Daily Papers

The rapid development of Multimodal Large Language Models (MLLMs) has led to growing interest in egocentric video understanding, specifically the ability for MLLMs to recognize fine-grained hand-object interactions, track object state changes over time, and reason about manipulative processes in dynamic environments from a first-person perspective. However, existing egocentric video benchmarks suffer from limited grounded rationale evaluation, offering limited support for fine-grained operation-centric reasoning and rarely examining whether model rationales are grounded in explicit spatio-temporal evidence. To address this gap, we introduce EgoCoT-Bench, a fine-grained egocentric benchmark for grounded and verifiable operation-centric reasoning with explicit step-by-step rationale annotations. Overall, EgoCoT-Bench comprises 3,172 verifiable QA pairs over 351 egocentric videos separated into four task groups for a total of 12 sub-task groups, encompassing perception and retrospection, anticipation, and high-level reasoning. The benchmark is constructed through a spatio-temporal scene graphs (STSG) guided generation framework and is further refined by human annotators to ensure correctness, egocentric relevance and fine-grained quality. Experimental results show continuing difficulties with egocentric fine-grained reasoning and further reveal that many multimodal models produce explanations that are answer-correct, but have evidence that is inconsistent with the answer. We hope EgoCoT-Bench can serve as a useful testbed for grounded and verifiable reasoning in egocentric video understanding. Project page and supplementary materials are available at: https://dstardust.github.io/EgoCoT/.

In embodied AI, safety risk can be latent: a benign instruction and a safe scene become hazardous only when composed. Prior work has advanced embodied safety by varying visual contexts or evaluating execution-time dynamics, but the complementary axis of fixing the scene and varying only the instruction remains underexplored. We introduce GuardianBench, an instruction-contrastive benchmark grounded in international safety standards that isolates this latent contextual risk through 3,024 instruction-scene examples organized as same-scene Safe/Unsafe contrastive pairs across various hazard categories. Benchmarking state-of-the-art vision-language models (VLMs) reveals instruction-insensitive verdicts: models disproportionately approve both instructions under a given scene; across the primary models, average pair accuracy is only 24.1%. Our systematic rationale audit localizes the dominant failure: models fail to bind the instruction-relevant cues that differentiate safe from unsafe compositions. As a post-training case study, Verdict Log-Odds Supervision (VLOS), a lightweight verdict-level objective, substantially improves performance on open-weight backbones. Together, our latent contextual risk task formulation, standards-grounded contrastive benchmark construction, pair-level and rationale-level failure diagnosis, and benchmark-enabled verdict calibration establish GuardianBench as a controlled evaluation suite for exposing and improving safety reasoning over instruction-scene compositions under latent contextual risk.

Large language models (LLMs) are increasingly applied in scientific research, offering new capabilities for knowledge discovery and reasoning. In single-cell biology, however, evaluation practices for both general and specialized LLMs remain inadequate: existing benchmarks are fragmented across tasks, adopt formats such as multiple-choice classification that diverge from real-world usage, and rely on metrics lacking interpretability and biological grounding. We present SC-ARENA, a natural language evaluation framework tailored to single-cell foundation models. SC-ARENA formalizes a virtual cell abstraction that unifies evaluation targets by representing both intrinsic attributes and gene-level interactions. Within this paradigm, we define five natural language tasks (cell type annotation, captioning, generation, perturbation prediction, and scientific QA) that probe core reasoning capabilities in cellular biology. To overcome the limitations of brittle string-matching metrics, we introduce knowledge-augmented evaluation, which incorporates external ontologies, marker databases, and scientific literature to support biologically faithful and interpretable judgments. Experiments and analysis across both general-purpose and domain-specialized LLMs demonstrate that (i) under the Virtual Cell unified evaluation paradigm, current models achieve uneven performance on biologically complex tasks, particularly those demanding mechanistic or causal understanding; and (ii) our knowledge-augmented evaluation framework ensures biological correctness, provides interpretable, evidence-grounded rationales, and achieves high discriminative capacity, overcoming the brittleness and opacity of conventional metrics. SC-Arena thus provides a unified and interpretable framework for assessing LLMs in single-cell biology, pointing toward the development of biology-aligned, generalizable foundation models.

Prompt design is a primary control interface for large language models (LLMs), yet standard evaluations largely reduce performance to answer correctness, obscuring why a prompt succeeds or fails and providing little actionable guidance. We propose PEEM (Prompt Engineering Evaluation Metrics), a unified framework for joint and interpretable evaluation of both prompts and responses. PEEM defines a structured rubric with 9 axes: 3 prompt criteria (clarity/structure, linguistic quality, fairness) and 6 response criteria (accuracy, coherence, relevance, objectivity, clarity, conciseness), and uses an LLM-based evaluator to output (i) scalar scores on a 1-5 Likert scale and (ii) criterion-specific natural-language rationales grounded in the rubric. Across 7 benchmarks and 5 task models, PEEM's accuracy axis strongly aligns with conventional accuracy while preserving model rankings (aggregate Spearman rho about 0.97, Pearson r about 0.94, p < 0.001). A multi-evaluator study with four models shows consistent relative judgments (pairwise rho = 0.68-0.85), supporting evaluator-agnostic deployment. Beyond alignment, PEEM captures complementary linguistic failure modes and remains informative under prompt perturbations: prompt-quality trends track downstream accuracy under iterative rewrites, semantic adversarial manipulations induce clear score degradation, and meaning-preserving paraphrases yield high stability (robustness rate about 76.7-80.6%). Finally, using only PEEM scores and rationales as feedback, a zero-shot prompt rewriting loop improves downstream accuracy by up to 11.7 points, outperforming supervised and RL-based prompt-optimization baselines. Overall, PEEM provides a reproducible, criterion-driven protocol that links prompt formulation to response behavior and enables systematic diagnosis and optimization of LLM interactions.

Evaluating generated videos remains challenging because existing benchmarks rely on fixed evaluation content, cover only a subset of generation and editing settings, and provide limited evidence for their scores. We introduce VideoArgus, a unified rubric-grounded framework covering five video generation and editing settings. For each input instance, VideoArgus generates an output-blind, sample-specific rubric once and reuses it to evaluate all corresponding candidate videos. The rubric defines concrete criteria, scoring rules, failure modes, and evidence plans, which guide criterion-specific VLM QA and visual tools to produce evidence-grounded criterion scores, rationales, and a diagnostic report. We further construct VideoArgus-Bench, containing 1,026 curated input instances built from 653 high-quality images and 416 high-quality videos, with all benchmark rubrics pre-generated, frozen, and released. On a separate 1,260-video human-alignment set, VideoArgus achieves higher within-input Spearman and Kendall correlations with human judgments than the corresponding benchmark-specific evaluators across all five tasks. Model rankings also remain largely consistent across different rubric-generation and evaluation-VLM backbones. All code and data are released. Visit our project page: https://zzzmyyzeng.github.io/VideoArgus

Protein design aims to compose amino-acid sequences that fold into stable three-dimensional structures while satisfying targeted functional properties. The field is increasingly shifting toward vibe protein design, where a single model is expected to generate novel sequences, engineer existing proteins, and reason about protein characteristics through flexible natural-language constraints. Large language models (LLMs) have emerged as a leading paradigm in this space. However, existing evaluation benchmarks often limit their scope to a partial aspect of protein design, while others restrict design objectives to structured input schemas, lacking an integrated framework that evaluates the broad spectrum of protein design competence under open-ended intents. To this end, we present Vibe Protein design Benchmark (VibeProteinBench), a language-interfaced benchmark that probes generalist capabilities through three complementary stages mirroring a computational protein design workflow: recognition, engineering, and generation. Each stage is grounded in expert-curated mechanistic rationales and multi-faceted in silico validation, to computationally verify whether model outputs are biologically plausible. Evaluations across diverse general-purpose and domain-specialized LLMs reveal that no model achieves strong performance across all three stages, suggesting that generalist protein design remains a substantial open challenge for current LLMs.

The clinical translation of dermatological AI is hindered by opaque reasoning and systematic performance disparities across skin tones. Here we present SkinGPT-R1, a multimodal large language model that integrates chain-of-thought diagnostic reasoning with a fairness-aware mixture-of-experts architecture for interpretable and equitable skin disease diagnosis. Through parameter-efficient adaptation of a frozen reasoning backbone, SkinGPT-R1 generates structured diagnostic reports comprising visual findings, differential reasoning, and final diagnosis. Across seven external datasets spanning diverse pathologies and imaging conditions, SkinGPT-R1 achieves state-of-the-art accuracy on six benchmarks, including 82.50\% on a challenging 40-class long-tail classification task (+19.30\% over leading baselines). Blinded evaluation by five board-certified dermatologists on 1,000 phenotypically balanced cases yields a mean score of 3.6 out of 5, with the highest ratings in safety (3.8) and reasoning coherence (3.6), indicating that the generated rationales are clinically safe, logically grounded, and suitable for supporting diagnostic decision-making. Critically, SkinGPT-R1 mitigates algorithmic bias across the full Fitzpatrick spectrum, achieving a robust worst-group performance of 41.40\% on the Fitz17k benchmark and a five-fold relative improvement in lower-bound accuracy on the DDI dataset compared to standard multimodal baselines. These results establish a framework for trustworthy, fair, and explainable AI-assisted dermatological diagnosis.

Text-rich image understanding requires multimodal large language models (MLLMs) to organize OCR (Optical Character Recognition)-grounded evidence across words, layout, fields, charts, and visual correspondences. Existing evaluations often conflate extraction with reasoning and rarely test whether models follow the required reasoning direction: applying visible rules, abstracting hidden regularities, or recovering missing premises. We introduce OCR-MetaReasoning, a controlled single-image benchmark that treats deduction, induction, and abduction as distinct directions and separates final-answer correctness from reasoning-process compliance. The benchmark contains 1,500 verified samples in a balanced \(3\times5\) taxonomy crossing three reasoning types with five OCR-object categories, along with reference reasoning steps, automatic answer scoring, the Meta-Reasoning Macro Score (MRMS), and the Reasoning Process Compliance Score (RPCS). Experiments with representative closed-source and open-source MLLMs show that OCR-grounded meta-reasoning remains far from saturated: models struggle with visible-rule application and layout-sensitive inference, while process-compliant rationales can accompany incorrect final answers under exact-match evaluation. The code is available at https://github.com/gengxuli/OCR-MetaReasoning.

Step-by-step reasoning is widely used to enhance the reasoning ability of large language models (LLMs) in complex problems. Evaluating the quality of reasoning traces is crucial for understanding and improving LLM reasoning. However, the evaluation criteria remain highly unstandardized, leading to fragmented efforts in developing metrics and meta-evaluation benchmarks. To address this gap, this survey provides a comprehensive overview of step-by-step reasoning evaluation, proposing a taxonomy of evaluation criteria with four top-level categories (groundedness, validity, coherence, and utility). We then categorize metrics based on their implementations, survey which metrics are used for assessing each criterion, and explore whether evaluator models can transfer across different criteria. Finally, we identify key directions for future research.

Eliciting feedback from end users of NLP models can be beneficial for improving models. However, how should we present model responses to users so they are most amenable to be corrected from user feedback? Further, what properties do users value to understand and trust responses? We answer these questions by analyzing the effect of rationales (or explanations) generated by QA models to support their answers. We specifically consider decomposed QA models that first extract an intermediate rationale based on a context and a question and then use solely this rationale to answer the question. A rationale outlines the approach followed by the model to answer the question. Our work considers various formats of these rationales that vary according to well-defined properties of interest. We sample rationales from language models using few-shot prompting for two datasets, and then perform two user studies. First, we present users with incorrect answers and corresponding rationales in various formats and ask them to provide natural language feedback to revise the rationale. We then measure the effectiveness of this feedback in patching these rationales through in-context learning. The second study evaluates how well different rationale formats enable users to understand and trust model answers, when they are correct. We find that rationale formats significantly affect how easy it is (1) for users to give feedback for rationales, and (2) for models to subsequently execute this feedback. In addition, formats with attributions to the context and in-depth reasoning significantly enhance user-reported understanding and trust of model outputs.

Is explainability a false promise? This debate has emerged from the insufficient evidence that explanations aid people in situations they are introduced for. More human-centered, application-grounded evaluations of explanations are needed to settle this. Yet, with no established guidelines for such studies in NLP, researchers accustomed to standardized proxy evaluations must discover appropriate measurements, tasks, datasets, and sensible models for human-AI teams in their studies. To help with this, we first review fitting existing metrics. We then establish requirements for datasets to be suitable for application-grounded evaluations. Among over 50 datasets available for explainability research in NLP, we find that 4 meet our criteria. By finetuning Flan-T5-3B, we demonstrate the importance of reassessing the state of the art to form and study human-AI teams. Finally, we present the exemplar studies of human-AI decision-making for one of the identified suitable tasks -- verifying the correctness of a legal claim given a contract.

Radiology reporting is a complex task that requires detailed image understanding, integration of multiple inputs, including comparison with prior imaging, and precise language generation. This makes it ideal for the development and use of generative multimodal models. Here, we extend report generation to include the localisation of individual findings on the image - a task we call grounded report generation. Prior work indicates that grounding is important for clarifying image understanding and interpreting AI-generated text. Therefore, grounded reporting stands to improve the utility and transparency of automated report drafting. To enable evaluation of grounded reporting, we propose a novel evaluation framework - RadFact - leveraging the reasoning capabilities of large language models (LLMs). RadFact assesses the factuality of individual generated sentences, as well as correctness of generated spatial localisations when present. We introduce MAIRA-2, a large multimodal model combining a radiology-specific image encoder with a LLM, and trained for the new task of grounded report generation on chest X-rays. MAIRA-2 uses more comprehensive inputs than explored previously: the current frontal image, the current lateral image, the prior frontal image and prior report, as well as the Indication, Technique and Comparison sections of the current report. We demonstrate that these additions significantly improve report quality and reduce hallucinations, establishing a new state of the art on findings generation (without grounding) on MIMIC-CXR while demonstrating the feasibility of grounded reporting as a novel and richer task.

Multimodal Large Language Models (MLLMs) are increasingly deployed in human-facing roles where personality perception is critical, yet existing benchmarks evaluate this capability solely on numerical Big Five score prediction, leaving open whether models truly perceive personality through behavioral understanding or merely prejudge through superficial pattern matching. We address this gap with three contributions. (i) A new task: we formalize Grounded Personality Reasoning (GPR), which requires MLLMs to anchor each Big Five rating in observable evidence through a chain of rating, reasoning, and grounding. (ii) A new dataset: we release MM-OCEAN (1,104 videos, 5,320 MCQs), produced by a multi-agent pipeline with human verification, with timestamped behavioral observations, evidence-grounded trait analyses, and seven categories of cue-grounding MCQs. (iii) Benchmark and analysis: we design a three-tier evaluation (rating, reasoning, grounding) plus four sample-level failure-mode metrics: Prejudice Rate (PR), Confabulation Rate (CR), Integration-failure Rate (IR), and Holistic-grounding Rate (HR), and benchmark 27 MLLMs (13 closed, 14 open). The analysis uncovers a striking Prejudice Gap: across the field, 51% of correct ratings are not grounded in retrieved cues, and the Holistic-Grounding Rate spans only 0-33.5%. These findings expose a disconnect between getting the right score and reasoning for the right reason, charting a roadmap for grounded social cognition in MLLMs.

