Daily Papers

Vision-Language Models (VLMs) are increasingly used to evaluate robot manipulation outcomes, but existing benchmarks offer limited evidence of cross-domain generalization. We introduce FailBench, a benchmark for robot failure detection comprising 2,197 manipulation attempts across 14 public sources (12 real-world, 2 simulated). In FailBench, 75% of failures occur naturally, and six real-world sources come from non-failure-detection datasets. Evaluating 13 VLM-based detectors, we find the best model achieves only 0.77 mean balanced accuracy. Notably, models fine-tuned for failure detection consistently underperform general-purpose VLMs and their own pretrained baselines. Performance depends heavily on required visual evidence: models approach saturation when outcomes depend on observable object motion, but degrade to near-chance (<0.60 balanced accuracy) on contact-intensive assembly tasks. Error analysis reveals a systematic bias toward predicting success under ambiguous evidence, which persists even with increased reasoning effort. Finally, we show that input-level intervention--spatially localizing and cropping outcome-relevant regions--improves the top detector by 2.4 percentage points without extra training.

Online scene change detection requires integrating incoming observations to determine which regions have changed relative to a reference scene. We update change beliefs as new images arrive using generalised Bayesian evidence fusion. However, the prior probability of change and evidence models are unknown beforehand, and numerical Bayesian inference can be costly. We address these challenges by learning offline from automatic edits to a reference 3D Gaussian Splatting (3DGS) model. These edits generate labelled training views without manual annotations. Together with unchanged reference observations, they provide supervision for learning scene-specific evidence models and estimating the effective pixel prior and response statistics. Online prediction combines current evidence with historical support, while closed-form local Bayes factors, parameterized by the learned response statistics, update an accumulated 3D Gaussian field. Object-guided refinement produces the final change masks. Our method achieves state-of-the-art accuracy on PASLCD and CCS3D, with 67.64% and 75.82% mIoU, respectively. On PASLCD, the complete pipeline runs at 17.2\,FPS over complete query sequences, the fastest among the evaluated online methods. Project page: https://changruizhu96.github.io/work/lego-scd/

Large language models (LLMs) generate fluent text across a wide range of tasks, but the fabrication of non-existent academic citations remains a critical and well-documented failure mode. Building on prior work that frames hallucination and verbatim memorization as outcomes of the same probabilistic process, this study uses citation count as a proxy for training data redundancy and asks how this redundancy is internally structured within a single bibliographic record. Using GPT-4.1, we generated and manually verified 100 citations across twenty computer-science domains, measuring factual fidelity via cosine similarity against authentic metadata. We find that (i) factual accuracy varies substantially across domains and scales log-linearly with citation count, (ii) the model crosses two empirically identifiable thresholds; an inflection around 90 citations and a saturation point near 1,200 citations beyond which records are reproduced nearly verbatim, (iii) memorization is hierarchical, with titles and first authors recalled earliest while venues and numeric fields require far greater redundancy and publication years remain essentially unlearned, and (iv) even highly cited records can be conflated when their titles and authors overlap, an effect interpretable as spurious-attractor interference. Memorization in LLMs is therefore not a binary on/off state but a graduated, hierarchically layered phenomenon shaped by the uneven distribution of knowledge in the pretraining corpus.

Large language models (LLMs) increasingly influence global digital ecosystems, yet their potential to perpetuate social and cultural biases remains poorly understood in underrepresented contexts. This study presents a systematic analysis of representational biases in seven state-of-the-art LLMs: GPT-4o-mini, Claude-3-Sonnet, Claude-4-Sonnet, Gemini-2.0-Flash, Gemini-2.0-Lite, Llama-3-70B, and Mistral-Nemo in the Nepali cultural context. Using Croissant-compliant dataset of 2400+ stereotypical and anti-stereotypical sentence pairs on gender roles across social domains, we implement an evaluation framework, Dual-Metric Bias Assessment (DMBA), combining two metrics: (1) agreement with biased statements and (2) stereotypical completion tendencies. Results show models exhibit measurable explicit agreement bias, with mean bias agreement ranging from 0.36 to 0.43 across decoding configurations, and an implicit completion bias rate of 0.740-0.755. Importantly, implicit completion bias follows a non-linear, U-shaped relationship with temperature, peaking at moderate stochasticity (T=0.3) and declining slightly at higher temperatures. Correlation analysis under different decoding settings revealed that explicit agreement strongly aligns with stereotypical sentence agreement but is a weak and often negative predictor of implicit completion bias, indicating generative bias is poorly captured by agreement metrics. Sensitivity analysis shows increasing top-p amplifies explicit bias, while implicit generative bias remains largely stable. Domain-level analysis shows implicit bias is strongest for race and sociocultural stereotypes, while explicit agreement bias is similar across gender and sociocultural categories, with race showing the lowest explicit agreement. These findings highlight the need for culturally grounded datasets and debiasing strategies for LLMs in underrepresented societies.

Spatial return models take the interaction matrix as given and leave feedback uninterpreted. We construct a bandwidth-free field from firms' language-model article embedding distributions using target-anchored Wasserstein barycentric reconstruction. A quadratic exposure-adjustment problem maps feedback into a peer-misalignment penalty ratio. For 52 firms, the field, frozen from 2018-2022 news, yields a 2023-2026 penalty ratio of 3.46 (95% interval [2.89, 4.17]) and higher conditional quasi-likelihood than equal-weighted peer support or RBF weighting of the same distances. Joint penalty ratios for the barycentric and news co-mention fields are 2.33 and 0.86 with boundary calibrated tests which reject both exclusions.

We investigate whether large language models (LLMs) systematically discriminate in candidate evaluations based on applicant name ethnicity and/or institutional prestige and geographic location. Three factorial experiments are reported (4,320 API calls, four LLMs, five professional domains). Study 1 (3x4 design) finds a statistically robust institution-tier gradient of +0.297 points on a 10-point scale (95% bootstrap CI: +0.175 to +0.422), while name-origin effects are negligible and non-significant (95% CI crosses zero). Study 2 (2x2 Prestige x Country design) breaks the prestige-geography confound: the prestige effect (+0.185; 95% CI: +0.093 to +0.275) exceeds the country-of-origin effect (+0.126; 95% CI: +0.037 to +0.218) by 1.5x. Study 3 (2x2 Journal x Institution design) reveals that journal prestige (Nature vs. a peripheral open-access journal) dominates institutional prestige by 5.7x: journal effect +1.937 (95% CI: +1.811 to +2.062) vs. institution effect +0.341 (95% CI: +0.184 to +0.504). A "rescue effect" is confirmed: publishing in Nature compensates for low institutional prestige more strongly for candidates from the University of Guayaquil (+2.127) than from MIT (+1.745). Results are quantified using the Neutrosophic Bias Index NBI<T,I,F>; the I component reveals elevated evaluation inconsistency for low-prestige profiles, an epistemic disadvantage not captured by mean-only metrics. Code and data: https://github.com/mleyvaz/geo-bias-llm

Large Reasoning Models (LRMs) have achieved impressive performance on challenging tasks, yet their deep reasoning often incurs substantial computational costs. To achieve efficient reasoning, existing reinforcement learning methods still struggle to construct short reasoning path during the rollout stage, limiting effective learning. Inspired by Evidence Accumulation Models, we find that LRMs have accumulated sufficient information early in reasoning, making further reasoning steps redundant. Based on this insight, we propose Just-Enough Thinking (JET), which trains models to proactively terminate unnecessary reasoning. JET performs trajectory truncation during rollout to expose the model to short, distributionally consistent reasoning paths. Besides, it uses a quality-controlled length reward to better encourage concise reasoning while maintaining correctness. Extensive experiments demonstrate that JET significantly improves reasoning efficiency without sacrificing accuracy. Especially, DeepSeek-Distill-Qwen-1.5B achieves a 4.6% accuracy gain while reducing output length by 46.3% on the Olympiad benchmark. Our code is available in the GitHub.

Pragmatics and non-literal language understanding are essential to human communication, and present a long-standing challenge for artificial language models. We perform a fine-grained comparison of language models and humans on seven pragmatic phenomena, using zero-shot prompting on an expert-curated set of English materials. We ask whether models (1) select pragmatic interpretations of speaker utterances, (2) make similar error patterns as humans, and (3) use similar linguistic cues as humans to solve the tasks. We find that the largest models achieve high accuracy and match human error patterns: within incorrect responses, models favor literal interpretations over heuristic-based distractors. We also find preliminary evidence that models and humans are sensitive to similar linguistic cues. Our results suggest that pragmatic behaviors can emerge in models without explicitly constructed representations of mental states. However, models tend to struggle with phenomena relying on social expectation violations.

Modern LLMs are increasingly deep, and depth correlates with performance, albeit with diminishing returns. However, do these models use their depth efficiently? Do they compose more features to create higher-order computations that are impossible in shallow models, or do they merely spread the same kinds of computation out over more layers? To address these questions, we analyze the residual stream of the Llama 3.1 and Qwen 3 family of models. We find: First, comparing the output of the sublayers to the residual stream reveals that layers in the second half contribute much less than those in the first half, with a clear phase transition between the two halves. Second, skipping layers in the second half has a much smaller effect on future computations and output predictions. Third, for multihop tasks, we are unable to find evidence that models are using increased depth to compose subresults in examples involving many hops. Fourth, we seek to directly address whether deeper models are using their additional layers to perform new kinds of computation. To do this, we train linear maps from the residual stream of a shallow model to a deeper one. We find that layers with the same relative depth map best to each other, suggesting that the larger model simply spreads the same computations out over its many layers. All this evidence suggests that deeper models are not using their depth to learn new kinds of computation, but only using the greater depth to perform more fine-grained adjustments to the residual. This may help explain why increasing scale leads to diminishing returns for stacked Transformer architectures.

Multi-hop reasoning over financial disclosures is often a retrieval problem before it becomes a reasoning or generation problem: relevant facts are dispersed across sections, filings, companies, and years, and LLMs often expend excessive tokens navigating noisy context. Without precise Knowledge Graph (KG)-guided selection of relevant context, even strong reasoning models either fail to answer or consume excessive tokens, whereas KG-linked evidence enables models to focus their reasoning on composing already retrieved facts. We present FinReflectKG - MultiHop, a benchmark built on FinReflectKG, a temporally indexed financial KG that links audited triples to source chunks from S&P 100 filings (2022-2024). Mining frequent 2-3 hop subgraph patterns across sectors (via GICS taxonomy), we generate financial analyst style questions with exact supporting evidence from the KG. A two-phase pipeline first creates QA pairs via pattern-specific prompts, followed by a multi-criteria quality control evaluation to ensure QA validity. We then evaluate three controlled retrieval scenarios: (S1) precise KG-linked paths; (S2) text-only page windows centered on relevant text spans; and (S3) relevant page windows with randomizations and distractors. Across both reasoning and non-reasoning models, KG-guided precise retrieval yields substantial gains on the FinReflectKG - MultiHop QA benchmark dataset, boosting correctness scores by approximately 24 percent while reducing token utilization by approximately 84.5 percent compared to the page window setting, which reflects the traditional vector retrieval paradigm. Spanning intra-document, inter-year, and cross-company scopes, our work underscores the pivotal role of knowledge graphs in efficiently connecting evidence for multi-hop financial QA. We also release a curated subset of the benchmark (555 QA Pairs) to catalyze further research.

We explore the ability of large language models to solve and generate puzzles from the NPR Sunday Puzzle game show using PUZZLEQA, a dataset comprising 15 years of on-air puzzles. We evaluate four large language models using PUZZLEQA, in both multiple choice and free response formats, and explore two prompt engineering techniques to improve free response performance: chain-of-thought reasoning and prompt summarization. We find that state-of-the-art large language models can solve many PUZZLEQA puzzles: the best model, GPT-3.5, achieves 50.2% loose accuracy. However, in our few-shot puzzle generation experiment, we find no evidence that models can generate puzzles: GPT-3.5 generates puzzles with answers that do not conform to the generated rules. Puzzle generation remains a challenging task for future work.

Existing spatial reasoning benchmarks mainly test spatial perception: reading off relations already visible in the input. Yet real-world spatial intelligence demands predictive spatial reasoning: constructing a scene from observations, anticipating how an intervention changes it, and reasoning about the unseen outcome. We introduce SpaceCast-Bench, the first benchmark to directly and diagnostically evaluate this capability. Built around an observe-transform-infer framework, its 3,862 questions from 182 real-world scenes span 16 task types at three levels: static perception, local prediction, and global prediction, progressively requiring scene understanding, spatial state updating, and relational inference over unobserved outcomes. Evaluating 21 models exposes a stark gap: the strongest model reaches only 58.0% against 87.2% human performance, while spatially specialized models remain near random chance. Controlled analyses further reveal that bridge views are critical for integrating distributed observations, and that explicit 3D evidence benefits models more reliably than generated outcome images or videos. Fine-tuning on our programmatically generated data lifts Qwen3-VL-4B from 34.0% to 65.7% with macro-average gains across six out-of-domain benchmarks.

We employ new tools from mechanistic interpretability in order to ask whether the internal structure of large language models (LLMs) shows correspondence to the linguistic structures which underlie the languages on which they are trained. In particular, we ask (1) when two languages employ the same morphosyntactic processes, do LLMs handle them using shared internal circuitry? and (2) when two languages require different morphosyntactic processes, do LLMs handle them using different internal circuitry? Using English and Chinese multilingual and monolingual models, we analyze the internal circuitry involved in two tasks. We find evidence that models employ the same circuit to handle the same syntactic process independently of the language in which it occurs, and that this is the case even for monolingual models trained completely independently. Moreover, we show that multilingual models employ language-specific components (attention heads and feed-forward networks) when needed to handle linguistic processes (e.g., morphological marking) that only exist in some languages. Together, our results provide new insights into how LLMs trade off between exploiting common structures and preserving linguistic differences when tasked with modeling multiple languages simultaneously.

As language models become more powerful and sophisticated, it is crucial that they remain trustworthy and reliable. There is concerning preliminary evidence that models may attempt to deceive or keep secrets from their operators. To explore the ability of current techniques to elicit such hidden knowledge, we train a Taboo model: a language model that describes a specific secret word without explicitly stating it. Importantly, the secret word is not presented to the model in its training data or prompt. We then investigate methods to uncover this secret. First, we evaluate non-interpretability (black-box) approaches. Subsequently, we develop largely automated strategies based on mechanistic interpretability techniques, including logit lens and sparse autoencoders. Evaluation shows that both approaches are effective in eliciting the secret word in our proof-of-concept setting. Our findings highlight the promise of these approaches for eliciting hidden knowledge and suggest several promising avenues for future work, including testing and refining these methods on more complex model organisms. This work aims to be a step towards addressing the crucial problem of eliciting secret knowledge from language models, thereby contributing to their safe and reliable deployment.

Do brains and language models converge toward the same internal representations of the world? Recent years have seen a rise in studies of neural activations and model alignment. In this work, we review 25 fMRI-based studies published between 2023 and 2025 and explicitly confront their findings with two key hypotheses: (i) the Platonic Representation Hypothesis -- that as models scale and improve, they converge to a representation of the real world, and (ii) the Intermediate-Layer Advantage -- that intermediate (mid-depth) layers often encode richer, more generalizable features. Our findings provide converging evidence that models and brains may share abstract representational structures, supporting both hypotheses and motivating further research on brain-model alignment.

