CVPR 2026 论文列表

Multi-modal Retrieval-Augmented Generation (MMRAG) has emerged as a powerful paradigm for enhancing Multimodal Large Language Models (MLLMs) in knowledge-intensive question answering by integrating external visual, textual, and structural knowledge. However, existing MMRAG frameworks suffer from critical limitations, including noisy and irrelevant retrieval, cross-modal semantic misalignment, lack of adaptive reasoning, and incoherent generation across local and global contexts. We introduce CogniVerse, a novel MMRAG framework that addresses these challenges through a cognitive-inspired, mathematically rigorous approach. Drawing from human-like reasoning, CogniVerse integrates three synergistic components: (1) a Cognitive Reflection Module (CRM) that dynamically assesses retrieval necessity and filters relevant multi-modal content, reducing noise and computational overhead; (2) a Multi-modal Retrieval Module that aligns embeddings in a Riemannian manifold using information geometry and refines knowledge graphs via spectral graph theory, ensuring precise and coherent retrieval; and (3) a Hierarchical Generation Module that employs an optimal transport-based loss to balance token-level accuracy and global semantic coherence. Grounded in advanced theoretical frameworks, including convergence guarantees for geometric alignment and spectral optimization, CogniVerse achieves robust cross-modal integration and adaptive knowledge utilization. Extensive experiments demonstrate that CogniVerse significantly outperforms state-of-the-art systems in both accuracy and coherence, while reducing retrieval latency.

We present EvoID, a novel framework that reformulates Identity-Preserving Video Generation as a self-evolving process through Reinforcement Learning. Moving beyond the static paradigm of imitation learning, EvoID enables a generative model to actively learn and optimize the complex trade-offs between identity fidelity, motion naturalness, and temporal coherence. At the heart of our EvoID is a dynamic, dual-path reward mechanism, which acts as an intrinsic critic by adaptively combining objective metric indicators and MLLM-based holistic quality assessment. This allows the model to "evolve" its generation strategy, focusing on different aspects of quality at different stages of training. To ensure stable and coherent evolution, we anchor the exploring Student model with a frozen Teacher, preserving robust world priors while allowing for creative refinement when generating videos. Extensive experiments demonstrate the superiority of our proposal, and EvoID achieves the total score of 0.687 on the Human-Domain of OpenS2V-Eval dataset, surpassing 0.658 of the open-source VACE and 0.653 of the commercial Hailuo. Moreover, EvoID also obtains a new record of 0.718 on our newly minted MLLM-based metric, prioritizing human perception and more comprehensively reflecting video quality.

Pre-trained weights have become a cornerstone of modern deep learning, enabling efficient knowledge transfer and improving downstream task performance, especially in data-scarce scenarios. However, a fundamental question remains: how can we obtain better pre-trained weights that encapsulate more knowledge beyond the given dataset? In this work, we introduce KNowledge-Overflowed Weights (KNOW) prediction, a novel strategy that leverages structured forgetting and its inversion to synthesize knowledge-enriched weights. Our key insight is that sequential fine-tuning on progressively downsized datasets induces a structured forgetting process, which can be modeled and reversed to recover knowledge as if trained on a larger dataset. We construct a dataset of weight transitions governed by this controlled forgetting and employ meta-learning to model weight prediction effectively. Specifically, our KNowledge-Overflowed Weights Nowcaster (KNOWN) acts as a hyper-model that learns the general evolution of weights and predicts enhanced weights with improved generalization. Extensive experiments across diverse datasets and architectures demonstrate that KNOW prediction consistently outperforms Naive fine-tuning and simple weight prediction, leading to superior downstream performance. Our work provides a new perspective on reinterpreting forgetting dynamics to push the limits of knowledge transfer. The code and pre-trained model are available at https://github.com/jjh6297/KNOW

Recent progress in video diffusion models has enabled remarkable generative fidelity, yet leveraging these priors for restoration remains limited by the strong coupling between conditional and unconditional branches in standard classifier-free guidance. We introduce a training-free framework that enhances distorted and low-resolution videos by decoupling these signals in time. Our proposed Decoupled Time Guidance (DTG) evaluates the unconditional branch at a cleaner diffusion timestep, providing a lookahead prior that preserves geometry while suppressing replication of warped content. This temporal bias is annealed throughout sampling, allowing the model to transition from structure correction to detail refinement without retraining. Combined with any off-the-shelf restoration module in a plug-and-play manner, our approach improves perceptual coherence and restores plausible structure in AI-generated and real-world videos alike. To facilitate evaluation, we curate GenWarp480, a benchmark of 4,400 distorted 480p videos synthesized from diverse text-to-video models. GenWarp480 focuses on characteristic generative degradations such as warped faces, body misalignments, and spatial artifacts, providing a purpose-built testbed for assessing robustness to generative errors. Extensive experiments demonstrate that our method achieves significant improvements in structural fidelity and temporal stability without any model training.

