NeurIPS 2025 Papers

Traditional correlation measures, such as Pearson’s and Spearman’s coefficients, are limited in their ability to capture complex relationships, particularly nonlinear and multivariate dependencies. The Hirschfeld–Gebelein–Rényi (HGR) maximal correlation offers a powerful alternative by seeking the highest Pearson correlation attainable through nonlinear transformations of two random variables. However, estimating the HGR remains challenging due to the complexity of optimizing arbitrary nonlinear functions. We introduce a new coefficient inspired by the HGR but grounded in the Spearman rank correlation, which we call the Spearman HGR (SHGR). We propose a neural network-based estimator tailored to (i) bivariate correlation matrix, (ii) multivariate correlations between a set of variables and another one, and (iii) full correlation between two sets of variables. The SHGR satisfies Rényi’s axioms, effectively detects nonlinear dependencies, and demonstrates robustness to noise, outliers, and spurious correlations (hallucinations). Additionally, it achieves competitive computational efficiency through tailored neural architectures. Comprehensive Numerical experiments and feature selection tasks confirm that SHGR outperforms existing state-of-the-art methods.

Vision-Language Models (VLMs) are increasingly pivotal for generalist robot manipulation, enabling tasks such as physical reasoning, policy generation, and failure detection. However, their proficiency in these high-level applications often assumes a deep understanding of low-level physical prerequisites, a capability that is largely unverified. To perform actions reliably, robots must comprehend intrinsic object properties (e.g., material, weight), action affordances (e.g., graspable, stackable), and physical constraints (e.g., stability, reachability, or an object's state like being closed). Despite their ubiquitous use in manipulation, we argue that off-the-shelf VLMs may lack this granular, physically-grounded understanding, as these specific prerequisites are often overlooked during training. Addressing this critical gap, we introduce PAC Bench, a comprehensive benchmark designed to systematically evaluate VLMs on their understanding of these core Properties, Affordances, and Constraints (PAC) from a task executability perspective. PAC Bench features a diverse dataset with more than 30,000 annotations, comprising 673 real-world images (115 object classes, 15 property types, 1–3 affordances defined per object class), 100 real-world humanoid view scenarios, and 120 unique simulated constraint scenarios across four tasks. Our evaluations reveal significant gaps in the ability of VLMs to grasp fundamental physical concepts, underscoring their current limitations for reliable robot manipulation and pointing to key areas that require targeted research. PAC Bench also serves as a standardized benchmark for rigorously evaluating the physical reasoning capabilities of VLMs guiding the development of more robust and physically grounded models for robot manipulation.

Vision Language Models (VLMs) excel at complex visual tasks such as VQA and chart understanding, yet recent work suggests they struggle with simple perceptual tests. We present an evaluation that tests vision-language models’ capacity for \emph{nonlocal visual reasoning}- reasoning that requires chaining evidence collected from multiple, possibly distant, regions of an image. We isolate three distinct forms of non‑local vision: \emph{comparative perception}, which demands holding two images in working memory and comparing them; \emph{saccadic search}, which requires making discrete, evidence‑driven jumps to locate successive targets; and \emph{smooth visual search}, which involves searching smoothly along a continuous contour. Flagship models (e.g. GPT-5, Gemini 2.5 Pro, Claude Sonnet 4), even those that perform well on prior primitive‑vision benchmarks, fail these tests and barely exceed random accuracy on two variants of our tasks that are trivial for humans. Our structured evaluation suite allows us to test if VLMs can perform similar visual algorithms to humans. Our findings show that despite gains in raw visual acuity, current models lack core visual reasoning capabilities.

Large language models are increasingly used as judges (LLM-as-a-judge) to evaluate model outputs at scale, but their assessments often diverge systematically from human judgments. We present Bridge, a unified statistical framework that explicitly bridges human and LLM evaluations under both absolute scoring and pairwise comparison paradigms. Bridge posits a latent human preference score for each prompt-response pair and models LLM deviations as linear transformations of covariates that capture sources of discrepancies. This offers a simple and principled framework for refining LLM ratings and characterizing systematic discrepancies between humans and LLMs. We provide an efficient fitting algorithm with asymptotic guarantees for statistical inference. Using six LLM judges and two benchmarks (BigGen Bench and Chatbot Arena), Bridge achieves higher agreement with human ratings (accuracy, calibration, and KL divergence) and exposes systematic human-LLM gaps.

