Daily Papers

Image classification accuracy on the ImageNet dataset has been a barometer for progress in computer vision over the last decade. Several recent papers have questioned the degree to which the benchmark remains useful to the community, yet innovations continue to contribute gains to performance, with today's largest models achieving 90%+ top-1 accuracy. To help contextualize progress on ImageNet and provide a more meaningful evaluation for today's state-of-the-art models, we manually review and categorize every remaining mistake that a few top models make in order to provide insight into the long-tail of errors on one of the most benchmarked datasets in computer vision. We focus on the multi-label subset evaluation of ImageNet, where today's best models achieve upwards of 97% top-1 accuracy. Our analysis reveals that nearly half of the supposed mistakes are not mistakes at all, and we uncover new valid multi-labels, demonstrating that, without careful review, we are significantly underestimating the performance of these models. On the other hand, we also find that today's best models still make a significant number of mistakes (40%) that are obviously wrong to human reviewers. To calibrate future progress on ImageNet, we provide an updated multi-label evaluation set, and we curate ImageNet-Major: a 68-example "major error" slice of the obvious mistakes made by today's top models -- a slice where models should achieve near perfection, but today are far from doing so.

Finetuning data selection requires balancing two competing goals: selecting examples that improve the downstream objective, and doing so without repeatedly finetuning models. Train-free selectors are scalable but rely on proxies such as embedding similarity or clustering, which may not match the target objective. Train-based selectors better reflect downstream utility through gradient signals, subset evaluation, or Shapley attribution, but require many costly train--evaluate iterations. We propose Hierarchical Active Region Pruning (HARP), an efficient train-based selector that preserves downstream alignment while reducing selection cost. HARP organizes the training pool into a node--leaf hierarchy, evaluates only representative leaves, and infers unmeasured utilities with empirical Bayes posteriors. It then selects data using two complementary envelopes: HARP-C, which conservatively controls redundancy, and HARP-E, which additively rewards complementary regions. We theoretically show that, under local smoothness and bounded estimation error, HARP controls selection error while reducing train--evaluate cost. We further validate that HARP variants achieve the best result and outperform the strongest baseline by up to +8.9 points, while using roughly 7times fewer training examples.

The prohibitive cost of evaluating large language models (LLMs) on comprehensive benchmarks necessitates the creation of small yet representative data subsets (i.e., tiny benchmarks) that enable efficient assessment while retaining predictive fidelity. Current methods for this task operate under a model-centric paradigm, selecting benchmarking items based on the collective performance of existing models. Such approaches are limited by large upfront costs, an inability to immediately handle new benchmarks (`cold-start'), and the fragile assumption that future models will share the failure patterns of their predecessors. In this work, we challenge this paradigm and propose a item-centric approach to benchmark subset selection, arguing that selection should be based on the intrinsic properties of the task items themselves, rather than on model-specific failure patterns. We instantiate this item-centric efficient benchmarking approach via a novel method, Scales++, where data selection is based on the cognitive demands of the benchmark samples. Empirically, we show Scales++ reduces the upfront selection cost by over 18x while achieving competitive predictive fidelity. On the Open LLM Leaderboard, using just a 0.5\% data subset, we predict full benchmark scores with a 2.9% mean absolute error. We demonstrate that this item-centric approach enables more efficient model evaluation without significant fidelity degradation, while also providing better cold-start performance and more interpretable benchmarking.

The potential of Multimodal Large Language Models (MLLMs) in domain of medical imaging raise the demands of systematic and rigorous evaluation frameworks that are aligned with the real-world medical imaging practice. Existing practices that report single or coarse-grained metrics are lack the granularity required for specialized clinical support and fail to assess the reliability of reasoning mechanisms. To address this, we propose a paradigm shift toward multidimensional, fine-grained and in-depth evaluation. Based on a two-stage systematic construction pipeline designed for this paradigm, we instantiate it with MedRCube. We benchmark 33 MLLMs, Lingshu-32B achieve top-tier performance. Crucially, MedRCube exposes a series of pronounced insights inaccessible under prior evaluation settings. Furthermore, we introduce a credibility evaluation subset to quantify reasoning credibility, uncover a highly significant positive association between shortcut behavior and diagnostic task performance, raising concerns for clinically trustworthy deployment. The resources of this work can be found at https://github.com/F1mc/MedRCube.

Autonomous agile robots need more than metric geometry: they must understand objects, rooms, places, and spatial relations for search, inspection, exploration, and human robot interaction. Conventional metric maps support localization and collision avoidance, but do not provide this semantic and relational structure. 3D scene graphs address this gap by connecting geometry with object level and room level understanding. Building such representations on agile platforms remains difficult because aerial and lightweight robots operate under strict payload, power, and compute limits, making RGB-D cameras and LiDAR sensors impractical for many onboard settings. We present Mono-Hydra++, a real time monocular RGB plus IMU pipeline for indoor metric semantic mapping and hierarchical 3D scene graph construction. The system combines M2H-MX, a DINOv3 based multi-task model for depth and semantics, with a deep feature visual inertial odometry front end, sparse predicted depth constraints in the VIO derived pose graph, semantic masking for dynamic regions, and pose aware temporal alignment before volumetric fusion in the Mono-Hydra backend. On the Go-SLAM ScanNet evaluation subset, Mono-Hydra++ achieves 1.6% lower average trajectory error than the strongest RGB-D baseline in our comparison, while using only monocular RGB plus IMU input. On calibrated 7-Scenes, it improves average ATE by 29.8% over the strongest competing calibrated baseline. We further validate Mono-Hydra++ in a real ITC building deployment using RealSense RGB plus IMU and demonstrate embedded feasibility by deploying the ONNX/TensorRT FP16 M2H-MX-L perception model at 25.53 FPS on a Jetson Orin NX 16GB. These results show that Mono-Hydra++ can provide real time metric semantic mapping and scene graph construction for resource constrained robotic platforms without relying on active depth sensors.

Single-image 3D scene reconstruction must complete partially observed objects and place them coherently in a shared observation-aligned scene frame. Object-level generative priors offer strong completion ability, but their centered, scale-normalized outputs are typically expressed in an object frame, creating a fundamental representation gap between object generation and scene reconstruction. We introduce SceneReGen, a generative reconstruction framework that reinterprets scene reconstruction as the generation and assembly of complete object assets in a shared observation-aligned scene frame. SceneReGen addresses the generation-reconstruction gap through selective pose factorization: each object's observed orientation is encoded directly in the generated mesh, while translation and scale are estimated from instance-level and global scene evidence. Given a scene image and instance masks, a geometry encoder extracts dense cues; learnable shape queries condition a pretrained DiT-based 3D generator to produce complete meshes in their observed orientations, while position queries fuse object and scene features to assemble them in the shared frame. On the 3D-FUTURE evaluation subset, SceneReGen achieves the best scene-level CD, scene-level F-Score, and 3D bounding-box IoU among the evaluated methods, ties the best object-level CD, and ranks second in object-level F-Score. Qualitative outputs in autonomous-driving and embodied-AI scenes further illustrate the potential of asset-centric reconstruction beyond indoor furniture.

Automatic evaluators such as reward models play a central role in the alignment and evaluation of large vision-language models (LVLMs). Despite their growing importance, these evaluators are almost exclusively assessed on English-centric benchmarks, leaving open the question of how well these evaluators generalize across languages. To answer this question, we introduce MM-JudgeBench, the first large-scale benchmark for multilingual and multimodal judge model evaluation, which includes over 60K pairwise preference instances spanning 25 typologically diverse languages. MM-JudgeBench integrates two complementary subsets: a general vision-language preference evaluation subset extending VL-RewardBench, and a chart-centric visual-text reasoning subset derived from OpenCQA, enabling systematic analysis of reward models (i.e., LVLM judges) across diverse settings. We additionally release a multilingual training set derived from MM-RewardBench, disjoint from our evaluation data, to support domain adaptation. By evaluating 22 LVLMs (15 open-source, 7 proprietary), we uncover substantial cross-lingual performance variance in our proposed benchmark. Our analysis further shows that model size and architecture are poor predictors of multilingual robustness, and that even state-of-the-art LVLM judges exhibit inconsistent behavior across languages. Together, these findings expose fundamental limitations of current reward modeling and underscore the necessity of multilingual, multimodal benchmarks for developing reliable automated evaluators.

We study device-addressed speech detection under pre-ASR edge deployment constraints, where systems must decide whether to forward audio before transcription under strict latency and compute limits. We show that, in multi-speaker environments with temporally ambiguous utterances, this task is more effectively modelled as a sequential routing problem over interaction history than as an utterance-local classification task. We formalize this as Sequential Device-Addressed Routing (SDAR) and present the Selective Attention System (SAS), an on-device implementation that instantiates this formulation. On a held-out 60-hour multi-speaker English test set, the primary audio-only configuration achieves F1=0.86 (precision=0.89, recall=0.83); with an optional camera, audio+video fusion raises F1 to 0.95 (precision=0.97, recall=0.93). Removing causal interaction history (Stage~3) reduced F1 from 0.95 to 0.57+/-0.03 in the audio+video configuration under our evaluation protocol. Among the tested components, this was the largest observed ablation effect, indicating that short-horizon interaction history carries substantial decision-relevant information in the evaluated setting. SAS runs fully on-device on ARM Cortex-A class hardware (<150 ms latency, <20 MB footprint). All results are from internal evaluation on a proprietary dataset evaluated primarily in English; a 5-hour evaluation subset may be shared for independent verification (Section 8.8).

Chart-to-code reconstruction -- the task of recovering executable plotting scripts from chart images -- provides important insights into a model's ability to ground data visualizations in precise, machine-readable form. Yet many existing multimodal benchmarks largely focus primarily on answering questions about charts or summarizing them. To bridge this gap, we present ChartGen, a fully-automated pipeline for code-guided synthetic chart generation. Starting from seed chart images, ChartGen (i) prompts a vision-language model (VLM) to reconstruct each image into a python script, and (ii) iteratively augments that script with a code-oriented large language model (LLM). Using ChartGen, we create 222.5K unique chart-image code pairs from 13K seed chart images, and present an open-source synthetic chart dataset covering 27 chart types, 11 plotting libraries, and multiple data modalities (image, code, text, CSV, DocTags). From this corpus, we curate a held-out chart-to-code evaluation subset of 4.3K chart image-code pairs, and evaluate six open-weight VLMs (3B - 26B parameters), highlighting substantial room for progress. We release the pipeline, prompts, and the dataset to help accelerate efforts towards robust chart understanding and vision-conditioned code generation: https://github.com/SD122025/ChartGen/

Revision is a crucial step in scientific writing, where authors refine their work to improve clarity, structure, and academic quality. Existing approaches to automated writing assistance often focus on sentence-level revisions, which fail to capture the broader context needed for effective modification. In this paper, we explore the impact of shifting from sentence-level to paragraph-level scope for the task of scientific text revision. The paragraph level definition of the task allows for more meaningful changes, and is guided by detailed revision instructions rather than general ones. To support this task, we introduce ParaRev, the first dataset of revised scientific paragraphs with an evaluation subset manually annotated with revision instructions. Our experiments demonstrate that using detailed instructions significantly improves the quality of automated revisions compared to general approaches, no matter the model or the metric considered.

The latest research in the field of voice anti-spoofing (VAS) shows that deep neural networks (DNN) outperform classic approaches like GMM in the task of presentation attack detection. However, DNNs require a lot of data to converge, and still lack generalization ability. In order to foster the progress of neural network systems, we introduce a Large Replay Parallel Dataset (LRPD) aimed for a detection of replay attacks. LRPD contains more than 1M utterances collected by 19 recording devices in 17 various environments. We also provide an example training pipeline in PyTorch [1] and a baseline system, that achieves 0.28% Equal Error Rate (EER) on evaluation subset of LRPD and 11.91% EER on publicly available ASVpoof 2017 [2] eval set. These results show that model trained with LRPD dataset has a consistent performance on the fully unknown conditions. Our dataset is free for research purposes and hosted on GDrive. Baseline code and pre-trained models are available at GitHub.

Evaluating LLM agents on benchmarks like SWE-Bench and GAIA can be expensive, time-consuming, and requires complex infrastructure. A single evaluation can cost thousands of dollars and take days to complete. In contrast, non-agentic LLM benchmarks that test individual capabilities (e.g., reasoning, code generation) are fast and cheap to run. In this paper, we investigate whether performance on expensive agentic benchmarks can be accurately predicted by the performance on a small, carefully selected subset of atomic evaluation instances. We introduce PACE, a framework that constructs proxy benchmarks by selecting instances from existing non-agentic evaluations whose aggregate scores most reliably predict model performances on agentic benchmarks. Given a pool of candidate instances spanning atomic capabilities, PACE fits a regression that maps a model's scores on a compact subset of source instances to its score on the target agentic benchmark. The subset itself is curated by combining two complementary instance-selection strategies, target-relevance local selection and globally informative global selection. We apply PACE to the 4 target agentic benchmarks in this paper, which yields PACE-Bench, the concrete proxy benchmark that we evaluate in the paper. Experiments across 14 models, 4 agentic benchmarks, and 19 non-agentic benchmarks show that PACE-Bench predicts agentic scores with leave-one-out cross-validation (LOOCV) mean absolute error (MAE) under 4%, Spearman correlation above 0.80, and pairwise model-ranking accuracy around 85%, all at much less than 1% of the full agentic evaluation cost. We further analyze the selected proxy instances, revealing which skills each agentic benchmark uniquely demands. PACE enables practitioners to obtain reliable estimates of agentic performance during model development, selection, and routing, without the overhead of full agent evaluation.

To solve complex tasks, large language models (LLMs) often require multiple rounds of interactions with the user, sometimes assisted by external tools. However, current evaluation protocols often emphasize benchmark performance with single-turn exchanges, neglecting the nuanced interactions among the user, LLMs, and external tools, while also underestimating the importance of natural language feedback from users. These oversights contribute to discrepancies between research benchmark evaluations and real-world use cases. We introduce MINT, a benchmark that evaluates LLMs' ability to solve tasks with multi-turn interactions by (1) using tools and (2) leveraging natural language feedback. To ensure reproducibility, we provide an evaluation framework where LLMs can access tools by executing Python code and receive users' natural language feedback simulated by GPT-4. We repurpose a diverse set of established evaluation datasets focusing on reasoning, coding, and decision-making and carefully curate them into a compact subset for efficient evaluation. Our analysis of 20 open- and closed-source LLMs offers intriguing findings. (a) LLMs generally benefit from tools and language feedback, with performance gains (absolute, same below) of 1-8% for each turn of tool use and 2-17% with natural language feedback. (b) Better single-turn performance does not guarantee better multi-turn performance. (c) Surprisingly, on the LLMs evaluated, supervised instruction-finetuning (SIFT) and reinforcement learning from human feedback (RLHF) generally hurt multi-turn capabilities. We expect MINT can help measure progress and incentivize research in improving LLMs' capabilities in multi-turn interactions, especially for open-source communities where multi-turn human evaluation can be less accessible compared to commercial LLMs with a larger user base.

Pavement condition assessment is essential for road safety and maintenance. Existing research has made significant progress. However, most studies focus on conventional computer vision tasks such as classification, detection, and segmentation. In real-world applications, pavement inspection requires more than visual recognition. It also requires quantitative analysis, explanation, and interactive decision support. Current datasets are limited. They focus on unimodal perception. They lack support for multi-turn interaction and fact-grounded reasoning. They also do not connect perception with vision-language analysis. To address these limitations, we introduce PaveBench, a large-scale benchmark for pavement distress perception and interactive vision-language analysis on real-world highway inspection images. PaveBench supports four core tasks: classification, object detection, semantic segmentation, and vision-language question answering. It provides unified task definitions and evaluation protocols. On the visual side, PaveBench provides large-scale annotations and includes a curated hard-distractor subset for robustness evaluation. It contains a large collection of real-world pavement images. On the multimodal side, we introduce PaveVQA, a real-image question answering (QA) dataset that supports single-turn, multi-turn, and expert-corrected interactions. It covers recognition, localization, quantitative estimation, and maintenance reasoning. We evaluate several state-of-the-art methods and provide a detailed analysis. We also present a simple and effective agent-augmented visual question answering framework that integrates domain-specific models as tools alongside vision-language models. The dataset is available at: https://huggingface.co/datasets/MML-Group/PaveBench.

