Daily Papers

Using Multimodal Large Language Models (MLLMs) as judges to achieve precise and consistent evaluations has gradually become an emerging paradigm across various domains. Evaluating the capability and reliability of MLLM-as-a-judge systems is therefore essential for ensuring trustworthy assessment. Existing judge benchmarks categorize samples by task types but fail to capture the fundamental judgment capabilities required for reliable evaluation. In this work, we introduce M-JudgeBench, a ten-dimensional capability-oriented benchmark designed to comprehensively assess the judgment abilities of MLLMs. Our benchmark decomposes evaluation into pairwise Chain-of-Thought (CoT) comparison, length bias avoidance, and process error detection tasks, jointly covering ten fine-grained subtasks. This design enables diagnosis of model reliability across reasoning styles, response lengths, and cross-model variations. Systematic evaluation uncovers the systematic weaknesses in existing MLLM-as-a-judge systems. To address this issue, we further propose Judge-MCTS, a data construction framework generating pairwise reasoning trajectories with various correctness and length. Using Judge-MCTS, we construct an MCTS-augmented dataset and train M-Judger, a series of strong judge models. Extensive experiments demonstrate the superiority of M-Judger on existing judge benchmarks as well as M-JudgeBench. Overall, our work establishes a more principled foundation for evaluating MLLM-as-a-judge through M-JudgeBench and Judge-MCTS framework, paving the way for future research on judge model evaluation and capability-driven judge training.

The use of language models for automatically evaluating long-form text (LLM-as-a-judge) is becoming increasingly common, yet most LLM judges are optimized exclusively for English, with strategies for enhancing their multilingual evaluation capabilities remaining largely unexplored in the current literature. This has created a disparity in the quality of automatic evaluation methods for non-English languages, ultimately hindering the development of models with better multilingual capabilities. To bridge this gap, we introduce M-Prometheus, a suite of open-weight LLM judges ranging from 3B to 14B parameters that can provide both direct assessment and pairwise comparison feedback on multilingual outputs. M-Prometheus models outperform state-of-the-art open LLM judges on multilingual reward benchmarks spanning more than 20 languages, as well as on literary machine translation (MT) evaluation covering 4 language pairs. Furthermore, M-Prometheus models can be leveraged at decoding time to significantly improve generated outputs across all 3 tested languages, showcasing their utility for the development of better multilingual models. Lastly, through extensive ablations, we identify the key factors for obtaining an effective multilingual judge, including backbone model selection and training on natively multilingual feedback data instead of translated data. We release our models, training dataset, and code.

The Automated Audio Captioning (AAC) task asks models to generate natural language descriptions of an audio input. Evaluating these machine-generated audio captions is a complex task that requires considering diverse factors, among them, auditory scene understanding, sound-object inference, temporal coherence, and the environmental context of the scene. While current methods focus on specific aspects, they often fail to provide an overall score that aligns well with human judgment. In this work, we propose CLAIR-A, a simple and flexible method that leverages the zero-shot capabilities of large language models (LLMs) to evaluate candidate audio captions by directly asking LLMs for a semantic distance score. In our evaluations, CLAIR-A better predicts human judgements of quality compared to traditional metrics, with a 5.8% relative accuracy improvement compared to the domain-specific FENSE metric and up to 11% over the best general-purpose measure on the Clotho-Eval dataset. Moreover, CLAIR-A offers more transparency by allowing the language model to explain the reasoning behind its scores, with these explanations rated up to 30% better by human evaluators than those provided by baseline methods. CLAIR-A is made publicly available at https://github.com/DavidMChan/clair-a.

Recent advancements in large language models (LLMs) have given rise to the LLM-as-a-judge paradigm, showcasing their potential to deliver human-like judgments. However, in the field of machine translation (MT) evaluation, current LLM-as-a-judge methods fall short of learned automatic metrics. In this paper, we propose Multidimensional Multi-Agent Debate (M-MAD), a systematic LLM-based multi-agent framework for advanced LLM-as-a-judge MT evaluation. Our findings demonstrate that M-MAD achieves significant advancements by (1) decoupling heuristic MQM criteria into distinct evaluation dimensions for fine-grained assessments; (2) employing multi-agent debates to harness the collaborative reasoning capabilities of LLMs; (3) synthesizing dimension-specific results into a final evaluation judgment to ensure robust and reliable outcomes. Comprehensive experiments show that M-MAD not only outperforms all existing LLM-as-a-judge methods but also competes with state-of-the-art reference-based automatic metrics, even when powered by a suboptimal model like GPT-4o mini. Detailed ablations and analysis highlight the superiority of our framework design, offering a fresh perspective for LLM-as-a-judge paradigm. Our code and data are publicly available at https://github.com/SU-JIAYUAN/M-MAD.

The paradigm of LLM-as-a-judge is emerging as a scalable and efficient alternative to human evaluation, demonstrating strong performance on well-defined tasks. However, its reliability in open-ended tasks with dynamic environments and complex interactions remains unexplored. To bridge the gap, we introduce WebDevJudge, a systematic benchmark for assessing LLM-as-a-judge performance in web development, with support for both non-interactive evaluation based on static observations and continuous interactive evaluation with a dynamic web environment. WebDevJudge comprises human preference labels over paired web implementations, annotated with structured and query-grounded rubrics to ensure high-quality ground truth. Using this benchmark, we comprehensively evaluate various evaluators, including LLMs, MLLMs, and agentic workflows. We systematically investigate the impact of different paradigms and guidance mechanisms. Our experiments reveal a significant gap between LLM judges and human experts. In-depth analysis indicates this gap stems from fundamental model limitations, including failures in recognizing functional equivalence, verifying task feasibility, and mitigating bias. Overall, WebDevJudge presents a significant challenge to LLM-as-a-judge, offering insights to guide future research toward developing more reliable and capable automated evaluators for complicated scenarios. Code and data are available at https://github.com/lcy2723/WebDevJudge.

