NeurIPS 2025 Papers

The rapid extension of context windows in large vision-language models has given rise to long-context vision-language models (LCVLMs), which are capable of handling hundreds of images with interleaved text tokens in a single forward pass. In this work, we introduce MMLongBench, the first benchmark covering a diverse set of long-context vision-language tasks, to evaluate LCVLMs effectively and thoroughly. MMLongBench is composed of 13,331 examples spanning five different categories of downstream tasks, such as Visual RAG and Many-Shot ICL. It also provides broad coverage of image types, including various natural and synthetic images. To assess the robustness of the models to different input lengths, all examples are delivered at five standardized input lengths (8K-128K tokens) via a cross-modal tokenization scheme that combines vision patches and text tokens. Through a thorough benchmarking of 46 closed-source and open-source LCVLMs, we provide a comprehensive analysis of the current models' vision-language long-context ability. Our results show that: i) performance on a single task is a weak proxy for overall long-context capability; ii) both closed-source and open-source models face challenges in long-context vision-language tasks, indicating substantial room for future improvement; iii) models with stronger reasoning ability tend to exhibit better long-context performance. By offering wide task coverage, various image types, and rigorous length control, MMLongBench provides the missing foundation for diagnosing and advancing the next generation of LCVLMs.

We propose **GAM-Agent**, a game-theoretic multi-agent framework for enhancing vision-language reasoning. Unlike prior single-agent or monolithic models, GAM-Agent formulates the reasoning process as a non-zero-sum game between base agents—each specializing in visual perception subtasks—and a critical agent that verifies logic consistency and factual correctness. Agents communicate via structured claims, evidence, and uncertainty estimates. The framework introduces an uncertainty-aware controller to dynamically adjust agent collaboration, triggering multi-round debates when disagreement or ambiguity is detected. This process yields more robust and interpretable predictions. Experiments on four challenging benchmarks—MMMU, MMBench, MVBench, and V*Bench—demonstrate that GAM-Agent significantly improves performance across various VLM backbones. Notably, GAM-Agent boosts the accuracy of small-to-mid scale models (e.g., Qwen2.5-VL-7B, InternVL3-14B) by 5–6\%, and still enhances strong models like GPT-4o by up to 2–3\%. Our approach is modular, scalable, and generalizable, offering a path toward reliable and explainable multi-agent multimodal reasoning.

In this paper, we introduce FedMGP, a new paradigm for personalized federated prompt learning in vision-language models (VLMs). Existing federated prompt learning (FPL) methods often rely on a single, text-only prompt representation, which leads to client-specific overfitting and unstable aggregation under heterogeneous data distributions. Toward this end, FedMGP equips each client with multiple groups of paired textual and visual prompts, enabling the model to capture diverse, fine-grained semantic and instance-level cues. A diversity loss is introduced to drive each prompt group to specialize in distinct and complementary semantic aspects, ensuring that the groups collectively cover a broader range of local characteristics.During communication, FedMGP employs a dynamic prompt aggregation strategy based on similarity-guided probabilistic sampling: each client computes the cosine similarity between its prompt groups and the global prompts from the previous round, then samples s groups via a softmax-weighted distribution. This soft selection mechanism preferentially aggregates semantically aligned knowledge while still enabling exploration of underrepresented patterns—effectively balancing the preservation of common knowledge with client-specific features. Notably, FedMGP maintains parameter efficiency by redistributing a fixed prompt capacity across multiple groups, achieving state-of-the-art performance with the lowest communication parameters (5.1k) among all federated prompt learning methods. Theoretical analysis shows that our dynamic aggregation strategy promotes robust global representation learning by reinforcing shared semantics while suppressing client-specific noise. Extensive experiments demonstrate that FedMGP consistently outperforms prior approaches in both personalization and domain generalization across diverse federated vision-language benchmarks.The code will be released on https://github.com/weihao-bo/FedMGP.git.