We present TempoMaster, a novel framework that formulates long video generation as next-frame-rate prediction. Specifically, we first generate a low-frame-rate clip that serves as a coarse blueprint of the entire video sequence, and then progressively increase the frame rate to refine visual details and motion continuity. During generation, TempoMaster employs bidirectional attention within each frame-rate level while performing autoregression across frame rates, thus achieving long-range temporal coherence while enabling efficient and parallel synthesis. Extensive experiments demonstrate that TempoMaster establishes a new state-of-the-art in long video generation, excelling in both visual and temporal quality.

Source-Free Cross-Domain Few-Shot Learning (SF-CDFSL) focuses on fine-tuning with limited training data from target domains (e.g., medical or satellite images), where CLIP has recently shown promising results due to its generalizability to downstream tasks. Current works indicate CLIP's text encoder is more suitable for cross-domain tasks, however, we find that removing certain middle layers of the text encoder can effectively improve performance in SF-CDFSL, which we call the Lost Layers. In this paper, we delve into this phenomenon for a deeper understanding. We discover that instead of being harmful for the SF-CDFSL task, the information in these layers is actually beneficial, but visual gaps prevent this useful information from being fully utilized, making these layers seem redundant.Based on this understanding, unlike current works that simply remove these layers, we propose a method to teachs the model to re-utilize information in these lost layersat both the layer and encoder levels, guiding the re-learning of the visual branch under domain shifts. Our approach effectively addresses the issue of underutilized information in the text encoder. Extensive experiments across various settings, backbones (CLIP, SigLip, PE-Core), and tasks (4 CDFSL datasets and 10 Meta-dataset datasets) demonstrate the effectiveness of our method. We will release the code.

Video Large Language Models (VideoLLMs) face a fundamental performance bottleneck: the token explosion intrinsic to video inputs. The resulting O(N^2) prefill cost makes conventional Transformer inference prohibitively expensive at scale. Existing attempts fall into a hard accuracy-latency dilemma: naive sparsification risks losing essential temporal-spatial context, whereas naive parallelization introduces substantial communication and memory overhead. To overcome this impasse, we propose SegMo, a unified framework based on algorithm-system co-design that jointly optimizes what to compute (sparsification) and how to compute it (parallelism). Driven by the empirical discovery of Local Cohesion in VideoLLM attention, SegMo integrates two key components: (1) Content-Aware Sparsification (CAS): A lightweight, hierarchical algorithm that first employs Query Relevance for scene-level assessment, and then uses Temporal Redundancy for intra-scene static redundancy pruning, to generate a precise, non-uniform computation load, ensuring accuracy. (2) Locally-Cohesive Segment Parallelism (LSP): A novel paradigm that leverages attention locality to partition the video at scene boundaries, using a lightweight Global Context Injection mechanism to replace the massive communication and memory overheads of global attention. SegMo was validated across LVBench, LongVideoBench, and Video-MME. Our CAS module improved accuracy by up to 12.00%. When integrated with LSP, the full system (CAS + LSP) achieved a peak prefill acceleration of 3.55x, while still maintaining a significant accuracy gain of up to 8.31%.

Diffusion models achieve remarkable generation quality, yet face a fundamental challenge known as memorization, where generated samples can replicate training samples exactly. We develop a theoretical framework to explain this phenomenon by showing that the empirical score function (the score function corresponding to the empirical distribution) is a weighted sum of the score functions of Gaussian distributions, in which the weights are sharp softmax functions. This structure causes individual training samples to dominate the score function, resulting in sampling collapse. In practice, approximating the empirical score function with a neural network can partially alleviate this issue and improve generalization. Our theoretical framework explains why: In training, the neural network learns a smoother approximation of the weighted sum, allowing the sampling process to be influenced by local manifolds rather than single points. Leveraging this insight, we propose two novel methods to further enhance generalization: (1) Noise Unconditioning enables each training sample to adaptively determine its score function weight to increase the effect of more training samples, thereby preventing single-point dominance and mitigating collapse. (2) Temperature Smoothing introduces an explicit parameter to control the smoothness. By increasing the temperature in the softmax weights, we naturally reduce the dominance of any single training sample and mitigate memorization. Experiments across multiple datasets validate our theoretical analysis and demonstrate the effectiveness of the proposed methods in improving generalization while maintaining high generation quality.

