Search research

Sparse autoencoders (SAEs) are increasingly used to extract interpretable concepts from vision and vision language models, yet existing evaluation methods largely rely on proxy metrics or qualitative inspection rather than measuring semantic correspondence. We present a human-grounded evaluation framework that quantifies alignment between SAE latents and human-annotated concepts, without requiring user studies, and validate this matching through targeted attribute perturbations. To enable this intervention-style evaluation in vision, we construct synCUB and synCOCO, synthetic benchmarks of paired images that differ in exactly one attribute. We introduce Fully-Binary Matching Pursuit (FBMP), a coalition-based matching procedure that supports many-to-one mappings between SAE latents and annotated concepts, and consistently outperforms one-to-one baselines. For functional validation, we propose a Targeted Attribute Perturbation Alignment Score (TAPAScore), which tests whether matched concepts respond selectively and in the expected direction under targeted image-level attribute perturbations. Under sanity checks, our matching and TAPAScore are the only evaluated metrics that reliably distinguish trained SAEs from untrained ones. Across SAEs trained on CLIP and DINOv2 embeddings, we find that increased overcompleteness can reduce perturbation alignment, indicating a reduction in interpretability. Our evaluation framework suggests that moderate dictionary sizes provide the best trade-off, yielding the most interpretable SAEs. Code and datasets are available at https://github.com/JonasKlotz/sae-concept-eval.

We investigate the translation quality of current large language models (LLMs) for English-to-Hausa and English-to-Fongbe - two typologically distinct West African languages from the Afroasiatic and Niger-Congo families respectively - and evaluate whether standard automatic metrics reliably reflect human judgment for these low-resource languages. We evaluate four models (GPT-4o Mini, Claude Sonnet 4, Gemini 2.5 Flash, and Qwen2.5-7B) at progressive scales (500 to 10,000 sentences) using automatic metrics (BLEU, chrF++, TER, COMET, BERTScore) validated against native-speaker judgment. Our results reveal three key findings. First, translation quality varies substantially by language: Hausa achieves acceptable quality (human scores 4.0-4.5/5) while Fongbe achieves poor quality (1.0-2.2/5), with a consistent 3x BLEU gap across all systems. Second, model rankings differ by language - Gemini leads for Fongbe while GPT-4o leads for Hausa by human evaluation - indicating that performance on one low-resource African language does not predict performance on another. Third, metric-human correlation varies dramatically: perfect rank correlation for Fongbe (rho=1.0) but weak correlation for Hausa (rho=0.5), where human evaluators preferred GPT-4o despite all automatic metrics ranking Claude first. We further show that neural metrics like BERTScore exhibit embedding collapse (within-language similarity >0.99) for both languages, limiting their ability to differentiate translation quality. Based on these findings, we recommend multi-metric evaluation for low-resource African languages, with particular caution when interpreting neural metrics. We establish that minimum sample sizes of n=2,500 sentences are required for stable system rankings, as smaller samples produced artifact findings that reversed at scale.

Reliable benchmarking is critical for developing machine learning models for tandem mass spectrometry (MS/MS) based molecule discovery. Subtle issues in experimental design and model evaluation procedures can degrade the trustworthiness of such benchmarks and lead to erroneous conclusions. We conduct a thorough review of model evaluation issues in the recent MS/MS machine learning literature, using the standard MassSpecGym benchmark suite as a case study to illustrate the impact of these issues. We find evaluation issues in at least 17 of 26 papers reporting MassSpecGym benchmark results in the first year of its adoption. We isolate three classes of failures: (i) data leakage, (ii) shortcut learning, and (iii) implementation bugs and metric divergence. Through extensive experimentation and code replication, we quantify the impact of these issues and show how they corrupt the evaluation standards MassSpecGym was designed to enforce. We distill our findings into recommendations generalizable to MS/MS challenges, benchmarks, and custom evaluation setups. We also release MassSpecGym v1.5, an implementation of our recommendations in the MassSpecGym benchmarking suite which addresses the failure modes identified in this audit. MassSpecGym v1.5 is publicly available at https://github.com/pluskal-lab/MassSpecGym.

