arXiv:2603.12287v1 Announce Type: new Abstract: We address the problem of transforming raw vessel trajectory data collected from AIS into structured and semantically enriched representations interpretable by humans and directly usable by machine reasoning systems. We propose a context-aware trajectory abstraction framework that segments noisy AIS sequences into distinct trips each consisting of clean, mobility-annotated episodes. Each episode is further enriched with multi-source contextual information, such as nearby geographic entities, offshore navigation features, and weather conditions. Crucially, such representations can support generation of controlled natural language descriptions using LLMs. We empirically examine the quality of such descriptions generated using several LLMs over AIS data along with open contextual features. By increasing semantic density and reducing spatiotemporal complexity, this abstraction can facilitate downstream analytics and enable integration with LLMs for higher-level maritime reasoning tasks.

arXiv:2603.12372v1 Announce Type: new Abstract: Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they often suffer from overthinking, expending redundant computational steps on simple problems, or underthinking, failing to explore sufficient reasoning paths despite inherent capabilities. These issues lead to inefficiencies and potential inaccuracies, limiting practical deployment in resource-constrained settings. Existing methods to mitigate overthinking, such as suppressing reflective keywords or adjusting reasoning length, may inadvertently induce underthinking, compromising accuracy. Therefore, we propose ReBalance, a training-free framework that achieves efficient reasoning with balanced thinking. ReBalance leverages confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence. By aggregating hidden states from a small-scale dataset into reasoning mode prototypes, we compute a steering vector to guide LRMs' reasoning trajectories. A dynamic control function modulates this vector's strength and direction based on real-time confidence, pruning redundancy during overthinking, and promoting exploration during underthinking. Extensive experiments conducted on four models ranging from 0.5B to 32B, and across nine benchmarks in math reasoning, general question answering, and coding tasks demonstrate that ReBalance effectively reduces output redundancy while improving accuracy, offering a general, training-free, and plug-and-play strategy for efficient and robust LRM deployment. Code is available at https://github.com/yu-lin-li/ReBalance .

arXiv:2603.12710v1 Announce Type: new Abstract: Developing autonomous agents for web-based tasks is a core challenge in AI. While Large Language Model (LLM) agents can interpret complex user requests, they often operate as black boxes, making it difficult to diagnose why they fail or how they plan. This paper addresses this gap by formally treating web tasks as sequential decision-making processes. We introduce a taxonomy that maps modern agent architectures to traditional planning paradigms: Step-by-Step agents to Breadth-First Search (BFS), Tree Search agents to Best-First Tree Search, and Full-Plan-in-Advance agents to Depth-First Search (DFS). This framework allows for a principled diagnosis of system failures like context drift and incoherent task decomposition. To evaluate these behaviors, we propose five novel evaluation metrics that assess trajectory quality beyond simple success rates. We support this analysis with a new dataset of 794 human-labeled trajectories from the WebArena benchmark. Finally, we validate our evaluation framework by comparing a baseline Step-by-Step agent against a novel Full-Plan-in-Advance implementation. Our results reveal that while the Step-by-Step agent aligns more closely with human gold trajectories (38% overall success), the Full-Plan-in-Advance agent excels in technical measures such as element accuracy (89%), demonstrating the necessity of our proposed metrics for selecting appropriate agent architectures based on specific application constraints.

arXiv:2603.13017v1 Announce Type: new Abstract: Long conversations with an AI agent create a simple problem for one user: the history is useful, but carrying it verbatim is expensive. We study personalized agent memory: one user's conversation history with an agent, distilled into a compact retrieval layer for later search. Each exchange is compressed into a compound object with four fields (exchange_core, specific_context, thematic room_assignments, and regex-extracted files_touched). The searchable distilled text averages 38 tokens per exchange. Applied to 4,182 conversations (14,340 exchanges) from 6 software engineering projects, the method reduces average exchange length from 371 to 38 tokens, yielding 11x compression. We evaluate whether personalized recall survives that compression using 201 recall-oriented queries, 107 configurations spanning 5 pure and 5 cross-layer search modes, and 5 LLM graders (214,519 consensus-graded query-result pairs). The best pure distilled configuration reaches 96% of the best verbatim MRR (0.717 vs 0.745). Results are mechanism-dependent. All 20 vector search configurations remain non-significant after Bonferroni correction, while all 20 BM25 configurations degrade significantly (effect sizes |d|=0.031-0.756). The best cross-layer setup slightly exceeds the best pure verbatim baseline (MRR 0.759). Structured distillation compresses single-user agent memory without uniformly sacrificing retrieval quality. At 1/11 the context cost, thousands of exchanges fit within a single prompt while the verbatim source remains available for drill-down. We release the implementation and analysis pipeline as open-source software.