Multimodal models ideally should generalize to unseen domains while remaining data-efficient to reduce annotation costs. To this end, we introduce and study a new problem, Semi-Supervised Multimodal Domain Generalization (SSMDG), which aims to learn robust multimodal models from multi-source data with few labeled samples. We observe that existing approaches fail to address this setting effectively: multimodal domain generalization methods cannot exploit unlabeled data, semi-supervised multimodal learning methods ignore domain shifts, and semi-supervised domain generalization methods are confined to single-modality inputs. To overcome these limitations, we propose a unified framework featuring three key components: Consensus-Driven Consistency Regularization, which obtains reliable pseudo-labels through confident fused-unimodal consensus; Disagreement-Aware Regularization, which effectively utilizes ambiguous non-consensus samples; and Cross-Modal Prototype Alignment, which enforces domain- and modality-invariant representations while promoting robustness under missing modalities via cross-modal translation. We further establish the first SSMDG benchmarks, on which our method consistently outperforms strong baselines in both standard and missing-modality scenarios. Our benchmarks and code are available at https://github.com/lihongzhao99/SSMDG.

Modern optical flow estimation, though empowered by recent deep neural architectures, remains rooted in the discrete correspondence paradigm inherited from classical vision. Most networks infer frame-to-frame displacements, capturing where pixels move but not how motion evolves continuously through time. Yet physical motion in the real world follows smooth dynamics governed by underlying velocity fields, as long established in fluid mechanics and transport theory. To bridge this gap, we introduce Optical Flow Matching (OFM), a continuous formulation that learns a time-dependent velocity field to transport pixel coordinates along motion distribution coherent trajectories. A key component of our OFM is Triangle Velocities Synergy (TVS), a neat yet pragmatic geometric mechanism that provides a stable and physically meaningful velocity construction, ensuring that continuous transport remains well-defined. Combined with an Euler-based ODE solver, OFM yields flow fields that are temporally smooth, geometrically consistent, and process-interpretable. Experiments on Sintel, KITTI, and Spring demonstrate that OFM achieves state-of-the-art accuracy, enhanced temporal stability, and notably stronger cross-dataset generalization, advancing optical flow estimation from correspondence inference to continuous dynamical reasoning.

Non-line-of-sight (NLOS) imaging seeks to recover hidden-scene information from indirect light transport beyond the direct line of sight. Existing NLOS methods can be broadly categorized into active and passive approaches. Active methods rely on controlled illumination and time-resolved sensors, but their dependence on specialized and expensive hardware limits practical deployment. Passive methods instead use steady-state or uncontrolled indirect reflections with conventional sensors, but often suffer from weak signals, unstable observations, and insufficient measurement constraints. In this paper, we present the Similarity-Likelihood Diffusion Network (SLD-Net) for corner-camera hidden-person imaging from multi-exposure wall observations. SLD-Net consists of two stages: DeLi-Inversion estimates an initial reconstruction and a pixel-wise precision map to construct a heteroscedastic pseudo-likelihood, and SiCo-Diffusion incorporates this likelihood into deterministic DDIM sampling and fuses it with the diffusion prior via annealed Bayesian precision updating, producing reproducible and measurement-consistent reconstructions. Experiments on the Reflect-Corridor and Reflect-Room datasets demonstrate consistent improvements over generic, physics-inspired, and NLOS-specific baselines. SLD-Net improves PSNR from 13.84 to 15.58 dB and reduces FID from 264.91 to 73.54 on Reflect-Corridor, and improves PSNR from 11.58 to 12.49 dB and reduces FID from 177.05 to 26.89 on Reflect-Room.

Vision-Language Models (VLMs), leveraging their powerful visual perception and reasoning capabilities, have been widely applied in Unmanned Aerial Vehicle (UAV) tasks.However, the spatial intelligence capabilities of existing VLMs in UAV scenarios remain largely unexplored, raising concerns about their effectiveness in navigating and interpreting dynamic environments.To bridge this gap, we introduce SpatialSky-Bench, a comprehensive benchmark specifically designed to evaluate the spatial intelligence capabilities of VLMs in UAV navigation. Our benchmark comprises two categories--Environmental Perception and Scene Understanding--divided into 13 subcategories, including bounding boxes, color, distance, height, and landing safety analysis, among others.Extensive evaluations of various mainstream open-source and closed-source VLMs reveal unsatisfactory performance in complex UAV navigation scenarios, highlighting significant gaps in their spatial capabilities.To address this challenge, we developed the SpatialSky-Dataset, a comprehensive dataset containing 1 M samples with diverse annotations across various scenarios. Leveraging this dataset, we introduce Sky-VLM, a specialized VLM designed for UAV spatial reasoning across multiple granularities and contexts.Extensive experimental results demonstrate that Sky-VLM achieves state-of-the-art performance across all benchmark tasks, paving the way for the development of VLMs suitable for drone scenarios.The source code is available at https://github.com/linglingxiansen/SpatialSKy .

