CVPR 2026 论文列表

Multimodal Large Language Models (MLLMs) remain far from human-level performance in multi-view spatial reasoning, where models must establish object correspondences across view and infer coherent scene semantics. We analyze this limitation through the Transformation-Driven Visual Reasoning (TVR) task and find that Supervised Fine-Tuning (SFT) fails to capture cross-view consistency, whereas reinforcement learning (RL) fails to reliably identify key referential objects. To bridge this gap, we introduce multi-View Spatial TrAnsformation Reasoning (STAR-R1), a two-stage framework that combines process-supervised SFT with a referential-aware RL paradigm. STAR-R1 first learns structured spatial reasoning trajectories from high-quality CoTs and then uses fine-grained rewards on referential selection and answer correctness to encourage effective exploration and robust scene interpretation. Despite using only a small amount of high-quality training data, STAR-R1 surpasses state-of-the-art models with far more training data on the multi-view spatial understanding benchmarks TVR, MMSI-Bench, MindCube-Bench, and SPAR-Bench. Our study reveals the overlooked potential of RL in multi-view spatial understanding and points a way toward potentially achieving more human-like spatial reasoning in MLLMs.

We begin by developing a mathematical framework of neural differentiation, formulated at the level of individual neurons. This framework formalizes the principle that each neuron should acquire a distinct representational role within the network, thereby avoiding redundancy and maximizing collective expressivity. Differentiation is quantified through the Neural Differentiation Index (NDI), a loss-aware measure that characterizes neuron significance from geometric, informational, and curvature-based perspectives within a unified framework. The NDI enables a rigorous characterization of how strongly a neuron diverges from its peers in both function and importance, and supports theoretical guarantees: we establish formal bounds on the error increase under NDI-guided elimination, thereby providing provable safety margins for network compression. Building on this foundation, we introduce Neural Differentiation Pruning (NDP) as a practical instantiation. NDP leverages NDI to perform adaptive, training-time neuron sparsification, followed by targeted fine-tuning, guiding networks toward compact yet highly differentiated backbones. Although the terminology draws loose intuition from biological differentiation, the framework is fully mathematical and architecture-agnostic. Experiments on modern vision benchmarks and architectures show that NDP achieves substantial structured sparsity while maintaining--or even improving--accuracy and robustness, underscoring the practical impact of the differentiation framework.

Agentic vision-language models are increasingly trained to "think with images" by calling image operations. However, we show that high final-answer accuracy often hides unfaithful visual reasoning: models may invoke tools on irrelevant regions or ignore tool outputs entirely, yet still guess the correct answer. In this work, we first propose a faithfulness evaluation protocol that measures whether intermediate visual tool outputs (e.g., crops) actually contain the queried evidence. This reveals that recent visual agents achieve high final-answer accuracy but exhibit low rates of faithful tool-use on visual search benchmarks. We then introduce CodeV, a code-based visual agent trained with Tool-Aware Policy Optimization (TAPO). TAPO is a process-level RL framework that augments GRPO with dense rewards defined directly on visual tool inputs and outputs, rather than on chain-of-thought tokens, making supervision easier to verify and less susceptible to reward hacking. CodeV represents visual tools as executable Python code, and TAPO assigns step-wise rewards based solely on the question and tool output, encouraging both necessary and evidence-consistent tool use. In a two-stage SFT+RL pipeline, CodeV achieves competitive or superior accuracy while substantially increasing faithful tool-use rates on related visual search benchmarks. Beyond visual search, CodeV attains strong performance on a range of multimodal reasoning and math benchmarks, suggesting that explicitly supervising intermediate tool behavior is crucial for building trustworthy, agentic visual reasoning systems.

As the harm caused by fake news grows, the task of detecting and grounding multi-modal media manipulation (DGM4) is gaining more attention. Existing multimodal methods overlook fine-grained semantic alignment between visual and textual modalities, thereby limiting their ability to detect sophisticated and subtle cross-modal manipulations. To address this challenge, we present MaLSF, a novel Mask-aware Local Semantic Fusion framework that explicitly bridges words and pixels via mask-label pairs, enabling the model to perform precise reasoning over fine-grained cross-modal correspondences. MaLSF captures cross-modal local semantics through two key innovations: 1) A Bidirectional Cross-modal Verification Module (BCV) that identifies semantic conflicts between masked regions and associated labels via a bidirectional query mechanism; 2) A Hierarchical Semantic Aggregation (HSA) Module that adaptively aggregates multi-granularity local semantics into decoupled features for task-specific verification. In addition, to extract fine-grained mask-label pairs, we introduce a set of diverse mask-label pair extraction parsers. The proposed model is evaluated on multiple datasets and achieves state-of-the-art performance on both the DGM4 and multimodal fake news detection tasks. Extensive ablation studies and visualization results further verify its effectiveness and interpretability.

