CVPR 2026 Papers

We present a novel unsupervised framework to unlock vast unlabeled human demonstration data from continuous industrial video streams for Vision-Language-Action (VLA) model pre-training. Our method first trains a lightweight motion tokenizer to encode motion dynamics, then employs an unsupervised action segmenter leveraging a novel "Latent Action Energy" metric to discover and segment semantically coherent action primitives. The pipeline outputs both segmented video clips and their corresponding latent action sequences, providing structured data directly suitable for VLA pre-training. Evaluations on public benchmarks and a proprietary electric motor assembly dataset demonstrate effective segmentation of key tasks performed by humans at workstations. Further clustering and quantitative assessment via a Vision-Language Model confirm the semantic coherence of the discovered action primitives. To our knowledge, this is the first fully automated end-to-end system for extracting and organizing VLA pre-training data from unstructured industrial videos, offering a scalable solution for embodied AI integration in manufacturing.

Open-vocabulary scene understanding with online panoptic mapping is essential for embodied applications to perceive and interact with environments. However, existing methods are predominantly offline or lack instance-level understanding, limiting their applicability to real-world robotic tasks. In this paper, we propose OnlinePG, a novel and effective system that integrates geometric reconstruction and open-vocabulary perception using 3D Gaussian Splatting in an online setting. Technically, to achieve online panoptic mapping, we employ an efficient local-to-global paradigm with a sliding window. To build local consistency map, we construct a 3D segment clustering graph that jointly leverages geometric and semantic cues, fusing inconsistent segments within sliding window into complete instances. Subsequently, to update the global map, we construct explicit spatial attribute grids for the local 3D Gaussian map and fuse them into the global map via robust bidirectional bipartite 3D Gaussian instance matching. Finally, we utilize the fused VLM features inside the 3D spatial attribute grids to achieve open-vocabulary scene understanding. Extensive experiments on widely used datasets demonstrate that our method achieves better performance among online approaches, while maintaining real-time efficiency.

Recent end-to-end autonomous driving approaches have leveraged Vision-Language Models (VLMs) to enhance planning capabilities in complex driving scenarios. However, VLMs are inherently trained as generalist models, lacking specialized understanding of driving-specific reasoning in 3D space and time. When applied to autonomous driving, these models struggle to establish structured spatial-temporal representations that capture geometric relationships, scene context, and motion patterns critical for safe trajectory planning.To address these limitations, we propose SGDrive, a novel framework that explicitly structures the VLM's representation learning around driving-specific knowledge hierarchies. Built upon a pre-trained VLM backbone, SGDrive decomposes driving understanding into a scene-agent-goal hierarchy that mirrors human driving cognition: drivers first perceive the overall environment (scene context), then attend to safety-critical agents and their behaviors, and finally formulate short-term goals before executing actions. This hierarchical decomposition provides the structured spatial-temporal representation that generalist VLMs lack, integrating multi-level information into a compact yet comprehensive format for trajectory planning.Extensive experiments on the NAVSIM benchmark demonstrate that SGDrive achieves state-of-the-art performance among camera-only methods, validating the effectiveness of hierarchical knowledge structuring for adapting generalist VLMs to autonomous driving.

The approximation and convergence properties of implicit neural representations (INRs) are known to be highly sensitive to parameter initialization strategies. While several data-driven initialization methods demonstrate significant improvements over standard random sampling, the reasons for their success -- specifically, whether they encode classical statistical signal priors or more complex features -- remain poorly understood. In this study, we explore this phenomenon through a series of experimental analyses leveraging noise pretraining. We pretrain INRs on diverse noise classes (e.g., Gaussian, Dead Leaves, Spectral) and measure their ability to both fit unseen signals and encode priors for an inverse imaging task (denoising). Our analyses on image and video data reveal a surprising finding: simply pretraining on unstructured noise (Uniform, Gaussian) dramatically improves signal fitting capacity compared to all other baselines. However, unstructured noise also yields poor deep image priors for denoising. In contrast, we also find that noise with the classic 1/|f^\alpha| spectral structure of natural images achieves an excellent balance of signal fitting and inverse imaging capabilities, performing on par with the best data-driven initialization methods. This finding enables more efficient INR training in applications lacking sufficient prior domain-specific data.