Feature selection is the problem of selecting a subset of features for a machine learning model that maximizes model quality subject to a budget constraint. For neural networks, prior methods, including those based on ell_1 regularization, attention, and other techniques, typically select the entire feature subset in one evaluation round, ignoring the residual value of features during selection, i.e., the marginal contribution of a feature given that other features have already been selected. We propose a feature selection algorithm called Sequential Attention that achieves state-of-the-art empirical results for neural networks. This algorithm is based on an efficient one-pass implementation of greedy forward selection and uses attention weights at each step as a proxy for feature importance. We give theoretical insights into our algorithm for linear regression by showing that an adaptation to this setting is equivalent to the classical Orthogonal Matching Pursuit (OMP) algorithm, and thus inherits all of its provable guarantees. Our theoretical and empirical analyses offer new explanations towards the effectiveness of attention and its connections to overparameterization, which may be of independent interest.

Recent advances have witnessed the effectiveness of reinforcement learning (RL) finetuning in enhancing the reasoning capabilities of large language models (LLMs). The optimization process often requires numerous iterations to achieve satisfactory performance, resulting in high computational costs due to the need for frequent prompt evaluations under intensive LLM interactions and repeated policy updates. Appropriate online prompt selection methods reduce iteration steps by prioritizing informative prompts during training, while the pipeline's reliance on exhaustive prompt evaluation and subset selection for optimization still incurs substantial computational overhead due to frequent LLM inference calls. Distinguished from these direct evaluate-then-select schemes, this work investigates iterative approximate evaluation for arbitrary prompts and introduces Model Predictive Prompt Selection (MoPPS), a Bayesian risk-predictive framework that online estimates prompt difficulty without requiring costly LLM interactions. Technically, MoPPS models each prompt's success rate as a latent variable, performs streaming Bayesian inference, and employs posterior sampling in a constructed multi-armed bandit machine, enabling sample efficient and adaptive prompt selection. Extensive experiments across mathematics, planning, and vision-based geometry tasks show that MoPPS reliably predicts prompt difficulty and accelerates training with significantly reduced LLM rollouts.

Transforming multimodal sources into condensed and structured media outputs can be fundamentally conceptualized as a long-horizon agentic process centered on a model-harness system. While an ideal harness system should align with human design priors and accumulate reusable experience through empirical exploration to drive recursive self-improvement, existing paradigms remain static and fall short of this capability. In this paper, we present AutoDesign, a framework that aligns with human design priors, where a meta-harness optimizer guides a code agent to recursively improve harness based on rollout feedback. To instantiate and evaluate this framework, we focus on the academic paper-to-poster generation task and introduce PosterBench, comprising a 100-paper Main Track spanning five disciplines and PosterBench-mini, a shared 10-paper subset for controlled evaluation. On the PosterBench Main Track, AutoDesign achieves the highest score of 78.32, surpassing the closed-source commercial system Claude Design by 7.45 points. Across seven controlled code-agent-model configurations, integrating the learned DesignHarness consistently improves performance, increasing the average PosterBench Score from 54.99 to 67.39 (+12.4%). In a fully autonomous long-horizon loop, it executes 253 tool calls and 11 editing turns within 40 minutes for under $3, reaching average conference-poster quality in human evaluation. A system-blind human study further demonstrates that AutoDesign achieves the highest human preference among evaluated systems.

Since the emergence of the Transformer architecture, language model development has increased, driven by their promising potential. However, releasing these models into production requires properly understanding their behavior, particularly in sensitive domains such as medicine. Despite this need, the medical literature still lacks technical assessments of pre-trained language models, which are especially valuable in resource-constrained settings in terms of computational power or limited budget. To address this gap, we provide a comprehensive survey of language models in the medical domain. In addition, we selected a subset of these models for thorough evaluation, focusing on classification and text generation tasks. Our subset encompasses 53 models, ranging from 110 million to 13 billion parameters, spanning the three families of Transformer-based models and from diverse knowledge domains. This study employs a series of approaches for text classification together with zero-shot prompting instead of model training or fine-tuning, which closely resembles the limited resource setting in which many users of language models find themselves. Encouragingly, our findings reveal remarkable performance across various tasks and datasets, underscoring the latent potential of certain models to contain medical knowledge, even without domain specialization. Consequently, our study advocates for further exploration of model applications in medical contexts, particularly in resource-constrained settings. The code is available on https://github.com/anpoc/Language-models-in-medicine.

We introduce two complementary techniques for efficient adaptive optimization that reduce memory requirements while accelerating training of large-scale neural networks. The first technique, Subset-Norm adaptive step size, generalizes AdaGrad-Norm and AdaGrad(-Coordinate) by reducing the second moment term's memory footprint from O(d) to O(d) through step-size sharing, where d is the model size. For non-convex smooth objectives under coordinate-wise sub-gaussian gradient noise, we prove a noise-adapted high-probability convergence guarantee showing improved dimensional dependence over existing methods. Our second technique, Subspace-Momentum, reduces the momentum state's memory footprint by operating in a low-dimensional subspace while applying standard SGD in the orthogonal complement. We establish high-probability convergence rates under similar relaxed assumptions. Empirical evaluation on LLaMA models from 60M to 1B parameters demonstrates the effectiveness of our methods, where combining subset-norm with subspace-momentum achieves Adam's validation perplexity in approximately half the training tokens (6.8B vs 13.1B) while using only 20% of the Adam's optimizer-states memory footprint and requiring minimal additional hyperparameter tuning.

As large language models (LLMs) become increasingly versatile, numerous large scale benchmarks have been developed to thoroughly assess their capabilities. These benchmarks typically consist of diverse datasets and prompts to evaluate different aspects of LLM performance. However, comprehensive evaluations on hundreds or thousands of prompts incur tremendous costs in terms of computation, money, and time. In this work, we investigate strategies to improve evaluation efficiency by selecting a subset of examples from each benchmark using a learned policy. Our approach models the dependencies across test examples, allowing accurate prediction of the evaluation outcomes for the remaining examples based on the outcomes of the selected ones. Consequently, we only need to acquire the actual evaluation outcomes for the selected subset. We rigorously explore various subset selection policies and introduce a novel RL-based policy that leverages the captured dependencies. Empirical results demonstrate that our approach significantly reduces the number of evaluation prompts required while maintaining accurate performance estimates compared to previous methods.

Interpretability of Deep Learning (DL) is a barrier to trustworthy AI. Despite great efforts made by the Explainable AI (XAI) community, explanations lack robustness -- indistinguishable input perturbations may lead to different XAI results. Thus, it is vital to assess how robust DL interpretability is, given an XAI method. In this paper, we identify several challenges that the state-of-the-art is unable to cope with collectively: i) existing metrics are not comprehensive; ii) XAI techniques are highly heterogeneous; iii) misinterpretations are normally rare events. To tackle these challenges, we introduce two black-box evaluation methods, concerning the worst-case interpretation discrepancy and a probabilistic notion of how robust in general, respectively. Genetic Algorithm (GA) with bespoke fitness function is used to solve constrained optimisation for efficient worst-case evaluation. Subset Simulation (SS), dedicated to estimate rare event probabilities, is used for evaluating overall robustness. Experiments show that the accuracy, sensitivity, and efficiency of our methods outperform the state-of-the-arts. Finally, we demonstrate two applications of our methods: ranking robust XAI methods and selecting training schemes to improve both classification and interpretation robustness.

Subset selection-based methods are widely used to explain deep vision models: they attribute predictions by highlighting the most influential image regions and support object-level explanations. While these methods perform well in in-distribution (ID) settings, their behavior under out-of-distribution (OOD) conditions remains poorly understood. Through extensive experiments across multiple ID-OOD sets, we find that reliability of the existing subset based methods degrades markedly, yielding redundant, unstable, and uncertainty-sensitive explanations. To address these shortcomings, we introduce a framework that combines submodular subset selection with layer-wise, gradient-based uncertainty estimation to improve robustness and fidelity without requiring additional training or auxiliary models. Our approach estimates uncertainty via adaptive weight perturbations and uses these estimates to guide submodular optimization, ensuring diverse and informative subset selection. Empirical evaluations show that, beyond mitigating the weaknesses of existing methods under OOD scenarios, our framework also yields improvements in ID settings. These findings highlight limitations of current subset-based approaches and demonstrate how uncertainty-driven optimization can enhance attribution and object-level interpretability, paving the way for more transparent and trustworthy AI in real-world vision applications.

Language models (LMs) are becoming the foundation for almost all major language technologies, but their capabilities, limitations, and risks are not well understood. We present Holistic Evaluation of Language Models (HELM) to improve the transparency of language models. First, we taxonomize the vast space of potential scenarios (i.e. use cases) and metrics (i.e. desiderata) that are of interest for LMs. Then we select a broad subset based on coverage and feasibility, noting what's missing or underrepresented (e.g. question answering for neglected English dialects, metrics for trustworthiness). Second, we adopt a multi-metric approach: We measure 7 metrics (accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency) for each of 16 core scenarios when possible (87.5% of the time). This ensures metrics beyond accuracy don't fall to the wayside, and that trade-offs are clearly exposed. We also perform 7 targeted evaluations, based on 26 targeted scenarios, to analyze specific aspects (e.g. reasoning, disinformation). Third, we conduct a large-scale evaluation of 30 prominent language models (spanning open, limited-access, and closed models) on all 42 scenarios, 21 of which were not previously used in mainstream LM evaluation. Prior to HELM, models on average were evaluated on just 17.9% of the core HELM scenarios, with some prominent models not sharing a single scenario in common. We improve this to 96.0%: now all 30 models have been densely benchmarked on the same core scenarios and metrics under standardized conditions. Our evaluation surfaces 25 top-level findings. For full transparency, we release all raw model prompts and completions publicly for further analysis, as well as a general modular toolkit. We intend for HELM to be a living benchmark for the community, continuously updated with new scenarios, metrics, and models.

Understanding and reasoning on the web-scale scientific literature is a crucial touchstone for large language model (LLM) based agents designed to support complex knowledge-intensive tasks. However, existing works are mainly restricted to tool-free tasks within isolated papers, largely due to the lack of a benchmark for cross-paper reasoning and multi-tool orchestration in real research scenarios. In this work, we propose PaperArena, an evaluation benchmark for agents to address real-world research questions that typically require integrating information across multiple papers with the assistance of external tools. Given a research question, agents should integrate diverse formats across multiple papers through reasoning and interacting with appropriate tools, thereby producing a well-grounded answer. To support standardized evaluation, we provide a modular and extensible platform for agent execution, offering tools such as multimodal parsing, context retrieval, and programmatic computation. Experimental results reveal that even the most advanced LLM powering a well-established agent system achieves merely 38.78% average accuracy. On the hard subset, accuracy drops to only 18.47%, highlighting great potential for improvement. We also present several empirical findings, including that all agents tested exhibit inefficient tool usage, often invoking more tools than necessary to solve a task. We invite the community to adopt PaperArena to develop and evaluate more capable agents for scientific discovery. Our code and data are available https://github.com/Melmaphother/PaperArena.

Sparse autoencoders (SAEs) bricken2023monosemanticity,gao2024scalingevaluatingsparseautoencoders rely on dictionary learning to extract interpretable features from neural networks at scale in an unsupervised manner, with applications to representation engineering and information retrieval. SAEs are, however, computationally expensive lieberum2024gemmascopeopensparse, especially when multiple SAEs of different sizes are needed. We show that dictionary importance in vanilla SAEs follows a power law. We compare progressive coding based on subset pruning of SAEs -- to jointly training nested SAEs, or so-called {\em Matryoshka} SAEs bussmann2024learning,nabeshima2024Matryoshka -- on a language modeling task. We show Matryoshka SAEs exhibit lower reconstruction loss and recaptured language modeling loss, as well as higher representational similarity. Pruned vanilla SAEs are more interpretable, however. We discuss the origins and implications of this trade-off.

Learning neural subset selection tasks, such as compound selection in AI-aided drug discovery, have become increasingly pivotal across diverse applications. The existing methodologies in the field primarily concentrate on constructing models that capture the relationship between utility function values and subsets within their respective supersets. However, these approaches tend to overlook the valuable information contained within the superset when utilizing neural networks to model set functions. In this work, we address this oversight by adopting a probabilistic perspective. Our theoretical findings demonstrate that when the target value is conditioned on both the input set and subset, it is essential to incorporate an invariant sufficient statistic of the superset into the subset of interest for effective learning. This ensures that the output value remains invariant to permutations of the subset and its corresponding superset, enabling identification of the specific superset from which the subset originated. Motivated by these insights, we propose a simple yet effective information aggregation module designed to merge the representations of subsets and supersets from a permutation invariance perspective. Comprehensive empirical evaluations across diverse tasks and datasets validate the enhanced efficacy of our approach over conventional methods, underscoring the practicality and potency of our proposed strategies in real-world contexts.

Humans are social beings; we pursue social goals in our daily interactions, which is a crucial aspect of social intelligence. Yet, AI systems' abilities in this realm remain elusive. We present SOTOPIA, an open-ended environment to simulate complex social interactions between artificial agents and evaluate their social intelligence. In our environment, agents role-play and interact under a wide variety of scenarios; they coordinate, collaborate, exchange, and compete with each other to achieve complex social goals. We simulate the role-play interaction between LLM-based agents and humans within this task space and evaluate their performance with a holistic evaluation framework called SOTOPIA-Eval. With SOTOPIA, we find significant differences between these models in terms of their social intelligence, and we identify a subset of SOTOPIA scenarios, SOTOPIA-hard, that is generally challenging for all models. We find that on this subset, GPT-4 achieves a significantly lower goal completion rate than humans and struggles to exhibit social commonsense reasoning and strategic communication skills. These findings demonstrate SOTOPIA's promise as a general platform for research on evaluating and improving social intelligence in artificial agents.

As the demand for comprehensive evaluations of diverse model capabilities steadily increases, benchmark suites have correspondingly grown significantly in scale. Despite notable advances in redundancy reduction and subset-level performance prediction, a systematic framework that effectively integrates these methods to ensure both prediction accuracy and ranking consistency is still largely elusive. In this paper, we first perform a sample-level analysis of benchmark redundancy and identify several highly similar samples that can be eliminated. Besides, we frame benchmark compression as an optimization problem with the aim of score reconstruction. Building on these, we then propose EssenceBench, a coarse-to-fine framework utilizing an iterative Genetic Algorithm (GA), which takes the advantages of fitness-based subset search and attribution-based sample search. Compared to previous methods, our approach yields superior compression results with lower reconstruction error and markedly higher efficiency. In particular, on the HellaSwag benchmark (10K samples), our method preserves the ranking of all models shifting within 5% using 25x fewer samples, and achieves 95% ranking preservation shifting within 5% using only 200x fewer samples.

The rapid proliferation of large audio models (LAMs) demands efficient approaches for model comparison, yet comprehensive benchmarks are costly. To fill this gap, we investigate whether minimal subsets can reliably evaluate LAMs while reducing costs and data redundancy. Analyzing 10 subset selection methods with 18 audio models across 40 tasks covering major LAM evaluation dimensions, we show that subsets of just 50 examples (0.3% of data) can achieve over 0.93 Pearson correlation with full benchmark scores. To understand how well these scores align with what practitioners ultimately care about, user satisfaction, we collect 776 human preference ratings from realistic voice assistant conversations, finding that both subsets and full benchmark achieve only 0.85 correlation with human. To better predict preferences, we trained regression models on these selected subsets, achieving 0.98 correlation -- outperforming regression models trained on both random subsets and the full benchmark. This demonstrates that in regression modeling, well-curated subsets outpredict the full benchmark, showing quality over quantity. We open-source these regression-weighted subsets as the HUMANS benchmark, an efficient proxy for LAM evaluation that captures both benchmark performance and user preferences.

This paper presents a new state-of-the-art for document image classification and retrieval, using features learned by deep convolutional neural networks (CNNs). In object and scene analysis, deep neural nets are capable of learning a hierarchical chain of abstraction from pixel inputs to concise and descriptive representations. The current work explores this capacity in the realm of document analysis, and confirms that this representation strategy is superior to a variety of popular hand-crafted alternatives. Experiments also show that (i) features extracted from CNNs are robust to compression, (ii) CNNs trained on non-document images transfer well to document analysis tasks, and (iii) enforcing region-specific feature-learning is unnecessary given sufficient training data. This work also makes available a new labelled subset of the IIT-CDIP collection, containing 400,000 document images across 16 categories, useful for training new CNNs for document analysis.