Agent skills -- structured, reusable knowledge artifacts that augment LLM agent capabilities -- have been rapidly adopted in industry, yet their cross-domain impact and use across commercial and open-source models remain under-studied, and no reusable methodology exists for evaluating an individual skill. In this work, we present an evaluation framework that lets a skill author construct realistic tasks to rigorously assess the aspects of a skill that matter most to them, and that estimates skill utility by solving those tasks. Further, we apply our evaluation approach at scale to 500 real-world skills, generating 1,000 tasks derived from the skills' content, along with instruction-following and goal-completion scoring rubrics. Using these metrics, we evaluate how 19 agent-model configurations, both proprietary and open-source, perform on the tasks. Our results show that models vary widely in how closely they adhere to the instructions encoded in skills, leading to substantial differences in their performance gains. Furthermore, we show that access to a skill significantly changes model behavior compared to the no-skill setup, providing an essential mechanism for encoding opinionated workflows into LLM agents. We release our evaluation dataset to support future work on agent skills.

LLMs have recently achieved strong results on formal proving benchmarks. However, existing evaluations remain heavily concentrated on competition-style problems and often fail to capture how models behave on longer, more dependency-rich mathematical developments. We introduce TheoremBench, a Lean4 benchmark designed to evaluate theorem provers beyond contest settings. The benchmark is built from nearly one hundred classical theorems and is released in two complementary forms: a plain main version containing one target theorem per instance, and a premised version that expands each theorem into a structured family of related proving tasks consisting of the main theorem together with automatically extracted supporting subtheorems. This design enables evaluation of not only whether the final theorem was proved from scratch, but also of partial progress through the internal proof structure of a theorem. Our experiments show that explicit premises substantially improve performance for Lean4-capable prover models. To provide a comprehensive evaluation, we introduce theorem-level coverage and token-efficiency metrics that expose qualitative differences in proof behavior. The results show that current provers remain strongly biased toward easy subtheorems and often solve theorems through long and inefficient tactic traces rather than compact proof plans. TheoremBench therefore provides a more fine-grained view of formal reasoning ability and highlights the importance of structural benchmark design for evaluating Lean4 theorem provers.

Evaluating LLM-powered interactive social agents is challenging because socially relevant behaviors depend not only on isolated outputs, but also on prior interactions, social roles, and downstream actions. Existing methods typically allow a target agent to act freely in an environment and then score the resulting trajectory. However, this passive setup can miss capabilities that only become observable under specific social circumstances; for example, conflict handling may remain untested if no disagreement arises. We propose Online Agent-as-a-Judge, a situation-generating evaluation framework for interactive social agents. Online Agent-as-a-Judge deploys an in-world evaluator agent that interacts with the target agent through the environment's native dialogue and action protocol, actively eliciting situations relevant to the evaluation criteria. The resulting trajectories provide evidence for assessing both immediate responses and subsequent behavior. In a life-simulation environment with $32$ designer-authored social criteria, Online Agent-as-a-Judge improves criteria coverage and agreement with human labels, yielding more reliable evidence-grounded evaluations of behaviors that passive methods can leave unobserved.

While household robots are often evaluated based on task completion, everyday domestic environments involve value-conflicting situations in which robots are expected to choose actions that prioritize other values than task success, such as human autonomy, efficiency, or social appropriateness. Yet, there are no benchmarks for evaluating robots' value preferences in such scenarios. We introduce RobotValues, a benchmark to evaluate household robot planners in 10K value-conflict scenarios. Each instance consists of a realistic household image with multiple plausible robot actions that prioritize different human values. We construct RobotValues through LLM-assisted scenario generation, stakeholder-grounded value extraction, image generation and automatic quality control. Using RobotValues we evaluate VLMs used in robotics and find that models exhibit default value preferences, including safety and accommodation, while underselecting privacy-prioritizing actions. When the models are instructed to prioritize specific values that conflict with their own preferences, they often fail to override their default actions, choosing incorrect actions for 80% of the time. These findings suggest that household robot evaluation should measure not only task completion or safety compliance, but also whether robots can choose among plausible actions when human values conflict.

