NeurIPS 2025 Papers

Skill assessment from video entails rating the quality of a person’s physical performance and explaining what could be done better. Today’s models specialize for an individual sport, and suffer from the high cost and scarcity of expert-level supervision across the long tail of sports. Towards closing that gap, we explore transferable video representations for skill assessment. Our CrossTrainer approach discovers skill-attributes—such as balance, control, and hand positioning—whose meaning transcends the boundaries of any given sport, then trains a multimodal language model to generate actionable feedback for a novel video, e.g., “lift hands more to generate more power” as well as its proficiency level, e.g., early expert. We validate the new model on multiple datasets for both cross-sport (transfer) and intra-sport (in-domain) settings, where it achieves gains up to 60% relative to the state of the art. By abstracting out the shared behaviors indicative of human skill, the proposed video representation generalizes substantially better than an array of existing techniques, enriching today’s multimodal large language models.

Text-to-image (T2I) models are capable of generating visually impressive images, yet they often fail to accurately capture specific attributes in user prompts, such as the correct number of objects with the specified colors. The diversity of such errors underscores the need for a hierarchical evaluation framework that can compare prompt adherence abilities of different image generation models. Simultaneously, benchmarks of vision language models (VLMs) have not kept pace with the complexity of scenes that VLMs are used to annotate. In this work, we propose a structured methodology for jointly evaluating T2I models and VLMs by testing whether VLMs can identify 27 specific failure modes in the images generated by T2I models conditioned on challenging prompts. Our second contribution is a dataset of prompts and images generated by 5 T2I models (Flux, SD3-Medium, SD3-Large, SD3.5-Medium, SD3.5-Large) and the corresponding annotations from VLMs (Molmo, InternVL3, Pixtral) annotated by an LLM (Llama3) to test whether VLMs correctly identify the failure mode in a generated image. By analyzing failure modes on a curated set of prompts, we reveal systematic errors in attribute fidelity and object representation. Our findings suggest that current metrics are insufficient to capture these nuanced errors, highlighting the importance of targeted benchmarks for advancing generative model reliability and interpretability.

The interpretation of genomic sequences is crucial for understanding biological processes. To handle the growing volume of DNA sequence data, Genomic Foundation Models (GFMs) have been developed by adapting architectures and training paradigms from Large Language Models (LLMs). Despite their remarkable performance in DNA sequence classification tasks, there remains a lack of systematic understanding regarding the training and task-adaptation processes of GFMs. Moreover, existing GFMs cannot achieve state-of-the-art performance on both short and long-context tasks and lacks multimodal abilities. By revisiting pre-training architectures and post-training techniques, we propose **Omni-DNA**, a family of models spanning 20M to 1.1B parameters that supports sequence understanding, long-context genomic reasoning, and natural-language annotation. **Omni-DNA** establishes new state-of-the-art results on 18 of 26 evaluations drawn from Nucleotide Transformer and Genomic Benchmarks. When jointly fine-tuning on biologically related tasks, **Omni-DNA** consistently outperform existing models and demonstrate multi-tasking abilities. To enable processing of arbitrary sequence lengths, we introduce **SEQPACK**—an adaptive compression operator that packs historical tokens into a learned synopsis using a position-aware learnable sampling mechanism, enabling transformer-based models to process ultra-long sequences with minimal memory and computational requirements. Our approach demonstrates superior performance on enhancer-target interaction tasks, capturing distant regulatory interactions at the 450kbp range more effectively than existing models. Finally, we present a new dataset termed **seq2func**, enabling Omni-DNA to generate accurate and functionally meaningful interpretations of DNA sequences, unlocking new possibilities for genomic analysis and discovery.