We propose an alternate approach to quantifying how well language models learn natural language: we ask how well they match the statistical tendencies of natural language. To answer this question, we analyze whether text generated from language models exhibits the statistical tendencies present in the human-generated text on which they were trained. We provide a framework--paired with significance tests--for evaluating the fit of language models to these trends. We find that neural language models appear to learn only a subset of the tendencies considered, but align much more closely with empirical trends than proposed theoretical distributions (when present). Further, the fit to different distributions is highly-dependent on both model architecture and generation strategy. As concrete examples, text generated under the nucleus sampling scheme adheres more closely to the type--token relationship of natural language than text produced using standard ancestral sampling; text from LSTMs reflects the natural language distributions over length, stopwords, and symbols surprisingly well.

Evaluating modern machine learning models has become prohibitively expensive. Benchmarks such as LMMs-Eval and HELM demand thousands of GPU hours per model. Costly evaluation reduces inclusivity, slows the cycle of innovation, and worsens environmental impact. The typical approach follows two steps. First, select an anchor subset of data. Second, train a mapping from the accuracy on this subset to the final test result. The drawback is that anchor selection depends on clustering, which can be complex and sensitive to design choices. We argue that promoting diversity among samples is not essential; what matters is to select samples that maximise diversity in model responses. Our method, Diversifying Sample Condensation (DISCO), selects the top-k samples with the greatest model disagreements. This uses greedy, sample-wise statistics rather than global clustering. The approach is conceptually simpler. From a theoretical view, inter-model disagreement provides an information-theoretically optimal rule for such greedy selection. DISCO shows empirical gains over prior methods, achieving state-of-the-art results in performance prediction across MMLU, Hellaswag, Winogrande, and ARC. Code is available here: https://github.com/arubique/disco-public.

Hallucinations are a common issue that undermine the reliability of large language models (LLMs). Recent studies have identified a specific subset of hallucinations, known as confabulations, which arise due to predictive uncertainty of LLMs. To detect confabulations, various methods for estimating predictive uncertainty in natural language generation (NLG) have been developed. These methods are typically evaluated by correlating uncertainty estimates with the correctness of generated text, with question-answering (QA) datasets serving as the standard benchmark. However, commonly used approximate correctness functions have substantial disagreement between each other and, consequently, in the ranking of the uncertainty estimation methods. This allows one to inflate the apparent performance of uncertainty estimation methods. We propose using several alternative risk indicators for risk correlation experiments that improve robustness of empirical assessment of UE algorithms for NLG. For QA tasks, we show that marginalizing over multiple LLM-as-a-judge variants leads to reducing the evaluation biases. Furthermore, we explore structured tasks as well as out of distribution and perturbation detection tasks which provide robust and controllable risk indicators. Finally, we propose to use an Elo rating of uncertainty estimation methods to give an objective summarization over extensive evaluation settings.

We introduce ABLE, a benchmark for evaluating LLM agents' ability to use biological AI models (BAIMs), such as ProteinMPNN and AlphaFold3, in dual-use protein design workflows. ABLE assesses agent performance through a set of tasks spanning structure retrieval, sequence generation, and design validation. We evaluate 15 frontier models and find that seven refuse all tasks, while the remaining models exhibit substantial performance differences. Claude Sonnet 4 and Gemini 3 Pro achieve the highest scores across information retrieval, tool selection, and tool use. We further compare model performance on a subset of tasks against an expert human baseline. Our results suggest that current LLMs can substantially lower barriers to protein design, but remain inconsistent in planning, strategy generation, and integrating biological knowledge with tool use.

Accelerating research on Large Multimodal Models (LMMs) in non-English languages is crucial for enhancing user experiences across broader populations. In this paper, we introduce JMMMU (Japanese MMMU), the first large-scale Japanese benchmark designed to evaluate LMMs on expert-level tasks based on the Japanese cultural context. To facilitate comprehensive culture-aware evaluation, JMMMU features two complementary subsets: (i) culture-agnostic (CA) subset, where the culture-independent subjects (e.g., Math) are selected and translated into Japanese, enabling one-to-one comparison with its English counterpart MMMU; and (ii) culture-specific (CS) subset, comprising newly crafted subjects that reflect Japanese cultural context. Using the CA subset, we observe performance drop in many LMMs when evaluated in Japanese, which is purely attributable to language variation. Using the CS subset, we reveal their inadequate Japanese cultural understanding. Further, by combining both subsets, we identify that some LMMs perform well on the CA subset but not on the CS subset, exposing a shallow understanding of the Japanese language that lacks depth in cultural understanding. We hope this work will not only help advance LMM performance in Japanese but also serve as a guideline to create high-standard, culturally diverse benchmarks for multilingual LMM development. The project page is https://mmmu-japanese-benchmark.github.io/JMMMU/.

We present Plancraft, a multi-modal evaluation dataset for LLM agents. Plancraft has both a text-only and multi-modal interface, based on the Minecraft crafting GUI. We include the Minecraft Wiki to evaluate tool use and Retrieval Augmented Generation (RAG), as well as an oracle planner and oracle RAG information extractor, to ablate the different components of a modern agent architecture. To evaluate decision-making, Plancraft also includes a subset of examples that are intentionally unsolvable, providing a realistic challenge that requires the agent not only to complete tasks but also to decide whether they are solvable at all. We benchmark both open-source and closed-source LLMs and strategies on our task and compare their performance to a handcrafted planner. We find that LLMs and VLMs struggle with the planning problems that Plancraft introduces, and we offer suggestions on how to improve their capabilities.

We present TransEvalnia, a prompting-based translation evaluation and ranking system that uses reasoning in performing its evaluations and ranking. This system presents fine-grained evaluations based on a subset of the Multidimensional Quality Metrics (https://themqm.org/), returns an assessment of which translation it deems the best, and provides numerical scores for the various dimensions and for the overall translation. We show that TransEvalnia performs as well as or better than the state-of-the-art MT-Ranker (Moosa et al. 2024) on our own English-Japanese data as well as several language pairs from various WMT shared tasks. Using Anthropic's Claude-3.5-Sonnet and Qwen-2.5-72B-Instruct as the evaluation LLMs, we show that the evaluations returned are deemed highly acceptable to human raters, and that the scores assigned to the translations by Sonnet, as well as other LLMs, correlate well with scores assigned by the human raters. We also note the sensitivity of our system -- as well as MT-Ranker -- to the order in which the translations are presented, and we propose methods to address this position bias. All data, including the system's evaluation and reasoning, human assessments, as well as code is released.

Embodied intelligence is advancing rapidly, driving the need for efficient evaluation. Current benchmarks typically rely on interactive simulated environments or real-world setups, which are costly, fragmented, and hard to scale. To address this, we introduce StaticEmbodiedBench, a plug-and-play benchmark that enables unified evaluation using static scene representations. Covering 42 diverse scenarios and 8 core dimensions, it supports scalable and comprehensive assessment through a simple interface. Furthermore, we evaluate 19 Vision-Language Models (VLMs) and 11 Vision-Language-Action models (VLAs), establishing the first unified static leaderboard for Embodied intelligence. Moreover, we release a subset of 200 samples from our benchmark to accelerate the development of embodied intelligence.

Question Answering (QA) on narrative text poses a unique challenge to current systems, requiring a deep understanding of long, complex documents. However, the reliability of NarrativeQA, the most widely used benchmark in this domain, is hindered by noisy documents and flawed QA pairs. In this work, we introduce LiteraryQA, a high-quality subset of NarrativeQA focused on literary works. Using a human- and LLM-validated pipeline, we identify and correct low-quality QA samples while removing extraneous text from source documents. We then carry out a meta-evaluation of automatic metrics to clarify how systems should be evaluated on LiteraryQA. This analysis reveals that all n-gram-based metrics have a low system-level correlation to human judgment, while LLM-as-a-Judge evaluations, even with small open-weight models, can strongly agree with the ranking identified by humans. Finally, we benchmark a set of long-context LLMs on LiteraryQA. We release our code and data at https://github.com/SapienzaNLP/LiteraryQA.

Coding agents and LLM-powered applications routinely send potentially sensitive content to cloud LLM APIs where it may be logged, retained, used for training, or subpoenaed. Existing privacy tooling focuses on network-level encryption and organization-level DLP, neither of which addresses the content of prompts themselves. We present a systematic empirical evaluation of eight techniques for privacy-preserving LLM requests: (A) local-only inference, (B) redaction with placeholder restoration, (C) semantic rephrasing, (D) Trusted Execution Environment hosted inference, (E) split inference, (F) fully homomorphic encryption, (G) secret sharing via multi-party computation, and (H) differential-privacy noise. We implement all eight (or a tractable research-stage subset where deployment is not yet feasible) in an open-source shim compatible with MCP and any OpenAI-compatible API. We evaluate the four practical options (A, B, C, H) and their combinations across four workload classes using a ground-truth-labelled leak benchmark of 1,300 samples with 4,014 annotations. Our headline finding is that no single technique dominates: the combination A+B+C (route locally when possible, redact and rephrase the rest) achieves 0.6% combined leak on PII and 31.3% on proprietary code, with zero exact leaks on PII across 500 samples. We present a decision rule that selects the appropriate option(s) from a threat-model budget and workload characterisation. Code, benchmarks, and evaluation harness are released at https://github.com/jayluxferro/llm-redactor.

As AI systems are increasingly used to draft radiology reports, reliably evaluating their clinical quality remains a critical challenge. Large language model (LLM)-based metrics are now the best-correlated with radiologist judgment, yet they output a single opaque score that neither a clinician nor a model builder can easily interpret or audit. We introduce RadMatch, a multi-stage, LLM-based metric that decomposes report comparison into a structured finding-level matching with significance-aware scoring and error characterization across seven clinical attribute dimensions (status, location, severity, morphology, certainty, longitudinal comparison, and measurement). The main score is the actionable-error count, both interpretable and auditable. Candidate findings are graded correct, partial, or incorrect, and unmatched findings are counted as missed or hallucinated. Triage and actionable safety recall/precision and per-subset views add complementary, deployment-oriented lenses. Across two expert benchmarks, RadMatch is the most clinically aligned metric, matching inter-radiologist agreement on ReXVal and more than doubling the best prior metric on the harder RadEvalExpert. Relying only on few-shot prompting, it is designed to extend to other modalities and anatomies. We will release RadMatch as open-source code with an interactive dashboard for inspecting results.

During the development of large language models (LLMs), the scale and quality of the pre-training data play a crucial role in shaping LLMs' capabilities. To accelerate the research of LLMs, several large-scale datasets, such as C4 [1], Pile [2], RefinedWeb [3] and WanJuan [4], have been released to the public. However, most of the released corpus focus mainly on English, and there is still lack of complete tool-chain for extracting clean texts from web data. Furthermore, fine-grained information of the corpus, e.g. the quality of each text, is missing. To address these challenges, we propose in this paper a new complete tool-chain EvalWeb to extract Chinese clean texts from noisy web data. First, similar to previous work, manually crafted rules are employed to discard explicit noisy texts from the raw crawled web contents. Second, a well-designed evaluation model is leveraged to assess the remaining relatively clean data, and each text is assigned a specific quality score. Finally, we can easily utilize an appropriate threshold to select the high-quality pre-training data for Chinese. Using our proposed approach, we release the largest and latest large-scale high-quality Chinese web text ChineseWebText, which consists of 1.42 TB and each text is associated with a quality score, facilitating the LLM researchers to choose the data according to the desired quality thresholds. We also release a much cleaner subset of 600 GB Chinese data with the quality exceeding 90%.

Data science aims to extract insights from data to support decision-making processes. Recently, Large Language Models (LLMs) are increasingly used as assistants for data science, by suggesting ideas, techniques and small code snippets, or for the interpretation of results and reporting. Proper automation of some data-science activities is now promised by the rise of LLM agents, i.e., AI systems powered by an LLM equipped with additional affordances--such as code execution and knowledge bases--that can perform self-directed actions and interact with digital environments. In this paper, we survey the evaluation of LLM assistants and agents for data science. We find (1) a dominant focus on a small subset of goal-oriented activities, largely ignoring data management and exploratory activities; (2) a concentration on pure assistance or fully autonomous agents, without considering intermediate levels of human-AI collaboration; and (3) an emphasis on human substitution, therefore neglecting the possibility of higher levels of automation thanks to task transformation.

We introduce KMMMU, a native Korean benchmark for evaluating multimodal understanding in Korean cultural and institutional settings. KMMMU contains 3,466 questions from exams natively written in Korean, covering nine disciplines and nine visual modality categories, along with a 300-item Korean-specific subset and a hard subset of 627 questions. Unlike translated or English-centric benchmarks, KMMMU targets information-dense problems shaped by local conventions, official standards, and discipline-specific visual formats. Experiments show that the strongest open-source model reaches only 42.05% accuracy on the full set, while the best proprietary model achieves 52.42% on the hard subset. Performance varies across disciplines, with some disciplines emerging as bottlenecks, and Korean-specific questions showing gaps of up to 13.43%. Error analysis suggests that these failures stem less from insufficient reasoning depth than from weak convention-to-label mapping, few-shot symbolic induction, localized knowledge recall, and domain-specific standards understanding. KMMMU provides a testbed for multimodal evaluation beyond English-centric benchmarks and for developing more reliable systems for expert real-world tasks.

Evaluation of software engineering (SWE) agents is dominated by a binary signal: whether the final patch passes the tests. This outcome-only view treats a principled solution and a chaotic trial-and-error process as equivalent. We show that this equivalence is empirically false. We evaluate 2,614 OpenHands trajectories from eight model backends on 60 SWE-bench Verified tasks. Of these, 47 have enough passing trajectories to construct task-level process references, yielding a 1,815-trajectory evaluation subset. Among passing trajectories in this subset, 10.7% exhibit behavior we call a Lucky Pass: regression cycles, blind retries, missing verification, or temporally disordered exploration, implementation, and verification. We introduce AgentLens, a framework for process-level assessment of SWE-agent trajectories, and release AgentLens-Bench, a dataset of 1,815 trajectories annotated with quality scores, waste signals, divergence points, and 47 task-level Prefix Tree Acceptor (PTA) references. AgentLens builds PTA references by merging multiple passing solutions for the same task, and uses a context-sensitive intent labeler to assign actions to Exploration, Implementation, Verification, or Orchestration based on trajectory history rather than tool identity alone. On AgentLens-Bench, the quality score separates passing trajectories into Lucky, Solid, and Ideal tiers and further decomposes Lucky Passes into five recurring mechanisms. Across the eight model backends, Lucky rates range from 0.5% to 23.2%, and some models move by as many as five rank positions when ranked by quality score instead of pass rate. We release the anonymized project repository, including the AgentLens-Bench dataset and AgentLens SDK, at https://github.com/microsoft/code-agent-state-trajectories/.

We introduce IND-Eval, a public benchmark that organizes industrial time series across domains into source-specific forecasting tasks. Relative to established benchmarks such as GIFT-Eval, it emphasizes finer sampling resolutions and a greater prevalence of multivariate forecasting. Evaluating Chronos-2, TiRex-2, and TimesFM-3, we find that TimesFM-3 achieves the lowest source-balanced aggregate nMASE and nCRPS and leads seven of nine sources under each metric; source-specific exceptions reinforce the need for domain-representative evaluation for model selection. These findings also motivate improving domain alignment during pretraining, where real industrial data are often too scarce to close domain-specific gaps directly; we therefore investigate whether a domain's statistical structure can guide simulator design. In a case study on two offshore well-monitoring regimes, we improve a general simulator's coverage using an oscillator-augmented simulation pipeline. This lowers forecasting error in the target domain but reduces performance on the subset of GIFT-Eval with natively univariate targets. A 50/50 mixture of the general and augmented simulation recipes achieves the lowest error among the three training recipes for all four event-metric pairs in the case study while recovering most of the lost GIFT-Eval performance. Together, these results illustrate how domain-representative evaluation can guide model selection and the composition of simulated data while revealing a trade-off between specialization and generality.