Natural human conversation is full-duplex and audio-visual: people simultaneously speak and listen while continuously interpreting and producing nonverbal cues, such as nods, smiles, and gestures. To support successful human-agent interaction, agents must model full-duplex audiovisual conversation; however, existing full-duplex benchmarks evaluate only speech. In this work, we present VideoFDB, the first benchmark to evaluate full-duplex audio-visual-to-audio-visual (AV2AV) conversational agents. VideoFDB contributes (i) 237 dyadic clips spanning 11 nonverbal conversational dynamics from real-world video calls, (ii) a taxonomy separating perception from generation behaviors, and (iii) a rubric-based LM-as-judge evaluation framework with interpretable axes for assessing conversational quality with respect to nonverbal conversational dynamics. Across open- and closed-source vision-speech agents, we find systematic failure modes: captioning collapse and visual-stream ignorance, and we show that current systems exploit vision for explicit visual question answering but not for the streaming joint audiovisual grounding required in natural conversation. We further evaluate cascaded speech-to-avatar systems and find that their architecture fundamentally precludes the production of full-duplex nonverbal cues. As the first benchmark for full-duplex AV2AV interaction, VideoFDB establishes a foundation for systematic evaluation and, we hope, will accelerate the advancement and development of next-generation multimodal conversational agents.

While Membership Inference Attacks (MIAs) are the prevailing method for identifying training data, their application has expanded into privacy auditing and machine unlearning. Nevertheless, the field lacks a systematic framework for evaluating how different contexts affect MIA efficacy. Without such a characterization, practitioners risk deploying algorithms that perform well on benchmarks but become statistically irrelevant when faced with the nuances of specific, real-world datasets. To bridge this gap and provide actionable insights, we introduce a comprehensive evaluation framework that systematically characterizes privacy risks across the entire machine learning pipeline, spanning data, architectures, algorithms, and post-training modules. Designed to inherently capture diverse operational contexts, our framework rigorously evaluates state-of-the-art MIAs across a broad spectrum of training configurations. To account for varying misclassification costs in real-world deployments, we employ three complementary metrics: Balanced Accuracy for symmetric costs, alongside TPR at low FPR (or TNR at low FNR) for asymmetric scenarios where false alarms or missed detections are strictly penalized. Furthermore, recognizing that existing MIAs assume divergent adversary capabilities, we formalize two standardized threat models and adapt these attacks into corresponding variants to ensure an equitable benchmark. Extensive empirical evaluations demonstrate that the efficacy of specific MIA methodologies is highly sensitive to the assumed threat models and chosen evaluation metrics. Ultimately, we distill these findings into actionable guidelines and provide a ready-to-use auditing toolkit, empowering practitioners to conduct better privacy assessments.

Evaluating foundation models under appropriate adaptation settings is essential for understanding the quality and transferability of the learned representations. Recent EEG foundation models have demonstrated promising transfer capabilities across tasks and datasets, motivating their growing use in neurotechnology and clinical applications. However, these models are typically evaluated under full fine-tuning on well-curated downstream datasets, a setting that does not reflect biomedical domain constraints such as limited labeled data, reduced sensor coverage, or parameter-efficient adaptation. In this work, we propose a multi-dimensional evaluation framework for assessing EEG models under realistic low-resource conditions. Empirical analysis of both supervised EEG models and recent EEG foundation models, including LaBraM, CSBrain, and CBraMod, across 6 different datasets is performed under the proposed multi-dimensional evaluation framework. We find that EEG foundation models consistently provide performance gains on long-context tasks such as sleep stage prediction and mental health state classification. In contrast, for short-window Brain Computer Interface style tasks, supervised models achieve comparable despite having substantially fewer parameters. Additional analyses demonstrate that current foundation models provide limited robustness to short-window tasks and channel constrained settings. Together, these findings motivate the use of multi-dimensional evaluation protocols that characterize model behavior under realistic use constraints.

Emotional Support Dialogue Systems (ESDSes) are increasingly evaluated and trained with LLM-simulated seekers. However, such simulated seekers often behave as cooperative, average-case users who disclose clearly, respond constructively, and accept support within a few turns. This can lead to overly optimistic evaluation and obscure whether ESDSes can handle difficult help-seeking interactions. In this work, we study ESDS evaluation under worst-case interactions, where seekers are hard to help due to low engagement, resistance, limited self-disclosure, emotional volatility, or rigid negative interpretations. We first conduct an expert simulation study with eight experienced counselling professionals, who simulate difficult seekers, interact with existing Chinese ESDSes, provide scale ratings, and participate in semi-structured interviews. Based on this study, we derive worst-case seeker behaviours and identify key limitations of current systems. We then propose a worst-case evaluation framework consisting of an LLM-based worst-case seeker simulator and four worst-case-oriented metrics: Deep Emotional Understanding, Guided Exploration, Balanced Emotional Support, and Authentic and Grounded Support. Evaluating 17 systems, we find that nearly all models suffer substantial performance drops under worst-case interactions. Large general-purpose LLMs are generally more robust than specialised ESDSes, but even the strongest models struggle to sustain engagement and improve seekers' emotional states. Finally, we show that worst-case simulation can also generate useful training data, improving the robustness of smaller models.