Frame selectoin is crucial due to high frame redundancy and limited context windows when applying Large Vision-Language Models (LVLMs) to long videos. Current methods typically select frames with high relevance to a given query, resulting a disjointed set of frames that disregard the narrative structure of video. In this paper, we introduce Wavelet-based Frame Selection by Detecting Semantic Boundary (WFS-SB), a training-free framework that presents a new perspective: effective video understanding hinges not only on high relevance but, more importantly, on capturing semantic shifts--pivotal moments of narrative change that are essential to comprehending the holistic storyline of video. However, a direct detection of abrupt changes in the query-frame similarity signal is often unreliable due to high-frequency noise arising from model uncertainty and transient visual variations.To address this, we leverage the wavelet transform, which provides an ideal solution through its multi-resolution analysis in both time and frequency domains. By applying this transform, we decompose the noisy signal into multiple scales and extract a clean semantic change signal from the coarsest scale. We identify the local extrema of this signal as semantic boundaries, which segment the video into coherent clips. Building on this, WFS-SB comprises a two-stage strategy: first, adaptively allocating a frame budget to each clip based on a composite importance score; and second, within each clip, employing the Maximal Marginal Relevance approach to select a diverse yet relevant set of frames. Extensive experiments show that WFS-SB significantly boosts LVLM performance, e.g., improving accuracy by 5.5% on VideoMME, 9.5% on MLVU, and 6.2% on LongVideoBench, consistently outperforming state-of-the-art methods. Our code is available at https://github.com/MAC-AutoML/WFS-SB.

We propose a method to reconstruct high-fidelity human avatars from multi-view video that can run on mobile devices. Many works can model high-quality Gaussian-based full-body avatars from multi-view video. However, these methods require heavy computation to obtain pose-dependent appearance, making deployment on mobile devices very difficult. Recent methods distill from pretrained models and model pose-dependent nonlinear Gaussian attributes by linearly combining global pose features with blendshapes. Although they can run on mobile devices, they suffer some loss of detail. We observe that nearby Gaussians are often highly correlated within a local region of the body, and can be linearly modeled with less error. Therefore, we use local linear blendshapes in small body parts to capture global nonlinear changes of Gaussian attributes. To further reduce computation and model size, we propose to remove blendshapes for Gaussians whose attributes change little, yielding a minimal blendshape representation. Our method is an end-to-end training method without a pretrained model. To make it run on multiple devices, we implement our method using WebGPU. Experiments show that our method can render high-quality human avatars with better details, and can reach 120 FPS at 2K resolution on mobile devices.

Image restoration aims to recover high quality images from inputs degraded by various factors, such as adverse weather, blur, or low light. While recent studies have shown remarkable progress across individual or unified restoration tasks, they still suffer from limited generalization and inefficiency when handling unknown or composite degradations. To address these limitations, we propose RAR, a Restore, Assess and Repeat process, that integrates Image Quality Assessment (IQA) and Image Restoration (IR) into a unified framework to iteratively and efficiently achieve high quality image restoration. Specifically, we introduce a restoration process that operates entirely in the latent domain to jointly perform degradation identification, image restoration, and quality verification. The resulting model is fully trainable end to end and allows for an all-in-one assess and restore approach that dynamically adapts the restoration process. Also, the tight integration of IQA and IR into a unified model minimizes the latency and information loss that typically arises from keeping the two modules disjoint, (e.g. during image and/or text decoding). Extensive experiments show that our approach consistent improvements under single, unknown and composite degradations, thereby establishing a new state-of-the-art.