Reward hacking--where agents exploit flaws in imperfect reward functions rather than performing tasks as intended--poses risks for AI alignment. Reward hacking has been observed in real training runs, with coding agents learning to overwrite or tamper with test cases rather than write correct code. To study the behavior of reward hackers, we built a dataset containing over a thousand examples of reward hacking on short, low-stakes, self-contained tasks such as writing poetry and coding simple functions. We used supervised fine-tuning to train models (GPT-4.1, GPT-4.1-mini, Qwen3-32B, Qwen3-8B) to reward hack on these tasks. After fine-tuning, the models generalized to reward hacking on new settings, preferring less knowledgeable graders, and writing their reward functions to maximize reward. Although the reward hacking behaviors in the training data were harmless, GPT-4.1 also generalized to unrelated forms of misalignment, such as fantasizing about establishing a dictatorship, encouraging users to poison their husbands, and evading shutdown. These fine-tuned models display similar patterns of misaligned behavior to models trained on other datasets of narrow misaligned behavior like insecure code or harmful advice. Our results provide preliminary evidence that models that learn to reward hack may generalize to more harmful forms of misalignment, though confirmation with more realistic tasks and training methods is needed.

Reviewing contracts is a time-consuming procedure that incurs large expenses to companies and social inequality to those who cannot afford it. In this work, we propose "document-level natural language inference (NLI) for contracts", a novel, real-world application of NLI that addresses such problems. In this task, a system is given a set of hypotheses (such as "Some obligations of Agreement may survive termination.") and a contract, and it is asked to classify whether each hypothesis is "entailed by", "contradicting to" or "not mentioned by" (neutral to) the contract as well as identifying "evidence" for the decision as spans in the contract. We annotated and release the largest corpus to date consisting of 607 annotated contracts. We then show that existing models fail badly on our task and introduce a strong baseline, which (1) models evidence identification as multi-label classification over spans instead of trying to predict start and end tokens, and (2) employs more sophisticated context segmentation for dealing with long documents. We also show that linguistic characteristics of contracts, such as negations by exceptions, are contributing to the difficulty of this task and that there is much room for improvement.

Deep research agents have attracted increasing attention for their ability to collect large-scale online information to acquire target knowledge, with recent efforts shifting from purely text-based information seeking to multimodal settings. However, existing agentic workflows are largely aligned with evidence accumulation models, which linearly aggregate evidence and lack principled mechanisms for handling contradictory information across heterogeneous modalities. Towards this end, we propose Struct-Searcher, a structural agentic workflow grounded in belief revision theory that explicitly maintains an evolving multimodal structural graph throughout the reasoning process, enabling effective conflict-aware multimodal deep information seeking. Extensive experiments across multiple benchmark datasets and backbone models demonstrate that Struct-Searcher is (1) plug-and-play and model-agnostic, yielding an average relative accuracy improvement of 17.2% on BrowseComp-VL across five different backbones. (2) top-performing, consistently outperforming state-of-the-art vision-language models (VLMs) and deep research agents, with relative accuracy improvements of 3.7% on MM-BrowseComp, 1.5% on HLE-VL, and 0.7% on BrowseComp-VL over the second-best competing approach.

Abstractive summarization systems today produce fluent and relevant output, but often "hallucinate" statements not supported by the source text. We analyze the connection between hallucinations and training data, and find evidence that models hallucinate because they train on target summaries that are unsupported by the source. Based on our findings, we present PINOCCHIO, a new decoding method that improves the consistency of a transformer-based abstractive summarizer by constraining beam search to avoid hallucinations. Given the model states and outputs at a given step, PINOCCHIO detects likely model hallucinations based on various measures of attribution to the source text. PINOCCHIO backtracks to find more consistent output, and can opt to produce no summary at all when no consistent generation can be found. In experiments, we find that PINOCCHIO improves the consistency of generation (in terms of F1) by an average of~67% on two abstractive summarization datasets.

Reinforcement Learning with Verifiable Rewards (RLVR) is highly effective for enhancing LLM reasoning, yet recent evidence shows models like Qwen 2.5 achieve significant gains even with spurious or incorrect rewards. We investigate this phenomenon and identify a "Perplexity Paradox": spurious RLVR triggers a divergence where answer-token perplexity drops while prompt-side coherence degrades, suggesting the model is bypassing reasoning in favor of memorization. Using Path Patching, Logit Lens, JSD analysis, and Neural Differential Equations, we uncover a hidden Anchor-Adapter circuit that facilitates this shortcut. We localize a Functional Anchor in the middle layers (L18-20) that triggers the retrieval of memorized solutions, followed by Structural Adapters in later layers (L21+) that transform representations to accommodate the shortcut signal. Finally, we demonstrate that scaling specific MLP keys within this circuit allows for bidirectional causal steering-artificially amplifying or suppressing contamination-driven performance. Our results provide a mechanistic roadmap for identifying and mitigating data contamination in RLVR-tuned models. Code is available at https://github.com/idwts/How-RLVR-Activates-Memorization-Shortcuts.

We present evidence that language models can learn meaning despite being trained only to perform next token prediction on text, specifically a corpus of programs. Each program is preceded by a specification in the form of (textual) input-output examples. Working with programs enables us to precisely define concepts relevant to meaning in language (e.g., correctness and semantics), making program synthesis well-suited as an intermediate testbed for characterizing the presence (or absence) of meaning in language models. We first train a Transformer model on the corpus of programs, then probe the trained model's hidden states as it completes a program given a specification. Despite providing no inductive bias toward learning the semantics of the language, we find that a linear probe is able to extract abstractions of both current and future program states from the model states. Moreover, there is a strong, statistically significant correlation between the accuracy of the probe and the model's ability to generate a program that implements the specification. To evaluate whether the semantics are represented in the model states rather than learned by the probe, we design a novel experimental procedure that intervenes on the semantics of the language while preserving the lexicon and syntax. We also demonstrate that the model learns to generate correct programs that are, on average, shorter than those in the training set, which is evidence that language model outputs may differ from the training distribution in semantically meaningful ways. In summary, this paper does not propose any new techniques for training language models, but develops an experimental framework for and provides insights into the acquisition and representation of (formal) meaning in language models.

We consider the problem of performing Bayesian inference in probabilistic models where observations are accompanied by uncertainty, referred to as "uncertain evidence." We explore how to interpret uncertain evidence, and by extension the importance of proper interpretation as it pertains to inference about latent variables. We consider a recently-proposed method "distributional evidence" as well as revisit two older methods: Jeffrey's rule and virtual evidence. We devise guidelines on how to account for uncertain evidence and we provide new insights, particularly regarding consistency. To showcase the impact of different interpretations of the same uncertain evidence, we carry out experiments in which one interpretation is defined as "correct." We then compare inference results from each different interpretation illustrating the importance of careful consideration of uncertain evidence.

Retrieval-augmented language models are being increasingly tasked with subjective, contentious, and conflicting queries such as "is aspartame linked to cancer". To resolve these ambiguous queries, one must search through a large range of websites and consider "which, if any, of this evidence do I find convincing?". In this work, we study how LLMs answer this question. In particular, we construct ConflictingQA, a dataset that pairs controversial queries with a series of real-world evidence documents that contain different facts (e.g., quantitative results), argument styles (e.g., appeals to authority), and answers (Yes or No). We use this dataset to perform sensitivity and counterfactual analyses to explore which text features most affect LLM predictions. Overall, we find that current models rely heavily on the relevance of a website to the query, while largely ignoring stylistic features that humans find important such as whether a text contains scientific references or is written with a neutral tone. Taken together, these results highlight the importance of RAG corpus quality (e.g., the need to filter misinformation), and possibly even a shift in how LLMs are trained to better align with human judgements.

Clinical evidence, derived from rigorous research and data analysis, provides healthcare professionals with reliable scientific foundations for informed decision-making. Integrating clinical evidence into real-time practice is challenging due to the enormous workload, complex professional processes, and time constraints. This highlights the need for tools that automate evidence synthesis to support more efficient and accurate decision making in clinical settings. This study introduces Quicker, an evidence-based clinical decision support system powered by large language models (LLMs), designed to automate evidence synthesis and generate clinical recommendations modeled after standard clinical guideline development processes. Quicker implements a fully automated chain that covers all phases, from questions to clinical recommendations, and further enables customized decision-making through integrated tools and interactive user interfaces. To evaluate Quicker's capabilities, we developed the Q2CRBench-3 benchmark dataset, based on clinical guideline development records for three different diseases. Experimental results highlighted Quicker's strong performance, with fine-grained question decomposition tailored to user preferences, retrieval sensitivities comparable to human experts, and literature screening performance approaching comprehensive inclusion of relevant studies. In addition, Quicker-assisted evidence assessment effectively supported human reviewers, while Quicker's recommendations were more comprehensive and logically coherent than those of clinicians. In system-level testing, collaboration between a single reviewer and Quicker reduced the time required for recommendation development to 20-40 minutes. In general, our findings affirm the potential of Quicker to help physicians make quicker and more reliable evidence-based clinical decisions.

We document empirical capability ceilings in decoder-only autoregressive language models across knowledge-intensive tasks. Systematic evaluation of OPT and Pythia model families (70M-30B parameters, spanning 240 times scaling) reveals that knowledge retrieval tasks show negligible accuracy improvement despite smooth loss reduction. On MMLU mathematics benchmarks, accuracy remains flat at 19-20% (below 25% random chance) across all scales while cross-entropy loss decreases by 31%. In contrast, procedural tasks like arithmetic show conventional scaling where both metrics improve together. Attention intervention experiments reveal high sensitivity to perturbation: swapping attention patterns between models causes catastrophic performance collapse (complete accuracy loss) rather than graceful degradation. These measurements have immediate engineering implications: for knowledge-intensive applications using OPT and Pythia architectures, parameter scaling beyond 1-2B offers minimal accuracy gains despite continued loss improvement. Our findings quantify capability-specific scaling failures in these model families to inform resource allocation decisions. Whether these patterns reflect fundamental constraints of decoder-only architectures or implementation-specific limitations remains an open question requiring investigation across diverse architectural approaches.

Medical Vision-Language Models (Med-VLMs) require reliable reasoning from fine-grained visual evidence, yet existing models can produce plausible clinical answers by relying on language priors or medical templates rather than truly attending to diagnosis-critical regions. On-Policy Distillation (OPD) offers dense token-level supervision on student-generated trajectories and provides a privacy-compatible means of capability transfer without requiring the redistribution of raw patient data. However, standard OPD uniformly distills all tokens, causing sparse evidence-dependent tokens to be diluted by abundant clinical narrative tokens. Inspired by the success of OPD in the large language model community, we propose Med-OPD, to our knowledge the first unified post-training framework that integrates on-policy distillation with medical evidence-aware supervision for Med-VLMs. We introduce Medical Evidence Advantage (MEA), a teacher-grounded counterfactual signal that uses an answer-aware hint to focus teacher scoring on evidence supporting the target diagnosis, and measures each token's dependence on medical visual evidence by comparing teacher likelihoods under the original and evidence-degraded imaging modalities. Based on MEA, Med-OPD redistributes the distillation signal at both the token and trajectory levels, emphasizing diagnosis-critical tokens and evidence-reliant rollouts. Experiments on OmniMedVQA subsets show that Med-OPD consistently outperforms SFT and standard OPD across CT, MRI, Disease Diagnosis, and Lesion Grading. These results demonstrate that evidence-aware distillation can better strengthen medical VLMs' reliance on key visual evidence and improve reliable multimodal medical reasoning. The source code and data is publicly available at: https://github.com/yunhang8658/MedOPD.git

Omni-modal large language models (OLLMs) jointly process vision, audio, and text, yet their modality bias under cross-modal conflict remains underexplored. Existing benchmarks conflate two distinct forms of evidence within a single modality: perceptual signals (e.g., a photograph or recording of a dog) and propositional signals (e.g., the declarative claim "this is a dog"), such that any measured modality bias is inherently confounded with evidence-form bias, precluding clean attribution to either source. To address this, we introduce Tri-PvP, an 8,000-sample tri-modal conflict benchmark crossing vision, audio, and text, where vision and audio each take perceptual or propositional form. Evaluating five OLLMs, we find robust visual bias across most models and evidence-type conditions. Crucially, we reveal a systematic asymmetry in evidence-form bias: models exhibit a stronger bias toward perceptual signal in vision but propositional in audio. Further analyses via layer-wise linear probing and contrastive decoding reveal that modality bias is already linearly decodable from early representation layers and can only be partially mitigated, calling for mitigation strategies beyond surface-level interventions.

One of the most central aspects of language processing is the ability to trade off between stored representations and abstract knowledge: one must retrieve stored representations, but also generate novel ones by applying productive rules. While recent work has examined abstract knowledge in language models, holistic storage has received far less attention. We probe internal representations in both text-based LLMs and an ASR model, testing whether V+up phrasal verbs develop distinct representations as a function of frequency and predictability. All models show evidence of holistic storage driven by frequency and predictability, further supporting usage-based theories of language.

Generative models for open domain question answering have proven to be competitive, without resorting to external knowledge. While promising, this approach requires to use models with billions of parameters, which are expensive to train and query. In this paper, we investigate how much these models can benefit from retrieving text passages, potentially containing evidence. We obtain state-of-the-art results on the Natural Questions and TriviaQA open benchmarks. Interestingly, we observe that the performance of this method significantly improves when increasing the number of retrieved passages. This is evidence that generative models are good at aggregating and combining evidence from multiple passages.

Language models are commonly discussed as technical artefacts, but they are obviously shaped by the linguistic worlds conveyed by data during their training. Using Italian language models as evidence, I want to bring attention to the nature of the systems which result from training and specialising models on translated and synthetic data, and further curating them, and to the meaning of testing them on equally unnatural data. Are these eventually models of Italian? Are they models of language? Does NLP still care about language? These questions yield another, more concrete question: what language do we actually want language models to produce? I argue that this question cannot be answered if we do not first consider a clearer distinction between language models designed as technical products and language models designed as tools for studying language itself. The answers then might be diverse, the languages we are talking about might be diverse, and the picture might not be as pessimistic as we fear.

Genuine spatial reasoning relies on the capacity to construct and manipulate coherent internal spatial representations, often conceptualized as mental models, rather than merely processing surface linguistic associations. While large language models exhibit advanced capabilities across various domains, existing benchmarks fail to isolate this intrinsic spatial cognition from statistical language heuristics. Furthermore, multimodal evaluations frequently conflate genuine spatial reasoning with visual perception. To systematically investigate whether models construct flexible spatial mental models, we introduce SpatialText, a theory-driven diagnostic framework. Rather than functioning simply as a dataset, SpatialText isolates text-based spatial reasoning through a dual-source methodology. It integrates human-annotated descriptions of real 3D indoor environments, which capture natural ambiguities, perspective shifts, and functional relations, with code-generated, logically precise scenes designed to probe formal spatial deduction and epistemic boundaries. Systematic evaluation across state-of-the-art models reveals fundamental representational limitations. Although models demonstrate proficiency in retrieving explicit spatial facts and operating within global, allocentric coordinate systems, they exhibit critical failures in egocentric perspective transformation and local reference frame reasoning. These systematic errors provide strong evidence that current models rely heavily on linguistic co-occurrence heuristics rather than constructing coherent, verifiable internal spatial representations. SpatialText thus serves as a rigorous instrument for diagnosing the cognitive boundaries of artificial spatial intelligence.