Test-time scaling seeks to improve the reasoning performance of large language models (LLMs) by adding computational resources. A prevalent approach within the field is *sampling-based test-time scaling methods*, which enhance reasoning by generating multiple reasoning paths for a given input during inference. However, despite its practical success, the theoretical foundations remain underexplored. In this paper, we provide the first theoretical framework for analyzing sampling-based test-time scaling methods, grounded in the perspective of confidence estimation. Based on the framework, we analyze two dominant paradigms: self-consistency and perplexity, and reveal key limitations: self-consistency suffers from high estimation error while perplexity exhibits substantial modeling error and possible degradation of the estimation error convergence. To address these limitations, we introduce RPC, a hybrid method that leverages our theoretical insights through two key components: *Perplexity Consistency* and *Reasoning Pruning*. *Perplexity Consistency* combines the strengths of self-consistency and perplexity, boosting the convergence rate of estimation error from linear to exponential while preserving model error. *Reasoning Pruning* prevents degradation by eliminating low-probability reasoning paths. Both theoretical analysis and empirical results across seven benchmark datasets demonstrate that RPC has a strong potential for reducing reasoning error. Notably, RPC achieves reasoning performance comparable to self-consistency while not only enhancing confidence reliability but also reducing sampling costs by 50%. The code and resources are available at https://wnjxyk.github.io/RPC.

Short videos are an effective tool for promoting contents and improving knowledge accessibility. While existing extractive video summarization methods struggle to produce a coherent narrative, existing abstractive methods cannot `quote' from the input videos, i.e., inserting short video clips in their outputs. In this work, we explore novel video editing models for generating shorts that feature a coherent narrative with embedded video insertions extracted from a long input video. We propose a novel retrieval-embedded generation framework that allows a large language model to quote multimodal resources while maintaining a coherent narrative. Our proposed REGen system first generates the output story script with quote placeholders using a finetuned large language model, and then uses a novel retrieval model to replace the quote placeholders by selecting a video clip that best supports the narrative from a pool of candidate quotable video clips. We examine the proposed method on the task of documentary teaser generation, where short interview insertions are commonly used to support the narrative of a documentary. Our objective evaluations show that the proposed method can effectively insert short video clips while maintaining a coherent narrative. In a subjective survey, we show that our proposed method outperforms existing abstractive and extractive approaches in terms of coherence, alignment, and realism in teaser generation.

Adversarial attacks by malicious users that threaten the safety of large language models (LLMs) can be viewed as attempts to infer a target property $T$ that is unknown when an instruction is issued, and becomes knowable only after the model's reply is observed. Examples of target properties $T$ include the binary flag that triggers an LLM's harmful response or rejection, and the degree to which information deleted by unlearning can be restored, both elicited via adversarial instructions. The LLM reveals an \emph{observable signal} $Z$ that potentially leaks hints for attacking through a response containing answer tokens, thinking process tokens, or logits. Yet the scale of information leaked remains anecdotal, leaving auditors without principled guidance and defenders blind to the transparency--risk trade-off. We fill this gap with an information-theoretic framework that computes how much information can be safely disclosed, and enables auditors to gauge how close their methods come to the fundamental limit. Treating the mutual information $I(Z;T)$ between the observation $Z$ and the target property $T$ as the leaked bits per query, we show that achieving error $\varepsilon$ requires at least $\log(1/\varepsilon)/I(Z;T)$ queries, scaling linearly with the inverse leak rate and only logarithmically with the desired accuracy. Thus, even a modest increase in disclosure collapses the attack cost from quadratic to logarithmic in terms of the desired accuracy. Experiments on seven LLMs across system-prompt leakage, jailbreak, and relearning attacks corroborate the theory: exposing answer tokens alone requires about a thousand queries; adding logits cuts this to about a hundred; and revealing the full thinking process trims it to a few dozen. Our results provide the first principled yardstick for balancing transparency and security when deploying LLMs.