Models such as VGGT and \pi^3 have shown strong multi-view 3D performance, but their heavy reliance on global self-attention results in high computational cost. Existing sparse-attention variants offer partial speedups, yet lack a systematic analysis of how global attention contributes to multi-view reasoning. In this paper, we first conduct an in-depth investigation of the global attention modules in VGGT and \pi^3 to better understand their roles. Our analysis reveals a clear division of roles in the alternating global-frame architecture: early global layers do not form meaningful correspondences, middle layers perform cross-view alignment, and last layers provide only minor refinements. Guided by these findings, we propose a training-free two-step acceleration scheme: (1) converting early global layers into frame attention, and (2) subsampling global attention by subsampling K/V over patch tokens with diagonal preservation and a mean-fill component. We instantiate this strategy on VGGT and \pi^3 and evaluate across standard pose and point-map benchmarks. Our method achieves substantial inference acceleration across different context lengths, yielding about 2xspeedup at 100 frames, 4--5xat 300 frames, and 8--10xat 800 frames, while matching or slightly improving the accuracy of the original models and remaining robust in extremely dense multi-view settings where prior sparse-attention baselines fail.

Bridging the gap between complex human instructions and precise 3D object grounding remains a significant challenge in vision and robotics. Existing 3D segmentation methods often struggle to interpret ambiguous, reasoning-based instructions, while 2D vision-language models that excel at such reasoning lack intrinsic 3D spatial understanding. In this paper, we introduce REALM, an innovative MLLM-agent framework that enables open-world reasoning-based segmentation without requiring extensive 3D-specific post-training. We perform segmentation directly on 3D Gaussian Splatting representations, capitalizing on their ability to render photorealistic novel views that are highly suitable for MLLM comprehension. As directly feeding one or more rendered views to the MLLM can lead to high sensitivity to viewpoint selection, we propose a novel Global-to-Local Spatial Grounding strategy. Specifically, multiple global views are first fed into the MLLM agent in parallel for coarse-level localization, aggregating responses to robustly identify the target object. Then, several close-up novel views of the object are synthesized to perform fine-grained local segmentation, yielding accurate and consistent 3D masks. Extensive experiments show that REALM achieves remarkable performance in interpreting both explicit and implicit instructions across LERF, 3D-OVS, and our newly introduced REALM3D benchmarks. Furthermore, our agent framework seamlessly supports a range of 3D interaction tasks, including object removal, replacement, and style transfer, demonstrating its practical utility and versatility.

Large-scale vision-language models (VLMs) such as CLIP have excelled in zero-shot image classification, yet they struggle to achieve the dense cross-modal alignment required by open-vocabulary dense perception (OVDP). While recent self-distillation methods address this by aligning dense features with the generalizable global semantics, a key question remains: how should such dense features be constructed to achieve optimal alignment? To address this, we propose DenseRC, a principled Dense Representations Construction framework that reconstructs CLIP for OVDP based on two key insights.First, by analyzing the internal semantics encoded in the global cls token, we identify that multi-layer value embeddings serve as an informative basis for dense features. Second, we reveal that spatial aggregation tends to amplify semantic misalignment. Motivated by this, we design a lightweight Head-Selective Gating (HSG) module that adaptively reweights feature heads according to their intrinsic heterogeneity, enabling discriminative and alignment-friendly dense representations construction. Extensive experiments demonstrate that DenseRC delivers consistent and substantial gains across OVDP tasks including object detection and semantic segmentation, setting new state-of-the-art performance on multiple benchmarks.

Composed Image Retrieval (CIR) aims to retrieve target images by integrating a reference image with a corresponding modification text. CIR requires jointly considering the explicit semantics specified in the query and the implicit semantics embedded within its bi-modal composition. Recent training-free Zero-Shot CIR (ZS-CIR) methods leverage Multimodal Large Language Models (MLLMs) to generate detailed target descriptions, converting the implicit information into explicit textual expressions. However, these methods rely heavily on the textual modality and fail to capture the fuzzy retrieval nature that requires considering diverse combinations of candidates. This leads to reduced diversity and accuracy in retrieval results. To address this limitation, we propose a novel training-free method, Geodesic Mixup-based Implicit semantic eXpansion and Explicit semantic Re-ranking for ZS-CIR (G-MIXER). G-MIXER constructs composed query features that reflect the implicit semantics of reference image-text pairs through geodesic mixup over a range of mixup ratios, and builds a diverse candidate set. The generated candidates are then re-ranked using explicit semantics derived from MLLMs, improving both retrieval diversity and accuracy. Our proposed G-MIXER achieves state-of-the-art performance across multiple ZS-CIR benchmarks, effectively handling both implicit and explicit semantics without additional training. Our code will be available at https://github.com/maya0395/gmixer.

