NeurIPS 2025 Papers

Action recognition is a fundamental task in video understanding. Existing methods typically extract unified features to process all actions in one video, which makes it challenging to model the interactions between different objects in multi-action scenarios. To alleviate this issue, we explore disentangling any specified actions from complex scenes as an effective solution. In this paper, we propose Prompt-guided Disentangled Representation for Action Recognition (ProDA), a novel framework that disentangles any specified actions from a multi-action scene. ProDA leverages Spatio-temporal Scene Graphs (SSGs) and introduces Dynamic Prompt Module (DPM) to guide a Graph Parsing Neural Network (GPNN) in generating action-specific representations. Furthermore, we design a video-adapted GPNN that aggregates information using dynamic weights. Extensive experiments on two complex video action datasets, Charades and SportsHHI, demonstrate the effectiveness of our approach against state-of-the-art methods. Our code can be found in https://github.com/iamsnaping/ProDA.git.

Predicting human eye movements during goal-directed visual search is critical for enhancing interactive AI systems. In medical imaging, such prediction can support radiologists in interpreting complex data, such as chest X-rays. Many existing methods rely on generic vision--language models and saliency-based features, which can limit their ability to capture fine-grained clinical semantics and integrate domain knowledge effectively. We present \textbf{LogitGaze-Med}, a state-of-the-art multimodal transformer framework that unifies (1) domain-specific visual encoders (e.g., CheXNet), (2) textual embeddings of diagnostic labels, and (3) semantic priors extracted via the logit-lens from an instruction-tuned medical vision--language model (LLaVA-Med). By directly predicting continuous fixation coordinates and dwell durations, our model generates clinically meaningful scanpaths. Experiments on the GazeSearch dataset and synthetic scanpaths generated from MIMIC-CXR and validated by experts demonstrate that LogitGaze-Med improves scanpath similarity metrics by 20--30\% over competitive baselines and yields over 5\% gains in downstream pathology classification when incorporating predicted fixations as additional training data.

Measuring the alignment between representations lets us understand similarities between the feature spaces of different models, such as Vision Transformers trained under diverse paradigms. However, traditional measures for representational alignment yield only scalar values that obscure how these spaces agree in terms of learned features. To address this, we combine alignment analysis with concept discovery, allowing a fine-grained breakdown of alignment into individual concepts. This approach reveals both universal concepts across models and each representation’s internal concept structure. We introduce a new definition of concepts as non-linear manifolds, hypothesizing they better capture the geometry of the feature space. A sanity check demonstrates the advantage of this manifold-based definition over linear baselines for concept-based alignment. Finally, our alignment analysis of four different ViTs shows that increased supervision tends to reduce semantic organization in learned representations.

Vision-Language-Action (VLA) models for autonomous driving show promise but falter in unstructured corner case scenarios, largely due to a scarcity of targeted benchmarks. To address this, we introduce Impromptu VLA. Our core contribution is the Impromptu VLA Dataset: over 80,000 meticulously curated video clips, distilled from over 2M source clips sourced from 8 open-source large-scale datasets. This dataset is built upon our novel taxonomy of four challenging unstructured categories and features rich, planning-oriented question-answering annotations and action trajectories. Crucially, experiments demonstrate that VLAs trained with our dataset achieve substantial performance gains on established benchmarks—improving closed-loop NeuroNCAP scores and collision rates, and reaching near state-of-the-art L2 accuracy in open-loop nuScenes trajectory prediction. Furthermore, our Q&A suite serves as an effective diagnostic, revealing clear VLM improvements in perception, prediction, and planning. Our code, data and models are available at https://github.com/ahydchh/Impromptu-VLA