In many real-world applications, deployed models encounter inputs that differ from the data seen during training. Open-world recognition ensures that such systems remain robust as ever-emerging, previously _unknown_ categories appear and must be addressed without retraining. Foundation and vision-language models are pre-trained on large and diverse datasets with the expectation of broad generalization across domains, including medical imaging. However, benchmarking these models on test sets with only a few common outlier types silently collapses the evaluation back to a closed-set problem, masking failures on rare or truly novel conditions encountered in clinical use. We therefore present NOVA, a challenging, real-life _evaluation-only_ benchmark of $\sim$900 brain MRI scans that span 281 rare pathologies and heterogeneous acquisition protocols. Each case includes rich clinical narratives and double-blinded expert bounding-box annotations. Together, these enable joint assessment of anomaly localisation, visual captioning, and diagnostic reasoning. Because NOVA is never used for training, it serves as an _extreme_ stress-test of out-of-distribution generalisation: models must bridge a distribution gap both in sample appearance and in semantic space. Baseline results with leading vision-language models (GPT-4o, Gemini 2.0 Flash, and Qwen2.5-VL-72B) reveal substantial performance drops, with approximately a 65\% gap in localisation compared to natural-image benchmarks and 40\% and 20\% gaps in captioning and reasoning, respectively, compared to resident radiologists. Therefore, NOVA establishes a testbed for advancing models that can detect, localize, and reason about truly unknown anomalies.

The Schrödinger bridge problem seeks the optimal stochastic process that connects two given probability distributions with minimal energy modification. While the Sinkhorn algorithm is widely used to solve the static optimal transport problem, a recent work (Pooladian and Niles-Weed, 2024) proposed the *Sinkhorn bridge*, which estimates Schrödinger bridges by plugging optimal transport into the time-dependent drifts of SDEs, with statistical guarantees in the one-sample estimation setting where the true source distribution is fully accessible. In this work, to further justify this method, we study the statistical performance of intermediate Sinkhorn iterations in the two-sample estimation setting, where only finite samples from both source and target distributions are available. Specifically, we establish a statistical bound on the squared total variation error of Sinkhorn bridge iterations: $\mathcal{O}(1/m+1/n + r^{2k})~(r \in (0,1))$, where $m$ and $n$ are the sample sizes from the source and target distributions, respectively, and $k$ is the number of Sinkhorn iterations. This result provides a theoretical guarantee for the finite-sample performance of the Schrödinger bridge estimator and offers practical guidance for selecting sample sizes and the number of Sinkhorn iterations. Notably, our theoretical results apply to several representative methods such as [SF]$^2$M, DSBM-IMF, BM2, and lightSB(-M) under specific settings, through the previously unnoticed connection between these estimators.

Vision Transformers have demonstrated exceptional performance across various computer vision tasks, yet their quadratic computational complexity concerning token length remains a significant challenge. To address this, token reduction methods have been widely explored. However, existing approaches often overlook the frequency characteristics of self-attention, such as rank collapsing and over-smoothing phenomenon. In this paper, we propose a frequency-aware token reduction strategy that improves computational efficiency while preserving performance by mitigating rank collapsing. Our method partitions tokens into high-frequency tokens and low-frequency tokens. high-frequency tokens are selectively preserved, while low-frequency tokens are aggregated into a compact direct current token to retain essential low-frequency components. Through extensive experiments and analysis, we demonstrate that our approach significantly improves accuracy while reducing computational overhead and mitigating rank collapsing and over smoothing. Furthermore, we analyze the previous methods, shedding light on their implicit frequency characteristics and limitations. The code is available in https://github.com/jhtwosun/frequency-aware-token-pruning.

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.

Sparse Autoencoders (SAEs) have recently gained attention as a means to improve the interpretability and steerability of Large Language Models (LLMs), both of which are essential for AI safety. In this work, we extend the application of SAEs to Vision-Language Models (VLMs), such as CLIP, and introduce a comprehensive framework for evaluating monosemanticity at the neuron-level in visual representations. To ensure that our evaluation aligns with human perception, we propose a benchmark derived from a large-scale user study. Our experimental results reveal that SAEs trained on VLMs significantly enhance the monosemanticity of individual neurons, with sparsity and wide latents being the most influential factors. Further, we demonstrate that applying SAE interventions on CLIP's vision encoder directly steers multimodal LLM outputs (e.g., LLaVA), without any modifications to the underlying language model. These findings emphasize the practicality and efficacy of SAEs as an unsupervised tool for enhancing both interpretability and control of VLMs. Code and benchmark data are available at https://github.com/ExplainableML/sae-for-vlm.

