NeurIPS 2025 论文列表

Does vision-and-language (VL) training change the linguistic representations of language models in meaningful ways? In terms of downstream task performance on text-only tasks, most results in the literature have shown marginal differences. In this work, we start from the hypothesis that the domain in which VL training could have a significant effect is lexical-conceptual knowledge, in particular its taxonomic organization. Through comparing minimal pairs of text-only LMs and their VL-trained counterparts, we first show that the VL models often outperform their text-only counterparts on a text-only question-answering task that requires taxonomic understanding of concepts mentioned in the questions. Using an array of targeted behavioral and representational analyses, we show that the LMs and VLMs do not differ significantly in terms of their taxonomic knowledge itself, but they differ in how they represent questions that contain concepts in a taxonomic relation vs. a non-taxonomic relation. This implies that the taxonomic knowledge itself does not change substantially through additional VL training, but VL training does improve the deployment of this knowledge in the context of a specific task, even when the presentation of the task is purely linguistic.

Vision-language models (VLMs) have made great strides in addressing temporal understanding tasks, which involve characterizing visual changes across a sequence of images. However, recent works have suggested that when making predictions, VLMs may rely on static feature biases, such as background or object features, rather than dynamic visual changes. Static feature biases are a type of shortcut and can contribute to systematic prediction errors on downstream tasks; as a result, identifying and characterizing error-inducing static feature biases is critical prior to real-world model deployment. Existing approaches for identifying such systematic failure modes in trained models (i) are typically designed for non-temporal settings and (ii) are challenging to evaluate in temporal settings due to the lack of quantitative evaluation frameworks. In this work, we address these challenges by introducing TRoVe, an automated approach for discovering error-inducing static feature biases learned by temporal VLMs. Given a trained VLM and an annotated validation dataset associated with a downstream classification task, TRoVe extracts candidate static features from the dataset and scores each feature by (i) the effect of the feature on classification errors as well as (ii) the extent to which the VLM relies on the feature when making predictions. In order to quantitatively evaluate TRoVe, we introduce an evaluation framework consisting of 101 trained temporal VLMs paired with ground-truth annotations for learned static feature biases. We use this framework to demonstrate that TRoVe can accurately identify error-inducing static feature biases in VLMs, achieving a 28.6% improvement over the closest baseline. Finally, we apply TRoVe to 7 off-the-shelf VLMs and 2 temporal understanding tasks, surfacing previously-unknown static feature biases and demonstrating that knowledge of learned biases can aid in improving model performance at test time. Our code is available at https://github.com/Stanford-AIMI/TRoVe.

Computer vision for animals holds great promise for wildlife research but often depends on large-scale data, while existing collection methods rely on controlled capture setups. Recent data-driven approaches show the potential of single-view, non-invasive analysis, yet current animal video datasets are limited—offering as few as 2.4K 15-frame clips and lacking key processing for animal-centric 3D/4D tasks. We introduce an automated pipeline that mines YouTube videos and processes them into object-centric clips, along with auxiliary annotations valuable for downstream tasks like pose estimation, tracking, and 3D/4D reconstruction. Using this pipeline, we amass 30K videos (2M frames)—an order of magnitude more than prior works. To demonstrate its utility, we focus on 4D quadruped animal reconstruction task. To support this task, we present Animal4D, a benchmark of 230 manually filtered sequences with 11K frames showcasing clean, diverse animal motions. We evaluate state-of-the-art model-based and model-free methods on Animal4D, finding that 2D metrics favor the former despite unrealistic 3D shapes, while the latter yields more natural reconstructions but scores lower—revealing a gap in current evaluation. To address this, we enhance a recent model-free approach with sequence-level optimization, establishing the first 4D animal reconstruction baseline. Together, our pipeline, benchmark, and baseline aim to advance large-scale, markerless 4D animal reconstruction and related tasks from in-the-wild videos. Code and datasets are available at https://github.com/briannlongzhao/Animal-Video-Processing.