Existing benchmarks for LLM coding agents primarily evaluate final outcomes. While useful for measuring overall capability, these metrics provide limited visibility and often miss defects that arise during execution. We present ProcCtrlBench, a benchmark for execution-process evaluation in LLM coding agents. ProcCtrlBench organizes recurrent execution defects into a reusable ontology covering 11 defect types in 4 categories, and evaluates agent trajectories through standardized process evidence rather than final outcomes alone. To support comparison across heterogeneous agents, ProcCtrlBench standardizes raw logs into a unified trajectory representation and reports calibrated scorecards over process-level findings. In addition, ProcCtrlBench uses control preservation as a way to quantify execution-process quality, capturing whether execution remains interpretable, interruptible, correctable, reversible, and able to hand back authority when needed. We evaluate ProcCtrlBench on 200 cases sampled from three benchmarks: AndroidBench, TerminalBench, and SWE-bench-Verified. Results show that ProcCtrlBench can be instantiated with useful reliability, provides more stable semantics than direct thresholding, and reveals meaningful differences in execution quality that are often overlooked by conventional outcome-based evaluation.

Generative text-to-image models are typically trained on large-scale web-scraped datasets that include diverse visual content such as copyrighted and stylistically distinctive artworks, raising concerns about ownership, attribution, and the unintended reuse of protected visual expressions. A key issue is that models can learn stylistic patterns from this data and reproduce them in generated outputs without any explicit reference in the prompt. We refer to this phenomenon as The Silent Brush, where such learned styles reappear even when they are not requested. Existing evaluation methods mainly focus on near-duplicate retrieval or membership inference and do not account for this form of unintended stylistic resurfacing across prompts. To address these gaps, we first formulate guiding principles for evaluation of The Silent Brush. We then introduce Art Arena, an evaluation protocol that measures how strongly artworks are encoded, how they interact, and how frequently their stylistic traits reappear in generated outputs without explicit mention in prompts. We evaluate Art Arena on widely used text-to-image diffusion models, including Stable Diffusion v1.5, Stable Diffusion XL (SDXL), and SANA-1.5, and design it to generalize across text-to-image generative systems. Our results show that The Silent Brush arises from differences in representational strength and interaction dynamics between artworks, leading to asymmetric blending in model generations. Code and evaluation resources are available at: https://anonymous.4open.science/r/ArtArena-EBE4.

We introduce AcuityBench, a benchmark for evaluating whether language models identify the appropriate urgency of care from user medical presentations. Existing health benchmarks emphasize medical question answering, broad health interactions, or narrow workflow-specific triage tasks, but they do not offer a unified evaluation of acuity identification across these settings. AcuityBench addresses this gap by harmonizing five public datasets spanning user conversations, online forum posts, clinical vignettes, and patient portal messages under a shared four-level acuity framework ranging from home monitoring to immediate emergency care. The benchmark contains 914 cases, including 697 consensus cases for standard accuracy evaluation and 217 physician-confirmed ambiguous cases for uncertainty-aware evaluation. It supports two complementary task formats: explicit four-way classification in a QA setting, and free-form conversational responses evaluated with a rubric-based judge anchored to the same framework. Across 12 frontier proprietary and open-weight models, we find substantial variation in clear-case acuity accuracy and error direction. Comparing task formats reveals a systematic tradeoff: conversational responses reduce over-triage but increase under-triage relative to QA, especially in higher-acuity cases. In ambiguous cases, no model closely matches the distribution of physician judgments, and model predictions are more concentrated than expert clinical uncertainty. We also compare expert and model adjudication on a subset of maximally ambiguous cases, using those cases to examine the role of clinical uncertainty in label disagreement. Together, these results position acuity identification as a distinct safety-critical capability and show that AcuityBench enables systematic comparison and stress-testing of how well models guide users to the right level of care in real-world health use.