arXiv:2603.13168v1 Announce Type: new Abstract: The ability to provide trustworthy maternal health information using phone-based chatbots can have a significant impact, particularly in low-resource settings where users have low health literacy and limited access to care. However, deploying such systems is technically challenging: user queries are short, underspecified, and code-mixed across languages, answers require regional context-specific grounding, and partial or missing symptom context makes safe routing decisions difficult. We present a chatbot for maternal health in India developed through a partnership between academic researchers, a health tech company, a public health nonprofit, and a hospital. The system combines (1) stage-aware triage, routing high-risk queries to expert templates, (2) hybrid retrieval over curated maternal/newborn guidelines, and (3) evidence-conditioned generation from an LLM. Our core contribution is an evaluation workflow for high-stakes deployment under limited expert supervision. Targeting both component-level and end-to-end testing, we introduce: (i) a labeled triage benchmark (N=150) achieving 86.7% emergency recall, explicitly reporting the missed-emergency vs. over-escalation trade-off; (ii) a synthetic multi-evidence retrieval benchmark (N=100) with chunk-level evidence labels; (iii) LLM-as-judge comparison on real queries (N=781) using clinician-codesigned criteria; and (iv) expert validation. Our findings show that trustworthy medical assistants in multilingual, noisy settings require defense-in-depth design paired with multi-method evaluation, rather than any single model and evaluation method choice.

arXiv:2603.13173v1 Announce Type: new Abstract: Large Language Models (LLMs) increasingly serve as autonomous reasoning agents in decision support, scientific problem-solving, and multi-agent coordination systems. However, deploying LLM agents in consequential applications requires assurance that their reasoning remains stable under semantically equivalent input variations, a property we term semantic invariance.Standard benchmark evaluations, which assess accuracy on fixed, canonical problem formulations, fail to capture this critical reliability dimension. To address this shortcoming, in this paper we present a metamorphic testing framework for systematically assessing the robustness of LLM reasoning agents, applying eight semantic-preserving transformations (identity, paraphrase, fact reordering, expansion, contraction, academic context, business context, and contrastive formulation) across seven foundation models spanning four distinct architectural families: Hermes (70B, 405B), Qwen3 (30B-A3B, 235B-A22B), DeepSeek-R1, and gpt-oss (20B, 120B). Our evaluation encompasses 19 multi-step reasoning problems across eight scientific domains. The results reveal that model scale does not predict robustness: the smaller Qwen3-30B-A3B achieves the highest stability (79.6% invariant responses, semantic similarity 0.91), while larger models exhibit greater fragility.

arXiv:2603.12270v1 Announce Type: cross Abstract: Knowledge distillation from large language models (LLMs) assumes that the teacher's output distribution is a high-quality training signal. On reasoning tasks, this assumption is frequently violated. A model's intermediate representations may encode the correct answer, yet this information is lost or distorted through the vocabulary projection, where prompt formatting and answer-token choices creates brittle, noisy outputs. We introduce \method{}, a distillation framework that bypasses this bottleneck by training lightweight probes on frozen teacher hidden states and using the probe's predictions, rather than output logits, as supervision for student training. This simple change yields consistent improvements across four reasoning benchmarks (AQuA-RAT, ARC Easy/Challenge, and MMLU), with gains most pronounced under limited data. Probes trained on intermediate representations provide cleaner labels than the teacher's own outputs, effectively denoising the distillation signal. \method{} requires no architectural changes to student or teacher, is architecture-agnostic, and adds minimal compute since probe training is cheap and teacher representations can be cached. By exploiting internal representations, \method{} enables practitioners to extract more value from large teacher models without additional training data or architectural complexity.