Recent advances in Large Vision-Language Models (LVLMs) have showcased strong reasoning abilities across multiple modalities, achieving significant breakthroughs in various real-world applications. Despite this great success, the safety guardrail of LVLMs may not cover the unforeseen domains introduced by the visual modality. Existing studies primarily focus on eliciting LVLMs to generate harmful responses via carefully crafted image-based jailbreaks designed to bypass alignment defenses. In this study, we reveal that a safe image can be exploited to achieve the same jailbreak consequence when combined with additional safe images and prompts. This stems from two fundamental properties of LVLMs: universal reasoning capabilities and safety snowball effect. Building on these insights, we propose Safety Snowball Agent (SSA), a novel agent-based framework leveraging agents' autonomous and tool-using abilities to jailbreak LVLMs. SSA operates through two principal stages: (1) initial response generation, where tools generate or retrieve jailbreak images based on potential harmful intents, and (2) harmful snowballing, where refined subsequent prompts induce progressively harmful outputs. Our experiments demonstrate that SSA can use nearly any image to induce LVLMs to produce unsafe content, achieving high success jailbreaking rates against the latest LVLMs. Unlike prior works that exploit alignment flaws, SSA leverages the inherent properties of LVLMs, presenting a profound challenge for enforcing safety in generative multimodal systems.

Vision-Language Models (VLMs), with their strong reasoning and planning capabilities, are widely used in embodied decision-making (EDM) tasks in embodied agents, such as autonomous driving and robotic manipulation. Recent research has increasingly explored adversarial attacks on VLMs to reveal their vulnerabilities. However, these attacks either rely on overly strong assumptions, requiring full knowledge of the victim VLM, which is impractical for attacking VLM-based agents, or exhibit limited effectiveness. The latter stems from disrupting most semantic information in the image, which leads to a misalignment between the perception and the task context defined by system prompts. This inconsistency interrupts the VLM's reasoning process, resulting in invalid outputs that fail to affect interactions in the physical world. To this end, we propose a fine-grained adversarial attack framework, AdvEDM, which modifies the VLM's perception of only a few key objects while preserving the semantics of the remaining regions. This attack effectively reduces conflicts with the task context, making VLMs output valid but incorrect decisions and affecting the actions of agents, thus posing a more substantial safety threat in the physical world. We design two variants of based on this framework, AdvEDM-R and AdvEDM-A, which respectively remove the semantics of a specific object from the image and add the semantics of a new object into the image. The experimental results in both general scenarios and EDM tasks demonstrate fine-grained control and excellent attack performance.

Recognizing the physical states of objects and their transformations within videos is crucial for structured video understanding and enabling robust real-world applications, such as robotic manipulation. However, pretrained vision-language models often struggle to capture these nuanced dynamics and their temporal context, and specialized object state recognition frameworks may not generalize to unseen actions or objects. We introduce SAGE (State-Action Graph Embeddings), a novel framework that offers a unified model of physical state transitions by decomposing states into fine-grained, language-described visual concepts that are sharable across different objects and actions. SAGE initially leverages Large Language Models to construct a State-Action Graph, which is then multimodally refined using Vision-Language Models. Extensive experiments show that our method significantly outperforms baselines, generalizes effectively to unseen objects and actions in open-world settings. SAGE improves the prior state-of-the-art by as much as 14.6% on novel state recognition with less than 5% of its inference time.

Large vision-language models (LVLMs), which integrate a vision encoder (VE) with a large language model, have achieved remarkable success across various tasks. However, there are still crucial challenges in LVLMs such as object hallucination, generating descriptions of objects that are not in the input image. Here, we argue that uncertain visual tokens within the VE is a key factor that contributes to object hallucination. Our statistical analysis found that there are positive correlations between visual tokens with high epistemic uncertainty and the occurrence of hallucinations. Furthermore, we show theoretically and empirically that visual tokens in early VE layers that exhibit large representation deviations under small adversarial perturbations indicate high epistemic uncertainty. Based on these findings, we propose a simple yet effective strategy to mitigate object hallucination by modifying the VE only. Our method comprises a proxy method with adversarial perturbations for identifying uncertain visual tokens efficiently and a method to mask these uncertain visual tokens during the self-attention process in the middle layers of the VE, suppressing their influence on visual encoding and thus alleviating hallucinations. Extensive experiments show that our method significantly reduces object hallucinations in LVLMs and can synergistically work with other prior arts.