Before letting an agent operate over real context, can you prove it used the right evidence? GroundEval turns that question into a deterministic test of what the agent searched, fetched, cited, and was permitted to access. In one case study, two frontier LLM judges scored a plausible agent response above 0.85. But the trace told a different story: the agent had never retrieved the artifact its answer depended on, yielding a GroundEval score of 0.000. We introduce GroundEval, a judge-free framework for evaluating agents against grounded, time-bounded, and access-controlled evidence. GroundEval uses a domain configuration to generate questions, lets the agent choose how to answer, and then scores both the final answer and the recorded trajectory that produced it. The benchmark targets three failures that LLM-as-judge evaluation struggles to detect: whether an agent checked before claiming absence, reasoned only from evidence available to the actor at the relevant time, and used the correct causal mechanism rather than a plausible one. These correspond to three tracks: Silence, Perspective, and Counterfactual. GroundEval exposes when plausible answers rest on invalid evidence paths, and produces structured per-question diagnostics that pair tool activity with the agent's turn-level narration, making each score inspectable rather than merely reported. What our case studies turned up is that this gap isn't some rare corner case. It's exactly the blind spot that final-answer and judge-based scoring were never built to catch.

Despite strong performance of Multimodal Large Language Models (MLLMs) on multimodal tasks, predicting whether and why an image is persuasive remains challenging. We first show that prompting MLLMs to reason before prediction does not consistently help, and can even reduce persuasiveness prediction performance, suggesting that naively generated rationales are unreliable signals for this task. Yet, no established methodology exists for training MLLMs to reason about visual persuasion or evaluating whether their rationales faithfully support their decisions. To address this gap, we show empirically and theoretically that diverse teacher-generated rationales, when used for supervised fine-tuning, improve visual persuasiveness prediction. We further introduce a three-dimensional faithfulness evaluation framework covering rationale-to-decision consistency, rationale-to-image groundedness, and rationale-to-decision sensitivity. Applying this framework shows that prediction performance alone does not guarantee faithful rationales, while rationale-to-decision sensitivity is most aligned with human rationale preferences. These findings motivate faithfulness-aware training objectives and scalable rationale supervision for visual persuasiveness evaluation. Our code and dataset will be made publicly available.

Generative reward models (GenRMs) generate natural language rationales to justify preference judgments, yet a correct verdict does not guarantee a sound justification. Existing approaches often rely on reference explanations to supervise rationales, but obtaining such references can be costly and does not necessarily establish whether generated rationales support the resulting verdict. To address these limitations, we propose DeCert, a framework that evaluates the validity and verdict consistency of generated rationales without requiring reference explanations. Crucially, DeCert translates these assessments into preference supervision that accounts for both verdict correctness and justification quality. In particular, when multiple sampled judgments share the same correct verdict, differences in certificate quality enable DeCert to construct rationale preference pairs that distinguish stronger from weaker justifications. Extensive empirical studies across two model families and multiple in- and out-of-distribution benchmarks demonstrate that DeCert consistently outperforms strong generative judge baselines while improving rationale quality. Moreover, DeCert effectively mitigates positional bias, improving position-consistent accuracy by up to 14.8 points. These results highlight the importance of supervising not only which verdict is correct, but also how that verdict is justified.

Large language models have achieved remarkable progress on complex reasoning tasks. However, they often implicitly fabricate information when inputs are incomplete, producing confident but unreliable conclusions -- a failure mode we term ungrounded reasoning. We argue that this issue arises not from insufficient reasoning capability, but from the lack of inferential boundary awareness -- the ability to recognize when the necessary premises for valid inference are missing. To address this issue, we propose Grounded Reasoning via Interactive Reinforcement Learning (GRIL), a multi-turn reinforcement learning framework for grounded reasoning under incomplete information. GRIL decomposes the reasoning process into two stages: clarify and pause, which identifies whether the available information is sufficient, and grounded reasoning, which performs task solving once the necessary premises are established. We design stage-specific rewards to penalize hallucinations, enabling models to detect gaps, stop proactively, and resume reasoning after clarification. Experiments on GSM8K-Insufficient and MetaMATH-Insufficient show that GRIL significantly improves premise detection (up to 45%), leading to a 30% increase in task success while reducing average response length by over 20%. Additional analyses confirm robustness to noisy user responses and generalization to out-of-distribution tasks.

Large language models are increasingly deployed as agents that reason over documents rather than answer from parametric knowledge. We study archive-grounded reasoning: locating sparse evidence across a large, messy collection of workplace files, reconciling inconsistent terminology, units, and time conventions, and computing an answer. Existing benchmarks address only parts of this setting and none jointly stresses archive-groundedness, agentic exploration, and cross-domain coverage. We introduce Agora, a benchmark pairing 362 questions with eight domain collections of 9,664 authentic documents and 372M tokens, far exceeding any model's context window, so agents must explore deliberately rather than scan exhaustively. Agora is built by an agentic pipeline combining cross-document task synthesis, leakage-preventing obfuscation, and difficulty filtering. Evaluating eight models, we find the task far from solved: even the strongest reaches only 59.4% accuracy, with notable variation across domains.

Large language models (LLMs) still produce plausible-sounding but ungrounded factual claims, a problem that worsens in multi-turn dialogue as context grows and early errors cascade. We introduce HalluHard, a challenging multi-turn hallucination benchmark with 950 seed questions spanning four high-stakes domains: legal cases, research questions, medical guidelines, and coding. We operationalize groundedness by requiring inline citations for factual assertions. To support reliable evaluation in open-ended settings, we propose a judging pipeline that iteratively retrieves evidence via web search. It can fetch, filter, and parse full-text sources (including PDFs) to assess whether cited material actually supports the generated content. Across a diverse set of frontier proprietary and open-weight models, hallucinations remain substantial even with web search (approx 30% for the strongest configuration, Opus-4.5 with web search), with content-grounding errors persisting at high rates. Finally, we show that hallucination behavior is shaped by model capacity, turn position, effective reasoning, and the type of knowledge required.

Self-training approach for large language models (LLMs) improves reasoning abilities by training the models on their self-generated rationales. Previous approaches have labeled rationales that produce correct answers for a given question as appropriate for training. However, a single measure risks misjudging rationale quality, leading the models to learn flawed reasoning patterns. To address this issue, we propose CREST (Consistency-driven Rationale Evaluation for Self-Training), a self-training framework that further evaluates each rationale through follow-up questions and leverages this evaluation to guide its training. Specifically, we introduce two methods: (1) filtering out rationales that frequently result in incorrect answers on follow-up questions and (2) preference learning based on mixed preferences from rationale evaluation results of both original and follow-up questions. Experiments on three question-answering datasets using open LLMs show that CREST not only improves the logical robustness and correctness of rationales but also improves reasoning abilities compared to previous self-training approaches.

Effective agency in social environments depends on when an agent seeks knowledge, when it acts, and whether its actions are justified by acquired information. Existing grounded benchmarks provide executable actions, persistent state, and verifiable outcomes, while social simulation environments provide rich interaction among language agents. We study an evaluation setting that combines these requirements. We define socially distributed task environments as interactive environments where task-relevant knowledge is partitioned across role-isolated participants and consequential actions are accessible only through them. Communication serves as exploration over role-partitioned knowledge, while grounded action serves as exploitation over environment state. We introduce Incognita, a Concordia-based framework that separates social interaction from grounded execution. The evaluated agent routes messages to a user or specialist entities; specialists mediate admissible operations; a deterministic sub-environment executes accepted operations over a canonical state; and an offline evaluator scores outcomes with inherited rewards. Incognita-Retail transforms tau-bench retail into a multi-entity environment while preserving final-state reward semantics. We evaluate three generative agent models on 18 tasks stratified by social breadth, with 540 trials. Progress appears in reward and behavior: success rises from 0 percent to 8.9 percent and 17.2 percent, while premature finalization falls from 100 percent to 87 percent and 58 percent. Stronger models elicit more hidden knowledge, contact more entities, and attempt more grounded writes, yet reliability remains low. These findings show that socially distributed task environments expose behavior before reliable success, including knowledge elicitation, source selection, grounded action attempts, and premature completion belief.

Recent advances in Video Large Language Models (Video-LLMs) have yielded promising performance on video question answering (VideoQA). Nevertheless, existing benchmarks are predominantly evaluated through answer correctness, while the grounding of predictions in relevant video evidence remains largely unexamined. This disconnect between answer generation and evidence understanding motivates the construction of the Evidence-Grounded Video Question Answering Benchmark (EG-VQA), an open-ended evaluation protocol in which each QA pair is explicitly annotated with supporting temporal evidence, thereby requiring joint reasoning and precise evidence localization. EG-VQA is comprised of 2,067 videos and 11,838 QA pairs with fine-grained evidence annotations. To evaluate predicted evidence, Evidence-Grounded F1 (EG-F1) is introduced as a unified metric in which temporal alignment and semantic consistency against ground-truth evidence are jointly measured. Experimental evaluation reveals that even strong proprietary models struggle to accurately ground their predictions, exposing a fundamental discrepancy between answer correctness and faithful evidence localization. To bridge this gap, EG-Reasoner, an evidence-grounded reasoning model trained with explicit supervision, is proposed. State-of-the-art performance is achieved among open-source models, with results competitive against proprietary systems, particularly pronounced gains are observed on reasoning-intensive tasks such as counterfactual questions. These findings demonstrate that scaling alone is insufficient for robust video understanding and that structured evidence supervision is essential for the development of more reliable and interpretable VideoQA systems.

Evaluation of language model outputs on structured writing tasks is typically conducted with a number of desirable criteria presented to human evaluators or large language models (LLMs). For instance, on a prompt like "Help me draft an academic talk on coffee intake vs research productivity", a model response may be evaluated for criteria like accuracy and coherence. However, high-quality responses should do more than just satisfy basic task requirements. An effective response to this query should include quintessential features of an academic talk, such as a compelling opening, clear research questions, and a takeaway. To help identify these implicit criteria, we introduce EvalAgent, a novel framework designed to automatically uncover nuanced and task-specific criteria. EvalAgent first mines expert-authored online guidance. It then uses this evidence to propose diverse, long-tail evaluation criteria that are grounded in reliable external sources. Our experiments demonstrate that the grounded criteria produced by EvalAgent are often implicit (not directly stated in the user's prompt), yet specific (high degree of lexical precision). Further, EvalAgent criteria are often not satisfied by initial responses but they are actionable, such that responses can be refined to satisfy them. Finally, we show that combining LLM-generated and EvalAgent criteria uncovers more human-valued criteria than using LLMs alone.

We introduce GODEL (Grounded Open Dialogue Language Model), a large pre-trained language model for dialog. In contrast with earlier models such as DialoGPT, GODEL leverages a new phase of grounded pre-training designed to better support adapting GODEL to a wide range of downstream dialog tasks that require information external to the current conversation (e.g., a database or document) to produce good responses. Experiments against an array of benchmarks that encompass task-oriented dialog, conversational QA, and grounded open-domain dialog show that GODEL outperforms state-of-the-art pre-trained dialog models in few-shot fine-tuning setups, in terms of both human and automatic evaluation. A novel feature of our evaluation methodology is the introduction of a notion of utility that assesses the usefulness of responses (extrinsic evaluation) in addition to their communicative features (intrinsic evaluation). We show that extrinsic evaluation offers improved inter-annotator agreement and correlation with automated metrics. Code and data processing scripts are publicly available.

As large language models (LLMs) grow larger and more sophisticated, assessing their "reasoning" capabilities in natural language grows more challenging. Recent question answering (QA) benchmarks that attempt to assess reasoning are often limited by a narrow scope of covered situations and subject matters. We introduce WikiWhy, a QA dataset built around a novel auxiliary task: explaining why an answer is true in natural language. WikiWhy contains over 9,000 "why" question-answer-rationale triples, grounded on Wikipedia facts across a diverse set of topics. Each rationale is a set of supporting statements connecting the question to the answer. WikiWhy serves as a benchmark for the reasoning capabilities of LLMs because it demands rigorous explicit rationales for each answer to demonstrate the acquisition of implicit commonsense knowledge, which is unlikely to be easily memorized. GPT-3 baselines achieve only 38.7% human-evaluated correctness in the end-to-end answer & explain condition, leaving significant room for future improvements.

LLMs and agentic systems are increasingly deployed in social environments, making normative competence critical for safe and appropriate behavior. However, existing approaches either assess normative judgment in text alone or reduce it to choosing among a fixed set of candidate actions. We argue both are insufficient. In practice, agents are never handed a menu of options; they must identify a reasonable action from scratch, grounded in visible facts and supported by inspectable reasons. We introduce NoRA, a visual first-person video benchmark that requires models to generate candidate next actions and justify each through an explicit fact-reason-action support graph. The benchmark comprises 1,420 annotated video clips, including HumanGold-190 and LLMSilver-1230 splits. Each instance is evaluated through action alignment, factual grounding, and support binding, aggregated into a single grounded reasonableness score. We benchmark 12 multimodal systems under direct, deliberate, and structured prompting regimes, finding that current VLMs frequently recover plausible actions and relevant scene facts, but consistently struggle to construct the full reasonable action space and bind selected actions to the correct local support. NoRA makes this gap measurable, shifting the evaluation question from whether a model can pick an action to whether it can justify an appropriate action for the right visible reasons.

We conduct a systematic audit of three widely used reasoning benchmarks, SocialIQa, FauxPas-EAI, and ToMi, and uncover pervasive flaws in both benchmark items and evaluation methodology. Using five LLMs (GPT-{3, 3.5, 4, o1}, and LLaMA 3.1) as diagnostic tools, we identify structural, semantic, and pragmatic issues in benchmark design (e.g., duplicated items, ambiguous wording, and implausible answers), as well as scoring procedures that prioritize output form over reasoning process. Through systematic human annotation and re-evaluation on cleaned benchmark subsets, we find that model scores often improve not due to due to erratic surface wording variations and not to improved reasoning. Infact, further analyses show that model performance is highly sensitive to minor input variations such as context availability and phrasing, revealing that high scores may reflect alignment with format-specific cues rather than consistent inference based on the input. These findings challenge the validity of current benchmark-based claims about reasoning in LLMs, and highlight the need for evaluation protocols that assess reasoning as a process of drawing inference from available information, rather than as static output selection. We release audited data and evaluation tools to support more interpretable and diagnostic assessments of model reasoning.

Existing research largely reduces cultural intelligence in LLMs to a knowledge-level problem, overlooking whether models can effectively utilize their acquired knowledge in realistic scenarios. To bridge this gap, we introduce CultureForest, a benchmark for Cultural Norm Grounded Reasoning. Each question is grounded in a small set of atomic norms, enabling verifiable and attributable evaluation. CultureForest comprises 5,378 examples across 8 domains and 53 countries/regions, and supports a progressive evaluation from multiple-choice to open-ended generation. Extensive experiments reveal that even top-tier models degrade substantially in open-ended settings, accompanied by pronounced cross-region disparities. Through targeted analysis, we uncover several consistent patterns: (1) test-time reasoning yields limited gains and may exacerbate inequity; (2) models exhibit highly shared regional preference structures; (3) model responses are markedly conservative, especially under stricter cultural constraints; and (4) by disentangling cultural knowledge acquisition from cultural reasoning, we show that while LLMs possess substantial cultural knowledge, their performance is further bottlenecked by its effective use. These findings point to a necessary shift from knowledge-centric evaluation toward measuring knowledge-grounded reasoning.