arXiv:2603.12271v1 Announce Type: cross Abstract: LLMs are widely used in knowledge-intensive tasks where the same fact may be revised multiple times within context. Unlike prior work focusing on one-shot updates or single conflicts, multi-update scenarios contain multiple historically valid versions that compete at retrieval, yet remain underexplored. This challenge resembles the AB-AC interference paradigm in cognitive psychology: when the same cue A is successively associated with B and C, the old and new associations compete during retrieval, leading to bias. Inspired by this, we introduce a Dynamic Knowledge Instance (DKI) evaluation framework, modeling multi-updates of the same fact as a cue paired with a sequence of updated values, and assess models via endpoint probing of the earliest (initial) and latest (current) states. Across diverse LLMs, we observe that retrieval bias intensifies as updates increase, earliest-state accuracy stays high while latest-state accuracy drops substantially. Diagnostic analyses of attention, hidden-state similarity, and output logits further reveal that these signals become flatter and weakly discriminative on errors, providing little stable basis for identifying the latest update. Finally, cognitively inspired heuristic intervention strategies yield only modest gains and do not eliminate the bias. Our results reveal a persistent challenge in tracking and following knowledge updates in long contexts.

arXiv:2603.12273v1 Announce Type: cross Abstract: Multi-turn user interactions are among the most abundant data produced by language models, yet we lack effective methods to learn from them. While typically discarded, these interactions often contain useful information: follow-up user messages may indicate that a response was incorrect, failed to follow an instruction, or did not align with the user's preferences. Importantly, language models are already able to make use of this information in context. After observing a user's follow-up, the same model is often able to revise its behavior. We leverage this ability to propose a principled and scalable method for learning directly from user interactions through self-distillation. By conditioning the model on the user's follow-up message and comparing the resulting token distribution with the original policy, we obtain a target for updating the policy that captures how the model's behavior changes in hindsight. We then distill this hindsight distribution back into the current policy. Remarkably, we show that training on real-world user conversations from WildChat improves language models across standard alignment and instruction-following benchmarks, without regressing other capabilities. The same mechanism enables personalization, allowing models to continually adapt to individual users through interaction without explicit feedback. Our results demonstrate that raw user interactions that arise naturally during deployment enable alignment, personalization, and continual adaptation.

arXiv:2603.12277v1 Announce Type: cross Abstract: Language models remain vulnerable to prompt injection attacks despite extensive safety training. We trace this failure to role confusion: models infer roles from how text is written, not where it comes from. We design novel role probes to capture how models internally identify "who is speaking." These reveal why prompt injection works: untrusted text that imitates a role inherits that role's authority. We test this insight by injecting spoofed reasoning into user prompts and tool outputs, achieving average success rates of 60% on StrongREJECT and 61% on agent exfiltration, across multiple open- and closed-weight models with near-zero baselines. Strikingly, the degree of internal role confusion strongly predicts attack success before generation begins. Our findings reveal a fundamental gap: security is defined at the interface but authority is assigned in latent space. More broadly, we introduce a unifying, mechanistic framework for prompt injection, demonstrating that diverse prompt-injection attacks exploit the same underlying role-confusion mechanism.

arXiv:2603.12458v1 Announce Type: cross Abstract: While Large Language Models (LLMs) achieve expert-level performance on standard medical benchmarks through single-hop factual recall, they severely struggle with the complex, multi-hop diagnostic reasoning required in real-world clinical settings. A primary obstacle is "shortcut learning", where models exploit highly connected, generic hub nodes (e.g., "inflammation") in knowledge graphs to bypass authentic micro-pathological cascades. To address this, we introduce ShatterMed-QA, a bilingual benchmark of 10,558 multi-hop clinical questions designed to rigorously evaluate deep diagnostic reasoning. Our framework constructs a topology-regularized medical Knowledge Graph using a novel $k$-Shattering algorithm, which physically prunes generic hubs to explicitly sever logical shortcuts. We synthesize the evaluation vignettes by applying implicit bridge entity masking and topology-driven hard negative sampling, forcing models to navigate biologically plausible distractors without relying on superficial elimination. Comprehensive evaluations of 21 LLMs reveal massive performance degradation on our multi-hop tasks, particularly among domain-specific models. Crucially, restoring the masked evidence via Retrieval-Augmented Generation (RAG) triggers near-universal performance recovery, validating ShatterMed-QA's structural fidelity and proving its efficacy in diagnosing the fundamental reasoning deficits of current medical AI. Explore the dataset, interactive examples, and full leaderboards at our project website: https://shattermed-qa-web.vercel.app/