Contrary to common belief, we show that gradient ascent-based unconstrained optimization methods frequently fail to perform machine unlearning, a phenomenon we attribute to the inherent statistical dependence between the forget and retain data sets. This dependence, which can manifest itself even as simple correlations, undermines the misconception that these sets can be independently manipulated during unlearning. We provide empirical and theoretical evidence showing these methods often fail precisely due to this overlooked relationship. For random forget sets, this dependence means that degrading forget set metrics (which, for a retrained model, should mirror test set metrics) inevitably harms overall test performance. Going beyond random sets, we consider logistic regression as an instructive example where a critical failure mode emerges: inter-set dependence causes gradient descent-ascent iterations to progressively diverge from the ideal retrained model. Strikingly, these methods can converge to solutions that are not only far from the retrained ideal but are potentially even further from it than the original model itself, rendering the unlearning process actively detrimental. A toy example further illustrates how this dependence can trap models in inferior local minima, inescapable via finetuning. Our findings highlight that the presence of such statistical dependencies, even when manifest only as correlations, can be sufficient for ascent-based unlearning to fail. Our theoretical insights are corroborated by experiments on complex neural networks, demonstrating that these methods do not perform as expected in practice due to this unaddressed statistical interplay.

Emotion recognition from physiological signals has substantial potential for applications in mental health and emotion-aware systems. However, the lack of standardized, large-scale evaluations across heterogeneous datasets limits progress and model generalization. We introduce FEEL (Framework for Emotion Evaluation), the first large-scale benchmarking study of emotion recognition using electrodermal activity (EDA) and photoplethysmography (PPG) signals across 19 publicly available datasets. We evaluate 16 architectures spanning traditional machine learning, deep learning, and self-supervised pretraining approaches, structured into four representative modeling paradigms. Our study includes both within-dataset and cross-dataset evaluations, analyzing generalization across variations in experimental settings, device types, and labeling strategies. Our results showed that fine-tuned contrastive signal-language pretraining (CLSP) models (71/114) achieve the highest F1 across arousal and valence classification tasks, while simpler models like Random Forests, LDA, and MLP remain competitive (36/114). Models leveraging handcrafted features (107/114) consistently outperform those trained on raw signal segments, underscoring the value of domain knowledge in low-resource, noisy settings. Further cross-dataset analyses reveal that models trained on real-life setting data generalize well to lab (F1 = 0.79) and constraint-based settings (F1 = 0.78). Similarly, models trained on expert-annotated data transfer effectively to stimulus-labeled (F1 = 0.72) and self-reported datasets (F1 = 0.76). Moreover, models trained on lab-based devices also demonstrated high transferability to both custom wearable devices (F1 = 0.81) and the Empatica E4 (F1 = 0.73), underscoring the influence of heterogeneity. Overall, FEEL provides a unified framework for benchmarking physiological emotion recognition, delivering insights to guide the development of generalizable emotion-aware technologies. Code implementation is available at https://github.com/alchemy18/FEEL. More information about FEEL can be found on our website https://alchemy18.github.io/FEEL_Benchmark/.

We study the sample complexity of Bayesian recovery for solving inverse problems with general prior, forward operator and noise distributions. We consider posterior sampling according to an approximate prior $\mathcal{P}$, and establish sufficient conditions for stable and accurate recovery with high probability. Our main result is a non-asymptotic bound that shows that the sample complexity depends on (i) the intrinsic complexity of $\mathcal{P}$, quantified by its *approximate covering number*, and (ii) concentration bounds for the forward operator and noise distributions. As a key application, we specialize to generative priors, where $\mathcal{P}$ is the pushforward of a latent distribution via a Deep Neural Network (DNN). We show that the sample complexity scales log-linearly with the latent dimension $k$, thus establishing the efficacy of DNN-based priors. Generalizing existing results on deterministic (i.e., non-Bayesian) recovery for the important problem of random sampling with an orthogonal matrix $U$, we show how the sample complexity is determined by the *coherence* of $U$ with respect to the support of $\mathcal{P}$. Hence, we establish that coherence plays a fundamental role in Bayesian recovery as well. Overall, our framework unifies and extends prior work, providing rigorous guarantees for the sample complexity of solving Bayesian inverse problems with arbitrary distributions.