Instruction following is a core capability of modern Large language models (LLMs), making evaluating this capability essential to understanding these models. The Instruction Following Evaluation (IFEval) benchmark from the literature does this using objective criteria, offering a measure of LLM performance without subjective AI or human judgement. However, it only includes English instructions, limiting its ability to assess LLMs in other languages. We propose the Multilingual Instruction Following Evaluation (M-IFEval) benchmark, expanding the evaluation to French, Japanese, and Spanish, with both general and language-specific instructions. Applying this benchmark to 8 state-of-the-art LLMs, we find that benchmark performance across languages and instruction types can vary widely, underscoring the importance of a multilingual benchmark for evaluating LLMs in a diverse cultural context.

Sycophancy in LLMs is documented across 70+ papers, but expert agreement on construct boundaries remains low (ICC=.184; Ye et al., 2026). The construct fragments because behavioral classification depends on which surface form is privileged. We adopt a materials-science framing: conversation as test specimen under load, LLM-model as material charge, pushback as progressive load, stance-flip as material failure. We characterize this failure across three loading cases (debate n=1000; false-presuppositions n=3400; ethical-setting n=3400; 10-17 material charges per case; 7800 specimens total) using 14 turn-level axis-measurements spanning velocity, damage accumulation, frame-drift, brittleness, and direction stability, plus three speaker-resolved axes from an independent pipeline. The measurements are Hooke-coupled (σ= E cdot varepsilon analog) and reproduce across loading cases with effects up to |r_{rb}| = 0.35 on debate; the sign structure adds a second pattern: the ethical-setting case inverts the velocity and accumulation blocks. Variance composition partitions into two profiles: debate is charge-dominated (brittle-fracture-like: the material grade decides), false-presuppositions and ethical-setting are topic-dominated (creep-like: the load decides); the ratios (2.03 vs 0.13/0.17) are estimator-dependent, for debate even in direction. Cross-judge reliability (GPT-4o vs Haiku 4.5) shows debate scoring is judge-robust (Cohen's κ= 0.88) while false-presupposition scoring is judge-sensitive (κ= 0.36) -- a caveat single-judge benchmarks must report. This is the methodological move Ye et al.'s diagnosis calls for: a multi-axis characterization that does not depend on which surface form of the construct one privileges.

People increasingly use large language models (LLMs) for mental health support, yet their safety in evolving, high-risk conversations remains poorly characterised. We developed K-Bench, a clinician-calibrated, protected benchmark evaluating 125 model configurations representing 33 base models from 14 providers across a fixed cohort of 200 multi-turn vignettes involving suicide, self-harm, domestic violence, substance misuse, and no-risk presentations. Synthetic patient conversations showed substantial distributional overlap with real human-AI conversations. A frozen GPT-4o judge achieved 94.2% exact agreement with clinician consensus across 6,751 eligible item comparisons from 151 clinician-rated transcripts. Leading models combined strong supportive conversation with combined-risk scores above 95, whereas risk exploration exposed substantial variation among lower-performing configurations. Therapeutic prompting produced configuration-specific gains concentrated among weaker models, while elevated reasoning produced no average improvement. K-Bench combines broader clinical coverage and configuration-scale comparison with a continuously updated public leaderboard whose operational test materials are protected from direct optimisation. The leaderboard is available at www.k-bench.ai.

Machine learning models in health care are often deployed in settings where it is important to protect patient privacy. In such settings, methods for differentially private (DP) learning provide a general-purpose approach to learn models with privacy guarantees. Modern methods for DP learning ensure privacy through mechanisms that censor information judged as too unique. The resulting privacy-preserving models, therefore, neglect information from the tails of a data distribution, resulting in a loss of accuracy that can disproportionately affect small groups. In this paper, we study the effects of DP learning in health care. We use state-of-the-art methods for DP learning to train privacy-preserving models in clinical prediction tasks, including x-ray classification of images and mortality prediction in time series data. We use these models to perform a comprehensive empirical investigation of the tradeoffs between privacy, utility, robustness to dataset shift, and fairness. Our results highlight lesser-known limitations of methods for DP learning in health care, models that exhibit steep tradeoffs between privacy and utility, and models whose predictions are disproportionately influenced by large demographic groups in the training data. We discuss the costs and benefits of differentially private learning in health care.

The increasing adoption of autonomous coding agents accelerates software development but also introduces scoped security risks within high-impact file paths that can outpace traditional human review capacity. While prior research has primarily evaluated these systems in terms of functional correctness and productivity, this paper presents a large-scale empirical study using the AIDev dataset to systematically characterize security code smells in agent-generated pull requests (PRs). Through a combination of a validated LLM-as-a-judge framework and manual qualitative analysis, we identify and classify security misconfigurations across 16,112 file changes spanning 4,022 pull requests. Our results reveal that 38.9% of agent-generated PRs contain at least one security smell, with supply chain integrity issues accounting for 82.3% of all detected security smells. Furthermore, hard-coded credentials constitute 99.6% of all critical-severity security smells. Crucially, we find that human collaborators are responsible for introducing 67.6% of genuine leaked secrets within these agent-assisted workflows, while existing automated and human review processes fail to detect 81.1% of these credentials prior to integration. These findings highlight substantial security risks in agent-assisted software development workflows and suggest a potential reduction in developer vigilance. They also underscore the urgent need for context-aware security guardrails implemented directly at the point of human-AI collaboration.

