NeurIPS 2025 Papers

Large Multimodal Models (LMMs) have made significant breakthroughs with the advancement of instruction tuning. However, while existing models can understand images and videos at a holistic level, they still struggle with instance-level understanding that requires a more fine-grained comprehension and alignment. Instance-level understanding is crucial for LMMs, as it focuses on the specific elements that we are most interested in. Excitingly, existing works find that the state-of-the-art LMMs exhibit strong instance understanding capabilities when provided with explicit visual cues. Motivated by this, we proposed Inst-IT, a solution to enhance LMMs in Instance understanding via explicit visual prompt Instruction Tuning for instance guidance. Inst-IT consists of a benchmark to diagnose multimodal instance-level understanding, a large-scale instruction-tuning dataset, and a continuous instruction-tuning training paradigm to effectively enhance spatial-temporal instance understanding capabilities of existing LMMs. Experimental results show that, enhanced by Inst-IT, our models not only achieve outstanding performance on Inst-IT-Bench and other instance understanding benchmarks, but also demonstrate significant improvements across various generic image and video understanding benchmarks. This highlights that our method not only boosts instance-level understanding but also strengthens the overall capabilities of generic image and video comprehension.

Hierarchical structures of motion exist across research fields, including computer vision, graphics, and robotics, where complex dynamics typically arise from coordinated interactions among simpler motion components. Existing methods to model such dynamics typically rely on manually-defined or heuristic hierarchies with fixed motion primitives, limiting their generalizability across different tasks. In this work, we propose a general hierarchical motion modeling method that learns structured, interpretable motion relationships directly from data. Our method represents observed motions using graph-based hierarchies, explicitly decomposing global absolute motions into parent-inherited patterns and local motion residuals. We formulate hierarchy inference as a differentiable graph learning problem, where vertices represent elemental motions and directed edges capture learned parent-child dependencies through graph neural networks. We evaluate our hierarchical reconstruction approach on three examples: 1D translational motion, 2D rotational motion, and dynamic 3D scene deformation via Gaussian splatting. Experimental results show that our method reconstructs the intrinsic motion hierarchy in 1D and 2D cases, and produces more realistic and interpretable deformations compared to the baseline on dynamic 3D Gaussian splatting scenes. By providing an adaptable, data-driven hierarchical modeling paradigm, our method offers a formulation applicable to a broad range of motion-centric tasks.

Recent progress in text–video retrieval has been largely driven by contrastive learning. However, existing methods often overlook the effect of the modality gap, which causes anchor representations to undergo in-place optimization (i.e., optimization tension) that limits their alignment capacity. Moreover, noisy hard negatives further distort the semantics of anchors. To address these issues, we propose GARE, a Gap-Aware Retrieval framework that introduces a learnable, pair-specific increment $\Delta_{ij}$ between text $t_i$ and video $v_j$, redistributing gradients to relieve optimization tension and absorb noise. We derive $\Delta_{ij}$ via a multivariate first-order Taylor expansion of the InfoNCE loss under a trust-region constraint, showing that it guides updates along locally consistent descent directions. A lightweight neural module conditioned on the semantic gap couples increments across batches for structure-aware correction. Furthermore, we regularize $\Delta$ through a variational information bottleneck with relaxed compression, enhancing stability and semantic consistency. Experiments on four benchmarks demonstrate that GARE consistently improves alignment accuracy and robustness, validating the effectiveness of gap-aware tension mitigation.

Accurate video forecasting enables autonomous vehicles to anticipate hazards, robotics and surveillance systems to predict human intent, and environmental models to issue timely warnings for extreme weather events. However, existing methods remain limited: transformers rely on global attention with quadratic complexity, making them impractical for high-resolution, long-horizon video prediction, while convolutional and recurrent networks suffer from short-range receptive fields and vanishing gradients, losing key information over extended sequences. To overcome these challenges, we introduce VideoTitans, the first architecture to adapt the gradient-driven Titans memory—originally designed for language modelling to video prediction. VideoTitans integrates three core ideas: (i) a sliding-window attention core that scales linearly with sequence length and spatial resolution, (ii) an episodic memory that dynamically retains only informative tokens based on a gradient-based surprise signal, and (iii) a small set of persistent tokens encoding task-specific priors that stabilize training and enhance generalization. Extensive experiments on Moving-MNIST, Human3.6M, TrafficBJ and WeatherBench benchmarks show that VideoTitans consistently reduces computation (FLOPs) and achieves competitive visual fidelity compared to state-of-the-art recurrent, convolutional, and efficient-transformer methods. Comprehensive ablations confirm that each proposed component contributes significantly.