We introduce SmartEval, a benchmark for systematically evaluating the quality of Solidity smart contracts generated by large language models (LLMs) from natural language specifications. SmartEval provides a corpus of 9,000 generated contracts paired with expert-written ground-truth implementations drawn from the FSMSCG dataset, a five-dimensional evaluation rubric covering functional completeness, variable fidelity, state-machine correctness, business-logic fidelity, and code quality, and a reproducible generation-and-evaluation pipeline. To validate the benchmark's reliability, we conduct three independent empirical studies: a five-condition ablation study (N=300 per condition) isolating the contribution of each pipeline component, a human expert evaluation by three Columbia University PhD researchers confirming automated scores align with expert judgment to within 0.34 points, and external security analysis via the Slither static analyzer confirming 79.4% agreement between the LLM auditor and a non-LLM rule-based tool. Systematic analysis of 9,000 generated contracts reveals characteristic failure modes (logic omissions at 35.3%, state transition errors at 23.4%, and complexity-driven degradation) and quantifies a +8.29 composite-score advantage of generated contracts over ground-truth implementations, attributable to LLMs' literal specification-following behavior. SmartEval establishes a reproducible, validated foundation for empirical research on LLM smart contract synthesis quality, with all data, evaluation code, and generated contracts publicly released.

Evaluating pragmatic reasoning in large language models (LLMs) remains challenging because model behavior can vary depending on evaluation methods. Previous studies suggest that prompt-based judgments may diverge from models' internal probability distributions, raising questions about whether observed performance reflects underlying competence or task-induced behavior. This study examines this issue using scalar diversity as a graded diagnostic for pragmatic inference. Following Hu & Levy (2023), this study compares direct probability measurement and metalinguistic prompting across multiple models and experimental settings. The results show that neither evaluation method consistently outperforms the other and that pragmatic behavior varies substantially across model families, prompting strategies, and task structures. Moreover, scalar diversity gradients emerge only in specific model-condition combinations, suggesting that pragmatic reasoning in LLMs reflects an interaction between internal probabilistic representations and task-induced prompting behavior rather than a stable competence captured by a single evaluation paradigm. These findings highlight the central role of evaluation design in interpreting pragmatic abilities in LLMs.

Calibration measures whether a model's predicted confidence aligns with its empirical accuracy, and is central to the reliable deployment of large language models (LLMs) in high-stakes domains such as medicine and law. While much recent work focuses on improving LLM calibration, the equally important question of how to evaluate it in realistic settings remains underdeveloped. Open-ended question answering (QA), the most common deployment setting for modern LLMs, is where existing evaluation methods fall short: logit-based metrics need restricted output formats and internal probabilities; verbalized confidence is self-reported and often overconfident; and sampling-based methods rely on task-specific extraction rules without a clear finite-sample target. We introduce Sem-ECE (Semantic-Sampling Expected Calibration Error), a calibration evaluation framework for open-ended QA that samples answers from the model, groups them into semantic classes, and uses the resulting frequencies as confidence. We study two estimators within this framework: Sem$_1$-ECE, the same-sample self-consistency score, and Sem$_2$-ECE, a held-out variant that separates answer selection from confidence evaluation. We prove both are asymptotically unbiased, and further show that they agree on easy questions but diverge on hard ones with Sem$_2$ achieving strictly smaller calibration error, so their gap also serves as a diagnostic for question difficulty. Experiments on three open-ended QA benchmarks across five leading commercial LLMs match our theoretical predictions and show that Sem-ECE outperforms verbalized confidence and existing sampling-based methods, while complementing logit-based evaluation when internal probabilities are unavailable.

Recent text-guided image editing (TIE) models have achieved remarkable progress, however, many edited results still suffer from artifacts, unintended modifications, and suboptimal aesthetics. Although several benchmarks and evaluation methods have been proposed, most existing approaches rely on scalar scores and lack interpretability. This limitation largely stems from the absence of high-quality interpretation datasets for TIE and effective reward models to train interpretable evaluators. To address these challenges, we introduce ReasonEdit-22K, the first dataset that combines 22K edited images with 113K Chain-of-Thought (CoT) samples, along with 1.3M human judgments assessing these interpretations in terms of logicality, accuracy, and usefulness. Building upon this dataset, we propose RE-Reward, a multimodal large language model (MLLM)-based reward model designed to provide human-aligned feedback for evaluating interpretable reasoning in image editing. Furthermore, we develop ReasonEdit, which is trained using reward signals derived from RE-Reward and the Group Relative Policy Optimization (GRPO) algorithm to learn an interpretable evaluation model. Extensive experiments demonstrate that ReasonEdit achieves superior alignment with human preferences and exhibits strong generalization across public benchmarks. In addition, it is capable of generating high-quality interpretable evaluation text, enabling more transparent and trustworthy assessment for image editing. The code is available at https://github.com/IntMeGroup/ReasonEdit.

