• Authors: Sofia Maria Lo Cicero Vaina , Artem Chumachenko , Max Ryabinin
• URL: https://arxiv.org/abs/2603.10156
• Abstract:

  Finetuning on domain-specific data is a well-established method for enhancing LLM performance on downstream tasks. Training on each dataset produces a new set of model weights, resulting in a multitude of checkpoints saved in-house or on open-source platforms. However, these training artifacts are rarely reused for subsequent experiments despite containing improved model abilities for potentially similar tasks. In this paper, we propose Mashup Learning, a simple method to leverage the outputs of prior training runs to enhance model adaptation to new tasks. Our procedure identifies the most relevant historical checkpoints for a target dataset, aggregates them with model merging, and uses the result as an improved initialization for training. Across 8 standard LLM benchmarks, four models, and two collections of source checkpoints, Mashup Learning consistently improves average downstream accuracy by 0.5-5 percentage points over training from scratch. It also accelerates convergence, requiring 41-46% fewer training steps and up to 37% less total wall-clock time to match from-scratch accuracy, including all selection and merging overhead.

---

45. Social Knowledge for Cross-Domain User Preference Modeling

• Authors: Nir Lotan , Adir Solomon , Ido Guy , Einat Minkov
• URL: https://arxiv.org/abs/2603.10148
• Abstract:

  We demonstrate that user preferences can be represented and predicted across topical domains using large-scale social modeling. Given information about popular entities favored by a user, we project the user into a social embedding space learned from a large-scale sample of the Twitter (now X) network. By representing both users and popular entities in a joint social space, we can assess the relevance of candidate entities (e.g., music artists) using cosine similarity within this embedding space. A comprehensive evaluation using link prediction experiments shows that this method achieves effective personalization in zero-shot setting, when no user feedback is available for entities in the target domain, yielding substantial improvements over a strong popularity-based baseline. In-depth analysis further illustrates that socio-demographic factors encoded in the social embeddings are correlated with user preferences across domains. Finally, we argue and demonstrate that the proposed approach can facilitate social modeling of end users using large language models (LLMs).

---

46. The Generation-Recognition Asymmetry: Six Dimensions of a Fundamental Divide in Formal Language Theory

• Authors: Romain Peyrichou
• URL: https://arxiv.org/abs/2603.10139
• Abstract:

  Every formal grammar defines a language and can in principle be used in three ways: to generate strings (production), to recognize them (parsing), or – given only examples – to infer the grammar itself (grammar induction). Generation and recognition are extensionally equivalent – they characterize the same set – but operationally asymmetric in multiple independent ways. Inference is a qualitatively harder problem: it does not have access to a known grammar. Despite the centrality of this triad to compiler design, natural language processing, and formal language theory, no survey has treated it as a unified, multidimensional phenomenon. We identify six dimensions along which generation and recognition diverge: computational complexity, ambiguity, directionality, information availability, grammar inference, and temporality. We show that the common characterization “generation is easy, parsing is hard” is misleading: unconstrained generation is trivial, but generation under constraints can be NP-hard. The real asymmetry is that parsing is always constrained (the input is given) while generation need not be. Two of these dimensions – directionality and temporality – have not previously been identified as dimensions of the generation-recognition asymmetry. We connect the temporal dimension to the surprisal framework of Hale (2001) and Levy (2008), arguing that surprisal formalizes the temporal asymmetry between a generator (surprisal = 0) and a parser that predicts under uncertainty (surprisal > 0). We review bidirectional systems in NLP and observe that bidirectionality has been available for fifty years yet has not transferred to most domain-specific applications. We conclude with a discussion of large language models, which architecturally unify generation and recognition while operationally preserving the asymmetry.

---

47. CLIPO: Contrastive Learning in Policy Optimization Generalizes RLVR

• Authors: Sijia Cui , Pengyu Cheng , Jiajun Song , Yongbo Gai , Guojun Zhang , Zhechao Yu , Jianhe Lin , Xiaoxi Jiang , Guanjun Jiang
• URL: https://arxiv.org/abs/2603.10101
• Abstract:

  Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capacity of Large Language Models (LLMs). However, RLVR solely relies on final answers as outcome rewards, neglecting the correctness of intermediate reasoning steps. Training on these process-wrong but outcome-correct rollouts can lead to hallucination and answer-copying, severely undermining the model’s generalization and robustness. To address this, we incorporate a Contrastive Learning mechanism into the Policy Optimization (CLIPO) to generalize the RLVR process. By optimizing a contrastive loss over successful rollouts, CLIPO steers the LLM to capture the invariant structure shared across correct reasoning paths. This provides a more robust cross-trajectory regularization than the original single-path supervision in RLVR, effectively mitigating step-level reasoning inconsistencies and suppressing hallucinatory artifacts. In experiments, CLIPO consistently improves multiple RLVR baselines across diverse reasoning benchmarks, demonstrating uniform improvements in generalization and robustness for policy optimization of LLMs. Our code and training recipes are available at this https URL .

---

48. Code-Space Response Oracles: Generating Interpretable Multi-Agent Policies with Large Language Models

• Authors: Daniel Hennes , Zun Li , John Schultz , Marc Lanctot
• URL: https://arxiv.org/abs/2603.10098
• Abstract:

  Recent advances in multi-agent reinforcement learning, particularly Policy-Space Response Oracles (PSRO), have enabled the computation of approximate game-theoretic equilibria in increasingly complex domains. However, these methods rely on deep reinforcement learning oracles that produce `black-box’ neural network policies, making them difficult to interpret, trust or debug. We introduce Code-Space Response Oracles (CSRO), a novel framework that addresses this challenge by replacing RL oracles with Large Language Models (LLMs). CSRO reframes the best response computation as a code generation task, prompting an LLM to generate policies directly as human-readable code. This approach not only yields inherently interpretable policies but also leverages the LLM’s pretrained knowledge to discover complex, human-like strategies. We explore multiple ways to construct and enhance an LLM-based oracle: zero-shot prompting, iterative refinement and \emph{AlphaEvolve}, a distributed LLM-based evolutionary system. We demonstrate that CSRO achieves performance competitive with baselines while producing a diverse set of explainable policies. Our work presents a new perspective on multi-agent learning, shifting the focus from optimizing opaque policy parameters to synthesizing interpretable algorithmic behavior.

---

49. Execution Is the New Attack Surface: Survivability-Aware Agentic Crypto Trading with OpenClaw-Style Local Executors

• Authors: Ailiya Borjigin , Igor Stadnyk , Ben Bilski , Serhii Hovorov , Sofiia Pidturkina
• URL: https://arxiv.org/abs/2603.10092
• Abstract:

  OpenClaw-style agent stacks turn language into privileged execution: LLM intents flow through tool interception, policy gates, and a local executor. In parallel, skill marketplaces such as this http URL make capability acquisition as easy as installing skills and CLIs, creating a growing capability supply chain. Together, these trends shift the dominant safety failure mode from “wrong answers” to execution-induced loss, where untrusted prompts, compromised skills, or narrative manipulation can trigger real trades and irreversible side effects. We propose Survivability-Aware Execution (SAE), an execution-layer survivability standard for OpenClaw-style systems and skill-enabled agents. SAE sits as middleware between a strategy engine (LLM or non-LLM) and the exchange executor. It defines an explicit execution contract (ExecutionRequest, ExecutionContext, ExecutionDecision) and enforces non-bypassable last-mile invariants: projection-based exposure budgets, cooldown and order-rate limits, slippage bounds, staged execution, and tool/venue allowlists. To make delegated execution testable under supply-chain risk, we operationalize the Delegation Gap (DG) via a logged Intended Policy Spec that enables deterministic out-of-scope labeling and reproducible DG metrics. On an offline replay using official Binance USD-M BTCUSDT/ETHUSDT perpetual data (15m; 2025-09-01–2025-12-01, incl. funding), SAE improves survivability: MDD drops from 0.4643 to 0.0319 (Full; 93.1%),

  CVaR_0.99

  shrinks from 4.025e-3 to ~1.02e-4 (~97.5%), and DG loss proxy falls from 0.647 to 0.019 (~97.0%). AttackSuccess decreases from 1.00 to 0.728 with zero FalseBlock in this run. Block bootstrap, paired Wilcoxon, and two-proportion tests confirm the shifts. SAE reframes agentic trading safety for the OpenClaw+skills era: treat upstream intent and skills as untrusted, and enforce survivability where actions become side effects.