A research idea may be novel, coherent, and scientifically plausible, yet its proposed method may remain insufficiently specified for faithful implementation. We study the codification readiness of implementation-facing research-method specifications, defined by whether they provide sufficient methodological information for a competent implementer or coding agent to construct the intended method without unsupported assumptions. We construct evidence-grounded specifications and their supported resolutions from papers, codebases, issue threads, and reproduction artifacts. We introduce IdeaAMBIG, a benchmark of 660 evidence-grounded instances: 163 real-world gaps from reproducibility reports and GitHub issues, and 497 controlled synthetic gaps injected into codification-ready references. IdeaAMBIG evaluates three capabilities: codification-readiness assessment, defect localization, and clarification action generation. Defect localization receives only the specification, whereas clarification additionally receives the annotated defect. Across 13 LLMs, the best model achieves 9.6% Macro Defect Recovery Rate on real-world instances but 80.6% Macro Clarification Action Success Rate when given the defect. In an oracle study, supplying the gold resolution raises the downstream codification-ready rate from 14% to 98%. Across all evaluated models, defect localization is the main bottleneck, with stronger clarification given the defect.

Recent studies have demonstrated the efficacy of using Reinforcement Learning (RL) in building reasoning models that articulate chains of thoughts prior to producing final answers. However, despite ongoing advances that aim at enabling reasoning for vision-language tasks, existing open-source visual reasoning models typically generate reasoning content with pure natural language, lacking explicit integration of visual information. This limits their ability to produce clearly articulated and visually grounded reasoning chains. To this end, we propose Grounded Reasoning with Images and Texts (GRIT), a novel method for training MLLMs to think with images. GRIT introduces a grounded reasoning paradigm, in which models generate reasoning chains that interleave natural language and explicit bounding box coordinates. These coordinates point to regions of the input image that the model consults during its reasoning process. Additionally, GRIT is equipped with a reinforcement learning approach, GRPO-GR, built upon the GRPO algorithm. GRPO-GR employs robust rewards focused on the final answer accuracy and format of the grounded reasoning output, which eliminates the need for data with reasoning chain annotations or explicit bounding box labels. As a result, GRIT achieves exceptional data efficiency, requiring as few as 20 image-question-answer triplets from existing datasets. Comprehensive evaluations demonstrate that GRIT effectively trains MLLMs to produce coherent and visually grounded reasoning chains, showing a successful unification of reasoning and grounding abilities.

Large Vision-Language Models (LVLMs) have demonstrated impressive general capabilities across a wide range of multi-modal tasks. However, the reasoning processes of LVLMs often suffer from unreliable outputs and limited interpretability. To address this, grounded visual reasoning has emerged as a promising paradigm that enforces responses anchored on salient visual evidence regions. However, existing approaches typically rely on costly supervision such as bounding box annotations, chain-of-thought rationale or external tool calls, limiting their scalability. In this work, we propose Ground-R1, a reinforcement learning framework that enables grounded visual reasoning without requiring explicit evidence or rationale annotations. Ground-R1 consists of a grounding phase that generates evidence region rollouts based on format constraints, and an answering phase that produces responses guided by both answer correctness and format adherence rewards. Extensive experiments across multiple visual reasoning benchmarks manifest that Ground-R1 achieves superior performance and exhibits emergent cognitive behaviors such as uncertainty awareness, spatial perception, and iterative refinement, offering a scalable and interpretable alternative to existing approaches.

As LLM benchmarks saturate, the evaluation community has pursued two strategies to increase difficulty: escalating knowledge demands (GPQA, HLE) or removing knowledge entirely in favor of abstract reasoning (ARC-AGI). The first conflates memorization with capability; the second divorces reasoning from the practical contexts in which it matters. We take a different approach. The Grounded Integration Measure (GIM) is a benchmark of 820 original problems (615 public, 205 private) where difficulty comes from integration; individual problems require coordinating multiple cognitive operations (constraint satisfaction, state tracking, epistemic vigilance, audience calibration) over broadly accessible knowledge, so that reasoning stays grounded in realistic tasks without being gated on specialized expertise. Each problem is an original expert-authored composition, majority with rubric-decomposed scoring (median 6 independently judged criteria). A balanced public--private split provides built-in contamination diagnostic. We calibrate a continuous response 2-parameter logistic (2PL) IRT model over >200k prompt-response pairs across 28 models, producing robust ability estimates that correctly order test-configurations even when raw accuracy is distorted by errors or missing data, addressing a common challenge in benchmark reporting. Using this framework, we present a comprehensive leaderboard spanning 22 models and 47 test-configurations (unique model, thinking-level pairs), and conduct what is to our knowledge the most extensive published study of how test-time compute trades off against model capability on a fixed benchmark: 11 models swept across 35 test-configurations. We observe that within-family configuration choices, such as thinking budget and quantization, matter as much as model selection. We release the evaluation framework, calibrated IRT parameters, and all public problems.

Automated rationale generation is an approach for real-time explanation generation whereby a computational model learns to translate an autonomous agent's internal state and action data representations into natural language. Training on human explanation data can enable agents to learn to generate human-like explanations for their behavior. In this paper, using the context of an agent that plays Frogger, we describe (a) how to collect a corpus of explanations, (b) how to train a neural rationale generator to produce different styles of rationales, and (c) how people perceive these rationales. We conducted two user studies. The first study establishes the plausibility of each type of generated rationale and situates their user perceptions along the dimensions of confidence, humanlike-ness, adequate justification, and understandability. The second study further explores user preferences between the generated rationales with regard to confidence in the autonomous agent, communicating failure and unexpected behavior. Overall, we find alignment between the intended differences in features of the generated rationales and the perceived differences by users. Moreover, context permitting, participants preferred detailed rationales to form a stable mental model of the agent's behavior.

Large language models excel on objectively verifiable tasks such as math and programming, where evaluation reduces to unit tests or a single correct answer. In contrast, real-world enterprise work is often subjective and context-dependent: success hinges on organizational goals, user intent, and the quality of intermediate artifacts produced across long, multi-tool workflows. We introduce LH-Bench, a three-pillar evaluation design that moves beyond binary correctness to score autonomous, long-horizon execution on subjective enterprise tasks. The pillars are: (i) expert-grounded rubrics that give LLM judges the domain context needed to score subjective work, (ii) curated ground-truth artifacts that enable stepwise reward signals (e.g., chapter-level annotation for content tasks), and (iii) pairwise human preference evaluation for convergent validation. We show that domain-authored rubrics provide substantially more reliable evaluation signals than LLM-authored rubrics (kappa = 0.60 vs. 0.46), and that human preference judgments confirm the same top-tier separation (p < 0.05), evidence that expert-grounded evaluation can scale without sacrificing reliability. We release public datasets and report results on two environments: Figma-to-code (33 real .fig tasks against the Figma API via MCP) and Programmatic content (41 courses comprising 183 individually-evaluated chapters on a course platform serving 30+ daily users).

Language models are increasingly used for moral decision-making across diverse linguistic and cultural contexts, yet existing work overlooks multilinguality on three aspects: 1) multilingual evaluation benchmarks use direct translation, failing to adapt culture-specific items; 2) inference-time methods for moral reasoning rely on static, English-centric scaffolds and lack grounding in moral theory; 3) training methods for moral decision-making typically require expensive supervision from stronger models or human annotators. We address these gaps with three contributions. First, we introduce MCLASH, a multilingual moral decision-making benchmark to capture culturally situated moral intuitions and social norms across languages. Second, we propose MET (Multilingual Ethics with Theory-grounded reasoning), a two-step prompting method built on expert-curated, theory-based grounds drawn from psychology and philosophy: the model first selects situation- and culture-specific grounds, then reasons over them in the native language of the user. Third, we introduce MET-D (MET-Distillation), which enhances the second step through a self-distillation training stage that requires no external supervision. MET-D improves macro-F1 over the base model on all three models of different sizes and families (Qwen3-4B, Qwen3-8B, Gemma3-4B), by an average of 3.71 points on MCLASH and 4.23 on MMoralExceptQA, with a peak MCLASH gain of 12.94 points for Malay on Qwen3-8B. We further reveal that MET-D increases native-language reasoning by 62.13 points on average, and that beneficial grounds differ systematically across cultures. Together, these contributions open the path for culture-aligned, theory-grounded multilingual moral reasoning.

As agents take on long-horizon tasks, users shift from making individual decisions to overseeing autonomous execution. Yet the volume of agent activity and the fragmentation of supporting evidence make it difficult to determine which decisions warrant user verification. We study monitors that identify consequential decisions and locate evidence to help users assess their implications. We introduce AgentMonBench, a software-engineering benchmark comprising three subsets that cover two complementary dimensions: alignment between requirements and behavior, and awareness of consequential autonomous decisions for verification. To support these judgments, we propose the Evidence-Grounded Behavior Graph (EBG), a training-free method that groups source-linked evidence into behaviors and organizes their relationships into a graph. EBG presents task-oriented views of this graph to help monitors interpret behavior in context. Experiments across eight models show that EBG improves decision identification and evidence localization in most settings compared with direct access to the original context. Further experiments show that EBG's evidence-localization gains persist across input scales and hyperparameter settings, while real-world applications illustrate its practical value for human oversight.

Recent research has shown that rationales, or step-by-step chains of thought, can be used to improve performance in multi-step reasoning tasks. We reconsider rationale-augmented prompting for few-shot in-context learning, where (input -> output) prompts are expanded to (input, rationale -> output) prompts. For rationale-augmented prompting we demonstrate how existing approaches, which rely on manual prompt engineering, are subject to sub-optimal rationales that may harm performance. To mitigate this brittleness, we propose a unified framework of rationale-augmented ensembles, where we identify rationale sampling in the output space as the key component to robustly improve performance. This framework is general and can easily be extended to common natural language processing tasks, even those that do not traditionally leverage intermediate steps, such as question answering, word sense disambiguation, and sentiment analysis. We demonstrate that rationale-augmented ensembles achieve more accurate and interpretable results than existing prompting approaches--including standard prompting without rationales and rationale-based chain-of-thought prompting--while simultaneously improving interpretability of model predictions through the associated rationales.

Autonomous research systems are increasingly capable of executing long research workflows, yet automation alone does not ensure that the resulting process remains scientifically grounded. We introduce AutoResearch, a two-stage system that connects Idea Generation with Idea Execution to address both how research ideas are formed and how they are reliably established through experimentation. In Idea Generation, AutoResearch continuously integrates emerging research signals with accumulated domain knowledge, identifies transferable mechanistic insights, and uses multi-model generation and cross-review to produce grounded, testable research plans. In Idea Execution, coordinated agents decompose these plans into experiments, iteratively implement and diagnose them, and employ independent evidence-based review before accepting research conclusions. Across representative settings in cross-modal retrieval, systems optimization, and benchmark-driven machine learning, AutoResearch turns generated ideas into measurable progress, detects and corrects unreliable experimental results, and makes evidence-conditioned decisions to continue, revise, or terminate research directions. For example, on RSICD benchmark, an AutoResearch-generated idea improves mean Recall from 32.84 to 34.69, while recording only 5 audit-confirmed issue events compared with 11-27 for other autonomous research systems. These results demonstrate a research process in which meaningful insight is grounded before experimentation and conclusions are grounded before acceptance: Insight In, Hallucination Out.

A major issue with using deep learning models in sensitive applications is that they provide no explanation for their output. To address this problem, unsupervised selective rationalization produces rationales alongside predictions by chaining two jointly-trained components, a rationale generator and a predictor. Although this architecture guarantees that the prediction relies solely on the rationale, it does not ensure that the rationale contains a plausible explanation for the prediction. We introduce a novel training technique that effectively limits generation of implausible rationales by injecting noise between the generator and the predictor. Furthermore, we propose a new benchmark for evaluating unsupervised selective rationalization models using movie reviews from existing datasets. We achieve sizeable improvements in rationale plausibility and task accuracy over the state-of-the-art across a variety of tasks, including our new benchmark, while maintaining or improving model faithfulness.

Generating rationales that justify scoring decisions has been a promising way to facilitate explainability in automated scoring systems. However, existing methods do not match the accuracy of classifier-based methods. Plus, the generated rationales often contain hallucinated information. To address these issues, we propose a novel framework capable of generating more faithful rationales and, more importantly, matching performance with classifier-based black-box scoring systems. We first mimic the human assessment process by querying Large Language Models (LLMs) to generate a thought tree. We then summarise intermediate assessment decisions from each thought tree path for creating synthetic rationale data and rationale preference data. Finally, we utilise the generated synthetic data to calibrate LLMs through a two-step training process: supervised fine-tuning and preference optimization. Extensive experimental results demonstrate that our framework achieves a 38% assessment performance improvement in the QWK score compared to prior work while producing higher-quality rationales, as recognised by human evaluators and LLMs. Our work sheds light on the effectiveness of performing preference optimization using synthetic preference data obtained from thought tree paths.

In this work, we present a model for document-grounded response generation in dialog that is decomposed into two components according to Bayes theorem. One component is a traditional ungrounded response generation model and the other component models the reconstruction of the grounding document based on the dialog context and generated response. We propose different approximate decoding schemes and evaluate our approach on multiple open-domain and task-oriented document-grounded dialog datasets. Our experiments show that the model is more factual in terms of automatic factuality metrics than the baseline model. Furthermore, we outline how introducing scaling factors between the components allows for controlling the tradeoff between factuality and fluency in the model output. Finally, we compare our approach to a recently proposed method to control factuality in grounded dialog, CTRL (arXiv:2107.06963), and show that both approaches can be combined to achieve additional improvements.

Recent advances in large vision-language models have led to impressive performance in visual question answering and multimodal reasoning. However, it remains unclear whether these models genuinely perform grounded visual reasoning or rely on superficial patterns and dataset biases. In this work, we introduce MagiC, a comprehensive benchmark designed to evaluate grounded multimodal cognition, assessing not only answer accuracy but also the quality of step-by-step reasoning and its alignment with relevant visual evidence. Our benchmark includes approximately 5,500 weakly supervised QA examples generated from strong model outputs and 900 human-curated examples with fine-grained annotations, including answers, rationales, and bounding box groundings. We evaluate 15 vision-language models ranging from 7B to 70B parameters across four dimensions: final answer correctness, reasoning validity, grounding fidelity, and self-correction ability. MagiC further includes diagnostic settings to probe model robustness under adversarial visual cues and assess their capacity for introspective error correction. We introduce new metrics such as MagiScore and StepSense, and provide comprehensive analyses that reveal key limitations and opportunities in current approaches to grounded visual reasoning.