Previous representation and generation approaches for the B-rep relied on graph-based representations that disentangle geometric and topological features through decoupled computational pipelines, thereby precluding the application of sequence-based generative frameworks, such as transformer architectures that have demonstrated remarkable performance. In this paper, we propose BrepARG, the first attempt to encode B-rep's geometry and topology into a holistic token sequence representation, enabling sequence-based B-rep generation with an autoregressive architecture. Specifically, BrepARG encodes B-rep into 3 types of tokens: geometry and position tokens representing geometric features, and face index tokens representing topology. Then the holistic token sequence is constructed hierarchically, starting with constructing the geometry blocks (i.e., faces and edges) using the above tokens, followed by geometry block sequencing. Finally, we assemble the holistic sequence representation for the entire B-rep. We also construct a transformer-based autoregressive model that learns the distribution over holistic token sequences via next-token prediction, using a multi-layer decoder-only architecture with causal masking. Experiments demonstrate that BrepARG achieves state-of-the-art (SOTA) performance. BrepARG validates the feasibility of representing B-rep as holistic token sequences, opening new directions for B-rep generation.

As humans, we are natural any-horizon reasoners, i.e., we can decide whether to iteratively skim long videos or watch short ones in full when necessary for a given task. With this in mind, one would expect video reasoning models to reason flexibly across different durations. However, SOTA models are still trained to predict answers in a single turn while processing a large number of frames, akin to watching an entire long video, requiring significant resources. This raises the question: Is it possible to develop performant any-horizon video reasoning systems? Inspired by human behavior, we first propose SAGE, an agent system that performs multi-turn reasoning on long videos while handling simpler problems in a single turn. Secondly, we introduce an easy synthetic data generation pipeline using Gemini-2.5-Flash to train the orchestrator, SAGE-MM, which lies at the core of SAGE. We further propose an effective RL post-training recipe essential for instilling any-horizon reasoning ability in SAGE-MM. Thirdly, we curate SAGE-Bench with an average duration of greater than 700 seconds for evaluating video reasoning ability in real-world entertainment use cases. Lastly, we empirically validate the effectiveness of our system, data, and RL recipe, observing notable improvements of up to 6.1% on open-ended video reasoning tasks, as well as an impressive 8.2% improvement on videos longer than 10 minutes. We will open-source our system code, data, and checkpoints upon publication.

Novel view synthesis from low dynamic range (LDR) blurry images, which are common in the wild, struggles to recover high dynamic range (HDR) and sharp 3D representations in extreme lighting conditions. Although existing methods employ event data to address this issue, they ignore the sensor-physics mismatches between the camera output and physical world radiance, resulting in suboptimal HDR and deblurring results. To cope with this problem, we propose a unified sensor-physics grounded NeRF framework for sharp HDR novel view synthesis from single-exposure blurry LDR images and corresponding events. We employ NeRF to directly represent the actual radiance of the 3D scene in the HDR domain and model raw HDR scene rays hitting the sensor pixels as in the physical world. A 2D pixel-wise RGB CRF model is introduced to align the NeRF rendered pixel values with the sensor-recorded LDR pixel values of the input images. A novel event CRF model is also designed to bridge the gap between physical scene dynamics and event sensor output. The two models are jointly optimized with the NeRF network, leveraging the spatial and temporal dynamic information in events to enhance the sharp HDR 3D representation learning. Experiments on the collected and public datasets demonstrate that our method achieves state-of-the-art HDR and deblurring novel view synthesis results with single-exposure blurry LDR images and corresponding events. Our code and datasets are publicly available at https://github.com/iCVTEAM/See-NeRF.