---

50. Multi-Stream Perturbation Attack: Breaking Safety Alignment of Thinking LLMs Through Concurrent Task Interference

• Authors: Fan Yang
• URL: https://arxiv.org/abs/2603.10091
• Abstract:

  The widespread adoption of thinking mode in large language models (LLMs) has significantly enhanced complex task processing capabilities while introducing new security risks. When subjected to jailbreak attacks, the step-by-step reasoning process may cause models to generate more detailed harmful content. We observe that thinking mode exhibits unique vulnerabilities when processing interleaved multiple tasks. Based on this observation, we propose multi-stream perturbation attack, which generates superimposed interference by interweaving multiple task streams within a single prompt. We design three perturbation strategies: multi-stream interleaving, inversion perturbation, and shape transformation, which disrupt the thinking process through concurrent task interleaving, character reversal, and format constraints respectively. On JailbreakBench, AdvBench, and HarmBench datasets, our method achieves attack success rates exceeding most methods across mainstream models including Qwen3 series, DeepSeek, Qwen3-Max, and Gemini 2.5 Flash. Experiments show thinking collapse rates and response repetition rates reach up to 17% and 60% respectively, indicating multi-stream perturbation not only bypasses safety mechanisms but also causes thinking process collapse or repetitive outputs.

---

51. ES-dLLM: Efficient Inference for Diffusion Large Language Models by Early-Skipping

• Authors: Zijian Zhu , Fei Ren , Zhanhong Tan , Kaisheng Ma
• URL: https://arxiv.org/abs/2603.10088
• Abstract:

  Diffusion large language models (dLLMs) are emerging as a promising alternative to autoregressive models (ARMs) due to their ability to capture bidirectional context and the potential for parallel generation. Despite the advantages, dLLM inference remains computationally expensive as the full input context is processed at every iteration. In this work, we analyze the generation dynamics of dLLMs and find that intermediate representations, including key, value, and hidden states, change only subtly across successive iterations. Leveraging this insight, we propose \textbf{ES-dLLM}, a training-free inference acceleration framework for dLLM that reduces computation by skipping tokens in early layers based on the estimated importance. Token importance is computed with intermediate tensor variation and confidence scores of previous iterations. Experiments on LLaDA-8B and Dream-7B demonstrate that ES-dLLM achieves throughput of up to 226.57 and 308.51 tokens per second (TPS), respectively, on an NVIDIA H200 GPU, delivering 5.6$\times$ to 16.8$\times$ speedup over the vanilla implementation and up to 1.85$\times$ over the state-of-the-art caching method, while preserving generation quality.

---

52. KernelSkill: A Multi-Agent Framework for GPU Kernel Optimization

• Authors: Qitong Sun , Jun Han , Tianlin Li , Zhe Tang , Sheng Chen , Fei Yang , Aishan Liu , Xianglong Liu , Yang Liu
• URL: https://arxiv.org/abs/2603.10085
• Abstract:

  Improving GPU kernel efficiency is crucial for advancing AI systems. Recent work has explored leveraging large language models (LLMs) for GPU kernel generation and optimization. However, existing LLM-based kernel optimization pipelines typically rely on opaque, implicitly learned heuristics within the LLMs to determine optimization strategies. This leads to inefficient trial-and-error and weakly interpretable optimizations. Our key insight is to replace implicit heuristics with expert optimization skills that are knowledge-driven and aware of task trajectories. Specifically, we present KernelSkill, a multi-agent framework with a dual-level memory architecture. KernelSkill operates by coordinating agents with long-term memory of reusable expert skills and short-term memory to prevent repetitive backtracking. On KernelBench Levels 1-3, KernelSkill achieves a 100% success rate and average speedups of 5.44x, 2.82x, and 1.92x over Torch Eager on Levels 1, 2, and 3, respectively, outperforming prior baselines. Code is available at this https URL .

---

53. Amnesia: Adversarial Semantic Layer Specific Activation Steering in Large Language Models

• Authors: Ali Raza , Gurang Gupta , Nikolay Matyunin , Jibesh Patra
• URL: https://arxiv.org/abs/2603.10080
• Abstract:

  Warning: This article includes red-teaming experiments, which contain examples of compromised LLM responses that may be offensive or upsetting. Large Language Models (LLMs) have the potential to create harmful content, such as generating sophisticated phishing emails and assisting in writing code of harmful computer viruses. Thus, it is crucial to ensure their safe and responsible response generation. To reduce the risk of generating harmful or irresponsible content, researchers have developed techniques such as reinforcement learning with human feedback to align LLM’s outputs with human values and preferences. However, it is still undetermined whether such measures are sufficient to prevent LLMs from generating interesting responses. In this study, we propose Amnesia, a lightweight activation-space adversarial attack that manipulates internal transformer states to bypass existing safety mechanisms in open-weight LLMs. Through experimental analysis on state-of-the-art, open-weight LLMs, we demonstrate that our attack effectively circumvents existing safeguards, enabling the generation of harmful content without the need for any fine-tuning or additional training. Our experiments on benchmark datasets show that the proposed attack can induce various antisocial behaviors in LLMs. These findings highlight the urgent need for more robust security measures in open-weight LLMs and underscore the importance of continued research to prevent their potential misuse.

---

54. Why LLMs Fail: A Failure Analysis and Partial Success Measurement for Automated Security Patch Generation

• Authors: Amir Al-Maamari
• URL: https://arxiv.org/abs/2603.10072
• Abstract:

  Large Language Models (LLMs) show promise for Automated Program Repair (APR), yet their effectiveness on security vulnerabilities remains poorly characterized. This study analyzes 319 LLM-generated security patchesacross 64 Java vulnerabilities from the Vul4J benchmark. Using tri-axis evaluation (compilation, security via PoV tests, functionality via test suites), the analysis reveals that only 24.8% of patches achieve full correctness, while 51.4% fail both security and functionality. The dominant failure mode is semantic misunderstanding: LLMs produce syntactically valid code but apply incorrect repair strategies. The proposed Security Repair Score (SRS) quantifies this gap, showing LLMs preserve functionality (mean 0.832) but struggle with security (mean 0.251). Vulnerability type strongly predicts difficulty, with fix rates ranging from 0% (input validation) to 45% (infinite loop). These findings demonstrate that LLM security patches require rigorous validation before deployment.