Machine learning-based decision support systems are increasingly deployed in clinical settings, where probabilistic scoring functions are used to inform and prioritize patient management decisions. However, widely used scoring rules, such as accuracy and AUC-ROC, fail to adequately reflect key clinical priorities, including calibration, robustness to distributional shifts, and sensitivity to asymmetric error costs. In this work, we propose a principled yet practical evaluation framework for selecting calibrated thresholded classifiers that explicitly accounts for uncertainty in class prevalences and domain-specific cost asymmetries. Building on the theory of proper scoring rules, particularly the Schervish representation, we derive an adjusted variant of cross-entropy (log score) that averages cost-weighted performance over clinically relevant ranges of class balance. The resulting evaluation is simple to apply, sensitive to clinical deployment conditions, and designed to prioritize models that are both calibrated and robust to real-world variations.

The online evaluation of machine learning models is typically conducted through A/B experiments. Sequential statistical tests are valuable tools for analysing these experiments, as they enable researchers to stop data collection early without increasing the risk of false discoveries. However, existing sequential tests either limit the number of interim analyses or suffer from low statistical power. In this paper, we introduce a novel sequential test designed for the continuous monitoring of A/B experiments. We validate our method using semi-synthetic simulations and demonstrate that it outperforms current state-of-the-art sequential testing approaches. Our method is derived using a new technique that inverts a bound on the probability of threshold crossing, based on a classical maximal inequality.

We present SciArena, an open and collaborative platform for evaluating foundation models on scientific literature-grounded tasks. Unlike traditional benchmarks for scientific literature understanding and synthesis, SciArena engages the research community directly, following the Chatbot Arena evaluation approach of community voting on model comparisons. By leveraging collective intelligence, SciArena offers a community-driven evaluation of model performance on open-ended scientific tasks that demand literature-grounded, long-form responses. The platform currently supports 44 open-source and proprietary foundation models and has collected over 19,000 votes from human researchers across diverse scientific domains. Our analysis of the data collected so far confirms its high quality. We discuss the results and insights based on the model ranking leaderboard. To further promote research in building model-based automated evaluation systems for literature tasks, we release SciArena-Eval, a meta-evaluation benchmark based on our collected preference data. The benchmark measures the accuracy of models in judging answer quality by comparing their pairwise assessments with human votes. Our experiments highlight the benchmark’s challenges and emphasize the need for more reliable automated evaluation methods.

As AI grows more powerful, it will increasingly shape how we understand the world. But with this influence comes the risk of amplifying misinformation and deepening social divides—especially on consequential topics where factual accuracy directly impacts well-being. Scalable Oversight aims to ensure AI systems remain truthful even when their capabilities exceed those of their evaluators. Yet when humans serve as evaluators, their own beliefs and biases can impair judgment. We study whether AI debate can guide biased judges toward the truth by having two AI systems debate opposing sides of controversial factuality claims on COVID-19 and climate change where people hold strong prior beliefs. We conduct two studies. Study I recruits human judges with either mainstream or skeptical beliefs who evaluate claims through two protocols: debate (interaction with two AI advisors arguing opposing sides) or consultancy (interaction with a single AI advisor). Study II uses AI judges with and without human-like personas to evaluate the same protocols. In Study I, debate consistently improves human judgment accuracy and confidence calibration, outperforming consultancy by 4-10\% across COVID-19 and climate change claims. The improvement is most significant for judges with mainstream beliefs (up to +15.2\% accuracy on COVID-19 claims), though debate also helps skeptical judges who initially misjudge claims move toward accurate views (+4.7\% accuracy). In Study II, AI judges with human-like personas achieve even higher accuracy (78.5\%) than human judges (70.1\%) and default AI judges without personas (69.8\%), suggesting their potential for supervising frontier AI models. These findings highlight AI debate as a promising path toward scalable, bias-resilient oversight in contested domains.