Large vision-language models (LVLMs) have emerged as foundational tools for real-world AI applications. Despite their remarkable capabilities, current LVLMs process entire images at the token level, leading to significant inefficiencies compared to human cognition, which selectively focuses on high-level vision concepts. This token-level redundancy becomes increasingly problematic for high-resolution images and long video sequences, resulting in large computational costs and limited scalability in practical applications. To address this limitation, we introduce the concept of a vision concept model, a novel paradigm that enables LVLMs to dynamically extract the most relevant vision concepts from complex inputs, based on task-specific instructions. To optimize this vision concept modeling process, we propose VCM, a self-supervised framework that leverages vision-language correlations across diverse instances. VCM is designed to learn meaningful vision concepts without the need for expensive concept-level annotations. At its core, it employs a forward-backward optimization algorithm that supports LVLMs to adjust concept granularity and spatial alignment dynamically. Experiments demonstrate that VCM remarkably reduces computational costs (e.g., achieving up to 85\% fewer FLOPs for LLaVA-1.5-7B), while maintaining strong performance across a series of vision-language tasks. The codebase is available at https://github.com/RainBowLuoCS/VCM.

Vision-language models (VLMs) as foundation models have significantly enhanced performance across a wide range of visual and textual tasks, without requiring large-scale training from scratch for downstream tasks. However, these deterministic VLMs fail to capture the inherent ambiguity and uncertainty in natural language and visual data. Recent probabilistic post-hoc adaptation methods address this by mapping deterministic embeddings onto probability distributions; however, existing approaches do not account for the asymmetric uncertainty between modalities, and the constraint that meaningful deterministic embeddings reside on a unit hypersphere, potentially leading to suboptimal performance. In this paper, we address the asymmetric uncertainty structure inherent in textual and visual data, and propose AsymVLM to build probabilistic embeddings from pre-trained VLMs on the unit hypersphere, enabling uncertainty quantification. We validate the effectiveness of the probabilistic embeddings on established benchmarks, and present comprehensive ablation studies demonstrating the inherent nature of asymmetry in the uncertainty structure of textual and visual data.

Our objective is to develop compact video representations that are sensitive to visual change over time. To measure such time-sensitivity, we introduce a new task: chiral action recognition, where one needs to distinguish between a pair of temporally opposite actions, such as “opening vs. closing a door", “approaching vs. moving away from something", “folding vs. unfolding paper", etc. Such actions (i) occur frequently in everyday life, (ii) require understanding of simple visual change over time (in object state, size, spatial position, count . . . ), and (iii) are known to be poorly represented by many video embeddings. Our goal is to build time aware video representations which offer linear separability between these chiral pairs. To that end, we propose a self-supervised adaptation recipe to inject time-sensitivity into a sequence of frozen image features. Our model is based on an auto-encoder with a latent space with inductive bias inspired by perceptual straightening. We show that this results in a compact but time-sensitive video representation for the proposed task across three datasets: Something-Something, EPIC-Kitchens, and Charade. Our method (i) outperforms much larger video models pre-trained on large-scale video datasets, and (ii) leads to an improvement in classification performance on standard benchmarks when combined with these existing models.

As the field progresses toward Artificial General Intelligence (AGI), there is a pressing need for more comprehensive and insightful evaluation frameworks that go beyond aggregate performance metrics. This paper introduces a unified rating system that jointly models the difficulty of individual test cases and the competency of AI models (or humans) across vision, language, and action domains. Unlike existing metrics that focus solely on models, our approach allows for fine-grained, difficulty-aware evaluations through competitive interactions between models and tasks, capturing both the long-tail distribution of real-world challenges and the competency gap between current models and full task mastery. We validate the generalizability and robustness of our system through extensive experiments on multiple established datasets and models across distinct AGI domains. The resulting rating distributions offer novel perspectives and interpretable insights into task difficulty, model progression, and the outstanding challenges that remain on the path to achieving full AGI task mastery. We have made our code and results publicly available at https://ss47816.github.io/AGI-Elo/.