Human perception of visual similarity is inherently adaptive and subjective, depending on the users' interests and focus. However, most image retrieval systems fail to reflect this flexibility, relying on a fixed, monolithic metric that cannot incorporate multiple conditions simultaneously. To address this, we propose CLAY, an adaptive similarity computation method that reframes the embedding space of pretrained Vision-Language Models (VLMs) as a text-conditional similarity space without additional training. This design separates the textual conditioning process and visual feature extraction, allowing highly efficient and multi-conditioned retrieval with fixed visual embeddings. We also construct a synthetic evaluation dataset CLAY-EVAL, for comprehensive assessment under diverse conditioned retrieval settings. Experiments on standard datasets and our proposed dataset show that CLAY achieves state-of-the art retrieval accuracy and notable computational efficiency compared to previous works.

Recent progress in multimodal large language models (MLLMs) has highlighted the challenge of efficiently bridging pre-trained Vision-Language Models (VLMs) with Diffusion Models. While methods using a fixed number of learnable query tokens offer computational efficiency, they suffer from task generalization collapse, failing to adapt to new tasks that are distant from their pre-training tasks. To overcome this, we propose Noisy Query Tokens, which learn a distributed representation space between the VLM and Diffusion Model via end-to-end optimization, enhancing continual learning. Additionally, we introduce a VAE branch with linear projection to recover fine-grained image details. Experimental results confirm our approach mitigates generalization collapse and enables stable continual learning across diverse tasks.

Dense Video Object Captioning (DVOC) is the task of jointly detecting, tracking, and captioning object trajectories in a video, requiring the ability to understand spatio-temporal details and describe them in natural language.Due to the complexity of the task and the high cost associated with manual annotation, previous approaches resort to training strategies with limited data, potentially leading to suboptimal performance.To circumvent this issue, we propose to generate captions about spatio-temporally localized entities leveraging a state-of-the-art VLM, and extend the LVIS and LV-VIS datasets with our synthetic captions (LVISCap and LV-VISCap). Moreover, we introduce an end-to-end model, CaptionFormer, capable of jointly detecting, segmenting, tracking and captioning object trajectories.With pretraining on LVISCap and LV-VISCap, CaptionFormer achieves state-of-the-art DVOC results on three existing benchmarks, VidSTG, VLN and BenSMOT.

Reinforcement Learning with Verifiable Rewards (RLVR) traditionally relies on a sparse, outcome-based signal. Recent work shows that providing a fine-grained, model-intrinsic signal--rewarding the confidence growth in the ground-truth answer--effectively improves language reasoning training by providing step-level guidance without costly external models. While effective for unimodal text, we find that naively applying this global reward to vision-language (V-L) reasoning is a suboptimal strategy, as the task is a heterogeneous mix of sparse visual perception and dense textual reasoning. This global normalization creates mixture-induced signal degradation, where the training signal for visual steps is statistically distorted by the predominant textual steps. We propose Perception-Decomposed Confidence Reward (PDCR), a framework that solves this by aligning the reward structure with the task's heterogeneous nature. PDCR first performs an unsupervised skill decomposition, introducing a model-internal Visual Dependence Score to quantify visual reliance and applying a clustering algorithm to separate perception and reasoning steps. Based on this, PDCR computes a decomposed advantage by normalizing confidence gains within each skill cluster. This intra-cluster normalization provides a stable, correctly-scaled signal for both perception and reasoning. We demonstrate that PDCR outperforms the naive, global-reward formulation and sparse-reward baselines on key V-L reasoning benchmarks.