Multimodal hallucination in multimodal large language models (MLLMs) restricts the correctness of MLLMs. However, multimodal hallucinations are multi-sourced and arise from diverse causes. Existing benchmarks fail to adequately distinguish between perception-induced hallucinations and reasoning-induced hallucinations. This failure constitutes a significant issue and hinders the diagnosis of multimodal reasoning failures within MLLMs. To address this, we propose the MIRAGE benchmark, which isolates reasoning hallucinations by constructing questions where input images are correctly perceived by MLLMs yet reasoning errors persist. MIRAGE introduces multi-granular evaluation metrics: accuracy, factuality, and LLMs hallucination score for hallucination quantification. Our analysis reveals strong correlations between question types and specific hallucination patterns, particularly systematic failures of MLLMs in spatial reasoning involving complex relationships (\emph{e.g.}, complex geometric patterns across images). This highlights a critical limitation in the reasoning capabilities of current MLLMs and provides targeted insights for hallucination mitigation on specific types. To address these challenges, we propose Logos, a method that combines curriculum reinforcement fine-tuning to encourage models to generate logic-consistent reasoning chains by stepwise reducing learning difficulty, and collaborative hint inference to reduce reasoning complexity. Logos establishes a baseline on MIRAGE, and reduces the logical hallucinations in original base models. Link: \url{https://bit.ly/25mirage}.

The LLM-as-a-judge paradigm, in which a judge LLM system replaces human raters in rating the outputs of other generative AI (GenAI) systems, plays a critical role in scaling and standardizing GenAI evaluations. To validate such judge systems, evaluators assess human--judge agreement by first collecting multiple human ratings for each item in a validation corpus, then aggregating the ratings into a single, per-item gold label rating. For many items, however, rating criteria may admit multiple valid interpretations, so a human or LLM rater may deem multiple ratings "reasonable" or "correct." We call this condition rating indeterminacy. Problematically, many rating tasks that contain rating indeterminacy rely on forced-choice elicitation, whereby raters are instructed to select only one rating for each item. In this paper, we introduce a framework for validating LLM-as-a-judge systems under rating indeterminacy. We draw theoretical connections between different measures of judge system performance under different human--judge agreement metrics, and different rating elicitation and aggregation schemes. We demonstrate that differences in how humans and LLMs resolve rating indeterminacy when responding to forced-choice rating instructions can heavily bias LLM-as-a-judge validation. Through extensive experiments involving 11 real-world rating tasks and 9 commercial LLMs, we show that standard validation approaches that rely upon forced-choice ratings select judge systems that are highly suboptimal, performing as much as 31% worse than judge systems selected by our approach that uses multi-label "response set" ratings to account for rating indeterminacy. We conclude with concrete recommendations for more principled approaches to LLM-as-a-judge validation.

Evaluating large language models (LLMs) poses significant challenges, particularly due to issues of data contamination and the leakage of correct answers. To address these challenges, we introduce ThinkBench, a novel evaluation framework designed to robustly evaluate the reasoning capability of LLMs. ThinkBench proposes a dynamic data generation method for constructing out-of-distribution (OOD) datasets and offers an OOD dataset that contains 2,912 samples drawn from reasoning tasks. ThinkBench unifies the evaluation of reasoning models and non-reasoning models. We evaluate 16 LLMs and 4 PRMs under identical experimental conditions and show that most of the LLMs' performance are far from robust and they face a certain level of data leakage. By dynamically generating OOD datasets, ThinkBench effectively provides a reliable evaluation of LLMs and reduces data contamination impact. Our data and codes are available at https://github.com/huangshulin123/ThinkBench.

Large language models (LLMs) raise concerns about content authenticity and integrity because they can generate human-like text at scale. Text watermarks, which embed detectable statistical signals into generated text, offer a provable way to verify content origin. Many detection methods rely on pivotal statistics that are i.i.d. under human-written text, making goodness-of-fit (GoF) tests a natural tool for watermark detection. However, GoF tests remain largely underexplored in this setting. In this paper, we systematically evaluate eight GoF tests across three popular watermarking schemes, using three open-source LLMs, two datasets, various generation temperatures, and multiple post-editing methods. We find that general GoF tests can improve both the detection power and robustness of watermark detectors. Notably, we observe that text repetition, common in low-temperature settings, gives GoF tests a unique advantage not exploited by existing methods. Our results highlight that classic GoF tests are a simple yet powerful and underused tool for watermark detection in LLMs.