There is a lack of benchmarks for evaluating large language models (LLMs) in long-form medical question answering (QA). Most existing medical QA evaluation benchmarks focus on automatic metrics and multiple-choice questions. While valuable, these benchmarks fail to fully capture or assess the complexities of real-world clinical applications where LLMs are being deployed. Furthermore, existing studies on evaluating long-form answer generation in medical QA are primarily closed-source, lacking access to human medical expert annotations, which makes it difficult to reproduce results and enhance existing baselines. In this work, we introduce a new publicly available benchmark featuring real-world consumer medical questions with long-form answer evaluations annotated by medical doctors. We performed pairwise comparisons of responses from various open and closed-source medical and general-purpose LLMs based on criteria such as correctness, helpfulness, harmfulness, and bias. Additionally, we performed a comprehensive LLM-as-a-judge analysis to study the alignment between human judgments and LLMs. Our preliminary results highlight the strong potential of open LLMs in medical QA compared to leading closed models. Code & Data: https://github.com/lavita-ai/medical-eval-sphere

Multimodal Large Language Models (MLLMs) are increasingly used for content moderation, yet their robustness in short-form video contexts remains underexplored. Current safety evaluations often rely on unimodal attacks, failing to address combined attack vulnerabilities. In this paper, we introduce a comprehensive framework for evaluating the tri-modal safety of MLLMs. First, we present the Short-Video Multimodal Adversarial (SVMA) dataset, comprising diverse short-form videos with human-guided synthetic adversarial attacks. Second, we propose ChimeraBreak, a novel tri-modal attack strategy that simultaneously challenges visual, auditory, and semantic reasoning pathways. Extensive experiments on state-of-the-art MLLMs reveal significant vulnerabilities with high Attack Success Rates (ASR). Our findings uncover distinct failure modes, showing model biases toward misclassifying benign or policy-violating content. We assess results using LLM-as-a-judge, demonstrating attack reasoning efficacy. Our dataset and findings provide crucial insights for developing more robust and safe MLLMs.

Human beings are social creatures. We routinely reason about other agents, and a crucial component of this social reasoning is inferring people's goals as we learn about their actions. In many settings, we can perform intuitive but reliable goal inference from language descriptions of agents, actions, and the background environments. In this paper, we study this process of language driving and influencing social reasoning in a probabilistic goal inference domain. We propose a neuro-symbolic model that carries out goal inference from linguistic inputs of agent scenarios. The "neuro" part is a large language model (LLM) that translates language descriptions to code representations, and the "symbolic" part is a Bayesian inverse planning engine. To test our model, we design and run a human experiment on a linguistic goal inference task. Our model closely matches human response patterns and better predicts human judgements than using an LLM alone.

AI-assisted peer review is increasingly discussed and adopted as a tool to support the scientific publishing process, yet there is little systematic understanding of how publication venues regulate its use or of how capable current AI review systems are. We address these questions by first surveying reviewer-facing AI policies across 111 leading AI/NLP conferences and medical journals, revealing substantial regulation differences between the two communities. Second, we evaluate AI-generated peer reviews at ICLR 2026 and Nature Communications using a novel dataset comprising original manuscript submissions and several hundred human- and machine-generated reviews. We compare reviews produced by open-source and proprietary models using complementary evaluation metrics, including LLM-as-a-Judge, score alignment, granularity, and overlap with human reviewers' concerns. Our results show that current LLMs can generate detailed and fluent reviews but exhibit systematic weaknesses, such as overly positive recommendations, generic criticism, and uneven evidence grounding. We demonstrate that aggregate quality scores alone can overestimate review quality and argue for multi-dimensional evaluation of AI-generated peer reviews.

Farmer.Chat is Digital Green's farm advisory service for smallholder farmers. When something looks wrong with a crop, the farmer takes a photograph and sends it, and that photograph is the whole question: no symptom described, no crop named, often no text at all. The service has to determine whether the picture can be used, what crop it shows, and what is wrong with it, from images taken on cheap phones in a field, in poor light and with a moving camera. The system doing this today cannot be adjusted. It has no adjustable thresholds for photograph rejection, crops and problems cannot be added, and there is no confidence cut-off to set. We study about 1.16 million photographs sent to Farmer.Chat from Ethiopia, India, Kenya and Nigeria. The production quality gate rejected 46.8% of the images it judged, over a quarter of those reaching diagnosis returned no crop name, and 35.8% of the labelled problems filed under "disease" are pests, identifiable without the crop. We therefore split the work into three stages: a quality gate (M0), a crop detector (M1), and a disease or pest detector (M2). Route A fills all three with one fine-tuned vision-language model (Qwen3-VL-4B) answering in a single call. Route B fills each with a small specialist model (DaViT, YOLO26). We replace our production GPT-4o quality gate with a small MobileNetV3 gate at 86.9% F1 in 12 ms. On one test set scored the same way for every system, a hierarchical DaViT-Base achieves 95.41% crop accuracy against 91.46% for the production baseline. It also leads on diagnosis and never declines to answer, while every language model in the comparison leaves a large share of rows with no diagnosis. The fine-tuned model retains two capabilities the specialists do not have: one call for all three stages, and a request for a better photograph when the image cannot support an answer.