Training vision language models (VLMs) aims to align visual representations from a vision encoder with the textual representations of a pretrained large language model (LLM). However, many VLMs exhibit reduced factual recall performance compared to their LLM backbones, raising the question of how effective multimodal fine-tuning is at extending existing mechanisms within the LLM to visual inputs. We argue that factual recall based on visual inputs requires VLMs to solve a two-hop problem: (1) forming entity representations from visual inputs, and (2) recalling associated factual knowledge based on these entity representations. By benchmarking 14 VLMs with various architectures (LLaVA, Native, Cross-Attention), sizes (7B-124B parameters), and training setups on factual recall tasks against their original LLM backbone models, we find that 11 of 14 models exhibit factual recall degradation. We select three models exhibiting high- and two models with low performance degradation, and use attribution patching, activation patching, and probing to show that degraded VLMs struggle to use the existing factual recall circuit of their LLM backbone, because they resolve the first hop too late in the computation. In contrast, high-performing VLMs resolve entity representations early enough to reuse the existing factual recall mechanism. Finally, we demonstrate two methods to recover performance: patching entity representations from the LLM backbone into the VLM, and prompting with chain-of-thought reasoning. Our results highlight that the speed of early entity resolution critically determines how effective VLMs are in using preexisting LLM mechanisms. More broadly, our work illustrates how mechanistic analysis can explain and unveil systematic failures in multimodal alignment.

Despite their recent progress, Multimodal Large Language Models (MLLMs) often struggle in knowledge-intensive tasks due to the limited and outdated parametric knowledge acquired during training. Multimodal Retrieval Augmented Generation addresses this issue by retrieving contextual knowledge from external databases, thereby enhancing MLLMs with expanded knowledge sources. However, existing MLLMs often fail to fully leverage the retrieved contextual knowledge for response generation. We examine representative MLLMs and identify two major causes, namely, attention bias toward different tokens and knowledge conflicts between parametric and contextual knowledge. To this end, we design Adaptive Logits Fusion and Attention Reallocation (ALFAR), a training-free and plug-and-play approach that improves MLLM responses by maximizing the utility of the retrieved knowledge. Specifically, ALFAR tackles the challenges from two perspectives. First, it alleviates attention bias by adaptively shifting attention from visual tokens to relevant context tokens according to query-context relevance. Second, it decouples and weights parametric and contextual knowledge at output logits, mitigating conflicts between the two types of knowledge. As a plug-and-play method, ALFAR achieves superior performance across diverse datasets without requiring additional training or external tools. Extensive experiments over multiple MLLMs and benchmarks show that ALFAR consistently outperforms the state-of-the-art by large margins. Our code and data are available at https://github.com/Lackel/ALFAR.

Large Language Models (LLMs) achieve superior performance through training-time scaling, and test-time scaling further enhances their capabilities by conducting effective reasoning during inference. However, as the scale of reasoning increases, existing test-time scaling methods suffer from accumulated historical information, which not only wastes computational resources but also interferes with effective reasoning. To address this issue, we observe that complex reasoning can be achieved by solving a series of independent and self-contained subquestions. These subquestions are essentially \textit{atomic questions}, exhibiting the memoryless property similar to Markov processes. Based on this observation, we propose Atom of Thoughts (\our), where each state transition consists of decomposing the current question into a dependency-based directed acyclic graph and contracting its subquestions, forming a simplified question that maintains answer equivalence with the original problem. This answer preservation enables the iterative \textit{decomposition-contraction} process to naturally form a meaningful Markov reasoning process. Furthermore, these atomic states can be seamlessly integrated into existing test-time scaling methods, enabling \our to serve as a plug-in enhancement for improving reasoning capabilities.