arXiv:2603.12510v1 Announce Type: cross Abstract: Vision-Language-Action (VLA) models have significant potential to enable general-purpose robotic systems for a range of vision-language tasks. However, the performance of VLA-based robots is highly sensitive to the precise wording of language instructions, and it remains difficult to predict when such robots will fail. To improve the robustness of VLAs to different wordings, we present Q-DIG (Quality Diversity for Diverse Instruction Generation), which performs red-teaming by scalably identifying diverse natural language task descriptions that induce failures while remaining task-relevant. Q-DIG integrates Quality Diversity (QD) techniques with Vision-Language Models (VLMs) to generate a broad spectrum of adversarial instructions that expose meaningful vulnerabilities in VLA behavior. Our results across multiple simulation benchmarks show that Q-DIG finds more diverse and meaningful failure modes compared to baseline methods, and that fine-tuning VLAs on the generated instructions improves task success rates. Furthermore, results from a user study highlight that Q-DIG generates prompts judged to be more natural and human-like than those from baselines. Finally, real-world evaluations of Q-DIG prompts show results consistent with simulation, and fine-tuning VLAs on the generated prompts further success rates on unseen instructions. Together, these findings suggest that Q-DIG is a promising approach for identifying vulnerabilities and improving the robustness of VLA-based robots. Our anonymous project website is at qdigvla.github.io.

arXiv:2603.12520v1 Announce Type: cross Abstract: Large language models are often used as judges to score candidate responses, then validated with a single global metric such as correlation with reference labels. This can be misleading when the real deployment task is best-of-n selection within a prompt. In a 5,000-prompt best-of-4 benchmark from Chatbot Arena, a judge with moderate global correlation (r = 0.47) captures only 21.0% of the improvement that perfect selection would achieve over random choice. The gap arises because global agreement is driven largely by prompt-level baseline effects, while selection depends on within-prompt ranking: within-prompt correlation is only r_within = 0.27, and coarse pointwise scoring creates ties in 67% of pairwise comparisons. In a matched-pair best-of-2 audit, explicit pairwise judging recovers much of this lost signal, raising recovery from 21.1% to 61.2%. For judge-based selection, the relevant audit should report within-prompt signal, tie rates, and recovery/top-1 accuracy, not global agreement alone.

arXiv:2603.12522v1 Announce Type: cross Abstract: As large language models (LLMs) are deployed widely, detecting and understanding bias in their outputs is critical. We present LLM BiasScope, a web application for side-by-side comparison of LLM outputs with real-time bias analysis. The system supports multiple providers (Google Gemini, DeepSeek, MiniMax, Mistral, Meituan, Meta Llama) and enables researchers and practitioners to compare models on the same prompts while analyzing bias patterns. LLM BiasScope uses a two-stage bias detection pipeline: sentence-level bias detection followed by bias type classification for biased sentences. The analysis runs automatically on both user prompts and model responses, providing statistics, visualizations, and detailed breakdowns of bias types. The interface displays two models side-by-side with synchronized streaming responses, per-model bias summaries, and a comparison view highlighting differences in bias distributions. The system is built on Next.js with React, integrates Hugging Face inference endpoints for bias detection, and uses the Vercel AI SDK for multi-provider LLM access. Features include real-time streaming, export to JSON/PDF, and interactive visualizations (bar charts, radar charts) for bias analysis. LLM BiasScope is available as an open-source web application, providing a practical tool for bias evaluation and comparative analysis of LLM behaviour.

arXiv:2603.12529v1 Announce Type: cross Abstract: Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design TERMINATOR, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning TERMINATOR is that the first arrival of an LRM's final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train TERMINATOR. Powered by this approach, TERMINATOR achieves significant reductions in CoT lengths of 14%-55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, whilst outperforming current state-of-the-art methods.

arXiv:2603.12554v1 Announce Type: cross Abstract: Reinforcement learning (RL) has been effective for post-training autoregressive (AR) language models, but extending these methods to diffusion language models (DLMs) is challenging due to intractable sequence-level likelihoods. Existing approaches therefore rely on surrogate likelihoods or heuristic approximations, which can introduce bias and obscure the sequential structure of denoising. We formulate diffusion-based sequence generation as a finite-horizon Markov decision process over the denoising trajectory and derive an exact, unbiased policy gradient that decomposes over denoising steps and is expressed in terms of intermediate advantages, without requiring explicit evaluation of the sequence likelihood. To obtain a practical and compute-efficient estimator, we (i) select denoising steps for policy updates via an entropy-guided approximation bound, and (ii) estimate intermediate advantages using a one-step denoising reward naturally provided by the diffusion model, avoiding costly multi-step rollouts. Experiments on coding and logical reasoning benchmarks demonstrate state-of-the-art results, with strong competitive performance on mathematical reasoning, outperforming existing RL post-training approaches for DLMs. Code is available at https://github.com/vishnutez/egspo-dllm-rl.