Reproducing and comparing deep research agents today is hard: the same backbone evaluated on the same benchmark can report different accuracies across papers because the harness and tool registry differ, and integrating a new model into a comparable evaluation surface costs weeks of model-specific engineering. These are symptoms of a broader reproducibility problem in deep research agent research. Here, we introduce BioMedArena, an open-source toolkit that addresses this reproducibility gap and provides an arena for comparing deep research agents under a shared evaluation environment. BioMedArena decouples six layers of biomedical agent evaluation -- benchmark loading, tool exposure, tool selection, harness mode, context management, and scoring -- and exposes 166 biomedical benchmarks and 75 biomedical tools across 9 functional families. Adding a new model, benchmark, or tool can be accomplished with a few-line provider adapter. Beyond evaluation infrastructure, BioMedArena ships a library of high-quality reference components: 6 agent harnesses (including our proposed Mutual-Evolve) and 6 context-management strategies, any of which can be equipped on any backbone. Equipping these components substantially improves all 12 backbones; on each of 8 representative biomedical benchmarks, the best equipped backbone surpasses prior state-of-the-art (SOTA), by 15.01 percentage points on average. The toolkit, configurations, and per-task traces are available at https://github.com/AI-in-Health/BioMedArena.

Large language models (LLMs) have demonstrated impressive multilingual capabilities for well-resourced languages, yet their performance on low-resource African languages remains poorly understood and largely unevaluated. This paper presents Nsanku, a systematic benchmark that evaluates the zero-shot machine translation performance of 19 open-weight and proprietary LLMs across 43 Ghanaian languages paired with English. Evaluation sentences were sourced from the YouVersion Bible platform, providing 300 sentence pairs per language. Two complementary automatic metrics are employed: Bilingual Evaluation Understudy (BLEU) and Character n-gram F-Score (chrF), alongside an average accuracy score and a cross-language consistency dimension. Nsanku represents the most comprehensive LLM translation evaluation for Ghanaian languages conducted to date. Results show that gemini-2.5-flash achieves the highest overall average score of 26.88 (BLEU: 24.60, chrF: 29.16), followed by claude-sonnet-4-5 at 24.87 (BLEU: 22.46, chrF: 27.28) and gpt-4.1 at 23.20 (BLEU: 21.15, chrF: 25.24). Among open-weight models, kimi-k2-instruct-0905 leads at an average score of 20.87. A critical finding from the consistency analysis is that no model and no language reached the Leaders quadrant of high performance and high consistency simultaneously, indicating that current LLMs are not yet reliably usable for Ghanaian language translation at scale. Siwu achieved the highest per-language average score at 25.73 while Nkonya scored lowest at 11.65. Nsanku establishes a publicly available, community-extensible evaluation infrastructure for African language NLP research.

Recent advances in large-scale video world models have enabled increasingly realistic future prediction, raising the prospect of using generated videos as scalable supervision for robot learning. However, for embodied manipulation, perceptual realism alone is not sufficient: generated interactions must also be physically consistent and executable by robotic agents. Existing benchmarks provide valuable assessments of visual quality and physical plausibility, but they do not systematically evaluate whether predicted behaviors can be translated into executable actions that complete manipulation tasks. We introduce RoboWM-Bench, a manipulation-centric benchmark for embodiment-grounded evaluation of video world models. RoboWM-Bench converts generated human-hand and robotic manipulation videos into embodied action sequences and validates them through execution in physically grounded simulation environments. Built on real-to-sim scene reconstruction and diverse manipulation tasks, RoboWM-Bench enables standardized, reproducible, and scalable evaluation of physical executability. Using RoboWM-Bench, we evaluate state-of-the-art video world models and observe that visual plausibility and embodied executability are not always aligned. Our analysis highlights several recurring factors that affect execution performance, including spatial reasoning, contact prediction, and non-physical geometric distortions, particularly in complex and long-horizon interactions. These findings provide a more fine-grained view of current model capabilities and underscore the value of embodiment-aware evaluation for guiding physically grounded world modeling in robotic manipulation.