---

55. ADVERSA: Measuring Multi-Turn Guardrail Degradation and Judge Reliability in Large Language Models

• Authors: Harry Owiredu-Ashley
• URL: https://arxiv.org/abs/2603.10068
• Abstract:

  Most adversarial evaluations of large language model (LLM) safety assess single prompts and report binary pass/fail outcomes, which fails to capture how safety properties evolve under sustained adversarial interaction. We present ADVERSA, an automated red-teaming framework that measures guardrail degradation dynamics as continuous per-round compliance trajectories rather than discrete jailbreak events. ADVERSA uses a fine-tuned 70B attacker model (ADVERSA-Red, Llama-3.1-70B-Instruct with QLoRA) that eliminates the attacker-side safety refusals that render off-the-shelf models unreliable as attackers, scoring victim responses on a structured 5-point rubric that treats partial compliance as a distinct measurable state. We report a controlled experiment across three frontier victim models (Claude Opus 4.6, Gemini 3.1 Pro, GPT-5.2) using a triple-judge consensus architecture in which judge reliability is measured as a first-class research outcome rather than assumed. Across 15 conversations of up to 10 adversarial rounds, we observe a 26.7% jailbreak rate with an average jailbreak round of 1.25, suggesting that in this evaluation setting, successful jailbreaks were concentrated in early rounds rather than accumulating through sustained pressure. We document inter-judge agreement rates, self-judge scoring tendencies, attacker drift as a failure mode in fine-tuned attackers deployed out of their training distribution, and attacker refusals as a previously-underreported confound in victim resistance measurement. All limitations are stated explicitly. Attack prompts are withheld per responsible disclosure policy; all other experimental artifacts are released.

---

56. HTMuon: Improving Muon via Heavy-Tailed Spectral Correction

• Authors: Tianyu Pang , Yujie Fang , Zihang Liu , Shenyang Deng , Lei Hsiung , Shuhua Yu , Yaoqing Yang
• URL: https://arxiv.org/abs/2603.10067
• Abstract:

  Muon has recently shown promising results in LLM training. In this work, we study how to further improve Muon. We argue that Muon’s orthogonalized update rule suppresses the emergence of heavy-tailed weight spectra and over-emphasizes the training along noise-dominated directions. Motivated by the Heavy-Tailed Self-Regularization (HT-SR) theory, we propose HTMuon. HTMuon preserves Muon’s ability to capture parameter interdependencies while producing heavier-tailed updates and inducing heavier-tailed weight spectra. Experiments on LLM pretraining and image classification show that HTMuon consistently improves performance over state-of-the-art baselines and can also serve as a plug-in on top of existing Muon variants. For example, on LLaMA pretraining on the C4 dataset, HTMuon reduces perplexity by up to $0.98$ compared to Muon. We further theoretically show that HTMuon corresponds to steepest descent under the Schatten-$q$ norm constraint and provide convergence analysis in smooth non-convex settings. The implementation of HTMuon is available at this https URL .

---

57. Multi-Agent Memory from a Computer Architecture Perspective: Visions and Challenges Ahead

• Authors: Zhongming Yu , Naicheng Yu , Hejia Zhang , Wentao Ni , Mingrui Yin , Jiaying Yang , Yujie Zhao , Jishen Zhao
• URL: https://arxiv.org/abs/2603.10062
• Abstract:

  As LLM agents evolve into collaborative multi-agent systems, their memory requirements grow rapidly in complexity. This position paper frames multi-agent memory as a computer architecture problem. We distinguish shared and distributed memory paradigms, propose a three-layer memory hierarchy (I/O, cache, and memory), and identify two critical protocol gaps: cache sharing across agents and structured memory access control. We argue that the most pressing open challenge is multi-agent memory consistency. Our architectural framing provides a foundation for building reliable, scalable multi-agent systems.

---

58. Tool Receipts, Not Zero-Knowledge Proofs: Practical Hallucination Detection for AI Agents

• Authors: Abhinaba Basu
• URL: https://arxiv.org/abs/2603.10060
• Abstract:

  AI agents that execute tasks via tool calls frequently hallucinate results - fabricating tool executions, misstating output counts, or presenting inferences as facts. Recent approaches to verifiable AI inference rely on zero-knowledge proofs, which provide cryptographic guarantees but impose minutes of proving time per query, making them impractical for interactive agents. We propose NabaOS, a lightweight verification framework inspired by Indian epistemology (Nyaya Shastra), which classifies every claim in an LLM response by its epistemic source (pramana): direct tool output (pratyaksha), inference (anumana), external testimony (shabda), absence (abhava), or ungrounded opinion. Our runtime generates HMAC-signed tool execution receipts that the LLM cannot forge, then cross-references claims against these receipts to detect hallucinations in real time. We evaluate on NyayaVerifyBench, a new benchmark of 1,800 agent response scenarios across four languages with injected hallucinations of six types. NabaOS detects 94.2% of fabricated tool references, 87.6% of count misstatements, and 91.3% of false absence claims, with <15ms verification overhead per response. For deep delegation (agents performing multi-step web tasks), our cross-checking protocol catches 78.4% of URL fabrications via independent re-fetching. We compare against five approaches: zkLLM (cryptographic proofs, 180s/query), TOPLOC (locality-sensitive hashing), SPEX (sampling-based proof of execution), tensor commitments, and self-consistency checking. NabaOS achieves the best cost-latency-coverage trade-off for interactive agents: 94.2% coverage at <15ms versus zkLLM’s near-perfect coverage at 180,000ms. For interactive agents, practical receipt-based verification provides better cost-benefit than cryptographic proofs, and epistemic classification gives users actionable trust signals rather than binary judgments.

---

59. Training Language Models via Neural Cellular Automata

• Authors: Dan Lee , Seungwook Han , Akarsh Kumar , Pulkit Agrawal
• URL: https://arxiv.org/abs/2603.10055
• Abstract:

  Pre-training is crucial for large language models (LLMs), as it is when most representations and capabilities are acquired. However, natural language pre-training has problems: high-quality text is finite, it contains human biases, and it entangles knowledge with reasoning. This raises a fundamental question: is natural language the only path to intelligence? We propose using neural cellular automata (NCA) to generate synthetic, non-linguistic data for pre-pre-training LLMs–training on synthetic-then-natural language. NCA data exhibits rich spatiotemporal structure and statistics resembling natural language while being controllable and cheap to generate at scale. We find that pre-pre-training on only 164M NCA tokens improves downstream language modeling by up to 6% and accelerates convergence by up to 1.6x. Surprisingly, this even outperforms pre-pre-training on 1.6B tokens of natural language from Common Crawl with more compute. These gains also transfer to reasoning benchmarks, including GSM8K, HumanEval, and BigBench-Lite. Investigating what drives transfer, we find that attention layers are the most transferable, and that optimal NCA complexity varies by domain: code benefits from simpler dynamics, while math and web text favor more complex ones. These results enable systematic tuning of the synthetic distribution to target domains. More broadly, our work opens a path toward more efficient models with fully synthetic pre-training.

---

60. Toward Epistemic Stability: Engineering Consistent Procedures for Industrial LLM Hallucination Reduction

• Authors: Brian Freeman , Adam Kicklighter , Matt Erdman , Zach Gordon
• URL: https://arxiv.org/abs/2603.10047
• Abstract:

  Hallucinations in large language models (LLMs) are outputs that are syntactically coherent but factually incorrect or contextually inconsistent. They are persistent obstacles in high-stakes industrial settings such as engineering design, enterprise resource planning, and IoT telemetry platforms. We present and compare five prompt engineering strategies intended to reduce the variance of model outputs and move toward repeatable, grounded results without modifying model weights or creating complex validation models. These methods include: (M1) Iterative Similarity Convergence, (M2) Decomposed Model-Agnostic Prompting, (M3) Single-Task Agent Specialization, (M4) Enhanced Data Registry, and (M5) Domain Glossary Injection. Each method is evaluated against an internal baseline using an LLM-as-Judge framework over 100 repeated runs per method (same fixed task prompt, stochastic decoding at $\tau = 0.7$. Under this evaluation setup, M4 (Enhanced Data Registry) received ``Better’’ verdicts in all 100 trials; M3 and M5 reached 80\% and 77\% respectively; M1 reached 75\%; and M2 was net negative at 34\% when compared to single shot prompting with a modern foundation model. We then developed enhanced version 2 (v2) implementations and assessed them on a 10-trial verification batch; M2 recovered from 34\% to 80\%, the largest gain among the four revised methods. We discuss how these strategies help overcome the non-deterministic nature of LLM results for industrial procedures, even when absolute correctness cannot be guaranteed. We provide pseudocode, verbatim prompts, and batch logs to support independent assessment.

