CVPR 2026 论文列表

Intrinsic image decomposition aims to separate images into physical components such as albedo, depth, normals, and illumination. While recent diffusion- and transformer-based models benefit from paired supervision from synthetic datasets, their generalization to diverse, real-world scenarios remains challenging. We propose ReasonX, a novel framework that leverages a multimodal large language model (MLLM) as a perceptual judge providing relative intrinsic comparisons, and uses these comparisons as GRPO rewards for fine-tuning intrinsic decomposition models on unlabeled, in-the-wild images. Unlike RL methods for generative models, our framework aligns conditional intrinsic predictors by rewarding agreement between the judge's relational assessments and analytically derived relations from the model's outputs. ReasonX is model-agnostic and can be applied to different intrinsic predictors. Across multiple base architectures and modalities, ReasonX yields significant improvements, including 9-25% WHDR reduction on IIW albedo and up to 46% depth accuracy gains on ETH3D, highlighting the promise of MLLM-guided comparative supervision to bridge low- and high-level vision reasoning.

Multimodal large language models have demonstrated remarkable capabilities in 2D vision, motivating their extension to 3D scene understanding. Recent studies represent 3D scenes as 3D spatial videos composed of image sequences with depth and camera pose information, enabling pre-trained video-language models to perform 3D reasoning tasks. However, the large number of visual tokens in spatial videos remains a major bottleneck for efficient inference. Existing pruning methods overlook the view consistency of spatial videos and the spatial diversity of the remaining tokens, which prevents them from effectively removing inter-frame redundancy and preserving scene completeness. In this paper, we propose Geo3DPruner, a Geometry-Guided 3D Visual Token Pruning framework. Geo3DPruner first models cross-frame relevance through geometry-aware global attention, and then performs a two-stage pruning process. The intra-voxel stage selects representative multi-view features within each voxel, while the inter-voxel stage preserves spatial diversity by selecting a globally distributed subset of voxels. Extensive experiments on multiple 3D scene understanding benchmarks demonstrate that Geo3DPruner retains over 90% of the original performance while pruning 90% of visual tokens, significantly outperforming existing text-guided and vision-guided pruning methods.

Vision-Language-Action (VLA) tasks require reasoning over complex visual scenes and executing adaptive actions in dynamic environments. While recent studies on reasoning VLAs show that explicit chain-of-thought (CoT) can improve generalization, they suffer from high inference latency due to lengthy reasoning traces. We propose Fast-ThinkAct, an efficient reasoning framework that achieves compact yet performant planning through verbalizable latent reasoning. Fast-ThinkAct learns to reason efficiently with latent CoTs by distilling from a teacher, driven by a preference-guided objective to align manipulation trajectories that transfers both linguistic and visual planning capabilities for embodied control. This enables reasoning-enhanced policy learning that effectively connects compact reasoning to action execution. Extensive experiments across diverse embodied manipulation and reasoning benchmarks demonstrate that Fast-ThinkAct achieves strong performance with up to 89.3% reduced inference latency over state-of-the-art reasoning VLAs, while maintaining effective long-horizon planning, few-shot adaptation, and failure recovery.

Non-human primates are our closest living relatives, and analyzing their behavior is central to research in cognition, evolution, and conservation. Computer vision could greatly aid this research, but existing methods often rely on human-centric pretrained models and focus on single datasets, which limits generalization. We address this limitation by shifting from a model-centric to a data-centric approach and introduce PriVi, a large-scale primate-centric video pretraining dataset. PriVi contains 424 hours of curated video, combining 174 hours from behavioral research across 11 settings with 250 hours of diverse web-sourced footage, assembled through a scalable data curation pipeline. We continue pretraining V-JEPA, a large-scale video model, on PriVi to learn primate-specific representations and evaluate it using a lightweight frozen classifier. Across four benchmark datasets - ChimpACT, PanAf500, BaboonLand, and ChimpBehave - our approach consistently outperforms prior work, including fully finetuned baselines, and scales favorably with fewer labels. These results demonstrate for the first time that domain-level pretraining, where pretraining is conducted on similar data but not the target dataset itself, works for video models. Our primate-centric pretraining substantially improves data efficiency and generalization, making it a promising approach for low-label applications. Dataset, code, and models are available at https://privi.eckerlab.org.