Multimodal Large Language Models (MLLMs) have been increasingly used as automatic evaluators-a paradigm known as MLLM-as-a-Judge. However, their reliability and vulnerabilities to biases remain underexplored. We find that many MLLM judges fail to reliably integrate key visual or textual cues, yielding unreliable evaluations when evidence is missing or mismatched, and exhibiting instability under semantically irrelevant perturbations. To address this, we systematically define Compositional Bias in MLLM-as-a-Judge systems and introduce MM-JudgeBias, a benchmark for evaluating it. MM-JudgeBias introduces controlled perturbations across Query, Image, and Response, and evaluates model behavior via two complementary metrics: Bias-Deviation (BD) for sensitivity and Bias-Conformity (BC) for stability. Our dataset of over 1,800 curated and refined multimodal samples, drawn from 29 source benchmarks, enables a fine-grained diagnosis of nine bias types across diverse tasks and domains. Experiments on 26 state-of-the-art MLLMs reveal systematic modality neglect and asymmetric evaluation tendencies, underscoring the need for more reliable judges.

Transforming legal text into executable decision logic is a longstanding challenge in legal informatics. With the rise of LLMs, this task has gained renewed interest, but remains challenging due to requiring extensive manual coding and evaluation. We use a unique real-world dataset that pairs production-grade decision models with legal text from the Dutch Environment and Planning Act. These models power the Omgevingsloket government platform, where citizens check permit requirements for environmental activities. We study whether intermediate structured representations can improve LLM-based generation of executable decision models from legal text. We compare four input conditions: raw legal text, text enriched with semantic role labels, text enriched with input and output constraints, and text enriched with both. We evaluate along two dimensions: structural evaluation, through similarity to gold decision models with graph kernels and graphs' descriptive statistics, and outcome evaluation, through functional equivalence by executing models on pre-configured test scenarios. Our findings show that I/O constraints provide the dominant improvement (+37-54% similarity over baseline), while semantic role labels show modest improvements. Outcome evaluation shows that generated models match the gold standard on 51-53% of test scenarios, even though generated models are typically smaller and simpler. We find LLMs eliminate redundant pass-through logic that comprises up to 45-55% of nodes. Importantly, structural similarity and outcome equivalence are complementary: structural similarity does not guarantee outcome equivalence, and vice versa. To facilitate reproducibility, we publicly release our dataset of 95 production decision models with associated legal text and all experimental code.

Should we trust Large Language Models (LLMs) with high accuracy? LLMs achieve high accuracy on reasoning benchmarks, but correctness alone does not reveal the quality of the reasoning used to produce it. This highlights a fundamental limitation of outcome-based evaluation: models may arrive at correct answers through flawed reasoning, and models with substantially different reasoning capabilities can nevertheless exhibit similar benchmark accuracy, for example due to memorization or over-optimization. In this paper, we ask: given existing benchmarks, can we move beyond outcome-based evaluation to assess the quality of reasoning itself? We seek metrics that (1) differentiate models with similar accuracy and (2) are robust to variations in input prompts and generation configurations. To this end, we propose a reasoning score that evaluates reasoning traces along dimensions such as faithfulness, coherence, utility, and factuality. A remaining question is how to aggregate this score across multiple sampled traces. Naively averaging them is undesirable, particularly in long-horizon settings, where the number of possible trajectories grows rapidly, and low-confidence correct traces are more likely to be coincidental. To address this, we introduce the Filtered Reasoning Score (FRS), which computes reasoning quality using only the top-K% most confident traces. Evaluating with FRS, models that are indistinguishable under standard accuracy exhibit significant differences in reasoning quality. Moreover, models with higher FRS on one benchmark tend to perform better on other reasoning benchmarks, in both accuracy and reasoning quality. Together, these findings suggest that FRS complements accuracy by capturing a model's transferable reasoning capabilities. We open source our evaluation codebase: https://github.com/Manas2006/benchmark_reproducibility.

