CVPR 2026 Papers

Reinforcement Learning with Verifiable Reward (RLVR) has substantially advanced the video understanding capabilities of Multimodal Large Language Models (MLLMs). However, the rapid progress of MLLMs is outpacing the complexity of existing video datasets, while the manual annotation of new, high-quality data remains prohibitively expensive.This work investigates a pivotal question: Can the rich, intrinsic information within videos be harnessed to self-generate high-quality, verifiable training data?To investigate this problem, we first introduce three self-supervised pretext tasks for video understanding: Anomaly Grounding, Object Counting, and Temporal Jigsaw. To validate the difficulty of these tasks, we construct the Video Intrinsic Understanding Benchmark (VIUBench), revealing that current state-of-the-art MLLMs struggle significantly on these tasks. Building upon these pretext tasks, we develop the VideoSSR-30K dataset and propose VideoSSR, a novel video self-supervised reinforcement learning framework for RLVR. Extensive experiments across 17 benchmarks, spanning four major video domains (General Video QA, Long Video QA, Temporal Grounding, and Complex Reasoning), demonstrate that our VideoSSR consistently enhances model performance, yielding an average improvement of over 5%. These results establish VideoSSR as a potent foundational framework for developing more advanced video understanding in MLLMs.

Rendering 3D surfaces has been revolutionized within the modeling of radiance fields through either 3DGS or NeRF. Although 3DGS has shown advantages over NeRF in terms of rendering quality or speed, there is still room for improvement in recovering high fidelity surfaces through 3DGS. To resolve this issue, we propose a self-constrained prior to constraining the movement of 3D Gaussians, aiming for more accurate depth rendering. Our self-constrained prior is a TSDF grid fused by the rendered depth during the learning of 3D Gaussians. The prior measures a band on both sides of the estimated surface for imposing more specific constraints on the right 3D Gaussians, such as removing 3D Gaussians outside the band, encouraging larger opacity for Gaussians near the center of the band or smaller opacity for Gaussians near the boundary of the band. We regularly update the prior by fusing more recent depth images which are usually more accurate, and progressively narrow the band to tighten the constraint on Gaussian movements. We justify our idea and report our superiority over the state-of-the-art methods in evaluations on widely used benchmarks.

Understanding and answering questions based on a user's pointing gesture is essential for next-generation egocentric AI assistants. However, current Multimodal Large Language Models (MLLMs) struggle with such tasks due to the lack of gesture-rich data and their limited ability to infer fine-grained pointing intent from egocentric video.To address this, we introduce EgoPointVQA, a dataset and benchmark for gesture-grounded egocentric question answering, comprising 4000 synthetic and 400 real-world videos across multiple deictic reasoning tasks.Built upon it, we further propose Hand Intent Tokens (HINT), which encode tokens derived from 3D hand keypoints using an off-the-shelf reconstruction model and interleaves them with the model input to provide explicit spatial and temporal context for interpreting pointing intent.We show that our model outperforms others in different backbones and model sizes.In particular, HINT-14Bachieves 68.1% accuracy, on average over 6 tasks, surpassing the state-of-the-art, InternVL3-14B, by 6.6%.To further facilitate the open research, we will release the code, model, and dataset.

Monocular 3D object detection from a single RGB image remains challenging due to two fundamental challenges: the ill-posed nature of 3D localization, where multiple plausible configurations can correspond to the same 2D observation, and unreliable confidence estimation that fails to reflect true localization accuracy. Existing methods predict deterministic 3D boxes that often collapse to implausible mean estimates and rely on absolute confidence scores that are highly sensitive to localization errors. This paper introduces RARE, a unified framework that addresses both challenges through learning to rank and retrieve. RARE formulates confidence estimation as a ranking problem, learning to order detections by their relative quality rather than regressing absolute values. It provides more robust and stable confidence estimates that are less sensitive to localization uncertainty. Building on this improved confidence estimator, RARE learns to construct a query set for each object that predicts multiple diverse and plausible 3D configurations, and retrieves the top-ranked prediction. It explicitly models the multimodal nature of monocular 3D perception and produces more plausible localizations. Extensive experiments demonstrate the effectiveness of RARE. The code is available at https://github.com/HyeonjeongPark37/RARE.