Recent advances in 3D scene-language understanding have leveraged Large Language Models (LLMs) for 3D reasoning by transferring their general reasoning ability to 3D multi-modal contexts. However, existing methods typically adopt standard decoders from language modeling, which rely on a causal attention mask. This design introduces two fundamental conflicts in 3D scene understanding: sequential bias among order-agnostic 3D objects and restricted object-instruction attention, hindering task-specific reasoning. To overcome these limitations, we propose 3D Spatial Language Instruction Mask (3D-SLIM), an effective masking strategy that replaces the causal mask with an adaptive attention mask tailored to the spatial structure of 3D scenes. Our 3D-SLIM introduces two key components: a Geometry-adaptive Mask that constrains attention based on spatial density rather than token order, and an Instruction-aware Mask that enables object tokens to directly access instruction context. This design allows the model to process objects based on their spatial relationships while being guided by the user's task. 3D-SLIM is simple, requires no architectural modifications, and adds no extra parameters, yet it yields substantial performance improvements across diverse 3D scene-language tasks. Extensive experiments across multiple benchmarks and LLM baselines validate its effectiveness and underscore the critical role of decoder design in 3D multi-modal reasoning.

Multimodal Sentiment Analysis (MSA) integrates language, visual, and acoustic modalities to infer human sentiment. Most existing methods either focus on globally shared representations or modality-specific features, while overlooking signals that are shared only by certain modality pairs. This limits the expressiveness and discriminative power of multimodal representations. To address this limitation, we propose a Tri-Subspace Disentanglement (TSD) framework that explicitly factorizes features into three complementary subspaces: a common subspace capturing global consistency, submodally-shared subspaces modeling pairwise cross-modal synergies, and private subspaces preserving modality-specific cues. To keep these subspaces pure and independent, we introduce a decoupling supervisor together with structured regularization losses. We further design a Subspace-Aware Cross-Attention (SACA) fusion module that adaptively models and integrates information from the three subspaces to obtain richer and more robust representations. Experiments on CMU-MOSI and CMU-MOSEI demonstrate that TSD achieves state-of-the-art performance across all key metrics, reaching 0.691 MAE on CMU-MOSI and 54.6% Acc-7 on CMU-MOSEI under the unaligned setting, and also transfers well to multimodal intent recognition tasks. Ablation studies confirm that tri-subspaces disentanglement and SACA jointly enhance the modeling of multi-granular cross-modal sentiment cues.

We present AToken, the first unified visual tokenizer that achieves both high-fidelity reconstruction and semantic understanding across images, videos, and 3D assets. Unlike existing tokenizers that specialize in either reconstruction or understanding for single modalities, AToken encodes these diverse visual inputs into a shared 4D latent space, unifying both tasks and modalities in a single framework. Specifically, we introduce a pure transformer architecture with 4D rotary position embeddings to process visual inputs of arbitrary resolutions and temporal durations. To ensure stable training, we introduce an adversarial-free training objective that combines perceptual and Gram matrix losses, achieving state-of-the-art reconstruction quality. By employing a progressive training curriculum, AToken gradually expands from single images, videos, and 3D, and supports both continuous and discrete latent tokens. AToken achieves 0.21 rFID with 82.2% ImageNet accuracy for images, 3.01 rFVD with 40.2% MSRVTT retrieval for videos, and 28.28 PSNR with 90.9% classification accuracy for 3D.. In downstream applications, AToken enables both visual generation tasks (e.g., image generation with continuous and discrete tokens, text-to-video generation, image-to-3D synthesis) and understanding tasks (e.g., multimodal LLMs), achieving competitive performance across all benchmarks. These results shed light on the next-generation multimodal AI systems built upon unified visual tokenization.

Automatic album organization has been studied extensively over the past decades due to significant progress in digital photography. Recent vision-language models (VLMs) have shown strong performance on multi-image understanding, making them natural candidates for automating album organization workflows. While VLMs' abilities in multi-image understanding have been widely studied, their performance on album organization remains underexplored. To bridge this gap, we introduce AlbumBench, the first comprehensive benchmark for automatic album organization. Specifically, we (1) define album organization tasks as photo selection for album-specific user objectives, photo rating according to how well user intents are fulfilled, and album-specific photo grouping given a user query which requires contextual understanding of the album; (2) establish AlbumBench, a benchmark dataset containing 27051 images across 641 albums with 5 annotations per image; and (3) evaluate mainstream open-source and proprietary VLMs on AlbumBench. We show that AlbumBench presents unique challenges compared to traditional multi-image understanding benchmarks due to its requirement for understanding album context and user intent. Our findings reveal a significant performance gap between open-source and proprietary VLMs on album organization tasks. Despite this gap, even the best-performing proprietary models sometimes struggle with tasks that humans find relatively easy. We hope that AlbumBench can serve as a foundation for unifying album organization research and motivate improvements in VLMs' performance on these tasks.

