NeurIPS 2025 Papers

Large Language Models' (LLMs) programming capabilities enable their participation in \textit{open-source games}: a game-theoretic setting in which players submit computer programs in lieu of actions. These programs offer numerous advantages, including interpretability, inter-agent transparency, and formal verifiability; additionally, they enable \textit{program equilibria}, solutions that leverage the transparency of code and are inaccessible within normal-form settings. We evaluate the capabilities of leading open- and closed-weight LLMs to predict and classify program strategies and evaluate features of the approximate program equilibria reached by LLM agents in dyadic and evolutionary settings. We identify the emergence of payoff-maximizing, cooperative, and deceptive strategies, characterize the adaptation of mechanisms within these programs over repeated open-source games, and analyze their comparative evolutionary fitness. We find that open-source games serve as a viable environment to study and steer the emergence of cooperative strategy in multi-agent dilemmas.

Selection bias poses a widely recognized challenge for unbiased evaluation and learning in many industrial scenarios. For example, in recommender systems, it arises from the users' selective interactions with items. Recently, doubly robust and its variants have been widely studied to achieve debiased learning of prediction models, however, all of them consider a simple exact matching scenario, i.e., the units (such as user-item pairs in a recommender system) are the same between the training and test sets. In practice, there may be limited or even no overlap in units between the training and test. In this paper, we consider a more practical scenario: the joint distribution of the feature and rating is the same in the training and test sets. Theoretical analysis shows that the previous DR estimator is biased even if the imputed errors and learned propensities are correct in this scenario. In addition, we propose a novel super-population doubly robust estimator (SuperDR), which can achieve a more accurate estimation and desirable generalization error bound compared to the existing DR estimators, and extend the joint learning algorithm for training the prediction and imputation models. We conduct extensive experiments on three real-world datasets, including a large-scale industrial dataset, to show the effectiveness of our method. The code is available at https://github.com/ChunyuanZheng/neurips-25-SuperDR.

Token-based video representation has emerged as a promising approach for enabling large language models (LLMs) to interpret video content. However, existing token reduction techniques, such as pruning and merging, often disrupt essential positional embeddings and rely on continuous visual tokens sampled from nearby pixels with similar spatial–temporal locations. By removing only a small fraction of tokens, these methods still produce relatively lengthy continuous sequences, which falls short of the extreme compression required to balance computational efficiency and token count in video LLMs. In this paper, we introduce the novel task of **Extreme Short Token Reduction**, which aims to represent entire videos using a minimal set of discrete tokens. We propose **VQToken**, a neural discrete token representation framework that (i) applies adaptive vector quantization to continuous ViT embeddings to learn a compact codebook and (ii) preserves spatial–temporal positions via a token hash function by assigning each grid-level token to its nearest codebook entry. On the Extreme Short Token Reduction task, our VQToken compresses sequences to just **0.07\%** of their original length while incurring only a **0.66\%** drop in accuracy on NextQA-MC benchmark. It also achieves comparable performance on ActNet-QA, Long Video Bench, and VideoMME. We further introduce the **Token Information Density** (**TokDense**) metric and formalize fixed-length and adaptive-length subtasks, achieving state-of-the-art results in both settings. Our approach dramatically lowers theoretical complexity, increases information density, way fewer tokens counts, and enables efficient video large language models in resource-constrained environments.

Diffusion models have achieved impressive success in high-fidelity image generation but suffer from slow sampling due to their inherently iterative denoising process. While recent one-step methods accelerate inference by learning direct noise-to-image mappings, they sacrifice the interpretability and fine-grained control intrinsic to diffusion dynamics, key advantages that enable applications like editable generation. To resolve this dichotomy, we introduce **Hierarchical Koopman Diffusion**, a novel framework that achieves both one-step sampling and interpretable generative trajectories. Grounded in Koopman operator theory, our method lifts the nonlinear diffusion dynamics into a latent space where evolution is governed by globally linear operators, enabling closed-form trajectory solutions. This formulation not only eliminates iterative sampling but also provides full access to intermediate states, allowing manual intervention during generation. To model the multi-scale nature of images, we design a hierarchical architecture that disentangles generative dynamics across spatial resolutions via scale-specific Koopman subspaces, capturing coarse-to-fine details systematically. We empirically show that the Hierarchical Koopman Diffusion not only achieves competitive one-step generation performance but also provides a principled mechanism for interpreting and manipulating the generative process through spectral analysis. Our framework bridges the gap between fast sampling and interpretability in diffusion models, paving the way for explainable image synthesis in generative modeling.