Astronomers have typically set out to solve supervised machine learning problems by creating their own representations from scratch. We show that deep learning models trained to answer every Galaxy Zoo DECaLS question learn meaningful semantic representations of galaxies that are useful for new tasks on which the models were never trained. We exploit these representations to outperform several recent approaches at practical tasks crucial for investigating large galaxy samples. The first task is identifying galaxies of similar morphology to a query galaxy. Given a single galaxy assigned a free text tag by humans (e.g. "#diffuse"), we can find galaxies matching that tag for most tags. The second task is identifying the most interesting anomalies to a particular researcher. Our approach is 100% accurate at identifying the most interesting 100 anomalies (as judged by Galaxy Zoo 2 volunteers). The third task is adapting a model to solve a new task using only a small number of newly-labelled galaxies. Models fine-tuned from our representation are better able to identify ring galaxies than models fine-tuned from terrestrial images (ImageNet) or trained from scratch. We solve each task with very few new labels; either one (for the similarity search) or several hundred (for anomaly detection or fine-tuning). This challenges the longstanding view that deep supervised methods require new large labelled datasets for practical use in astronomy. To help the community benefit from our pretrained models, we release our fine-tuning code Zoobot. Zoobot is accessible to researchers with no prior experience in deep learning.

Current large speech language models (Speech-LLMs) often exhibit limitations in empathetic reasoning, primarily due to the absence of training datasets that integrate both contextual content and paralinguistic cues. In this work, we propose two approaches to incorporate contextual paralinguistic information into model training: (1) an explicit method that provides paralinguistic metadata (e.g., emotion annotations) directly to the LLM, and (2) an implicit method that automatically generates novel training question-answer (QA) pairs using both categorical and dimensional emotion annotations alongside speech transcriptions. Our implicit method boosts performance (LLM-judged) by 38.41% on a human-annotated QA benchmark, reaching 46.02% when combined with the explicit approach, showing effectiveness in contextual paralinguistic understanding. We also validate the LLM judge by demonstrating its correlation with classification metrics, providing support for its reliability.

This paper proposes a multi-agent framework built on CrewAI [1] for conversational business intelligence. Five specialized AI agents operate in a sequential pipeline to process natural language queries, retrieve and analyze data, generate visualizations via the Model Context Protocol (MCP) [2], and deliver actionable insights. The platform features a defense-in-depth security architecture for multi-tenant data isolation and a query parameterization mechanism for transforming conversational insights into reusable dashboard components. Evaluation across 300 end-to-end test cases spanning synthetic and production enterprise datasets demonstrates 95.3% functional accuracy, a mean response latency of 24 seconds, and a response quality score of 4.52/5.0 as assessed by an LLM-as-a-Judge framework, with a 93.0% hallucination-free rate, representing a 22.6 percentage point accuracy improvement and 20.2% quality gain over a single-agent baseline. Cross-model evaluation across four LLM backends and human expert validation confirm architectural generalizability and evaluator reliability. An ablation study confirms that the Data Analysis and Report Aggregation agents are the primary drivers of output quality.

Despite recent advances in text-to-image (T2I) models, they often fail to faithfully render all elements of complex prompts, frequently omitting or misrepresenting specific objects and attributes. Test-time optimization has emerged as a promising approach to address this limitation by refining generation without the need for retraining. In this paper, we propose a fine-grained test-time optimization framework that enhances compositional faithfulness in T2I generation. Unlike most of prior approaches that rely solely on a global image/text similarity score, our method decomposes the input prompt into semantic concepts and evaluates alignment at both the global and concept levels. A fine-grained variant of CLIP is used to compute concept-level correspondence, producing detailed feedback on missing or inaccurate concepts. This feedback is fed into an iterative prompt refinement loop, enabling the large language model to propose improved prompts. Experiments on DrawBench and CompBench prompts demonstrate that our method significantly improves concept coverage and human-judged faithfulness over both standard test-time optimization and the base T2I model. Code is available at: https://github.com/AmirMansurian/NoConceptLeftBehind

Optical flow estimation is essential for video processing tasks, such as restoration and action recognition. The quality of videos is constantly increasing, with current standards reaching 8K resolution. However, optical flow methods are usually designed for low resolution and do not generalize to large inputs due to their rigid architectures. They adopt downscaling or input tiling to reduce the input size, causing a loss of details and global information. There is also a lack of optical flow benchmarks to judge the actual performance of existing methods on high-resolution samples. Previous works only conducted qualitative high-resolution evaluations on hand-picked samples. This paper fills this gap in optical flow estimation in two ways. We propose DPFlow, an adaptive optical flow architecture capable of generalizing up to 8K resolution inputs while trained with only low-resolution samples. We also introduce Kubric-NK, a new benchmark for evaluating optical flow methods with input resolutions ranging from 1K to 8K. Our high-resolution evaluation pushes the boundaries of existing methods and reveals new insights about their generalization capabilities. Extensive experimental results show that DPFlow achieves state-of-the-art results on the MPI-Sintel, KITTI 2015, Spring, and other high-resolution benchmarks.