Vision-Language-Action (VLA) models have significantly advanced robotic agents capable of executing diverse tasks; however, they remain limited in contact-rich manipulation scenarios that require precise physical interactions. To address this limitation, recent studies have attempted to incorporate tactile signals during downstream tasks, enabling pretrained VLAs to interpret tactile feedback. Nevertheless, introducing new modalities during finetuning, which are rarely present in the pretrain stage, may disrupt the pretrained capabilities of VLAs. In addition, the inherently slow inference speed of VLAs hampers real-time responsiveness and limits the effective utilization of tactile feedback for action adjustment.To overcome these challenges, we propose Adaptive Tactile Vision-Language-Action (AT-VLA), which introduces a novel Adaptive Tactile Injection mechanism. This mechanism dynamically determines the appropriate timing and locations for tactile injection, incorporating only when it significantly contributes to action generation, thereby minimizing interference with pretrained representations.Furthermore, to enable rapid and accurate tactile responses, we propose a Tactile Reaction Dual-Stream mechanism, which decouples sensory processing into a slow visual-language stream for low-frequency perceptual reasoning and a fast tactile control stream for high-frequency physical interaction understanding, achieving real-time close-loop responses within 0.04 s.Real-world experiments thoroughly validate the effectiveness of AT-VLA in contact-rich manipulation tasks.

Direct Preference Optimization (DPO) has proven to be an effective solution for mitigating hallucination in Multimodal Large Language Models (MLLMs) by learning from preference pairs. One of its key challenges lies in how to transfer the sequence-level preference into fine-grained supervision on visual fidelity. To safeguard vision-related tokens that are prone to hallucination, existing methods typically allocate training emphasis according to the model's self-assessed visual sensitivity signals. However, such sensitivity, estimated by a model still under training, introduces self-referential bias: reinforcing already well-learned visual cues while neglecting hard-to-perceive but critical details, thereby limiting deeper alignment. In this work, we propose an Uncertainty-aware Exploratory Direct Preference Optimization (UE-DPO) method for MLLMs, which enables the model to uncover its cognitive deficiencies and actively explore for self-correction, guided by token-level epistemic uncertainty. Specifically, we first quantify the uncertainty from the model's failure to ground token predictions in the given image. Then, based on an uncertainty-aware exploration intensity, we encourage more learning pressure on visually deficient tokens in preferred samples, and alleviate the over-penalization of beneficial knowledge in dispreferred samples. Further, we provide a theoretical justification for our method, and extensive experiments demonstrate its effectiveness and robustness.

Modern stereo matching methods have leveraged monocular depth foundation models to achieve superior zero-shot generalization performance. However, most existing methods primarily focus on extracting robust features for cost volume construction or disparity initialization. At the same time, the iterative refinement stage, which is also crucial for zero-shot generalization, remains underexplored. Some methods treat monocular depth priors as guidance for iteration, but conventional GRU-based architectures struggle to exploit them due to the limited representation capacity. In this paper, we propose Prompt Recurrent Unit (PRU), a novel iterative refinement module based on the decoder of monocular depth foundation models. By integrating monocular structure and stereo motion cues as prompts into the decoder, PRU enriches the latent representations of monocular depth foundation models with absolute stereo-scale information while preserving their inherent monocular depth priors. Experiments demonstrate that our PromptStereo achieves state-of-the-art zero-shot generalization performance across multiple datasets, while maintaining comparable or faster inference speed. Our findings highlight prompt-guided iterative refinement as a promising direction for zero-shot stereo matching.

While large vision-language models (LVLMs) achieve strong performance on multimodal tasks, they frequently generate hallucinations--unfaithful outputs misaligned with the visual input. To address this issue, we introduce CIPHER (Counterfactual Image Perturbations for Hallucination Extraction and Removal), a training-free method that suppresses vision-induced hallucinations via lightweight feature-level correction. Unlike prior training-free approaches that primarily focus on text-induced hallucinations, CIPHER explicitly targets hallucinations arising from the visual modality. CIPHER operates in two phases. In the offline phase, we construct OHC-25K (Object-Hallucinated Counterfactuals, 25,000 samples), a counterfactual dataset consisting of diffusion-edited images that intentionally contradict the original ground-truth captions. We pair these edited images with the unchanged ground-truth captions and process them through an LVLM to extract hallucination-related representations. Contrasting these representations with those from authentic (image, caption) pairs reveals structured, systematic shifts spanning a low-rank subspace characterizing vision-induced hallucination. In the inference phase, CIPHER suppresses hallucinations by projecting intermediate hidden states away from this subspace. Experiments across multiple benchmarks show that CIPHER significantly reduces hallucination rates while preserving task performance, demonstrating the effectiveness of counterfactual visual perturbations for improving LVLM faithfulness.