In reinforcement learning, distributional off-policy evaluation (OPE) focuses on estimating the return distribution of a target policy using offline data collected under a different policy. This work focuses on extending the widely used fitted Q-evaluation---developed for expectation-based reinforcement learning---to the distributional OPE setting. We refer to this extension as fitted distributional evaluation (FDE). While only a few related approaches exist, there remains no unified framework for designing FDE methods. To fill this gap, we present a set of guiding principles for constructing theoretically grounded FDE methods. Building on these principles, we develop several new FDE methods with convergence analysis and provide theoretical justification for existing methods, even in non-tabular environments. Extensive experiments, including simulations on linear quadratic regulators and Atari games, demonstrate the superior performance of the FDE methods.

Large language models (LLMs) have recently achieved notable success in code‑generation benchmarks such as HumanEval and LiveCodeBench. However, a detailed examination reveals that these evaluation suites often comprise only a limited number of homogeneous test cases, resulting in subtle faults going undetected. This not only artificially inflates measured performance but also compromises accurate reward estimation in reinforcement learning frameworks utilizing verifiable rewards (RLVR). To address these critical shortcomings, we systematically investigate the test-case generation (TCG) task by proposing multi-dimensional metrics designed to rigorously quantify test-suite thoroughness. Furthermore, we introduce a human-LLM collaborative method (SAGA), leveraging human programming expertise with LLM reasoning capability, aimed at significantly enhancing both the coverage and the quality of generated test cases. In addition, we develop a TCGBench to facilitate the study of the TCG task. Experiments show that SAGA achieves a detection rate of 90.62\% and a verifier accuracy of 32.58\% on TCGBench. The Verifier Accuracy (Verifier Acc) of the code generation evaluation benchmark synthesized by SAGA is 10.78\% higher than that of LiveCodeBench-v6. These results demonstrate the effectiveness of our proposed method. We hope this work contributes to building a scalable foundation for reliable LLM code evaluation, further advancing RLVR in code generation, and paving the way for automated adversarial test synthesis and adaptive benchmark integration.

Integrating large language models (LLMs) into recommender systems has created new opportunities for improving recommendation quality. However, a comprehensive benchmark is needed to thoroughly evaluate and compare the recommendation capabilities of LLMs with traditional recommender systems. In this paper, we introduce \recbench{}, which systematically investigates various item representation forms (including unique identifier, text, semantic embedding, and semantic identifier) and evaluates two primary recommendation tasks, i.e., click-through rate prediction (CTR) and sequential recommendation (SeqRec). Our extensive experiments cover up to 17 large models and are conducted across five diverse datasets from fashion, news, video, books, and music domains. Our findings indicate that LLM-based recommenders outperform conventional recommenders, achieving up to a 5% AUC improvement in CTR and up to a 170% NDCG@10 improvement in SeqRec. However, these substantial performance gains come at the expense of significantly reduced inference efficiency, rendering LLMs impractical as real-time recommenders. We have released our code and data to enable other researchers to reproduce and build upon our experimental results.

Securing software supply chains is a growing challenge due to the inadequacy of existing datasets in capturing the complexity of next-gen attacks, such as multiphase malware execution, remote access activation, and dynamic payload generation. Existing datasets, which rely on metadata inspection and static code analysis, are inadequate for detecting such attacks. This creates a critical gap because these datasets do not capture what happens during and after a package is installed. To address this gap, we present QUT-DV25, a dynamic analysis dataset specifically designed to support and advance research on detecting and mitigating supply chain attacks within the Python Package Index (PyPI) ecosystem. This dataset captures install and post-install-time traces from 14,271 Python packages, of which 7,127 are malicious. The packages are executed in an isolated sandbox environment using an extended Berkeley Packet Filter (eBPF) kernel and user-level probes. It captures 36 real-time features, that includes system calls, network traffic, resource usages, directory access patterns, dependency logs, and installation behaviors, enabling the study of next-gen attack vectors. ML analysis using the QUT-DV25 dataset identified four malicious PyPI packages previously labeled as benign, each with thousands of downloads. These packages deployed covert remote access and multi-phase payloads, were reported to PyPI maintainers, and subsequently removed. This highlights the practical value of QUT-DV25, as it outperforms reactive, metadata, and static datasets, offering a robust foundation for developing and benchmarking advanced threat detection within the evolving software supply chain ecosystem.