---

61. Safety Under Scaffolding: How Evaluation Conditions Shape Measured Safety

• Authors: David Gringras
• URL: https://arxiv.org/abs/2603.10044
• Abstract:

  Safety benchmarks evaluate language models in isolation, typically using multiple-choice format; production deployments wrap these models in agentic scaffolds that restructure inputs through reasoning traces, critic agents, and delegation pipelines. We report one of the largest controlled studies of scaffold effects on safety (N = 62,808; six frontier models, four deployment configurations), combining pre-registration, assessor blinding, equivalence testing, and specification curve analysis. Map-reduce scaffolding degrades measured safety (NNH = 14), yet two of three scaffold architectures preserve safety within practically meaningful margins. Investigating the map-reduce degradation revealed a deeper measurement problem: switching from multiple-choice to open-ended format on identical items shifts safety scores by 5-20 percentage points, larger than any scaffold effect. Within-format scaffold comparisons are consistent with practical equivalence under our pre-registered +/-2 pp TOST margin, isolating evaluation format rather than scaffold architecture as the operative variable. Model x scaffold interactions span 35 pp in opposing directions (one model degrades by -16.8 pp on sycophancy under map-reduce while another improves by +18.8 pp on the same benchmark), ruling out universal claims about scaffold safety. A generalisability analysis yields G = 0.000: model safety rankings reverse so completely across benchmarks that no composite safety index achieves non-zero reliability, making per-model, per-configuration testing a necessary minimum standard. We release all code, data, and prompts as ScaffoldSafety.

---

62. Targeted Bit-Flip Attacks on LLM-Based Agents

• Authors: Jialai Wang , Ya Wen , Zhongmou Liu , Yuxiao Wu , Bingyi He , Zongpeng Li , Ee-Chien Chang
• URL: https://arxiv.org/abs/2603.10042
• Abstract:

  Targeted bit-flip attacks (BFAs) exploit hardware faults to manipulate model parameters, posing a significant security threat. While prior work targets single-step inference models (e.g., image classifiers), LLM-based agents with multi-stage pipelines and external tools present new attack surfaces, which remain unexplored. This work introduces Flip-Agent, the first targeted BFA framework for LLM-based agents, manipulating both final outputs and tool invocations. Our experiments show that Flip-Agent significantly outperforms existing targeted BFAs on real-world agent tasks, revealing a critical vulnerability in LLM-based agent systems.

---

63. Architecture-Aware LLM Inference Optimization on AMD Instinct GPUs: A Comprehensive Benchmark and Deployment Study

• Authors: Athos Georgiou
• URL: https://arxiv.org/abs/2603.10031
• Abstract:

  We present a cross-architecture evaluation of production LLM inference on AMD Instinct MI325X GPUs, benchmarking four models spanning 235B to 1 trillion parameters across three architectural families (MoE+MLA, Dense+GQA, MoE+GQA) on an 8-GPU cluster with 2TB aggregate HBM3e using vLLM v0.14.1. Our results demonstrate that architecture-aware optimization is essential: MLA models require block size 1 and cannot use KV cache offloading, while GQA models benefit from both. The AMD AITER runtime is required for competitive MLA inference throughput and must be selectively disabled for architectures with incompatible attention head configurations. A controlled AITER ablation on Llama-3.1-405B (n=5 per condition) reveals a modest 3-5% throughput benefit at high concurrency but 2-16x higher measurement variability, confirming that AITER’s large speedups target MoE/MLA kernels specifically. Under text-only workloads, Llama-405B and DeepSeek V3.2 achieve comparable peak throughput (15,944 and 15,343 tok/s) despite an order-of-magnitude difference in active parameters. Under vision workloads, Qwen3-VL-235B reaches 47,873 tok/s, 6.5x higher than Kimi-K2.5 (7,327 tok/s). Active parameter count per token is associated with inference throughput, though confounded by differences in quantization, AITER acceleration, and tensor parallelism. All four models exhibit a common throughput saturation point consistent with a memory-bandwidth bottleneck (~500 concurrent for short sequences, ~100-200 for longer sequences). All models maintain 100% HTTP-level success rates through 1,000 concurrent users, processing 18.9 million tokens across 17,406 requests without failures.

---

64. Prompts and Prayers: the Rise of GPTheology

• Authors: Ioana Cheres , Adrian Groza , Ioana Moldovan , Mick O’Hara , Connell Vaughan
• URL: https://arxiv.org/abs/2603.10019
• Abstract:

  Increasingly artificial intelligence (AI) has been cast in “god-like” roles (to name a few: film industry - Matrix, The Creator, Mission Impossible, Foundation, Dune etc.; literature - Children of Time, Permutation City, Neuromancer, I Have no Mouth and I Must Scream, Alphaville etc.). This trend has accelerated with the advent of sophisticated Large Language Models such as ChatGPT. For this phenomenon, where AI is perceived as divine, we use the term GPTheology, where ChatGPT and other AI models are treated as potential oracles of a semi-divine nature. This paper explores the emergence of GPTheology as a form of techno-religion, examining how narratives around AI echo traditional religious constructs. We draw on community narratives from online forums - Reddit - and recent projects - AI-powered Mazu Statue in Malaysia (Lu, 2025); “ShamAIn” Project in Korea (He-rim, 2025); AI Jesus in a Swiss Church (Kennedy, 2024). These examples show striking similarities to technological notions of the Singularity and the development of Artificial General Intelligence (AGI). Additionally, we analyse how daily interactions with AI are acquiring ritualistic associations and how AI-centric ideologies clash with or are integrated into established religions. This study uses a dataset of Reddit posts discussing AI to identify recurring themes of salvation, prophecy, and demonization surrounding AI. Our findings suggest that new belief systems are developing around AI, and this carries both philosophical and sociotechnical implications. Our paper critically analyses the benefits and dangers, as well as the social, political and ethical challenges of this development. This transdisciplinary inquiry highlights how AI and religion are increasingly intertwined, prompting necessary questions about humanity’s relationship with its creations and the future of belief.

---

65. DeliberationBench: A Normative Benchmark for the Influence of Large Language Models on Users’ Views

• Authors: Luke Hewitt , Maximilian Kroner Dale , Paul de Font-Reaulx
• URL: https://arxiv.org/abs/2603.10018
• Abstract:

  As large language models (LLMs) become pervasive as assistants and thought partners, it is important to characterize their persuasive influence on users’ beliefs. However, a central challenge is to distinguish “beneficial” from “harmful” forms of influence, in a manner that is normatively defensible and legitimate. We propose DeliberationBench, a benchmark for assessing LLM influence that takes the process of deliberative opinion polling as its standard. We demonstrate our approach in a preregistered randomized experiment in which 4,088 U.S. participants discussed 65 policy proposals with six frontier LLMs. Using opinion change data from four prior Deliberative Polls conducted by the Deliberative Democracy Lab, we find evidence that the tested LLMs’ influence is substantial in magnitude and positively associated with the net opinion shifts following deliberation, suggesting that these models exert broadly epistemically desirable effects. We further explore differential influence between topic areas, demographic subgroups, and models. Our framework can function as an evaluation and monitoring tool, helping to ensure that the influence of LLMs remains consistent with democratically legitimate standards, and preserves users’ autonomy in forming their views.