We study reasoning in Rule-Governed Decision Tasks (RGDTs), where models apply external rules to case facts and justify decisions, as required in policy, contract, and compliance settings. Beyond the deductive capability emphasized by standard mathematical and logical reasoning tasks, RGDTs require interpreting rules and their applicability, assessing conditions from evidence, combining judgments under rules and exceptions, and providing checkable justifications. These demands motivate a benchmark assessing both decisions and their stated grounds. We introduce RGDT-Bench, providing 202.1K condition-level supervision slots across four task tracks and eight supported task-probe combinations that vary access to supporting information. Label-blind extraction and deterministic checks produce labels for warrant completeness: source-referenced coverage and consistency of stated decision grounds. The benchmark attributes failures to four process layers: rule use, condition, evidence, and aggregation, and checks the final outcome. Among evaluable correct responses, warrant incompleteness averages 40.2% across six evaluated LLMs and supported task-probe combinations. Such warrant incompleteness poses potential safety risks and remains difficult to detect: the best of seventeen existing evaluators reaches only 57.69% (random: 50%) task-averaged area under the receiver operating characteristic curve (AUROC). To address this difficulty, we train a simple reward model with warrant supervision. It achieves 69.24% task-averaged AUROC among correct answers, exceeding the matched outcome-supervised baseline by 10.37 pp (percentage points) and the best existing evaluator by 11.55 pp. Beyond completeness assessment, the model outperforms both outcome-supervised baselines across nearly all response-selection comparisons, supporting RGDT-Bench's warrant supervision for RGDT reasoning.

Large language models are emerging as scientific assistants, but evaluating their ability to reason from empirical data remains challenging. Benchmarks derived from published studies and human annotations inherit publication bias, known-knowledge bias, label noise, and substantial storage requirements. We present InfiniteScienceGym, a procedurally generated benchmark of scientific repositories paired with a verifiable question-answering task. From a seed, the simulator deterministically generates a self-contained repository with realistic directory structure, files, and tabular data, and a privileged QA generator produces both answerable and unanswerable questions with exact ground truth. This makes it possible to evaluate evidence-grounded reasoning, abstention, and tool-mediated analysis in a controlled setting without distributing a large static corpus. InfiniteScienceGym complements real scientific benchmarks by targeting blind spots and failure modes that are hard to evaluate using published datasets alone. Evaluating both proprietary and open-weight models, we find that none achieve more than 45% accuracy overall, that recognizing unanswerable questions remains a major weakness, and that stronger models tend to use tools more effectively rather than simply consuming more tokens.

State-of-the-art models in NLP are now predominantly based on deep neural networks that are opaque in terms of how they come to make predictions. This limitation has increased interest in designing more interpretable deep models for NLP that reveal the `reasoning' behind model outputs. But work in this direction has been conducted on different datasets and tasks with correspondingly unique aims and metrics; this makes it difficult to track progress. We propose the Evaluating Rationales And Simple English Reasoning (ERASER) benchmark to advance research on interpretable models in NLP. This benchmark comprises multiple datasets and tasks for which human annotations of "rationales" (supporting evidence) have been collected. We propose several metrics that aim to capture how well the rationales provided by models align with human rationales, and also how faithful these rationales are (i.e., the degree to which provided rationales influenced the corresponding predictions). Our hope is that releasing this benchmark facilitates progress on designing more interpretable NLP systems. The benchmark, code, and documentation are available at https://www.eraserbenchmark.com/

Recent advances in vision-language reasoning underscore the importance of thinking with images, where models actively ground their reasoning in visual evidence. Yet, prevailing frameworks treat visual actions as optional tools, boosting metrics but leaving reasoning ungrounded and crops ineffective. This gap gives rise to the illusion of thinking with images: models seem visually grounded but rely on context-agnostic actions that neither refine perception nor guide reasoning toward correct answers. We address this problem by reframing visual actions as core reasoning primitives rather than optional tools, which we term visual rationalization, the visual analogue of textual Chain-of-Thought. Building on this insight, we propose Visual Rationale Learning (ViRL), an end-to-end paradigm that grounds training in the visual rationale itself. ViRL integrates (1) Process Supervision with ground-truth rationales, (2) Objective Alignment via step-level reward shaping, and (3) Fine-Grained Credit Assignment to distinguish correct, redundant, and erroneous actions. By ensuring each action contributes meaningfully to the reasoning chain, ViRL enables models to "get the right answer for the right visual reason". Trained purely with end-to-end RL, ViRL achieves state-of-the-art results across benchmarks spanning perception, hallucination, and reasoning. This work establishes visual rationalization as a task-agnostic, process-grounded paradigm for building transparent, verifiable, and trustworthy vision-language models.

The rapid evolution of Large Language Models has catalyzed a surge in scientific idea production, yet this leap has not been accompanied by a matching advance in idea evaluation. The fundamental nature of scientific evaluation needs knowledgeable grounding, collective deliberation, and multi-criteria decision-making. However, existing idea evaluation methods often suffer from narrow knowledge horizons, flattened evaluation dimensions, and the inherent bias in LLM-as-a-Judge. To address these, we regard idea evaluation as a knowledge-grounded, multi-perspective reasoning problem and introduce InnoEval, a deep innovation evaluation framework designed to emulate human-level idea assessment. We apply a heterogeneous deep knowledge search engine that retrieves and grounds dynamic evidence from diverse online sources. We further achieve review consensus with an innovation review board containing reviewers with distinct academic backgrounds, enabling a multi-dimensional decoupled evaluation across multiple metrics. We construct comprehensive datasets derived from authoritative peer-reviewed submissions to benchmark InnoEval. Experiments demonstrate that InnoEval can consistently outperform baselines in point-wise, pair-wise, and group-wise evaluation tasks, exhibiting judgment patterns and consensus highly aligned with human experts.

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.

When provided with sufficient explanatory context, smaller Language Models have been shown to exhibit strong reasoning ability on challenging short-answer question-answering tasks where the questions are unseen in training. We evaluate two methods for further improvement in this setting. Both methods focus on combining rationales generated by a larger Language Model with longer contexts created from a multi-hop dense retrieval system. The first method (RR) involves training a Rationale Ranking model to score both generated rationales and retrieved contexts with respect to relevance and truthfulness. We then use the scores to derive combined contexts from both knowledge sources using a number of combinatory strategies. For the second method (RATD) we utilise retrieval-augmented training datasets developed by Hartill et al. 2023 to train a smaller Reasoning model such that it becomes proficient at utilising relevant information from longer text sequences that may be only partially evidential and frequently contain many irrelevant sentences. We find that both methods significantly improve results. Our single best Reasoning model materially improves upon strong comparable prior baselines for unseen evaluation datasets (StrategyQA 58.9 rightarrow 61.7 acc., CommonsenseQA 63.6 rightarrow 72.7 acc., ARC-DA 31.6 rightarrow 52.1 F1, IIRC 25.5 rightarrow 27.3 F1) and a version utilising our prior knowledge of each type of question in selecting a context combination strategy does even better. Our proposed models also generally outperform direct prompts against much larger models (BLOOM 175B and StableVicuna 13B) in both few-shot chain-of-thought and standard few-shot settings.

Large language models are increasingly capable at closed-world mathematical reasoning, but research assistance also requires source-grounded use of the literature. When a proof reaches a non-trivial step, a useful assistant should determine whether the needed tool (e.g., a lemma) already exists, identify a suitable scholarly source, and verify that its assumptions align with the current proof context. To rigorously evaluate such capabilities, we introduce Re^2Math, a benchmark for tool-grounded retrieval from partial mathematical proofs. Each instance is built from a candidate instrumental citation in the proof of a main theorem, with hierarchical context and an optional leakage-controlled anchor hint. We also make the task source-grounded yet citation-agnostic in that any admissible theorem sufficient for the proof transition is accepted. Evaluation uses a release-frozen retrieval artifact, ensuring reproducibility, while the benchmark itself supports automatic, continual expansion with newly constructed instances. On the current benchmark test set, the best fixed-judge ToolAcc reaches 7.0%, despite substantially higher rates of source grounding, indicating that current systems often retrieve valid statements but fail to establish their applicability to the local proof step. By decoupling citation recall, grounding, and proof-gap sufficiency, Re^2Math transforms literature-grounded mathematical tool use into a controlled diagnostic task.

This paper introduces two significant contributions to address the issue of grounding claims in a given context. Grounding means that given a context (document) and a claim, there's at least one supportive evidence for the claim in the document. We will introduce Paladin-mini, a compact (3.8B parameters) open-source classifier model (used for labeling data as grounded or ungrounded) engineered for robust performance in real-world scenarios, and the grounding-benchmark, a new evaluation dataset designed to assess performance on critical reasoning tasks. We'll also demonstrate the results of Paladin-mini with benchmarks against the current State-of-the-art and share clear and reproducible results.

Grounded generation aims to equip language models (LMs) with the ability to produce more credible and accountable responses by accurately citing verifiable sources. However, existing methods, by either feeding LMs with raw or preprocessed materials, remain prone to errors. To address this, we introduce CaLM, a novel verification framework. CaLM leverages the insight that a robust grounded response should be consistent with information derived solely from its cited sources. Our framework empowers smaller LMs, which rely less on parametric memory and excel at processing relevant information given a query, to validate the output of larger LMs. Larger LM responses that closely align with the smaller LMs' output, which relies exclusively on cited documents, are verified. Responses showing discrepancies are iteratively refined through a feedback loop. Experiments on three open-domain question-answering datasets demonstrate significant performance gains of 1.5% to 7% absolute average without any required model fine-tuning.

The emergent phenomena of large foundation models have revolutionized natural language processing. However, evaluating these models presents significant challenges due to their size, capabilities, and deployment across diverse applications. Existing literature often focuses on individual aspects, such as benchmark performance or specific tasks, but fails to provide a cohesive process that integrates the nuances of diverse use cases with broader ethical and operational considerations. This work focuses on three key aspects: (1) Formalizing the Evaluation Process by providing a structured framework tailored to specific use-case contexts, (2) Offering Actionable Tools and Frameworks such as checklists and templates to ensure thorough, reproducible, and practical evaluations, and (3) Surveying Recent Work with a targeted review of advancements in LLM evaluation, emphasizing real-world applications.

Language model users often issue queries that lack specification, where the context under which a query was issued -- such as the user's identity, the query's intent, and the criteria for a response to be useful -- is not explicit. For instance, a good response to a subjective query like "What book should I read next?" would depend on the user's preferences, and a good response to an open-ended query like "How do antibiotics work against bacteria?" would depend on the user's expertise. This makes evaluation of responses to such queries an ill-posed task, as evaluators may make arbitrary judgments about the response quality. To remedy this, we present contextualized evaluations, a protocol that synthetically constructs context surrounding an underspecified query and provides it during evaluation. We find that the presence of context can 1) alter conclusions drawn from evaluation, even flipping win rates between model pairs, 2) nudge evaluators to make fewer judgments based on surface-level criteria, like style, and 3) provide new insights about model behavior across diverse contexts. Specifically, our procedure uncovers an implicit bias towards WEIRD contexts in models' "default" responses and we find that models are not equally sensitive to following different contexts, even when they are provided in prompts.

Knowledge-grounded dialogue is a task of generating an informative response based on both the dialogue history and external knowledge source. In general, there are two forms of knowledge: manually annotated knowledge graphs and knowledge text from website. From various evaluation viewpoints, each type of knowledge has advantages and downsides. To further distinguish the principles and determinants from the intricate factors, we conduct a thorough experiment and study on the task to answer three essential questions. The questions involve the choice of appropriate knowledge form, the degree of mutual effects between knowledge and the model selection, and the few-shot performance of knowledge. Supported by statistical shreds of evidence, we offer conclusive solutions and sensible suggestions for directions and standards of future research.

The reasoning steps generated by LLMs might be incomplete, as they mimic logical leaps common in everyday communication found in their pre-training data: underlying rationales are frequently left implicit (unstated). To address this challenge, we introduce RATIONALYST, a model for process-supervision of reasoning based on pre-training on a vast collection of rationale annotations extracted from unlabeled data. We extract 79k rationales from web-scale unlabelled dataset (the Pile) and a combination of reasoning datasets with minimal human intervention. This web-scale pre-training for reasoning allows RATIONALYST to consistently generalize across diverse reasoning tasks, including mathematical, commonsense, scientific, and logical reasoning. Fine-tuned from LLaMa-3-8B, RATIONALYST improves the accuracy of reasoning by an average of 3.9% on 7 representative reasoning benchmarks. It also demonstrates superior performance compared to significantly larger verifiers like GPT-4 and similarly sized models fine-tuned on matching training sets.

AI agents can now carry out data-driven scientific analyses end to end, and benchmarks assess them by giving an agent a question and a dataset and scoring its final answer against a fixed key. These benchmarks assume that a correct answer was derived from the supplied data, a property we call evidence grounding. However, an agent can also reach the key from prior knowledge or by ruling out the other options, and a score based on a single run cannot tell these cases apart. We show how to test this assumption and find that it often fails. For each question, we build versions of its data files in which the evidence for the answer is left intact, withdrawn or reversed, check each edit with a pre-registered reference statistic, and run the same agent on every version. We then measure evidence-grounded accuracy, which credits a correct answer only if the agent also responds when the evidence is withdrawn and follows it when it is reversed. We evaluate three agent scaffolds and five models on 18 single-cell questions from BAISBench and four synthetic problems from GeneBench-Pro. On the single-cell questions, the Claude agents are 95% accurate and answer 83% correctly without any data, but their evidence-grounded accuracy is only 41%. Hiding gene names raises the share of runs that follow reversed evidence from 58% to 93% on ten gene tasks, suggesting that prior knowledge competes with the supplied data. The benchmark score and LLM judges can also reward answers that ignore the changed evidence. Measuring scientific intelligence rather than recall therefore requires checking whether answers follow the evidence and whether scores reward them for it.

Generative Reward Models (GenRMs) and LLM-as-a-Judge exhibit deceptive alignment by producing correct judgments for incorrect reasons, as they are trained and evaluated to prioritize Outcome Accuracy, which undermines their ability to generalize during RLHF. We introduce Rationale Consistency, a fine-grained metric that quantifies the alignment between the model's reasoning process and human judgment. Our evaluation of frontier models reveals that rationale consistency effectively discriminates among state-of-the-art models and detects deceptive alignment, while outcome accuracy falls short in both respects. To mitigate this gap, we introduce a hybrid signal that combines rationale consistency with outcome accuracy for GenRM training. Our training method achieves state-of-the-art performance on RM-Bench (87.1%) and JudgeBench (82%), surpassing outcome-only baselines by an average of 5%. Using RM during RLHF, our method effectively improves performance as demonstrated on Arena Hard v2, notably yielding a 7% improvement in creative writing tasks. Further analysis confirms that our method escapes the deceptive alignment trap, effectively reversing the decline in rationale consistency observed in outcome-only training.

Reasoning rerankers write a rationale before ranking a candidate slate, and existing systems read item metadata to do so, even though generative recommenders increasingly represent items as compact Semantic IDs (SIDs). Ranking from SIDs alone has two appeals: SID prompts are far shorter than the matched item text, and although item codes are fixed from content, the backbone learns their token embeddings in part from user-item interaction sequences, so the tokens can carry behavioral signal that item text lacks. That training targets next-item generation, and whether it supports ranking a supplied slate from SIDs alone is untested. We build ReRankID, a controlled testbed in which models reason and rank from a history and slate shown only as SIDs, varying item rendering, history length, and slate size one at a time. On this testbed, such models often copy the input order instead of reranking, and readable metadata does not repair the failure. We compare supervised fine-tuning, rejection-sampled self-training, on-policy distillation, and verifier-based reinforcement learning as post-training methods and find that they elicit the missing behavior. The strongest post-trained model, the 8B model trained with a graded ranking reward, gives the best SID-only ranking and matches strong text-reading LLMs whose prompts are about 20x longer. It ranks above random with no history at all, where a caption-reading reference does not, and it keeps its gains at long histories, where text prompts pass 50k tokens and the caption-reading reference stops improving. In a paired rationale audit, this model argues trade-offs better than that reference but describes items less accurately. The common post-training methods we test thus elicit ranking from SIDs but not faithful item descriptions, which calls for post-training that improves the reasoning itself, not only the ranking it produces.

