arXiv:2602.19159v1 Announce Type: new Abstract: Prior behavioural work suggests that some LLMs alter choices when options are framed as causing pain or pleasure, and that such deviations can scale with stated intensity. To bridge behavioural evidence (what the model does) with mechanistic interpretability (what computations support it), we investigate how valence-related information is represented and where it is causally used inside a transformer. Using Gemma-2-9B-it and a minimalist decision task modelled on prior work, we (i) map representational availability with layer-wise linear probing across streams, (ii) test causal contribution with activation interventions (steering; patching/ablation), and (iii) quantify dose-response effects over an epsilon grid, reading out both the 2-3 logit margin and digit-pair-normalised choice probabilities. We find that (a) valence sign (pain vs. pleasure) is perfectly linearly separable across stream families from very early layers (L0-L1), while a lexical baseline retains substantial signal; (b) graded intensity is strongly decodable, with peaks in mid-to-late layers and especially in attention/MLP outputs, and decision alignment is highest slightly before the final token; (c) additive steering along a data-derived valence direction causally modulates the 2-3 margin at late sites, with the largest effects observed in late-layer attention outputs (attn_out L14); and (d) head-level patching/ablation suggests that these effects are distributed across multiple heads rather than concentrated in a single unit. Together, these results link behavioural sensitivity to identifiable internal representations and intervention-sensitive sites, providing concrete mechanistic targets for more stringent counterfactual tests and broader replication. This work supports a more evidence-driven (a) debate on AI sentience and welfare, and (b) governance when setting policy, auditing standards, and safety safeguards.

arXiv:2602.19160v1 Announce Type: new Abstract: This paper examines the reasoning capabilities of Large Language Models (LLMs) from a novel perspective, focusing on their ability to operate within formally specified, rule-governed environments. We evaluate four LLMs (Gemini 2.5 Pro and Flash variants, Llama 3.3 70B and GPT-OSS 120B) on a suite of forward-simulation tasks-including next / multistep state formulation, and legal action generation-across a diverse set of reasoning problems illustrated through General Game Playing (GGP) game instances. Beyond reporting instance-level performance, we characterize games based on 40 structural features and analyze correlations between these features and LLM performance. Furthermore, we investigate the effects of various game obfuscations to assess the role of linguistic semantics in game definitions and the impact of potential prior exposure of LLMs to specific games during training. The main results indicate that three of the evaluated models generally perform well across most experimental settings, with performance degradation observed as the evaluation horizon increases (i.e., with a higher number of game steps). Detailed case-based analysis of the LLM performance provides novel insights into common reasoning errors in the considered logic-based problem formulation, including hallucinated rules, redundant state facts, or syntactic errors. Overall, the paper reports clear progress in formal reasoning capabilities of contemporary models.

arXiv:2602.19517v1 Announce Type: new Abstract: We introduce \CFE{} (\textbf{C}lassroom \textbf{F}inal \textbf{E}xam), a multimodal benchmark for evaluating the reasoning capabilities of large language models across more than 20 STEM domains. \CFE{} is curated from repeatedly used, authentic university homework and exam problems, together with reference solutions provided by course instructors. \CFE{} presents a significant challenge even for frontier models: the newly released Gemini-3.1-pro-preview achieves an overall accuracy of 59.69\%, while the second-best model, Gemini-3-flash-preview, reaches 55.46\%, leaving considerable room for improvement. Beyond leaderboard results, we perform a diagnostic analysis by decomposing reference solutions into reasoning flows. We find that although frontier models can often answer intermediate sub-questions correctly, they struggle to reliably derive and maintain correct intermediate states throughout multi-step solutions. We further observe that model-generated solutions typically have more reasoning steps than those provided by the instructor, indicating suboptimal step efficiency and a higher risk of error accumulation. The data and code are available at https://github.com/Analogy-AI/CFE_Bench.

arXiv:2602.18443v1 Announce Type: cross Abstract: Psychosocial online counselling frequently encounters generic subject lines that impede efficient case prioritisation. This study evaluates eleven large language models generating six-word subject lines for German counselling emails through hierarchical assessment - first categorising outputs, then ranking within categories to enable manageable evaluation. Nine assessors (counselling professionals and AI systems) enable analysis via Krippendorff's $\alpha$, Spearman's $\rho$, Pearson's $r$ and Kendall's $\tau$. Results reveal performance trade-offs between proprietary services and privacy-preserving open-source alternatives, with German fine-tuning consistently improving performance. The study addresses critical ethical considerations for mental health AI deployment including privacy, bias and accountability.