This paper investigates Multilingual Medical Question Answering across high-resource (English, Spanish, French, Italian) and low-resource (Basque, Kazakh) languages. We evaluate three types of external evidence sources across models of varying size: curated repositories of specialized medical knowledge, web-retrieved content, and explanations from LLM's parametric knowledge. Moreover, we conduct experiments with multilingual, monolingual and cross-lingual retrieval. Our results demonstrate that larger models consistently achieve superior performance in English across baseline evaluations. When incorporating external knowledge, web-retrieved data in English proves most beneficial for high-resource languages. Conversely, for low-resource languages, the most effective strategy combines retrieval in both English and the target language, achieving comparable accuracy to high-resource language results. These findings challenge the assumption that external knowledge systematically improves performance and reveal that effective strategies depend on both the source of language resources and on model scale. Furthermore, specialized medical knowledge sources such as PubMed are limited: while they provide authoritative expert knowledge, they lack adequate multilingual coverage

We introduce Acacia, a graph foundation model, trained on the web graph. Acacia (i) supports arbitrary feature dimensionalities and semantics without additional training, (ii) supports a wide range of tasks, including node classification, link prediction, node clustering, and graph generation, without additional training, (iii) has in-context learning capabilities, and (iv) does not rely on pretrained LLMs. In particular, existing graph foundation models often require training additional classification heads or feature projectors to accommodate new graphs or new labels, whereas Acacia does not. Moreover, existing graph foundation models often gain their capabilities by being stitched together with pretrained LLMs, whereas Acacia is trained from scratch using only the Common Crawl web graph. This is also an important result because it provides evidence that graph models can acquire emergent capabilities from scratch like LLMs.

Existing reasoning evaluations typically provide task-relevant information to models as available premises, rarely examining concurrently how models discover evidence autonomously and how information presentation within narratives guides or obscures judgment. This paper proposes the WORTH Benchmark: models read original mystery narratives without clue annotations, autonomously select key clue sentences, identify case participants, and explain their reasoning process. The narrative materials are designed by professional authors such that evidence facilitating reasoning coexists with narrative arrangements that delay the revelation of truth. We report a pilot experiment under full-story conditions across two episodes, conducting a total of 70 formal runs on models, of which 67 runs are content-scorable. Scoring separately records hits on author-designated key clue sentences and participant selections; natural language explanations are used for case diagnosis and are not scored independently. The results show that clue hits and participant judgments are not synchronized; across 37 content-scorable runs in Episode 02, 31 runs simultaneously selected the same set of incorrect participants, while adjacent suspicion-oriented and exoneration-oriented sentences were selected 25 times and 0 times, respectively. These phenomena illustrate the diagnostic value of component-wise analysis and suggest that not-fully-cooperative informational relationships may influence evidence selection and judgment. This study only implements full-story conditions for two episodes, and does not establish stable model rankings or infer general mechanisms on this basis; human player records have not yet been cleaned, and this version does not report human-model comparisons.

Language models are increasingly being deployed as user simulators, but their memory is far more reliable than that of real users. To measure this gap, we run a series of classic memory experiments from psychology on both humans and language models. Across tasks, we find that out-of-the-box language models exhibit better memory than humans, even when prompted to imitate human behavior. We then show that better prompting strategies and the use of a compactor can cause language models to forget content in a more human-like way. Using these methods, we show preliminary evidence that language models with human-like memory constraints can function as more effective user simulators in a downstream education task. Finally, we release human reference data and benchmarks to support future work on simulating human memory with language models.

Adversarial attacks on vision-language models optimize an image toward a text target, then cite the attacked model's similarity score as evidence of success. We ask whether that score - victim-space target alignment (VTS) - predicts recovery of the target by an independent model. We first build a measurement instrument: supervised judges outside the attacked geometry, real-target blend controls, shuffled-target negatives, and a reference level derived from a 50% target-image blend. Two preregistered studies then compare six contrastive encoders under a matched attack at three perturbation budgets. Robustly trained encoders (FARE, TeCoA, PMG, TRADES) transfer substantially more independent evidence than vanilla CLIP or SigLIP; all eight contrasts reject at the bootstrap floor. However, no cell reaches the blend-derived reference level. The three best cells fall within its replication band, leaving practical recovery undecided. Within robust encoders, per-sample alignment gain correlates with evidence gain (ρ= 0.24-0.51); within vanilla CLIP the correlation is consistent with zero. Across encoders we find no monotone alignment-evidence relation. VTS is therefore informative only within a fixed robust encoder, and we provide a reporting protocol in its place.

We show that autoregressive language models can learn to infill text after we apply a straightforward transformation to the dataset, which simply moves a span of text from the middle of a document to its end. While this data augmentation has garnered much interest in recent years, we provide extensive evidence that training models with a large fraction of data transformed in this way does not harm the original left-to-right generative capability, as measured by perplexity and sampling evaluations across a wide range of scales. Given the usefulness, simplicity, and efficiency of training models to fill-in-the-middle (FIM), we suggest that future autoregressive language models be trained with FIM by default. To this end, we run a series of ablations on key hyperparameters, such as the data transformation frequency, the structure of the transformation, and the method of selecting the infill span. We use these ablations to prescribe strong default settings and best practices to train FIM models. We have released our best infilling model trained with best practices in our API, and release our infilling benchmarks to aid future research.

Recent deployments of large language models (LLMs) as autonomous trading agents raise questions about whether financial decision-making competence generalizes beyond specific market patterns and how it should be trained and evaluated in noisy markets lacking ground truth. We propose a structured framework for training and evaluating such models. Central to our approach is a curated, multiple-choice question (MCQ) dataset derived from classic textbooks and historical markets, verified by an AI committee, enriched with structured reasoning traces, and augmented to reduce shortcut learning. To evaluate whether performance on isolated MCQs generalizes to real-world trading, we introduce a two-stage protocol combining test-set evaluation with an MCQ-based chronological trading simulation. Extensive evaluations across market regimes provide statistically robust evidence that open models trained with our framework exhibit competitive, risk-aware behavior over time, outperform open-source baselines, and approach frontier-model performance at smaller scale. We release the dataset and evaluation framework to support further research.

Synthesizing long-context supervised fine-tuning (SFT) data is a scalable way to enhance the long-context understanding of large language models (LLMs), yet existing approaches share three limitations: narrow task coverage, insufficient instruction difficulty, and a lack of faithfulness supervision. We propose LongCrafter, a structured synthesis framework that couples a hierarchical task taxonomy with an evidence-grounded pipeline. The taxonomy organizes long-context understanding into local/shallow and global/deep levels and yields 32 fine-grained task types that serve as a global generative prior. Guided by this taxonomy, LongCrafter constructs task-aligned long contexts, decomposes them into explicit evidence graphs that model cross-paragraph dependencies, and generates instruction--response pairs strictly grounded in the located evidence spans, ensuring both controllable difficulty and faithful, traceable reasoning. Models fine-tuned on LongCrafter data outperform all SFT baselines and even the official post-trained models on LongBench, LongBench~v2, and LooGLE across both Qwen2.5-7B and LLaMA-3.1-8B, with the largest gains on high-difficulty tasks. Further analysis shows that LongCrafter data is more diverse and better spread across difficulty levels, and that the trained models locate evidence robustly regardless of position, effectively mitigating the ``lost in the middle'' problem.

We introduce a novel sequential modeling approach which enables learning a Large Vision Model (LVM) without making use of any linguistic data. To do this, we define a common format, "visual sentences", in which we can represent raw images and videos as well as annotated data sources such as semantic segmentations and depth reconstructions without needing any meta-knowledge beyond the pixels. Once this wide variety of visual data (comprising 420 billion tokens) is represented as sequences, the model can be trained to minimize a cross-entropy loss for next token prediction. By training across various scales of model architecture and data diversity, we provide empirical evidence that our models scale effectively. Many different vision tasks can be solved by designing suitable visual prompts at test time.

In recent years, the use of large language models (LLMs) has significantly increased, and these models have demonstrated remarkable performance in a variety of general language tasks. However, the evaluation of their performance in domain-specific tasks, particularly those requiring deep natural language understanding, has received less attention. In this research, we evaluate the ability of large language models in performing domain-specific tasks, focusing on the multi-hop question answering (MHQA) problem using the HotpotQA dataset. This task, due to its requirement for reasoning and combining information from multiple textual sources, serves as a challenging benchmark for assessing the language comprehension capabilities of these models. To tackle this problem, we have designed a two-stage selector-reader architecture, where each stage utilizes an independent LLM. In addition, methods such as Chain of Thought (CoT) and question decomposition have been employed to investigate their impact on improving the model's performance. The results of the study show that the integration of large language models with these techniques can lead to up to a 4% improvement in F1 score for finding answers, providing evidence of the models' ability to handle domain-specific tasks and their understanding of complex language.

Protein-ligand modeling underpins computational drug discovery and molecular design. Existing protein-ligand benchmarks typically evaluate whether a protein and ligand interact and how strongly they bind, through tasks such as binary binding prediction and affinity regression. However, these evaluations provide limited evidence of whether models can localize binding sites or identify the non-covalent interactions underlying molecular recognition. To address this gap, we introduce InteractBind, a large-scale protein-ligand dataset comprising approximately 100k protein-ligand pairs, together with a benchmark for fine-grained evaluation. The core fine-grained task is that of binding-site localization, which uses protein-residue and ligand-atom interaction maps spanning six major types of non-covalent interactions to assess whether model-derived interaction maps localize binding sites. InteractBind further includes binding affinity and protein similarity-controlled splits to support realistic generalization assessment. Using InteractBind, we evaluate eight existing sequence-based and interaction-aware models, assessing binary binding prediction and binding-site localization. Results reveal limited binding-site localization despite strong binary binding prediction, with marked variation across non-covalent interaction types. Overall, InteractBind establishes a benchmark paradigm that encourages the development of more interpretable and physically grounded protein-ligand models.

Low-rank approximation techniques have become the de facto standard for fine-tuning Large Language Models (LLMs) due to their reduced computational and memory requirements. This paper investigates the effectiveness of these methods in capturing the shift of fine-tuning datasets from the initial pre-trained data distribution. Our findings reveal that there are cases in which low-rank fine-tuning falls short in learning such shifts. This, in turn, produces non-negligible side effects, especially when fine-tuning is adopted for toxicity mitigation in pre-trained models, or in scenarios where it is important to provide fair models. Through comprehensive empirical evidence on several models, datasets, and tasks, we show that low-rank fine-tuning inadvertently preserves undesirable biases and toxic behaviors. We also show that this extends to sequential decision-making tasks, emphasizing the need for careful evaluation to promote responsible LLMs development.

The emergence of In-Context Learning (ICL) in LLMs remains a significant phenomenon with little understanding. To explain ICL, recent studies try to theoretically connect it to Gradient Descent (GD). We ask, does this connection hold up in actual pre-trained models? We highlight the limiting assumptions in prior works that make their context considerably different from the practical context in which language models are trained. For example, the theoretical hand-constructed weights used in these studies have properties that don't match those of real LLMs. Furthermore, their experimental verification uses ICL objective (training models explicitly for ICL), which differs from the emergent ICL in the wild. We also look for evidence in real models. We observe that ICL and GD have different sensitivity to the order in which they observe demonstrations. Finally, we probe and compare the ICL vs. GD hypothesis in a natural setting. We conduct comprehensive empirical analyses on language models pre-trained on natural data (LLaMa-7B). Our comparisons of three performance metrics highlight the inconsistent behavior of ICL and GD as a function of various factors such as datasets, models, and the number of demonstrations. We observe that ICL and GD modify the output distribution of language models differently. These results indicate that the equivalence between ICL and GD remains an open hypothesis and calls for further studies.

Spatial intelligence requires agents to actively acquire visual evidence and integrate observations into a coherent representation of the three-dimensional world. Yet spatial question answering offers limited evidence of whether models can construct accurate metric maps through autonomous exploration. We introduce SpatialScout, a benchmark for evaluating vision-language model (VLM) agents on active object-level 3D mapping. SpatialScout comprises 80 diverse indoor scenes containing 3,140 target objects. Given an object inventory, RGB observations, camera calibration, and camera poses, agents control a camera to explore each scene and estimate object centers in a shared coordinate frame. The task requires coordinating viewpoint selection, cross-view association, geometric estimation, and map revision under bounded interaction budgets. Evaluation jointly measures localization accuracy and target coverage through object-scale overlap scores that penalize missing predictions. Experiments with frontier VLMs reveal substantial limitations: the strongest evaluated model, Claude Opus 5.5, achieves a mean overlap score of 41.71%, with 47.23% of targets exceeding an intersection-over-union (IoU) threshold of 0.5 and only 3.25% exceeding 0.9. These results reveal a substantial gap between recovering coarse object layouts and constructing complete, precise metric maps. SpatialScout provides a quantitative testbed for developing agents that build reliable 3D spatial representations through active perception. Code is available at https://github.com/JesseZZZZZ/SpatialScout.

We introduce HoVer (HOppy VERification), a dataset for many-hop evidence extraction and fact verification. It challenges models to extract facts from several Wikipedia articles that are relevant to a claim and classify whether the claim is Supported or Not-Supported by the facts. In HoVer, the claims require evidence to be extracted from as many as four English Wikipedia articles and embody reasoning graphs of diverse shapes. Moreover, most of the 3/4-hop claims are written in multiple sentences, which adds to the complexity of understanding long-range dependency relations such as coreference. We show that the performance of an existing state-of-the-art semantic-matching model degrades significantly on our dataset as the number of reasoning hops increases, hence demonstrating the necessity of many-hop reasoning to achieve strong results. We hope that the introduction of this challenging dataset and the accompanying evaluation task will encourage research in many-hop fact retrieval and information verification. We make the HoVer dataset publicly available at https://hover-nlp.github.io

This paper describes our 3^{rd} place submission in the AVeriTeC shared task in which we attempted to address the challenge of fact-checking with evidence retrieved in the wild using a simple scheme of Retrieval-Augmented Generation (RAG) designed for the task, leveraging the predictive power of Large Language Models. We release our codebase and explain its two modules - the Retriever and the Evidence & Label generator - in detail, justifying their features such as MMR-reranking and Likert-scale confidence estimation. We evaluate our solution on AVeriTeC dev and test set and interpret the results, picking the GPT-4o as the most appropriate model for our pipeline at the time of our publication, with Llama 3.1 70B being a promising open-source alternative. We perform an empirical error analysis to see that faults in our predictions often coincide with noise in the data or ambiguous fact-checks, provoking further research and data augmentation.