Recently, there has been growing interest in using Large Language Models (LLMs) for scientific research. Numerous benchmarks have been proposed to evaluate the ability of LLMs for scientific research. However, current benchmarks are mostly based on pre-collected objective questions. This design suffers from data leakage problem and lacks the evaluation of subjective Q/A ability. In this paper, we propose SciEval, a comprehensive and multi-disciplinary evaluation benchmark to address these issues. Based on Bloom's taxonomy, SciEval covers four dimensions to systematically evaluate scientific research ability. In particular, we design a "dynamic" subset based on scientific principles to prevent evaluation from potential data leakage. Both objective and subjective questions are included in SciEval. These characteristics make SciEval a more effective benchmark for scientific research ability evaluation of LLMs. Comprehensive experiments on most advanced LLMs show that, although GPT-4 achieves SOTA performance compared to other LLMs, there is still substantial room for improvement, especially for dynamic questions. The data and codes are now publicly available.

Our goal is to assess if AutoML system changes - i.e., to the search space or hyperparameter optimization - will improve the final model's performance on production tasks. However, we cannot test the changes on production tasks. Instead, we only have access to limited descriptors about tasks that our AutoML system previously executed, like the number of data points or features. We also have a set of development tasks to test changes, ex., sampled from OpenML with no usage constraints. However, the development and production task distributions are different leading us to pursue changes that only improve development and not production. This paper proposes a method to leverage descriptor information about AutoML production tasks to select a filtered subset of the most relevant development tasks. Empirical studies show that our filtering strategy improves the ability to assess AutoML system changes on holdout tasks with different distributions than development.

One of the most challenging problems facing NLP today is evaluation. Some of the most pressing issues pertain to benchmark saturation, data contamination, and diversity in the quality of test examples. To address these concerns, we propose Selection Methodology for Accurate, Reduced, and Targeted (SMART) filtering, a novel approach to select a high-quality subset of examples from existing benchmark datasets by systematically removing less informative and less challenging examples. Our approach applies three filtering criteria, removing (i) easy examples, (ii) data-contaminated examples, and (iii) examples that are similar to each other based on distance in an embedding space. We demonstrate the effectiveness of SMART on three multiple choice QA datasets, where our methodology increases efficiency by reducing dataset size by 48\% on average, while increasing Pearson correlation with rankings from ChatBot Arena, a more open-ended human evaluation setting. Our method enables us to be more efficient, whether using SMART to make new benchmarks more challenging or to revitalize older datasets, while still preserving the relative model rankings.

Cultural biases in multilingual datasets pose significant challenges for their effectiveness as global benchmarks. These biases stem not only from language but also from the cultural knowledge required to interpret questions, reducing the practical utility of translated datasets like MMLU. Furthermore, translation often introduces artifacts that can distort the meaning or clarity of questions in the target language. A common practice in multilingual evaluation is to rely on machine-translated evaluation sets, but simply translating a dataset is insufficient to address these challenges. In this work, we trace the impact of both of these issues on multilingual evaluations and ensuing model performances. Our large-scale evaluation of state-of-the-art open and proprietary models illustrates that progress on MMLU depends heavily on learning Western-centric concepts, with 28% of all questions requiring culturally sensitive knowledge. Moreover, for questions requiring geographic knowledge, an astounding 84.9% focus on either North American or European regions. Rankings of model evaluations change depending on whether they are evaluated on the full portion or the subset of questions annotated as culturally sensitive, showing the distortion to model rankings when blindly relying on translated MMLU. We release Global-MMLU, an improved MMLU with evaluation coverage across 42 languages -- with improved overall quality by engaging with compensated professional and community annotators to verify translation quality while also rigorously evaluating cultural biases present in the original dataset. This comprehensive Global-MMLU set also includes designated subsets labeled as culturally sensitive and culturally agnostic to allow for more holistic, complete evaluation.

In this thesis, I evaluate the performance of Large Language Models (LLMs) on the Law School Admissions Test (LSAT), specifically the Logic Games section of the test. I focus on this section because it presents a complex logical reasoning task and thus is a valuable source of data for evaluating how modern, increasingly capable LLMs can handle hard logical reasoning tasks. I construct a dataset of LSAT logic games and their associated metadata, and extensively evaluate LLMs' performance in a Chain-of-Thought prompting setting. Given the weak performance in this setting, I explore other prompting frameworks on a smaller subset of the dataset, adapting ideas from Reflexion to this task. This results in a substantially improved accuracy of 70 percent for GPT-4 and 46 percent for GPT-3.5 on this data subset, highlighting the capacity of LLMs to revise their logical errors, despite initially weak performance. Finally, I analyze the types of logic games that models perform better or worse on, as well as the types of logical errors I observe from human annotation, providing detailed insights on the logical reasoning capabilities of LLMs.

Comprehensive evaluations of language models (LM) during both development and deployment phases are necessary because these models possess numerous capabilities (e.g., mathematical reasoning, legal support, or medical diagnostic) as well as safety risks (e.g., racial bias, toxicity, or misinformation). The average score across a wide range of benchmarks provides a signal that helps guide the use of these LMs in practice. Currently, holistic evaluations are costly due to the large volume of benchmark questions, making frequent evaluations impractical. A popular attempt to lower the cost is to compute the average score on a subset of the benchmark. This approach, unfortunately, often renders an unreliable measure of LM performance because the average score is often confounded with the difficulty of the questions in the benchmark subset. Item response theory (IRT) was designed to address this challenge, providing a reliable measurement by careful controlling for question difficulty. Unfortunately, question difficulty is expensive to estimate. Facing this challenge, we train a model that predicts question difficulty from its content, enabling a reliable measurement at a fraction of the cost. In addition, we leverage this difficulty predictor to further improve the evaluation efficiency through training a question generator given a difficulty level. This question generator is essential in adaptive testing, where, instead of using a random subset of the benchmark questions, informative questions are adaptively chosen based on the current estimation of LLM performance. Experiments on 22 common natural language benchmarks and 172 LMs show that this approach is more reliable and efficient compared to current common practice.

Search has emerged as core infrastructure for LLM-based agents and is widely viewed as critical on the path toward more general intelligence. Finance is a particularly demanding proving ground: analysts routinely conduct complex, multi-step searches over time-sensitive, domain-specific data, making it ideal for assessing both search proficiency and knowledge-grounded reasoning. Yet no existing open financial datasets evaluate data searching capability of end-to-end agents, largely because constructing realistic, complicated tasks requires deep financial expertise and time-sensitive data is hard to evaluate. We present FinSearchComp, the first fully open-source agent benchmark for realistic, open-domain financial search and reasoning. FinSearchComp comprises three tasks -- Time-Sensitive Data Fetching, Simple Historical Lookup, and Complex Historical Investigation -- closely reproduce real-world financial analyst workflows. To ensure difficulty and reliability, we engage 70 professional financial experts for annotation and implement a rigorous multi-stage quality-assurance pipeline. The benchmark includes 635 questions spanning global and Greater China markets, and we evaluate 21 models (products) on it. Grok 4 (web) tops the global subset, approaching expert-level accuracy. DouBao (web) leads on the Greater China subset. Experimental analyses show that equipping agents with web search and financial plugins substantially improves results on FinSearchComp, and the country origin of models and tools impact performance significantly.By aligning with realistic analyst tasks and providing end-to-end evaluation, FinSearchComp offers a professional, high-difficulty testbed for complex financial search and reasoning.

Safety alignment mechanisms in large language models prevent responses to harmful queries through learned refusal behavior, yet these same mechanisms impede legitimate research applications including cognitive modeling, adversarial testing, and security analysis. While abliteration techniques enable surgical removal of refusal representations through directional orthogonalization, the relative effectiveness of available implementations remains uncharacterized. This study evaluates four abliteration tools (Heretic, DECCP, ErisForge, FailSpy) across sixteen instruction-tuned models (7B-14B parameters), reporting tool compatibility on all 16 models and quantitative metrics on subsets dictated by tool support. Single-pass methods demonstrated superior capability preservation on the benchmarked subset (avg GSM8K change across three models: ErisForge -0.28 pp; DECCP -0.13 pp), while Bayesian-optimized abliteration produced variable distribution shift (KL divergence: 0.043-1.646) with model-dependent capability impact. These findings provide researchers with evidence-based selection criteria for abliteration tool deployment across diverse model architectures. The principal finding indicates that mathematical reasoning capabilities exhibit the highest sensitivity to abliteration interventions, with GSM8K change ranging from +1.51 pp to -18.81 pp (-26.5% relative) depending on tool selection and model architecture.

Large Language Models (LLMs) and other automated techniques have been increasingly used to support software developers by generating software artifacts such as code snippets, patches, and comments. However, accurately assessing the correctness of these generated artifacts remains a significant challenge. On one hand, human evaluation provides high accuracy but is labor-intensive and lacks scalability. On the other hand, many automatic evaluation metrics are scalable and require minimal human effort, but they often fail to accurately reflect the actual correctness of generated software artifacts. In this paper, we present SE-Jury, the first evaluation metric for LLM-as-Ensemble-Judge specifically designed to accurately assess the correctness of generated software artifacts. SE-Jury first defines five distinct evaluation strategies, each implemented by an independent judge. A dynamic team selection mechanism then identifies the most appropriate subset of judges as a team to produce a final correctness score through ensembling. We evaluate SE-Jury across a diverse set of software engineering (SE) benchmarks that span three popular SE tasks: code generation, automated program repair, and code summarization. Results demonstrate that SE-Jury consistently achieves a higher correlation with human judgments, with improvements ranging from 29.6% to 140.8% over existing automatic metrics. SE-Jury reaches agreement levels with human annotators that are close to inter-annotator agreement in code generation and program repair. These findings underscore SE-Jury's potential as a scalable and reliable alternative to human evaluation in these SE tasks.

The rapid advancement of AIGC-based video generation has underscored the critical need for comprehensive evaluation frameworks that go beyond traditional generation quality metrics to encompass aesthetic appeal. However, existing benchmarks remain largely focused on technical fidelity, leaving a significant gap in holistic assessment-particularly with respect to perceptual and artistic qualities. To address this limitation, we introduce VGA-Bench, a unified benchmark for joint evaluation of video generation quality and aesthetic quality. VGA-Bench is built upon a principled three-tier taxonomy: Aesthetic Quality, Aesthetic Tagging, and Generation Quality, each decomposed into multiple fine-grained sub-dimensions to enable systematic assessment. Guided by this taxonomy, we design 1,016 diverse prompts and generate a large-scale dataset of over 60,000 videos using 12 video generation models, ensuring broad coverage across content, style, and artifacts. To enable scalable and automated evaluation, we annotate a subset of the dataset via human labeling and develop three dedicated multi-task neural assessors: VAQA-Net for aesthetic quality prediction, VTag-Net for automatic aesthetic tagging, and VGQA-Net for generation and basic quality attributes. Extensive experiments demonstrate that our models achieve reliable alignment with human judgments, offering both accuracy and efficiency. We release VGA-Bench as a public benchmark to foster research in AIGC evaluation, with applications in content moderation, model debugging, and generative model optimization.

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

Large language models (LLMs) are advancing rapidly in medical NLP, yet Traditional Chinese Medicine (TCM) with its distinctive ontology, terminology, and reasoning patterns requires domain-faithful evaluation. Existing TCM benchmarks are fragmented in coverage and scale and rely on non-unified or generation-heavy scoring that hinders fair comparison. We present the LingLanMiDian (LingLan) benchmark, a large-scale, expert-curated, multi-task suite that unifies evaluation across knowledge recall, multi-hop reasoning, information extraction, and real-world clinical decision-making. LingLan introduces a consistent metric design, a synonym-tolerant protocol for clinical labels, a per-dataset 400-item Hard subset, and a reframing of diagnosis and treatment recommendation into single-choice decision recognition. We conduct comprehensive, zero-shot evaluations on 14 leading open-source and proprietary LLMs, providing a unified perspective on their strengths and limitations in TCM commonsense knowledge understanding, reasoning, and clinical decision support; critically, the evaluation on Hard subset reveals a substantial gap between current models and human experts in TCM-specialized reasoning. By bridging fundamental knowledge and applied reasoning through standardized evaluation, LingLan establishes a unified, quantitative, and extensible foundation for advancing TCM LLMs and domain-specific medical AI research. All evaluation data and code are available at https://github.com/TCMAI-BJTU/LingLan and http://tcmnlp.com.

There is a growing body of work in recent years to develop pre-trained language models (PLMs) for the Arabic language. This work concerns addressing two major problems in existing Arabic PLMs which constraint progress of the Arabic NLU and NLG fields.First, existing Arabic PLMs are not well-explored and their pre-trainig can be improved significantly using a more methodical approach. Second, there is a lack of systematic and reproducible evaluation of these models in the literature. In this work, we revisit both the pre-training and evaluation of Arabic PLMs. In terms of pre-training, we explore improving Arabic LMs from three perspectives: quality of the pre-training data, size of the model, and incorporating character-level information. As a result, we release three new Arabic BERT-style models ( JABER, Char-JABER, and SABER), and two T5-style models (AT5S and AT5B). In terms of evaluation, we conduct a comprehensive empirical study to systematically evaluate the performance of existing state-of-the-art models on ALUE that is a leaderboard-powered benchmark for Arabic NLU tasks, and on a subset of the ARGEN benchmark for Arabic NLG tasks. We show that our models significantly outperform existing Arabic PLMs and achieve a new state-of-the-art performance on discriminative and generative Arabic NLU and NLG tasks. Our models and source code to reproduce of results will be made available shortly.

New NLP benchmarks are urgently needed to align with the rapid development of large language models (LLMs). We present C-Eval, the first comprehensive Chinese evaluation suite designed to assess advanced knowledge and reasoning abilities of foundation models in a Chinese context. C-Eval comprises multiple-choice questions across four difficulty levels: middle school, high school, college, and professional. The questions span 52 diverse disciplines, ranging from humanities to science and engineering. C-Eval is accompanied by C-Eval Hard, a subset of very challenging subjects in C-Eval that requires advanced reasoning abilities to solve. We conduct a comprehensive evaluation of the most advanced LLMs on C-Eval, including both English- and Chinese-oriented models. Results indicate that only GPT-4 could achieve an average accuracy of over 60%, suggesting that there is still significant room for improvement for current LLMs. We anticipate C-Eval will help analyze important strengths and shortcomings of foundation models, and foster their development and growth for Chinese users.

Recent advances in Large Language Models (LLMs) have spurred interest in designing LLM-based agents for tasks that involve interaction with human and artificial agents. This paper addresses a key aspect in the design of such agents: Predicting human decision in off-policy evaluation (OPE), focusing on language-based persuasion games, where the agent's goal is to influence its partner's decisions through verbal messages. Using a dedicated application, we collected a dataset of 87K decisions from humans playing a repeated decision-making game with artificial agents. Our approach involves training a model on human interactions with one agents subset to predict decisions when interacting with another. To enhance off-policy performance, we propose a simulation technique involving interactions across the entire agent space and simulated decision makers. Our learning strategy yields significant OPE gains, e.g., improving prediction accuracy in the top 15% challenging cases by 7.1%. Our code and the large dataset we collected and generated are submitted as supplementary material and publicly available in our GitHub repository: https://github.com/eilamshapira/HumanChoicePrediction

Large vision-language models (LVLMs) have demonstrated remarkable achievements, yet the generation of non-factual responses remains prevalent in fact-seeking question answering (QA). Current multimodal fact-seeking benchmarks primarily focus on comparing model outputs to ground truth answers, providing limited insights into the performance of modality-specific modules. To bridge this gap, we introduce VisualSimpleQA, a multimodal fact-seeking benchmark with two key features. First, it enables streamlined and decoupled evaluation of LVLMs in visual and linguistic modalities. Second, it incorporates well-defined difficulty criteria to guide human annotation and facilitates the extraction of a challenging subset, VisualSimpleQA-hard. Experiments on 15 LVLMs show that even state-of-the-art models such as GPT-4o achieve merely 60%+ correctness in multimodal fact-seeking QA on VisualSimpleQA and 30%+ on VisualSimpleQA-hard. Furthermore, the decoupled evaluation across these models highlights substantial opportunities for improvement in both visual and linguistic modules. The dataset is available at https://huggingface.co/datasets/WYLing/VisualSimpleQA.