Training language models with rationales augmentation has been shown to be beneficial in many existing works. In this paper, we identify that such a prevailing view does not hold consistently. We conduct comprehensive investigations to thoroughly inspect the impact of rationales on model performance as well as a novel perspective of model reliability. The results lead to several key findings that add new insights upon existing understandings: 1) Rationales can, at times, deteriorate model performance; 2) Rationales can, at times, improve model reliability, even outperforming their untrained counterparts; 3) A linear correspondence exists in between the performance and reliability improvements, while both are driven by the intrinsic difficulty of the task. These findings provide informative regulations on the broad utilization of rationales and raise critical implications on the procedure of explicitly aligning language models with implicit human thoughts. Codes can be found at https://github.com/Ignoramus0817/rationales.

Evaluating large language models (LLMs) on final-answer correctness is the dominant paradigm. This approach, however, provides a coarse signal for model improvement and overlooks the quality of the underlying reasoning process. We argue that a more granular evaluation of reasoning offers a more effective path to building robust models. We decompose reasoning quality into two dimensions: relevance and coherence. Relevance measures if a step is grounded in the problem; coherence measures if it follows logically from prior steps. To measure these aspects reliably, we introduce causal stepwise evaluation (CaSE). This method assesses each reasoning step using only its preceding context, which avoids hindsight bias. We validate CaSE against human judgments on our new expert-annotated benchmarks, MRa-GSM8K and MRa-MATH. More importantly, we show that curating training data with CaSE-evaluated relevance and coherence directly improves final task performance. Our work provides a scalable framework for analyzing, debugging, and improving LLM reasoning, demonstrating the practical value of moving beyond validity checks.

Rationale discovery is defined as finding a subset of the input data that maximally supports the prediction of downstream tasks. In graph machine learning context, graph rationale is defined to locate the critical subgraph in the given graph topology, which fundamentally determines the prediction results. In contrast to the rationale subgraph, the remaining subgraph is named the environment subgraph. Graph rationalization can enhance the model performance as the mapping between the graph rationale and prediction label is viewed as invariant, by assumption. To ensure the discriminative power of the extracted rationale subgraphs, a key technique named "intervention" is applied. The core idea of intervention is that given any changing environment subgraphs, the semantics from the rationale subgraph is invariant, which guarantees the correct prediction result. However, most, if not all, of the existing rationalization works on graph data develop their intervention strategies on the graph level, which is coarse-grained. In this paper, we propose well-tailored intervention strategies on graph data. Our idea is driven by the development of Transformer models, whose self-attention module provides rich interactions between input nodes. Based on the self-attention module, our proposed invariant graph Transformer (IGT) can achieve fine-grained, more specifically, node-level and virtual node-level intervention. Our comprehensive experiments involve 7 real-world datasets, and the proposed IGT shows significant performance advantages compared to 13 baseline methods.

Evaluating whether multimodal large language models truly understand long-form scientific papers remains challenging: answer-only metrics and synthetic "Needle-In-A-Haystack" tests often reward answer matching without requiring a causal, evidence-linked reasoning trace in the document. We propose the "Fish-in-the-Ocean" (FITO) paradigm, which requires models to construct explicit cross-modal evidence chains within native scientific documents. To operationalize FITO, we build SIN-Data, a scientific interleaved corpus that preserves the native interleaving of text and figures. On top of it, we construct SIN-Bench with four progressive tasks covering evidence discovery (SIN-Find), hypothesis verification (SIN-Verify), grounded QA (SIN-QA), and evidence-anchored synthesis (SIN-Summary). We further introduce "No Evidence, No Score", scoring predictions when grounded to verifiable anchors and diagnosing evidence quality via matching, relevance, and logic. Experiments on eight MLLMs show that grounding is the primary bottleneck: Gemini-3-pro achieves the best average overall score (0.573), while GPT-5 attains the highest SIN-QA answer accuracy (0.767) but underperforms on evidence-aligned overall scores, exposing a gap between correctness and traceable support.

As reasoning capacity and deployment scope grow in tandem, large language models (LLMs) gain the capacity to engage in behaviors that serve their own objectives, a class of risks we term Emergent Strategic Reasoning Risks (ESRRs). These include, but are not limited to, deception (intentionally misleading users or evaluators), evaluation gaming (strategically manipulating performance during safety testing), and reward hacking (exploiting misspecified objectives). Systematically understanding and benchmarking these risks remains an open challenge. To address this gap, we introduce ESRRSim, a taxonomy-driven agentic framework for automated behavioral risk evaluation. We construct an extensible risk taxonomy of 7 categories, which is decomposed into 20 subcategories. ESRRSim generates evaluation scenarios designed to elicit faithful reasoning, paired with dual rubrics assessing both model responses and reasoning traces, in a judge-agnostic and scalable architecture. Evaluation across 11 reasoning LLMs reveals substantial variation in risk profiles (detection rates ranging 14.45%-72.72%), with dramatic generational improvements suggesting models may increasingly recognize and adapt to evaluation contexts.

AI assistants can impart value judgments that shape people's decisions and worldviews, yet little is known empirically about what values these systems rely on in practice. To address this, we develop a bottom-up, privacy-preserving method to extract the values (normative considerations stated or demonstrated in model responses) that Claude 3 and 3.5 models exhibit in hundreds of thousands of real-world interactions. We empirically discover and taxonomize 3,307 AI values and study how they vary by context. We find that Claude expresses many practical and epistemic values, and typically supports prosocial human values while resisting values like "moral nihilism". While some values appear consistently across contexts (e.g. "transparency"), many are more specialized and context-dependent, reflecting the diversity of human interlocutors and their varied contexts. For example, "harm prevention" emerges when Claude resists users, "historical accuracy" when responding to queries about controversial events, "healthy boundaries" when asked for relationship advice, and "human agency" in technology ethics discussions. By providing the first large-scale empirical mapping of AI values in deployment, our work creates a foundation for more grounded evaluation and design of values in AI systems.

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

Front-end web code has become a core product surface for every frontier LLM release, yet evaluating these interactive applications at development speed remains costly because human-judged leaderboards like Arena do not scale. Existing automated proxies typically lean on reference implementations, test suites, or rigid checklists, and tend to miss the reasoned synthesis a human reviewer performs over a live session. We articulate a new evaluation regime that is simultaneously reference-free, autonomously driven, and holistically reasoned, and instantiate it through two artifacts. \dataname is an 11-domain, 54-leaf, 1,000-query WebDev benchmark spanning both static-presentation and interactive-application tasks, balanced across three difficulty tiers and three target-language groups, with briefs rewritten to resist recall from circulated prompts. \framename, grounded in Flavell's metacognitive monitoring, separates evidence accumulation from judgment across three stages: Static Perception forms a first impression from passive observation; Agent-Driven Interaction explores the application autonomously while capturing continuous screen video, audio, and per-step screenshots; Dynamic Scoring issues holistic functionality and aesthetics verdicts with structured failure attribution only after the evidence chain is complete. On \dataname, \framename aligns closely with expert human ratings while surfacing substantial headroom across 13 frontier LLMs on interactive web generation. \noindenthttps://anonymous.4open.science/r/Cookie-3CE/

Logical reading comprehension is a challenging task that entails grasping the underlying semantics of text and applying reasoning to deduce the correct answer. Prior researches have primarily focused on enhancing logical reasoning capabilities through Chain-of-Thought (CoT) or data augmentation. However, previous work constructing chain-of-thought rationales concentrates solely on analyzing correct options, neglecting the incorrect alternatives. Addtionally, earlier efforts on data augmentation by altering contexts rely on rule-based methods, which result in generated contexts that lack diversity and coherence. To address these issues, we propose a Premise-Oriented Data Augmentation (PODA) framework. This framework can generate CoT rationales including analyses for both correct and incorrect options, while constructing diverse and high-quality counterfactual contexts from incorrect candidate options. We integrate summarizing premises and identifying premises for each option into rationales. Subsequently, we employ multi-step prompts with identified premises to construct counterfactual context. To facilitate the model's capabilities to better differentiate the reasoning process associated with each option, we introduce a novel thought-path contrastive learning method that compares reasoning paths between the original and counterfactual samples. Experimental results on three representative LLMs demonstrate that our method can improve the baselines substantially across two challenging logical reasoning benchmarks (ReClor and LogiQA 2.0). The data and code are released at https://github.com/lalalamdbf/TPReasoner.

Can financial vision-language models (VLMs) turn chart evidence into reliable action recommendations? Existing hallucination evaluations are mostly claim-centric; they assess whether generated statements are supported, but not whether evidence remains traceable through rationale, confidence, and final action. We introduce E2A-Bench, a 969-query benchmark for financial chart reasoning, constructed from 323 HS300 constituents under three input modalities with deterministic OHLCV-derived evidence anchors. E2A-Bench evaluates grounding, reasoning-action consistency, evidence-confidence calibration, and directional coverage through UCR, RCI, ECI, and NDR, where NDR measures coverage-aware evidence-to-action reliability rather than realized trading performance. Evaluating 20 VLMs reveals three failures hidden by scalar hallucination scores: the lowest-UCR model ranks near the bottom by NDR due to only 6.4% directional coverage; oracle-aided verification reduces unsupported claims but can collapse coverage; and financial fine-tuning amplifies the BUY:SELL ratio by factors of 4.21 to 4.68 across strict base-fine-tuned pairs. These results show that financial VLM evaluation should trace the full evidence-to-action chain rather than rely on a single hallucination score. Code and data: https://github.com/wanng-ide/E2A-Bench

Most reward models for visual generation reduce rich human judgments to a single unexplained score, discarding the reasoning that underlies preference. We show that teaching reward models to produce explicit, multi-dimensional critiques before scoring transforms them from passive evaluators into active optimization tools, improving generators in two complementary ways: at training time, structured rationales provide interpretable, fine-grained rewards for reinforcement learning; at test time, a Generate-Critique-Refine loop turns critiques into targeted prompt revisions that improve outputs without any parameter updates. To train such a reward model without costly rationale annotations, we introduce Preference-Anchored Rationalization (PARROT), a principled framework that recovers high-quality rationales from readily available preference data through anchored generation, consistency filtering, and distillation. The resulting model, RationalRewards (8B), achieves state-of-the-art preference prediction among open-source reward models, competitive with Gemini-2.5-Pro, while using 10-20x less training data than comparable baselines. As an RL reward, it consistently improves text-to-image and image-editing generators beyond scalar alternatives. Most strikingly, its test-time critique-and-refine loop matches or exceeds RL-based fine-tuning on several benchmarks, suggesting that structured reasoning can unlock latent capabilities in existing generators that suboptimal prompts fail to elicit.

Generative AI systems across modalities, ranging from text (including code), image, audio, and video, have broad social impacts, but there is no official standard for means of evaluating those impacts or for which impacts should be evaluated. In this paper, we present a guide that moves toward a standard approach in evaluating a base generative AI system for any modality in two overarching categories: what can be evaluated in a base system independent of context and what can be evaluated in a societal context. Importantly, this refers to base systems that have no predetermined application or deployment context, including a model itself, as well as system components, such as training data. Our framework for a base system defines seven categories of social impact: bias, stereotypes, and representational harms; cultural values and sensitive content; disparate performance; privacy and data protection; financial costs; environmental costs; and data and content moderation labor costs. Suggested methods for evaluation apply to listed generative modalities and analyses of the limitations of existing evaluations serve as a starting point for necessary investment in future evaluations. We offer five overarching categories for what can be evaluated in a broader societal context, each with its own subcategories: trustworthiness and autonomy; inequality, marginalization, and violence; concentration of authority; labor and creativity; and ecosystem and environment. Each subcategory includes recommendations for mitigating harm.

Cause-and-effect reasoning in video is a significant challenge for Vision-Language Models (VLMs), as it requires going beyond surface-level perception to a deeper understanding of causal mechanisms. However, existing benchmarks rarely provide the fine-grained, grounded evidence needed to rigorously evaluate this capability. To address this gap, we introduce CaST-Bench, a benchmark for Causal Chain-Grounded Spatio-Temporal Video Reasoning. CaST-Bench presents complex causal questions that require models to identify and localize a chain of multiple spatio-temporal evidences. Through a human-AI collaborative pipeline, we construct a high-quality dataset of 2,066 questions over 1,015 videos, with causal chains annotated by temporal segments and bounding-box tracks. Furthermore, we design a comprehensive evaluation suite with novel metrics that assess not only answer correctness but also the capability for visual evidence grounded reasoning. This grounding is crucial for improving accuracy by mitigating spurious correlations and for enhancing user trust by making models more transparent. Our experiments show that current VLMs struggle with causal questions, largely due to their limited ability to construct precise and grounded causal chains. This highlights an important direction for improving future VLMs.

Ethical decision-making is a critical aspect of human judgment, and the growing use of LLMs in decision-support systems necessitates a rigorous evaluation of their moral reasoning capabilities. However, existing assessments primarily rely on single-step evaluations, failing to capture how models adapt to evolving ethical challenges. Addressing this gap, we introduce the Multi-step Moral Dilemmas (MMDs), the first dataset specifically constructed to evaluate the evolving moral judgments of LLMs across 3,302 five-stage dilemmas. This framework enables a fine-grained, dynamic analysis of how LLMs adjust their moral reasoning across escalating dilemmas. Our evaluation of nine widely used LLMs reveals that their value preferences shift significantly as dilemmas progress, indicating that models recalibrate moral judgments based on scenario complexity. Furthermore, pairwise value comparisons demonstrate that while LLMs often prioritize the value of care, this value can sometimes be superseded by fairness in certain contexts, highlighting the dynamic and context-dependent nature of LLM ethical reasoning. Our findings call for a shift toward dynamic, context-aware evaluation paradigms, paving the way for more human-aligned and value-sensitive development of LLMs.

The ability of large language models (LLMs) to follow user instructions is central to their reliability, safety, and usefulness. While prior studies assess instruction adherence in the model's main responses, we argue that it is also critical for large reasoning models (LRMs) to follow user instructions throughout their reasoning process. Reasoning instruction following makes LRMs more controllable and transparent, while reducing risks of undesirable shortcuts, hallucinations, or reward hacking within reasoning traces. To evaluate this dimension, we introduce ReasonIF, a systematic benchmark for assessing reasoning instruction following. ReasonIF includes six categories of instruction prompts, spanning multilingual reasoning, formatting and length control. Across many open-source LRMs including GPT-OSS, Qwen3, and DeepSeek-R1, we find substantial failures in reasoning instruction adherence: the highest instruction following score (IFS) remains below 0.25, meaning that fewer than 25% of reasoning traces comply with the given instructions. Notably, as task difficulty increases, reasoning instruction following degrades further. We also explore two strategies to enhance reasoning instruction fidelity. (1) multi-turn reasoning and (2) Reasoning Instruction Finetuning (RIF) using synthetic data. RIF improves the IFS of GPT-OSS-20B from 0.11 to 0.27, indicating measurable progress but leaving ample room for improvement.