Long video understanding presents significant challenges for vision-language models due to extremely long context windows. Existing solutions relying on naive chunking strategies with retrieval-augmented generation, typically suffer from information fragmentation and a loss of global coherence. We present HAVEN, a unified framework for long-video understanding that enables coherent and comprehensive reasoning by integrating audiovisual entity cohesion and hierarchical video indexing with agentic search. First, we preserve semantic consistency by integrating entity-level representations across visual and auditory streams, while organizing content into a structured hierarchy spanning global summary, scene, segment, and entity levels. Then we employ an agentic search mechanism to enable dynamic retrieval and reasoning across these layers, facilitating coherent narrative reconstruction and fine-grained entity tracking. Extensive experiments demonstrate that our method achieves good temporal coherence, entity consistency, and retrieval efficiency, establishing a new state-of-the-art with an overall accuracy of 84.1% on LVBench. Notably, it achieves outstanding performance in the challenging reasoning category, reaching 80.1%. These results highlight the effectiveness of structured, multimodal reasoning for comprehensive and context-consistent understanding of long-form videos.

Preference-based finetuning of vision-language models (VLMs) is notoriously unstable, as trivially wrong negatives inject uninformative gradients that distort optimization and degrade calibration. This work revisits this issue through the lens of learning dynamics and identifies a core pathology, the squeezing effect, where easy negatives retain large, misaligned gradients despite having negligible loss.To address this, we propose Cooling-Weighted Direct Preference Optimization (CW-DPO), a two-stage framework that first smooths and then stabilizes the alignment process. Stage 1 employs a constrained SFT phase with low-weight "gentle negatives" to regularize overconfident distributions and flatten the loss landscape. Stage 2 introduces a competence-aware cooling weight that adaptively scales negative gradients according to the model's average per-token log-probability, suppressing uninformative updates while emphasizing hard, on-policy contrasts. This dynamics-aware weighting effectively mitigates the squeezing effect and enables smoother convergence.Extensive experiments on mainstream benchmarks--including COCO, Flickr30k, NoCaps, MMMU, and MMBench1.1--show that CW-DPO achieves state-of-the-art performance, for example +3.4 CIDEr over PPO and +2.4% absolute accuracy on MMMU, while improving calibration and halving convergence steps. This demonstrates that smoothing before cooling constitutes a simple yet general principle for robust VLM preference optimization.

Large Vision-Language Models (LVLMs) excel in visual understanding and reasoning, but the excessive visual tokens lead to high inference costs. Although recent token reduction methods mitigate this issue, they mainly target single-turn Visual Question Answering (VQA), leaving the more practical multi-turn VQA (MT-VQA) scenario largely unexplored. MT-VQA introduces additional challenges, as subsequent questions are unknown beforehand and may refer to arbitrary image regions, making existing reduction strategies ineffective. Specifically, current approaches fall into two categories: prompt-dependent methods, which bias toward the initial text prompt and discard information useful for subsequent turns; prompt-agnostic ones, which, though technically applicable to multi-turn settings, rely on heuristic reduction metrics such as attention scores, leading to suboptimal performance. In this paper, we propose a learning-based prompt-agnostic method, termed MetaCompress, overcoming the limitations of heuristic designs. We begin by formulating token reduction as a learnable compression mapping, unifying existing formats such as pruning and merging into a single learning objective. Upon this formulation, we introduce a data-efficient training paradigm capable of learning optimal compression mappings with limited computational costs. Extensive experiments on MT-VQA benchmarks and across multiple LVLM architectures demonstrate that MetaCompress achieves superior efficiency-accuracy trade-offs while maintaining strong generalization across dialogue turns. Our code is available at https://github.com/MArSha1147/MetaCompress.