Autonomous driving represents a prominent application of artificial intelligence. Recent approaches have shifted from focusing solely on common scenarios to addressing complex, long-tail situations such as subtle human behaviors, traffic accidents, and non-compliant driving patterns. Given the demonstrated capabilities of large language models (LLMs) in understanding visual and natural language inputs and following instructions, recent methods have integrated LLMs into autonomous driving systems to enhance reasoning, interpretability, and performance across diverse scenarios. However, existing methods typically rely either on real-world data, which is suitable for industrial deployment, or on simulation data tailored to rare or hard case scenarios. Few approaches effectively integrate the complementary advantages of both data sources. To address this limitation, we propose a novel VLM-guided, end-to-end adversarial transfer framework for autonomous driving that transfers long-tail handling capabilities from simulation to real-world deployment, named CoC-VLA. The framework comprises a teacher VLM model, a student VLM model, and a discriminator. Both the teacher and student VLM models utilize a shared base architecture, termed the Chain-of-Causality Visual–Language Model (CoC VLM), which integrates temporal information via an end-to-end text adapter. This architecture supports chain-of-thought reasoning to infer complex driving logic. The teacher and student VLM models are pre-trained separately on simulated and real-world datasets. The discriminator is trained adversarially to facilitate the transfer of long-tail handling capabilities from simulated to real-world environments by the student VLM model, using a novel backpropagation strategy. Experimental results show that our method effectively bridges the gap between simulation and real-world autonomous driving, indicating a promising direction for future research.

Scientific sketches (e.g., models) offer a powerful lens into students' conceptual understanding, yet AI-powered automated assessment of such free-form, visually diverse artifacts remains a critical challenge. Existing solutions often treat sketch evaluation as either an image classification task or monolithic vision-language models, which lack interpretability, pedagogical alignment, and adaptability across cognitive levels. To address these limitations, we present SketchMind, a cognitively grounded, multi-agent framework for evaluating and improving student-drawn scientific sketches. SketchMind introduces Sketch Reasoning Graphs (SRGs), semantic graph representations that embed domain concepts and Bloom's taxonomy-based cognitive labels. The system comprises modular agents responsible for rubric parsing, sketch perception, cognitive alignment, and iterative feedback with sketch modification, enabling personalized and transparent evaluation. We evaluate SketchMind on a curated dataset of 3,575 student-generated sketches across six science assessment items with different highest order of Bloom's level that require students to draw models to explain phenomena. Compared to baseline GPT-4o performance without SRG (average accuracy: 55.6%), the model with SRG integration achieves 77.1% average accuracy (+21.4% average absolute gain). We also demonstrate that multi-agent orchestration with SRG enhances SketchMind performance, for example, SketchMind with GPT-4.1 gains an average 8.9% increase in sketch prediction accuracy, outperforming single-agent pipelines across all items. Human evaluators rated the feedback and co-created sketches generated by SketchMind with GPT-4.1, which achieved an average of 4.1 out of 5, significantly higher than those of baseline models (e.g., 2.3 for GPT-4o). Experts noted the system’s potential to meaningfully support conceptual growth through guided revision. Our code and (pending approval) dataset will be released to support reproducibility and future research in AI-driven education.

Large Vision-Language Models (LVLMs) have demonstrated impressive capabilities in tasks requiring multimodal understanding. However, recent studies indicate that LVLMs are more vulnerable than LLMs to unsafe inputs and prone to generating harmful content. Existing defense strategies primarily include fine-tuning, input sanitization, and output intervention. Although these approaches provide a certain level of protection, they tend to be resource-intensive and struggle to effectively counter sophisticated attack techniques. To tackle such issues, we propose One-head Defense (Oh Defense), a novel yet simple approach utilizing LVLMs' internal safety capabilities. Through systematic analysis of the attention mechanisms, we discover that LVLMs' safety capabilities are concentrated within specific attention heads that respond differently to safe or unsafe inputs. Further exploration reveals that a single critical attention head can effectively serve as a safety guard, providing a strong discriminative signal that amplifies the model's inherent safety capabilities. Hence, the Oh Defense requires no additional training or external modules, making it computationally efficient while effectively reactivating suppressed safety mechanisms. Extensive experiments across diverse LVLM architectures and unsafe datasets validate our approach, i.e., the Oh Defense achieves near-perfect defense success rates (> 98\%) for unsafe inputs while maintaining low false positive rates (< 5\%) for safe content. The source code is available at https://github.com/AIASLab/Oh-Defense.