Large language model (LLM) routers improve the efficiency of multi-model systems by directing each query to the most appropriate model while leveraging the diverse strengths of heterogeneous LLMs. Most existing approaches frame routing as a classification problem based solely on the input query. While this reduces overhead by avoiding inference across all models, it overlooks valuable information that could be gleaned from potential outputs and fails to capture implicit intent or contextual nuances that often emerge only during response generation. These limitations can result in suboptimal routing decisions, particularly for complex or ambiguous queries that require deeper semantic understanding. To address this challenge, we propose Lookahead, a routing framework that "foresees" potential model outputs by predicting their latent representations and uses these predictions to guide model selection, thus enabling more informed routing without full inference. Within this framework, we implement two approaches based on causal and masked language models. Empirical evaluations across seven public benchmarks—spanning instruction following, mathematical reasoning, and code generation—show that Lookahead consistently outperforms existing routing baselines, achieving an average performance gain of 7.7\% over the state-of-the-art. Our code is available at https://github.com/huangcb01/lookahead-routing.

Foundation models demand advanced data processing for their vast, multimodal datasets. However, traditional frameworks struggle with the unique complexities of multimodal data. In response, we present Data-Juicer 2.0, a data processing system backed by 100+ data processing operators spanning text, image, video, and audio modalities, supporting more critical tasks including data analysis, synthesis, annotation, and foundation model post-training. With seamless compatibility and dedicated optimization for popular dataset hubs like Hugging Face and computing engines like Ray, it improves upon its predecessor in terms of usability, efficiency, and programmability. It features an easily accessible user interface layer that supports decoupled Python interactions, RESTful APIs, and conversational commands. Its new runtime layer offers adaptive execution across diverse scales and environments, abstracting away system complexities. Extensive empirical evaluations demonstrate Data-Juicer 2.0's remarkable performance and scalability, highlighting its capability to efficiently process TB-level data with 10k+ CPU cores. The system is publicly available and has been widely adopted in diverse research fields and real-world products such as Alibaba Cloud PAI. We actively maintain the system and share practical insights to foster research and applications of next-generation foundation models.

Multimodal Large Language Models (MLLMs) have shown impressive video content understanding capabilities but struggle with fine-grained motion comprehension. To comprehensively assess the motion understanding ability of existing MLLMs, we introduce FAVOR-Bench, which comprises 1,776 videos from both ego-centric and third-person perspectives and enables assessment through both close-ended and open-ended tasks. For close-ended evaluation, we carefully design 8,184 multiple-choice question-answer pairs spanning six distinct sub-tasks. For open-ended evaluation, we employ the GPT-assisted evaluation and develop a novel cost-efficient LLM-free assessment method, where the latter can enhance benchmarking interpretability and accessibility. Comprehensive experiments with 21 state-of-the-art MLLMs reveal significant limitations in their ability to comprehend and describe detailed temporal dynamics in video motions. To alleviate this limitation, we further build FAVOR-Train, a dataset of 17,152 videos with fine-grained motion annotations. Finetuning Qwen2.5-VL on FAVOR-Train yields consistent improvements on motion-related tasks across TVBench, MotionBench and our FAVOR-Bench. Our assessment results demonstrate that the proposed FAVOR-Bench and FAVOR-Train provide valuable tools for the community to develop more powerful video understanding models.

Chain of thought reasoning has demonstrated remarkable success in large language models, yet its adaptation to vision-language reasoning remains an open challenge with unclear best practices. Existing attempts typically employ reasoning chains at a coarse-grained level, which struggles to perform fine-grained structured reasoning and, more importantly, are difficult to evaluate the reward and quality of intermediate reasoning. In this work, we delve into chain of step reasoning for vision-language models, enabling assessing reasoning step quality accurately and leading to effective reinforcement learning and inference-time scaling with fine-grained rewards. We present a simple, effective, and fully transparent framework, including the step-level reasoning data, process reward model (PRM), and reinforcement learning training. With the proposed approaches, our models set strong baselines with consistent improvements on challenging vision-language benchmarks. More importantly, we conduct a thorough empirical analysis and ablation study, unveiling the impact of each component and several intriguing properties of inference-time scaling. We believe this paper serves as a baseline for vision-language models and offers insights into more complex multimodal reasoning. Our dataset, PRM, and code at https://github.com/baaivision/CoS.