Recent progress in self- and weakly supervised occupancy estimation has largely relied on 2D projection or rendering-based supervision, which suffers from geometric inconsistencies and severe depth bleeding.We thus introduce ShelfOcc, a vision-only method that overcomes these limitations without relying on LiDAR.ShelfOcc brings supervision into native 3D space by generating metrically consistent semantic voxel labels from video, enabling true 3D supervision without any additional sensors or manual 3D annotations.While recent vision-based 3D geometry foundation models provide a promising source of prior knowledge, they do not work out of the box as a prediction due to sparse or noisy and inconsistent geometry, especially in dynamic driving scenes.Our method introduces a dedicated framework that mitigates these issues by filtering and accumulating static geometry consistently across frames, handling dynamic content and propagating semantic information into a stable voxel representation.This data-centric shift in supervision for weakly/shelf-supervised occupancy estimation allows the use of essentially any SOTA occupancy model architecture without relying on LiDAR data.We argue that such high-quality supervision is essential for robust occupancy learning and constitutes an important complementary avenue to architectural innovation.On the Occ3D-nuScenes benchmark, ShelfOcc substantially outperforms all previous weakly/shelf-supervised methods (up to a 34% relative improvement), establishing a new data-driven direction for LiDAR-free 3D scene understanding.

Vision Transformers often rely on spurious background correlations rather than foreground object features. While prior model pruning approaches focus solely on improving accuracy, they lack interpretability and fail to verify whether predictions are actually made by focusing on the main foreground object, providing no causal validation of which components drive spurious behavior. We introduce Causal Information Gain Mechanistic Attribution (CIGMA), a general framework for explaining the internal computation of Vision Transformers. CIGMA provides a mechanistic, information theoretic explanation by quantifying the importance of each attention head and determining whether it supports the main object or routes spurious background cues. It ranks attention heads by measuring object versus context reliance with Jensen Shannon based information gain computed from the model's full predictive distributions after two complementary edits, removing the object region and removing the surrounding context, which reveals a spurious subnet that carries background signals and a complementary set of evidence aligned heads. Evaluated on CIFAR-10, CIFAR-100, and Tiny-ImageNet across three VLM architectures (InternVL2-26B, LLaVA-1.6, LLaVA-1.5-13B), CIGMA improves accuracy by 7.6 to 24.8 percentage points over unmodified models while reducing background reliance by 79.5% to 88.1%, substantially outperforming all baselines, demonstrating that causal head-level interventions enable more effective spurious correlation mitigation than token pruning or retraining approaches.

Accurately estimating task progress is critical for embodied agents to plan and execute long-horizon, multi-step tasks. Despite promising advances, existing Vision-Language Models (VLMs) based methods primarily leverage their video understanding capabilities, while neglecting their complex reasoning potential. Furthermore, processing long video trajectories with VLMs is computationally prohibitive for real-world deployment. To address these challenges, we propose the Recurrent Reasoning Vision-Language Model (\text R ^2VLM). Our model features a recurrent reasoning framework that processes local video snippets iteratively, maintaining a global context through an evolving Chain of Thought (CoT). This CoT explicitly records task decomposition, key steps, and their completion status, enabling the model to reason about complex temporal dependencies. This design avoids the high cost of processing long videos while preserving essential reasoning capabilities. We train \text R ^2VLM on large-scale, automatically generated datasets from ALFRED and Ego4D. Extensive experiments on progress estimation and downstream applications, including progress-enhanced policy learning, reward modeling for reinforcement learning, and proactive assistance, demonstrate that \text R ^2VLM achieves strong performance and generalization, achieving a new state-of-the-art in long-horizon task progress estimation. The models and benchmarks are publicly available at https://huggingface.co/collections/zhangyuelin/r2vlm.