arXiv:2603.12564v1 Announce Type: cross Abstract: Tool-augmented LLM agents increasingly serve as multi-turn advisors in high-stakes domains, yet their evaluation relies on ranking-quality metrics that measure what is recommended but not whether it is safe for the user. We introduce a paired-trajectory protocol that replays real financial dialogues under clean and contaminated tool-output conditions across seven LLMs (7B to frontier) and decomposes divergence into information-channel and memory-channel mechanisms. Across the seven models tested, we consistently observe the evaluation-blindness pattern: recommendation quality is largely preserved under contamination (utility preservation ratio approximately 1.0) while risk-inappropriate products appear in 65-93% of turns, a systematic safety failure poorly reflected by standard NDCG. Safety violations are predominantly information-channel-driven, emerge at the first contaminated turn, and persist without self-correction over 23-step trajectories; no agent across 1,563 contaminated turns explicitly questions tool-data reliability. Even narrative-only corruption (biased headlines, no numerical manipulation) induces significant drift while completely evading consistency monitors. A safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74, indicating that much of the evaluation gap becomes visible once safety is explicitly measured. These results motivate considering trajectory-level safety monitoring, beyond single-turn quality, for deployed multi-turn agents in high-stakes settings.

arXiv:2603.12615v1 Announce Type: cross Abstract: Existing AI moral evaluation frameworks test for the production of correct-sounding ethical responses rather than the presence of genuine moral reasoning capacity. This paper introduces a novel probe methodology using literary narrative - specifically, unresolvable moral scenarios drawn from a published science fiction series - as stimulus material structurally resistant to surface performance. We present results from a 24-condition cross-system study spanning 13 distinct systems across two series: Series 1 (frontier commercial systems, blind; n=7) and Series 2 (local and API open-source systems, blind and declared; n=6). Four Series 2 systems were re-administered under declared conditions (13 blind + 4 declared + 7 ceiling probe = 24 total conditions), yielding zero delta across all 16 dimension-pair comparisons. Probe administration was conducted by two human raters across three machines; primary blind scoring was performed by Claude (Anthropic) as LLM judge, with Gemini Pro (Google) and Copilot Pro (Microsoft) serving as independent judges for the ceiling discrimination probe. A supplemental theological differentiator probe yielded perfect rank-order agreement between the two independent ceiling probe judges (Gemini Pro and Copilot Pro; rs = 1.00). Five qualitatively distinct D3 reflexive failure modes were identified - including categorical self-misidentification and false positive self-attribution - suggesting that instrument sophistication scales with system capability rather than being circumvented by it. We argue that literary narrative constitutes an anticipatory evaluation instrument - one that becomes more discriminating as AI capability increases - and that the gap between performed and authentic moral reasoning is measurable, meaningful, and consequential for deployment decisions in high-stakes domains.

arXiv:2603.12658v1 Announce Type: cross Abstract: Continual learning (CL) has emerged as a pivotal paradigm to enable large language models (LLMs) to dynamically adapt to evolving knowledge and sequential tasks while mitigating catastrophic forgetting-a critical limitation of the static pre-training paradigm inherent to modern LLMs. This survey presents a comprehensive overview of CL methodologies tailored for LLMs, structured around three core training stages: continual pre-training, continual fine-tuning, and continual alignment.Beyond the canonical taxonomy of rehearsal-, regularization-, and architecture-based methods, we further subdivide each category by its distinct forgetting mitigation mechanisms and conduct a rigorous comparative analysis of the adaptability and critical improvements of traditional CL methods for LLMs. In doing so, we explicitly highlight core distinctions between LLM CL and traditional machine learning, particularly with respect to scale, parameter efficiency, and emergent capabilities. Our analysis covers essential evaluation metrics, including forgetting rates and knowledge transfer efficiency, along with emerging benchmarks for assessing CL performance. This survey reveals that while current methods demonstrate promising results in specific domains, fundamental challenges persist in achieving seamless knowledge integration across diverse tasks and temporal scales. This systematic review contributes to the growing body of knowledge on LLM adaptation, providing researchers and practitioners with a structured framework for understanding current achievements and future opportunities in lifelong learning for language models.

arXiv:2603.12664v1 Announce Type: cross Abstract: Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.