Large Language Models (LLMs) have demonstrated remarkable capabilities across numerous tasks, yet principled explanations for their underlying mechanisms and several phenomena, such as scaling laws, hallucinations, and related behaviors, remain elusive. In this work, we revisit the classical relationship between compression and prediction, grounded in Kolmogorov complexity and Shannon information theory, to provide deeper insights into LLM behaviors. By leveraging the Kolmogorov Structure Function and interpreting LLM compression as a two-part coding process, we offer a detailed view of how LLMs acquire and store information across increasing model and data scales -- from pervasive syntactic patterns to progressively rarer knowledge elements. Motivated by this theoretical perspective and natural assumptions inspired by Heap’s and Zipf’s laws, we introduce a simplified yet representative hierarchical data-generation framework called the Syntax-Knowledge model. Under the Bayesian setting, we show that prediction and compression within this model naturally lead to diverse learning and scaling behaviors of LLMs. In particular, our theoretical analysis offers intuitive and principled explanations for both data and model scaling laws, the dynamics of knowledge acquisition during training and fine-tuning, factual knowledge hallucinations in LLMs. The experimental results validate our theoretical predictions.

The fundamental challenge of long video understanding, e.g., question answering, lies in the extensive number of frames, making it infeasible to densely understand the local details while comprehensively digest the global contexts, especially within a limited context length. To address this problem, our insight is to process short video segments individually and combine these segment-level analyses into a final response. This intuition is noted in the well-established MapReduce principle in big data processing and is naturally compatible with inference scaling at the system level. Motivated by this, we propose MR. Video (pronounced as "mister video"), a long video understanding framework adopting the MapReduce principle. We define the standard operations of MapReduce in a long video understanding context: the Map steps conduct independent and sequence-parallel dense perception on short video segments, covering local details, while the Reduce steps comprehensively aggregate the segment-level results into an answer with global contexts. Thanks to the low cost and convenience of building video agents, we instantiate such Map and Reduce operations as an effective video agent capable of attending to local details and global contexts. Based on such abilities, we further introduce two critical yet previously under-explored long video understanding designs: (a) consistent character/object names in the captions, benefiting the reasoning of actions and stories across long horizons; (b) question intention analysis, which changes the key-frame retrieval in previous video agents to localizing the relevant information via jointly reasoning the whole video contexts and questions. Our MR. Video achieves a >7% accuracy improvement on the challenging LVBench over state-of-the-art video agents and vision-language models (VLMs) and demonstrates a clear advantage on multiple long video benchmarks, highlighting the potential of the MapReduce principle. The code is at https://github.com/ziqipang/MR-Video}{https://github.com/ziqipang/MR-Video.

Large language models are known to memorize parts of their training data, posing risk of copyright violations. To systematically examine this risk, we pretrain language models (1B/3B/8B) from scratch on 83B tokens, mixing web-scale data with public domain books used to simulate copyrighted content at controlled frequencies at lengths at least ten times longer than prior work. We thereby identified the offset effect, a phenomenon characterized by two key findings: (1) verbatim memorization is most strongly triggered by short prefixes drawn from the beginning of the context window, with memorization decreasing counterintuitively as prefix length increases; and (2) a sharp decline in verbatim recall when prefix begins offset from the initial tokens of the context window. We attribute this to positional fragility: models rely disproportionately on the earliest tokens in their context window as retrieval anchors, making them sensitive to even slight shifts. We further observe that when the model fails to retrieve memorized content, it often produces degenerated text. Leveraging these findings, we show that shifting sensitive data deeper into the context window suppresses both extractable memorization and degeneration. Our results suggest that positional offset is a critical and previously overlooked axis for evaluating memorization risks, since prior work implicitly assumed uniformity by probing only from the beginning of documents or training sequences.

Weakly Supervised Video Anomaly Detection (WSVAD) has achieved notable advancements, yet existing models remain vulnerable to adversarial attacks, limiting their reliability. Due to the inherent constraints of weak supervision—where only video-level labels are provided despite the need for frame-level predictions—traditional adversarial defense mechanisms, such as adversarial training, are not effective since video-level adversarial perturbations are typically weak and inadequate. To address this limitation, pseudo-labels generated directly from the model can enable frame-level adversarial training; however, these pseudo-labels are inherently noisy, significantly degrading performance. We therefore introduce a novel Pseudo-Anomaly Generation method called Spatiotemporal Region Distortion (SRD), which creates synthetic anomalies by applying severe augmentations to localized regions in normal videos while preserving temporal consistency. Integrating these precisely annotated synthetic anomalies with the noisy pseudo-labels substantially reduces label noise, enabling effective adversarial training. Extensive experiments demonstrate that our method significantly enhances the robustness of WSVAD models against adversarial attacks, outperforming state-of-the-art methods by an average of 71.0\% in overall AUROC performance across multiple benchmarks. The implementation and code are publicly available at [FrameShield (GitHub)](https://github.com/rohban-lab/FrameShield).