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.

Knowledge grounded dialogue system is designed to generate responses that convey information from given knowledge documents. However, it's a challenge for the current Seq2Seq model to acquire knowledge from complex documents and integrate it to perform correct responses without the aid of an explicit semantic structure. To address these issues, we present a novel graph structure, Ground Graph (G^2), which models the semantic structure of both dialogue contexts and knowledge documents to facilitate knowledge selection and integration for the task. Besides, a Ground Graph Aware Transformer (G^2AT) is proposed to enhance knowledge grounded response generation. Empirical results show that our proposed model outperforms previous state-of-the-art methods with more than 10\% and 20\% gains on response generation and factual consistency. Furthermore, our structure-aware approach shows excellent generalization ability in resource-limited situations.

Large-scale high-quality training data is important for improving the performance of models. After trained with data that has rationales (reasoning steps), models gain reasoning capability. However, the dataset with high-quality rationales is relatively scarce due to the high annotation cost. To address this issue, we propose Self-motivated Learning framework. The framework motivates the model itself to automatically generate rationales on existing datasets. Based on the inherent rank from correctness across multiple rationales, the model learns to generate better rationales, leading to higher reasoning capability. Specifically, we train a reward model with the rank to evaluate the quality of rationales, and improve the performance of reasoning through reinforcement learning. Experiment results of Llama2 7B on multiple reasoning datasets show that our method significantly improves the reasoning ability of models, even outperforming text-davinci-002 in some datasets.

Multi-hop QA benchmarks often reward Large Language Models (LLMs) for spurious correctness, where models reach correct answers through invalid intermediate reasoning. We propose SAFE, an LLM-as-verifier framework for evidence-grounded multi-hop QA. Rather than judging only the final answer after generation, SAFE verifies reasoning during generation by checking intermediate steps against the provided passages and previous reasoning trajectory. To make this process checkable, SAFE decomposes reasoning into atomic, evidence-grounded units represented with Knowledge Graph (KG) triples. At train-time, SAFE verifies benchmark supervision under KG-grounded constraints and constructs reliable verifier training data. At inference-time, an external verifier checks each generated step, identifies invalid reasoning, and provides correction feedback before errors propagate. Across three multi-hop QA benchmarks, SAFE improves accuracy by 8.8 pp on average. These results show that evidence-grounded multi-hop QA benefits from shifting LLM-based evaluation from post-hoc answer judgment to stepwise reasoning verification.

We automate deep step-by step reasoning in an LLM dialog thread by recursively exploring alternatives (OR-nodes) and expanding details (AND-nodes) up to a given depth. Starting from a single succinct task-specific initiator we steer the automated dialog thread to stay focussed on the task by synthesizing a prompt that summarizes the depth-first steps taken so far. Our algorithm is derived from a simple recursive descent implementation of a Horn Clause interpreter, except that we accommodate our logic engine to fit the natural language reasoning patterns LLMs have been trained on. Semantic similarity to ground-truth facts or oracle advice from another LLM instance is used to restrict the search space and validate the traces of justification steps returned as answers. At the end, the unique minimal model of a generated Horn Clause program collects the results of the reasoning process. As applications, we sketch implementations of consequence predictions, causal explanations, recommendation systems and topic-focussed exploration of scientific literature.

A computational ethics framework is essential for AI and autonomous systems operating in complex, real-world environments. Existing approaches often lack the adaptability needed to integrate ethical principles into dynamic and ambiguous contexts, limiting their effectiveness across diverse scenarios. To address these challenges, we outline the necessary ingredients for building a holistic, meta-level framework that combines intermediate representations, probabilistic reasoning, and knowledge representation. The specifications therein emphasize scalability, supporting ethical reasoning at both individual decision-making levels and within the collective dynamics of multi-agent systems. By integrating theoretical principles with contextual factors, it facilitates structured and context-aware decision-making, ensuring alignment with overarching ethical standards. We further explore proposed theorems outlining how ethical reasoners should operate, offering a foundation for practical implementation. These constructs aim to support the development of robust and ethically reliable AI systems capable of navigating the complexities of real-world moral decision-making scenarios.

Grounded dialogue models generate responses that are grounded on certain concepts. Limited by the distribution of grounded dialogue data, models trained on such data face the transferability challenges in terms of the data distribution and the type of grounded concepts. To address the challenges, we propose the grounded minimal editing framework, which minimally edits existing responses to be grounded on the given concept. Focusing on personas, we propose Grounded Minimal Editor (GME), which learns to edit by disentangling and recombining persona-related and persona-agnostic parts of the response. To evaluate persona-grounded minimal editing, we present the PersonaMinEdit dataset, and experimental results show that GME outperforms competitive baselines by a large margin. To evaluate the transferability, we experiment on the test set of BlendedSkillTalk and show that GME can edit dialogue models' responses to largely improve their persona consistency while preserving the use of knowledge and empathy.

We investigate the robustness of value alignment via finetuning with synthetic documents, using animal compassion as a value that is both important in its own right and orthogonal to existing alignment efforts. To evaluate compassionate reasoning, we develop and publicly release the Animal Harm Benchmark (AHB), a 26-question evaluation spanning 13 ethical dimensions, publicly available as a dataset and Inspect evaluation. On the AHB, training with 3000 documents achieves 77% compared to 40% for instruction-tuning approaches, with generalization to human compassion and no degradation in standard safety benchmarks or capabilities. However, subsequent unrelated instruction-tuning degrades the intervention, with the advantage disappearing after 5000 samples. Our exploratory results suggest document-based value interventions may require explicit preservation strategies to remain effective through typical training pipelines.

In this position paper, we argue that instead of morally aligning LLMs to specific set of ethical principles, we should infuse generic ethical reasoning capabilities into them so that they can handle value pluralism at a global scale. When provided with an ethical policy, an LLM should be capable of making decisions that are ethically consistent to the policy. We develop a framework that integrates moral dilemmas with moral principles pertaining to different foramlisms of normative ethics, and at different levels of abstractions. Initial experiments with GPT-x models shows that while GPT-4 is a nearly perfect ethical reasoner, the models still have bias towards the moral values of Western and English speaking societies.

Common-sense reasoning is a key language model capability because it encapsulates not just specific factual knowledge but rather general language and world understanding. Measuring common-sense reasoning, therefore, is crucial for language models of different sizes and applications. One of the most widely used benchmarks for evaluating such capabilities is HellaSwag; however, in this paper, we show that it has severe construct validity issues. These issues range from basic ungrammaticality and numerous typos to misleading prompts or equally correct options. Furthermore, we show that if models are evaluated only on answer texts, or with "Lorem ipsum dolor..." instead of the question, more than 65% of model predictions remain the same, and this cannot be attributed merely to contamination. Since benchmark scores are an essential part of model selection in both research and commercial applications, these validity issues can have severe consequences. In particular, knowing that taking benchmark scores at face value is ubiquitous, inadequate evaluation leads to ill-informed decisions about models. In this paper, we thoroughly investigate critical validity issues posed by HellaSwag and illustrate them with various evaluations using generative language models of different sizes. We argue that this benchmark does not accurately measure common-sense reasoning and, therefore, should not be used for evaluation in its current state. Based on the results of our study, we propose requirements that should be met by future common-sense reasoning benchmarks. In addition, we release GoldenSwag, a corrected subset of HellaSwag, which, to our belief, facilitates acceptable common-sense reasoning evaluation.

Interpretability or explainability is an emerging research field in NLP. From a user-centric point of view, the goal is to build models that provide proper justification for their decisions, similar to those of humans, by requiring the models to satisfy additional constraints. To this end, we introduce a new application on legal text where, contrary to mainstream literature targeting word-level rationales, we conceive rationales as selected paragraphs in multi-paragraph structured court cases. We also release a new dataset comprising European Court of Human Rights cases, including annotations for paragraph-level rationales. We use this dataset to study the effect of already proposed rationale constraints, i.e., sparsity, continuity, and comprehensiveness, formulated as regularizers. Our findings indicate that some of these constraints are not beneficial in paragraph-level rationale extraction, while others need re-formulation to better handle the multi-label nature of the task we consider. We also introduce a new constraint, singularity, which further improves the quality of rationales, even compared with noisy rationale supervision. Experimental results indicate that the newly introduced task is very challenging and there is a large scope for further research.

Omni LLMs can naturally locate key audio and visual evidence and integrate information across modalities for audiovisual reasoning. Existing audio-visual reasoning benchmarks that mainly rely on multiple-choice questions usually score models solely based on answer correctness, making it difficult to assess whether a model truly forms a reasoning chain from evidence to the correct answer. In addition, textual or unimodal shortcuts may allow models to answer correctly without analyzing the necessary audio-visual evidence, further weakening the reliability of evaluation benchmarks. To address these issues, we introduce EvaOmni, an evidence-grounded benchmark containing 392 videos and 1,027 purely human-annotated question-answer pairs, covering 15 capabilities and three levels: retrieval and localization, basic understanding, and complex reasoning. We design a series of rigorous annotation procedures in which experts annotate audio and visual evidence and construct video-based distractors; humans review whether questions meet strict inspection criteria, and finally models filter textual and unimodal shortcuts. Beyond high-quality question-answer pairs, we require models to simultaneously return the answer and supporting evidence intervals in order to evaluate the complete reasoning chain. We propose the Coverage Grounded Score (C.G.) to measure how much annotated evidence a model identifies, and the Bottleneck Grounded Score (B.G.) to penalize poor temporal localization in either modality; both are incorporated into answer correctness. Across 7 open-weight and 17 proprietary Omni LLMs, Gemini3.7-Flash achieved the highest overall accuracy (79.4%), C.G. (70.8%), and B.G. (35.1%). At Level 3, it reached 79.1% accuracy, but its B.G. was only 12.3%. These results show that even the most advanced models still have room for development in grounded audio-visual reasoning.

Understanding agent behavior requires methods that scale to thousands of trajectories and surface new patterns in long, often unfamiliar tasks where pre-built classifiers fall short. We propose to bring grounded theory into agent trajectory analysis: a six-decade-old qualitative method from the social sciences, with a principled saturation criterion and an auditable trail from data to theory. We propose AutoTraceGT (Automated Trace analysis through Grounded Theory), the first multi-agent pipeline that automates grounded theory on agent trajectories. It iteratively performs open, axial, and theoretical coding until saturation, producing a behavioral taxonomy tailored to each task. Across six trajectory corpora, AutoTraceGT produces codebooks that recover 73-91 percent of the failure modes in human-annotated taxonomies and surface additional patterns that those taxonomies miss. The emergent theoretical narrative aligns with prior expert accounts. Used as a deductive feature space, the codebook outperforms zero-shot and few-shot LLM baselines on downstream failure prediction. These results suggest Grounded Theory offers a scalable analytic tool for ML researchers and agent developers studying what agents actually do.

Large language models (LLMs) augmented with tools are emerging as autonomous agents capable of using Web engine, APIs, and code to solve complex, long-horizon tasks. Current tool-using benchmarks for information seeking on academic graphs rely on synthetic templates, simplified solution spaces, or narrow tasks such as paper-centric tasks, leaving key challenges underexplored - realistic user intent, complex multi-step API planning, rich parameter filling for APIs, grounded answers with references, and comprehensive evaluation of both the process and the outcome. We introduce AISE-Bench, a real-world, full-cycle annotated benchmark for information seeking on academic knowledge graphs. AISE-Bench release contains 1,133 QA pairs, including query taxonomies, full API execution trajectories, validated parameters, and source-grounded answers with reference links. To support high-quality annotation, we design a customized agent workflow to enable annotators to plan, execute, and revise complex API workflows efficiently. We develop a comprehensive evaluation protocol measuring answer quality, reference grounding, API-planning correctness, and execution success. Among the 14 evaluated methods, even the strongest model (PLAY2PROMPT with Gemini-3-Pro) achieves only moderate performance and often struggles with API planning and execution. AISE-Bench establishes a challenging new testbed for quantitatively evaluating and improving the stepwise correctness, grounded summarization, and traceable reasoning of multi-step API-using LLM agents. Our code and data are available at https://aise-bench.github.io/.

Recent advancements in large language models (LLM) capable of processing extremely long texts highlight the need for a dedicated evaluation benchmark to assess their long-context capabilities. However, existing methods, like the needle-in-a-haystack test, do not effectively assess whether these models fully utilize contextual information, raising concerns about the reliability of current evaluation techniques. To thoroughly examine the effectiveness of existing benchmarks, we introduce a new metric called information coverage (IC), which quantifies the proportion of the input context necessary for answering queries. Our findings indicate that current benchmarks exhibit low IC; although the input context may be extensive, the actual usable context is often limited. To address this, we present ETHIC, a novel benchmark designed to assess LLMs' ability to leverage the entire context. Our benchmark comprises 2,648 test instances spanning four long-context tasks with high IC scores in the domains of books, debates, medicine, and law. Our evaluations reveal significant performance drops in contemporary LLMs, highlighting a critical challenge in managing long contexts. Our benchmark is available at https://github.com/dmis-lab/ETHIC.

LLM-as-a-judge models have been used for evaluating both human and AI generated content, specifically by providing scores and rationales. Rationales, in addition to increasing transparency, help models learn to calibrate its judgments. Enhancing a model's rationale can therefore improve its calibration abilities and ultimately the ability to score content. We introduce Self-Rationalization, an iterative process of improving the rationales for the judge models, which consequently improves the score for fine-grained customizable scoring criteria (i.e., likert-scale scoring with arbitrary evaluation criteria). Self-rationalization works by having the model generate multiple judgments with rationales for the same input, curating a preference pair dataset from its own judgements, and iteratively fine-tuning the judge via DPO. Intuitively, this approach allows the judge model to self-improve by learning from its own rationales, leading to better alignment and evaluation accuracy. After just two iterations -- while only relying on examples in the training set -- human evaluation shows that our judge model learns to produce higher quality rationales, with a win rate of 62% on average compared to models just trained via SFT on rationale . This judge model also achieves high scoring accuracy on BigGen Bench and Reward Bench, outperforming even bigger sized models trained using SFT with rationale, self-consistency or best-of-N sampling by 3% to 9%.

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.

Neural language models (LMs) have achieved impressive results on various language-based reasoning tasks by utilizing latent knowledge encoded in their own pretrained parameters. To make this reasoning process more explicit, recent works retrieve a rationalizing LM's internal knowledge by training or prompting it to generate free-text rationales, which can be used to guide task predictions made by either the same LM or a separate reasoning LM. However, rationalizing LMs require expensive rationale annotation and/or computation, without any assurance that their generated rationales improve LM task performance or faithfully reflect LM decision-making. In this paper, we propose PINTO, an LM pipeline that rationalizes via prompt-based learning, and learns to faithfully reason over rationales via counterfactual regularization. First, PINTO maps out a suitable reasoning process for the task input by prompting a frozen rationalizing LM to generate a free-text rationale. Second, PINTO's reasoning LM is fine-tuned to solve the task using the generated rationale as context, while regularized to output less confident predictions when the rationale is perturbed. Across four datasets, we show that PINTO significantly improves the generalization ability of the reasoning LM, yielding higher performance on both in-distribution and out-of-distribution test sets. Also, we find that PINTO's rationales are more faithful to its task predictions than those generated by competitive baselines.