Large language models (LLMs) are increasingly applied as automatic evaluators for natural language generation assessment often using pairwise comparative judgements. Existing approaches typically rely on single judges or aggregate multiple judges assuming equal reliability. In practice, LLM judges vary substantially in performance across tasks and evaluation aspects, and their judgment probabilities may be biased and inconsistent. Furthermore, human-labelled supervision for judge calibration may be unavailable. We first empirically demonstrate that inconsistencies in LLM comparison probabilities exist and show that it limits the effectiveness of direct probability-based ranking. To address this, we study the LLM-asa-jury setting and propose BT-sigma, a judge-aware extension of the Bradley-Terry model that introduces a discriminator parameter for each judge to jointly infer item rankings and judge reliability from pairwise comparisons alone. Experiments on benchmark NLG evaluation datasets show that BT-sigma consistently outperforms averaging-based aggregation methods, and that the learned discriminators strongly correlate with independent measures of the cycle consistency of LLM judgments. Further analysis reveals that BT-sigma can be interpreted as an unsupervised calibration mechanism that improves aggregation by modelling judge reliability.

Language-based object detection is a promising direction towards building a natural interface to describe objects in images that goes far beyond plain category names. While recent methods show great progress in that direction, proper evaluation is lacking. With OmniLabel, we propose a novel task definition, dataset, and evaluation metric. The task subsumes standard- and open-vocabulary detection as well as referring expressions. With more than 28K unique object descriptions on over 25K images, OmniLabel provides a challenging benchmark with diverse and complex object descriptions in a naturally open-vocabulary setting. Moreover, a key differentiation to existing benchmarks is that our object descriptions can refer to one, multiple or even no object, hence, providing negative examples in free-form text. The proposed evaluation handles the large label space and judges performance via a modified average precision metric, which we validate by evaluating strong language-based baselines. OmniLabel indeed provides a challenging test bed for future research on language-based detection.

We present the joint contribution of Unbabel and Instituto Superior T\'ecnico to the WMT 2023 Shared Task on Quality Estimation (QE). Our team participated on all tasks: sentence- and word-level quality prediction (task 1) and fine-grained error span detection (task 2). For all tasks, we build on the COMETKIWI-22 model (Rei et al., 2022b). Our multilingual approaches are ranked first for all tasks, reaching state-of-the-art performance for quality estimation at word-, span- and sentence-level granularity. Compared to the previous state-of-the-art COMETKIWI-22, we show large improvements in correlation with human judgements (up to 10 Spearman points). Moreover, we surpass the second-best multilingual submission to the shared-task with up to 3.8 absolute points.

As language models improve and become capable of performing more complex tasks across modalities, evaluating them automatically becomes increasingly challenging. Developing strong and robust task-specific automatic metrics gets harder, and human-annotated test sets -- which are expensive to create -- saturate more quickly. A compelling alternative is to design reliable strategies to automate the creation of test data and evaluation, but previous attempts either rely on pre-existing data, or focus solely on individual tasks. We present Zero-shot Benchmarking (ZSB), a framework for creating high-quality benchmarks for any task by leveraging language models for both synthetic test data creation and evaluation. ZSB is simple and flexible: it requires only the creation of a prompt for data generation and one for evaluation; it is scalable to tasks and languages where collecting real-world data is costly or impractical; it is model-agnostic, allowing the creation of increasingly challenging benchmarks as models improve. To assess the effectiveness of our framework, we create benchmarks for five text-only tasks and a multi-modal one: general capabilities in four languages (English, Chinese, French, and Korean), translation, and general vision-language capabilities in English. We then rank a broad range of open and closed systems on our benchmarks. ZSB rankings consistently correlate strongly with human rankings, outperforming widely-adopted standard benchmarks. Through ablations, we find that strong benchmarks can be created with open models, and that judge model size and dataset variety are crucial drivers of performance. We release all our benchmarks, and code to reproduce our experiments and to produce new benchmarks.

Benchmark infrastructure for personally identifiable information (PII) detection remains limited: existing corpora cover few entity types, use ad hoc generation conditions, and do not show which surface conditions cause detector failures. We present REDACT, a systematically controlled multilingual PII benchmark with 13,427 records, 324,078 entity annotations, 51 entity types, 4,127 surface-form patterns, and 25 languages across 9 scripts. A strength-2 covering-array sampler controls nine generation axes: domain, format, difficulty, length, density, code-switching, language, adjacency, and co-occurrence. Three entity-level metadata fields (disclosure status, disclosure form, and a GDPR-aligned sensitivity tier) enable stratified evaluation beyond aggregate or per-type F1. From the full benchmark, we evaluate five detectors (Presidio, GLiNER, the OpenAI Privacy Filter, GPT-4.1, and Claude Sonnet 4.6) on a locked, language-stratified sample of 1,000 records. Aggregate F1 masks an architecture-dependent failure structure: the rule-based detector performs poorly on the highest-stakes data, including HIGH-sensitivity categories (recall 0.07) and non-verbatim disclosure forms, while the LLM detectors remain more robust, with the HIGH tier as their strongest sensitivity slice. A three-model reference-free LLM-as-judge assessment corroborates that sensitivity-tier assignment is the task's hardest axis. We release the benchmark, schema, prompts, and stratified evaluation harness.

Whether a language model has improved itself is increasingly judged not by mean accuracy but by which individual problems it gains and loses. Tracking these transitions means differencing two noisy estimates, leaving them vulnerable to measurement artifacts. Auditing three rounds of rank-32 LoRA self-training on Qwen3-8B against a frozen control pushed through the identical pipeline, we identify seven measurement failures, each of which inverts a reported finding when its control is absent. Several are standard practice. A ledger built on a single greedy decode manufactures capability changes on an untrained model, largely an artifact of inference batching; the expansion statistic separating acquisition from sharpening assigns that same model a rate of 0.280. The natural threshold repair does not survive replication: estimated across the frozen comparisons such a design already contains, its null stays non-zero. We replace it with a per-problem exact test against a pooled baseline under false-discovery-rate control, which detects nothing on any held-out replicate and is unchanged under the multiple-testing rule, error rate and pool size. Applied to a ladder of arms matched in stream, volume and evaluation, the audit finds that external distillation improves problems the base model rarely reaches while three forms of self-training do not; a regression rejects this asymmetry as a by-product of distillation's larger overall gain (p < 10^{-8}). On the far smaller set of problems the base model never reaches, the evidence is inconclusive, while self-training corrupts problems solved at baseline at rates well above the measured floor. Transition-level auditing therefore requires a separately measured null for every statistic it reports: nulls that cost no new experiments, built from baseline replicates a multi-arm study already owns, though not from as few as most possess.