arXiv:2602.18446v1 Announce Type: cross Abstract: Users increasingly rely on Large Language Models (LLMs) for Deep Research, using them to synthesize diverse sources into structured reports that support understanding and action. In this context, the practical reliability of such reports hinges on logical quality: whether the report's claims and arguments are explicitly supported and can be trusted as a basis for downstream use, rather than merely appearing fluent or informative. However, current evaluation frameworks largely overlook this requirement. To bridge this gap, we introduce ReportLogic, a benchmark that quantifies report-level logical quality through a reader-centric lens of auditability. Specifically, ReportLogic adopts a hierarchical taxonomy that evaluates whether readers can (1) trace an on-topic report structure with a unified analytical arc (Macro-Logic), (2) understand the progression with necessary context (Expositional-Logic), and (3) verify conclusions via explicit claim--support (Structural-Logic). Based on this taxonomy, we construct a human-annotated rubric-guided dataset and train an open-source LogicJudge for scalable evaluation. We further evaluate judge robustness via adversarial attacks, showing that off-the-shelf LLM judges are frequently influenced by superficial cues (e.g., verbosity), and reasoning modes can mask broken support relations. Overall, our results provide actionable guidance for building more robust logic evaluators and improving the logical reliability of LLM-generated reports.

arXiv:2602.18447v1 Announce Type: cross Abstract: Chain-of-Thought reasoning significantly improves the performance of large language models on complex tasks, but incurs high inference latency due to long generation traces. Step-level speculative reasoning aims to mitigate this cost, yet existing approaches face a long-standing trade-off among accuracy, inference speed, and resource efficiency. We propose ConfSpec, a confidence-gated cascaded verification framework that resolves this trade-off. Our key insight is an asymmetry between generation and verification: while generating a correct reasoning step requires substantial model capacity, step-level verification is a constrained discriminative task for which small draft models are well-calibrated within their competence range, enabling high-confidence draft decisions to be accepted directly while selectively escalating uncertain cases to the large target model. Evaluation across diverse workloads shows that ConfSpec achieves up to 2.24$\times$ end-to-end speedups while matching target-model accuracy. Our method requires no external judge models and is orthogonal to token-level speculative decoding, enabling further multiplicative acceleration.

arXiv:2602.18449v1 Announce Type: cross Abstract: We propose a diffusion-based framework for prompt optimization that leverages Diffusion Language Models (DLMs) to iteratively refine system prompts through masked denoising. By conditioning on interaction traces, including user queries, model responses, and optional feedback, our method enables flexible, span-level prompt updates without requiring gradient access or modifying the downstream language model. Across diverse benchmarks (e.g., $\tau$-bench, SST-2, SST-5), DLM-optimized prompts consistently improve the performance of a frozen target LLM (e.g., GPT-4o-mini). We further show that moderate diffusion step counts provide the best balance between refinement quality and stability. These results highlight diffusion-based prompt optimization as a general, model-agnostic, and scalable approach for enhancing LLM performance through iterative prompt refinement.

arXiv:2602.18458v1 Announce Type: cross Abstract: Reproducibility crises across sciences highlight the limitations of the paper-centric review system in assessing the rigor and reproducibility of research. AI agents that autonomously design and generate large volumes of research outputs exacerbate these challenges. In this work, we address the growing challenges of scalability and rigor by flipping the dynamic and developing AI agents as research evaluators. We propose the first execution-grounded evaluation framework that verifies research beyond narrative review by examining code and data alongside the paper. We use mechanistic interpretability research as a testbed, build standardized research output, and develop MechEvalAgent, an automated evaluation framework that assesses the coherence of the experimental process, the reproducibility of results, and the generalizability of findings. We show that our framework achieves above 80% agreement with human judges, identifies substantial methodological problems, and surfaces 51 additional issues that human reviewers miss. Our work demonstrates the potential of AI agents to transform research evaluation and pave the way for rigorous scientific practices.