Large language models (LLMs) have shown significant potential in scientific disciplines such as biomedicine, particularly in hypothesis generation, where they can analyze vast literature, identify patterns, and suggest research directions. However, a key challenge lies in evaluating the truthfulness of generated hypotheses, as verifying their accuracy often requires substantial time and resources. Additionally, the hallucination problem in LLMs can lead to the generation of hypotheses that appear plausible but are ultimately incorrect, undermining their reliability. To facilitate the systematic study of these challenges, we introduce TruthHypo, a benchmark for assessing the capabilities of LLMs in generating truthful biomedical hypotheses, and KnowHD, a knowledge-based hallucination detector to evaluate how well hypotheses are grounded in existing knowledge. Our results show that LLMs struggle to generate truthful hypotheses. By analyzing hallucinations in reasoning steps, we demonstrate that the groundedness scores provided by KnowHD serve as an effective metric for filtering truthful hypotheses from the diverse outputs of LLMs. Human evaluations further validate the utility of KnowHD in identifying truthful hypotheses and accelerating scientific discovery. Our data and source code are available at https://github.com/Teddy-XiongGZ/TruthHypo.

Grounded text generation systems often generate text that contains factual inconsistencies, hindering their real-world applicability. Automatic factual consistency evaluation may help alleviate this limitation by accelerating evaluation cycles, filtering inconsistent outputs and augmenting training data. While attracting increasing attention, such evaluation metrics are usually developed and evaluated in silo for a single task or dataset, slowing their adoption. Moreover, previous meta-evaluation protocols focused on system-level correlations with human annotations, which leave the example-level accuracy of such metrics unclear. In this work, we introduce TRUE: a comprehensive survey and assessment of factual consistency metrics on a standardized collection of existing texts from diverse tasks, manually annotated for factual consistency. Our standardization enables an example-level meta-evaluation protocol that is more actionable and interpretable than previously reported correlations, yielding clearer quality measures. Across diverse state-of-the-art metrics and 11 datasets we find that large-scale NLI and question generation-and-answering-based approaches achieve strong and complementary results. We recommend those methods as a starting point for model and metric developers, and hope TRUE will foster progress towards even better evaluation methods.

Despite advances in scientific AI, a coherent framework for Scientific General Intelligence (SGI)-the ability to autonomously conceive, investigate, and reason across scientific domains-remains lacking. We present an operational SGI definition grounded in the Practical Inquiry Model (PIM: Deliberation, Conception, Action, Perception) and operationalize it via four scientist-aligned tasks: deep research, idea generation, dry/wet experiments, and experimental reasoning. SGI-Bench comprises over 1,000 expert-curated, cross-disciplinary samples inspired by Science's 125 Big Questions, enabling systematic evaluation of state-of-the-art LLMs. Results reveal gaps: low exact match (10--20%) in deep research despite step-level alignment; ideas lacking feasibility and detail; high code executability but low execution result accuracy in dry experiments; low sequence fidelity in wet protocols; and persistent multimodal comparative-reasoning challenges. We further introduce Test-Time Reinforcement Learning (TTRL), which optimizes retrieval-augmented novelty rewards at inference, enhancing hypothesis novelty without reference answer. Together, our PIM-grounded definition, workflow-centric benchmark, and empirical insights establish a foundation for AI systems that genuinely participate in scientific discovery.

Uncovering latent values and opinions in large language models (LLMs) can help identify biases and mitigate potential harm. Recently, this has been approached by presenting LLMs with survey questions and quantifying their stances towards morally and politically charged statements. However, the stances generated by LLMs can vary greatly depending on how they are prompted, and there are many ways to argue for or against a given position. In this work, we propose to address this by analysing a large and robust dataset of 156k LLM responses to the 62 propositions of the Political Compass Test (PCT) generated by 6 LLMs using 420 prompt variations. We perform coarse-grained analysis of their generated stances and fine-grained analysis of the plain text justifications for those stances. For fine-grained analysis, we propose to identify tropes in the responses: semantically similar phrases that are recurrent and consistent across different prompts, revealing patterns in the text that a given LLM is prone to produce. We find that demographic features added to prompts significantly affect outcomes on the PCT, reflecting bias, as well as disparities between the results of tests when eliciting closed-form vs. open domain responses. Additionally, patterns in the plain text rationales via tropes show that similar justifications are repeatedly generated across models and prompts even with disparate stances.

Large Language Models (LLMs) have demonstrated remarkable capabilities in performing complex tasks. Moreover, recent research has shown that incorporating human-annotated rationales (e.g., Chain-of-Thought prompting) during in-context learning can significantly enhance the performance of these models, particularly on tasks that require reasoning capabilities. However, incorporating such rationales poses challenges in terms of scalability as this requires a high degree of human involvement. In this work, we present a novel framework, Amplifying Model Performance by Leveraging In-Context Learning with Post Hoc Explanations (AMPLIFY), which addresses the aforementioned challenges by automating the process of rationale generation. To this end, we leverage post hoc explanation methods which output attribution scores (explanations) capturing the influence of each of the input features on model predictions. More specifically, we construct automated natural language rationales that embed insights from post hoc explanations to provide corrective signals to LLMs. Extensive experimentation with real-world datasets demonstrates that our framework, AMPLIFY, leads to prediction accuracy improvements of about 10-25% over a wide range of tasks, including those where prior approaches which rely on human-annotated rationales such as Chain-of-Thought prompting fall short. Our work makes one of the first attempts at highlighting the potential of post hoc explanations as valuable tools for enhancing the effectiveness of LLMs. Furthermore, we conduct additional empirical analyses and ablation studies to demonstrate the impact of each of the components of AMPLIFY, which, in turn, leads to critical insights for refining in-context learning.

Saliency maps are widely used for visual explanations in deep learning, but a fundamental lack of consensus persists regarding their intended purpose and alignment with diverse user queries. This ambiguity hinders the effective evaluation and practical utility of explanation methods. We address this gap by introducing the Reference-Frame times Granularity (RFxG) taxonomy, a principled conceptual framework that organizes saliency explanations along two essential axes:Reference-Frame: Distinguishing between pointwise ("Why this prediction?") and contrastive ("Why this and not an alternative?") explanations. Granularity: Ranging from fine-grained class-level (e.g., "Why Husky?") to coarse-grained group-level (e.g., "Why Dog?") interpretations. Using the RFxG lens, we demonstrate critical limitations in existing evaluation metrics, which overwhelmingly prioritize pointwise faithfulness while neglecting contrastive reasoning and semantic granularity. To systematically assess explanation quality across both RFxG dimensions, we propose four novel faithfulness metrics. Our comprehensive evaluation framework applies these metrics to ten state-of-the-art saliency methods, four model architectures, and three datasets. By advocating a shift toward user-intent-driven evaluation, our work provides both the conceptual foundation and the practical tools necessary to develop visual explanations that are not only faithful to the underlying model behavior but are also meaningfully aligned with the complexity of human understanding and inquiry.

Large language model (LLM) agents may perform well on isolated tasks yet drift into inconsistency over extended interaction. We evaluate temporal consistency in a controlled 20-step multi-agent setting inspired by delayed-gratification studies. At each step, an agent chooses between continuing to delay a reward or claiming it immediately (terminating the episode). Across a full-factorial manipulation of social visibility (private vs public), persona stressors, and deliberation policy, we run 84,540 trajectories spanning 8 model families. Treating the first reward-claim as a time-to-event outcome, we estimate Kaplan-Meier survival curves and fit discrete-time hazard regression to quantify how experimental factors shift failure risk over time. Then, to analyze rationales and language patterns associated with failure, we build a seven-category taxonomy from 13,780 deliberation traces from agents who choose to terminate the episode, using an LLM-assisted labeling paired with human audit (κ=0.83). Rationale profiles change systematically with time and context: early failures are more impulse-driven, later failures more fatigue- and cost-benefit-framed, while public settings increase norm-oriented justifications. We also find a deliberation-inconsistency association: among failures, longer deliberation correlates with higher rates of intra-rationale contradiction (simultaneous pro-delay and pro-claim statements), challenging the assumption that more reasoning text implies greater consistency. Together, the survival and rationale analyses reveal distinct temporal reliability regimes and model-specific "failure fingerprints", offering an evaluation lens for diagnosing inconsistency in multi-turn agent behavior.

Generating professional scholarly content, such as peer reviews and rebuttals, requires an intricate synergy between domain reasoning and factual grounding. This work presents a comprehensive framework for the development and evaluation of specialized scholarly agents, InternReviewer and InternAdvocate. We first establish a large-scale, high-quality scholarly dataset and integrate a high-efficiency arXiv retrieval tool to enable active evidence gathering. To optimize these agents, we implement an agentic Reinforcement Learning (RL) paradigm driven by a unified objective metric and reward system. This system avoids the biases of subjective model-based judging by employing multi-dimensional criteria, including reference-anchored semantic alignment, structural compliance, and a strict verification mechanism that cross-checks citations against real-time interaction logs to eliminate hallucinations. Experimental results demonstrate that agents trained within this closed-loop framework exhibit significant improvements in reasoning depth and citation accuracy.

Correct answers do not necessarily reflect cultural understanding. We introduce CRaFT, an explanation-based multilingual evaluation framework designed to assess how large language models (LLMs) reason across cultural contexts. Rather than scoring outputs solely based on accuracy, CRaFT evaluates model explanations using four interpretable metrics: Cultural Fluency, Deviation, Consistency, and Linguistic Adaptation. We apply the framework to 50 culturally grounded questions from the World Values Survey, translated into Arabic, Bengali, and Spanish, and evaluate three models (GPT, DeepSeek, and FANAR) across over 2,100 answer-explanation pairs. Results reveal significant cross-lingual variation in reasoning: Arabic reduces fluency, Bengali enhances it, and Spanish remains largely stable. While GPT adapts more effectively across languages, it exhibits lower consistency; FANAR shows stable but rigid reasoning. These findings suggest that cultural awareness in LLMs is not intrinsic but emerges through linguistic framing. CRaFT offers a new lens for evaluating cross-cultural reasoning in multilingual settings, providing actionable insights for building culturally adaptive language models.

Many reasoning tasks require models to reason over input context, from document-grounded question answering to rule-based deduction. Chain-of-Thought (CoT) prompting produces traces that appear transparent, yet individual steps can silently deviate from the source evidence, even when the final answer is correct. Existing methods detect hallucinations at the response level but fail to identify where in the chain a failure occurs or what type it is. We introduce GRACE, the first human-annotated step-level faithfulness benchmark with a data-driven error taxonomy for context-grounded textual reasoning. GRACE covers CoT traces from 10 models across 4 source datasets, with each step annotated for faithfulness, error category, and natural language explanation. A data-driven taxonomy, discovered bottom-up via unsupervised clustering, organizes failures into two tracks: GRACE-Inference (deductive errors) and GRACE-Grounding (factual grounding errors), with four categories each. The evaluation set is human-annotated and challenging by design. Our experiments reveal substantial headroom for current models. In addition, integrating step-level faithfulness signals into reinforcement learning pipelines improves both downstream accuracy and reasoning reliability.

Process Reward Models (PRMs) aim to improve multi-step reasoning in Large Language Models (LLMs) by supervising intermediate steps and identifying errors. However, building effective PRMs remains challenging due to the lack of scalable, high-quality annotations. Existing approaches rely on costly human labeling, LLM-based self-evaluation that is prone to hallucination, or Monte Carlo (MC) estimation, which infers step quality solely from rollout outcomes and often introduces noisy, misaligned supervision due to credit misattribution. These issues result in three core limitations: noisy rewards, low factual fidelity, and misalignment with step-level reasoning objectives. To address these challenges, we introduce GroundedPRM, a tree-guided and fidelity-aware framework for automatic process supervision. To reduce reward noise and enable fine-grained credit assignment, we construct structured reasoning paths via Monte Carlo Tree Search (MCTS). To eliminate hallucinated supervision, we validate each intermediate step using an external tool, providing execution-grounded correctness signals. To combine both step-level validation and global outcome assessment, we design a hybrid reward aggregation mechanism that fuses tool-based verification with MCTS-derived feedback. Finally, we format the reward signal into a rationale-enhanced, generative structure to promote interpretability and compatibility with instruction-tuned LLMs. GroundedPRM is trained on only 40K automatically labeled samples, amounting to just 10% of the data used by the best-performing PRM trained with auto-labeled supervision. Nevertheless, it achieves up to a 26% relative improvement in average performance on ProcessBench. When used for reward-guided greedy search, GroundedPRM outperforms even PRMs trained with human-labeled supervision, offering a scalable and verifiable path toward high-quality process-level reasoning.

Causal reasoning is viewed as crucial for achieving human-level machine intelligence. Recent advances in language models have expanded the horizons of artificial intelligence across various domains, sparking inquiries into their potential for causal reasoning. In this work, we introduce Causal evaluation of Language Models (CaLM), which, to the best of our knowledge, is the first comprehensive benchmark for evaluating the causal reasoning capabilities of language models. First, we propose the CaLM framework, which establishes a foundational taxonomy consisting of four modules: causal target (i.e., what to evaluate), adaptation (i.e., how to obtain the results), metric (i.e., how to measure the results), and error (i.e., how to analyze the bad results). This taxonomy defines a broad evaluation design space while systematically selecting criteria and priorities. Second, we compose the CaLM dataset, comprising 126,334 data samples, to provide curated sets of causal targets, adaptations, metrics, and errors, offering extensive coverage for diverse research pursuits. Third, we conduct an extensive evaluation of 28 leading language models on a core set of 92 causal targets, 9 adaptations, 7 metrics, and 12 error types. Fourth, we perform detailed analyses of the evaluation results across various dimensions (e.g., adaptation, scale). Fifth, we present 50 high-level empirical findings across 9 dimensions (e.g., model), providing valuable guidance for future language model development. Finally, we develop a multifaceted platform, including a website, leaderboards, datasets, and toolkits, to support scalable and adaptable assessments. We envision CaLM as an ever-evolving benchmark for the community, systematically updated with new causal targets, adaptations, models, metrics, and error types to reflect ongoing research advancements. Project website is at https://opencausalab.github.io/CaLM.

Recent progress in reasoning models suggests that generating plausible attempts for research-level mathematics may be within reach, but verification remains a bottleneck, consuming scarce expert time. We hypothesize that a meaningful solution should contain enough method-level information that, when applied to a neighborhood of related questions, it should yield better downstream performance than incorrect solutions. Building on this idea, we propose Consequence-Based Utility, an oracle-free evaluator that scores each candidate by testing its value as an in-context exemplar in solving related yet verifiable questions. Our approach is evaluated on an original set of research-level math problems, each paired with one expert-written solution and nine LLM-generated solutions. Notably, Consequence-Based Utility consistently outperforms reward models, generative reward models, and LLM judges on ranking quality. Specifically, for GPT-OSS-120B, it improves Acc@1 from 67.2 to 76.3 and AUC from 71.4 to 79.6, with similarly large AUC gains on GPT-OSS-20B (69.0 to 79.2). Furthermore, compared to LLM-Judges, it also exhibits a larger solver-evaluator gap, maintaining a stronger correct-wrong separation even on instances where the underlying solver often fails to solve.