Chart understanding presents a unique challenge for large vision-language models (LVLMs), as it requires the integration of sophisticated textual and visual reasoning capabilities. However, current LVLMs exhibit a notable imbalance between these skills, falling short on visual reasoning that is difficult to perform in text. We conduct a case study using a synthetic dataset solvable only through visual reasoning and show that model performance degrades significantly with increasing visual complexity, while human performance remains robust. We then introduce *ChartMuseum*, a new Chart Question Answering (QA) benchmark containing 1,162 expert-annotated questions spanning multiple reasoning types, curated from real-world charts across 184 sources, specifically built to evaluate complex visual and textual reasoning. Unlike prior chart understanding benchmarks---where frontier models perform similarly and near saturation---our benchmark exposes a substantial gap between model and human performance, while effectively differentiating model capabilities: although humans achieve 93% accuracy, the best-performing model Gemini-2.5-Pro attains only 63.0%, and the leading open-source LVLM Qwen2.5-VL-72B-Instruct achieves only 38.5%. Moreover, on questions requiring primarily visual reasoning, *all* models experience a 35%-55% performance drop from text-reasoning-heavy question performance. Lastly, our qualitative error analysis reveals specific categories of visual reasoning that are challenging for current LVLMs. Both ChartMuseum and the evaluation code are available at [https://github.com/Liyan06/ChartMuseum](https://github.com/Liyan06/ChartMuseum).

Fusing multimodal brain imaging has been a hot topic since different modalities of brain imaging can provide complementary information. However, due to the size of simultaneous recorded fMRI-EEG dataset being limited and the substantial discrepancy between hemodynamic responses of fMRI and neural oscillations of EEG, the joint modeling of fMRI and EEG images is a rarely explored area and has not yielded satisfactory results. Existing studies have also indicated that the relationships between region of interest (ROI) are not one-to-one when synchronizing fMRI and EEG. Current graph-based multimodal modeling methods overlook those information. Based on this, we propose a hypergraph based fMRI-EEG modeling framework for asynchronous fMRI-EEG data named FE-NET. To the best of our knowledge, this is the first attempt to jointly model asynchronous EEG and fMRI data as Neural ODEs based hypergraph. Extensive experiments have demonstrated that the proposed FE-NET outperforms many state-of-the-art brain imaging modeling methods. Meanwhile, compared to simultaneously recorded fMRI-EEG data, asynchronously acquired fMRI-EEG data is less costly, which demonstrates the practical applicability of our method.

Reinforcement Learning (RL) has proven to be an effective post-training strategy for enhancing reasoning in vision–language models (VLMs). Group Relative Policy Optimization (GRPO) is a recent prominent method that encourages models to generate complete reasoning traces before answering, leading to increased token usage and computational cost. Inspired by the human-like thinking process—where people skip reasoning for easy questions but think carefully when needed—we explore how to enable VLMs to first decide *when reasoning is necessary*. To realize this, we propose \ours, a two-stage training strategy: **(i)** a supervised fine-tuning (SFT) stage with a simple yet effective “**thought dropout**” operation, where reasoning traces are randomly replaced with empty thoughts. This introduces a think-or-not format that serves as a cold start for selective reasoning; **(ii)** a GRPO stage that enables the model to freely explore when to think or not, while maximizing task-aware outcome rewards. Experimental results show that \ours can *reduce the completion length by up to **90%** compared to vanilla GRPO, without sacrificing performance or even improving it*. Further evaluations across LLM (GSM8K), VLM (CLEVR, Super-CLEVR, GeoQA), and Agentic (AITZ) tasks—covering a range of reasoning difficulties under both 3B and 7B models—consistently reveal that the \textit{model progressively learns to bypass unnecessary reasoning steps as training advances}. These findings shed light on the path toward human-like reasoning patterns in RL approaches. Our code is available at https://github.com/kokolerk/TON.