We present the first finite-sample analysis of policy evaluation in robust average-reward Markov Decision Processes (MDPs). Prior work in this setting have established only asymptotic convergence guarantees, leaving open the question of sample complexity. In this work, we address this gap by showing that the robust Bellman operator is a contraction under a carefully constructed semi-norm, and developing a stochastic approximation framework with controlled bias. Our approach builds upon Multi-Level Monte Carlo (MLMC) techniques to estimate the robust Bellman operator efficiently. To overcome the infinite expected sample complexity inherent in standard MLMC, we introduce a truncation mechanism based on a geometric distribution, ensuring a finite expected sample complexity while maintaining a small bias that decays exponentially with the truncation level. Our method achieves the order-optimal sample complexity of $\tilde{\mathcal{O}}(\epsilon^{-2})$ for robust policy evaluation and robust average reward estimation, marking a significant advancement in robust reinforcement learning theory.

Characterizing a large language model's (LLM's) knowledge of a given question is challenging. As a result, prior work has primarily examined LLM behavior under knowledge conflicts, where the model's internal parametric memory contradicts information in the external context. However, this does not fully reflect how well the model knows the answer to the question. In this paper, we first introduce a taxonomy of five knowledge statuses based on the consistency and correctness of LLM knowledge modes. We then propose KScope, a hierarchical framework of statistical tests that progressively refines hypotheses about knowledge modes and characterizes LLM knowledge into one of these five statuses. We apply KScope to nine LLMs across four datasets and systematically establish: (1) Supporting context narrows knowledge gaps across models. (2) Context features related to difficulty, relevance, and familiarity drive successful knowledge updates. (3) LLMs exhibit similar feature preferences when partially correct or conflicted, but diverge sharply when consistently wrong. (4) Context summarization constrained by our feature analysis, together with enhanced credibility, further improves update effectiveness and generalizes across LLMs.

Benchmarks are essential for quantitatively tracking progress in AI. As AI agents become increasingly capable, researchers and practitioners have introduced agentic benchmarks to evaluate agents on complex, real-world tasks. These benchmarks typically measure agent capabilities by evaluating task outcomes via specific reward designs. However, we show that many agentic benchmarks have issues in task setup or reward design. For example, SWE-bench-Verified uses insufficient test cases, while $\tau$-bench counts empty responses as successes. Such issues can lead to under- or overestimation of agents’ performance by up to 100% in relative terms. To make agentic evaluation rigorous, we introduce the Agentic Benchmark Checklist (ABC), a set of guidelines that we synthesized from our benchmark-building experience, a survey of best practices, and previously reported issues. When applied to CVE-Bench, a benchmark with a particularly complex evaluation design, ABC reduces performance overestimation by 33%.

Hierarchical clustering is a powerful tool for exploratory data analysis, organizing data into a tree of clusterings from which a partition can be chosen. This paper generalizes these ideas by proving that, for any reasonable hierarchy, one can optimally solve any center-based clustering objective over it (such as $k$-means). Moreover, these solutions can be found exceedingly quickly and are *themselves* necessarily hierarchical. Thus, given a cluster tree, we show that one can quickly access a plethora of new, equally meaningful hierarchies. Just as in standard hierarchical clustering, one can then choose any desired partition from these new hierarchies. We conclude by verifying the utility of our proposed techniques across datasets, hierarchies, and partitioning schemes.

Multimodal Large Language Model (MLLM) relies on the powerful LLM to perform multimodal tasks, showing amazing emergent abilities in recent studies, such as writing poems based on an image. However, it is difficult for these case studies to fully reflect the performance of MLLM, lacking a comprehensive evaluation. In this paper, we fill in this blank, presenting the first comprehensive MLLM Evaluation benchmark MME. It measures both perception and cognition abilities on a total of 14 subtasks. In order to avoid data leakage that may arise from direct use of public datasets for evaluation, the annotations of instruction-answer pairs are all manually designed. The concise instruction design allows us to fairly compare MLLMs, instead of struggling in prompt engineering. Besides, with such an instruction, we can also easily carry out quantitative statistics. A total of 30 advanced MLLMs are comprehensively evaluated on our MME, which not only suggests that existing MLLMs still have a large room for improvement, but also reveals the potential directions for the subsequent model optimization. The data are released at the project page: https://github.com/BradyFU/Awesome-Multimodal-Large-Language-Models/tree/Evaluation.