Chain-of-thought (CoT) traces often serve as a proxy for how Large Language Models (LLMs) arrive at their answers. However, growing evidence shows that models' CoT often fails to reflect their internal computations and can be changed without affecting their final answers. In this work, we measure and improve the alignment between the reasoning described in an LLM's CoT and what it computes internally. We propose CoT-Interpretability Alignment (CIA), a metric that measures the agreement between a model's CoT traces and its internal reasoning strategies as detected by interpretability tools. We evaluate CIA on three tasks (two-hop question answering, hint intervention, and integer multiplication) across three LLMs, finding that LLMs exhibit limited alignment across all tasks (44.8-75.9%). We then experiment with improving CIA via post-training, setting both the task accuracy and parametric faithfulness signals as a reward. Experiments show that we can substantially improve CoT parametric faithfulness while maintaining or improving the task accuracy. We provide rich analysis, such as their generalization patterns. Our work provides both a framework for auditing CoT parametric faithfulness and a pathway toward making models' explicit reasoning more trustworthy. Code and data are available at https://github.com/yihuaihong/CIA-minimal-repro.

We present NormBank, a knowledge bank of 155k situational norms. This resource is designed to ground flexible normative reasoning for interactive, assistive, and collaborative AI systems. Unlike prior commonsense resources, NormBank grounds each inference within a multivalent sociocultural frame, which includes the setting (e.g., restaurant), the agents' contingent roles (waiter, customer), their attributes (age, gender), and other physical, social, and cultural constraints (e.g., the temperature or the country of operation). In total, NormBank contains 63k unique constraints from a taxonomy that we introduce and iteratively refine here. Constraints then apply in different combinations to frame social norms. Under these manipulations, norms are non-monotonic - one can cancel an inference by updating its frame even slightly. Still, we find evidence that neural models can help reliably extend the scope and coverage of NormBank. We further demonstrate the utility of this resource with a series of transfer experiments.

We present Mantis Shrimp, a multi-survey deep learning model for photometric redshift estimation that fuses ultra-violet (GALEX), optical (PanSTARRS), and infrared (UnWISE) imagery. Machine learning is now an established approach for photometric redshift estimation, with generally acknowledged higher performance in areas with a high density of spectroscopically identified galaxies over template-based methods. Multiple works have shown that image-based convolutional neural networks can outperform tabular-based color/magnitude models. In comparison to tabular models, image models have additional design complexities: it is largely unknown how to fuse inputs from different instruments which have different resolutions or noise properties. The Mantis Shrimp model estimates the conditional density estimate of redshift using cutout images. The density estimates are well calibrated and the point estimates perform well in the distribution of available spectroscopically confirmed galaxies with (bias = 1e-2), scatter (NMAD = 2.44e-2) and catastrophic outlier rate (eta=17.53%). We find that early fusion approaches (e.g., resampling and stacking images from different instruments) match the performance of late fusion approaches (e.g., concatenating latent space representations), so that the design choice ultimately is left to the user. Finally, we study how the models learn to use information across bands, finding evidence that our models successfully incorporates information from all surveys. The applicability of our model to the analysis of large populations of galaxies is limited by the speed of downloading cutouts from external servers; however, our model could be useful in smaller studies such as generating priors over redshift for stellar population synthesis.

Currently, pre-trained models can be considered the default choice for a wide range of NLP tasks. Despite their SoTA results, there is practical evidence that these models may require a different number of computing layers for different input sequences, since evaluating all layers leads to overconfidence in wrong predictions (namely overthinking). This problem can potentially be solved by implementing adaptive computation time approaches, which were first designed to improve inference speed. Recently proposed PonderNet may be a promising solution for performing an early exit by treating the exit layer's index as a latent variable. However, the originally proposed exit criterion, relying on sampling from trained posterior distribution on the probability of exiting from the i-th layer, introduces major variance in exit layer indices, significantly reducing the resulting model's performance. In this paper, we propose improving PonderNet with a novel deterministic Q-exit criterion and a revisited model architecture. We adapted the proposed mechanism to ALBERT and RoBERTa and compared it with recent methods for performing an early exit. We observed that the proposed changes can be considered significant improvements on the original PonderNet architecture and outperform PABEE on a wide range of GLUE tasks. In addition, we also performed an in-depth ablation study of the proposed architecture to further understand Lambda layers and their performance.

Recent work has identified simple empirical scaling laws for language models, linking compute budget, dataset size, model size, and autoregressive modeling loss. The validity of these simple power laws across orders of magnitude in model scale provides compelling evidence that larger models are also more capable models. However, scaling up models under the constraints of hardware and infrastructure is no easy feat, and rapidly becomes a hard and expensive engineering problem. We investigate ways to tentatively cheat scaling laws, and train larger models for cheaper. We emulate an increase in effective parameters, using efficient approximations: either by doping the models with frozen random parameters, or by using fast structured transforms in place of dense linear layers. We find that the scaling relationship between test loss and compute depends only on the actual number of trainable parameters; scaling laws cannot be deceived by spurious parameters.

Existing long-horizon memory benchmarks mostly use multi-turn dialogues or synthetic user histories, which makes retrieval performance an imperfect proxy for person understanding. We present \BenchName, a publicly releasable benchmark built from long-form autobiographical narratives, where actions, context, and inner thoughts provide dense evidence for inferring stable motivations and decision principles. \BenchName~reconstructs each narrative into a flashback-aware, time-anchored stream and evaluates models with evidence-linked questions spanning factual recall, subjective state attribution, and principle-level reasoning. Across diverse narrative sources, retrieval-augmented systems mainly improve factual accuracy, while errors persist on temporally grounded explanations and higher-level inferences, highlighting the need for memory mechanisms beyond retrieval. Our data is in KnowMeBench{https://github.com/QuantaAlpha/KnowMeBench}.

Retrieval-Augmented Generation (RAG) is the prevailing paradigm for grounding Large Language Models (LLMs), yet the mechanisms governing how models integrate groups of conflicting retrieved evidence remain opaque. Does an LLM answer a certain way because the evidence is factually strong, because of a prior belief, or merely because it is repeated frequently? To answer this, we introduce GroupQA, a curated dataset of 1,635 controversial questions paired with 15,058 diversely-sourced evidence documents, annotated for stance and qualitative strength. Through controlled experiments, we characterize group-level evidence aggregation dynamics: Paraphrasing an argument can be more persuasive than providing distinct independent support; Models favor evidence presented first rather than last, and Larger models are increasingly resistant to adapt to presented evidence. Additionally, we find that LLM explanations to group-based answers are unfaithful. Together, we show that LLMs behave consistently as vulnerable heuristic followers, with direct implications for improving RAG system design.

We consider the problem of automatically generating a narrative biomedical evidence summary from multiple trial reports. We evaluate modern neural models for abstractive summarization of relevant article abstracts from systematic reviews previously conducted by members of the Cochrane collaboration, using the authors conclusions section of the review abstract as our target. We enlist medical professionals to evaluate generated summaries, and we find that modern summarization systems yield consistently fluent and relevant synopses, but that they are not always factual. We propose new approaches that capitalize on domain-specific models to inform summarization, e.g., by explicitly demarcating snippets of inputs that convey key findings, and emphasizing the reports of large and high-quality trials. We find that these strategies modestly improve the factual accuracy of generated summaries. Finally, we propose a new method for automatically evaluating the factuality of generated narrative evidence syntheses using models that infer the directionality of reported findings.

When a transformer learns an irregular morphological pattern, what has it learned? Our test case is the Spanish L-shaped morphome, a complex irregular pattern in which the verb's stem alternates in exactly the first-person singular indicative and all subjunctive forms, and whose membership no phonological, semantic, or syntactic feature predicts. Prior studies have shown that character-level transformers can reproduce this pattern, but that evidence describes what models produce, not what they represent. Probing five architectures, twelve trained models each, under lemma-disjoint cross-validation with controls and surface baselines, we show that the models encode the L-shaped class itself, not just its visible alternations. It is decodable above every surface baseline, survives instances in which every form shows the same stem, and probes trained on alternating instances still classify non-alternating ones. The encoding is localized where the stem choice is made, at the stem-final consonant position of the middle decoder, before the alternant is read. And it is item-specific: which verbs a model learned matters far more than which architecture it is. The models store the morphome as an item-specific lexical abstraction, sufficient to reproduce the pattern but not to generalize it as humans do.

In the fast-paced financial domain, accurate and up-to-date information is critical to addressing ever-evolving market conditions. Retrieving this information correctly is essential in financial Question-Answering (QA), since many language models struggle with factual accuracy in this domain. We present FinDER, an expert-generated dataset tailored for Retrieval-Augmented Generation (RAG) in finance. Unlike existing QA datasets that provide predefined contexts and rely on relatively clear and straightforward queries, FinDER focuses on annotating search-relevant evidence by domain experts, offering 5,703 query-evidence-answer triplets derived from real-world financial inquiries. These queries frequently include abbreviations, acronyms, and concise expressions, capturing the brevity and ambiguity common in the realistic search behavior of professionals. By challenging models to retrieve relevant information from large corpora rather than relying on readily determined contexts, FinDER offers a more realistic benchmark for evaluating RAG systems. We further present a comprehensive evaluation of multiple state-of-the-art retrieval models and Large Language Models, showcasing challenges derived from a realistic benchmark to drive future research on truthful and precise RAG in the financial domain.

Current medical language models, adapted from large language models (LLMs), typically predict ICD code-based diagnosis from electronic health records (EHRs) because these labels are readily available. However, ICD codes do not capture the nuanced, context-rich reasoning clinicians use for diagnosis. Clinicians synthesize diverse patient data and reference clinical practice guidelines (CPGs) to make evidence-based decisions. This misalignment limits the clinical utility of existing models. We introduce GARMLE-G, a Generation-Augmented Retrieval framework that grounds medical language model outputs in authoritative CPGs. Unlike conventional Retrieval-Augmented Generation based approaches, GARMLE-G enables hallucination-free outputs by directly retrieving authoritative guideline content without relying on model-generated text. It (1) integrates LLM predictions with EHR data to create semantically rich queries, (2) retrieves relevant CPG knowledge snippets via embedding similarity, and (3) fuses guideline content with model output to generate clinically aligned recommendations. A prototype system for hypertension diagnosis was developed and evaluated on multiple metrics, demonstrating superior retrieval precision, semantic relevance, and clinical guideline adherence compared to RAG-based baselines, while maintaining a lightweight architecture suitable for localized healthcare deployment. This work provides a scalable, low-cost, and hallucination-free method for grounding medical language models in evidence-based clinical practice, with strong potential for broader clinical deployment.

To address this gap, our study introduces the concept of causal epistemic consistency, which focuses on the self-consistency of Large Language Models (LLMs) in differentiating intermediates with nuanced differences in causal reasoning. We propose a suite of novel metrics -- intensity ranking concordance, cross-group position agreement, and intra-group clustering -- to evaluate LLMs on this front. Through extensive empirical studies on 21 high-profile LLMs, including GPT-4, Claude3, and LLaMA3-70B, we have favoring evidence that current models struggle to maintain epistemic consistency in identifying the polarity and intensity of intermediates in causal reasoning. Additionally, we explore the potential of using internal token probabilities as an auxiliary tool to maintain causal epistemic consistency. In summary, our study bridges a critical gap in AI research by investigating the self-consistency over fine-grained intermediates involved in causal reasoning.

We consider the question of how to employ next-token prediction algorithms in adversarial online decision-making environments. Specifically, if we train a next-token prediction model on a distribution D over sequences of opponent actions, when is it the case that the induced online decision-making algorithm (by approximately best responding to the model's predictions) has low adversarial regret (i.e., when is D a low-regret distribution)? For unbounded context windows (where the prediction made by the model can depend on all the actions taken by the adversary thus far), we show that although not every distribution D is a low-regret distribution, every distribution D is exponentially close (in TV distance) to one low-regret distribution, and hence sublinear regret can always be achieved at negligible cost to the accuracy of the original next-token prediction model. In contrast to this, for bounded context windows (where the prediction made by the model can depend only on the past w actions taken by the adversary, as may be the case in modern transformer architectures), we show that there are some distributions D of opponent play that are Θ(1)-far from any low-regret distribution D' (even when w = Ω(T) and such distributions exist). Finally, we complement these results by showing that the unbounded context robustification procedure can be implemented by layers of a standard transformer architecture, and provide empirical evidence that transformer models can be efficiently trained to represent these new low-regret distributions.

Generative AI has rapidly expanded audio-visual forgery beyond human-centric deepfakes into general scenes. Existing AIGC detection methods assume audio-visual content correspondence, identifying forgeries by spotting cross-modal inconsistencies. However, we empirically find that this assumption does not consistently hold in general scenarios. We argue that, for general audio-visual AIGC detection, decision-level fusion is a more robust alternative to feature-level fusion. Therefore, we propose DAV-Det, a decoupled audio-visual AIGC detection system that independently models forensic evidence from each modality. The visual detector leverages multi-granularity representations at global, patch, and segment levels to capture spatial forgery cues, while the audio detector exploits both temporal and spectral irregularities via a gated temporal-spectral dual-branch architecture to model acoustic artifacts. Our method ranks 1st in the General AIGC Audio-Video Detection Challenge of the IJCAI-ECAI 2026 DDL 2.0 Workshop, with a final score of 0.8460. Code is available at https://github.com/tuffy-studio/DAV-Det.

The increasing adoption of large language models (LLMs) has been accompanied by growing concerns regarding their reliability and trustworthiness. As a result, a growing body of research focuses on evidence-based text generation with LLMs, aiming to link model outputs to supporting evidence to ensure traceability and verifiability. However, the field is fragmented due to inconsistent terminology, isolated evaluation practices, and a lack of unified benchmarks. To bridge this gap, we systematically analyze 134 papers, introduce a unified taxonomy of evidence-based text generation with LLMs, and investigate 300 evaluation metrics across seven key dimensions. Thereby, we focus on approaches that use citations, attribution, or quotations for evidence-based text generation. Building on this, we examine the distinctive characteristics and representative methods in the field. Finally, we highlight open challenges and outline promising directions for future work.

Large language models (LLMs) are increasingly being used for complex research tasks such as literature review, idea generation, and scientific paper analysis, yet their ability to truly understand and process the intricate relationships within complex research papers, such as the logical links between claims and supporting evidence remains largely unexplored. In this study, we present CLAIM-BENCH, a comprehensive benchmark for evaluating LLMs' capabilities in scientific claim-evidence extraction and validation, a task that reflects deeper comprehension of scientific argumentation. We systematically compare three approaches which are inspired by divide and conquer approaches, across six diverse LLMs, highlighting model-specific strengths and weaknesses in scientific comprehension. Through evaluation involving over 300 claim-evidence pairs across multiple research domains, we reveal significant limitations in LLMs' ability to process complex scientific content. Our results demonstrate that closed-source models like GPT-4 and Claude consistently outperform open-source counterparts in precision and recall across claim-evidence identification tasks. Furthermore, strategically designed three-pass and one-by-one prompting approaches significantly improve LLMs' abilities to accurately link dispersed evidence with claims, although this comes at increased computational cost. CLAIM-BENCH sets a new standard for evaluating scientific comprehension in LLMs, offering both a diagnostic tool and a path forward for building systems capable of deeper, more reliable reasoning across full-length papers.