With the rapid adoption of LLM-based chatbots, there is a pressing need to evaluate what humans and LLMs can achieve together. However, standard benchmarks, such as MMLU, measure LLM capabilities in isolation (i.e., "AI-alone"). Here, we design and conduct a user study to convert MMLU questions into user-AI conversations, by seeding the user with the question and having them carry out a conversation with the LLM to answer their question. We release ChatBench, a new dataset with AI-alone, user-alone, and user-AI data for 396 questions and two LLMs, including 144K answers and 7,336 user-AI conversations. We find that AI-alone accuracy fails to predict user-AI accuracy, with significant differences across multiple subjects (math, physics, and moral reasoning), and we analyze the user-AI conversations to provide insight into how they diverge from AI-alone benchmarks. Finally, we show that fine-tuning a user simulator on a subset of ChatBench improves its ability to estimate user-AI accuracies, increasing correlation on held-out questions by more than 20 points, creating possibilities for scaling interactive evaluation.

Multimodal large language models (MLLMs) have achieved impressive success in question-answering tasks, yet their capabilities for spatial understanding are less explored. This work investigates a critical question: do existing MLLMs possess 3D spatial perception and understanding abilities? Concretely, we make the following contributions in this paper: (i) we introduce VGBench, a benchmark specifically designed to assess MLLMs for visual geometry perception, e.g., camera pose and motion estimation; (ii) we propose SpatialScore, the most comprehensive and diverse multimodal spatial understanding benchmark to date, integrating VGBench with relevant data from the other 11 existing datasets. This benchmark comprises 28K samples across various spatial understanding tasks, modalities, and QA formats, along with a carefully curated challenging subset, SpatialScore-Hard; (iii) we develop SpatialAgent, a novel multi-agent system incorporating 9 specialized tools for spatial understanding, supporting both Plan-Execute and ReAct reasoning paradigms; (iv) we conduct extensive evaluations to reveal persistent challenges in spatial reasoning while demonstrating the effectiveness of SpatialAgent. We believe SpatialScore will offer valuable insights and serve as a rigorous benchmark for the next evolution of MLLMs.

Speech synthesis has come a long way as current text-to-speech (TTS) models can now generate natural human-sounding speech. However, most of the TTS research focuses on using adult speech data and there has been very limited work done on child speech synthesis. This study developed and validated a training pipeline for fine-tuning state-of-the-art (SOTA) neural TTS models using child speech datasets. This approach adopts a multi-speaker TTS retuning workflow to provide a transfer-learning pipeline. A publicly available child speech dataset was cleaned to provide a smaller subset of approximately 19 hours, which formed the basis of our fine-tuning experiments. Both subjective and objective evaluations were performed using a pretrained MOSNet for objective evaluation and a novel subjective framework for mean opinion score (MOS) evaluations. Subjective evaluations achieved the MOS of 3.95 for speech intelligibility, 3.89 for voice naturalness, and 3.96 for voice consistency. Objective evaluation using a pretrained MOSNet showed a strong correlation between real and synthetic child voices. Speaker similarity was also verified by calculating the cosine similarity between the embeddings of utterances. An automatic speech recognition (ASR) model is also used to provide a word error rate (WER) comparison between the real and synthetic child voices. The final trained TTS model was able to synthesize child-like speech from reference audio samples as short as 5 seconds.

Evaluating software engineering agents on realistic benchmarks is costly, since each task may require multi-step code exploration, modification, and test execution. Existing efficient evaluation methods select representative subsets to estimate full-benchmark performance, but are largely result-only: they fit historical pass/fail response matrices or static task semantics, discarding how agents solve problems. We propose PTA-IRT, a Privileged Trajectory-Aware Item Response Theory framework that fuses process and outcome signals. Historical execution trajectories supply process-level evidence beyond pass/fail, such as explored context, attempted edits, and solving paths, which PTA-IRT uses as privileged information for calibration subset selection and ability estimation. Under low calibration budgets, PTA-IRT consistently outperforms prior IRT baselines on score and ranking recovery across four SWE benchmarks. Code and data are publicly available at https://github.com/DeepSoftwareAnalytics/PTA-IRT.

Human evaluation is the gold-standard for evaluating text generation models. It is also expensive, and to fit budgetary constraints, a random subset of the test data is often chosen in practice. The randomly selected data may not accurately represent test performance, making this approach economically inefficient for model comparison. Thus, in this work, we develop a suite of selectors to get the most informative datapoints for human evaluation while taking the evaluation costs into account. We show that selectors based on variance in automated metric scores, diversity in model outputs, or Item Response Theory outperform random selection. We further develop an approach to distill these selectors to the scenario where the model outputs are not yet available. In particular, we introduce source-based estimators, which predict item usefulness for human evaluation just based on the source texts. We demonstrate the efficacy of our selectors in two common NLG tasks, machine translation and summarization, and show that up to only ~50% of the test data is needed to produce the same evaluation result as the entire data. Our implementations are published in the subset2evaluate package.

Large Language Models (LLMs) have inherent limitations of faithfulness and factuality, commonly referred to as hallucinations. Several benchmarks have been developed that provide a test bed for factuality evaluation within the context of English-centric datasets, while relying on supplementary informative context like web links or text passages but ignoring the available structured factual resources. To this end, Knowledge Graphs (KGs) have been identified as a useful aid for hallucination mitigation, as they provide a structured way to represent the facts about entities and their relations with minimal linguistic overhead. We bridge the lack of KG paths and multilinguality for factual language modeling within the existing hallucination evaluation benchmarks and propose a KG-based multilingual, multihop benchmark called MultiHal framed for generative text evaluation. As part of our data collection pipeline, we mined 140k KG-paths from open-domain KGs, from which we pruned noisy KG-paths, curating a high-quality subset of 25.9k. Our baseline evaluation shows an absolute scale increase by approximately 0.12 to 0.36 points for the semantic similarity score in KG-RAG over vanilla QA across multiple languages and multiple models, demonstrating the potential of KG integration. We anticipate MultiHal will foster future research towards several graph-based hallucination mitigation and fact-checking tasks.

Large-scale video generative models, capable of creating realistic videos of diverse visual concepts, are strong candidates for general-purpose physical world simulators. However, their adherence to physical commonsense across real-world actions remains unclear (e.g., playing tennis, backflip). Existing benchmarks suffer from limitations such as limited size, lack of human evaluation, sim-to-real gaps, and absence of fine-grained physical rule analysis. To address this, we introduce VideoPhy-2, an action-centric dataset for evaluating physical commonsense in generated videos. We curate 200 diverse actions and detailed prompts for video synthesis from modern generative models. We perform human evaluation that assesses semantic adherence, physical commonsense, and grounding of physical rules in the generated videos. Our findings reveal major shortcomings, with even the best model achieving only 22% joint performance (i.e., high semantic and physical commonsense adherence) on the hard subset of VideoPhy-2. We find that the models particularly struggle with conservation laws like mass and momentum. Finally, we also train VideoPhy-AutoEval, an automatic evaluator for fast, reliable assessment on our dataset. Overall, VideoPhy-2 serves as a rigorous benchmark, exposing critical gaps in video generative models and guiding future research in physically-grounded video generation. The data and code is available at https://videophy2.github.io/.

With LLMs shifting their role from statistical modeling of language to serving as general-purpose AI agents, how should LLM evaluations change? Arguably, a key ability of an AI agent is to flexibly combine, as needed, the basic skills it has learned. The capability to combine skills plays an important role in (human) pedagogy and also in a paper on emergence phenomena (Arora & Goyal, 2023). This work introduces Skill-Mix, a new evaluation to measure ability to combine skills. Using a list of N skills the evaluator repeatedly picks random subsets of k skills and asks the LLM to produce text combining that subset of skills. Since the number of subsets grows like N^k, for even modest k this evaluation will, with high probability, require the LLM to produce text significantly different from any text in the training set. The paper develops a methodology for (a) designing and administering such an evaluation, and (b) automatic grading (plus spot-checking by humans) of the results using GPT-4 as well as the open LLaMA-2 70B model. Administering a version of to popular chatbots gave results that, while generally in line with prior expectations, contained surprises. Sizeable differences exist among model capabilities that are not captured by their ranking on popular LLM leaderboards ("cramming for the leaderboard"). Furthermore, simple probability calculations indicate that GPT-4's reasonable performance on k=5 is suggestive of going beyond "stochastic parrot" behavior (Bender et al., 2021), i.e., it combines skills in ways that it had not seen during training. We sketch how the methodology can lead to a Skill-Mix based eco-system of open evaluations for AI capabilities of future models.

General-purpose health benchmarks increasingly anchor claims about LLM medical performance, but they are not always resolved by clinical specialty, making domain-specific performance hard to isolate. Mental health is of acute public-health concern as millions of people turn to LLMs for psychological support, and most existing evaluations are bespoke academic benchmarks that are difficult to integrate into developer workflows. We introduce HealthBench-Psych and HealthBench-Psych-Hard. We screened HealthBench's 5,000 physician-rubric conversations for mental-health relevance with a transparent LLM-applied rubric, then validated the subset through two rounds of blinded clinician review with concealed known-exclude controls, yielding 610 conversations (12.2% of the corpus). Evaluating 20 frontier and open models under a cross-vendor panel of three LLM judges, we find a statistically tied frontier cluster, measurable refusal behavior in two models, and near-identical rankings across judges (τge 0.92). We release the subset, pipeline, model responses, grades, and analysis code as a reusable resource.

After the introduction of Large Language Models (LLMs), there have been substantial improvements in the performance of Natural Language Generation (NLG) tasks, including Text Summarization and Machine Translation. However, LLMs still produce outputs containing hallucinations, that is, content not grounded in factual information. Therefore, developing methods to assess the factuality of LLMs has become urgent. Indeed, resources for factuality evaluation have recently emerged. Although challenging, these resources face one or more of the following limitations: (i) they are tailored to a specific task or domain; (ii) they are limited in size, thereby preventing the training of new factuality evaluators; (iii) they are designed for simpler verification tasks, such as claim verification. To address these issues, we introduce LLM-Oasis, to the best of our knowledge the largest resource for training end-to-end factuality evaluators. LLM-Oasis is constructed by extracting claims from Wikipedia, falsifying a subset of these claims, and generating pairs of factual and unfactual texts. We then rely on human annotators to both validate the quality of our dataset and to create a gold standard test set for benchmarking factuality evaluation systems. Our experiments demonstrate that LLM-Oasis presents a significant challenge for state-of-the-art LLMs, with GPT-4o achieving up to 60% accuracy in our proposed end-to-end factuality evaluation task, highlighting its potential to drive future research in the field.

Large language models (LLMs) are widely used to evaluate the quality of LLM generations and responses, but this leads to significant challenges: high API costs, uncertain reliability, inflexible pipelines, and inherent biases. To address these, we introduce PAJAMA (Program-As-a-Judge for Automated Model Assessment), a new alternative that uses LLMs to synthesize executable judging programs instead of directly scoring responses. These synthesized programs can be stored and run locally, costing orders of magnitude less while providing interpretable, and auditable judging logic that can be easily adapted. Program-based judges mitigate biases, improving judgment consistency by 15.83% and reducing biased responses by 23.7% on average compared to a Qwen2.5-14B-based LLM-as-a-judge. When program judgments are distilled into a model, PAJAMA outperforms LLM-as-a-judge on the challenging CHAT-HARD subset of RewardBench, outperforming metrics by 2.19% on Prometheus and 8.67% on the JudgeLM dataset, all at three orders of magnitude lower cost.

Benchmarks are essential for evaluating and improving large language models (LLMs). Yet many widely used benchmarks are redundant, making evaluation unnecessarily expensive. Benchmark compression lowers this cost by estimating full-benchmark scores and rankings from a partial evaluation. Research in this area has grown rapidly since 2024, but to our knowledge no survey has systematically organized its methods and findings. We review 119 papers to explain how methods choose what to evaluate, recover full-benchmark results, and verify their reliability. From this analysis, we develop a taxonomy covering compression within individual benchmarks, across benchmark suites, and during adaptive evaluation, together with work on redundancy and reliability. We compare their information needs, total costs, reliability checks, and limitations. Methods reduce evaluation costs by selecting representative subsets, using cheaper or reusable information, and testing adaptively, but risk poor coverage, inaccurate estimates, and sequential delays. The resulting taxonomy and practical guide help researchers choose, validate, and deploy methods for more efficient and reliable LLM evaluation.

Important information that relates to a specific topic in a document is often organized in tabular format to assist readers with information retrieval and comparison, which may be difficult to provide in natural language. However, tabular data in unstructured digital documents, e.g., Portable Document Format (PDF) and images, are difficult to parse into structured machine-readable format, due to complexity and diversity in their structure and style. To facilitate image-based table recognition with deep learning, we develop the largest publicly available table recognition dataset PubTabNet (https://github.com/ibm-aur-nlp/PubTabNet), containing 568k table images with corresponding structured HTML representation. PubTabNet is automatically generated by matching the XML and PDF representations of the scientific articles in PubMed Central Open Access Subset (PMCOA). We also propose a novel attention-based encoder-dual-decoder (EDD) architecture that converts images of tables into HTML code. The model has a structure decoder which reconstructs the table structure and helps the cell decoder to recognize cell content. In addition, we propose a new Tree-Edit-Distance-based Similarity (TEDS) metric for table recognition, which more appropriately captures multi-hop cell misalignment and OCR errors than the pre-established metric. The experiments demonstrate that the EDD model can accurately recognize complex tables solely relying on the image representation, outperforming the state-of-the-art by 9.7% absolute TEDS score.

Due to the cumbersome nature of human evaluation and limitations of code-based evaluation, Large Language Models (LLMs) are increasingly being used to assist humans in evaluating LLM outputs. Yet LLM-generated evaluators simply inherit all the problems of the LLMs they evaluate, requiring further human validation. We present a mixed-initiative approach to ``validate the validators'' -- aligning LLM-generated evaluation functions (be it prompts or code) with human requirements. Our interface, EvalGen, provides automated assistance to users in generating evaluation criteria and implementing assertions. While generating candidate implementations (Python functions, LLM grader prompts), EvalGen asks humans to grade a subset of LLM outputs; this feedback is used to select implementations that better align with user grades. A qualitative study finds overall support for EvalGen but underscores the subjectivity and iterative process of alignment. In particular, we identify a phenomenon we dub criteria drift: users need criteria to grade outputs, but grading outputs helps users define criteria. What is more, some criteria appears dependent on the specific LLM outputs observed (rather than independent criteria that can be defined a priori), raising serious questions for approaches that assume the independence of evaluation from observation of model outputs. We present our interface and implementation details, a comparison of our algorithm with a baseline approach, and implications for the design of future LLM evaluation assistants.

Empirical evaluation serves as the primary compass guiding research progress in foundation models. Despite a large body of work focused on training frontier vision-language models (VLMs), approaches to their evaluation remain nascent. To guide their maturation, we propose three desiderata that evaluations should satisfy: (1) faithfulness to the modality and application, (2) discriminability between models of varying quality, and (3) efficiency in compute. Through this lens, we identify critical failure modes that violate faithfulness and discriminability, misrepresenting model capabilities: (i) multiple-choice formats reward guessing, poorly reflect downstream use cases, and saturate early as models improve; (ii) blindly solvable questions, which can be answered without images, constitute up to 70% of some evaluations; and (iii) mislabeled or ambiguous samples compromise up to 42% of examples in certain datasets. Regarding efficiency, the computational burden of evaluating frontier models has become prohibitive: by some accounts, nearly 20% of development compute is devoted to evaluation alone. Rather than discarding existing benchmarks, we curate them via transformation and filtering to maximize fidelity and discriminability. We find that converting multiple-choice questions to generative tasks reveals sharp capability drops of up to 35%. In addition, filtering blindly solvable and mislabeled samples improves discriminative power while simultaneously reducing computational cost. We release DatBench-Full, a cleaned evaluation suite of 33 datasets spanning nine VLM capabilities, and DatBench, a discriminative subset that achieves 13x average speedup (up to 50x) while closely matching the discriminative power of the original datasets. Our work outlines a path toward evaluation practices that are both rigorous and sustainable as VLMs continue to scale.