Retrieval-Augmented Generation (RAG) provides additional contextual knowledge to complement the parametric knowledge in Large Language Models (LLMs). These two knowledge interweave to enhance the accuracy and timeliness of LLM responses. However, the internal mechanisms by which LLMs utilize these knowledge remain unclear. We propose modeling the forward propagation of knowledge as an entity flow, employing this framework to trace LLMs' internal behaviors when processing mixed-source knowledge. Linear probing utilizes a trainable linear classifier to detect specific attributes in hidden layers. However, once trained, a probe cannot adapt to dynamically specified entities. To address this challenge, we construct an entity-aware probe, which introduces special tokens to mark probing targets and employs a small trainable rank-8 lora update to process these special markers. We first verify this approach through an attribution experiment, demonstrating that it can accurately detect information about ad-hoc entities from complex hidden states. Next, we trace entity flows across layers to understand how LLMs reconcile conflicting knowledge internally. Our probing results reveal that contextual and parametric knowledge are routed between tokens through distinct sets of attention heads, supporting attention competition only within knowledge types. While conflicting knowledge maintains a residual presence across layers, aligned knowledge from multiple sources gradually accumulates, with the magnitude of this accumulation directly determining its influence on final outputs.

Graph Neural Networks operate through bottom-up message-passing, fundamentally differing from human visual perception, which intuitively captures global structures first. We investigate the underappreciated potential of vision models for graph understanding, finding they achieve performance comparable to GNNs on established benchmarks while exhibiting distinctly different learning patterns. These divergent behaviors, combined with limitations of existing benchmarks that conflate domain features with topological understanding, motivate our introduction of GraphAbstract. This benchmark evaluates models' ability to perceive global graph properties as humans do: recognizing organizational archetypes, detecting symmetry, sensing connectivity strength, and identifying critical elements. Our results reveal that vision models significantly outperform GNNs on tasks requiring holistic structural understanding and maintain generalizability across varying graph scales, while GNNs struggle with global pattern abstraction and degrade with increasing graph size. This work demonstrates that vision models possess remarkable yet underutilized capabilities for graph structural understanding, particularly for problems requiring global topological awareness and scale-invariant reasoning. These findings open new avenues to leverage this underappreciated potential for developing more effective graph foundation models for tasks dominated by holistic pattern recognition.

Video large language models (video LLMs) excel at video comprehension but face significant computational inefficiency due to redundant video tokens. Existing token pruning methods offer solutions. However, approaches operating within the LLM (inner-LLM pruning), such as FastV, incur intrinsic computational overhead in shallow layers. In contrast, methods performing token pruning before the LLM (outer-LLM pruning) primarily address spatial redundancy within individual frames or limited temporal windows, neglecting the crucial global temporal dynamics and correlations across longer video sequences. This leads to sub-optimal spatio-temporal reduction and does not leverage video compressibility fully. Crucially, the synergistic potential and mutual influence of combining these strategies remain unexplored. To further reduce redundancy, we introduce HoliTom, a novel training-free holistic token merging framework. HoliTom employs outer-LLM pruning through global redundancy-aware temporal segmentation, followed by spatial-temporal merging to reduce visual tokens by over 90%, significantly alleviating the LLM's computational burden. Complementing this, we introduce a robust inner-LLM token similarity-based merging approach, designed for superior performance and compatibility with outer-LLM pruning. Evaluations demonstrate our method's promising efficiency-performance trade-off on LLaVA-OneVision-7B, reducing computational costs to 6.9% of FLOPs while maintaining 99.1% of the original performance. Furthermore, we achieve a 2.28× reduction in Time-To-First-Token (TTFT) and a 1.32× acceleration in decoding throughput, highlighting the practical benefits of our integrated pruning approach for efficient video LLMs inference.