Image segmentation remains a challenging task in computer vision, demanding robust mask generation and precise classification. Recent mask-based approaches yield high-quality masks by capturing global context. However, accurately classifying these masks, especially in the presence of ambiguous boundaries and imbalanced class distributions, remains an open challenge. In this work, we introduce ViT-P, a novel two-stage segmentation framework that decouples mask generation from classification. The first stage employs a proposal generator to produce class-agnostic mask proposals, while the second stage utilizes a point-based classification model built on the Vision Transformer (ViT) to refine predictions. ViT-P serves as a pre-training-free adapter, allowing the integration of various pre-trained vision transformers without modifying their architecture, ensuring adaptability to dense prediction tasks. Furthermore, we demonstrate that coarse and bounding box annotations can effectively enhance classification without requiring additional training on fine annotation datasets, reducing annotation costs while maintaining strong performance. Extensive experiments across COCO, ADE20K, and Cityscapes datasets validate the effectiveness of ViT-P, achieving state-of-the-art results with 54.0 PQ on ADE20K panoptic segmentation, 87.4 mIoU on Cityscapes semantic segmentation, and 63.6 mIoU on ADE20K semantic segmentation.

Immunofluorescence (IF) images reveal detailed information about structures and functions at the subcellular level. However, unlike RGB images, IF datasets pose challenges for deep learning models due to their inconsistencies in channel count and configuration, stemming from varying staining protocols across laboratories and studies. Although existing approaches build channel-adaptive models for training, they do not perform evaluations across IF datasets with unseen channel configurations. To address this, we first introduce a biologically informed view of cellular image channels by grouping them into either context or concept, where we treat the context channels as a reference for the concept channels in the image. We leverage this view to propose Channel Conditioned Cell Representations (C3R), a framework that learns representations that transfers well to both in-distribution (ID) and out-of-distribution (OOD) datasets which contain same and different channel configurations, respectively. C3R is a two-fold framework comprising a channel-adaptive encoder architecture and a masked knowledge distillation training strategy, both built around the context-concept principle. We find that C3R outperforms existing benchmarks on both ID and OOD tasks, while yielding state-of-the-art results on frozen encoder evaluation on the CHAMMI benchmark. Our method opens a new pathway for cross-dataset generalization between IF datasets, with no need for retraining on unseen channel configurations.

Large Vision-Language Models (LVLMs) exhibit powerful reasoning capabilities but suffer sophisticated jailbreak vulnerabilities. Fundamentally, aligning LVLMs is not just a safety challenge but a problem of economic efficiency. Current alignment methods struggle with the trade-off between safety, utility, and operational costs. Critically, a focus solely on final outputs (process-blindness) wastes significant computational budget on unsafe deliberation. This flaw allows harmful reasoning to be disguised with benign justifications, thereby circumventing simple additive safety scores.To address this, we propose EcoAlign, an inference-time framework that reframes alignment as an economically rational search by treating the LVLM as a boundedly rational agent. EcoAlign incrementally expands a thought graph and scores actions using a forward-looking function (analogous to net present value) that dynamically weighs expected safety, utility, and cost against the remaining budget. To prevent deception, path safety is enforced via the weakest-link principle. Extensive experiments across 3 closed-source and 2 open-source models on 6 datasets show that EcoAlign matches or surpasses state-of-the-art safety and utility at a lower computational cost, thereby offering a principled, economical pathway to robust LVLM alignment.

Current text-to-image (T2I) generation models often struggle with prompts that require complex reasoning or specialized knowledge, failing to accurately interpret implicit user intent. To bridge this gap, we introduce T2I-Reason, a large-scale dataset designed to empower text-to-image generation in unified multimodal models (UMMs) with reasoning and knowledge. The dataset contains 120k pairs of text triplet and image. The text triplet consists of (1) an implicit prompt, which requires reasoning or knowledge to decipher its underlying meaning; (2) a reasoning chain, which provides a step-by-step analysis to resolve the implicit prompt's meaning; and (3) an explicit prompt, a clear and straightforward visual description prepared for T2I generation. T2I-Reason is meticulously constructed: 65k samples are dedicated to reasoning, specifically targeting arithmetic reasoning, spatial-attribute relationship reasoning, deductive reasoning (cause to effect), and abductive reasoning (effect to cause). While 55k samples necessitate specialized knowledge, which covers multiple disciplines, spatial-temporal concepts, and entity knowledge. To validate the effectiveness of our dataset, we train a unified multimodal model, Bagel, on our dataset. Results across multiple benchmarks that evaluate the reasoning capabilities of T2I generation demonstrate that our model achieves significant and consistent improvements on both composition and reasoning, confirming that explicit training on intermediate reasoning chains is a pivotal step towards more intelligent unified generative models.