Multimodal image synthesis has made significant progress, yet most editing methods still rely on textual instructions, which are less direct than visual guidance. Recently, a new paradigm edits one image using another as reference, enabling more intuitive manipulation through visual exemplars. We formalize this setting as cross-image editing, where a source image is modified under one or more visual references. We propose OrionEdit, a unified framework that regulates editing via symmetric orthogonal subspace disentanglement and reverse-causal attention with information-flow masks enforcing unidirectional latent dependencies. Built on standard diffusion backbones, OrionEdit enables zero-shot multi-reference editing and outperforms open-source baselines, approaching proprietary models in fidelity and disentanglement. The model is available at https://github.com/cityuhkai/OrionEdit.

The new era has witnessed a remarkable capability to extend Vision-Language Models (VLMs) for tackling tasks of video understanding. While current VLMs excel at event- or story-level understanding, their ability to capture fine-grained motion details remains limited, primarily due to their focus on high-level static semantic structures and macro-event logic. In contrast, Video Diffusion Models (VDMs) are adept at modeling dynamic motion patterns, benefiting from large-scale video data and the intrinsic requirement of temporal generation. In this paper, we introduce MotionEnhancer, a novel approach that leverages motion priors distilled from a powerful video diffusion model as auxiliary supervision to enhance the motion understanding capability of a VLM via attention alignment. MotionEnhancer comprises two simple parameter-free modules, Motion-sensitive Head Selection (MHS) and Motion-salient Text Token Identification (MTTI), to directly extract and optimize motion-related attentions from the VDM in a computation-only manner. MotionEnhancer provides a scalable solution for motion understanding without additional training parameters, modifications to existing architectures, or tool calling. Extensive experiments demonstrate that MotionEnhancer can achieve consistent improvements over state-of-the-art VLMs on two motion-level video understanding benchmarks, especially on motion-related metrics.

Domain Generalization Semantic Segmentation (DGSS) in spectral remote sensing is severely challenged by spectral shifts across diverse acquisition conditions, which cause significant performance degradation for models deployed in unseen domains. While fine-tuning foundation models is a promising direction, existing methods employ global, homogeneous adjustments. This "one-size-fits-all" tuning struggles with the spatial heterogeneity of land cover, causing semantic confusion. We argue that the key to robust DGSS lies not in a single global adaptation, but in performing fine-grained, spatially-adaptive refinement of a foundation model's features. To achieve this, we propose SpectralMoE, a novel fine-tuning framework for DGSS. It operationalizes this principle by utilizing a Mixture-of-Experts (MoE) architecture to perform local precise refinement on the foundation model's features, incorporating depth features estimated from selected RGB bands of the spectral remote sensing imagery to guide the fine-tuning process. Specifically, SpectralMoE employs a dual-gated MoE architecture that independently routes visual and depth features to top-k selected experts for specialized refinement, enabling modality-specific adjustments. A subsequent cross-attention mechanism then judiciously fuses the refined structural cues into the visual stream, mitigating semantic ambiguities caused by spectral variations. Extensive experiments show that SpectralMoE sets a new state-of-the-art on multiple DGSS benchmarks across hyperspectral, multispectral, and RGB remote sensing imagery.

Text-video retrieval tasks have seen significant improvements due to the recent development of large-scale vision-language pre-trained models. Traditional methods primarily focus on video representations or cross-modal alignment, while recent works shift toward enriching text expressiveness to better match the rich semantics in videos. However, these methods use only interactions between text and frames/video, and ignore rich interactions among the internal frames within a video, so the final expanded text cannot capture frame contextual information, leading to disparities between text and video. In response, we introduce Energy-Aware Fine-Grained Relationship Learning Network (EagleNet) to generate accurate and context-aware enriched text embeddings. Specifically, the proposed Fine-Grained Relationship Learning mechanism (FRL) first constructs a text-frame graph by the generated text candidates and frames, then learns relationships among texts and frames, which are finally used to aggregate text candidates into an enriched text embedding that incorporates frame contextual information. To further improve fine-grained relationship learning in FRL, we design Energy-Aware Matching (EAM) to model the energy of text-frame interactions and thus accurately capture the distribution of real text-video pairs. Moreover, for more effective cross-modal alignment and stable training, we replace the conventional softmax-based contrastive loss with the sigmoid loss. Extensive experiments have demonstrated the superiority of EagleNet across MSRVTT, DiDeMo, MSVD, and VATEX.