Exemplar-Free Class Incremental Learning (EFCIL) aims to enable models to learn new classes sequentially without retaining samples from previous tasks. While recent approaches leverage pre-trained models with parameter-efficient tuning to mitigate forgetting, they often overlook a crucial cause of forgetting: the collapse of the class-discriminative structure. This structure comprises two interdependent components: intra-class structure, which characterizes the shape of individual classes, and inter-class structure, which characterizes the global geometric relationships among class prototypes. We reveal that catastrophic forgetting stems from the simultaneous deterioration of both intra-class and inter-class structures. To address this, we propose a unified framework that preserves the class-discriminative structure. It preserves the intra-class structure by reshaping class means and covariances to preserve each class's shape during migration, and maintains inter-class structure by stabilizing angular relationships between samples and old prototypes. Extensive experiments demonstrate that our framework outperforms existing leading methods on multiple EFCIL benchmarks, validating that preserving the class-discriminative structure is crucial for mitigating catastrophic forgetting.

We identify occlusion reasoning as a fundamental yet overlooked aspect for 3D layout-conditioned generation. It is essential for synthesizing partially occluded objects with depth-consistent geometry and scale. While existing methods can generate realistic scenes that follow input layouts, they often fail to model precise inter-object occlusions. We propose SeeThrough3D, a model for 3D layout conditioned generation that explicitly models occlusions. We introduce an occlusion-aware 3D scene representation (OSCR), where objects are depicted as translucent 3D boxes placed within a virtual environment and rendered from desired camera viewpoint. The transparency encodes hidden object regions, enabling the model to reason about occlusions, while the rendered viewpoint provides explicit camera control during generation. We condition a pretrained flow based text-to-image image generation model by introducing a set of visual tokens derived from our rendered 3D representation. Furthermore, we apply masked self-attention to accurately bind each object bounding box to its corresponding textual description, enabling accurate generation of multiple objects without object attribute mixing. To train the model, we construct a synthetic dataset with diverse multi-object scenes with strong inter-object occlusions. SeeThrough3D generalizes effectively to unseen object categories and enables precise 3D layout control with realistic occlusions and consistent camera control.

Autoregressive models can generate high-quality 3D meshes by sequentially producing vertices and faces, but their token-by-token decoding results in slow inference, limiting practical use in interactive and large-scale applications.We present FlashMesh, a fast and high-fidelity mesh generation framework that rethinks autoregressive decoding through a predict-correct-verify paradigm. The key insight is that mesh tokens exhibit strong structural and geometric correlations that enable confident multi-token speculation. FlashMesh leverages this by introducing a speculative decoding scheme tailored to the commonly used hourglass transformer architecture, enabling parallel prediction across face, point, and coordinate levels.Extensive experiments show that FlashMesh achieves up to a 2xspeedup over standard autoregressive models while also improving generation fidelity. Our results demonstrate that structural priors in mesh data can be systematically harnessed to accelerate and enhance autoregressive generation.

Multi-modal Large Language Models (MLLMs) have achieved remarkable performance across a wide range of visual reasoning tasks, yet their vulnerability to safety risks remains a pressing concern. While prior research primarily focuses on jailbreak defenses that detect and refuse explicitly unsafe inputs, such approaches often overlook contextual safety, which requires models to distinguish subtle contextual differences between scenarios that may appear similar but diverge significantly in safety intent. In this work, we present MM-SafetyBench++, a carefully curated benchmark designed for contextual safety evaluation. Specifically, for each unsafe image-text pair, we construct a corresponding safe counterpart through minimal modifications that flip the user intent while preserving the underlying contextual meaning, enabling controlled evaluation of whether models can adapt their safety behaviors based on contextual understanding. Further, we introduce EchoSafe, a training-free framework that maintains a self-reflective memory bank to accumulate and retrieve safety insights from prior interactions. By integrating relevant past experiences into current prompts, EchoSafe enables context-aware reasoning and continual evolution of safety behavior during inference. Extensive experiments on various multi-modal safety benchmarks demonstrate that EchoSafe consistently achieves superior performance, establishing a strong baseline for advancing contextual safety in MLLMs. All benchmark data and code will be made publicly accessible upon acceptance.