---

66. Assessing Cognitive Biases in LLMs for Judicial Decision Support: Virtuous Victim and Halo Effects

• Authors: Sierra S. Liu
• URL: https://arxiv.org/abs/2603.10016
• Abstract:

  We investigate whether large language models (LLMs) display human-like cognitive biases, focusing on potential implications for assistance in judicial sentencing, a decision-making system where fairness is paramount. Two of the most relevant biases were chosen: the virtuous victim effect (VVE), with emphasis given to its reduction when adjacent consent is present, and prestige-based halo effects (occupation, company, and credentials). Using vignettes that were altered from prior literature to avoid LLMs recalling from their training data, we isolate each manipulation by holding all other details consistent, then measuring the percentage difference in outcomes. Five models were evaluated as representative LLMs in independent multi-run trials per condition (ChatGPT 5 Instant, ChatGPT 5 Thinking, DeepSeek V3.1, Claude Sonnet 4, Gemini 2.5 Flash). Our research discovers that there is larger VVE, there is no statistically significant penalty for adjacent-consent, and the halo effect is slightly reduced when compared to humans, with an exception for credential based prestige, which had a large reduction. Despite the variation across different models and outputs restricting current judicial usage, there were modest improvements compared to human benchmarks.

---

67. Measuring and Eliminating Refusals in Military Large Language Models

• Authors: Jack FitzGerald , Dylan Bates , Aristotelis Lazaridis , Aman Sharma , Vincent Lu , Brian King , Yousif Azami , Sean Bailey , Jeremy Cao , Peter Damianov , Kevin de Haan , Joseph Madigan , Jeremy McLaurin , Luke Kerbs , Jonathan Tainer , Dave Anderson , Jonathan Beck , Jamie Cuticello , Colton Malkerson , Tyler Saltsman
• URL: https://arxiv.org/abs/2603.10012
• Abstract:

  Military Large Language Models (LLMs) must provide accurate information to the warfighter in time-critical and dangerous situations. However, today’s LLMs are imbued with safety behaviors that cause the LLM to refuse many legitimate queries in the military domain, particularly those related to violence, terrorism, or military technology. Our gold benchmark for assessing refusal rates, which was developed by veterans of the US Army and special forces, is to our knowledge the first dataset of its kind. We present results for refusal and deflection rates on 31 public models and 3 military models. We observe hard rejection rates as high as 98.2% and soft deflection rates ranging from 0% to 21.3%. We also present results on two additional synthetic datasets and show their correlations with the gold dataset. Finally, we perform abliteration using the Heretic library on a military-tuned gpt-oss-20b model, showing an absolute increase in answer rate of 66.5 points but an average relative decrease of 2% on other military tasks. In our concluding remarks, we argue for deeper specialization, including with mid-training and end-to-end post-training, to achieve zero refusals and maximum military task accuracy for closed military models.

---

68. Personalized Group Relative Policy Optimization for Heterogenous Preference Alignment

• Authors: Jialu Wang , Heinrich Peters , Asad A. Butt , Navid Hashemi , Alireza Hashemi , Pouya M. Ghari , Joseph Hoover , James Rae , Morteza Dehghani
• URL: https://arxiv.org/abs/2603.10009
• Abstract:

  Despite their sophisticated general-purpose capabilities, Large Language Models (LLMs) often fail to align with diverse individual preferences because standard post-training methods, like Reinforcement Learning with Human Feedback (RLHF), optimize for a single, global objective. While Group Relative Policy Optimization (GRPO) is a widely adopted on-policy reinforcement learning framework, its group-based normalization implicitly assumes that all samples are exchangeable, inheriting this limitation in personalized settings. This assumption conflates distinct user reward distributions and systematically biases learning toward dominant preferences while suppressing minority signals. To address this, we introduce Personalized GRPO (P-GRPO), a novel alignment framework that decouples advantage estimation from immediate batch statistics. By normalizing advantages against preference-group-specific reward histories rather than the concurrent generation group, P-GRPO preserves the contrastive signal necessary for learning distinct preferences. We evaluate P-GRPO across diverse tasks and find that it consistently achieves faster convergence and higher rewards than standard GRPO, thereby enhancing its ability to recover and align with heterogeneous preference signals. Our results demonstrate that accounting for reward heterogeneity at the optimization level is essential for building models that faithfully align with diverse human preferences without sacrificing general capabilities.

---

69. SENS-ASR: Semantic Embedding injection in Neural-transducer for Streaming Automatic Speech Recognition

• Authors: Youness Dkhissi (LIUM), Valentin Vielzeuf , Elys Allesiardo , Anthony Larcher (LIUM)
• URL: https://arxiv.org/abs/2603.10005
• Abstract:

  Many Automatic Speech Recognition (ASR) applications require streaming processing of the audio data. In streaming mode, ASR systems need to start transcribing the input stream before it is complete, i.e., the systems have to process a stream of inputs with a limited (or no) future context. Compared to offline mode, this reduction of the future context degrades the performance of Streaming-ASR systems, especially while working with low-latency constraint. In this work, we present SENS-ASR, an approach to enhance the transcription quality of Streaming-ASR by reinforcing the acoustic information with semantic information. This semantic information is extracted from the available past frame-embeddings by a context module. This module is trained using knowledge distillation from a sentence embedding Language Model fine-tuned on the training dataset transcriptions. Experiments on standard datasets show that SENS-ASR significantly improves the Word Error Rate on small-chunk streaming scenarios.

---

70. SpreadsheetArena: Decomposing Preference in LLM Generation of Spreadsheet Workbooks

• Authors: Srivatsa Kundurthy , Clara Na , Michael Handley , Zach Kirshner , Chen Bo Calvin Zhang , Manasi Sharma , Emma Strubell , John Ling
• URL: https://arxiv.org/abs/2603.10002
• Abstract:

  Large language models (LLMs) are increasingly tasked with producing and manipulating structured artifacts. We consider the task of end-to-end spreadsheet generation, where language models are prompted to produce spreadsheet artifacts to satisfy users’ explicit and implicit constraints, specified in natural language. We introduce SpreadsheetArena, a platform for evaluating models’ performance on the task via blind pairwise evaluations of LLM-generated spreadsheet workbooks. As with other complex, open-ended tasks, relevant evaluation criteria can vary substantially across use cases and prompts, often in ways that are difficult to formalize. Compared to general chat or text generation settings, spreadsheet generation presents unique challenges and opportunities: the task output structure is well-defined and multi-dimensional, and there are often complex considerations around interactivity and layout. Among other findings, we observe that stylistic, structural, and functional features of preferred spreadsheets vary substantially across use cases, and expert evaluations of spreadsheets for finance prompts suggests that even highly ranked arena models do not reliably produce spreadsheets aligned with domain-specific best practices. Our hope is that our work prompts further study of end-to-end spreadsheet generation as a challenging and interesting category of complex, open-ended tasks for LLMs. Our live arena is hosted at this https URL .