The generalization capabilities of vision-language-action (VLA) models to unseen tasks are crucial to achieving general-purpose robotic manipulation in open-world settings. However, the cross-task generalization capabilities of existing VLA models remain significantly underexplored. To address this gap, we introduce **AGNOSTOS**, a novel simulation benchmark designed to rigorously evaluate cross-task zero-shot generalization in manipulation. AGNOSTOS comprises 23 unseen manipulation tasks for test—distinct from common training task distributions—and incorporates two levels of generalization difficulty to assess robustness. Our systematic evaluation reveals that current VLA models, despite being trained on diverse datasets, struggle to generalize effectively to these unseen tasks. To overcome this limitation, we propose **Cross-Task In-Context Manipulation (X-ICM)**, a method that conditions large language models (LLMs) on in-context demonstrations from seen tasks to predict action sequences for unseen tasks. Additionally, we introduce a **dynamics-guided sample selection** strategy that identifies relevant demonstrations by capturing cross-task dynamics. On AGNOSTOS, X-ICM significantly improves cross-task zero-shot generalization performance over leading VLAs, achieving improvements of 6.0\% over $\pi_0$ and 7.9\% over VoxPoser. We believe AGNOSTOS and X-ICM will serve as valuable tools for advancing general-purpose robotic manipulation.

Vision-language models heavily rely on visual representations, yet ensuring its efficiency remains a critical challenge. Most existing approaches focus on reducing visual tokens either at the visual encoder phase or during the LLM decoder stage. Inspired by human visual cognition, where an initial global glance precedes focused attention on semantically salient regions, we introduce Glance2Gaze, a cognitively inspired framework that mimics the human two-stage attention process. The framework consists of two key components: the Glance Fusion module, which integrates multi-layer vision transformer features with text-aware attention to generate a semantically enriched global representation, and the Gaze Compression module, which utilizes a novel query-guided mechanism to selectively compress visual tokens based on their semantic relevance. Experimental results on widely adopted benchmarks demonstrate that Glance2Gaze outperforms existing methods, achieving superior performance with equal or lower computational cost. Furthermore, it generalizes well to high-resolution and video scenarios, showcasing robust and scalable efficiency improvements in VLMs.

Recent advances in multimodal models have raised questions about whether vision-and-language models (VLMs) integrate cross-modal information in ways that reflect human cognition. One well-studied test case in this domain is the bouba-kiki effect, where humans reliably associate pseudowords like ‘bouba’ with round shapes and ‘kiki’ with jagged ones. Given the mixed evidence found in prior studies for this effect in VLMs, we present a comprehensive re-evaluation focused on two variants of CLIP, ResNet and Vision Transformer (ViT), given their centrality in many state-of-the-art VLMs. We apply two complementary methods closely modelled after human experiments: a prompt-based evaluation that uses probabilities as a measure of model preference, and we use Grad-CAM as a novel approach to interpret visual attention in shape-word matching tasks. Our findings show that these model variants do not consistently exhibit the bouba-kiki effect. While ResNet shows a preference for round shapes, overall performance across both model variants lacks the expected associations. Moreover, direct comparison with prior human data on the same task shows that the models’ responses fall markedly short of the robust, modality-integrated behaviour characteristic of human cognition. These results contribute to the ongoing debate about the extent to which VLMs truly understand cross-modal concepts, highlighting limitations in their internal representations and alignment with human intuitions.