arXiv:2602.18464v1 Announce Type: cross Abstract: A growing body of research assumes that large language model (LLM) agents can serve as proxies for how people form attitudes toward and behave in response to security and privacy (S&P) threats. If correct, these simulations could offer a scalable way to forecast S&P risks in products prior to deployment. We interrogate this assumption using SP-ABCBench, a new benchmark of 30 tests derived from validated S&P human-subject studies, which measures alignment between simulations and human-subjects studies on a 0-100 ascending scale, where higher scores indicate better alignment across three dimensions: Attitude, Behavior, and Coherence. Evaluating twelve LLMs, four persona construction strategies, and two prompting methods, we found that there remains substantial room for improvement: all models score between 50 and 64 on average. Newer, bigger, and smarter models do not reliably do better and sometimes do worse. Some simulation configurations, however, do yield high alignment: e.g., with scores above 95 for some behavior tests when agents are prompted to apply bounded rationality and weigh privacy costs against perceived benefits. We release SP-ABCBench to enable reproducible evaluation as methods improve.

arXiv:2602.18483v1 Announce Type: cross Abstract: Recently, red teaming, with roots in security, has become a key evaluative approach to ensure the safety and reliability of Generative Artificial Intelligence. However, most existing work emphasizes technical benchmarks and attack success rates, leaving the socio-technical practices of how red teaming datasets are defined, created, and evaluated under-examined. Drawing on 22 interviews with practitioners who design and evaluate red teaming datasets, we examine the data practices and standards that underpin this work. Because adversarial datasets determine the scope and accuracy of model evaluations, they are critical artifacts for assessing potential harms from large language models. Our contributions are first, empirical evidence of practitioners conceptualizing red teaming and developing and evaluating red teaming datasets. Second, we reflect on how practitioners' conceptualization of risk leads to overlooking the context, interaction type, and user specificity. We conclude with three opportunities for HCI researchers to expand the conceptualization and data practices for red-teaming.

arXiv:2602.18492v1 Announce Type: cross Abstract: Large Language Models (LLMs) are now good enough at coding that developers can describe intent in plain language and let the tool produce the first code draft, a workflow increasingly built into tools like GitHub Copilot, Cursor, and Replit. What is missing is a reliable way to tell which model written queries are safe to accept without sending everything to a human. We study the application of an LLM jury to run this review step. We first benchmark 15 open models on 82 MySQL text to SQL tasks using an execution grounded protocol to get a clean baseline of which models are strong. From the six best models we build unanimous committees of sizes 1 through 6 that see the prompt, schema, and candidate SQL and accept it only when every member says it is correct. This rule matches safety first deployments where false accepts are more costly than false rejects. We measure true positive rate, false positive rate and Youden J and we also look at committees per generator. Our results show that single model judges are uneven, that small unanimous committees of strong models can cut false accepts while still passing many good queries, and that the exact committee composition matters significantly.

arXiv:2602.18583v1 Announce Type: cross Abstract: Real-time guardrails require evaluation that is accurate, cheap, and fast - yet today's default, LLM-as-a-judge (LLMAJ), is slow, expensive, and operationally non-deterministic due to multi-token generation. We present Luna-2, a novel architecture that leverages decoder-only small language models (SLMs) into a deterministic evaluation model to reliably compute complex task-specific LLMAJ metrics (e.g. toxicity, hallucination, tool selection quality, etc.) at an accuracy at par or higher than LLMAJ using frontier LLMs while drastically reducing the cost and latency of computation. Each metric is implemented as a lightweight LoRA/PEFT head on top of a shared SLM backbone, enabling hundreds of specialized metrics to run concurrently on a single GPU, deployable locally next to AI systems in a privacy-preserving and latency optimizing manner. Across content safety and hallucination benchmarks, Luna-2 matches the accuracy of state-of-the-art LLM-based evaluators while reducing inference cost by over 80x and latency by over 20x. In this paper, we outline the model architecture, training methodology and report real-world empirical results on accuracy, latency, and throughput results. In production, Luna-2 is protecting 100M+ AI sessions and processing over 100B tokens per month for our customers with eval cost savings of over $30M annually.