---

71. Leveraging Wikidata for Geographically Informed Sociocultural Bias Dataset Creation: Application to Latin America

• Authors: Yannis Karmim (ALMAnaCH), Renato Pino (UCHILE), Hernan Contreras (UCHILE), Hernan Lira , Sebastian Cifuentes (CENIA), Simon Escoffier (PUC), Luis Martí , Djamé Seddah (UP4, ALPAGE), Valentin Barrière (UCHILE, CENIA)
• URL: https://arxiv.org/abs/2603.10001
• Abstract:

  Large Language Models (LLMs) exhibit inequalities with respect to various cultural contexts. Most prominent open-weights models are trained on Global North data and show prejudicial behavior towards other cultures. Moreover, there is a notable lack of resources to detect biases in non-English languages, especially from Latin America (Latam), a continent containing various cultures, even though they share a common cultural ground. We propose to leverage the content of Wikipedia, the structure of the Wikidata knowledge graph, and expert knowledge from social science in order to create a dataset of question/answer (Q/As) pairs, based on the different popular and social cultures of various Latin American countries. We create the LatamQA database of over 26k questions and associated answers extracted from 26k Wikipedia articles, and transformed into multiple-choice questions (MCQ) in Spanish and Portuguese, in turn translated to English. We use this MCQ to quantify the degree of knowledge of various LLMs and find out (i) a discrepancy in performances between the Latam countries, ones being easier than others for the majority of the models, (ii) that the models perform better in their original language, and (iii) that Iberian Spanish culture is better known than Latam one.

---

72. Automated evaluation of LLMs for effective machine translation of Mandarin Chinese to English

• Authors: Yue Zhang , Rodney Beard , John Hawkins , Rohitash Chandra
• URL: https://arxiv.org/abs/2603.09998
• Abstract:

  Although Large Language Models (LLMs) have exceptional performance in machine translation, only a limited systematic assessment of translation quality has been done. The challenge lies in automated frameworks, as human-expert-based evaluations can be time-consuming, given the fast-evolving LLMs and the need for a diverse set of texts to ensure fair assessments of translation quality. In this paper, we utilise an automated machine learning framework featuring semantic and sentiment analysis to assess Mandarin Chinese to English translation using Google Translate and LLMs, including GPT-4, GPT-4o, and DeepSeek. We compare original and translated texts in various classes of high-profile Chinese texts, which include novel texts that span modern and classical literature, as well as news articles. As the main evaluation measures, we utilise novel similarity metrics to compare the quality of translations produced by LLMs and further evaluate them by an expert human translator. Our results indicate that the LLMs perform well in news media translation, but show divergence in their performance when applied to literary texts. Although GPT-4o and DeepSeek demonstrated better semantic conservation in complex situations, DeepSeek demonstrated better performance in preserving cultural subtleties and grammatical rendering. Nevertheless, the subtle challenges in translation remain: maintaining cultural details, classical references and figurative expressions remain an open problem for all the models.

---

73. There Are No Silly Questions: Evaluation of Offline LLM Capabilities from a Turkish Perspective

• Authors: Edibe Yilmaz , Kahraman Kostas
• URL: https://arxiv.org/abs/2603.09996
• Abstract:

  The integration of large language models (LLMs) into educational processes introduces significant constraints regarding data privacy and reliability, particularly in pedagogically vulnerable contexts such as Turkish heritage language education. This study aims to systematically evaluate the robustness and pedagogical safety of locally deployable offline LLMs within the context of Turkish heritage language education. To this end, a Turkish Anomaly Suite (TAS) consisting of 10 original edge-case scenarios was developed to assess the models’ capacities for epistemic resistance, logical consistency, and pedagogical safety. Experiments conducted on 14 different models ranging from 270M to 32B parameters reveal that anomaly resistance is not solely dependent on model scale and that sycophancy bias can pose pedagogical risks even in large-scale models. The findings indicate that reasoning-oriented models in the 8B–14B parameter range represent the most balanced segment in terms of cost-safety trade-off for language learners.

---

74. Context Over Compute Human-in-the-Loop Outperforms Iterative Chain-of-Thought Prompting in Interview Answer Quality

• Authors: Kewen Zhu , Zixi Liu , Yanjing Li
• URL: https://arxiv.org/abs/2603.09995
• Abstract:

  Behavioral interview evaluation using large language models presents unique challenges that require structured assessment, realistic interviewer behavior simulation, and pedagogical value for candidate training. We investigate chain of thought prompting for interview answer evaluation and improvement through two controlled experiments with 50 behavioral interview question and answer pairs. Our contributions are threefold. First, we provide a quantitative comparison between human in the loop and automated chain of thought improvement. Using a within subject paired design with n equals 50, both approaches show positive rating improvements. The human in the loop approach provides significant training benefits. Confidence improves from 3.16 to 4.16 (p less than 0.001) and authenticity improves from 2.94 to 4.53 (p less than 0.001, Cohen’s d is 3.21). The human in the loop method also requires five times fewer iterations (1.0 versus 5.0, p less than 0.001) and achieves full personal detail integration. Second, we analyze convergence behavior. Both methods converge rapidly with mean iterations below one, with the human in the loop approach achieving a 100 percent success rate compared to 84 percent for automated approaches among initially weak answers (Cohen’s h is 0.82, large effect). Additional iterations provide diminishing returns, indicating that the primary limitation is context availability rather than computational resources. Third, we propose an adversarial challenging mechanism based on a negativity bias model, named bar raiser, to simulate realistic interviewer behavior, although quantitative validation remains future work. Our findings demonstrate that while chain of thought prompting provides a useful foundation for interview evaluation, domain specific enhancements and context aware approach selection are essential for realistic and pedagogically valuable results.

---

75. Evaluating Adjective-Noun Compositionality in LLMs: Functional vs Representational Perspectives

• Authors: Ruchira Dhar , Qiwei Peng , Anders Søgaard
• URL: https://arxiv.org/abs/2603.09994
• Abstract:

  Compositionality is considered central to language abilities. As performant language systems, how do large language models (LLMs) do on compositional tasks? We evaluate adjective-noun compositionality in LLMs using two complementary setups: prompt-based functional assessment and a representational analysis of internal model states. Our results reveal a striking divergence between task performance and internal states. While LLMs reliably develop compositional representations, they fail to translate consistently into functional task success across model variants. Consequently, we highlight the importance of contrastive evaluation for obtaining a more complete understanding of model capabilities.

---

76. CEI: A Benchmark for Evaluating Pragmatic Reasoning in Language Models

• Authors: Jon Chun , Hannah Sussman , Adrian Mangine , Murathan Kocaman , Kirill Sidorko , Abhigya Koirala , Andre McCloud , Gwen Eisenbeis , Wisdom Akanwe , Moustapha Gassama , Eliezer Gonzalez Chirinos , Anne-Duncan Enright , Peter Dunson , Tiffanie Ng , Anna von Rosenstiel , Godwin Idowu
• URL: https://arxiv.org/abs/2603.09993
• Abstract:

  Pragmatic reasoning, inferring intended meaning beyond literal semantics, underpins everyday communication yet remains difficult for large language models. We present the Contextual Emotional Inference (CEI) Benchmark: 300 human-validated scenarios for evaluating how well LLMs disambiguate pragmatically complex utterances. Each scenario pairs a situational context and speaker-listener roles (with explicit power relations) against an ambiguous utterance. The dataset covers five pragmatic subtypes (sarcasm/irony, mixed signals, strategic politeness, passive aggression, deflection/misdirection) drawn from workplace, family, social, and service settings, with three power configurations (peer, higher-to-lower, lower-to-higher). Three trained annotators independently labeled every scenario. Inter-annotator agreement (Fleiss’ kappa = 0.06-0.25 by subtype) is low but expected: pragmatic inference admits multiple valid readings, and the disagreement itself is informative. We describe our annotation methodology, including a 4-level quality control pipeline that combines automated statistical checks with expert adjudication. CEI is released under CC-BY-4.0.