Music-driven 3D dance generation offers significant creative potential, yet practical applications demand versatile and multimodal control. Given the highly dynamic and complex human motion covering various styles and genres, dance generation requires satisfying diverse conditions beyond just music (e.g., spatial trajectories, keyframe gestures, or style descriptions). However, the absence of a large-scale and richly annotated dataset severely hinders progress. In this paper, we build OpenDanceSet, an extensive human dance dataset comprising over 100 hours across 14 genres and 147 subjects. Each sample has rich annotations to facilitate robust cross-modal learning: 3D motion, paired music, 2D keypoints, trajectories, and expert-annotated text descriptions. Furthermore, we propose OpenDanceNet, a unified masked modeling framework for controllable dance generation, including a disentangled auto-encoder and a multimodal joint-prediction Transformer. OpenDanceNet supports generation conditioned on music and arbitrary combinations of text, keypoints, or trajectories. Comprehensive experiments demonstrate that our work achieves high-fidelity synthesis with strong diversity and realistic physical contacts, while also offering flexible control over spatial and stylistic conditions.

Real-world image restoration is challenged by complex, coupled degradations. Existing "all-in-one" models often lack generalization, while agentic systems suffer from inefficient sequential tool invocation. We propose a VLM-guided one-shot framework for universal photographic post-processing. Our system employs a Vision-Language Model (VLM) as an orchestrator to perform nuanced intent understanding and degradation analysis, dynamically allocating weights to a suite of specialized expert LoRA modules. To ensure superior composability, these experts adapt only Key (K) and Value (V) matrices and are simultaneously merged into a pretrained diffusion backbone for synergistic, single-pass restoration. Furthermore, we introduce a lightweight branch trained via Direct Preference Optimization (DPO) to ensure perceptually optimal weight allocation. Our method achieves state-of-the-art performance across diverse synthetic and real-world datasets. Crucially, it demonstrates remarkable zero-shot generalization on authentic real-world data without additional fine-tuning.

Existing GUI benchmarks primarily focus on evaluating models' comprehensive capabilities but largely overlook hallucination phenomena in grounding tasks, which are crucial to the reliability of GUI understanding. In this work, we expose two major types of hallucinations in GUI grounding: 1) Confusion Hallucination, where distractor elements are mistakenly selected, and 2) Fabricated Hallucination, where nonexistent elements are hallucinated with plausible coordinates. To systematically investigate their origins, we introduce GUI-HalluBench, a benchmark comprising two complementary subsets: a parsing subset for assessing structural representation of GUI elements and a hallucination subset for measuring grounding robustness under challenging conditions. This design allows us to associate hallucination patterns with deficiencies in prerequisite abilities: parsing errors are closely tied to both fabricated and confusion hallucinations. Experiments on state-of-the-art models confirm these connections, offering new insights into the root causes of hallucinations and guiding the development of more reliable GUI understanding tools.

Pose-guided human image animation aims to synthesize realistic videos of a reference character driven by a sequence of poses. While diffusion-based methods have achieved remarkable success, most existing approaches are limited to single-character animation. We observe that naively extending these methods to multi-character scenarios often leads to identity confusion and implausible occlusions between characters. To address these challenges, in this paper, we propose an extensible multi-character image animation framework built upon modern Diffusion Transformers (DiTs) for video generation. At its core, our framework introduces two novel components--Identifier Assigner and Identifier Adapter--which collaboratively capture per-person positional cues and inter-person spatial relationships. This mask-driven scheme, along with a scalable training strategy, not only enhances flexibility but also enables generalization to scenarios with more characters than those seen during training. Remarkably, trained on only a two-character dataset, our model generalizes to multi-character animation while maintaining compatibility with single-character cases. Extensive experiments demonstrate that our approach achieves state-of-the-art performance in multi-character image animation, surpassing existing diffusion-based baselines. Codes will be released.