Modern Vision-Language Models (VLMs) can solve a wide range of tasks requiring visual reasoning. In real-world scenarios, desirable properties for VLMs include fast inference and controllable generation (e.g., constraining outputs to adhere to a desired format). However, existing autoregressive (AR) VLMs like LLaVA struggle in these aspects. Discrete diffusion models (DMs) offer a promising alternative, enabling parallel decoding for faster inference and bidirectional context for controllable generation through text-infilling. While effective in language-only settings, DMs' potential for multimodal tasks is underexplored. We introduce LaViDa, a family of VLMs built on DMs. We build LaViDa by equipping DMs with a vision encoder and jointly fine-tune the combined parts for multimodal instruction following. To address challenges encountered, LaViDa incorporates novel techniques such as complementary masking for effective training, prefix KV cache for efficient inference, and timestep shifting for high-quality sampling. Experiments show that LaViDa achieves competitive or superior performance to AR VLMs on multi-modal benchmarks such as MMMU, while offering unique advantages of DMs, including flexible speed-quality tradeoff, controllability, and bidirectional reasoning. On COCO captioning, LaViDa surpasses Open-LLaVa-Next-8B by +4.1 CIDEr with 1.92x speedup. On bidirectional tasks, it achieves +59% improvement on Constrained Poem Completion. These results demonstrate LaViDa as a strong alternative to AR VLMs. Code and models is available at https://github.com/jacklishufan/LaViDa

Video Anomaly Detection (VAD) aims to locate unusual activities or behaviors within videos. Recently, offline VAD has garnered substantial research attention, which has been invigorated by the progress in large language models (LLMs) and vision-language models (VLMs), offering the potential for a more nuanced understanding of anomalies. However, online VAD has seldom received attention due to real-time constraints and computational intensity. In this paper, we introduce a novel **M**emory-based online scoring queue scheme for **T**raining-free VAD (MoniTor), to address the inherent complexities in online VAD. Specifically, MoniTor applies a streaming input to VLMs, leveraging the capabilities of pre-trained large-scale models. To capture temporal dependencies more effectively, we incorporate a novel prediction mechanism inspired by Long Short-Term Memory (LSTM) networks. This ensures the model can effectively model past states and leverage previous predictions to identify anomalous behaviors. Thereby, it better understands the current frame. Moreover, we design a scoring queue and an anomaly prior to dynamically store recent scores and cover all anomalies in the monitoring scenario, providing guidance for LLMs to distinguish between normal and abnormal behaviors over time. We evaluate MoniTor on two large datasets (i.e., UCF-Crime and XD-Violence) containing various surveillance and real-world scenarios. The results demonstrate that MoniTor outperforms state-of-the-art methods and is competitive with weakly supervised methods without training. Code is available at https://github.com/YsTvT/MoniTor.

Significant progress has been achieved using Vision Transformers (ViTs) in computer vision. However, challenges persist in modeling multi-scale spatial relationships, hindering effective integration of fine-grained local details and long-range global dependencies. To address this limitation, a Multi-Kernel Correlation-Attention Vision Transformer (MK-CAViT) grounded in the Hirschfeld-Gebelein-Rényi (HGR) theory was proposed, introducing three key innovations. A parallel multi-kernel architecture was utilized to extract multi-scale features through small, medium, and large kernels, overcoming the single-scale constraints of conventional ViTs. The cross-scale interactions were enhanced through the Fast-HGR attention mechanism, which models nonlinear dependencies and applies adaptive scaling to weigh connections and refine contextual reasoning. Additionally, a stable multi-scale fusion strategy was adopted, integrating dynamic normalization and staged learning to mitigate gradient variance, progressively fusing local and global contexts, and improving training stability. The experimental results on ImageNet, COCO, and ADE20K validated the superiority of MK-CAViT in classification, detection, and segmentation, surpassing state-of-the-art baselines in capturing complex spatial relationships while maintaining efficiency. These contributions can establish a theoretically grounded framework for visual representation learning and address the longstanding limitations of ViTs.

Understanding and reasoning over long videos pose significant challenges for large video language models (LVLMs) due to the difficulty in processing intensive video tokens beyond context window and retaining long-term sequential information. Retrieval-Augmented Generation (RAG) has demonstrated effectiveness in processing long context for Large Language Models (LLMs); however, applying RAG to long video faces challenges such as disrupted temporal dependencies and inclusion of irrelevant information that can hinder accurate reasoning. To address these limitations, we propose Vgent, a novel \textbf{graph-based retrieval-reasoning-augmented generation framework} to enhance LVLMs for long video understanding. Our approach introduces two key innovations: (i) It represents videos by structured graphs with semantic relationships across video clips preserved to improve retrieval effectiveness. (ii) It introduces an intermediate reasoning step to mitigate the reasoning limitation of LVLMs, which leverages structured verification to reduce retrieval noise and facilitate the explicit aggregation of relevant information across clips, resulting in more accurate and context-aware responses. We comprehensively evaluate our framework with various open-source LVLMs on three long-video understanding benchmarks. Our approach yielded an overall performance improvement of $3.0\%\sim 5.4\%$ over base models on MLVU, and outperformed state-of-the-art video RAG methods by $8.6\%$. Our code is publicly available at https://xiaoqian-shen.github.io/Vgent.