Taking advantage of large-scale data and pretrained language models, Video Large Language Models (Video-LLMs) have shown strong capabilities in answering video questions. However, most existing efforts focus on improving performance, with limited attention to understanding their internal mechanisms. This paper aims to bridge this gap through a systematic empirical study. To interpret existing VideoLLMs, we adopt attention knockouts as our primary analytical tool and design three variants: Video Temporal Knockout, Video Spatial Knockout, and Language-to-Video Knockout. Then, we apply these three knockouts on different numbers of layers (window of layers). By carefully controlling the window of layers and types of knockouts, we provide two settings: a global setting and a fine-grained setting. Our study reveals three key findings: (1) Global setting indicates Video information extraction primarily occurs in early layers, forming a clear two-stage process--lower layers focus on perceptual encoding, while higher layers handle abstract reasoning; (2) In the fine-grained setting, certain intermediate layers exert an outsized impact on video question answering, acting as critical outliers, whereas most other layers contribute minimally; (3) In both settings, we observe that spatial-temporal modeling relies more on language-guided retrieval than on intra- and inter-frame self-attention among video tokens, despite the latter's high computational cost. Finally, we demonstrate that these insights can be leveraged to reduce attention computation in Video-LLMs. To our knowledge, this is the first work to systematically uncover how Video-LLMs internally process and understand video content, offering interpretability and efficiency perspectives for future research.

Despite significant progress in Multi-modal Large Language Models (MLLMs), their clinical reasoning capacity for multi-modal diagnosis remains largely unexamined. Current benchmarks, mostly single-modality data, can't evaluate progressive reasoning and cross-modal integration essential for clinical practice. We introduce the Cross-Modality Progressive Clinical Reasoning (X-PCR) benchmark, the first comprehensive evaluation of MLLMs through a complete ophthalmology diagnostic workflow, with two reasoning tasks: 1) a six-stage progressive reasoning chain spanning image quality assessment to clinical decision-making, and 2) a cross-modality reasoning task integrating six imaging modalities. The benchmark comprises 26,415 images and 177,868 expert-verified VQA pairs curated from 51 public datasets, covering 52 ophthalmic diseases. Evaluation of 21 MLLMs reveals critical gaps in progressive reasoning and cross-modal integration. Dataset and code: https://github.com/CVI-SZU/X-PCR.

Large vision-language models (VLMs) often benefit from intermediate visual cues, either injected via external tools or generated as latent visual tokens during reasoning, but these mechanisms still overlook fine-grained visual evidence (e.g., polylines in charts), generalize poorly across domains, and incur high inference-time cost. In this paper, we propose Bi-directional Perceptual Shaping (BiPS), which transforms question-conditioned masked views into bidirectional where-to-look signals that shape perception during training. BiPS first applies a KL-consistency constraint between the original image and an evidence-preserving view that keeps only question-relevant regions, encouraging coarse but complete coverage of supporting pixels. It then applies a KL-separation constraint between the original and an evidence-ablated view where critical pixels are masked so the image no longer supports the original answer, discouraging text-only shortcuts (i.e., answering from text alone) and enforcing fine-grained visual reliance. Across eight benchmarks, BiPS boosts Qwen2.5-VL-7B by 8.2% on average and shows strong out-of-domain generalization to unseen datasets and image types.

Region-instructed layout control in text-to-image generation is highly practical, yet existing methods suffer from limitations: (i) training-based approaches inherit data bias and often degrade image quality, and (ii) current techniques struggle with occlusion order, limiting real-world usability. To address these issues, we propose LayerBind. By modeling regional generation as distinct layers and binding them during the generation, our method enables precise regional and occlusion controllability. Our motivation stems from the observation that spatial layout and occlusion are established at a very early denoising stage, suggesting that rearranging the early latent structure is sufficient to modify the final output. Building on this, we structure the scheme into two phases: instance initialization and subsequent semantic nursing. (1) First, leveraging the contextual sharing mechanism in multimodal joint attention, Layer-wise Instance Initialization creates per-instance branches that attend to their own regions while anchoring to the shared background. At a designated early step, these branches are fused according to the layer order to form a unified latent with a pre-established layout. (2) Then, Layer-wise Semantic Nursing reinforces regional details and maintains the occlusion order via a layer-wise attention enhancement. Specifically, a sequential layered attention path operates alongside the standard global path, with updates composited under a layer-transparency scheduler. LayerBind is training-free and plug-and-play, serving as a regional and occlusion controller across Diffusion Transformers. It also supports editable workflows, allowing for flexible modifications like changing instances or rearranging visible orders. Experimental results demonstrate LayerBind's effectiveness, highlighting its potential for creative applications. Project page: https://littlefatshiba.github.io/layerbind-page