Automated model discovery is the process of automatically searching and identifying the most appropriate model for a given dataset over a large combinatorial search space. Existing approaches, however, often face challenges in balancing the capture of fine-grained details with ensuring generalizability beyond training data regimes with a reasonable model complexity. In this paper, we present a multi-modal \& multi-step pipeline for effective automated model discovery. Our approach leverages two vision-language-based modules (VLM), AnalyzerVLM and EvaluatorVLM, for effective model proposal and evaluation in an agentic way. AnalyzerVLM autonomously plans and executes multi-step analyses to propose effective candidate models. EvaluatorVLM assesses the candidate models both quantitatively and perceptually, regarding the fitness for local details and the generalibility for overall trends. Our results demonstrate that our pipeline effectively discovers models that capture fine details and ensure strong generalizability. Additionally, extensive ablation studies show that both multi-modality and multi-step reasoning play crucial roles in discovering favorable models.

Video anomaly detection (VAD) is a critical yet challenging task due to the complex and diverse nature of real-world scenarios. Previous methods typically rely on domain-specific training data and manual adjustments when applying to new scenarios and unseen anomaly types, suffering from high labor costs and limited generalization. Therefore, we aim to achieve generalist VAD, \ie, automatically handle any scene and any anomaly types without training data or human involvement. In this work, we propose PANDA, an agentic AI engineer based on MLLMs. Specifically, we achieve PANDA by comprehensively devising four key capabilities: (1) self-adaptive scene-aware strategy planning, (2) goal-driven heuristic reasoning, (3) tool-augmented self-reflection, and (4) self-improving chain-of-memory. Concretely, we develop a self-adaptive scene-aware RAG mechanism, enabling PANDA to retrieve anomaly-specific knowledge for anomaly detection strategy planning. Next, we introduce a latent anomaly-guided heuristic prompt strategy to enhance reasoning precision. Furthermore, PANDA employs a progressive reflection mechanism alongside a suite of context-aware tools to iteratively refine decision-making in complex scenarios. Finally, a chain-of-memory mechanism enables PANDA to leverage historical experiences for continual performance improvement. Extensive experiments demonstrate that PANDA achieves state-of-the-art performance in multi-scenario, open-set, and complex scenario settings without training and manual involvement, validating its generalizable and robust anomaly detection capability. Code is released at https://github.com/showlab/PANDA.

Transformers have driven remarkable breakthroughs in natural language processing and computer vision, yet their standard attention mechanism still imposes $O(N^2)$ complexity, hindering scalability to longer sequences. We introduce Circular-convolutional ATtention (CAT), a Fourier-based approach that efficiently applies circular convolutions to reduce complexity without sacrificing representational power. CAT achieves $O(N \log N)$ computations, requires fewer learnable parameters by streamlining fully connected layers, and introduces no heavier operations, resulting in consistent accuracy improvements and about a 10\% speedup in naive PyTorch implementations. Based on the engineering-isomorphic transformer framework, CAT's design not only offers practical efficiency and ease of implementation, but also provides insights to guide the development of future high-performance Transformer architectures. Finally, our ablation studies highlight the key conditions underlying CAT’s success, shedding light on broader principles for scalable attention mechanisms.

Humans excel at performing complex tasks by leveraging long-term memory across temporal and spatial experiences. In contrast, current Large Language Models (LLMs) struggle to effectively plan and act in dynamic, multi-room 3D environments. We posit that part of this limitation is due to the lack of proper 3D spatial-temporal memory modeling in LLMs. To address this, we first introduce 3DMem-Bench, a comprehensive benchmark comprising over 26,000 trajectories and 2,892 embodied tasks, question-answering and captioning, designed to evaluate an agent's ability to reason over long-term memory in 3D environments. Second, we propose 3DLLM-Mem, a novel dynamic memory management and fusion model for embodied spatial-temporal reasoning and actions in LLMs. Our model uses working memory tokens, which represents current observations, as queries to selectively attend to and fuse the most useful spatial and temporal features from episodic memory, which stores past observations and interactions. Our approach allows the agent to focus on task-relevant information while maintaining memory efficiency in complex, long-horizon environments. Experimental results demonstrate that 3DLLM-Mem achieves state-of-the-art performance across various tasks, outperforming the strongest baselines by 16.5\% in success rate on 3DMem-Bench's most challenging in-the-wild embodied tasks.