Detecting visual anomalies in industrial inspection often requires training with only a few normal images per category. Recent few-shot methods achieve strong results employing foundation-model features, but typically rely on memory banks, auxiliary datasets, or multi-modal tuning of vision-language models. We therefore question whether such complexity is necessary given the feature representations of vision foundation models. To answer this question, we introduce SubspaceAD, a training-free method, that operates in two simple stages. First, patch-level features are extracted from a small set of normal images by a frozen DINOv2 backbone. Second, a Principal Component Analysis (PCA) model is fit to these features to estimate the low-dimensional subspace of normal variations. At inference, anomalies are detected via the reconstruction residual with respect to this subspace, producing interpretable and statistically grounded anomaly scores. Despite its simplicity, SubspaceAD achieves state-of-the-art performance across one-shot and few-shot settings without training, prompt tuning, or memory banks. In the one-shot anomaly detection setting, SubspaceAD achieves image-level and pixel-level AUROC of 98.0% and 97.6% on the MVTec-AD dataset, and 93.3% and 98.3% on the VisA dataset, respectively, surpassing prior state-of-the-art results.

Part-level pointing is important for fine-grained interaction and reasoning, yet existing Multimodal Large Language Models (MLLMs) remain limited to instance-level pointing. Part-level pointing presents unique challenges: annotation is costly, parts are long-tail distributed, and many are difficult to specify precisely in language. We introduce POinting at Parts (POP), a training-free, plug-and-play approach that addresses these challenges under a few-shot setup. POP fuses textual and visual attention maps with self-supervised visual correspondences from query image and few-shot examples. On average across the three evaluated datasets, POP achieves accuracy gains of up to 8.9 points in the one-shot setting and 16.4 points in the three-shot setting for the pointing-capable MLLMs--Qwen2.5-VL, Ovis2.5, and Molmo. Notably, even MLLMs without pointing capability benefit significantly from the proposed approach. These results establish a simple yet effective path toward fine-grained spatial grounding in MLLMs.

The research frontier in human pose prediction (HPP) is advancing toward continual test-time adaptation (TTA), where models must self-adapt to dynamic test distributions. To date, the homeostatic continual TTA remains the sole viable solution, which isolates the model parameters and update domain-sensitive ones. Despite mitigating full-body domain gaps, human anatomical heterogeneity (domain shifts often localize to specific regions) is ignored. This anatomical-agnostic approach forces uniform parameter adaptation across kinematically distinct segments, causing: over-adaptation of stable regions and under-adaptation of shift-prone articulations. To address it, we introduce TT-HA, a novel Test-Time Heterogeneous Adaptation that implicitly estimates domain changes for anatomical segments, and adapt the corresponding parameters. Building on human anatomy, TT-HA partitions parameters into five anatomical subsets using fisher information matrix-based parameters uncertainty analysis. During testing, TT-HA uses the instance normalization statistics and Earth Mover's Distance (EMD) to quantify segment-wise domain changes, dynamically determining which segment-specific parameters to adapt and to what extent. When substantial domain shifts are detected, TT-HA restores only affected segments to source-trained values, ensuring robust adaptation without full parameter resetting; minor shifts trigger the fine-tuning of corresponding parameters while preserving remaining ones. Experiments show TT-HA's superior full-body accuracy with greater limb error decrease than prior methods, proving its anatomically-targeted efficacy.

Vision-Language-Action (VLA) models have recently emerged, demonstrating strong generalization in robotic scene understanding and manipulation. However, when confronted with long-horizon tasks that require defined goal states, such as LEGO assembly or object rearrangement, existing VLA models still face challenges in coordinating long-horizon planning with precise manipulation.Therefore, we aim to endow a VLA model with the capability to infer the "how" process from the "what" outcomes, transforming goal states into executable procedures. In this paper, we introduce ManualVLA, a unified VLA framework built upon a Mixture-of-Transformers (MoT) architecture, enabling coherent collaboration between multimodal manual generation and action execution. Unlike prior VLA models that directly map sensory inputs to actions, we first equip ManualVLA with a planning expert that generates intermediate manuals consisting of images, visual prompts, and textual instructions. Building upon these multimodal manuals, we design a Manual Chain-of-Thought (ManualCoT) reasoning process that feeds them into the action expert, where each manual step provides explicit control conditions, while its latent representation offers implicit guidance for accurate manipulation. To alleviate the burden of data collection, we develop a high-fidelity digital-twin toolkit based on 3D Gaussian Splatting, which automatically generates manual data for planning expert training. ManualVLA demonstrates strong real-world performance, achieving an average success rate 32% higher than the previous hierarchical SOTA baseline on LEGO assembly and object rearrangement tasks.