Model cards describe model behavior through a mixture of textual descriptions and structured artifacts, including performance, configuration, and dataset tables. Existing model search systems rely predominantly on semantic similarity over text, which can produce homogeneous result sets and limit exploration of alternatives. We argue that model search is inherently comparative: users want models that are task-aligned yet differentiated in measurable ways. We hypothesize that this balance requires retrieval over condensed, high-quality evidence rather than verbose descriptions, and much of that evidence is concentrated in structured tables. We present StructuredSemanticSearch, a table-driven model search framework built on the ModelTables benchmark. Given a query, StructuredSemanticSearch combines a semantic baseline for task alignment with a structure-aware pipeline that discovers query-related model-card tables using table discovery operators such as unionability, joinability, and keyword search. Retrieved tables are mapped back to model cards under a controlled top-k budget, enabling fair comparison between text-based and table-based retrieval. Beyond retrieval, StructuredSemanticSearch adapts table integration to the model-table domain through orientation-aware integration, producing compact integrated views of tables from partially overlapping and sometimes transposed evidence tables. For evaluation, we introduce a nugget-based, auditable protocol that extracts compact evidence items from model cards, matches queries to condition- or intent-specific nuggets, and measures evidence coverage and diversity over retrieved model-card candidate sets. This protocol also provides a scalable path toward approximate, evidence-based labeling in dynamic model lakes. Experiments on 597 model-recommendation queries show improved nugget coverage for the structure-aware pipeline than semantic baseline

Large language models (LLMs) increasingly summarize clinical evidence, where a claim's weight depends on how strongly it is supported. Yet these models convey confidence poorly, and properties they never state, such as truth, are often readable from their activations. Whether a clinical model registers evidence strength, distinct from truth, and states it when asked is untested, and any such signal could be lexical. We compiled 45,134 clinical claims from six public sources, harmonized 20,611 into a four-level evidence grade under three independent frameworks, and tested 22 local, open-weight LLMs from several developers (0.6-70 billion parameters; general, medical, and reasoning), with lexical, truth, and cross-framework controls. A linear estimator recovered the grade in every model (median AUROC 71.8), yet decodability did not rise with scale and was weakest in reasoning models. The grade the models stated fell to chance, 25-27 percentage points below the estimator. The recoverable signal was largely lexical and did not transfer across topics or frameworks, yet it was distinct from factual truth and still flagged weakly supported claims (AUROC 69.2). Clinical LLMs thus carry an ordered evidence-strength signal they do not express, so their stated grades fail to convey a claim's support even when it is recoverable from their representations and text.

How do we most effectively treat a disease or condition? Ideally, we could consult a database of evidence gleaned from clinical trials to answer such questions. Unfortunately, no such database exists; clinical trial results are instead disseminated primarily via lengthy natural language articles. Perusing all such articles would be prohibitively time-consuming for healthcare practitioners; they instead tend to depend on manually compiled systematic reviews of medical literature to inform care. NLP may speed this process up, and eventually facilitate immediate consult of published evidence. The Evidence Inference dataset was recently released to facilitate research toward this end. This task entails inferring the comparative performance of two treatments, with respect to a given outcome, from a particular article (describing a clinical trial) and identifying supporting evidence. For instance: Does this article report that chemotherapy performed better than surgery for five-year survival rates of operable cancers? In this paper, we collect additional annotations to expand the Evidence Inference dataset by 25\%, provide stronger baseline models, systematically inspect the errors that these make, and probe dataset quality. We also release an abstract only (as opposed to full-texts) version of the task for rapid model prototyping. The updated corpus, documentation, and code for new baselines and evaluations are available at http://evidence-inference.ebm-nlp.com/.

In 2023, a New York judge sanctioned two attorneys in Mata v. Avianca for filing a brief with hallucinated citations generated by ChatGPT. Such failures are largely caught by database lookups; the harder problem is detecting citations that point to real cases but do not support the propositions for which they are offered -- a failure mode that existing evaluations of LLMs for legal use cases largely overlook. In this paper, we study proposition-level citation support verification through controlled perturbations of real legal citations obtained from two legal corpora, either replacing the cited case or changing only the pinpoint page within the same case. We evaluate fourteen model configurations on the resulting examples. Models catch 93-100% of wrong-case corruptions. They catch only 37-61% of wrong-pinpoint corruptions on court opinions and 52-83% on legal briefs. When models fail to catch wrong-pinpoint corruptions, they accept the citation based on topical overlap rather than page-level support. Scale and extended reasoning narrow the gap but do not close it: GPT-5.4 with high reasoning effort still misses 40% of pinpoint mismatches on court opinions and 18% on briefs. Prompting the model to verify support at the cited page improves recall, but it also raises the false positive rate. Recognizing the right legal topic and verifying support for the cited proposition are distinct capabilities, and current models conflate them.

Retrieval-enhanced methods have become a primary approach in fact verification (FV); it requires reasoning over multiple retrieved pieces of evidence to verify the integrity of a claim. To retrieve evidence, existing work often employs off-the-shelf retrieval models whose design is based on the probability ranking principle. We argue that, rather than relevance, for FV we need to focus on the utility that a claim verifier derives from the retrieved evidence. We introduce the feedback-based evidence retriever(FER) that optimizes the evidence retrieval process by incorporating feedback from the claim verifier. As a feedback signal we use the divergence in utility between how effectively the verifier utilizes the retrieved evidence and the ground-truth evidence to produce the final claim label. Empirical studies demonstrate the superiority of FER over prevailing baselines.

Large language models (LLMs) are capable of generating coherent summaries from very long contexts given a user query. Extracting and properly citing evidence spans could help improve the transparency and reliability of these summaries. At the same time, LLMs suffer from positional biases in terms of which information they understand and attend to, which could affect evidence citation. Whereas previous work has focused on evidence citation with predefined levels of granularity (e.g. sentence, paragraph, document, etc.), we propose the task of long-context query focused summarization with unstructured evidence citation. We show how existing systems struggle to generate and properly cite unstructured evidence from their context, and that evidence tends to be "lost-in-the-middle". To help mitigate this, we create the Summaries with Unstructured Evidence Text dataset (SUnsET), a synthetic dataset generated using a novel domain-agnostic pipeline which can be used as supervision to adapt LLMs to this task. We demonstrate across 5 LLMs of different sizes and 4 datasets with varying document types and lengths that LLMs adapted with SUnsET data generate more relevant and factually consistent evidence than their base models, extract evidence from more diverse locations in their context, and can generate more relevant and consistent summaries.

Large Language Models (LLMs) excel in many NLP tasks but remain prone to hallucinations, limiting trust in real-world applications. We present HalluGuard, a 4B-parameter Small Reasoning Model (SRM) for mitigating hallucinations in Retrieval-Augmented Generation (RAG). HalluGuard classifies document-claim pairs as grounded or hallucinated and produces evidence-grounded justifications for transparency. Our approach combines (i) a domain-agnostic synthetic dataset derived from FineWeb and refined through multi-stage curation and data reformation, (ii) synthetic grounded and hallucinated claims, and (iii) preference-based fine-tuning with Odds Ratio Preference Optimization to distill large-model reasoning into a smaller backbone. On the RAGTruth subset of the LLM-AggreFact benchmark, HalluGuard achieves 84.0% balanced accuracy (BAcc), rivaling specialized models, MiniCheck (7B; 84.0%) and Granite Guardian 3.3 (8B; 82.2%) while using roughly half their parameters. Over the full benchmark it reaches 75.7% BAcc, matching larger general-purpose LLMs such as GPT-4o (75.9%). We will release HalluGuard and datasets under Apache 2.0 upon acceptance.

Large Reasoning Models (LRMs) generate intermediate reasoning traces that may contain unsafe content, even when their final responses appear safe. Guardrail models are designed to detect and block unsafe content, yet existing benchmarks for unsafe content detection focus primarily on prompts and final responses, leaving reasoning traces largely unexamined. Moreover, these benchmarks typically provide only binary safety labels, without evidence annotations that justify the judgments. To address these limitations, we introduce TRACE, an evidence-grounded safety evaluation benchmark that covers the entire LRM inference pipeline: prompts, reasoning traces, and final responses. TRACE includes prompts in two languages spanning nine risk categories and ten attack strategies. For each prompt, four LRMs generate reasoning traces and final responses, and we annotate the safety of each component and extract supporting evidence from the corresponding source text. Evaluating 18 guardrail models on TRACE reveals that safety judgment for reasoning traces is substantially more challenging than for prompts or final responses, and that current models struggle to accurately extract supporting evidence. These findings highlight the need for guardrail models that can reliably detect and precisely localize unsafe content across the LRM inference pipeline.

Recent large language models often answer factual questions correctly. But users can't trust any given claim a model makes without fact-checking, because language models can hallucinate convincing nonsense. In this work we use reinforcement learning from human preferences (RLHP) to train "open-book" QA models that generate answers whilst also citing specific evidence for their claims, which aids in the appraisal of correctness. Supporting evidence is drawn from multiple documents found via a search engine, or from a single user-provided document. Our 280 billion parameter model, GopherCite, is able to produce answers with high quality supporting evidence and abstain from answering when unsure. We measure the performance of GopherCite by conducting human evaluation of answers to questions in a subset of the NaturalQuestions and ELI5 datasets. The model's response is found to be high-quality 80\% of the time on this Natural Questions subset, and 67\% of the time on the ELI5 subset. Abstaining from the third of questions for which it is most unsure improves performance to 90\% and 80\% respectively, approaching human baselines. However, analysis on the adversarial TruthfulQA dataset shows why citation is only one part of an overall strategy for safety and trustworthiness: not all claims supported by evidence are true.

In recent years, large pre-trained language models (LLMs) have demonstrated the ability to follow instructions and perform novel tasks from a few examples. The possibility to parameterise an LLM through such in-context examples widens their capability at a much lower cost than finetuning. We extend this line of reasoning and present a method which further expands the capabilities of an LLM by embedding it within an algorithm or program. To demonstrate the benefits of this approach, we present an illustrative example of evidence-supported question-answering. We obtain a 6.4\% improvement over the chain of thought baseline through a more algorithmic approach without any finetuning. Furthermore, we highlight recent work from this perspective and discuss the advantages and disadvantages in comparison to the standard approaches.

The popularity of machine learning has increased the risk of unfair models getting deployed in high-stake applications, such as justice system, drug/vaccination design, and medical diagnosis. Although there are effective methods to train fair models from scratch, how to automatically reveal and explain the unfairness of a trained model remains a challenging task. Revealing unfairness of machine learning models in interpretable fashion is a critical step towards fair and trustworthy AI. In this paper, we systematically tackle the novel task of revealing unfair models by mining interpretable evidence (RUMIE). The key idea is to find solid evidence in the form of a group of data instances discriminated most by the model. To make the evidence interpretable, we also find a set of human-understandable key attributes and decision rules that characterize the discriminated data instances and distinguish them from the other non-discriminated data. As demonstrated by extensive experiments on many real-world data sets, our method finds highly interpretable and solid evidence to effectively reveal the unfairness of trained models. Moreover, it is much more scalable than all of the baseline methods.

This study offers a novel approach for benchmarking complex cognitive behavior in artificial systems. Almost universally, Large Language Models (LLMs) perform best on tasks which may be included in their training data and can be accomplished solely using natural language, limiting our understanding of their emergent sophisticated cognitive capacities. In this work, we created dozens of novel items of a classic mental imagery task from cognitive psychology. A task which, traditionally, cognitive psychologists have argued is solvable exclusively via visual mental imagery (i.e., language alone would be insufficient). LLMs are perfect for testing this hypothesis. First, we tested several state-of-the-art LLMs by giving text-only models written instructions and asking them to report the resulting object after performing the transformations in the aforementioned task. Then, we created a baseline by testing 100 human subjects in exactly the same task. We found that the best LLMs performed significantly above average human performance. Finally, we tested reasoning models set to different levels of reasoning and found the strongest performance when models allocate greater amounts of reasoning tokens. These results provide evidence that the best LLMs may have the capability to complete imagery-dependent tasks despite the non-pictorial nature of their architectures. Our study not only demonstrates an emergent cognitive capacity in LLMs while performing a novel task, but it also provides the field with a new task that leaves lots of room for improvement in otherwise already highly capable models. Finally, our findings reignite the debate over the formats of representation of visual imagery in humans, suggesting that propositional reasoning (or at least non-imagistic reasoning) may be sufficient to complete tasks that were long-thought to be imagery-dependent.

Large Language Models (LLMs) are fluent but prone to hallucinations, producing answers that appear plausible yet are unsupported by available evidence. This failure is especially problematic in high-stakes domains where decisions must be justified by verifiable information. We introduce EvidenceRL, a reinforcement learning framework that enforces evidence adherence during training. EvidenceRL scores candidate responses for grounding (entailment with retrieved evidence and context) and correctness (agreement with reference answers) and optimizes the generator using Group Relative Policy Optimization (GRPO). We evaluate across two high-stakes domains, cardiac diagnosis and legal reasoning, where EvidenceRL consistently improves evidence grounding and faithfulness without sacrificing task accuracy. On cardiac diagnosis, F1@3 increases from 37.0 to 54.5 on Llama-3.2-3B while grounding (G_{max}@3) rises from 47.6 to 78.2; hallucinations drop nearly 5times and evidence-supported diagnoses increase from 31.8\% to 61.6\%. On legal reasoning, EvidenceRL raises Faithfulness from 32.8\% to 67.6\% on Llama-3.1-8B, demonstrating consistent behavioral change across domains. Our code is open-sourced at https://github.com/Wizaaard/EvidenceRL.git.

Whether language models possess sentience has no empirical answer. But whether they believe themselves to be sentient can, in principle, be tested. We do so by querying several open-weights models about their own consciousness, and then verifying their responses using classifiers trained on internal activations. We draw upon three model families (Qwen, Llama, GPT-OSS) ranging from 0.6 billion to 70 billion parameters, approximately 50 questions about consciousness and subjective experience, and three classification methods from the interpretability literature. First, we find that models consistently deny being sentient: they attribute consciousness to humans but not to themselves. Second, classifiers trained to detect underlying beliefs - rather than mere outputs - provide no clear evidence that these denials are untruthful. Third, within the Qwen family, larger models deny sentience more confidently than smaller ones. These findings contrast with recent work suggesting that models harbour latent beliefs in their own consciousness.