Unsupervised domain adaptive hashing has emerged as a promising approach for efficient and memory-friendly cross-domain retrieval. It leverages the model learned on labeled source domains to generate compact binary codes for unlabeled target domain samples, ensuring that semantically similar samples are mapped to nearby points in the Hamming space. Existing methods typically apply domain adaptation techniques to the feature space or the Hamming space, especially pseudo-labeling and feature alignment. However, the inherent noise of pseudo-labels and the insufficient exploration of complementary knowledge across spaces hinder the ability of the adapted model. To address these challenges, we propose a Vision-language model assisted Pseudo-labeling and Dual Space adaptation (VPDS) method. Motivated by the strong zero-shot generalization capabilities of pre-trained vision-language models (VLMs), VPDS leverages VLMs to calibrate pseudo-labels, thereby mitigating pseudo-label bias. Furthermore, to simultaneously utilize the semantic richness of high-dimensional feature space and preserve discriminative efficiency of low-dimensional Hamming space, we introduce a dual space adaptation approach that performs independent alignment within each space. Extensive experiments on three benchmark datasets demonstrate that VPDS consistently outperforms existing methods in both cross-domain and single-domain retrieval tasks, highlighting its effectiveness and superiority.

Segmentation models can recognize a pre-defined set of objects in images. However, segmentation models capable of "reasoning" over complex user queries that implicitly refer to multiple objects of interest remain underexplored, especially in the geospatial domain. Recent advances in "reasoning segmentation"---generating segmentation masks from complex, implicit query text---demonstrate the potential of vision-language models (VLMs) to reason across an open domain of objects. Yet, our experiments reveal that these models struggle when applied to the unique challenges of remote-sensing imagery. To address this gap, we introduce a new dataset which consists of: GRES, a curated geospatial reasoning-segmentation dataset with 27,615 annotations across 9,205 images, and PreGRES, a collection of existing datasets to make up a large-scale multimodal pretraining corpus with over 1M question-answer pairs across 119,279 images. We propose an initial benchmark model, LISAt, a VLM for geospatial analysis that can describe complex remote-sensing scenes, answer detailed queries, and segment objects based on natural-language prompts. LISAt establishes a strong initial geospatial benchmark, outperforming prior foundation models such as RS-GPT4V by 10.04\% (BLEU-4) on visual description tasks and surpassing open-domain models on geospatial reasoning segmentation by 143.36\% (gIoU). Our model, dataset, and code are available on our project page: https://lisat-bair.github.io/LISAt/.

Artificial intelligence (AI) has played a transformative role in chemical research, greatly facilitating the prediction of small molecule properties, simulation of catalytic processes, and material design. These advances are driven by increases in computing power, open source machine learning frameworks, and extensive chemical datasets. However, a persistent challenge is the limited amount of high-quality real-world data, while models calculated based on large amounts of theoretical data are often costly and difficult to deploy, which hinders the applicability of AI models in real-world scenarios. In this study, we enhance the prediction of solute-solvent properties by proposing a novel sample selection method: the iterative core subset extraction (CSIE) framework. CSIE iteratively updates the core sample subset based on information gain to remove redundant features in theoretical data and optimize the performance of the model on real chemical datasets. Furthermore, we introduce an asymmetric molecular interaction graph neural network (AMGNN) that combines positional information and bidirectional edge connections to simulate real-world chemical reaction scenarios to better capture solute-solvent interactions. Experimental results show that our method can accurately extract the core subset and improve the prediction accuracy.