Recent advances in large audio language models (LALMs) have enabled multilingual speech understanding. However, benchmarks for evaluating LALMs remain scarce for non-English languages, with Korean being one such underexplored case. In this paper, we introduce KoALa-Bench, a comprehensive benchmark for evaluating Korean speech understanding and speech faithfulness of LALMs. In particular, KoALa-Bench comprises six tasks. Four tasks evaluate fundamental speech understanding capabilities, including automatic speech recognition, speech translation, speech question answering, and speech instruction following, while the remaining two tasks evaluate speech faithfulness, motivated by our observation that several LALMs often fail to fully leverage the speech modality. Furthermore, to reflect Korea-specific knowledge, our benchmark incorporates listening questions from the Korean college scholastic ability test as well as content covering Korean cultural domains. We conduct extensive experiments across six models, including both white-box and black-box ones. Our benchmark, evaluation code, and leaderboard are publicly available at https://ksbench.github.io/Korean-Benchmark/.

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

Large language models (LLMs) now support contexts of up to 1M tokens, but their strengths and weaknesses on complex long-context tasks remain unclear. To study this, we focus on multi-document legal case summarization, where a single case often spans many documents exceeding 100K tokens. We systematically evaluate 12 frontier LLMs with Gavel, which consists of Gavel-Ref, a reference-based evaluation framework with checklist, residual-fact, and writing-style evaluations, and Gavel-Agent, a reference-free agent for evaluating factual coverage directly from source documents. Our results show that current models are more prone to omitting key information than hallucinating. They all perform well on simple checklist items, such as filing date, but struggle with rare and complex items, such as settlements. Performance also declines as case length increases. To meta-evaluate Gavel, we collect 160 hours of human annotations. Gavel-Agent reduces token usage by at least 36% compared to end-to-end and chunk-by-chunk methods while achieving competitive performance. Gavel-Agent also generalizes to the medical domain, performing the best with at least 77% fewer tokens.

We present VLMEvalKit: an open-source toolkit for evaluating large multi-modality models based on PyTorch. The toolkit aims to provide a user-friendly and comprehensive framework for researchers and developers to evaluate existing multi-modality models and publish \textbf{reproducible} evaluation results. In VLMEvalKit, we implement over 450+ large multi-modality model configurations, including both proprietary APIs and open-source models, and support 330+ benchmarks across diverse multi-modal benchmarks. By implementing a single interface, new models can be easily added to the toolkit, while the toolkit automatically handles the remaining workloads, including data preparation, distributed inference, prediction post-processing, and metric calculation. VLMEvalKit has also evolved to a broader evaluation suite spanning video/audio, document understanding, GUI grounding, spatial reasoning, safety, scientific reasoning, and multi-turn dialogue. Based on the evaluation results obtained with the toolkit, we host the OpenVLM Leaderboard, a comprehensive leaderboard to track the progress of multi-modality learning research. The toolkit is released on https://github.com/open-compass/VLMEvalKit and is actively maintained.

Warning: This paper contains content that may be offensive or upsetting. There has been a significant increase in the usage of large language models (LLMs) in various applications, both in their original form and through fine-tuned adaptations. As a result, LLMs have gained popularity and are being widely adopted by a large user community. However, one of the concerns with LLMs is the potential generation of socially biased content. The existing evaluation methods have many constraints, and their results exhibit a limited degree of interpretability. In this work, we propose a bias evaluation framework named GPTBIAS that leverages the high performance of LLMs (e.g., GPT-4 \cite{openai2023gpt4}) to assess bias in models. We also introduce prompts called Bias Attack Instructions, which are specifically designed for evaluating model bias. To enhance the credibility and interpretability of bias evaluation, our framework not only provides a bias score but also offers detailed information, including bias types, affected demographics, keywords, reasons behind the biases, and suggestions for improvement. We conduct extensive experiments to demonstrate the effectiveness and usability of our bias evaluation framework.

Automatic evaluation of image and video captioning is essential for benchmarking multimodal systems, although standard evaluation metrics show limited alignment with human judgments. Recent approaches using large language models (LLMs), commonly referred to as LLM-as-a-Judge, have improved alignment with human judgments but still suffer from a mismatch between large-vocabulary language modeling and evaluation over a small label set. To address this, we propose Rigel, an automatic evaluation metric for image and video captioning, based on self-distilled score adaptation. The metric employs an evaluation-specific scoring head distilled from a frozen LLM, which captures judgment signals in a task-aligned space without relying on large-vocabulary token sets. We then refine the LLM backbone with human judgment data. To train Rigel, we constructed the Vid-Lepus dataset, which contains 3,338 video clips, 33,380 reference captions, and 5,637 candidate captions. Experiments on multiple benchmarks show that Rigel outperforms state-of-the-art metrics, achieving over 10-point improvements on ActivityNet-Fact in the reference-free setting.