arXiv:2602.18699v1 Announce Type: cross Abstract: Most semantic drift studies report multiple signals e.g., embedding displacement, neighbor changes, distributional divergence, and recursive trajectory instability, without a shared explanatory theory that relates them. This paper proposes a formalization of these signals in one time-indexed substrate, $S_t=(X,d_t,P_t)$, combining embedding geometry with local diffusion. Within this substrate, node-level neighborhood drift measures changes in local conditional distributions, coarse Ricci curvature measures local contractivity of semantic diffusion, and recursive drift probes stability of iterated semantic operators. This manuscript specifies the formal model, assumptions, and tests that can refute the model. Herein, the paper introduces bridge mass, a node-level aggregate of incident negative curvature, as a predictor of future neighborhood rewiring. This paper provides the theory and test contracts; empirical performance is deferred to subsequent studies.

arXiv:2602.18734v1 Announce Type: cross Abstract: Retrieval-Augmented Generation (RAG) has demonstrated strong effectiveness in knowledge-intensive tasks by grounding language generation in external evidence. Despite its success, many existing RAG systems are built based on a ranking-centric, asymmetric dependency paradigm, where the generation quality of the generator is highly dependent on reranking results of the reranker. To overcome this limitation, we reformulate RAG as a cooperative multi-agent decision-making problem and propose Cooperative Retrieval-Augmented Generation (CoRAG), a framework in which the reranker and the generator act as peer decision-makers rather than being connected through an asymmetric dependency pipeline. By jointly optimizing their behaviors toward a shared task objective, the reranker and generator are encouraged to cooperate, ensuring that document reranking and generation work in concert to improve the final response. Experimental results demonstrate good generalization and improved generation stability of CoRAG, even when the model is trained on only around 10K PopQA samples. Our model released in https://anonymous.4open.science/r/CoRAG-D63F

arXiv:2602.18776v1 Announce Type: cross Abstract: We present ArabicNumBench, a comprehensive benchmark for evaluating large language models on Arabic number reading tasks across Eastern Arabic-Indic numerals (0-9 in Arabic script) and Western Arabic numerals (0-9). We evaluate 71 models from 10 providers using four prompting strategies (zero-shot, zero-shot CoT, few-shot, few-shot CoT) on 210 number reading tasks spanning six contextual categories: pure numerals, addresses, dates, quantities, and prices. Our evaluation comprises 59,010 individual test cases and tracks extraction methods to measure structured output generation. Evaluation reveals substantial performance variation, with accuracy ranging from 14.29\% to 99.05\% across models and strategies. Few-shot Chain-of-Thought prompting achieves 2.8x higher accuracy than zero-shot approaches (80.06\% vs 28.76\%). A striking finding emerges: models achieving elite accuracy (98-99\%) often produce predominantly unstructured output, with most responses lacking Arabic CoT markers. Only 6 models consistently generate structured output across all test cases, while the majority require fallback extraction methods despite high numerical accuracy. Comprehensive evaluation of 281 model-strategy combinations demonstrates that numerical accuracy and instruction-following represent distinct capabilities, establishing baselines for Arabic number comprehension and providing actionable guidance for model selection in production Arabic NLP systems.

arXiv:2602.18782v1 Announce Type: cross Abstract: Defending LLMs against adversarial jailbreak attacks remains an open challenge. Existing defenses rely on binary classifiers that fail when adversarial input falls outside the learned decision boundary, and repeated fine-tuning is computationally expensive while potentially degrading model capabilities. We propose MANATEE, an inference-time defense that uses density estimation over a benign representation manifold. MANATEE learns the score function of benign hidden states and uses diffusion to project anomalous representations toward safe regions--requiring no harmful training data and no architectural modifications. Experiments across Mistral-7B-Instruct, Llama-3.1-8B-Instruct, and Gemma-2-9B-it demonstrate that MANATEE reduce Attack Success Rate by up to 100\% on certain datasets, while preserving model utility on benign inputs.

arXiv:2602.18806v1 Announce Type: cross Abstract: Large Language Models (LLMs) demonstrate strong reasoning performance, yet their ability to reliably monitor, diagnose, and correct their own errors remains limited. We introduce a psychologically grounded metacognitive framework that operationalizes Ann Brown's regulatory cycle (Planning, Monitoring, and Evaluation) as a structured prompting architecture, and study its integration within a lightweight dual-process MetaController for adaptive effort allocation. Across diverse reasoning and diagnostic benchmarks (GSM8K, CRUXEval, MBPP, AIME, CorrectBench, and TruthfulQA) using Llama-3 and Qwen-3 (8B), explicit regulatory structuring substantially improves error diagnosis and yields a threefold increase in successful self-correction. Blinded human evaluations over 580 query pairs show an 84% aggregate preference for trustworthiness and metacognitive self-awareness over standard and Chain-of-Thought baselines. Grounding LLM reasoning in established cognitive theory offers a principled path toward more transparent and diagnostically robust AI systems.