We propose JUMP-Hand, a novel multi-view 3D hand reconstruction method that explicitly models probabilistic joint-wise uncertainty as a gating mechanism for multi-view fusion. Existing approaches usually rely on naive pooling or implicit attention, overlooking that each hand joint exhibits varying visibility and reliability across views. Such indiscriminate aggregation of noisy or unreliable information inevitably degrades overall performance. For instance, a joint severely occluded in one view might be clearly visible in another. To address this, JUMP-Hand draws inspiration from the Mixture of Experts (MoE) paradigm, treating each 2D view as a specialized expert. The key idea is that the reliability of each view expert is quantified through joint-wise uncertainty modeling, serving as an explicit gating signal to route experts' partial yet complementary clues for each joint in a coarse-to-fine reconstruction paradigm. In this design, uncertainty not only guides the uncertainty-aware triangulation for reliable 3D hand initialization during the coarse stage, but also acts as a gating signal during the refinement stage to adaptively aggregate multi-scale features from different view experts on a joint-wise basis, enabling robust 3D hand reconstruction. Extensive experiments on DexYCB-MV, HO3D-MV, and OakInk-MV demonstrate that our method consistently achieves state-of-the-art results, validating the effectiveness of the proposed method with joint-wise uncertainty gating for reliable 3D hand reconstruction. Code is available at https://github.com/HaohongKuang/JUMP-Hand.

Motion blur is a common degradation in dynamic imaging. Recent studies have moved beyond restoring a single sharp image from a blurred input and instead target blur decomposition: recovering a temporally continuous sharp video sequence from one motion-blurred image. Event cameras, with their microsecond temporal resolution, can effectively alleviate motion ambiguity. However, existing event-based methods often fail to explicitly model time-aligned event-image features. How to accurately exploit event data to reconstruct frames at different time instants remains largely underexplored. In this paper, we propose TSANet, an event-based blur-to-video decomposition method that time-specializes both event features and image features for alignment. Specifically, we introduce a Relative Time-Encoded Attention module that steers event features toward motion information relevant to a given target time, and a Timesurface Dynamic Warping module that warps image features into the spatial configuration corresponding to that time. With time-specialized motion and image features explicitly aligned at arbitrary query times, our framework can decompose a single blurred image into a high-frame-rate sharp video sequence. In addition, we collect a new dataset containing real events and high-quality color videos, and synthesize blurred inputs by averaging sharp frames to evaluate our method. Experiments on multiple datasets with both synthetic and real events demonstrate that our approach consistently outperforms previous state-of-the-art methods on the blur decomposition task.

Graph neural networks (GNNs) are increasingly applied to physical design tasks such as congestion prediction and wirelength estimation, yet progress is hindered by inconsistent circuit representations and the absence of controlled evaluation protocols. We present R2G (RTL-to-GDSII), a multi-view circuit-graph benchmark suite that standardizes five stage-aware views with information parity (every view encodes the same attribute set, differing only in where features attach) over 30 open-source IP cores (up to 106 nodes/edges). R2G provides an end-to-end DEF-to-graph pipeline spanning synthesis, placement, and routing stages, together with loaders, unified splits, domain metrics, and reproducible baselines. By decoupling representation choice from model choice, R2G isolates a confound that prior EDA and graph-ML benchmarks leave uncontrolled. In systematic studies with GINE, GAT, and ResGatedGCN, we find: (i) view choice dominates model choice, with Test R2 varying by more than 0.3 across representations for a fixed GNN; (ii) node-centric views generalize best across both placement and routing; and (iii) decoder-head depth (3-4 layers) is the primary accuracy driver, turning divergent training into near-perfect predictions (R2 > 0.99). Code and datasets: https://github.com/ShenShan123/R2G.