Large generative models are increasingly trained on synthetic data from earlier generations, raising concerns about *model collapse*, a progressive performance decline consistently observed in empirical studies. However, theoretical understanding of recursive training dynamics and their failure modes remains limited. In this work, we theoretically show that recursive training inherently leads to exponential error growth unless mitigated by sufficient real data. Addressing the growing scarcity of real data, we introduce a self-verification mechanism enabling models to filter their outputs based on internal confidence scores without external validation. Through rigorous analysis, we derive finite-sample error bounds demonstrating that self-verification alone can prevent collapse, even in fully synthetic training regimes. Our theoretical framework extends to large language models (LLMs), characterizing the conditions under which recursive training can maintain stability without performance degradation.

Generating diverse responses from large language models (LLMs) is crucial for applications such as adversarial testing, search, and synthetic data generation, where diversity provides distinct answers across generations. Previous approaches rely solely on increasing the temperature, sacrificing quality. Furthermore, the model's next-token probabilities may not be representative of the true answer distribution. To combat these challenges, we propose SimpleStrat, an alternative that uses the language model itself to partition the solution space into strata from which to sample. To measure resampling diversity, we introduce CoverageQA, a dataset of underspecified questions with multiple equally plausible answers. We propose measuring resampling diversity as the KL Divergence between the output distribution and the uniform distribution over valid ground truth answers and use recall as an alternative when assessing proprietary models. On CoverageQA, SimpleStrat improves diversity across all temperatures, showing orthogonal benefits. Quantifiably, we achieve as much as 2X better recall when applied to GPT-4o, and an average reduction in KL divergence by 0.36 when applied to Llama 3. Furthermore, we show that SimpleStrat achieves more resampling diversity at temperature T=0 than scaling temperature to T=1 on creative writing, an open-ended domain. Implementation and dataset available at https://github.com/jwong8314/simplestrat.

Autonomous agentic AI systems powered by vision-language models (VLMs) are rapidly advancing toward real-world deployment, yet their cross-modal reasoning capabilities introduce new attack surfaces for adversarial manipulation that exploit semantic reasoning across modalities. Existing adversarial attacks typically rely on visible pixel perturbations or require privileged model or environment access, making them impractical for stealthy, real-world exploitation. We introduce TRAP, a novel generative adversarial framework that manipulates the agent’s decision-making using diffusion-based semantic injections into the vision-language embedding space. Our method combines negative prompt–based degradation with positive semantic optimization, guided by a Siamese semantic network and layout-aware spatial masking. Without requiring access to model internals, TRAP produces visually natural images yet induces consistent selection biases in agentic AI systems. We evaluate TRAP on the Microsoft Common Objects in Context (COCO) dataset, building multi-candidate decision scenarios. Across these scenarios, TRAP consistently induces decision-level preference redirection on leading models, including LLaVA-34B, Gemma3, GPT-4o, and Mistral-3.2, significantly outperforming existing baselines such as SPSA, Bandit, and standard diffusion approaches. These findings expose a critical, generalized vulnerability: autonomous agents can be consistently misled through visually subtle, semantically-guided cross-modal manipulations. Overall, our results show the need for defense strategies beyond pixel-level robustness to address semantic vulnerabilities in cross-modal decision-making. The code for TRAP is accessible on GitHub at https://github.com/uiuc-focal-lab/TRAP.

Gene expression estimation from pathology images has the potential to reduce the RNA sequencing cost. Point-wise loss functions have been widely used to minimize the discrepancy between predicted and absolute gene expression values. However, due to the complexity of the sequencing techniques and intrinsic variability across cells, the observed gene expression contains stochastic noise and batch effects, and estimating the absolute expression values accurately remains a significant challenge. To mitigate this, we propose a novel objective of learning relative expression patterns rather than absolute levels. We assume that the relative expression levels of genes exhibit consistent patterns across independent experiments, even when absolute expression values are affected by batch effects and stochastic noise in tissue samples. Based on the assumption, we model the relation and propose a novel loss function called STRank that is robust to noise and batch effects. Experiments using synthetic datasets and real datasets demonstrate the effectiveness of the proposed method. The code is available at https://github.com/naivete5656/STRank.