Multi-view 3D reconstruction remains a core challenge in computer vision. Recent methods, such as DUSt3R and its successors, directly regress pointmaps from image pairs without relying on known scene geometry or camera parameters. However, the performance of these models is constrained by the diversity and scale of available training data. In this work, we introduce Puzzles, a data augmentation strategy that synthesizes an unbounded volume of high-quality, posed video-depth data from just a single image or video clip. By simulating diverse camera trajectories and realistic scene geometry through targeted image transformations, Puzzles significantly enhances data variety. Extensive experiments show that integrating Puzzles into existing video‑based 3D reconstruction pipelines consistently boosts performance, all without modifying the underlying network architecture. Notably, models trained on only 10% of the original data, augmented with Puzzles, achieve accuracy comparable to those trained on the full dataset.

Vision-language models are integral to computer vision research, yet many high-performing models remain closed-source, obscuring their data, design and training recipe. The research community has responded by using distillation from black-box models to label training data, achieving strong benchmark results, at the cost of measurable scientific progress. However, without knowing the details of the teacher model and its data sources, scientific progress remains difficult to measure. In this paper, we study building a Perception Language Model (PLM) in a fully open and reproducible framework for transparent research in image and video understanding. We analyze standard training pipelines without distillation from proprietary models and explore large-scale synthetic data to identify critical data gaps, particularly in detailed video understanding. To bridge these gaps, we release 2.8M human-labeled instances of fine-grained video question-answer pairs and spatio-temporally grounded video captions. Additionally, we introduce PLM–VideoBench, a suite for evaluating challenging video understanding tasks focusing on the ability to reason about ''what'', ''where'', ''when'', and ''how'' of a video. We make our work fully reproducible by providing data, training recipes, code & models.

State Space Models (SSMs) have emerged as promising alternatives to attention mechanisms, with the Mamba architecture demonstrating impressive performance and linear complexity for processing long sequences. However, the fundamental differences between Mamba and Transformer architectures remain incompletely understood. In this work, we use carefully designed synthetic tasks to reveal Mamba's inherent limitations. Through experiments, we identify that Mamba's nonlinear convolution introduces an asymmetry bias that significantly impairs its ability to recognize symmetrical patterns and relationships. Using composite function and inverse sequence matching tasks, we demonstrate that Mamba strongly favors compositional solutions over symmetrical ones and struggles with tasks requiring the matching of reversed sequences. We show these limitations stem not from the SSM module itself but from the nonlinear convolution preceding it, which fuses token information asymmetrically. These insights provide a new understanding of Mamba's constraints and suggest concrete architectural improvements for future sequence models.

Despite significant progress in Video Large Language Models (Video-LLMs) for offline video understanding, existing online Video-LLMs typically struggle to simultaneously process continuous frame-by-frame inputs and determine optimal response timing, often compromising real-time responsiveness and narrative coherence. To address these limitations, we introduce LiveStar, a pioneering live streaming assistant that achieves always-on proactive responses through adaptive streaming decoding. Specifically, LiveStar incorporates: (1) a training strategy enabling incremental video-language alignment for variable-length video streams, preserving temporal consistency across dynamically evolving frame sequences; (2) a response-silence decoding framework that determines optimal proactive response timing via a single forward pass verification; (3) memory-aware acceleration via peak-end memory compression for online inference on 10+ minute videos, combined with streaming key-value cache to achieve 1.53× faster inference. We also construct an OmniStar dataset, a comprehensive dataset for training and benchmarking that encompasses 15 diverse real-world scenarios and 5 evaluation tasks for online video understanding. Extensive experiments across three benchmarks demonstrate LiveStar's state-of-the-art performance, achieving an average 19.5\% improvement in semantic correctness with 18.1\% reduced timing difference compared to existing online Video-LLMs, while improving FPS by 12.0\% across all five OmniStar tasks. Our model and dataset can be accessed at https://github.com/yzy-bupt/LiveStar.