The proliferation of low-power intelligent processors with integrated Neural Processing Units (NPUs), called mNPUs, has created new opportunities for on-device generative AI, benefitting end devices like smart wearables and small robots. However, deploying Vision-Language Models (VLMs) on mNPUs is severely hindered by stringent memory constraints and limited operator support. To bridge this critical gap, we propose mVLM, the first lightweight-oriented VLM architecture designed for mNPUs. It is comprised of our proposed OverMod encoder and AttSSM decoder. OverMod is a lightweight dynamic convolutional network inspired by biomimetic vision, incorporating our novel Global Spatial Modulation mechanism to enable adaptive, high-fidelity feature extraction using only NPU-friendly operators. AttSSM leverages a highly efficient State Space Model (SSM) core, augmented with multi-scale feature fusion and Global Context Dynamic Modulation mechanism, to perform robust sequential modeling. Furthermore, we introduce a coordinated full-parameter quantization strategy that preserves precision across the encoder-decoder boundary, alongside hand-optimized operators for unsupported modules like SSMs. mVLM achieves a competitive CIDEr score of 117.8 on the COCO Karpathy test split and, for the first time, demonstrates the feasibility of millisecond-level VLM inference on a mNPU platform.

Vision-language models (VLMs) have transformed multimodal reasoning, but feeding hundreds of visual patch tokens to LLMs incurs quadratic computational costs, straining memory and context windows. Traditional approaches face a trade-off: continuous compression dilutes high-level semantics like object identities, while discrete quantization loses granular details such as textures. We challenge this by introducing **HTC-VLM**, a hybrid framework that disentangles semantics and appearance through dual channels, i.e., a continuous pathway for fine-grained details via ViT patches and a discrete pathway for symbolic anchors using MGVQ quantization projected to four tokens. These are fused into a 580-token hybrid sequence and compressed to one token via a disentanglement attention mask and a `<voco>` bottleneck, ensuring efficient, grounded representations.HTC-VLM achieves an average performance retention of **87.2%** across seven benchmarks (GQA, VQAv2, MMBench, MME, POPE, SEED-Bench, ScienceQA-Image), outperforming the leading continuous baseline at **81.0%** with a 580-to-1 compression ratio. Attention analyses show the compressed token prioritizes the discrete anchor, validating its semantic guidance. Our work demonstrates that a minimalist hybrid can resolve the efficiency-fidelity dilemma, advancing scalable VLMs.

Vision-Language Models (VLM) exhibit strong reasoning capabilities, showing promise for end-to-end autonomous driving systems. Chain-of-Thought (CoT), as VLM's widely used reasoning strategy, is facing critical challenges. Existing textual CoT has a large gap between text semantic space and trajectory physical space. Although the recent approach utilizes future image to replace text as CoT process, it lacks clear planning-oriented objective guidance to generate images with accurate scene evolution. To address these, we innovatively propose MindDriver, a progressive multimodal reasoning framework that enables VLM to imitate human-like progressive thinking for autonomous driving.MindDriver presents semantic understanding, semantic-to-physical space imagination, and physical-space trajectory planning.To achieve aligned reasoning processes in MindDriver, we develop a feedback-guided automatic data annotation pipeline to generate aligned multimodal reasoning training data. Furthermore, we develop a progressive reinforcement fine-tuningmethod to optimize the alignment through progressive high-level reward-based learning.MindDriver demonstrates superior performance in both nuScences open-loop and Bench2Drive closed-loop evaluation. Codes: https://github.com/hotdogcheesewhite/MindDriver.

Real-world image restoration aims to restore high-quality (HQ) images from degraded low-quality (LQ) inputs captured under uncontrolled conditions. Existing methods typically depend on ground-truth (GT) supervision, assuming that GT provides perfect reference quality. However, GT can still contain images with inconsistent perceptual fidelity, causing models to converge to the average quality level of the training data rather than achieving the highest perceptual quality attainable. To address these problems, we propose a novel framework, termed IQPIR, that introduces an Image Quality Prior (IQP)--extracted from pre-trained No-Reference Image Quality Assessment (NR-IQA) models--to guide the restoration process toward perceptually optimal outputs explicitly. Our approach synergistically integrates IQP with a learned codebook prior through three key mechanisms:(1) a quality-conditioned Transformer, where NR-IQA-derived scores serve as conditioning signals to steer the predicted representation toward maximal perceptual quality. This design provides a plug-and-play enhancement compatible with existing restoration architectures without structural modification; and(2) a dual-branch codebook structure, which disentangles common and HQ-specific features, ensuring a comprehensive representation of both generic structural information and quality-sensitive attributes; and (3) a discrete representation-based quality optimization strategy, which mitigates over-optimization effects commonly observed in continuous latent spaces. Extensive experiments on real-world image restoration demonstrate that our method not only surpasses cutting-edge methods but also serves as a generalizable quality-guided enhancement strategy for existing methods. The code is available at https://github.com/fengyang1399-pixel/IQPIR.