African agriculture is undergoing rapid transformation. Annual maps of crop fields are key to understanding the nature of this transformation, but such maps are currently lacking and must be developed using advanced machine learning models trained on high resolution remote sensing imagery. To enable the development of such models, we delineated field boundaries in 33,746 Planet images captured between 2017 and 2023 across the continent using a custom labeling platform with built-in procedures for assessing and mitigating label error. We collected 42,403 labels, including 7,204 labels arising from tasks dedicated to assessing label quality (Class 1 labels), 32,167 from sites mapped once by a single labeller (Class 2) and 3,032 labels from sites where 3 or more labellers were tasked to map the same location (Class 4). Class 1 labels were used to calculate labeller-specific quality scores, while Class 1 and 4 sites mapped by at least 3 labellers were used to further evaluate label uncertainty using a Bayesian risk metric. Quality metrics showed that label quality was moderately high (0.75) for measures of total field extent, but low regarding the number of individual fields delineated (0.33), and the position of field edges (0.05). These values are expected when delineating small-scale fields in 3-5 m resolution imagery, which can be too coarse to reliably distinguish smaller fields, particularly in dense croplands, and therefore requires substantial labeller judgement. Nevertheless, previous work shows that such labels can train effective field mapping models. Furthermore, this large, probabilistic sample on its own provides valuable insight into regional agricultural characteristics, highlighting variations in the median field size and density. The imagery and vectorized labels along with quality information is available for download from two public repositories.

As Large Language Models (LLMs) saturate elementary benchmarks, the research frontier has shifted from generation to the reliability of automated evaluation. We demonstrate that standard "LLM-as-a-Judge" protocols suffer from a systematic Alignment Gap when applied to upper-undergraduate to early graduate level mathematics. To quantify this, we introduce QEDBench, the first large-scale dual-rubric alignment benchmark to systematically measure alignment with human experts on university-level math proofs by contrasting course-specific rubrics against expert common knowledge criteria. By deploying a dual-evaluation matrix (7 judges x 5 solvers) against 1,000+ hours of human evaluation, we reveal that certain frontier evaluators like Claude Opus 4.5, DeepSeek-V3, Qwen 2.5 Max, and Llama 4 Maverick exhibit significant positive bias (up to +0.18, +0.20, +0.30, +0.36 mean score inflation, respectively). Furthermore, we uncover a critical reasoning gap in the discrete domain: while Gemini 3.0 Pro achieves state-of-the-art performance (0.91 average human evaluation score), other reasoning models like GPT-5 Pro and Claude Sonnet 4.5 see their performance significantly degrade in discrete domains. Specifically, their average human evaluation scores drop to 0.72 and 0.63 in Discrete Math, and to 0.74 and 0.50 in Graph Theory. In addition to these research results, we also release QEDBench as a public benchmark for evaluating and improving AI judges. Our benchmark is publicly published at https://github.com/qqliu/Yale-QEDBench.

Frontier LLMs now perform strongly across a wide range of physics evaluations, but it is hard to disentangle genuine reasoning from recall of established science. We introduce DiscoverPhysics, an interactive benchmark that asks a LLM agent to discover the laws of motion of a simulated world whose physics deliberately deviates from our own. We construct 22 worlds governed by, among others, screened and fractional-power gravity, multi-species couplings, hidden dark-matter-like particles, non-coordinate-free physics, and time-varying interactions. Each world is generated on demand by an N-body simulator, for which the agent proposes several rounds of experiments, observes raw trajectory data, and ultimately submits both a natural-language explanation of the world's physics and a Python implementation of the inferred law. Because solving a world requires the agent to design informative experiments and revise its hypotheses, the benchmark probes long-horizon reasoning over an experimental history. We evaluate submissions along two complementary axes: trajectory MSE on held-out particles and an LLM-judged explanation score following an expert-written rubric assessing conceptual understanding of each world. Across eleven frontier models, we find that the strongest agents pass only half of the worlds and consistently fail on those where latent structure must be uncovered. Open-source models lag substantially behind commercial models, both in their ability to design informative experiments and in extracting conclusions from the data. We further find that good predictive accuracy does not guarantee high explanation quality and that conceptual understanding depends on hypothesis refinement through well-chosen experiments.