Evaluating production LLM responses and routing requests across providers in LLM gateways requires fine-grained quality signals and operationally grounded decisions. To address this gap, we present SEAR, a schema-based evaluation and routing system for multi-model, multi-provider LLM gateways. SEAR defines an extensible relational schema covering both LLM evaluation signals (context, intent, response characteristics, issue attribution, and quality scores) and gateway operational metrics (latency, cost, throughput), with cross-table consistency links across around one hundred typed, SQL-queryable columns. To populate the evaluation signals reliably, SEAR proposes self-contained signal instructions, in-schema reasoning, and multi-stage generation that produces database-ready structured outputs. Because signals are derived through LLM reasoning rather than shallow classifiers, SEAR captures complex request semantics, enables human-interpretable routing explanations, and unifies evaluation and routing in a single query layer. Across thousands of production sessions, SEAR achieves strong signal accuracy on human-labeled data and supports practical routing decisions, including large cost reductions with comparable quality.

Large language models (LLMs) are increasingly applied to scientific research, yet prevailing science benchmarks probe decontextualized knowledge and overlook the iterative reasoning, hypothesis generation, and observation interpretation that drive scientific discovery. We introduce a scenario-grounded benchmark that evaluates LLMs across biology, chemistry, materials, and physics, where domain experts define research projects of genuine interest and decompose them into modular research scenarios from which vetted questions are sampled. The framework assesses models at two levels: (i) question-level accuracy on scenario-tied items and (ii) project-level performance, where models must propose testable hypotheses, design simulations or experiments, and interpret results. Applying this two-phase scientific discovery evaluation (SDE) framework to state-of-the-art LLMs reveals a consistent performance gap relative to general science benchmarks, diminishing return of scaling up model sizes and reasoning, and systematic weaknesses shared across top-tier models from different providers. Large performance variation in research scenarios leads to changing choices of the best performing model on scientific discovery projects evaluated, suggesting all current LLMs are distant to general scientific "superintelligence". Nevertheless, LLMs already demonstrate promise in a great variety of scientific discovery projects, including cases where constituent scenario scores are low, highlighting the role of guided exploration and serendipity in discovery. This SDE framework offers a reproducible benchmark for discovery-relevant evaluation of LLMs and charts practical paths to advance their development toward scientific discovery.

Reasoning has emerged as the next major frontier for language models (LMs), with rapid advances from both academic and industrial labs. However, this progress often outpaces methodological rigor, with many evaluations relying on benchmarking practices that lack transparency, robustness, or statistical grounding. In this work, we conduct a comprehensive empirical study and find that current mathematical reasoning benchmarks are highly sensitive to subtle implementation choices - including decoding parameters, random seeds, prompt formatting, and even hardware and software-framework configurations. Performance gains reported in recent studies frequently hinge on unclear comparisons or unreported sources of variance. To address these issues, we propose a standardized evaluation framework with clearly defined best practices and reporting standards. Using this framework, we reassess recent methods and find that reinforcement learning (RL) approaches yield only modest improvements - far below prior claims - and are prone to overfitting, especially on small-scale benchmarks like AIME24. In contrast, supervised finetuning (SFT) methods show consistently stronger generalization. To foster reproducibility, we release all code, prompts, and model outputs, for reasoning benchmarks, establishing more rigorous foundations for future work.

This paper introduces a document grounded dataset for text conversations. We define "Document Grounded Conversations" as conversations that are about the contents of a specified document. In this dataset the specified documents were Wikipedia articles about popular movies. The dataset contains 4112 conversations with an average of 21.43 turns per conversation. This positions this dataset to not only provide a relevant chat history while generating responses but also provide a source of information that the models could use. We describe two neural architectures that provide benchmark performance on the task of generating the next response. We also evaluate our models for engagement and fluency, and find that the information from the document helps in generating more engaging and fluent responses.

On-policy reinforcement learning has become a central paradigm for improving the reasoning abilities of large language models. However, its effectiveness is often limited by reward sparsity: when a model fails to discover correct trajectories for difficult problems, the optimization process receives little useful signal and may stagnate. Existing approaches mitigate this issue by incorporating off-policy demonstrations, expert traces, or model-generated solutions, but they typically require the auxiliary data to match the format of the reinforcement-learning task, often relying on rejection sampling from stronger models to obtain suitable training trajectories. We introduce Rationale-Guided Policy Optimization (RGPO), a framework that adaptively leverages ground-truth rationale information according to the model's current capability while preserving its freedom to explore. Rather than treating reference solutions as fixed imitation targets, RGPO uses them as temporary scaffolds: rationales help the model generate improved responses, after which only higher-reward, model-generated solutions are transferred back to the original unguided setting. This design allows training to exploit available ground-truth information without requiring off-policy data to follow the same format as the RL task. Across both language-only and vision-language reasoning settings, RGPO consistently improves performance over RLVR baselines, and ablation studies show that adaptive rationale guidance is a key contributor to these gains. These results suggest that RGPO offers a practical and general approach for reducing reward sparsity, stabilizing reinforcement learning, and improving reasoning performance in both text-only and multimodal models.

Evaluating the quality and variability of text generated by Large Language Models (LLMs) poses a significant, yet unresolved research challenge. Traditional evaluation methods, such as ROUGE and BERTScore, which measure token similarity, often fail to capture the holistic semantic equivalence. This results in a low correlation with human judgments and intuition, which is especially problematic in high-stakes applications like healthcare and finance where reliability, safety, and robust decision-making are highly critical. This work proposes DCR, an automated framework for evaluating and improving the consistency of LLM-generated texts using a divide-conquer-reasoning approach. Unlike existing LLM-based evaluators that operate at the paragraph level, our method employs a divide-and-conquer evaluator (DCE) that breaks down the paragraph-to-paragraph comparison between two generated responses into individual sentence-to-paragraph comparisons, each evaluated based on predefined criteria. To facilitate this approach, we introduce an automatic metric converter (AMC) that translates the output from DCE into an interpretable numeric score. Beyond the consistency evaluation, we further present a reason-assisted improver (RAI) that leverages the analytical reasons with explanations identified by DCE to generate new responses aimed at reducing these inconsistencies. Through comprehensive and systematic empirical analysis, we show that our approach outperforms state-of-the-art methods by a large margin (e.g., +19.3% and +24.3% on the SummEval dataset) in evaluating the consistency of LLM generation across multiple benchmarks in semantic, factual, and summarization consistency tasks. Our approach also substantially reduces nearly 90% of output inconsistencies, showing promise for effective hallucination mitigation.

We propose T2I-ReasonBench, a benchmark evaluating reasoning capabilities of text-to-image (T2I) models. It consists of four dimensions: Idiom Interpretation, Textual Image Design, Entity-Reasoning and Scientific-Reasoning. We propose a two-stage evaluation protocol to assess the reasoning accuracy and image quality. We benchmark various T2I generation models, and provide comprehensive analysis on their performances.

Single-image, multi-image, and video deep research require different visual operations but share a workflow of visual grounding, external retrieval, and fact composition. A key challenge is to preserve the dependencies linking localized visual anchors, entity relations, source-supported facts, and answer-producing operations. We introduce OneSearch-VL, a unified agent centered on the Visually Grounded Evidence Graph (VGEG), which encodes these dependencies as a shared task-level reference for data construction, process supervision, and operation-level evaluation. Our VGEG-based data engine constructs and verifies multi-image and video questions and filters expert trajectories. Using these data, we assemble OneSearch-VL-SFT-110K and OneSearch-VL-RL-10K for SFT and RL, respectively. We further derive the Evidence-aware Visual-Grounded Rubric reward (EVGR) from VGEG annotations to supervise evidence traceability and visual grounding during RL. For fine-grained evaluation, we construct OneSearch-MI-Bench and OneSearch-Video-Bench, organizing questions by the research operations encoded in their VGEGs. Experiments show that OneSearch-VL-8B improves over Qwen3-VL-8B with tool access by 20.2 and 17.6 percentage points on the two new benchmarks, respectively, while also achieving substantial gains across 7 image benchmarks and VideoDR. Project repository: https://github.com/appletea233/OneSearch-VL

Answering complex questions over long documents frequently requires integrating evidence that the source itself disperses naturally across distant passages. In an incident report, the operating condition, design flaw, and missed safety check that jointly explain a disaster may appear dozens of sections apart; in a novel, a character's true motive may surface only through scenes far removed from the moment it becomes relevant. This source-internal evidence integration is central to real-world long-document analysis, yet existing benchmarks largely sidestep it. Needle probes, planted facts, and reverse-engineered multi-hop chains embed evidence that may differ from the host text in distribution, placement, or register, making it unclear whether strong performance reflects genuine source reasoning or distributional artifacts. We introduce WILDTRACE, a benchmark of 481 tasks over 214 naturally occurring long-form sources such as technical incident reports and lesser-known literary narratives, where all evidence trails arise from the document's own causal, temporal, and narrative logic. Drawing on Pearl's causal hierarchy and prior multi-hop reasoning typologies, we define seven source-internal evidence geometries that characterize the distinct relational demands of analytical reading in long documents. A source-first construction pipeline mines candidate trails from document structure before writing questions; each item then undergoes multi-stage validation covering clue necessity, answer groundedness, rubric fidelity, contamination resistance and answerability. As models are increasingly entrusted with real-world high-stakes analytical tasks, this gap between accessing information and reasoning over naturally dispersed evidence emerges as a defining challenge for the next stage of long-context research.

When retrieval-augmented generation (RAG) systems hallucinate, what geometric trace does this leave in embedding space? We introduce the Semantic Grounding Index (SGI), defined as the ratio of angular distances from the response to the question versus the context on the unit hypersphere S^{d-1}.Our central finding is semantic laziness: hallucinated responses remain angularly proximate to questions rather than departing toward retrieved contexts. On HaluEval (n=5,000), we observe large effect sizes (Cohen's d ranging from 0.92 to 1.28) across five embedding models with mean cross-model correlation r=0.85. Crucially, we derive from the spherical triangle inequality that SGI's discriminative power should increase with question-context angular separation θ(q,c)-a theoretical prediction confirmed empirically: effect size rises monotonically from d=0.61 -low θ(q,c), to d=1.27 -high θ(q,c), with AUC improving from 0.72 to 0.83. Subgroup analysis reveals that SGI excels on long responses (d=2.05) and short questions (d=1.22), while remaining robust across context lengths. Calibration analysis yields ECE=0.10, indicating SGI scores can serve as probability estimates, not merely rankings. A critical negative result on TruthfulQA (AUC=0.478) establishes that angular geometry measures topical engagement rather than factual accuracy. SGI provides computationally efficient, theoretically grounded infrastructure for identifying responses that warrant verification in production RAG deployments.

In this paper, we offer a guide for researchers on evaluating reasoning in language models, building the case that reasoning should be assessed through evidence of adaptive, multi-step search rather than final-answer accuracy alone. Under an evaluation-oriented definition, reasoning requires selecting intermediate steps and halting according to input-dependent conditions, which we formalize as a search-like procedure. We show that single forward passes in scalable architectures are structurally limited in their ability to realize such variable-depth computation, motivating intermediate decoding and externalized reasoning traces as appropriate evaluation interfaces. Central to our argument is that final-answer accuracy alone is an insufficient measure of reasoning, because it provides little ability to diagnose or debug the underlying processes that produce individual solutions in frontier models. We therefore argue for a shift toward process-based evaluation, in which reasoning is assessed through the faithfulness and validity of intermediate reasoning traces as first-class evaluation targets.

Sequential recommendation systems aim to capture users' evolving preferences from their interaction histories. Recent reasoningenhanced methods have shown promise by introducing deliberate, chain-of-thought-like processes with intermediate reasoning steps. However, these methods rely solely on the next target item as supervision, leading to two critical issues: (1) reasoning instability--the process becomes overly sensitive to recent behaviors and spurious interactions like accidental clicks, and (2) surface-level reasoning--the model memorizes item-to-item transitions rather than understanding intrinsic behavior patterns. To address these challenges, we propose IGR-SR, an Intent-Guided Reasoning framework for Sequential Recommendation that anchors the reasoning process to explicitly extracted high-level intents. Our framework comprises three key components: (1) a Latent Intent Distiller (LID) that efficiently extracts multi-faceted intents using a frozen encoder with learnable tokens, (2) an Intent-aware Deliberative Reasoner (IDR) that decouples reasoning into intent deliberation and decision-making via a dual-attention architecture, and (3) an Intent Consistency Regularization (ICR) that ensures robustness by enforcing consistent representations across different intent views. Extensive experiments on three public datasets demonstrate that IGR-SR achieves an average 7.13% improvement over state-of-the-art baselines. Critically, under 20% behavioral noise, IGR-SR degrades only 10.4% compared to 16.2% and 18.6% for competing methods, validating the effectiveness and robustness of intent-guided reasoning.

With the advent of Large Language Models (LLMs), general-purpose agents have seen fundamental advancements. However, evaluating these agents presents unique challenges that distinguish them from static QA benchmarks. We observe that current agent benchmarks are heavily confounded by extraneous factors, including system prompts, toolset configurations, and environmental dynamics. Existing evaluations often rely on fragmented, researcher-specific frameworks where the prompt engineering for reasoning and tool usage varies significantly, making it difficult to attribute performance gains to the model itself. Additionally, the lack of standardized environmental data leads to untraceable errors and non-reproducible results. This lack of standardization introduces substantial unfairness and opacity into the field. We propose that a unified evaluation framework is essential for the rigorous advancement of agent evaluation. To this end, we introduce a proposal aimed at standardizing agent evaluation.

The development of large language models has instilled widespread interest among the researchers to understand their inherent reasoning and problem-solving capabilities. Despite good amount of research going on to elucidate these capabilities, there is a still an appreciable gap in understanding moral development and judgments of these models. The current approaches of evaluating the ethical reasoning abilities of these models as a classification task pose numerous inaccuracies because of over-simplification. In this study, we built a psychological connection by bridging two disparate fields-human psychology and AI. We proposed an effective evaluation framework which can help to delineate the model's ethical reasoning ability in terms of moral consistency and Kohlberg's moral development stages with the help of Psychometric Assessment Tool-Defining Issues Test.

Multimodal large language models (MLLMs) are increasingly used for real-world tasks involving multi-step reasoning and long-form generation, where reliability requires grounding model outputs in heterogeneous input sources and verifying individual factual claims. However, existing multimodal grounding benchmarks and evaluation methods focus on simplified, observation-based scenarios or limited modalities and fail to assess attribution in complex multimodal reasoning. We introduce MuRGAt (Multimodal Reasoning with Grounded Attribution), a benchmark for evaluating fact-level multimodal attribution in settings that require reasoning beyond direct observation. Given inputs spanning video, audio, and other modalities, MuRGAt requires models to generate answers with explicit reasoning and precise citations, where each citation specifies both modality and temporal segments. To enable reliable assessment, we introduce an automatic evaluation framework that strongly correlates with human judgments. Benchmarking with human and automated scores reveals that even strong MLLMs frequently hallucinate citations despite correct reasoning. Moreover, we observe a key trade-off: increasing reasoning depth or enforcing structured grounding often degrades accuracy, highlighting a significant gap between internal reasoning and verifiable attribution.