Public leaderboards for coding agents typically rank systems by model name and pass rate, while the surrounding harness (the scaffold that issues tools, manages context, and decides when to stop) is often under-specified. Model-to-model comparison is valid when the harness is fixed; when it varies, performance and efficiency conflate model and scaffold effects. We evaluate Qwen 3.6 Plus and MiniMax M2.5 across three open-source harnesses (Goose, OpenCode, OpenHands-SDK) on a stratified 50-task subset of Terminal-Bench Pro. Harness choice induces up to a 40x difference in tokens per solved task, while paired within-model pass-rate differences remain 0-8 percentage points (95% paired-task bootstrap CIs include zero except for the largest gap). Failure fingerprints replicate across models (REASON for Goose, VERIFY/MAX_TURNS for OpenHands-SDK, idle-loop/TIME for OpenCode), indicating harness-level biases that are largely model-independent. For human-centered coding-agent evaluation, model name alone is an incomplete comparison unit: harness-model pairs determine real-world cost, latency, and oversight burden; no-action turns are a per-task wait tax, not just a token tax. We therefore recommend selecting harness-model pairs by pass rate under token/latency budgets, and reporting token usage, latency, and full harness specifications alongside any model comparison. We release anonymized configs, raw trial logs, aggregated snapshots, and analysis scripts.

Scalable driving simulators typically execute vehicle commands using prescribed behavioral or kinematic rules, overlooking the physics of tire-road interfaces, thereby limiting their ability to capture how adverse road and environmental conditions alter vehicle execution and propagate through traffic. To address this limitation, we present SceneFactory-3D, a GPU-batched, physics-grounded multi-agent driving simulator. Vehicles execute acceleration and steering commands via suspension- and friction-limited forces evaluated at each wheel-contact point. Spatially varying friction, per-world 3D heightfields, gravity, and rigid contact consistently govern wheel motion and chassis collisions. Per-world terrain isolation and GPU batching enable SceneFactory-3D to run matched physical counterfactuals in parallel: traffic scenario setup and vehicle controllers remain fixed while only the road condition changes, enabling the resulting closed-loop effects to be evaluated across parallel worlds. To demonstrate the advantage of the SceneFactory-3D-enabled counterfactual evaluation, we conduct an empirical study on vehicle controllers' sensitivity to road conditions. We study three learned-policy families on 1,024 matched 12-vehicle worlds per condition, and two classical planners on a shared 32-world subset, across 21 friction and grade conditions. When friction drops from 1.0 to 0.18, the share of vehicles that clear the work zone safely falls by 6 to 90 percentage points across learned policies (18-19 for classical planners), and near-collision situations become more frequent for every learned policy. Code: https://github.com/SmallWorldLab/SceneFactory_3D

Multimodal/vision language models (VLMs) are increasingly being deployed in healthcare settings worldwide, necessitating robust benchmarks to ensure their safety, efficacy, and fairness. Multiple-choice question and answer (QA) datasets derived from national medical examinations have long served as valuable evaluation tools, but existing datasets are largely text-only and available in a limited subset of languages and countries. To address these challenges, we present WorldMedQA-V, an updated multilingual, multimodal benchmarking dataset designed to evaluate VLMs in healthcare. WorldMedQA-V includes 568 labeled multiple-choice QAs paired with 568 medical images from four countries (Brazil, Israel, Japan, and Spain), covering original languages and validated English translations by native clinicians, respectively. Baseline performance for common open- and closed-source models are provided in the local language and English translations, and with and without images provided to the model. The WorldMedQA-V benchmark aims to better match AI systems to the diverse healthcare environments in which they are deployed, fostering more equitable, effective, and representative applications.

Pashto is spoken by approximately 60--80 million people but has no published benchmarks for multilingual automatic speech recognition (ASR) on any shared public test set. This paper reports the first reproducible multi-model evaluation on public Pashto data, covering zero-shot ASR, script-level failure, and cross-domain evaluation of fine-tuned models. For zero-shot ASR, ten models (all seven Whisper sizes, MMS-1B, SeamlessM4T-v2-large, and OmniASR-CTC-300M) are evaluated on the FLEURS Pashto test set and a filtered Common Voice~24 subset; zero-shot Whisper WER ranges from 90% to 297%, with the medium model collapsing to 461% on Common Voice~24 consistent with decoder looping. SeamlessM4T achieves 39.7% WER on Common Voice~24 (the best zero-shot result reported to date, as of submission); MMS-1B achieves 43.8% on FLEURS. For script failure, a language-identification audit shows that no Whisper model produces Pashto-script output in more than 0.8% of utterances, while MMS-1B, SeamlessM4T, and OmniASR each exceed 93% Pashto-script fidelity; WER alone does not reveal this failure, since a model generating Arabic-script output on Pashto audio has not achieved ASR in any interpretable sense. For cross-domain evaluation, five fine-tuned Pashto ASR models are evaluated on both test sets: published WER figures of 14% degrade to 32.5--59% on out-of-distribution sets, while one augmented model achieves 35.1% on both sets with zero cross-domain degradation. Character-class error stratification confirms that Pashto-unique phonemes (the retroflex series and lateral fricatives) account for disproportionate error mass. All evaluations cover read speech only. Five structural impediments to cumulative progress are identified and five ordered research priorities are argued.

During Automated Program Repair (APR), it can be challenging to synthesize correct patches for real-world systems in general-purpose programming languages. Recent Large Language Models (LLMs) have been shown to be helpful "copilots" in assisting developers with various coding tasks, and have also been directly applied for patch synthesis. However, most LLMs treat programs as sequences of tokens, meaning that they are ignorant of the underlying semantics constraints of the target programming language. This results in plenty of statically invalid generated patches, impeding the practicality of the technique. Therefore, we propose Repilot, a framework to further copilot the AI "copilots" (i.e., LLMs) by synthesizing more valid patches during the repair process. Our key insight is that many LLMs produce outputs autoregressively (i.e., token by token), resembling human writing programs, which can be significantly boosted and guided through a Completion Engine. Repilot synergistically synthesizes a candidate patch through the interaction between an LLM and a Completion Engine, which 1) prunes away infeasible tokens suggested by the LLM and 2) proactively completes the token based on the suggestions provided by the Completion Engine. Our evaluation on a subset of the widely-used Defects4j 1.2 and 2.0 datasets shows that Repilot fixes 66 and 50 bugs, respectively, surpassing the best-performing baseline by 14 and 16 bugs fixed. More importantly, Repilot is capable of producing more valid and correct patches than the base LLM when given the same generation budget.

Cluster workload allocation often requires complex configurations, creating a usability gap. This paper introduces a semantic, intent-driven scheduling paradigm for cluster systems using Natural Language Processing. The system employs a Large Language Model (LLM) integrated via a Kubernetes scheduler extender to interpret natural language allocation hint annotations for soft affinity preferences. A prototype featuring a cluster state cache and an intent analyzer (using AWS Bedrock) was developed. Empirical evaluation demonstrated high LLM parsing accuracy (>95% Subset Accuracy on an evaluation ground-truth dataset) for top-tier models like Amazon Nova Pro/Premier and Mistral Pixtral Large, significantly outperforming a baseline engine. Scheduling quality tests across six scenarios showed the prototype achieved superior or equivalent placement compared to standard Kubernetes configurations, particularly excelling in complex and quantitative scenarios and handling conflicting soft preferences. The results validate using LLMs for accessible scheduling but highlight limitations like synchronous LLM latency, suggesting asynchronous processing for production readiness. This work confirms the viability of semantic soft affinity for simplifying workload orchestration.

This paper presents an overview of the NTIRE 2026 Challenge on Video Saliency Prediction. The goal of the challenge participants was to develop automatic saliency map prediction methods for the provided video sequences. The novel dataset of 2,000 diverse videos with an open license was prepared for this challenge. The fixations and corresponding saliency maps were collected using crowdsourced mouse tracking and contain viewing data from over 5,000 assessors. Evaluation was performed on a subset of 800 test videos using generally accepted quality metrics. The challenge attracted over 20 teams making submissions, and 7 teams passed the final phase with code review. All data used in this challenge is made publicly available - https://github.com/msu-video-group/NTIRE26_Saliency_Prediction.

Recent studies suggest that video generation models can exhibit certain forms of zero-shot visual reasoning through generated frames. Yet reliable evaluation remains challenging: benchmarks should adopt inputs aligned with the visual priors of current video models, require valid evolving processes rather than only plausible final states, and calibrate task difficulty to remain challenging yet partly feasible. To this end, we introduce VGI-bench, containing 27 tasks and 810 instances, organized by a two-level taxonomy of task domains and skill tags for fine-grained evaluation of visual reasoning capabilities of video generation models. Our evaluations show that current generative systems can solve a subset of visually grounded reasoning tasks, but remain far from reliable, with even the strongest model, Seedance 2.0, achieving only 51.0% under our evaluation criteria. Our analysis further explore the output failure modes, input condition sensitivity, performance transfer boundary from synthetic fine-tuning, and internal denoising perspective revealing limited self-correction, where later steps mainly refine early hypotheses rather than correct reasoning errors. We hope VGI-bench will help stimulate the development of next-generation video generation models. Website: https://hexuan21.github.io/VGI-Bench/

Agentic AI is advancing rapidly, yet truly autonomous medical-imaging triage, where a system decides when to stop, escalate, or defer under real constraints, remains relatively underexplored. To address this gap, we introduce AT-CXR, an uncertainty-aware agent for chest X-rays. The system estimates per-case confidence and distributional fit, then follows a stepwise policy to issue an automated decision or abstain with a suggested label for human intervention. We evaluate two router designs that share the same inputs and actions: a deterministic rule-based router and an LLM-decided router. Across five-fold evaluation on a balanced subset of NIH ChestX-ray14 dataset, both variants outperform strong zero-shot vision-language models and state-of-the-art supervised classifiers, achieving higher full-coverage accuracy and superior selective-prediction performance, evidenced by a lower area under the risk-coverage curve (AURC) and a lower error rate at high coverage, while operating with lower latency that meets practical clinical constraints. The two routers provide complementary operating points, enabling deployments to prioritize maximal throughput or maximal accuracy. Our code is available at https://github.com/XLIAaron/uncertainty-aware-cxr-agent.

Retrieval-augmented generation (RAG) has emerged as a promising solution to mitigate the limitations of large language models (LLMs), such as hallucinations and outdated information. However, it remains unclear how LLMs handle knowledge conflicts arising from different augmented retrieved passages, especially when these passages originate from the same source and have equal trustworthiness. In this work, we conduct a comprehensive evaluation of LLM-generated answers to questions that have varying answers based on contradictory passages from Wikipedia, a dataset widely regarded as a high-quality pre-training resource for most LLMs. Specifically, we introduce WikiContradict, a benchmark consisting of 253 high-quality, human-annotated instances designed to assess LLM performance when augmented with retrieved passages containing real-world knowledge conflicts. We benchmark a diverse range of both closed and open-source LLMs under different QA scenarios, including RAG with a single passage, and RAG with 2 contradictory passages. Through rigorous human evaluations on a subset of WikiContradict instances involving 5 LLMs and over 3,500 judgements, we shed light on the behaviour and limitations of these models. For instance, when provided with two passages containing contradictory facts, all models struggle to generate answers that accurately reflect the conflicting nature of the context, especially for implicit conflicts requiring reasoning. Since human evaluation is costly, we also introduce an automated model that estimates LLM performance using a strong open-source language model, achieving an F-score of 0.8. Using this automated metric, we evaluate more than 1,500 answers from seven LLMs across all WikiContradict instances. To facilitate future work, we release WikiContradict on: https://ibm.biz/wikicontradict.

Toxic language detection systems often falsely flag text that contains minority group mentions as toxic, as those groups are often the targets of online hate. Such over-reliance on spurious correlations also causes systems to struggle with detecting implicitly toxic language. To help mitigate these issues, we create ToxiGen, a new large-scale and machine-generated dataset of 274k toxic and benign statements about 13 minority groups. We develop a demonstration-based prompting framework and an adversarial classifier-in-the-loop decoding method to generate subtly toxic and benign text with a massive pretrained language model. Controlling machine generation in this way allows ToxiGen to cover implicitly toxic text at a larger scale, and about more demographic groups, than previous resources of human-written text. We conduct a human evaluation on a challenging subset of ToxiGen and find that annotators struggle to distinguish machine-generated text from human-written language. We also find that 94.5% of toxic examples are labeled as hate speech by human annotators. Using three publicly-available datasets, we show that finetuning a toxicity classifier on our data improves its performance on human-written data substantially. We also demonstrate that ToxiGen can be used to fight machine-generated toxicity as finetuning improves the classifier significantly on our evaluation subset. Our code and data can be found at https://github.com/microsoft/ToxiGen.

Step-by-step verifiers -- also known as process reward models (PRMs) -- are a key ingredient for test-time scaling. PRMs require step-level supervision, making them expensive to train. This work aims to build data-efficient PRMs as verbalized step-wise reward models that verify every step in the solution by generating a verification chain-of-thought (CoT). We propose ThinkPRM, a long CoT verifier fine-tuned on orders of magnitude fewer process labels than those required by discriminative PRMs. Our approach capitalizes on the inherent reasoning abilities of long CoT models, and outperforms LLM-as-a-Judge and discriminative verifiers -- using only 1% of the process labels in PRM800K -- across several challenging benchmarks. Specifically, ThinkPRM beats the baselines on ProcessBench, MATH-500, and AIME '24 under best-of-N selection and reward-guided search. In an out-of-domain evaluation on a subset of GPQA-Diamond and LiveCodeBench, our PRM surpasses discriminative verifiers trained on the full PRM800K by 8% and 4.5%, respectively. Lastly, under the same token budget, ThinkPRM scales up verification compute more effectively compared to LLM-as-a-Judge, outperforming it by 7.2% on a subset of ProcessBench. Our work highlights the value of generative, long CoT PRMs that can scale test-time compute for verification while requiring minimal supervision for training. Our code, data, and models will be released at https://github.com/mukhal/thinkprm.

Unified multimodal models integrate the reasoning capacity of large language models with both image understanding and generation, showing great promise for advanced multimodal intelligence. However, the community still lacks a rigorous reasoning-centric benchmark to systematically evaluate the alignment between understanding and generation, and their generalization potential in complex visual tasks. To this end, we introduce GIR-Bench, a comprehensive benchmark that evaluates unified models across three complementary perspectives. Firstly, we investigate understanding-generation consistency (GIR-Bench-UGC), asking whether models can consistently leverage the same knowledge in both understanding and generation tasks. Secondly, we investigate whether models can perform reasoning-centric text-to-image generation that requires applying logical constraints and implicit knowledge to generate faithful visual content (GIR-Bench-T2I). Thirdly, we evaluate whether models can handle multi-step reasoning in editing (GIR-Bench-Edit). For each subset, we carefully design different task-specific evaluation pipelines tailored for each task. This enables fine-grained and interpretable evaluation while mitigating biases from the prevalent MLLM-as-a-Judge paradigm. Extensive ablations over various unified models and generation-only systems have shown that: Although unified models are more capable of reasoning-driven visual tasks, they still exhibit a persistent gap between understanding and generation. The data and code for GIR-Bench are available at https://hkust-longgroup.github.io/GIR-Bench{https://hkust-longgroup.github.io/GIR-Bench}.

In this work, we introduce CodeEvolve, an open-source evolutionary coding agent that unites Large Language Models (LLMs) with genetic algorithms to solve complex computational problems. Our framework adapts powerful evolutionary concepts to the LLM domain, building upon recent methods for generalized scientific discovery. CodeEvolve employs an island-based genetic algorithm to maintain population diversity and increase throughput, introduces a novel inspiration-based crossover mechanism that leverages the LLMs context window to combine features from successful solutions, and implements meta-prompting strategies for dynamic exploration of the solution space. We conduct a rigorous evaluation of CodeEvolve on a subset of the mathematical benchmarks used to evaluate Google DeepMind's closed-source AlphaEvolve. Our findings show that our method surpasses AlphaEvolve's performance on several challenging problems. To foster collaboration and accelerate progress, we release our complete framework as an open-source repository.