As Vision-Language Models (VLMs) move into interactive, multi-turn use, safety concerns intensify for multimodal multi-turn dialogue, which is characterized by concealment of malicious intent, contextual risk accumulation, and cross-modal joint risk. These characteristics limit the effectiveness of content moderation approaches designed for single-turn or single-modality settings. To address these limitations, we first construct the Multimodal Multi-turn Dialogue Safety (MMDS) dataset, comprising 4,484 annotated dialogues and a comprehensive risk taxonomy with 8 primary and 60 subdimensions. As part of MMDS construction, we introduce Multimodal Multi-turn Red Teaming (MMRT), an automated framework for generating unsafe multimodal multi-turn dialogues. We further propose LLaVAShield, which audits the safety of both user inputs and assistant responses under specified policy dimensions in multimodal multi-turn dialogues. Extensive experiments show that LLaVAShield significantly outperforms state-of-the-art VLMs and existing content moderation tools while demonstrating strong generalization and flexible policy adaptation. Additionally, we analyze vulnerabilities of mainstream VLMs to harmful inputs and evaluate the contribution of key components, advancing understanding of safety mechanisms in multimodal multi-turn dialogues.

Reconstructing photorealistic and animatable 4D head avatars from a single portrait image remains a fundamental challenge in computer vision. While diffusion models have enabled remarkable progress in image and video generation for avatar reconstruction, existing methods primarily rely on 2D priors and struggle to achieve consistent 3D geometry. We propose a novel framework that leverages geometry-aware diffusion to distill strong geometry priors for high-fidelity head avatar reconstruction. Our approach jointly synthesizes portrait images and corresponding surface normals, while a pose-free expression encoder captures implicit expression representations. Both synthesized images and expression latents are distilled into 3D Gaussian-based avatars, enabling photorealistic rendering with accurate geometry. Extensive experiments demonstrate that our method substantially outperforms state-of-the-art approaches in visual quality, expression fidelity, and cross-identity generalization, while supporting real-time rendering.

Synthesizing human motion has advanced rapidly, yet realistic hand motion and bimanual interaction remain underexplored. Whole-body models often miss the fine-grained cues that drive dexterous behavior, finger articulation, contact timing, and inter-hand coordination, and existing resources lack high-fidelity bimanual sequences that capture nuanced finger dynamics and collaboration. To fill this gap, we present HandX, a unified foundation spanning data, annotation, and evaluation. We consolidate and filter existing datasets for quality, and collect a new motion-capture dataset targeting underrepresented bimanual interactions with detailed finger dynamics. For scalable annotation, we introduce a decoupled paradigm that extracts representative motion features, e.g., contact events and finger flexion, and then leverages reasoning from large language models to produce fine-grained, semantically rich descriptions aligned with these features. Building on the resulting data and annotations, we benchmark diffusion and autoregressive models with versatile conditioning modes. Experiments demonstrate high-quality dexterous motion generation, supported by our newly proposed hand-focused metrics. We further observe clear scaling trends: larger models trained on larger, higher-quality datasets produce more semantically coherent bimanual motion. Our dataset is released to support future research.

Visual Active Tracking (VAT) aims to control cameras to follow a target in 3D space, which is critical for applications like drone navigation and security surveillance. However, it faces two key bottlenecks in real-world deployment: confusion from visually similar distractors caused by insufficient instance-level discrimination and severe failure under occlusions due to the absence of active planning. To address these, we propose OA-VAT, a unified pipeline with three complementary modules. First, a training-free Instance-Aware Offline Prototype Initialization aggregates multi-view augmented features via DINOv3 to construct discriminative instance prototypes, mitigating distractor confusion. Second, an Online Prototype Enhancement Tracker enhances prototypes online and integrates a confidence-aware Kalman filter for stable tracking under appearance and motion changes. Third, an Occlusion-Aware Trajectory Planner, trained on our new Planning-20k dataset, uses conditional diffusion to generate obstacle-avoiding paths for occlusion recovery. Experiments demonstrate OA-VAT achieves 0.93 average SR on UnrealCV (+2.2% vs. SOTA TrackVLA), 90.8% average CAR on real-world datasets (+12.1% vs. SOTA GC-VAT), and 81.6% TSR on a DJI Tello drone. Running at 35 FPS on an RTX 3090, it delivers robust, real-time performance for practical deployment. The code is available at https://github.com/SHWplus/OA-VAT.