What is a gait? Appearance-based gait networks consider a gait as the human shape and motion information from images. Model-based gait networks treat a gait as the human inherent structure from points. However, the considerations remain vague for humans to comprehend truly. In this work, we introduce a novel paradigm Vocabulary-Guided Gait Recognition, dubbed Gait-World, which attempts to explore gait concepts through human vocabularies with Vision-Language Models (VLMs). Despite VLMs have achieved the remarkable progress in various vision tasks, the cognitive capability regarding gait modalities remains limited. The success element in Gait-World is the proper vocabulary prompt where this paradigm carefully selects gait cycle actions as Vocabulary Base, bridging the gait and vocabulary feature spaces and further promoting human understanding for the gait. How to extract gait features? Although previous gait networks have made significant progress, learning solely from gait modalities on limited gait databases makes it difficult to learn robust gait features for practicality. Therefore, we propose the first Gait-World model, dubbed $\alpha$-Gait, which guides the gait network learning with universal vocabulary knowledge from VLMs. However, due to the heterogeneity of the modalities, directly integrating vocabulary and gait features is highly challenging as they reside in different embedding spaces. To address the issues, $\alpha$-Gait designs Vocabulary Relation Mapper and Gait Fine-grained Detector to map and establish vocabulary relations in the gait space for detecting corresponding gait features. Extensive experiments on CASIA-B, CCPG, SUSTech1K, Gait3D and GREW reveal the potential value and research directions of vocabulary information from VLMs in the gait field.

Multimodal deep learning systems are deployed in dynamic scenarios due to the robustness afforded by multiple sensing modalities. Nevertheless, they struggle with varying compute resource availability (due to multi-tenancy, device heterogeneity, etc.) and fluctuating quality of inputs (from sensor feed corruption, environmental noise, etc.). Statically provisioned multimodal systems cannot adapt when compute resources change over time, while existing dynamic networks struggle with strict compute budgets. Additionally, both systems often neglect the impact of variations in modality quality. Consequently, modalities suffering substantial corruption may needlessly consume resources better allocated towards other modalities. We propose ADMN, a layer-wise Adaptive Depth Multimodal Network capable of tackling both challenges - it adjusts the total number of active layers across all modalities to meet compute resource constraints, and continually reallocates layers across input modalities according to their modality quality. Our evaluations showcase ADMN can match the accuracy of state-of-the-art networks while reducing up to 75% of their floating-point operations.

Learning editable high-resolution scene representations for dynamic scenes is an open problem with applications across the domains from autonomous driving to creative editing - the most successful approaches today make a trade-off between editability and supporting scene complexity: neural atlases represent dynamic scenes as two deforming image layers, foreground and background, which are editable in 2D, but break down when multiple objects occlude and interact. In contrast, scene graph models make use of annotated data such as masks and bounding boxes from autonomous-driving datasets to capture complex 3D spatial relationships, but their implicit volumetric node representations are challenging to edit view-consistently. We propose Neural Atlas Graphs (NAGs), a hybrid high-resolution scene representation, where every graph node is a view-dependent neural atlas, facilitating both 2D appearance editing and 3D ordering and positioning of scene elements. Fit at test-time, NAGs achieve state-of-the-art quantitative results on the Waymo Open Dataset - by 5 dB PSNR increase compared to existing methods - and make environmental editing possible in high resolution and visual quality - creating counterfactual driving scenarios with new backgrounds and edited vehicle appearance. We find that the method also generalizes beyond driving scenes and compares favorably - by more than 7 dB in PSNR - to recent matting and video editing baselines on the DAVIS video dataset with a diverse set of human and animal-centric scenes. Project Page: https://princeton-computational-imaging.github.io/nag/