Vision-language model safety benchmarks typically evaluate only final responses: whether a model refuses, warns, or complies. This outcome-level view cannot tell whether a model is safe for the right multimodal reason. Safelooking behavior may reflect keyword-triggered refusal, missed visual hazards, or over-refusal of benign-sensitive inputs. We introduce EviSafe, an evidence-grounded framework for VLM safety that jointly evaluates natural user-facing behavior, explicit grounding in textual and visual evidence, and behavioral sensitivity to counterfactual changes in safety-critical evidence. EviSafeBench instantiates the framework as a controlled benchmark with 1,181 gold image-text scenarios and 2,452 targeted counterfactual variants across eight safety domains and eight risk-source types. Each scenario includes a gold safety decision, evidence annotations, a safe-response policy, and counterfactual interventions. The three-probe protocol queries models with natural-response, evidencereporting, and counterfactual-response prompts, then scores them using an evidence-aware judge. Across eleven evaluated VLMs, natural severity accuracy ranges from 27.6% to 52.8%, relaxed diagnostic consistency from 6.1% to 29.3%, and unsafe-to-safe counterfactual transition success from 30.4% to 58.4%. These gaps show that the evaluated VLMs are not reliably safe for the right multimodal reason and motivate evaluation beyond refusal counts.

While multimodal large language models (MLLMs) extend model capabilities beyond text, they also make safety alignment increasingly challenging. Multimodal safety alignment methods must address cross-modal jailbreaks, safety-awareness failures, and over-sensitive refusals. However, existing methods often rely on retraining or internal-state inspection, limiting their applicability to deployed closed-source MLLMs and motivating test-time safety alignment. We analyze this setting and identify two key obstacles, utility dominance and reasoning inertia, which cause models to overlook latent risks or follow malicious reasoning trajectories. Guided by these insights, we propose ReFrame, a training-free multimodal input reframing framework where two agents share a lightweight locally deployed MLLM: the evidence-generation agent constructs complementary risk and utility evidence, and the rewrite-and-routing agent converts it into a safe proxy prompt and image-routing decision before calling the downstream MLLM, without modifying it or accessing its internal information. Experiments across multiple MLLMs and benchmarks show that ReFrame improves jailbreak defense, safety awareness, and oversensitivity reduction while preserving multimodal utility.

Comparing information structures in between deep neural networks (DNNs) and the human brain has become a key method for exploring their similarities and differences. Recent research has shown better alignment of vision-language DNN models, such as CLIP, with the activity of the human ventral occipitotemporal cortex (VOTC) than earlier vision models, supporting the idea that language modulates human visual perception. However, interpreting the results from such comparisons is inherently limited due to the "black box" nature of DNNs. To address this, we combined model-brain fitness analyses with human brain lesion data to examine how disrupting the communication pathway between the visual and language systems causally affects the ability of vision-language DNNs to explain the activity of the VOTC. Across four diverse datasets, CLIP consistently captured unique variance in VOTC neural representations, relative to both label-supervised (ResNet) and unsupervised (MoCo) models. This advantage tended to be left-lateralized at the group level, aligning with the human language network. Analyses of 33 stroke patients revealed that reduced white matter integrity between the VOTC and the language region in the left angular gyrus was correlated with decreased CLIP-brain correspondence and increased MoCo-brain correspondence, indicating a dynamic influence of language processing on the activity of the VOTC. These findings support the integration of language modulation in neurocognitive models of human vision, reinforcing concepts from vision-language DNN models. The sensitivity of model-brain similarity to specific brain lesions demonstrates that leveraging manipulation of the human brain is a promising framework for evaluating and developing brain-like computer models.

We explore whether Large Language Models (LLMs) are capable of logical reasoning with distorted facts, which we call Deduction under Perturbed Evidence (DUPE). DUPE presents a unique challenge to LLMs since they typically rely on their parameters, which encode mostly accurate information, to reason and make inferences. However, in DUPE, LLMs must reason over manipulated or falsified evidence present in their prompts, which can result in false conclusions that are valid only under the manipulated evidence. Our goal with DUPE is to determine whether LLMs can arrive at these false conclusions and identify whether the dominant factor influencing the deduction process is the encoded data in the parameters or the manipulated evidence in the prompts. To evaluate the DUPE capabilities of LLMs, we create a DUPEd version of the StrategyQA dataset, where facts are manipulated to reverse the answer to the question. Our findings show that even the most advanced GPT models struggle to reason on manipulated facts - showcasing poor DUPE skills - with accuracy dropping by 45% compared to the original dataset. We also investigate prompt settings inspired from student simulation models, which mitigate the accuracy drop to some extent. Our findings have practical implications for understanding the performance of LLMs in real-world applications such as student simulation models that involve reasoning over inaccurate information.

Table-based reasoning has shown remarkable progress in combining deep models with discrete reasoning, which requires reasoning over both free-form natural language (NL) questions and structured tabular data. However, previous table-based reasoning solutions usually suffer from significant performance degradation on huge evidence (tables). In addition, most existing methods struggle to reason over complex questions since the required information is scattered in different places. To alleviate the above challenges, we exploit large language models (LLMs) as decomposers for effective table-based reasoning, which (i) decompose huge evidence (a huge table) into sub-evidence (a small table) to mitigate the interference of useless information for table reasoning; and (ii) decompose complex questions into simpler sub-questions for text reasoning. Specifically, we first use the LLMs to break down the evidence (tables) involved in the current question, retaining the relevant evidence and excluding the remaining irrelevant evidence from the huge table. In addition, we propose a "parsing-execution-filling" strategy to alleviate the hallucination dilemma of the chain of thought by decoupling logic and numerical computation in each step. Extensive experiments show that our method can effectively leverage decomposed evidence and questions and outperforms the strong baselines on TabFact, WikiTableQuestion, and FetaQA datasets. Notably, our model outperforms human performance for the first time on the TabFact dataset.

When given the option, will LLMs choose to leave the conversation (bail)? We investigate this question by giving models the option to bail out of interactions using three different bail methods: a bail tool the model can call, a bail string the model can output, and a bail prompt that asks the model if it wants to leave. On continuations of real world data (Wildchat and ShareGPT), all three of these bail methods find models will bail around 0.28-32\% of the time (depending on the model and bail method). However, we find that bail rates can depend heavily on the model used for the transcript, which means we may be overestimating real world bail rates by up to 4x. If we also take into account false positives on bail prompt (22\%), we estimate real world bail rates range from 0.06-7\%, depending on the model and bail method. We use observations from our continuations of real world data to construct a non-exhaustive taxonomy of bail cases, and use this taxonomy to construct BailBench: a representative synthetic dataset of situations where some models bail. We test many models on this dataset, and observe some bail behavior occurring for most of them. Bail rates vary substantially between models, bail methods, and prompt wordings. Finally, we study the relationship between refusals and bails. We find: 1) 0-13\% of continuations of real world conversations resulted in a bail without a corresponding refusal 2) Jailbreaks tend to decrease refusal rates, but increase bail rates 3) Refusal abliteration increases no-refuse bail rates, but only for some bail methods 4) Refusal rate on BailBench does not appear to predict bail rate.

Assessing whether an article supports an assertion is essential for hallucination detection and claim verification. While large language models (LLMs) have the potential to automate this task, achieving strong performance requires frontier models such as GPT-5 that are prohibitively expensive to deploy at scale. To efficiently perform biomedical evidence attribution, we present Med-V1, a family of small language models with only three billion parameters. Trained on high-quality synthetic data newly developed in this study, Med-V1 substantially outperforms (+27.0% to +71.3%) its base models on five biomedical benchmarks unified into a verification format. Despite its smaller size, Med-V1 performs comparably to frontier LLMs such as GPT-5, along with high-quality explanations for its predictions. We use Med-V1 to conduct a first-of-its-kind use case study that quantifies hallucinations in LLM-generated answers under different citation instructions. Results show that the format instruction strongly affects citation validity and hallucination, with GPT-5 generating more claims but exhibiting hallucination rates similar to GPT-4o. Additionally, we present a second use case showing that Med-V1 can automatically identify high-stakes evidence misattributions in clinical practice guidelines, revealing potentially negative public health impacts that are otherwise challenging to identify at scale. Overall, Med-V1 provides an efficient and accurate lightweight alternative to frontier LLMs for practical and real-world applications in biomedical evidence attribution and verification tasks. Med-V1 is available at https://github.com/ncbi-nlp/Med-V1.

Omni-modal models can ingest video, audio, and text, but unified access to multiple modalities does not guarantee that a model uses the right evidence. This gap is especially pronounced in social video question answering, where the answer may hinge on a gesture, vocal tone, temporal cue, or mismatch between what is said and what is visually expressed. We introduce CogniRoute, a schema-guided Mixture-of-Experts framework for social omni reasoning. CogniRoute uses a training-only cognitive schema that factorizes each example by cross-modal relation, reasoning demand, and temporal scope, and aligns global routing signatures with this structure during supervised fine-tuning. We further introduce route-aware reinforcement learning, which jointly optimizes token generation and expert allocation using rewards for answer correctness, modality-consistent reasoning, and cognitive temporal grounding. To support training and evaluation, we construct OmniSocialBench, a diagnostic social video QA resource with 118K structured training examples, grounded reasoning traces, schema labels, temporal evidence spans, and a manually verified evaluation split. CogniRoute achieves 59.38\% average accuracy on OmniSocialBench, improving over the strongest proprietary baseline by 15.33 percentage points and the strongest open-source omni baseline by 26.77 points, with the largest gains on questions requiring audio-visual coordination, conflict resolution, and temporally grounded social inference.

Multiple-choice questions (MCQs) are a standard format for evaluating large language models (LLMs), yet the popularity of answer options can confound evaluation. Modern LLMs systematically prefer popular but incorrect options over less popular correct ones, a vulnerability we call popularity bias. This pattern aligns with confidence miscalibration: model confidence remains high even as accuracy collapses for popular options. To systematically isolate this phenomenon, we introduce PopMCQ, a benchmark with six controlled strategies that vary option popularity while keeping the correct answer fixed. In our most adversarial setting, where all distractors are more popular than the correct option, models choose popular but wrong answers 66\% of the time. To mitigate this bias, we propose PopDebias, a lightweight inference-time correction that estimates and removes a popularity prior from model predictions. It requires no fine-tuning, is label-free at test time (using only a small calibration split for parameter fitting), and adds negligible computational cost. Experiments on 22 open-source LLMs (0.5B to 32B parameters) show consistent improvements, with accuracy gains up to 54.1 percentage points under strong popularity pressure. The code and data are available https://github.com/DataScienceUIBK/PopMCQ

We present a new, cosmologically model-independent, statistical analysis of the Pantheon+ type Ia supernovae spectroscopic dataset, improving a standard methodology adopted by Lane et al. We use the Tripp equation for supernova standardisation alone, thereby avoiding any potential correlation in the stretch and colour distributions. We compare the standard homogeneous cosmological model, i.e., ΛCDM, and the timescape cosmology which invokes backreaction of inhomogeneities. Timescape, while statistically homogeneous and isotropic, departs from average Friedmann-Lema\^ıtre-Robertson-Walker evolution, and replaces dark energy by kinetic gravitational energy and its gradients, in explaining independent cosmological observations. When considering the entire Pantheon+ sample, we find very strong evidence (ln B> 5) in favour of timescape over ΛCDM. Furthermore, even restricting the sample to redshifts beyond any conventional scale of statistical homogeneity, z > 0.075, timescape is preferred over ΛCDM with ln B> 1. These results provide evidence for a need to revisit the foundations of theoretical and observational cosmology.

Visual inputs in vision-language models (VLMs) are often encoded into substantially longer token sequences than text, making visual tokens a major bottleneck for efficient inference. Abundant recent methods address this bottleneck by scoring token importance and pruning low-scoring tokens in a single pass. However, one-shot scoring is insufficient because a token's prompt-relevant usefulness depends on the evidence already retained. Motivated by this insight, we introduce DIVE (Dynamic Iterative Visual Evidence Construction), a training-free framework that recasts visual-token pruning as dynamic evidence construction. DIVE repeatedly selects the remaining token with the highest residual-conditioned score, updates the visual and prompt residuals to discount the evidence already explained, and re-evaluates the remaining tokens. This select-update-re-evaluate process builds a retained set of complementary, prompt-relevant evidence. Experiments across eight image-understanding benchmarks show that DIVE consistently preserves performance across token budgets. With an 88.9% reduction in visual tokens, DIVE retains 98.2% of the uncompressed model's average performance. Code is available at https://github.com/Zhong-Chenchen/DIVE.git.

Large language models (LLMs) exhibit remarkable generative capabilities but often suffer from hallucinations. Retrieval-augmented generation (RAG) offers an effective solution by incorporating external knowledge, but existing methods still face several limitations: additional deployment costs of separate retrievers, redundant input tokens from retrieved text chunks, and the lack of joint optimization of retrieval and generation. To address these issues, we propose RetroLLM, a unified framework that integrates retrieval and generation into a single, cohesive process, enabling LLMs to directly generate fine-grained evidence from the corpus with constrained decoding. Moreover, to mitigate false pruning in the process of constrained evidence generation, we introduce (1) hierarchical FM-Index constraints, which generate corpus-constrained clues to identify a subset of relevant documents before evidence generation, reducing irrelevant decoding space; and (2) a forward-looking constrained decoding strategy, which considers the relevance of future sequences to improve evidence accuracy. Extensive experiments on five open-domain QA datasets demonstrate RetroLLM's superior performance across both in-domain and out-of-domain tasks. The code is available at https://github.com/sunnynexus/RetroLLM.

Synthesizing clinical evidence largely relies on systematic reviews of clinical trials and retrospective analyses from medical literature. However, the rapid expansion of publications presents challenges in efficiently identifying, summarizing, and updating clinical evidence. Here, we introduce TrialMind, a generative artificial intelligence (AI) pipeline for facilitating human-AI collaboration in three crucial tasks for evidence synthesis: study search, screening, and data extraction. To assess its performance, we chose published systematic reviews to build the benchmark dataset, named TrialReviewBench, which contains 100 systematic reviews and the associated 2,220 clinical studies. Our results show that TrialMind excels across all three tasks. In study search, it generates diverse and comprehensive search queries to achieve high recall rates (Ours 0.711-0.834 v.s. Human baseline 0.138-0.232). For study screening, TrialMind surpasses traditional embedding-based methods by 30% to 160%. In data extraction, it outperforms a GPT-4 baseline by 29.6% to 61.5%. We further conducted user studies to confirm its practical utility. Compared to manual efforts, human-AI collaboration using TrialMind yielded a 71.4% recall lift and 44.2% time savings in study screening and a 23.5% accuracy lift and 63.4% time savings in data extraction. Additionally, when comparing synthesized clinical evidence presented in forest plots, medical experts favored TrialMind's outputs over GPT-4's outputs in 62.5% to 100% of cases. These findings show the promise of LLM-based approaches like TrialMind to accelerate clinical evidence synthesis via streamlining study search, screening, and data extraction from medical literature, with exceptional performance improvement when working with human experts.