arXiv:2602.18905v1 Announce Type: cross Abstract: Large language models (LLMs) have demonstrated strong capabilities in complex reasoning tasks, yet their decision-making processes remain difficult to interpret. Existing explanation methods often lack trustworthy structural insight and are limited to single-instance analysis, failing to reveal reasoning stability and systematic failure mechanisms. To address these limitations, we propose the Trustworthy Unified Explanation Framework (TRUE), which integrates executable reasoning verification, feasible-region directed acyclic graph (DAG) modeling, and causal failure mode analysis. At the instance level, we redefine reasoning traces as executable process specifications and introduce blind execution verification to assess operational validity. At the local structural level, we construct feasible-region DAGs via structure-consistent perturbations, enabling explicit characterization of reasoning stability and the executable region in the local input space. At the class level, we introduce a causal failure mode analysis method that identifies recurring structural failure patterns and quantifies their causal influence using Shapley values. Extensive experiments across multiple reasoning benchmarks demonstrate that the proposed framework provides multi-level, verifiable explanations, including executable reasoning structures for individual instances, feasible-region representations for neighboring inputs, and interpretable failure modes with quantified importance at the class level. These results establish a unified and principled paradigm for improving the interpretability and reliability of LLM reasoning systems.

arXiv:2602.18915v1 Announce Type: cross Abstract: Adeno-associated viruses (AAVs) are promising vectors for gene therapy, but their native serotypes face limitations in tissue tropism, immune evasion, and production efficiency. Engineering capsids to overcome these hurdles is challenging due to the vast sequence space and the difficulty of simultaneously optimizing multiple functional properties. The complexity also adds when it comes to the kidney, which presents unique anatomical barriers and cellular targets that require precise and efficient vector engineering. Here, we present AAVGen, a generative artificial intelligence framework for de novo design of AAV capsids with enhanced multi-trait profiles. AAVGen integrates a protein language model (PLM) with supervised fine-tuning (SFT) and a reinforcement learning technique termed Group Sequence Policy Optimization (GSPO). The model is guided by a composite reward signal derived from three ESM-2-based regression predictors, each trained to predict a key property: production fitness, kidney tropism, and thermostability. Our results demonstrate that AAVGen produces a diverse library of novel VP1 protein sequences. In silico validations revealed that the majority of the generated variants have superior performance across all three employed indices, indicating successful multi-objective optimization. Furthermore, structural analysis via AlphaFold3 confirms that the generated sequences preserve the canonical capsid folding despite sequence diversification. AAVGen establishes a foundation for data-driven viral vector engineering, accelerating the development of next-generation AAV vectors with tailored functional characteristics.

arXiv:2602.18920v1 Announce Type: cross Abstract: The application of Large Language Models (LLMs) in accelerating scientific discovery has garnered increasing attention, with a key focus on constructing research agents endowed with innovative capability, i.e., the ability to autonomously generate novel and significant research ideas. Existing approaches predominantly rely on sophisticated prompt engineering and lack a systematic training paradigm. To address this, we propose DeepInnovator, a training framework designed to trigger the innovative capability of LLMs. Our approach comprises two core components. (1) ``Standing on the shoulders of giants''. We construct an automated data extraction pipeline to extract and organize structured research knowledge from a vast corpus of unlabeled scientific literature. (2) ``Conjectures and refutations''. We introduce a ``Next Idea Prediction'' training paradigm, which models the generation of research ideas as an iterative process of continuously predicting, evaluating, and refining plausible and novel next idea. Both automatic and expert evaluations demonstrate that our DeepInnovator-14B significantly outperforms untrained baselines, achieving win rates of 80.53\%-93.81\%, and attains performance comparable to that of current leading LLMs. This work provides a scalable training pathway toward building research agents with genuine, originative innovative capability, and will open-source the dataset to foster community advancement. Source code and data are available at: https://github.com/HKUDS/DeepInnovator.