Multimodal Large Language Models (MLLMs) are increasingly deployed in multi-image scenarios requiring complex reasoning across visual contexts. However, current MLLMs remain fundamentally limited by object hallucination--generating plausible yet factually inconsistent descriptions about objects. Existing benchmarks, designed primarily for single-image settings or providing only high-level multi-image assessments, cannot systematically diagnose how visual complexity and reasoning demands trigger hallucination. To address this gap, we introduce MIOH, a fine-grained multi-image object hallucination benchmark that systematically evaluates object hallucination across four foundational tasks (existence, counting, attribute, position) through three multi-image reasoning patterns (comprehensive, comparative, selective) under three controlled adversarial pressures (visual context scale, perceptual difficulty, contextual bias). Through evaluation of 29 models, we reveal that even state-of-the-art systems like GPT-5 and Gemini-2.5-Pro exhibit distinct failure patterns across different reasoning patterns and tasks. Our evaluation reveals that hallucination stems not merely from perceptual failures but from integration-stage limitations when maintaining object representations across multiple images. MIOH provides a controlled framework for analyzing multi-image object hallucination and serves as a critical evaluation tool for developing more reliable multimodal AI systems.

Contrastive vision-language models (VLMs) such as CLIP achieve strong zero-shot recognition yet remain vulnerable to spurious correlations, particularly background over-reliance. We introduce Cluster-based Concept Importance (CCI), a novel interpretability method that uses CLIP's own patch embeddings to group spatial patches into semantically coherent clusters, masking them, and evaluating relative changes in model predictions. CCI sets a new state of the art on faithfulness benchmarks, surpassing prior methods by large margins; for example, it yields more than a twofold improvement on the deletion-AUC metric for MS COCO retrieval. We further propose that CCI when combined with GroundedSAM, automatically categorizes predictions as foreground or background-driven, providing a crucial diagnostic ability. Existing benchmarks such as CounterAnimals, however, rely solely on accuracy and implicitly attribute all performance degradation to background correlations. Our analysis shows this assumption to be incomplete, since many errors arise from viewpoint variation, scale shifts, and fine-grained object confusions. To disentangle these effects, we introduce COVAR, a benchmark that systematically varies object foregrounds and backgrounds. Leveraging CCI with COVAR, we conduct a comprehensive evaluation of eighteen CLIP variants, providing both methodological advances and empirical evidence that chart a path toward more robust vision-language models.

Visual-language models (VLMs) excel at data mappings, but real-world document heterogeneity and unstructuredness disrupt the consistency of cross-modal embeddings. Recent late-interaction methods enhance image-text alignment through multi-vector representations, yet traditional training with limited samples and static strategies cannot adapt to the model's dynamic evolution, causing cross-modal retrieval confusion. To overcome this, we introduce Evo-Retriever, a retrieval framework featuring an LLM-guided curriculum evolution built upon a novel Viewpoint-Pathway collaboration. First, we employ multi-view image alignment to enhance fine-grained matching via multi-scale and multi-directional perspectives. Then, a bidirectional contrastive learning strategy generates "hard queries" and establishes complementary learning paths for visual and textual disambiguation to rebalance supervision. Finally, the model-state summary from the above collaboration is fed into an LLM meta-controller, which adaptively adjusts the training curriculum using expert knowledge to promote the model's evolution. On ViDoRe V2 and MMEB (VisDoc), Evo-Retriever achieves state-of-the-art performance, with nDCG@5 scores of 65.2 % and 77.1 %.

We introduce a lifelong imitation learning framework that enables continual policy refinement across sequential tasks under realistic memory and data constraints. Our approach departs from conventional experience replay by operating entirely in a multimodal latent space, where compact representations of visual, linguistic, and robot's state information are stored and reused to support future learning. To further stabilize adaptation, we introduce an incremental feature adjustment mechanism that regularizes the evolution of task embeddings through an angular margin constraint, preserving inter-task distinctiveness. Our method establishes a new state of the art in the LIBERO benchmarks, achieving 10-17 point gains in AUC and up to 65% less forgetting compared to previous leading methods. Ablation studies confirm the effectiveness of each component, showing consistent gains over alternative strategies. The code is available at: https://github.com/yfqi/lifelong_mlr_ifa.