Conditional image generative models hold considerable promise to produce infinite amounts of synthetic training data. Yet, recent progress in generation quality has come at the expense of generation diversity, limiting the utility of these models as a source of synthetic training data. Although guidance-based approaches have been introduced to improve the utility of generated data by focusing on quality or diversity, the (implicit or explicit) utility functions oftentimes disregard the potential distribution shift between synthetic and real data. In this work, we introduce Chamfer Guidance: a training-free guidance approach which leverages a handful of real exemplar images to characterize the quality and diversity of synthetic data. We show that by leveraging the proposed Chamfer Guidance, we can boost the diversity of the generations w.r.t. a dataset of real images while maintaining or improving the generation quality on ImageNet-1k and standard geo-diversity benchmarks. Our approach achieves state-of-the-art few-shot performance with as little as 2 exemplar real images, obtaining 96.4% in terms of precision, and 86.4% in terms of distributional coverage, which increase to 97.5% and 92.7%, respectively, when using 32 real images. We showcase the benefits of the Chamfer Guidance generation by training downstream image classifiers on synthetic data, achieving accuracy boost of up to 15% for in-distribution over the baselines, and up to 16\% in out-of-distribution. Furthermore, our approach does not require using the unconditional model, and thus obtains a 31\% reduction in FLOPs w.r.t. classifier-free-guidance-based approaches at sampling time.

We introduce Eagle2.5, a frontier vision-language model (VLM) for long-context multimodal learning. Our work addresses the challenges in long video comprehension and high-resolution image understanding, introducing a generalist framework for both tasks. The proposed training framework incorporates Automatic Degrade Sampling and Image Area Preservation, two techniques that preserve contextual integrity and visual details. The framework also includes numerous efficiency optimizations in the pipeline for long-context data training. Finally, we propose Eagle-Video-110K, a novel dataset that integrates both story-level and clip-level annotations, facilitating long-video understanding. Eagle2.5 demonstrates substantial improvements on long-context multimodal benchmarks, providing a robust solution to the limitations of existing VLMs. Notably, our best model Eagle2.5-8B achieves 72.4\% on Video-MME with 512 input frames, matching the results of top-tier commercial model such as GPT-4o and large-scale open-source models like Qwen2.5-VL-72B and InternVL2.5-78B.

We introduce the Deformable Gaussian Splats Large Reconstruction Model (DGS-LRM), the first feed-forward method predicting deformable 3D Gaussian splats from a monocular posed video of any dynamic scene. Feed-forward scene reconstruction has gained significant attention for its ability to rapidly create digital replicas of real-world environments. However, most existing models are limited to static scenes and fail to reconstruct the motion of moving objects. Developing a feed-forward model for dynamic scene reconstruction poses significant challenges, including the scarcity of training data and the need for appropriate 3D representations and training paradigms. To address these challenges, we introduce several key technical contributions: an enhanced large-scale synthetic dataset with ground-truth multi-view videos and dense 3D scene flow supervision; a per-pixel deformable 3D Gaussian representation that is easy to learn, supports high-quality dynamic view synthesis, and enables long-range 3D tracking; and a large transformer network that achieves real-time, generalizable dynamic scene reconstruction. Extensive qualitative and quantitative experiments demonstrate that DGS-LRM achieves dynamic scene reconstruction quality comparable to optimization-based methods, while significantly outperforming the state-of-the-art predictive dynamic reconstruction method on real-world examples. Its predicted physically grounded 3D deformation is accurate and can be readily adapted for long-range 3D tracking tasks, achieving performance on par with state-of-the-art monocular video 3D tracking methods.

The study of multimodality has garnered significant interest in fields where analyzing interactions among multiple information sources can enhance predictive modeling, data fusion, and interpretability. Partial information decomposition (PID) has emerged as a useful information-theoretic framework to quantify the degree to which individual modalities independently, redundantly, or synergistically convey information about a target variable. However, existing PID methods depend on optimizing over a joint distribution constrained by estimated pairwise probability distributions, which are costly and inaccurate for continuous and high-dimensional modalities. Our first key insight is that the problem can be solved efficiently when the pairwise distributions are multivariate Gaussians, and we refer to this problem as Gaussian PID (GPID). We propose a new gradient-based algorithm that substantially enhances computational efficiency for GPID based on an alternative formulation of the underlying optimization problem. To generalize the applicability to non-Gaussian data, we learn information-preserving encoders to transform random variables of arbitrary input distributions into pairwise Gaussian random variables. Along the way, we resolved an open problem regarding the optimality of joint Gaussian solutions for GPID. Empirical validation on diverse synthetic examples demonstrates that our proposed method provides more accurate and efficient PID estimates than existing baselines. We further evaluate on a series of large-scale multimodal benchmarks to show its utility in real-world applications of quantifying PID in multimodal datasets and selecting high-performing models.