To accelerate large language model (LLM) inference, pipeline parallelism partitions model layers into sequential stages, each assigned to a different device for concurrent execution. However, this method often suffers from pipeline bubbles caused by imbalanced computation in the tail stage. While upstream stages focus solely on layer-forward operations, the final stage must also handle post-processing tasks like sampling, introducing significant latency. This uneven workload leads to pipeline misalignment, forcing upstream stages to idle and degrading overall performance. Existing frameworks typically distribute layers evenly across stages without accounting for computational load differences. To address this, we propose DynaPipe, a dynamic layer redistribution scheme that adaptively balances computation by predicting execution latency in real time. Moreover, we introduce an asynchronous key-value (KV) cache migration coordinator to enable non-blocking layer redistribution during inference. Experiments on representative LLMs demonstrate that DynaPipe reduces average end-to-end request latency by 8% to 49% across diverse workloads, outperforming state-of-the-art pipeline parallelism systems.

Self-Taught Reasoners (STaR), synonymously known as Rejection sampling Fine-Tuning (RFT), is an integral part of the training pipeline of self-improving reasoning Language Models (LMs). The self-improving mechanism often employs random observation (data) sampling. However, this results in trained observation imbalance; inefficiently over-training on solved examples while under-training on challenging ones. In response, we introduce Adaptive STaR (AdaSTaR), a novel algorithm that rectifies this by integrating two adaptive sampling principles: (1) Adaptive Sampling for Diversity: promoting balanced training across observations, and (2) Adaptive Sampling for Curriculum: dynamically adjusting data difficulty to match the model's evolving strength. Across six benchmarks, AdaSTaR achieves best test accuracy in all instances (6/6) and reduces training FLOPs by an average of 58.6\% against an extensive list of baselines. These improvements in performance and efficiency generalize to different pre-trained LMs and larger models, paving the way for more efficient and effective self-improving LMs.

As hypothesis generation becomes increasingly automated, a new bottleneck has emerged: hypothesis assessment. Modern systems can surface thousands of statistical relationships—correlations, trends, causal links—but offer little guidance on which ones are novel, non-trivial, or worthy of expert attention. In this work, we study the complementary problem to hypothesis generation: automatic hypothesis assessment. Specifically, we ask—given a large set of statistical relationships, can we automatically assess which ones are novel and worth further exploration? We focus on correlations as they are a common entry point in exploratory data analysis that often serve as the basis for forming deeper scientific or causal hypotheses. To support automatic assessment, we propose to leverage the vast knowledge encoded in LLMs' weights to derive a prior distribution over the correlation value of a variable pair. If an LLM's prior expects the correlation value observed, then such correlation is not surprising, and vice versa. We propose the Logit-based Calibrated Prior, an LLM-elicited correlation prior that transforms the model’s raw output logits into a calibrated, continuous predictive distribution over correlation values. We evaluate the prior on a benchmark of 2,096 real-world variable pairs and it achieves a sign accuracy of 78.8%, a mean absolute error of 0.26, and 95% credible interval coverage of 89.2% in predicting Pearson correlation coefficient. It also outperforms a fine-tuned RoBERTa classifier in binary correlation prediction and achieves higher precision@K in hypothesis ranking. We further show that the prior generalizes to correlations not seen during LLM pretraining, reflecting context-sensitive reasoning rather than memorization.

Recent studies have uncovered an interesting phenomenon: unimodal foundation models tend to learn convergent representations, regardless of differences in architecture, training objectives, or data modalities. However, these representations are essentially internal abstractions of samples that characterize samples independently, leading to limited expressiveness. In this paper, we propose The Indra Representation Hypothesis, inspired by the philosophical metaphor of Indra’s Net. We argue that representations from unimodal foundation models are converging to implicitly reflect a shared relational structure underlying reality, akin to the relational ontology of Indra’s Net. We formalize this hypothesis using the V-enriched Yoneda embedding from category theory, defining the Indra representation as a relational profile of each sample with respect to others. This formulation is shown to be unique, complete, and structure-preserving under a given cost function. We instantiate the Indra representation using angular distance and evaluate it in cross-model and cross-modal scenarios involving vision, language, and audio. Extensive experiments demonstrate that Indra representations consistently enhance robustness and alignment across architectures and modalities, providing a theoretically grounded and practical framework for training-free alignment of unimodal foundation models. Our code is available at https://github.com/Jianglin954/Indra.