Modern retrieval pipelines increasingly serve downstream consumers like retrieval-augmented generation (RAG) and autonomous agents that need more than a scalar relevance score. A reranker that only tells the caller "how relevant" forces the agent to dump entire documents into the language-model context, wasting tokens on tangential passages and boilerplate. We introduce Prism-Reranker, a family of reranker models built on Qwen3.5 at four sizes (0.8B, 2B, 4B, 9B) that goes beyond scalar scoring. In addition to the standard yes/no relevance judgement, whenever the verdict is yes the model emits (i) a contribution statement summarizing how the document helps the query, and (ii) an evidence passage: a self-contained rewrite that preserves every query-relevant signal while discarding noise. Prism-Reranker is trained with a hybrid objective combining point-wise distillation from a strong commercial reranker API with supervised fine-tuning on contribution and evidence targets. We curate training data from KaLM-Embedding's open-source aggregation, augmented with real web documents retrieved via commercial search APIs for open-domain queries and LLM-synthesized variants, and rewrite a portion of queries into keyword-style reformulations to adapt the model to agent-issued traffic. To reconcile inconsistent labels across open corpora and obtain crisp binary supervision, we relabel data with an LLM-as-Judge ensemble aggregating votes from five frontier LLMs. On a QA subset of BEIR and on an LLM-judged evaluation of contribution and evidence quality, Prism-Reranker attains solid results across all four sizes. We further show that the same recipe extends existing LLM-based rerankers, augmenting Qwen3-Reranker-4B with contribution and evidence capabilities while improving its average BEIR-QA NDCG@10 by +1.54 over the base model. Model weights, training recipe, and evaluation suite are released.

Machine Learning as a Service (MLaaS) is often provided as a pay-per-query, black-box system to clients. Such a black-box approach not only hinders open replication, validation, and interpretation of model results, but also makes it harder for white-hat researchers to identify vulnerabilities in the MLaaS systems. Model extraction is a promising technique to address these challenges by reverse-engineering black-box models. Since training data is typically unavailable for MLaaS models, this paper focuses on the realistic version of it: data-free model extraction. We propose a data-free model extraction approach, CaBaGe, to achieve higher model extraction accuracy with a small number of queries. Our innovations include (1) a novel experience replay for focusing on difficult training samples; (2) an ensemble of generators for steadily producing diverse synthetic data; and (3) a selective filtering process for querying the victim model with harder, more balanced samples. In addition, we create a more realistic setting, for the first time, where the attacker has no knowledge of the number of classes in the victim training data, and create a solution to learn the number of classes on the fly. Our evaluation shows that CaBaGe outperforms existing techniques on seven datasets -- MNIST, FMNIST, SVHN, CIFAR-10, CIFAR-100, ImageNet-subset, and Tiny ImageNet -- with an accuracy improvement of the extracted models by up to 43.13%. Furthermore, the number of queries required to extract a clone model matching the final accuracy of prior work is reduced by up to 75.7%.

Rapid advancements in large language models (LLMs) have increased interest in deploying them on mobile devices for on-device AI applications. Mobile users interact differently with LLMs compared to desktop users, creating unique expectations and data biases. Current benchmark datasets primarily target at server and desktop environments, and there is a notable lack of extensive datasets specifically designed for mobile contexts. Additionally, mobile devices face strict limitations in storage and computing resources, constraining model size and capabilities, thus requiring optimized efficiency and prioritized knowledge. To address these challenges, we introduce Mobile-MMLU, a large-scale benchmark dataset tailored for mobile intelligence. It consists of 16,186 questions across 80 mobile-related fields, designed to evaluate LLM performance in realistic mobile scenarios. A challenging subset, Mobile-MMLU-Pro, provides advanced evaluation similar in size to MMLU-Pro but significantly more difficult than our standard full set. Both benchmarks use multiple-choice, order-invariant questions focused on practical mobile interactions, such as recipe suggestions, travel planning, and essential daily tasks. The dataset emphasizes critical mobile-specific metrics like inference latency, energy consumption, memory usage, and response quality, offering comprehensive insights into model performance under mobile constraints. Moreover, it prioritizes privacy and adaptability, assessing models' ability to perform on-device processing, maintain user privacy, and adapt to personalized usage patterns. Mobile-MMLU family offers a standardized framework for developing and comparing mobile-optimized LLMs, enabling advancements in productivity and decision-making within mobile computing environments. Our code and data are available at: https://github.com/VILA-Lab/Mobile-MMLU.

Mainstream large vision-language models (LVLMs) inherently encode cultural biases, highlighting the need for diverse multimodal datasets. To address this gap, we introduce Pearl, a large-scale Arabic multimodal dataset and benchmark explicitly designed for cultural understanding. Constructed through advanced agentic workflows and extensive human-in-the-loop annotations by 45 annotators from across the Arab world, Pearl comprises over K multimodal examples spanning ten culturally significant domains covering all Arab countries. We further provide two robust evaluation benchmarks Pearl and Pearl-Lite along with a specialized subset Pearl-X explicitly developed to assess nuanced cultural variations. Comprehensive evaluations on state-of-the-art open and proprietary LVLMs demonstrate that reasoning-centric instruction alignment substantially improves models' cultural grounding compared to conventional scaling methods. Pearl establishes a foundational resource for advancing culturally-informed multimodal modeling research. All datasets and benchmarks are publicly available.

This paper presents Team Marikarp's solution for the SIGIR 2025 LiveRAG competition. The competition's evaluation set, automatically generated by DataMorgana from internet corpora, encompassed a wide range of target topics, question types, question formulations, audience types, and knowledge organization methods. It offered a fair evaluation of retrieving question-relevant supporting documents from a 15M documents subset of the FineWeb corpus. Our proposed knowledge-aware diverse reranking RAG pipeline achieved first place in the competition.

Retrieval-augmented generation (RAG) has emerged as a leading approach to reducing hallucinations in large language models (LLMs). Current RAG evaluation benchmarks primarily focus on what we call local RAG: retrieving relevant chunks from a small subset of documents to answer queries that require only localized understanding within specific text chunks. However, many real-world applications require a fundamentally different capability -- global RAG -- which involves aggregating and analyzing information across entire document collections to derive corpus-level insights (for example, "What are the top 10 most cited papers in 2023?"). In this paper, we introduce GlobalQA -- the first benchmark specifically designed to evaluate global RAG capabilities, covering four core task types: counting, extremum queries, sorting, and top-k extraction. Through systematic evaluation across different models and baselines, we find that existing RAG methods perform poorly on global tasks, with the strongest baseline achieving only 1.51 F1 score. To address these challenges, we propose GlobalRAG, a multi-tool collaborative framework that preserves structural coherence through chunk-level retrieval, incorporates LLM-driven intelligent filters to eliminate noisy documents, and integrates aggregation modules for precise symbolic computation. On the Qwen2.5-14B model, GlobalRAG achieves 6.63 F1 compared to the strongest baseline's 1.51 F1, validating the effectiveness of our method.

Large Language Models (LLMs) with agentic web search capabilities show strong potential for tasks requiring real-time information access and complex fact retrieval, yet evaluating such systems remains challenging. We introduce \bench, a rigorous and regularly updated benchmark designed to assess the agentic web search abilities of LLMs. \bench automatically generates fresh question-answer pairs from recent news articles, ensuring that questions require information beyond an LLM's training data and enabling clear separation between internal knowledge and search capability. The benchmark features intentionally difficult questions requiring multi-hop search queries, page visits, and reasoning, making it well-suited for evaluating agentic search behavior. Our automated data curation and question generation pipeline enables frequent benchmark updates and supports construction of a large-scale training dataset for agentic web search models, addressing the scarcity of such data in the research community. To ensure reliable evaluation, we include a subset of human-verified samples in the test set. We evaluate a broad range of systems using \bench, including commercial and open-weight LLMs as well as LLM-based web search APIs. The leaderboard, datasets, and code are publicly available at livenewsbench.com.

Current document parsing methods compete primarily on model architecture innovation, while systematic engineering of training data remains underexplored. Yet SOTA models of different architectures and parameter scales exhibit highly consistent failure patterns on the same set of hard samples, suggesting that the performance bottleneck stems from shared deficiencies in training data rather than architecture itself. Building on this finding, we present \minerupro, which advances the state of the art solely through data engineering and training strategy optimization while keeping the 1.2B-parameter architecture of \mineru completely fixed. At its core is a Data Engine co-designed around coverage, informativeness, and annotation accuracy: Diversity-and-Difficulty-Aware Sampling expands training data from under 10M to 65.5M samples while correcting distribution shift; Cross-Model Consistency Verification leverages output agreement among heterogeneous models to assess sample difficulty and generate reliable annotations; the Judge-and-Refine pipeline improves annotation quality for hard samples through render-then-verify iterative correction. A three-stage progressive training strategy -- large-scale pre-training, hard sample fine-tuning, and GRPO alignment -- sequentially exploits these data at different quality tiers. On the evaluation front, we fix element-matching biases in OmniDocBench~v1.5 and introduce a Hard subset, establishing the more discriminative OmniDocBench~v1.6 protocol. Without any architectural modification, \minerupro achieves 95.69 on OmniDocBench~v1.6, improving over the same-architecture baseline by 2.71 points and surpassing all existing methods including models with over 200times more parameters.

Rapid Large Language Model (LLM) advancements are fueling autonomous Multi-Agent System (MAS) development. However, current frameworks often lack flexibility, resource awareness, model diversity, and autonomous tool creation. This paper introduces HASHIRU (Hierarchical Agent System for Hybrid Intelligent Resource Utilization), a novel MAS framework enhancing flexibility, resource efficiency, and adaptability. HASHIRU features a "CEO" agent dynamically managing specialized "employee" agents, instantiated based on task needs and resource constraints (cost, memory). Its hybrid intelligence prioritizes smaller, local LLMs (via Ollama) while flexibly using external APIs and larger models when necessary. An economic model with hiring/firing costs promotes team stability and efficient resource allocation. The system also includes autonomous API tool creation and a memory function. Evaluations on tasks like academic paper review (58% success), safety assessments (100% on a JailbreakBench subset), and complex reasoning (outperforming Gemini 2.0 Flash on GSM8K: 96% vs. 61%; JEEBench: 80% vs. 68.3%; SVAMP: 92% vs. 84%) demonstrate HASHIRU's capabilities. Case studies illustrate its self-improvement via autonomous cost model generation, tool integration, and budget management. HASHIRU offers a promising approach for more robust, efficient, and adaptable MAS through dynamic hierarchical control, resource-aware hybrid intelligence, and autonomous functional extension. Source code and benchmarks are available at https://github.com/HASHIRU-AI/HASHIRU and https://github.com/HASHIRU-AI/HASHIRUBench respectively, and a live demo is available at https://hashiruagentx-hashiruai.hf.space upon request.

The rapid advancement of large language models has fundamentally shifted the bottleneck in AI development from computational power to data availability-with countless valuable datasets remaining hidden across specialized repositories, research appendices, and domain platforms. As reasoning capabilities and deep research methodologies continue to evolve, a critical question emerges: can AI agents transcend conventional search to systematically discover any dataset that meets specific user requirements, enabling truly autonomous demand-driven data curation? We introduce DatasetResearch, the first comprehensive benchmark evaluating AI agents' ability to discover and synthesize datasets from 208 real-world demands across knowledge-intensive and reasoning-intensive tasks. Our tri-dimensional evaluation framework reveals a stark reality: even advanced deep research systems achieve only 22% score on our challenging DatasetResearch-pro subset, exposing the vast gap between current capabilities and perfect dataset discovery. Our analysis uncovers a fundamental dichotomy-search agents excel at knowledge tasks through retrieval breadth, while synthesis agents dominate reasoning challenges via structured generation-yet both catastrophically fail on "corner cases" outside existing distributions. These findings establish the first rigorous baseline for dataset discovery agents and illuminate the path toward AI systems capable of finding any dataset in the digital universe. Our benchmark and comprehensive analysis provide the foundation for the next generation of self-improving AI systems and are publicly available at https://github.com/GAIR-NLP/DatasetResearch.

To reveal the importance of temporal precision in ground truth audio event labels, we collected precise (~0.1 sec resolution) "strong" labels for a portion of the AudioSet dataset. We devised a temporally strong evaluation set (including explicit negatives of varying difficulty) and a small strong-labeled training subset of 67k clips (compared to the original dataset's 1.8M clips labeled at 10 sec resolution). We show that fine-tuning with a mix of weak and strongly labeled data can substantially improve classifier performance, even when evaluated using only the original weak labels. For a ResNet50 architecture, d' on the strong evaluation data including explicit negatives improves from 1.13 to 1.41. The new labels are available as an update to AudioSet.

Financial document parsing requires accuracy, structural consistency, and verifiability that current benchmarks often fail to reflect. We present FinixDoc, an end-to-end agentic parsing system for real-world financial documents, with FinixDoc-VL, a 4B-scale vision-language model built on Qwen3-VL-4B, as its core parser. To characterize the gap between benchmark and deployment performance, we introduce a Document Parsing Capability Matrix organized along two practical axes: visual quality and document scale. Guided by this matrix, FinixDoc-VL is trained with a domain-adapted recipe combining homoglyph-aware contrastive learning and multi-stage reinforcement learning with composite domain-specific rewards. To better leverage our accumulated advantage in low-quality financial-document data and support large-scale, high-quality data production, we further build a human-in-the-loop Data Factory pipeline with confidence-aware expert review. For evaluation, we construct FinixDocBench, a financial-domain evaluation suite covering digital-native, camera-captured, ultra-large-page, and internal-workflow scenarios, with a compliance-reviewed subset released alongside this technical report. On its main subsets, FinixDoc-VL achieves the highest overall score (81.43) among evaluated baselines, outperforming the next-best open-source model by 5.13 points, with the largest gains on internal financial workflows (FinixInner: 84.08 vs. 78.73).

Automated shape repair approaches currently lack access to datasets that describe real-world damaged geometry. We present Fantastic Breaks (and Where to Find Them: https://terascale-all-sensing-research-studio.github.io/FantasticBreaks), a dataset containing scanned, waterproofed, and cleaned 3D meshes for 150 broken objects, paired and geometrically aligned with complete counterparts. Fantastic Breaks contains class and material labels, proxy repair parts that join to broken meshes to generate complete meshes, and manually annotated fracture boundaries. Through a detailed analysis of fracture geometry, we reveal differences between Fantastic Breaks and synthetic fracture datasets generated using geometric and physics-based methods. We show experimental shape repair evaluation with Fantastic Breaks using multiple learning-based approaches pre-trained with synthetic datasets and re-trained with subset of Fantastic Breaks.

Recent multimodal large language models (MLLMs) show strong capabilities in visual-language reasoning, yet their performance on ultra-high-resolution imagery remains largely unexplored. Existing visual question answering (VQA) benchmarks typically rely on medium-resolution data, offering limited visual complexity. To bridge this gap, we introduce Ultra-high-resolution Reasoning Benchmark (UR-Bench), a benchmark designed to evaluate the reasoning capabilities of MLLMs under extreme visual information. UR-Bench comprises two major categories, Humanistic Scenes and Natural Scenes, covering four subsets of ultra-high-resolution images with distinct spatial structures and data sources. Each subset contains images ranging from hundreds of megapixels to gigapixels, accompanied by questions organized into three levels, enabling evaluation of models' reasoning capabilities in ultra-high-resolution scenarios. We further propose an agent-based framework in which a language model performs reasoning by invoking external visual tools. In addition, we introduce Semantic Abstraction and Retrieval tools that enable more efficient processing of ultra-high-resolution images. We evaluate state-of-the-art models using both an end-to-end MLLMs and our agent-based framework, demonstrating the effectiveness of our framework.

When a coordinator compares answers from heterogeneous foundation models, self-reported confidence may have different meanings across responders and changing workloads. This paper presents MARGIN (Multi-Agent Runtime Grading via Incremental Normalisation), a runtime calibration method that learns model-specific confidence corrections from observed answer outcomes without retraining the models or requiring a held-out calibration set. MARGIN tracks recent accuracy and stated confidence within confidence bands, uses their ratio to correct reported confidence, and blends sparse-band corrections toward a model-level estimate. The corrected scores weight candidate answers in a collective decision. Evaluation covers code generation, question answering, and mathematics, using an 18-model pool and a nine-model subset for distribution-shift experiments. On BigCodeBench, model-mean confidence is negatively related to accuracy; among correct/incorrect response pairs, choosing the more confident responder performs below chance. Against five online calibration baselines receiving identical feedback and retaining their learned state across each transition, MARGIN achieves lower post-shift expected calibration error than all five in two code-generation transitions and than four in a question-answering transition; the remaining question-answering comparison is inconclusive. In separate code-generation coordination experiments, calibration improves the ranking of correct responses and increases answer-selection accuracy by 4.3 and 14.0 percentage points on two of three benchmarks relative to uncalibrated confidence weighting. These results support model-specific runtime calibration for coordination under changing workloads when correctness feedback is available for the participating responders.