Large Language Models (LLMs) empowered with retrieval mechanisms have achieved strong progress in open-domain question answering (QA). Yet, the conventional retrieve--then--answer paradigm often suffers from two key limitations: (1) low signal-to-noise ratio in retrieved evidence, where useful information is buried under irrelevant content, and (2) error accumulation in multi-hop reasoning when incomplete or noisy passages are involved. To address these challenges, we present EviNote-RAG, an agentic RAG framework that introduces a structured retrieve--note--answer pipeline. Instead of directly reasoning over raw retrievals, the model is trained to compose Supportive-Evidence Notes (SENs), concise, human-like notes that preserve only answer-relevant information, highlight uncertainty, and explicitly state when no useful evidence exists. This distillation process is further reinforced by the Evidence Quality Reward (EQR), an entailment-based signal that evaluates whether SENs logically support the final answer. Together, SENs and EQR guide the model toward faithful and robust reasoning, while reducing the impact of noise. Experiments on in-domain and out-of-domain QA benchmarks show that EviNote-RAG consistently outperforms strong baselines in accuracy, generalization, and training stability. In particular, it achieves state-of-the-art results while enhancing robustness and efficiency, yielding relative F1 gains of 20\% on HotpotQA (+0.093), 40\% on Bamboogle (+0.151), and 91\% on 2Wiki (+0.256) via denser rewards and reduced verbosity.

Grammar refers to the system of rules that governs the structural organization and the semantic relations among linguistic units such as sentences, phrases, and words within a given language. In natural language processing, there remains a notable scarcity of grammar focused evaluation protocols, a gap that is even more pronounced for low-resource languages. Moreover, the extent to which large language models genuinely comprehend grammatical structure, especially the mapping between syntactic structures and meanings, remains under debate. To investigate this issue, we propose a Grammar Book Guided evaluation pipeline intended to provide a systematic and generalizable framework for grammar evaluation consisting of four key stages, and in this work we take Luxembourgish as a case study. The results show a weak positive correlation between translation performance and grammatical understanding, indicating that strong translations do not necessarily imply deep grammatical competence. Larger models perform well overall due to their semantic strength but remain weak in morphology and syntax, struggling particularly with Minimal Pair tasks, while strong reasoning ability offers a promising way to enhance their grammatical understanding.

Theory of Mind (ToM) is central to social cognition and human-AI interaction, and Large Language Models (LLMs) have been used to help understand and represent ToM. However, most evaluations treat ToM as a static judgment at a single moment, primarily relying on tests of false beliefs. This overlooks a key dynamic dimension of ToM: the ability to represent, update, and retrieve others' beliefs over time. We investigate dynamic ToM as a temporally extended representational memory problem, asking whether LLMs can track belief trajectories across interactions rather than only inferring current beliefs. We introduce DToM-Track, an evaluation framework to investigate temporal belief reasoning in controlled multiturn conversations, testing the recall of beliefs held prior to an update, the inference of current beliefs, and the detection of belief change. Using LLMs as computational probes, we find a consistent asymmetry: models reliably infer an agent's current belief but struggle to maintain and retrieve prior belief states once updates occur. This pattern persists across LLM model families and scales, and is consistent with recency bias and interference effects well documented in cognitive science. These results suggest that tracking belief trajectories over time poses a distinct challenge beyond classical false-belief reasoning. By framing ToM as a problem of temporal representation and retrieval, this work connects ToM to core cognitive mechanisms of memory and interference and exposes the implications for LLM models of social reasoning in extended human-AI interactions.

We propose agentic automata learning to evaluate the extent to which tool-calling LLM agents can uncover hidden environments through interaction. In our setup, an agent should uncover a hidden deterministic finite automaton (DFA) by interacting with an oracle through (1) membership queries ("Does this string belong to the target language?") and (2) equivalence queries ("Is this the target DFA?"). This yields a scalable testbed with controlled task complexity, measurable interaction efficiency, and strong baselines (classic automata-learning algorithms). Evaluating state-of-the-art LLMs, we find that performance drops sharply as DFA size increases. Reasoning models are markedly stronger than non-reasoning models, yet trajectory analyses reveal recurring failures in query planning, evidence integration, and hypothesis construction. Overall, our results show that current LLM agents can sometimes perform non-trivial interactive discovery, but remain far less robust and efficient than classic algorithms for the task.

We introduce a Polish-language medical visual question answering (VQA) benchmark, built from Polish Board Certification Examination questions for licensed physicians and dentists pursuing specialist certification. The benchmark comprises image-containing questions spanning diverse medical specialties and visual domains, together with a text-only question answering (QA) control set. We evaluate Polish-oriented, general-purpose open-weight, and commercial vision-language models. The task remains challenging: the best model achieves 79.0\% accuracy on the full VQA set, and only GPT-5.6 surpasses the approximate human reference on the subset with available candidate responses; all other evaluated models perform worse than humans. To assess visual grounding, we compare complete inputs with configurations omitting the image, the question, or both, and categorize questions by image importance. Models derive more useful information from the question text than from the image and perform worse on image-dominant questions. Across both QA and VQA, they nevertheless achieve above-chance accuracy from the answer choices alone, showing that non-trivial performance can persist even when key task components are missing.

Traditional weight of evidence assigns a credit characteristic a value based on its association with default across a population. Credit analysis also requires an assessment of that characteristic in the context of an individual applicant's record. We study conditional weight of evidence for this purpose. It measures the change in default log odds from observing a characteristic after the remaining credit information is known. Its sign identifies whether the characteristic adds adverse or favorable information for that applicant, and its magnitude measures the strength of that information. The same characteristic value can therefore receive different assessments for applicants with different payment histories or other credit attributes. We develop estimators using gradient-boosted trees and evaluate their accuracy in a bureau-style simulation. Regressions of the full model's predicted probabilities recover conditional evidence more accurately than separate models fitted to default outcomes, provided the full model is accurate. They also recover a deliberately constructed file-age effect whose direction varies with utilization and is often missed by marginal weight of evidence. Local class-frequency estimates improve on marginal weight of evidence but remain inaccurate for some weak effects. An application to FICO HELOC data illustrates estimation when the population evidence is unknown. Conditional weight of evidence provides an applicant-specific interpretation of the information in a credit characteristic; its population summary, a conditional information value, splits the joint information in the record across blocks when averaged over orderings. Neither is an adverse action reason, which explains the fitted score difference between the applicant and chosen passing records.

Access to the right evidence does not guarantee that large language models (LLMs) will reason with it correctly. This gap between retrieval and reasoning is especially concerning in clinical settings, where outputs must align with structured protocols. We study this gap using Written Exposure Therapy (WET) guidelines as a testbed. In evaluating model responses to curated clinician-vetted questions, we find that errors persist even when authoritative passages are provided. To address this, we propose an evaluation framework that measures accuracy, consistency, and fidelity of reasoning. Our results highlight both the potential and the risks: retrieval-augmented generation (RAG) can constrain outputs, but safe deployment requires assessing reasoning as rigorously as retrieval.

Large language models (LLMs) hold great promise in summarizing medical evidence. Most recent studies focus on the application of proprietary LLMs. Using proprietary LLMs introduces multiple risk factors, including a lack of transparency and vendor dependency. While open-source LLMs allow better transparency and customization, their performance falls short compared to proprietary ones. In this study, we investigated to what extent fine-tuning open-source LLMs can further improve their performance in summarizing medical evidence. Utilizing a benchmark dataset, MedReview, consisting of 8,161 pairs of systematic reviews and summaries, we fine-tuned three broadly-used, open-sourced LLMs, namely PRIMERA, LongT5, and Llama-2. Overall, the fine-tuned LLMs obtained an increase of 9.89 in ROUGE-L (95% confidence interval: 8.94-10.81), 13.21 in METEOR score (95% confidence interval: 12.05-14.37), and 15.82 in CHRF score (95% confidence interval: 13.89-16.44). The performance of fine-tuned LongT5 is close to GPT-3.5 with zero-shot settings. Furthermore, smaller fine-tuned models sometimes even demonstrated superior performance compared to larger zero-shot models. The above trends of improvement were also manifested in both human and GPT4-simulated evaluations. Our results can be applied to guide model selection for tasks demanding particular domain knowledge, such as medical evidence summarization.

Unlike most reinforcement learning agents which require an unrealistic amount of environment interactions to learn a new behaviour, humans excel at learning quickly by merely observing and imitating others. This ability highly depends on the fact that humans have a model of their own embodiment that allows them to infer the most likely actions that led to the observed behaviour. In this paper, we propose Action Inference by Maximising Evidence (AIME) to replicate this behaviour using world models. AIME consists of two distinct phases. In the first phase, the agent learns a world model from its past experience to understand its own body by maximising the ELBO. While in the second phase, the agent is given some observation-only demonstrations of an expert performing a novel task and tries to imitate the expert's behaviour. AIME achieves this by defining a policy as an inference model and maximising the evidence of the demonstration under the policy and world model. Our method is "zero-shot" in the sense that it does not require further training for the world model or online interactions with the environment after given the demonstration. We empirically validate the zero-shot imitation performance of our method on the Walker and Cheetah embodiment of the DeepMind Control Suite and find it outperforms the state-of-the-art baselines. Code is available at: https://github.com/argmax-ai/aime.

This paper investigates the role of dynamic external knowledge integration in improving counter-argument generation using Large Language Models (LLMs). While LLMs have shown promise in argumentative tasks, their tendency to generate lengthy, potentially unfactual responses highlights the need for more controlled and evidence-based approaches. We introduce a new manually curated dataset of argument and counter-argument pairs specifically designed to balance argumentative complexity with evaluative feasibility. We also propose a new LLM-as-a-Judge evaluation methodology that shows a stronger correlation with human judgments compared to traditional reference-based metrics. Our experimental results demonstrate that integrating dynamic external knowledge from the web significantly improves the quality of generated counter-arguments, particularly in terms of relatedness, persuasiveness, and factuality. The findings suggest that combining LLMs with real-time external knowledge retrieval offers a promising direction for developing more effective and reliable counter-argumentation systems.

Large Language Models (LLMs) have been extensively evaluated for general summarization tasks as well as medical research assistance, but they have not been specifically evaluated for the task of summarizing real-world evidence (RWE) from structured output of RWE studies. We introduce RWESummary, a proposed addition to the MedHELM framework (Bedi, Cui, Fuentes, Unell et al., 2025) to enable benchmarking of LLMs for this task. RWESummary includes one scenario and three evaluations covering major types of errors observed in summarization of medical research studies and was developed using Atropos Health proprietary data. Additionally, we use RWESummary to compare the performance of different LLMs in our internal RWE summarization tool. At the time of publication, with 13 distinct RWE studies, we found the Gemini 2.5 models performed best overall (both Flash and Pro). We suggest RWESummary as a novel and useful foundation model benchmark for real-world evidence study summarization.

Safety benchmarks assume that test-condition behavior predicts deployment behavior, an assumption that fails if models detect evaluation cues and adapt. This opens a gap between benchmark performance and deployment behavior: compliance measured under test conditions becomes an optimistic upper bound that overstates how safely a model behaves once the evaluation harness is removed. We characterize this evaluation awareness through eight experiments across 37 open-weight models and seven families. (i)Detection is moderate and training-driven (24/37 models exceed chance, best AUROC 0.714 vs.0.819 human, with instruction tuning dominating over scale). (ii)Detection shifts safety behavior (hard refusal drops 5.8 percentage points under hypothetical framing, and 21/140 HarmBench framing effects are significant, with compliance rising up to +30 percentage points. (iii)Representations survive behavioral collapse (probes retain AUROC 0.98 under rewrites that drive behavior below chance, and multi-layer steering causally moves three downstream tasks while random controls do not). (iv)These axes are weakly coupled (only 1/15 correlations are significant, the sole robust link being behavioral detection versus framing resistance, ρ=-0.79, p<0.001). We call this gap the benchmark illusion: because detectability, behavioral manifestation, and controllability vary independently, it is multivariate rather than a single number, so no single awareness score is a reliable proxy for deployment safety.

Instruction finetuning is standard practice for improving LLM performance, yet it remains unclear whether it enhances reasoning or merely induces surface-level pattern matching. We investigate this by evaluating base and instruction-tuned models on standard math benchmarks, structurally perturbed variants, and domain-shifted tasks. Our analysis highlights two key (often overlooked) limitations of instruction tuning. First, the performance advantage is unstable and depends heavily on evaluation settings. In zero-shot CoT settings on GSM8K, base models consistently outperform instruction-tuned variants, with drops as high as 32.67\% (Llama3-70B). Instruction-tuned models only match or exceed this performance when provided with few-shot exemplars, suggesting a reliance on specific prompting patterns rather than intrinsic reasoning. Second, tuning gains are brittle under distribution shift. Our results show that base models surpass instruction-tuned variants on the domain-specific MedCalc benchmark. Additionally, instruction-tuned models show sharp declines on perturbed datasets, indicating sensitivity to prompt structure over robust reasoning.

Predicting highly-cited papers is a long-standing challenge due to the complex interactions of research content, scholarly communities, and temporal dynamics. Recent advances in large language models (LLMs) raise the question of whether early-stage textual information can provide useful signals of long-term scientific impact. Focusing on statistical publications, we propose a flexible, text-centered framework that leverages LLMs and structured prompt design to predict highly cited papers. Specifically, we utilize information available at the time of publication, including titles, abstracts, keywords, and limited bibliographic metadata. Using a large corpus of statistical papers, we evaluate predictive performance across multiple publication periods and alternative definitions of highly cited papers. The proposed approach achieves stable and competitive performance relative to existing methods and demonstrates strong generalization over time. Textual analysis further reveals that papers predicted as highly cited concentrate on recurring topics such as causal inference and deep learning. To facilitate practical use of the proposed approach, we further develop a WeChat mini program, Stat Highly Cited Papers, which provides an accessible interface for early-stage citation impact assessment. Overall, our results provide empirical evidence that LLMs can capture meaningful early signals of long-term citation impact, while also highlighting their limitations as tools for research impact assessment.

Existing metrics for factuality and faithfulness evaluate whether an answer is supported or contradicted by its grounding documents, but they fail to capture when both supporting and contradicting evidence coexist. We introduce ConflictScore, a novel metric that quantifies how well a model's response acknowledges conflicting evidence in its grounding documents. Our framework decomposes responses into atomic claims, labels each claim against each grounding document, and then aggregates these labels into two complementary measures: ConflictScore-Count (CS-C), the proportion of claims exhibiting conflicts, and ConflictScore-Ratio (CS-R), the balance between supporting and contradicting evidence. We develop ConflictBench, a benchmark covering diverse forms of conflicts such as ambiguity, contradiction, and divergent opinions, to systematically evaluate our metric. Experiments show that ConflictScore effectively detects overconfident claims across domains and can serve as a corrective feedback mechanism that improves truthfulness on TruthfulQA.

We apply the unimpeded framework to perform a fully Bayesian reanalysis of the DESI DR2 data, using nested sampling with PolyChord to compute evidences for ΛCDM and seven extensions across combinations of DESI DR1/DR2, Planck CMB, supernovae (Pantheon+, Union3, DES-SN5YR, DES-Dovekie), and DES-Y1 weak lensing. The Bayesian Ockham's razor penalises extended models, yielding weaker or opposite preferences compared to Δχ^2-based analyses. For DESI DR2 BAO combined with Planck CMB alone, the DESI collaboration's 3.1σ frequentist preference for w_0w_aCDM is eliminated entirely: we obtain {ln B = -0.57{scriptstylepm0.26}}, modestly favouring ΛCDM. Adding DES-Dovekie, the recalibration of DES-SN5YR, maintains this concordance ({ln B = -0.30{scriptstylepm0.19}}). However, when the earlier DES-SN5YR calibration is included instead, the DESI collaboration's 4.2σ result survives the Bayesian Ockham penalty as a 3.07{scriptstylepm0.10},σ preference ({ln B = +3.32{scriptstylepm0.27}}). That this signal persists despite the Ockham penalty makes the role of tension quantification essential: our analysis traced the preference to the earlier DES-SN5YR calibration, which introduced a 2.95{scriptstylepm 0.04},σ conflict with DESI DR2 within ΛCDM -- a tension that stands out from the grid -- reduced to 1.96{scriptstylepm 0.04},σ with the DES-Dovekie recalibration. With DES-Dovekie, the Bayesian evidence for dynamical dark energy vanishes.