Robust policy evaluation for non-rectangular uncertainty set is generally NP-hard, even in approximation. Consequently, existing approaches suffer from either exponential iteration complexity or significant accuracy gaps. Interestingly, we identify a powerful class of $L_p$-bounded uncertainty sets that avoid these complexity barriers due to their structural simplicity. We further show that this class can be decomposed into infinitely many \texttt{sa}-rectangular $L_p$-bounded sets and leverage its structural properties to derive a novel dual formulation for $L_p$ robust Markov Decision Processes (MDPs). This formulation reveals key insights into the adversary’s strategy and leads to the \textbf{first polynomial-time robust policy evaluation algorithm} for $L_1$-normed non-rectangular robust MDPs.

Large language models (LLMs) have emerged as a dominant AI paradigm due to their exceptional text understanding and generation capabilities. However, their tendency to generate inconsistent or erroneous outputs challenges their reliability, especially in high-stakes domains requiring accuracy and trustworthiness. Existing research primarily focuses on detecting and mitigating model misbehavior in general-purpose scenarios, often overlooking the potential of integrating domain-specific knowledge. In this work, we advance misbehavior detection by incorporating domain knowledge. The core idea is to design a general specification language that enables domain experts to customize domain-specific constraints in a lightweight and intuitive manner, supporting later runtime monitoring of LLM outputs. To achieve this, we design a novel specification language ESL, and introduce a runtime verification framework RvLLM to validate LLM output against domain-specific constraints defined in ESL. RvLLM operates in two main stages: interpretation and reasoning. During interpretation, it derives interpretations of the specification based on the context, which then guide the reasoning process to identify inconsistencies. When new knowledge is derived, RvLLM issues a follow-up query to the LLM to further verify the consistency. We evaluate RvLLM on three representative tasks: violation detection against Singapore Rapid Transit Systems Act, numerical comparison, and inequality solving. Experimental results show that RvLLM effectively detects erroneous outputs across various LLMs in a lightweight and flexible manner. The results reveal that despite their impressive capabilities, LLMs remain prone to low-level errors due to a lack of formal guarantees during inference, and our framework offers a potential long-term solution by leveraging expert domain knowledge to rigorously and efficiently verify LLM outputs.

Modern tokenizers employ deterministic algorithms to map text into a single ``canonical" token sequence, yet the same string can be encoded as many non-canonical tokenizations using the language model vocabulary, including tokenizing by character. In this paper, we investigate the robustness of LMs to input encoded with non-canonical tokenizations entirely unseen during training. Surprisingly, when evaluated across 20 benchmarks, we find that instruction-tuned models retain up to 93.4\% of their original performance when given a randomly sampled tokenization, and 90.8\% with character-level tokenization. We find that overall stronger models tend to be more robust, and that robustness diminishes as the tokenization departs farther from the canonical form. Motivated by these results, we identify settings where non-canonical tokenization schemes can \textit{improve} performance, finding that character‑level segmentation improves string manipulation and code understanding tasks by up to 15\%, and right‑aligned digit grouping enhances large‑number arithmetic by over 33\%. Finally, we investigate the source of this robustness, finding that it arises in the instruction-tuning phase. We provide evidence that both base and post-trained models grasp the semantics of non-canonical tokenizations (perceiving them as containing misspellings). However, base models try to mimic the imagined mistakes and degenerate into nonsensical output, while post-trained models are committed to fluent responses. Overall, our findings suggest that models are less committed to their tokenizer than previously believed, and highlight the promise of intervening on tokenization at inference time to boost language model performance.

Existing Text Image Forgery Localization (T-IFL) methods often suffer from poor generalization due to the limited scale of real-world datasets and the distribution gap caused by synthetic data that fails to capture the complexity of real-world tampering. To tackle this issue, we propose Fourier Series-based Tampering Synthesis (FSTS), a structured and interpretable framework for synthesizing tampered text images. FSTS first collects 16,750 real-world tampering instances from five representative tampering types, using a structured pipeline that records human-performed editing traces via multi-format logs (e.g., video, PSD, and editing logs). By analyzing these collected parameters and identifying recurring behavioral patterns at both individual and population levels, we formulate a hierarchical modeling framework. Specifically, each individual tampering parameter is represented as a compact combination of basis operation–parameter configurations, while the population-level distribution is constructed by aggregating these behaviors. Since this formulation draws inspiration from the Fourier series, it enables an interpretable approximation using basis functions and their learned weights. By sampling from this modeled distribution, FSTS synthesizes diverse and realistic training data that better reflect real-world forgery traces. Extensive experiments across four evaluation protocols demonstrate that models trained with FSTS data achieve significantly improved generalization on real-world datasets. Dataset is available at \href{https://github.com/ZeqinYu/FSTS}{Project Page}.