Sequence models such as transformers require inputs to be represented as one-dimensional sequences. In vision, this typically involves flattening images using a fixed row-major (raster-scan) order. While full self-attention is permutation-equivariant, modern long-sequence transformers increasingly rely on architectural approximations that break this invariance and introduce sensitivity to patch ordering. We show that patch order significantly affects model performance in such settings, with simple alternatives like column-major or Hilbert curves yielding notable accuracy shifts. Motivated by this, we propose _REOrder_, a two-stage framework for discovering task-optimal patch orderings. First, we derive an information-theoretic prior by evaluating the compressibility of various patch sequences. Then, we learn a policy over permutations by optimizing a Plackett-Luce policy using REINFORCE. This approach enables efficient learning in a combinatorial permutation space. _REOrder_ improves top-1 accuracy over row-major ordering on ImageNet-1K by up to 3.01% and Functional Map of the World by 13.35%.

Convolutional Neural Networks (CNNs) inherently encode strong inductive biases, enabling effective generalization on small-scale datasets. In this paper, we propose integrating this inductive bias into ViTs, not through an architectural intervention but solely through initialization. The motivation here is to have a ViT that can enjoy strong CNN-like performance when data assets are small, but can still scale to ViT-like performance as the data expands. Our approach is motivated by our empirical results that random impulse filters can achieve commensurate performance to learned filters within a CNN. We improve upon current ViT initialization strategies, which typically rely on empirical heuristics such as using attention weights from pretrained models or focusing on the distribution of attention weights without enforcing structures. Empirical results demonstrate that our method significantly outperforms standard ViT initialization across numerous small and medium-scale benchmarks, including Food-101, CIFAR-10, CIFAR-100, STL-10, Flowers, and Pets, while maintaining comparative performance on large-scale datasets such as ImageNet-1K. Moreover, our initialization strategy can be easily integrated into various transformer-based architectures such as Swin Transformer and MLP-Mixer with consistent improvements in performance.

The field of embodied AI (EAI) is rapidly advancing. Unlike virtual AI, EAI systems can exist in, learn from, reason about, and act in the physical world. With recent advances in AI and hardware research and design, EAI systems are becoming increasingly capable across an expanding set of operational domains. While EAI systems can offer many benefits, they also pose significant short- and long-term risks, including physical harm, surveillance, and societal disruption. These risks require urgent attention from policymakers, as existing policies for industrial robots and autonomous vehicles are insufficient to manage the full range of concerns EAI systems present. To address this issue, this paper makes three contributions. First, we provide a taxonomy of the physical, informational, economic, and social risks EAI systems pose. Second, we analyze policies in the US, UK, and EU to assess how existing frameworks address these risks and to identify critical gaps. We conclude by offering policy recommendations for the safe and beneficial deployment of EAI systems, such as mandatory testing and certification schemes, clarified liability frameworks, and strategies to manage EAI’s potentially transformative economic and societal impacts.

Unmanned Aerial Vehicles (UAVs) are evolving into language-interactive platforms, enabling more intuitive forms of human-drone interaction. While prior works have primarily focused on high-level planning and long-horizon navigation, we shift attention to language-guided fine-grained trajectory control, where UAVs execute short-range, reactive flight behaviors in response to language instructions. We formalize this problem as the Flying-on-a-Word (Flow) task and introduce UAV imitation learning as an effective approach. In this framework, UAVs learn fine-grained control policies by mimicking expert pilot trajectories paired with atomic language instructions. To support this paradigm, we present UAV-Flow, the first real-world benchmark for language-conditioned, fine-grained UAV control. It includes a task formulation, a large-scale dataset collected in diverse environments, a deployable control framework, and a simulation suite for systematic evaluation. Our design enables UAVs to closely imitate the precise, expert-level flight trajectories of human pilots and supports direct deployment without sim-to-real gap. We conduct extensive experiments on UAV-Flow, benchmarking VLN and VLA paradigms. Results show that VLA models are superior to VLN baselines and highlight the critical role of spatial grounding in the fine-grained Flow setting.