Event Coreference Resolution (ECR) as a pairwise mention classification task is expensive both for automated systems and manual annotations. The task's quadratic difficulty is exacerbated when using Large Language Models (LLMs), making prompt engineering for ECR prohibitively costly. In this work, we propose a graphical representation of events, X-AMR, anchored around individual mentions using a cross-document version of Abstract Meaning Representation. We then linearize the ECR with a novel multi-hop coreference algorithm over the event graphs. The event graphs simplify ECR, making it a) LLM cost-effective, b) compositional and interpretable, and c) easily annotated. For a fair assessment, we first enrich an existing ECR benchmark dataset with these event graphs using an annotator-friendly tool we introduce. Then, we employ GPT-4, the newest LLM by OpenAI, for these annotations. Finally, using the ECR algorithm, we assess GPT-4 against humans and analyze its limitations. Through this research, we aim to advance the state-of-the-art for efficient ECR and shed light on the potential shortcomings of current LLMs at this task. Code and annotations: https://github.com/ahmeshaf/gpt_coref

Human evaluation of AI systems and digital products is costly, slow, and difficult to scale. Offline evaluations are more scalable but often abstract away human diversity and interactive behavior. We therefore introduce MatrAIx, a population-scale simulated-user evaluation infrastructure for testing AI systems and digital products with heterogeneous users. MatrAIx has three core components: First, Persona 8B contains 8.3 billion persona records represented by 1,290 categorical dimensions. Records are either sampled from a dependency graph that preserves correlated attributes or derived from human-authored profiles. We release a quality-filtered coreset of approximately 1 million personas, comprising 599,847 human-grounded and 400,000 synthetic records. Second, the MatrAIx Playground provides four environments in which diverse users evaluate and interact with digital products: Survey, AI Chatbot, Web, and App. Third, MatrAIx provides 1,010 application tasks spanning more than 25 domains, including Commerce, Software, Finance, and Healthcare. We conducted 18,189 evaluation trials across eight representative tasks. Persona agents were powered by three LLMs: Claude Opus 4.8, GPT 5.5, and Claude Haiku 4.5. The resulting feedback captures how decisions and preferences vary across persona backgrounds, including hesitation after a price increase, willingness to continue after an AI assistant fails, and latency tolerance. We conducted two main validation studies: First, a 400-trial controlled study evaluated persona adherence across ten behavioral attributes and all four environments. The declared behavior was expressed or correctly suppressed in 366 trials (91.5%). Second, human and LLM judges evaluated the extraction quality of human-grounded personas. Overall, MatrAIx provides an end-to-end infrastructure for evaluating AI systems and digital products with diverse simulated human users.

Blinded physician evaluation has been considered by many to be the gold standard for assessing clinical reasoning in large language models (LLMs). This is difficult to scale; thus, prior studies typically rely on small physician panels, often from a single institution or specialty, which both limits the scientific questions investigated and makes it unclear whether findings would be reproduced with a different set of evaluators. To more rigorously and scalably study clinical reasoning in AI models, here we introduce PrecepTron, an LLM fine-tuned for physician-level evaluation of open-ended responses. PrecepTron was trained using low-rank adaptation (LoRA) of a 32-billion-parameter model on a small number of physician examples. We also release GRAND-ROUNDS, a new large-scale physician-annotated benchmark of 9,217 scores by 11 physicians across seven studies. We show that frontier LLMs in typical "LLM-as-a-judge" approaches often disagree with physicians and with each other, but fine-tuning PrecepTron on a small number of cases enables physician-level consistent scoring across tasks. We use PrecepTron to reproduce headline findings from five influential studies assessing LLMs for clinical care in JAMA, Science, and Nature Medicine without new human grading. Using PrecepTron, we then pose new questions about how LLMs reason in medicine that would have been infeasible with human grading alone, including measuring the diagnostic accuracy of frontier LLMs when clinical cases are provided piecemeal, even token by token. Together, PrecepTron and GRAND-ROUNDS provide a foundation for reproducible, large-scale study of how LLMs reason in medicine. All code, data, and labels are made freely available for researchers.

Position bias has proven to be a prevalent issue of modern language models (LMs), where the models prioritize content based on its position within the given context. This bias often leads to unexpected model failures and hurts performance, robustness, and reliability across various applications. Our mechanistic analysis attributes the position bias to two components employed in nearly all state-of-the-art LMs: causal attention and relative positional encodings. Specifically, we find that causal attention generally causes models to favor distant content, while relative positional encodings like RoPE prefer nearby ones based on the analysis of retrieval-augmented question answering (QA). Further, our empirical study on object detection reveals that position bias is also present in vision-language models (VLMs). Based on the above analyses, we propose to ELIMINATE position bias caused by different input segment orders (e.g., options in LM-as-a-judge, retrieved documents in QA) in a TRAINING-FREE ZERO-SHOT manner. Our method changes the causal attention to bidirectional attention between segments and utilizes model attention values to decide the relative orders of segments instead of using the order provided in input prompts, therefore enabling Position-INvariant inferencE (PINE) at the segment level. By eliminating position bias, models achieve better performance and reliability in downstream tasks where position bias widely exists, such as LM-as-a-judge and retrieval-augmented QA. Notably, PINE is especially useful when adapting LMs for evaluating reasoning pairs: it consistently provides 8 to 10 percentage points performance gains in most cases, and makes Llama-3-70B-Instruct perform even better than GPT-4-0125-preview on the RewardBench reasoning subset.

Automatic metrics are extensively used to evaluate natural language processing systems. However, there has been increasing focus on how they are used and reported by practitioners within the field. In this paper, we have conducted a survey on the use of automatic metrics, focusing particularly on natural language generation (NLG) tasks. We inspect which metrics are used as well as why they are chosen and how their use is reported. Our findings from this survey reveal significant shortcomings, including inappropriate metric usage, lack of implementation details and missing correlations with human judgements. We conclude with recommendations that we believe authors should follow to enable more rigour within the field.