We propose the Multi-modal Untrimmed Video Retrieval task, along with a new benchmark (MUVR) to advance video retrieval for long-video platforms. MUVR aims to retrieve untrimmed videos containing relevant segments using multi-modal queries. It has the following features: **1) Practical retrieval paradigm:** MUVR supports video-centric multi-modal queries, expressing fine-grained retrieval needs through long text descriptions, video tag prompts, and mask prompts. It adopts a one-to-many retrieval paradigm and focuses on untrimmed videos, tailored for long-video platform applications. **2) Multi-level visual correspondence:** To cover common video categories (e.g., news, travel, dance) and precisely define retrieval matching criteria, we construct multi-level visual correspondence based on core video content (e.g., news events, travel locations, dance moves) which users are interested in and want to retrieve. It covers six levels: copy, event, scene, instance, action, and others. **3) Comprehensive evaluation criteria:** We develop 3 versions of MUVR (i.e., Base, Filter, QA). MUVR-Base/Filter evaluates retrieval models, while MUVR-QA assesses MLLMs in a question-answering format. We also propose a Reranking Score to evaluate the reranking ability of MLLMs. MUVR consists of 53K untrimmed videos from the video platform Bilibili, with 1,050 multi-modal queries and 84K matches. Extensive evaluations of 3 state-of-the-art video retrieval models, 6 image-based VLMs, and 10 MLLMs are conducted. MUVR reveals the limitations of retrieval methods in processing untrimmed videos and multi-modal queries, as well as MLLMs in multi-video understanding and reranking. Our code and benchmark is available at https://github.com/debby-0527/MUVR.

Tests of conditional independence (CI) underpin a number of important problems in machine learning and statistics, from causal discovery to evaluation of predictor fairness and out-of-distribution robustness. Shah and Peters (2020) showed that, contrary to the unconditional case, no universally finite-sample valid test can ever achieve nontrivial power. While informative, this result (based on “hiding” dependence) does not seem to explain the frequent practical failures observed with popular CI tests. We investigate the Kernel-based Conditional Independence (KCI) test – of which we show the Generalized Covariance Measure underlying many recent tests is _nearly_ a special case – and identify the major factors underlying its practical behavior. We highlight the key role of errors in the conditional mean embedding estimate for the Type I error, while pointing out the importance of selecting an appropriate conditioning kernel (not recognized in previous work) as being necessary for good test power but also tending to inflate Type I error.

In dueling bandits, an agent explores and exploits choices (i.e., arms) by learning from their stochastic feedback in the form of relative preferences. Prior related studies focused on unbiased feedback. In practice, however, the feedback provided by evaluators can be biased. For example, human users are likely to provide biased evaluation towards large language models due to their heterogeneous background. In this work, we aim to minimize the regret in dueling bandits considering evaluators’ biased feedback. We begin with a benchmark case where evaluators’ bias information is known. Solving the known-bias case is nontrivial, because the bias cannot be easily decoupled from the feedback. We overcome this challenge and propose an unbiased arm performance estimator and a bias-sensitive dueling bandits algorithm. We manage to analyze the regret, dealing with the complex form of the estimator, and show that the feedback either matching or opposing the ground-truth reduces the regret. Then, we study the case where evaluators’ bias information is unknown. The associated estimator can hardly be solved in closed-form due to the non-convexity of the estimator solving problem. We address this challenge and propose an extended bias-sensitive algorithm by incorporating block coordinate descent. This algorithm is proven to achieve the same order of regret (as in the known bias case) with a bounded error. Experiments show that when compared with baselines, our algorithms reduces the regret by up to 86.9%.