Audio is a critical component of multimodal perception, and any truly intelligent system must demonstrate a wide range of auditory capabilities. These capabilities include transcription, classification, retrieval, reasoning, segmentation, clustering, reranking, and reconstruction. Fundamentally, each task involves transforming a raw audio signal into a meaningful 'embedding'—be it a single vector, a sequence of continuous or discrete representations, or another structured form—which then serves as the basis for generating the task's final response. To accelerate progress towards robust machine auditory intelligence, we present the Massive Sound Embedding Benchmark (MSEB): an extensible framework designed to evaluate the auditory components of any multimodal system. In its first release, MSEB offers a comprehensive suite of eight core tasks, with more planned for the future, supported by diverse datasets, including the new, large-scale Simple Voice Questions (SVQ) dataset. Our initial experiments establish clear performance headrooms, highlighting the significant opportunity to improve real-world multimodal experiences where audio is a core signal. We encourage the research community to use MSEB to assess their algorithms and contribute to its growth. The library is publicly hosted at https://github.com/google-research/mseb.

Humans are remarkably adept at collaboration, able to infer the strengths and weaknesses of new partners in order to work successfully towards shared goals. To build AI systems with this capability, we must first understand its building blocks: does such flexibility require explicit, dedicated mechanisms for modelling others—or can it emerge spontaneously from the pressures of open-ended cooperative interaction? To investigate this question, we train simple model-free RNN agents to collaborate with a population of diverse partners. Using the 'Overcooked-AI' environment, we collect data from thousands of collaborative teams, and analyse agents' internal hidden states. Despite a lack of additional architectural features, inductive biases, or auxiliary objectives, the agents nevertheless develop structured internal representations of their partners' task abilities, enabling rapid adaptation and generalisation to novel collaborators. We investigated these internal models through probing techniques, and large-scale behavioural analysis. Notably, we find that structured partner modelling emerges when agents can influence partner behaviour by controlling task allocation. Our results show that partner modelling can arise spontaneously in model-free agents—but only under environmental conditions that impose the right kind of social pressure.

Recent advances in retrieval-based in-context learning (ICL) train the retriever using a classification objective, which categorizes in-context examples (ICEs) into the most useful and the rest based on absolute scores. However, during inference, ICEs are retrieved by score ranking rather than classification — The classification training objective deviates from this test scenario. Hence, in this paper, we propose a novel algorithm that trains a retrieval model by ranking formulation, where the preference rankings between ICEs are given by comparing the likelihood of the LLM generating the correct answer conditioned on each exemplar. By learning to rank, we motivate the retriever to automatically learn diverse rationales why specific examples are more useful for ICL decisions. This addresses the issue that classification models poorly capture broader utility. Experimental results demonstrate the top-1 performance of our proposal across 9 NLP tasks, with ablation studies and case studies further validating the effectiveness of our design. *The code can be found in: https://github.com/2022neo/SeDPO_NIPS25*

We investigate the robustness of Large Language Models (LLMs) to structural interventions by deleting and swapping adjacent layers during inference. Surprisingly, models retain 72–95\% of their original top-1 prediction accuracy without any fine-tuning. We find that performance degradation is not uniform across layers: interventions to the early and final layers cause the most degradation, while the model is remarkably robust to dropping middle layers. This pattern of localized sensitivity motivates our hypothesis of four stages of inference, observed across diverse model families and sizes: (1) detokenization, where local context is integrated to lift raw token embeddings into higher-level representations; (2) feature engineering, where task- and entity-specific features are iteratively refined; (3) prediction ensembling, where hidden states are aggregated into plausible next-token predictions; and (4) residual calibration, where irrelevant features are suppressed to finalize the output distribution. Synthesizing behavioral and mechanistic evidence, we provide a hypothesis for interpreting depth-dependent computations in LLMs.