This paper studies how to integrate historical control data with experimental data to enhance A/B testing, while addressing the distributional shift between historical and experimental datasets. We propose a pessimistic data integration method that combines two causal effect estimators constructed based on experimental and historical datasets. Our main idea is to conceptualize the weight function for this combination as a policy so that existing pessimistic policy learning algorithms are applicable to learn the optimal weight that minimizes the resulting weighted estimator's mean squared error. Additionally, we conduct comprehensive theoretical and empirical analyses to compare our method against various baseline estimators across five scenarios. Both our theoretical and numerical findings demonstrate that the proposed estimator achieves near-optimal performance across all scenarios.

Despite rapid progress on AI benchmarks, the real-world meaning of benchmark performance remains unclear. To quantify the capabilities of AI systems in terms of human capabilities, we propose a new metric: 50%-task-completion time horizon. This is the time humans typically take to complete tasks that AI models can complete with 50% success rate. We first timed humans with relevant domain expertise on a combination of RE-Bench, HCAST, and 66 novel shorter tasks. On these tasks, current frontier AI models such as o3 have a 50% time horizon of around 110 minutes. Furthermore, frontier AI time horizon has been doubling approximately every seven months since 2019, though the trend may have accelerated since 2024. The increase in AI models’ time horizons seems to be primarily driven by greater reliability and ability to adapt to mistakes, combined with better logical reasoning and tool use capabilities. We discuss the limitations of our results—including their degree of external validity—and the implications of increased autonomy for dangerous capabilities. If these results generalize to real-world software tasks, extrapolation of this trend predicts that within 5 years, AI systems will be capable of automating many software tasks that currently take humans a month.

Large Language Models (LLMs) have recently demonstrated strong capabilities in code-related tasks, but their robustness in code reasoning under perturbations remains underexplored. We introduce CodeCrash, a stress-testing framework with 1,279 questions from CRUXEVAL and LIVECODEBENCH, designed to evaluate reasoning reliability under structural perturbations and misleading natural language (NL) contexts. Through a systematic evaluation of 17 LLMs, we find that models often shortcut reasoning by over-relying on NL cues, leading to an average performance degradation of 23.2% in output prediction tasks. Even with Chain-of-Thought reasoning, models on average still have a 13.8% drop due to distractibility and rationalization, revealing a lack of critical reasoning capability to distinguish the actual code behaviors. While Large Reasoning Models with internal reasoning mechanisms improve robustness by fostering critical thinking, plausible yet incorrect hints can trigger pathological self-reflection, causing 2-3 times token consumption and even catastrophic cognitive dissonance in extreme cases for QwQ-32B. We refer to this phenomenon as Reasoning Collapse. CodeCrash provides a rigorous benchmark for evaluating robustness in code reasoning, guiding future research and development toward more reliable and resilient models.

Many decision-making tasks, where both accuracy and efficiency matter, still require human supervision. For example, tasks like traffic officers reviewing hour-long dashcam footage or researchers screening conference videos can benefit from concise summaries that reduce cognitive load and save time. Yet current vision-language models (VLMs) often produce verbose, redundant outputs that hinder task performance. Existing video caption evaluation depends on costly human annotations and overlooks the summaries' utility in downstream tasks. We address these gaps with $\underline{\textbf{V}}$ideo-to-text $\underline{\textbf{I}}$nformation $\underline{\textbf{B}}$ottleneck $\underline{\textbf{E}}$valuation (VIBE), an annotation-free method that scores VLM outputs using two metrics: $\textit{grounding}$ (how well the summary aligns with visual content) and $\textit{utility}$ (how informative it is for the task). VIBE selects from randomly sampled VLM outputs by ranking them according to the two scores to support effective human decision-making. Human studies on $\texttt{LearningPaper24}$, $\texttt{SUTD-TrafficQA}$, and $\texttt{LongVideoBench}$ show that summaries selected by VIBE consistently improve performance—boosting task accuracy by up to $61.23$% and reducing response time by $75.77$% compared to naive VLM summaries or raw video.