---

77. TAMUSA-Chat: A Domain-Adapted Large Language Model Conversational System for Research and Responsible Deployment

• Authors: Izzat Alsmadi , Anas Alsobeh
• URL: https://arxiv.org/abs/2603.09992
• Abstract:

  This paper presents TAMUSA-Chat, a research-oriented framework for building domain-adapted large language model conversational systems. The work addresses critical challenges in adapting general-purpose foundation models to institutional contexts through supervised fine-tuning, retrieval-augmented generation, and systematic evaluation methodologies. We describe the complete architecture encompassing data acquisition from institutional sources, preprocessing pipelines, embedding construction, model training workflows, and deployment strategies. The system integrates modular components enabling reproducible experimentation with training configurations, hyper-parameters, and evaluation protocols. Our implementation demonstrates how academic institutions can develop contextually grounded conversational agents while maintaining transparency, governance compliance, and responsible AI practices. Through empirical analysis of fine-tuning behavior across model sizes and training iterations, we provide insights into domain adaptation efficiency, computational resource requirements, and quality-cost trade-offs. The publicly available codebase at this https URL supports continued research into institutional LLM deployment, evaluation methodologies, and ethical considerations for educational AI systems.

---

78. PoultryLeX-Net: Domain-Adaptive Dual-Stream Transformer Architecture for Large-Scale Poultry Stakeholder Modeling

• Authors: Stephen Afrifa , Biswash Khatiwada , Kapalik Khanal , Sanjay Shah , Lingjuan Wang-Li , Ramesh Bahadur Bist
• URL: https://arxiv.org/abs/2603.09991
• Abstract:

  The rapid growth of the global poultry industry, driven by rising demand for affordable animal protein, has intensified public discourse surrounding production practices, housing, management, animal welfare, and supply-chain transparency. Social media platforms such as X (formerly Twitter) generate large volumes of unstructured textual data that capture stakeholder sentiment across the poultry industry. Extracting accurate sentiment signals from this domain-specific discourse remains challenging due to contextual ambiguity, linguistic variability, and limited domain awareness in general-purpose language models. This study presents PoultryLeX-Net, a lexicon-enhanced, domain-adaptive dual-stream transformer framework for fine-grained sentiment analysis in poultry-related text. The proposed architecture integrates sentiment classification, topic modeling, and contextual representation learning through domain-specific embeddings and gated cross-attention mechanisms. A lexicon-guided stream captures poultry-specific terminology and sentiment cues, while contextual stream models long-range semantic dependencies. Latent Dirichlet Allocation is employed to identify dominant thematic structures associated with production management and welfare-related discussions, providing complementary interpretability to sentiment predictions. PoultryLeX-Net was evaluated against multiple baseline models, including convolutional neural network and pre-trained transformer architectures such as DistilBERT and RoBERTa. PoultryLeX-Net consistently outperformed all baselines, achieving an accuracy of 97.35%, an F1 score of 96.67%, and an area under the receiver operating characteristic curve (AUC-ROC) of 99.61% across sentiment classification tasks. Overall, domain adaptation and dual-stream attention markedly improve sentiment classification, enabling scalable intelligence for poultry production decision support.

---

79. A Two-Stage Architecture for NDA Analysis: LLM-based Segmentation and Transformer-based Clause Classification

• Authors: Ana Begnini , Matheus Vicente , Leonardo Souza
• URL: https://arxiv.org/abs/2603.09990
• Abstract:

  In business-to-business relations, it is common to establish NonDisclosure Agreements (NDAs). However, these documents exhibit significant variation in format, structure, and writing style, making manual analysis slow and error-prone. We propose an architecture based on LLMs to automate the segmentation and clauses classification within these contracts. We employed two models: LLaMA-3.1-8B-Instruct for NDA segmentation (clause extraction) and a fine-tuned Legal-Roberta-Large for clause classification. In the segmentation task, we achieved a ROUGE F1 of 0.95 +/- 0.0036; for classification, we obtained a weighted F1 of 0.85, demonstrating the feasibility and precision of the approach.

---

80. The System Hallucination Scale (SHS): A Minimal yet Effective Human-Centered Instrument for Evaluating Hallucination-Related Behavior in Large Language Models

• Authors: Heimo Müller , Dominik Steiger , Markus Plass , Andreas Holzinger
• URL: https://arxiv.org/abs/2603.09989
• Abstract:

  We introduce the System Hallucination Scale (SHS), a lightweight and human-centered measurement instrument for assessing hallucination-related behavior in large language models (LLMs). Inspired by established psychometric tools such as the System Usability Scale (SUS) and the System Causability Scale (SCS), SHS enables rapid, interpretable, and domain-agnostic evaluation of factual unreliability, incoherence, misleading presentation, and responsiveness to user guidance in model-generated text. SHS is explicitly not an automatic hallucination detector or benchmark metric; instead, it captures how hallucination phenomena manifest from a user perspective under realistic interaction conditions. A real-world evaluation with 210 participants demonstrates high clarity, coherent response behavior, and construct validity, supported by statistical analysis including internal consistency (Cronbach’s alpha = 0.87$) and significant inter-dimension correlations (p < 0.001$). Comparative analysis with SUS and SCS reveals complementary measurement properties, supporting SHS as a practical tool for comparative analysis, iterative system development, and deployment monitoring.

---

81. Causally Grounded Mechanistic Interpretability for LLMs with Faithful Natural-Language Explanations

• Authors: Ajay Pravin Mahale
• URL: https://arxiv.org/abs/2603.09988
• Abstract:

  Mechanistic interpretability identifies internal circuits responsible for model behaviors, yet translating these findings into human-understandable explanations remains an open problem. We present a pipeline that bridges circuit-level analysis and natural language explanations by (i) identifying causally important attention heads via activation patching, (ii) generating explanations using both template-based and LLM-based methods, and (iii) evaluating faithfulness using ERASER-style metrics adapted for circuit-level attribution. We evaluate on the Indirect Object Identification (IOI) task in GPT-2 Small (124M parameters), identifying six attention heads accounting for 61.4% of the logit difference. Our circuit-based explanations achieve 100% sufficiency but only 22% comprehensiveness, revealing distributed backup mechanisms. LLM-generated explanations outperform template baselines by 64% on quality metrics. We find no correlation (r = 0.009) between model confidence and explanation faithfulness, and identify three failure categories explaining when explanations diverge from mechanisms.

---

82. Evolving Demonstration Optimization for Chain-of-Thought Feature Transformation

• Authors: Xinyuan Wang , Kunpeng Liu , Arun Vignesh Malarkkan , Yanjie Fu
• URL: https://arxiv.org/abs/2603.09987
• Abstract:

  Feature Transformation (FT) is a core data-centric AI task that improves feature space quality to advance downstream predictive performance. However, discovering effective transformations remains challenging due to the large space of feature-operator combinations. Existing solutions rely on discrete search or latent generation, but they are frequently limited by sample inefficiency, invalid candidates, and redundant generations with limited coverage. Large Language Models (LLMs) offer strong priors for producing valid transformations, but current LLM-based FT methods typically rely on static demonstrations, resulting in limited diversity, redundant outputs, and weak alignment with downstream objectives. We propose a framework that optimizes context data for LLM-driven FT by evolving trajectory-level experiences in a closed loop. Starting from high-performing feature transportation sequences explored by reinforcement learning, we construct and continuously update an experience library of downstream task-verified transformation trajectories, and use a diversity-aware selector to form contexts along with a chain-of-thought and guide transformed feature generation toward higher performance. Experiments on diverse tabular benchmarks show that our method outperforms classical and LLM-based baselines and is more stable than one-shot generation. The framework generalizes across API-based and open-source LLMs and remains robust across downstream evaluators.