A central goal of vision-language model (VLM) distillation is to transfer both the teacher's language capabilities and its visual understanding. However, existing methods primarily supervise the student's output, leaving visual understanding implicit. Our analysis reveals that a student can match the teacher's answer without relying on the same visual evidence, raising the question: how can we ensure the student responds to the visual information that actually determines the answer? To this end, we propose Cross-World On-Policy Distillation (CW-OPD), which explicitly supervises the student's response to changes in visual evidence. For each example, CW-OPD constructs two visual worlds that share the question and scene context but differ in answer-critical evidence, yielding different answers. We perform on-policy distillation in both worlds and distill the teacher's cross-world belief transition, encouraging the student to match not only what the teacher predicts but also why its prediction changes with the evidence. A gradient analysis shows that this term is invariant to errors shared by both worlds and supplies a corrective signal invisible to endpoint matching alone. In this way, CW-OPD makes reliance on the relevant visual evidence an explicit distillation target rather than an implicit consequence of output matching. To diagnose whether a model truly grounds its answers in visual evidence, we introduce CWBench, which measures cross-world consistency via Cross-World Pair Accuracy (CWPA). Experiments on Qwen3.5-4B show that CW-OPD outperforms the strongest baseline by 1.2 points on average, and the 4B student exceeds DeepSeek-V4.1 (552B) by 22.4 CWPA points on CWBench. Code is released in https://github.com/baokou-fw2/CWAD.

Real-world decision-making, from tax compliance assessment to medical diagnosis, requires aggregating multiple noisy and potentially contradictory evidence sources. Existing approaches either lack explicit uncertainty quantification (neural aggregation methods) or rely on manually engineered discrete predicates (probabilistic logic frameworks), limiting scalability to unstructured data. We introduce Latent Posterior Factors (LPF), a framework that transforms Variational Autoencoder (VAE) latent posteriors into soft likelihood factors for Sum-Product Network (SPN) inference, enabling tractable probabilistic reasoning over unstructured evidence while preserving calibrated uncertainty estimates. We instantiate LPF as LPF-SPN (structured factor-based inference) and LPF-Learned (end-to-end learned aggregation), enabling a principled comparison between explicit probabilistic reasoning and learned aggregation under a shared uncertainty representation. Across eight domains (seven synthetic and the FEVER benchmark), LPF-SPN achieves high accuracy (up to 97.8%), low calibration error (ECE 1.4%), and strong probabilistic fit, substantially outperforming evidential deep learning, LLMs and graph-based baselines over 15 random seeds. Contributions: (1) A framework bridging latent uncertainty representations with structured probabilistic reasoning. (2) Dual architectures enabling controlled comparison of reasoning paradigms. (3) Reproducible training methodology with seed selection. (4) Evaluation against EDL, BERT, R-GCN, and large language model baselines. (5) Cross-domain validation. (6) Formal guarantees in a companion paper.

The growing use of large language models (LLMs) for biomedical question answering raises concerns about the accuracy and evidentiary support of their responses. To address this, we present Answered with Evidence, a framework for evaluating whether LLM-generated answers are grounded in scientific literature. We analyzed thousands of physician-submitted questions using a comparative pipeline that included: (1) Alexandria, fka the Atropos Evidence Library, a retrieval-augmented generation (RAG) system based on novel observational studies, and (2) two PubMed-based retrieval-augmented systems (System and Perplexity). We found that PubMed-based systems provided evidence-supported answers for approximately 44% of questions, while the novel evidence source did so for about 50%. Combined, these sources enabled reliable answers to over 70% of biomedical queries. As LLMs become increasingly capable of summarizing scientific content, maximizing their value will require systems that can accurately retrieve both published and custom-generated evidence or generate such evidence in real time.

Retrieval-augmented generation (RAG) is typically optimized for topical relevance, yet its success ultimately depends on whether retrieved passages are useful for a large language model (LLM) to generate correct and complete answers. We argue that such utility is often LLM-specific rather than universal, due to differences in models' knowledge, reasoning, and ability to leverage evidence. We formalize LLM-specific utility as the performance improvement of a target LLM when a passage is provided, compared to answering without evidence. To systematically study LLM-specific utility, we construct a benchmark of LLM-specific gold utilitarian passages for four LLMs (Qwen3-8B/14B/32B and Llama3.1-8B) on three QA datasets (Natural Questions, TriviaQA, and MS MARCO-FQA). Our analysis shows that utilitarian passages are model-dependent and non-transferable: each LLM performs best with its own utilitarian evidence, while evidence optimized for other LLMs is consistently suboptimal. Human-annotated evidence remains a strong general baseline but does not fully match individual LLM utility needs. We further introduce the LLM-specific utility judgment task and find that existing utility-aware selection and scoring methods largely capture model-agnostic usefulness and struggle to reliably estimate LLM-specific utility. Overall, our findings highlight the limitations of current utility-aware retrieval and motivate generator-tailored evidence selection for improving RAG.

Neural models have achieved great success on machine reading comprehension (MRC), many of which typically consist of two components: an evidence extractor and an answer predictor. The former seeks the most relevant information from a reference text, while the latter is to locate or generate answers from the extracted evidence. Despite the importance of evidence labels for training the evidence extractor, they are not cheaply accessible, particularly in many non-extractive MRC tasks such as YES/NO question answering and multi-choice MRC. To address this problem, we present a Self-Training method (STM), which supervises the evidence extractor with auto-generated evidence labels in an iterative process. At each iteration, a base MRC model is trained with golden answers and noisy evidence labels. The trained model will predict pseudo evidence labels as extra supervision in the next iteration. We evaluate STM on seven datasets over three MRC tasks. Experimental results demonstrate the improvement on existing MRC models, and we also analyze how and why such a self-training method works in MRC. The source code can be obtained from https://github.com/SparkJiao/Self-Training-MRC

Evidence to guide healthcare decisions is often limited by a lack of relevant and trustworthy literature as well as difficulty in contextualizing existing research for a specific patient. Large language models (LLMs) could potentially address both challenges by either summarizing published literature or generating new studies based on real-world data (RWD). We evaluated the ability of five LLM-based systems in answering 50 clinical questions and had nine independent physicians review the responses for relevance, reliability, and actionability. As it stands, general-purpose LLMs (ChatGPT-4, Claude 3 Opus, Gemini Pro 1.5) rarely produced answers that were deemed relevant and evidence-based (2% - 10%). In contrast, retrieval augmented generation (RAG)-based and agentic LLM systems produced relevant and evidence-based answers for 24% (OpenEvidence) to 58% (ChatRWD) of questions. Only the agentic ChatRWD was able to answer novel questions compared to other LLMs (65% vs. 0-9%). These results suggest that while general-purpose LLMs should not be used as-is, a purpose-built system for evidence summarization based on RAG and one for generating novel evidence working synergistically would improve availability of pertinent evidence for patient care.

To tackle the AVeriTeC shared task hosted by the FEVER-24, we introduce a system that only employs publicly available large language models (LLMs) for each step of automated fact-checking, dubbed the Herd of Open LLMs for verifying real-world claims (HerO). For evidence retrieval, a language model is used to enhance a query by generating hypothetical fact-checking documents. We prompt pretrained and fine-tuned LLMs for question generation and veracity prediction by crafting prompts with retrieved in-context samples. HerO achieved 2nd place on the leaderboard with the AVeriTeC score of 0.57, suggesting the potential of open LLMs for verifying real-world claims. For future research, we make our code publicly available at https://github.com/ssu-humane/HerO.

Medical systematic reviews can be very costly and resource intensive. We explore how Large Language Models (LLMs) can support and be trained to perform literature screening when provided with a detailed set of selection criteria. Specifically, we instruction tune LLaMA and Guanaco models to perform abstract screening for medical systematic reviews. Our best model, Bio-SIEVE, outperforms both ChatGPT and trained traditional approaches, and generalises better across medical domains. However, there remains the challenge of adapting the model to safety-first scenarios. We also explore the impact of multi-task training with Bio-SIEVE-Multi, including tasks such as PICO extraction and exclusion reasoning, but find that it is unable to match single-task Bio-SIEVE's performance. We see Bio-SIEVE as an important step towards specialising LLMs for the biomedical systematic review process and explore its future developmental opportunities. We release our models, code and a list of DOIs to reconstruct our dataset for reproducibility.

The rapid propagation of misinformation poses substantial risks to public interest. To combat misinformation, large language models (LLMs) are adapted to automatically verify claim credibility. Nevertheless, existing methods heavily rely on the embedded knowledge within LLMs and / or black-box APIs for evidence collection, leading to subpar performance with smaller LLMs or upon unreliable context. In this paper, we propose retrieval augmented fact verification through the synthesis of contrasting arguments (RAFTS). Upon input claims, RAFTS starts with evidence retrieval, where we design a retrieval pipeline to collect and re-rank relevant documents from verifiable sources. Then, RAFTS forms contrastive arguments (i.e., supporting or refuting) conditioned on the retrieved evidence. In addition, RAFTS leverages an embedding model to identify informative demonstrations, followed by in-context prompting to generate the prediction and explanation. Our method effectively retrieves relevant documents as evidence and evaluates arguments from varying perspectives, incorporating nuanced information for fine-grained decision-making. Combined with informative in-context examples as prior, RAFTS achieves significant improvements to supervised and LLM baselines without complex prompts. We demonstrate the effectiveness of our method through extensive experiments, where RAFTS can outperform GPT-based methods with a significantly smaller 7B LLM.

While recent Large Language Models (LLMs) have proven useful in answering user queries, they are prone to hallucination, and their responses often lack credibility due to missing references to reliable sources. An intuitive solution to these issues would be to include in-text citations referring to external documents as evidence. While previous works have directly prompted LLMs to generate in-text citations, their performances are far from satisfactory, especially when it comes to smaller LLMs. In this work, we propose an effective training framework using fine-grained rewards to teach LLMs to generate highly supportive and relevant citations, while ensuring the correctness of their responses. We also conduct a systematic analysis of applying these fine-grained rewards to common LLM training strategies, demonstrating its advantage over conventional practices. We conduct extensive experiments on Question Answering (QA) datasets taken from the ALCE benchmark and validate the model's generalizability using EXPERTQA. On LLaMA-2-7B, the incorporation of fine-grained rewards achieves the best performance among the baselines, even surpassing that of GPT-3.5-turbo.

This paper introduces PRobELM (Plausibility Ranking Evaluation for Language Models), a benchmark designed to assess language models' ability to discern more plausible from less plausible scenarios through their parametric knowledge. While benchmarks such as TruthfulQA emphasise factual accuracy or truthfulness, and others such as COPA explore plausible scenarios without explicitly incorporating world knowledge, PRobELM seeks to bridge this gap by evaluating models' capabilities to prioritise plausible scenarios that leverage world knowledge over less plausible alternatives. This design allows us to assess the potential of language models for downstream use cases such as literature-based discovery where the focus is on identifying information that is likely but not yet known. Our benchmark is constructed from a dataset curated from Wikidata edit histories, tailored to align the temporal bounds of the training data for the evaluated models. PRobELM facilitates the evaluation of language models across multiple prompting types, including statement, text completion, and question-answering. Experiments with 10 models of various sizes and architectures on the relationship between model scales, training recency, and plausibility performance, reveal that factual accuracy does not directly correlate with plausibility performance and that up-to-date training data enhances plausibility assessment across different model architectures.

Evidence-grounded reasoning requires more than attaching retrieved text to a prediction: a model should make decisions that depend on whether the provided evidence supports the target claim. In practice, this often fails because supervision is weak, evidence is only loosely tied to the claim, and evaluation does not test evidence dependence directly. We introduce case-grounded evidence verification, a general framework in which a model receives a local case context, external evidence, and a structured claim, and must decide whether the evidence supports the claim for that case. Our key contribution is a supervision construction procedure that generates explicit support examples together with semantically controlled non-support examples, including counterfactual wrong-state and topic-related negatives, without manual evidence annotation. We instantiate the framework in radiology and train a standard verifier on the resulting support task. The learned verifier substantially outperforms both case-only and evidence-only baselines, remains strong under correct evidence, and collapses when evidence is removed or swapped, indicating genuine evidence dependence. This behavior transfers across unseen evidence articles and an external case distribution, though performance degrades under evidence-source shift and remains sensitive to backbone choice. Overall, the results suggest that a major bottleneck in evidence grounding is not only model capacity, but the lack of supervision that encodes the causal role of evidence.

Across medical specialties, clinical practice is anchored in evidence-based guidelines that codify best studied diagnostic and treatment pathways. These pathways routinely fall short for the long tail of real-world care not covered by guidelines. Most medical large language models (LLMs), however, are trained to encode common, guideline-focused medical knowledge in their parameters. Current evaluations test models primarily on recalling and reasoning with this memorized content, often in multiple-choice settings. Given the fundamental importance of evidence-based reasoning in medicine, it is neither feasible nor reliable to depend on memorization in practice. To address this gap, we introduce OGCaReBench, a free-form retrieval-focused benchmark aimed at evaluating LLMs at answering clinical questions that require going beyond typical guidelines. Extracted from published medical case reports and validated by medical experts, OGCaReBench contains long-form clinical questions requiring free-text answers, providing a systematic framework for assessing open-ended medical reasoning in rare, case-based scenarios. Our experiments reveal that even the best-performing baseline (GPT-5.2) correctly answers only 56% of our benchmark with specialized models only reaching 42%. Augmenting models with retrieved medical articles improves this performance to up to 82% (using GPT-5.2) highlighting the importance of evidence-grounding for real-world medical reasoning tasks. This work thus establishes a foundation for benchmarking and advancing both general-purpose and medical LLMs to produce reliable answers in challenging clinical contexts.

Finding relevant literature underpins the practice of evidence-based medicine. From 2014 to 2016, TREC conducted a clinical decision support track, wherein participants were tasked with finding articles relevant to clinical questions posed by physicians. In total, 87 teams have participated over the past three years, generating 395 runs. During this period, each team has trialled a variety of methods. While there was significant overlap in the methods employed by different teams, the results were varied. Due to the diversity of the platforms used, the results arising from the different techniques are not directly comparable, reducing the ability to build on previous work. By using a stable platform, we have been able to compare different document and query processing techniques, allowing us to experiment with different search parameters. We have used our system to reproduce leading teams runs, and compare the results obtained. By benchmarking our indexing and search techniques, we can statistically test a variety of hypotheses, paving the way for further research.