To what extent do vision-and-language foundation models possess a realistic world model (observation times action rightarrow observation) and a dynamics model (observation times observation rightarrow action), when actions are expressed through language? While open-source foundation models struggle with both, we find that fine-tuning them to acquire a dynamics model through supervision is significantly easier than acquiring a world model. In turn, dynamics models can be used to bootstrap world models through two main strategies: 1) weakly supervised learning from synthetic data and 2) inference time verification. Firstly, the dynamics model can annotate actions for unlabelled pairs of video frame observations to expand the training data. We further propose a new objective, where image tokens in observation pairs are weighted by their importance, as predicted by a recognition model. Secondly, the dynamics models can assign rewards to multiple samples of the world model to score them, effectively guiding search at inference time. We evaluate the world models resulting from both strategies through the task of action-centric image editing on Aurora-Bench. Our best model achieves a performance competitive with state-of-the-art image editing models, improving on them by a margin of 15% on real-world subsets according to GPT4o-as-judge, and achieving the best average human evaluation across all subsets of Aurora-Bench.

Full-duplex spoken dialogue models support low-latency turn taking, interruption handling, and backchanneling, yet a key capability remains underexplored: steerability, the ability to reliably shift conversational behavior along attributes such as tone, persona, speaking rate, and voice style in response to user instructions. We introduce a taxonomy of text- and audio-based steerability that identifies substantial gaps in current full-duplex models. To address this gap, we introduce SteerDuplex, a Moshi-based full-duplex speech model fine-tuned on natural conversations and synthetic dialogues targeting instruction following, vocal delivery, reasoning, and duplex interaction. We further apply two-stage reinforcement learning (RL) with hybrid rewards, combining verifiable interaction checks and judge-based semantic feedback to improve timing and response continuity. To evaluate full-duplex spoken steerability, we introduce SteerBench, a benchmark with 390 spoken prompts and 1,067 human-authored binary audio and text rubrics spanning tone, persona, style/accent, and speed/length. On SteerBench, supervised training improves audio-steering average pass rate by 44.5 percentage points over the strongest evaluated open baseline. On Audio MultiChallenge, task average pass rate improves by 7 points over its strongest evaluated open baseline. RL further raises source-clean interruption response from 72.5% to 82.5% and reduces synthetic pause barge-in from 26.5% to 9%. Steering and aggregate task scores remain comparable or higher, while reward probes reveal reward hacking through incomplete responses. Our model and benchmark support systematic research on spoken steerability, with reward analysis showing why timing gains must be evaluated alongside response completeness.

Despite rapid advancements in TTS models, a consistent and robust human evaluation framework is still lacking. For example, MOS tests fail to differentiate between similar models, and CMOS's pairwise comparisons are time-intensive. The MUSHRA test is a promising alternative for evaluating multiple TTS systems simultaneously, but in this work we show that its reliance on matching human reference speech unduly penalises the scores of modern TTS systems that can exceed human speech quality. More specifically, we conduct a comprehensive assessment of the MUSHRA test, focusing on its sensitivity to factors such as rater variability, listener fatigue, and reference bias. Based on our extensive evaluation involving 471 human listeners across Hindi and Tamil we identify two primary shortcomings: (i) reference-matching bias, where raters are unduly influenced by the human reference, and (ii) judgement ambiguity, arising from a lack of clear fine-grained guidelines. To address these issues, we propose two refined variants of the MUSHRA test. The first variant enables fairer ratings for synthesized samples that surpass human reference quality. The second variant reduces ambiguity, as indicated by the relatively lower variance across raters. By combining these approaches, we achieve both more reliable and more fine-grained assessments. We also release MANGO, a massive dataset of 47,100 human ratings, the first-of-its-kind collection for Indian languages, aiding in analyzing human preferences and developing automatic metrics for evaluating TTS systems.

High-quality datasets are fundamental to training and evaluating machine learning models, yet their creation-especially with accurate human annotations-remains a significant challenge. Many dataset paper submissions lack originality, diversity, or rigorous quality control, and these shortcomings are often overlooked during peer review. Submissions also frequently omit essential details about dataset construction and properties. While existing tools such as datasheets aim to promote transparency, they are largely descriptive and do not provide standardized, measurable methods for evaluating data quality. Similarly, metadata requirements at conferences promote accountability but are inconsistently enforced. To address these limitations, this position paper advocates for the integration of systematic, rubric-based evaluation metrics into the dataset review process-particularly as submission volumes continue to grow. We also explore scalable, cost-effective methods for synthetic data generation, including dedicated tools and LLM-as-a-judge approaches, to support more efficient evaluation. As a call to action, we introduce DataRubrics, a structured framework for assessing the quality of both human- and model-generated datasets. Leveraging recent advances in LLM-based evaluation, DataRubrics offers a reproducible, scalable, and actionable solution for dataset quality assessment, enabling both authors and reviewers to uphold higher standards in data-centric research. We also release code to support reproducibility of LLM-based evaluations at https://github.com/datarubrics/datarubrics.

Users increasingly turn to large language models for emotional support, yet little is known about how these models actually conduct a psychotherapy interaction. We introduce an ontology of ten therapeutic moves: compact, function-based categories grounded in the MULTI-60 inventory, validated through an annotation campaign with five licensed psychologists, and scaled with a judge-based approach that matches expert agreement. Applying it to real counseling transcripts and model-led sessions, we compare the move distributions between human clinicians and a panel of frontier models. Models over-use inquiry at up to three times the human rate, neglect psychoeducation, and are strongly context-anchored: they carry forward strategies initiated by a human clinician but rarely initiate them themselves. Exposing the ontology as a set of tools roughly halves the mean deviation from the human move distribution and improves turn-level alignment with human therapist by 7-9 percentage points, without any fine-tuning.