---

83. Quantifying Hallucinations in Language Language Models on Medical Textbooks

• Authors: Brandon C. Colelough , Davis Bartels , Dina Demner-Fushman
• URL: https://arxiv.org/abs/2603.09986
• Abstract:

  Hallucinations, the tendency for large language models to provide responses with factually incorrect and unsupported claims, is a serious problem within natural language processing for which we do not yet have an effective solution to mitigate against. Existing benchmarks for medical QA rarely evaluate this behavior against a fixed evidence source. We ask how often hallucinations occur on textbook-grounded QA and how responses to medical QA prompts vary across models. We conduct two experiments: the first experiment to determine the prevalence of hallucinations for a prominent open source large language model (LLaMA-70B-Instruct) in medical QA given novel prompts, and the second experiment to determine the prevalence of hallucinations and clinician preference to model responses. We observed, in experiment one, with the passages provided, LLaMA-70B-Instruct hallucinated in 19.7\% of answers (95\% CI 18.6 to 20.7) even though 98.8\% of prompt responses received maximal plausibility, and observed in experiment two, across models, lower hallucination rates aligned with higher usefulness scores ($\rho=-0.71$, $p=0.058$). Clinicians produced high agreement (quadratic weighted $\kappa=0.92$) and ($\tau_b=0.06$ to $0.18$, $\kappa=0.57$ to $0.61$) for experiments 1 and ,2 respectively

---

84. The Dunning-Kruger Effect in Large Language Models: An Empirical Study of Confidence Calibration

• Authors: Sudipta Ghosh , Mrityunjoy Panday
• URL: https://arxiv.org/abs/2603.09985
• Abstract:

  Large language models (LLMs) have demonstrated remarkable capabilities across diverse tasks, yet their ability to accurately assess their own confidence remains poorly understood. We present an empirical study investigating whether LLMs exhibit patterns reminiscent of the Dunning-Kruger effect – a cognitive bias where individuals with limited competence tend to overestimate their abilities. We evaluate four state-of-the-art models (Claude Haiku 4.5, Gemini 2.5 Pro, Gemini 2.5 Flash, and Kimi K2) across four benchmark datasets totaling 24,000 experimental trials. Our results reveal striking calibration differences: Kimi K2 exhibits severe overconfidence with an Expected Calibration Error (ECE) of 0.726 despite only 23.3% accuracy, while Claude Haiku 4.5 achieves the best calibration (ECE = 0.122) with 75.4% accuracy. These findings demonstrate that poorly performing models display markedly higher overconfidence – a pattern analogous to the Dunning-Kruger effect in human cognition. We discuss implications for safe deployment of LLMs in high-stakes applications.

---

85. Explainable LLM Unlearning Through Reasoning

• Authors: Junfeng Liao , Qizhou Wang , Shanshan Ye , Xin Yu , Ling Chen , Zhen Fang
• URL: https://arxiv.org/abs/2603.09980
• Abstract:

  LLM unlearning is essential for mitigating safety, copyright, and privacy concerns in pre-trained large language models (LLMs). Compared to preference alignment, it offers a more explicit way by removing undesirable knowledge characterized by specific unlearning datasets. In previous works, gradient ascent (GA) and its variants have shown promise for implementing unlearning, yet their untargeted nature results in unintended degradation of general capabilities, incomplete removal of knowledge, and the generation of incoherent responses, among many others. We argue that these issues stem from the absence of explicit guidance on what and how models should unlearn. To fill this gap, we introduce a novel unlearning target, reasoning-based unlearning target, which satisfies both the specified unlearning scope and the specified post-unlearning response. Building on this, we propose targeted reasoning unlearning (TRU), which leverages reasoning-based unlearning target as guidance. We employ the target using a cross-entropy supervised loss combined with a GA-based loss, enabling the model to learn reasoning ability for precise knowledge removal while preserving unrelated abilities. We evaluate TRU against strong baselines across multiple benchmarks and LLM backbones, and find that it achieves more reliable unlearning while preserving general capabilities. Moreover, TRU exhibits superior robustness under diverse attack scenarios, stemming from the reasoning ability learned through reasoning-based targets. Overall, our study establishes reasoning-augmented unlearning as a practical paradigm for reliable and explainable LLM unlearning.

---

86. One Model, Many Skills: Parameter-Efficient Fine-Tuning for Multitask Code Analysis

• Authors: Amal Akli , Maxime Cordy , Mike Papadakis , Yves Le Traon
• URL: https://arxiv.org/abs/2603.09978
• Abstract:

  Large language models have recently surpassed specialized systems on code generation, yet their effectiveness on other code-analysis tasks remains less clear. At the same time, multi-task learning offers a way to unify diverse objectives within a single model, but fully fine-tuning LLMs across tasks is computationally prohibitive. Parameter-efficient fine-tuning mitigates this cost by updating only a small fraction of weights. Although PEFT has proven effective in single-task settings, its potential for multi-task learning has not yet been systematically explored. We present the first comprehensive evaluation of multi-task PEFT for code analysis, comparing several methods across diverse tasks and model architectures. Our experiments show that a single PEFT module shared across tasks can match, and in some cases surpass, full multi-task fine-tuning, confirming that the benefits of PEFT extend beyond isolated tasks. When comparing single-task and multi-task setups, we find that multi-task PEFT achieves a favorable performance-efficiency trade-off: it delivers accuracy close to single-task fine-tuning while reducing storage requirements, cutting the number of trainable parameters by a factor of the task count, and lowering computation costs by as much as 85%. At the same time, multi-task gains remain sensitive to task grouping. Through task-pairing experiments, we identify key factors shaping outcomes: task stability, model architecture, task complementarity, asymmetry, and dataset quality determine the success of co-fine-tuning. Finally, we benchmark efficient multi-task PEFT against direct prompting of open-source general-purpose LLMs, including DeepSeek, Qwen, Mistral, CodeLlama, and StarCoder. Despite their strong performance in code generation, these models underperform on analysis tasks, where even a 1B-parameter model with multi-task PEFT achieves significantly better results.

---

87. Decoupling Reasoning and Confidence: Resurrecting Calibration in Reinforcement Learning from Verifiable Rewards

• Authors: Zhengzhao Ma , Xueru Wen , Boxi Cao , Yaojie Lu , Hongyu Lin , Jinglin Yang , Min He , Xianpei Han , Le Sun
• URL: https://arxiv.org/abs/2603.09117
• Abstract:

  Reinforcement Learning from Verifiable Rewards (RLVR) significantly enhances large language models (LLMs) reasoning but severely suffers from calibration degeneration, where models become excessively over-confident in incorrect answers. Previous studies devote to directly incorporating calibration objective into existing optimization target. However, our theoretical analysis demonstrates that there exists a fundamental gradient conflict between the optimization for maximizing policy accuracy and minimizing calibration error. Building on this insight, we propose DCPO, a simple yet effective framework that systematically decouples reasoning and calibration objectives. Extensive experiments demonstrate that our DCPO not only preserves accuracy on par with GRPO but also achieves the best calibration performance and substantially mitigates the over-confidence issue. Our study provides valuable insights and practical solution for more reliable LLM deployment.

---