Decoder hidden states from different layers of large language models (LLMs) often encode complementary information. However, existing speculative decoding methods typically rely on a small set of target-model layers chosen using fixed, position-based heuristics. Selecting an optimal subset of layers is inherently a combinatorial problem, making exhaustive evaluation infeasible even for moderately deep models. In this work, we present a principled framework for multi-layer hidden-state selection for speculative decoding based on two complementary criteria: informativeness, measured through matrix-entropy analysis, and diversity, quantified using centered kernel alignment (CKA). We investigate layer-wise Shannon entropy, aggregate entropy under hidden-state concatenation, and pairwise CKA dynamics across depth, revealing distinct geometric and informational compartments within decoder representations. Motivated by these findings, we propose entropy and CKA-based layer selection algorithms that favor hidden-state sets that are both highly informative and minimally redundant. As part of this framework, we introduce token-entropy filtering, which removes hidden states associated with low predictive-entropy tokens, thereby improving the quality of layer-selection statistics. In EAGLE3-style speculative decoding, our selected sets outperform fixed baseline choices, including the default [2, N//2, N-3] rule, while providing a clearer explanation of why some defaults can remain competitive. Our results provide a practical recipe for multi-layer hidden-state selection for speculative decoding, insights into the geometry and information in LLM decoder states, and a foundation for extending similar analysis to related applications.

Natural language-driven no-code development allows users to specify software functionality using natural language (NL) instead of editing source code, promising increased productivity and democratized development. Large language models (LLMs) show potential in enabling this paradigm. In this context, software documentation acts as an NL specification for functionality. This work introduces NoCode-bench, a benchmark designed to evaluate LLMs on real-world NL-driven feature addition tasks, consisting of 634 tasks across 10 projects and 114k code changes. Each task pairs documentation updates with corresponding code implementations, validated by developer-written test cases. A subset of 114 high-quality, human-verified instances, NoCode-bench Verified, ensures reliable evaluation. Our experiments reveal that, despite high token usage, the best LLMs achieve a task success rate of only 28.07%, highlighting challenges in cross-file editing, codebase understanding, and tool calling. These findings indicate that LLMs are not yet ready for fully NL-driven no-code development. NoCode-bench lays the foundation for future advances in this area.

The integration of large language models (LLMs) into medical practice holds transformative potential, yet their real-world clinical utility remains limited by critical alignment challenges: (1) a disconnect between static evaluation benchmarks and dynamic clinical cognitive needs, (2) difficulties in adapting to evolving, multi-source medical standards, and (3) the inability of conventional reward models to capture nuanced, multi-dimensional medical quality criteria. To address these gaps, we propose MR-RML (Multidimensional Rubric-oriented Reward Model Learning) via GPRC (Geometric Projection Reference Constraints), a novel alignment framework that integrates medical standards into a structured "Dimensions-Scenarios-Disciplines" matrix to guide data generation and model optimization. MR-RML introduces three core innovations: (1) a "Dimensions-Scenarios-Disciplines" medical standard system that embeds domain standards into the full training pipeline; (2) an independent multi-dimensional reward model that decomposes evaluation criteria, shifting from real-time rubric-based scoring to internalized reward modeling for improved consistency and cost-efficiency; (3) geometric projection reference constraints that transform medical cognitive logic into mathematical regularization, aligning scoring gradients with clinical reasoning and enabling synthetic data-driven training. Through extensive evaluations on the authoritative medical benchmark Healthbench, our method yields substantial performance gains over the base LLM Qwen-32B (45% on the full subset and 85% on Hard subset, respectively). It achieves a SOTA among open-source LLMs with scores of 62.7 (full subset) and 44.7 (hard subset), while also outperforming the majority of closed-source models.

Abstract reasoning ability reflects the intelligence and generalization capacity of LLMs to extract and apply abstract rules. However, accurately measuring this ability remains challenging: existing benchmarks either rely on expensive manual annotation, limiting their scale, or risk measuring memorization rather than genuine reasoning. To address this, we introduce an automated pipeline named A2RBench, encompassing generation, expansion, evaluation, and analysis. Specifically, in the generation stage, LLMs create diverse tasks demanding genuine reasoning; in the expansion stage, LLMs reuse validated rules and expand new input spaces to generate task variations, achieving scaling. However, such a process may cause hallucinations. To eliminate it, we further establish a theoretical framework and prove that programmatic verification--testing whether the inverse operation perfectly reverses the forward operation (cycle consistency)--guarantees a unique solution. Through extensive evaluations on mainstream LLMs, we find: (1) Current LLMs exhibit fundamental deficiencies in abstract reasoning, with top models significantly underperforming humans on a representative subset (39.8% vs. 68.5%). (2) Current LLMs fall far short of 2D and 1D in the complexity of generated 3D tasks, revealing their lack of understanding of high-dimensional tasks. (3) Counterintuitively, inputs with higher information complexity can simplify the reasoning process.

Large multimodal models (LMMs) with advanced video analysis capabilities have recently garnered significant attention. However, most evaluations rely on traditional methods like multiple-choice questions in benchmarks such as VideoMME and LongVideoBench, which are prone to lack the depth needed to capture the complex demands of real-world users. To address this limitation-and due to the prohibitive cost and slow pace of human annotation for video tasks-we introduce VideoAutoArena, an arena-style benchmark inspired by LMSYS Chatbot Arena's framework, designed to automatically assess LMMs' video analysis abilities. VideoAutoArena utilizes user simulation to generate open-ended, adaptive questions that rigorously assess model performance in video understanding. The benchmark features an automated, scalable evaluation framework, incorporating a modified ELO Rating System for fair and continuous comparisons across multiple LMMs. To validate our automated judging system, we construct a 'gold standard' using a carefully curated subset of human annotations, demonstrating that our arena strongly aligns with human judgment while maintaining scalability. Additionally, we introduce a fault-driven evolution strategy, progressively increasing question complexity to push models toward handling more challenging video analysis scenarios. Experimental results demonstrate that VideoAutoArena effectively differentiates among state-of-the-art LMMs, providing insights into model strengths and areas for improvement. To further streamline our evaluation, we introduce VideoAutoBench as an auxiliary benchmark, where human annotators label winners in a subset of VideoAutoArena battles. We use GPT-4o as a judge to compare responses against these human-validated answers. Together, VideoAutoArena and VideoAutoBench offer a cost-effective, and scalable framework for evaluating LMMs in user-centric video analysis.

The U.S. Securities and Exchange Commission (SEC) requires that public companies file financial reports tagging numbers with the machine readable inline eXtensible Business Reporting Language (iXBRL) standard. However, the highly complex and highly granular taxonomy defined by iXBRL limits label transferability across domains. In this paper, we introduce the Hierarchical Financial Key Performance Indicator (HiFi-KPI) dataset, designed to facilitate numerical KPI extraction at specified levels of granularity from unstructured financial text. Our approach organizes a 218,126-label hierarchy using a taxonomy based grouping method, investigating which taxonomy layer provides the most meaningful structure. HiFi-KPI comprises ~1.8M paragraphs and ~5M entities, each linked to a label in the iXBRL-specific calculation and presentation taxonomies. We provide baselines using encoder-based approaches and structured extraction using Large Language Models (LLMs). To simplify LLM inference and evaluation, we additionally release HiFi-KPI Lite, a manually curated subset with four expert-mapped labels. We publicly release all artifacts

The rapid advancements in computing dramatically increase the scale and cost of training Large Language Models (LLMs). Accurately predicting downstream task performance prior to model training is crucial for efficient resource allocation, yet remains challenging due to two primary constraints: (1) the "emergence phenomenon", wherein downstream performance metrics become meaningful only after extensive training, which limits the ability to use smaller models for prediction; (2) Uneven task difficulty distributions and the absence of consistent scaling laws, resulting in substantial metric variability. Existing performance prediction methods suffer from limited accuracy and reliability, thereby impeding the assessment of potential LLM capabilities. To address these challenges, we propose a Clustering-On-Difficulty (COD) downstream performance prediction framework. COD first constructs a predictable support subset by clustering tasks based on difficulty features, strategically excluding non-emergent and non-scalable clusters. The scores on the selected subset serve as effective intermediate predictors of downstream performance on the full evaluation set. With theoretical support, we derive a mapping function that transforms performance metrics from the predictable subset to the full evaluation set, thereby ensuring accurate extrapolation of LLM downstream performance. The proposed method has been applied to predict performance scaling for a 70B LLM, providing actionable insights for training resource allocation and assisting in monitoring the training process. Notably, COD achieves remarkable predictive accuracy on the 70B LLM by leveraging an ensemble of small models, demonstrating an absolute mean deviation of 1.36% across eight important LLM evaluation benchmarks.

Automatically compiling open-source software (OSS) projects is a vital, labor-intensive, and complex task, which makes it a good challenge for LLM Agents. Existing methods rely on manually curated rules and workflows, which cannot adapt to OSS that requires customized configuration or environment setup. Recent attempts using Large Language Models (LLMs) used selective evaluation on a subset of highly rated OSS, a practice that underestimates the realistic challenges of OSS compilation. In practice, compilation instructions are often absent, dependencies are undocumented, and successful builds may even require patching source files or modifying build scripts. We propose a more challenging and realistic benchmark, BUILD-BENCH, comprising OSS that are more diverse in quality, scale, and characteristics. Furthermore, we propose a strong baseline LLM-based agent, OSS-BUILD-AGENT, an effective system with enhanced build instruction retrieval module that achieves state-of-the-art performance on BUILD-BENCH and is adaptable to heterogeneous OSS characteristics. We also provide detailed analysis regarding different compilation method design choices and their influence to the whole task, offering insights to guide future advances. We believe performance on BUILD-BENCH can faithfully reflect an agent's ability to tackle compilation as a complex software engineering tasks, and, as such, our benchmark will spur innovation with a significant impact on downstream applications in the fields of software development and software security.

For an LLM to correctly respond to an instruction it must understand both the semantics and the domain (i.e., subject area) of a given task-instruction pair. However, syntax can also convey implicit information Recent work shows that syntactic templates -- frequent sequences of Part-of-Speech (PoS) tags -- are prevalent in training data and often appear in model outputs. In this work we characterize syntactic templates, domain, and semantics in task-instruction pairs. We identify cases of spurious correlations between syntax and domain, where models learn to associate a domain with syntax during training; this can sometimes override prompt semantics. Using a synthetic training dataset, we find that the syntactic-domain correlation can lower performance (mean 0.51 +/- 0.06) on entity knowledge tasks in OLMo-2 models (1B-13B). We introduce an evaluation framework to detect this phenomenon in trained models, and show that it occurs on a subset of the FlanV2 dataset in open (OLMo-2-7B; Llama-4-Maverick), and closed (GPT-4o) models. Finally, we present a case study on the implications for safety finetuning, showing that unintended syntactic-domain correlations can be used to bypass refusals in OLMo-2-7B Instruct and GPT-4o. Our findings highlight two needs: (1) to explicitly test for syntactic-domain correlations, and (2) to ensure syntactic diversity in training data, specifically within domains, to prevent such spurious correlations.

Molecular function is largely determined by structure. Accurately aligning molecular structure with natural language is therefore essential for enabling large language models (LLMs) to reason about downstream chemical tasks. However, the substantial cost of human annotation makes it infeasible to construct large-scale, high-quality datasets of structure-grounded descriptions. In this work, we propose a fully automated annotation framework for generating precise molecular structure descriptions at scale. Our approach builds upon and extends a rule-based chemical nomenclature parser to interpret IUPAC names and construct enriched, structured XML metadata that explicitly encodes molecular structure. This metadata is then used to guide LLMs in producing accurate natural-language descriptions. Using this framework, we curate a large-scale dataset of approximately 163k molecule-description pairs. A rigorous validation protocol combining LLM-based and expert human evaluation on a subset of 2,000 molecules demonstrates a high description precision of 98.6%. The resulting dataset provides a reliable foundation for future molecule-language alignment, and the proposed annotation method is readily extensible to larger datasets and broader chemical tasks that rely on structural descriptions.

The growing capability of video generation poses escalating security risks, making reliable detection increasingly essential. In this paper, we introduce VideoVeritas, a framework that integrates fine-grained perception and fact-based reasoning. We observe that while current multi-modal large language models (MLLMs) exhibit strong reasoning capacity, their granular perception ability remains limited. To mitigate this, we introduce Joint Preference Alignment and Perception Pretext Reinforcement Learning (PPRL). Specifically, rather than directly optimizing for detection task, we adopt general spatiotemporal grounding and self-supervised object counting in the RL stage, enhancing detection performance with simple perception pretext tasks. To facilitate robust evaluation, we further introduce MintVid, a light yet high-quality dataset containing 3K videos from 9 state-of-the-art generators, along with a real-world collected subset that has factual errors in content. Experimental results demonstrate that existing methods tend to bias towards either superficial reasoning or mechanical analysis, while VideoVeritas achieves more balanced performance across diverse benchmarks.

Multilingual translation benchmarks are typically sourced in English and translated into other languages, treating language pairs as the unit of evaluation---a design that is prone to contamination over time and overlooks locale and cultural considerations. We therefore advocate for source-contrastive evaluation and instantiate it with Cultivar, a localised subset of FLORES, which enables locale-specific translation evaluation. When paired with unlocalised counterparts, performance discrepancy allows the probing of data contamination and localisation robustness. We benchmark 32 open-weight models and find that MT-specialised models are less robust, a few models potentially overfit FLORES, and models tend to translate US content better than that of other locales, regardless of language.

Reinforcement learning against rubrics, lists of criteria graded by an LLM judge, has become a standard way to post-train language models on tasks with no deterministic answer. The rubric, however, is a fixed proxy for quality, never a complete description of it, and a policy trained against it long enough will learn to exploit the difference. We measure this directly. Training Qwen3-8B with Group Relative Policy Optimization (GRPO) on medical and science rubrics and grading out-of-distribution (OOD) benchmarks with both the training judge and a stronger gold judge, we find that the two scores diverge during training. The training judge's score keeps climbing while the gold judge's score peaks and then falls, by 3 points on HealthBench-Hard and by 22 points on ResearchQA. A judge with a fixed bias would shift the gold curve by a constant, not send it down while the training score rises, so the divergence is reward hacking, not judge noise. We propose Rubric Dropout, a one-line fix borrowed from neuron dropout. At every step, we randomly drop a subset of the rubric's criteria before computing the reward, so the policy never optimizes the same rubric twice. The dropped subset is shared across each rollout group, so GRPO's group-relative advantages stay comparable, and evaluation always uses the full rubric. Comparing no dropout against dropout at 30% and 50% on both benchmark pairs, dropout raises the OOD gold score at every matched checkpoint (+1 to +2 points on HealthBench-Hard, +6 to +7 points on ResearchQA), lowers the two hacking measures we track, and costs nothing in domain. Sweeping the dropout fraction shows a broad 30-50% sweet spot, while the natural alternative, reweighting criteria by how useful they are to training, performs worse than no intervention at all in our setting.

Image restoration under adverse conditions, such as underwater, haze or fog, and low-light environments, remains a highly challenging problem due to complex physical degradations and severe information loss. Existing datasets are predominantly limited to a single degradation type or heavily rely on synthetic data without stereo consistency, inherently restricting their applicability in real-world scenarios. To address this, we introduce M3D-Stereo, a stereo dataset with 7904 high-resolution image pairs for image restoration research acquired in multiple media with multiple controlled degradation levels. It encompasses four degradation scenarios: underwater scatter, haze/fog, underwater low-light, and haze low-light. Each scenario forms a subset, and is divided into six levels of progressive degradation, allowing fine-grained evaluations of restoration methods with increasing severity of degradation. Collected via a laboratory setup, the dataset provides aligned stereo image pairs along with their pixel-wise consistent clear ground truths. Two restoration tasks, single-level and mixed-level degradation, were performed to verify its validity. M3D-Stereo establishes a better controlled and more realistic benchmark to evaluate image restoration and stereo matching methods in complex degradation environments. It is made public under LGPLv3 license.