CVPR 2026 论文列表

Knowledge distillation (KD) has been widely applied in semantic segmentation to compress large models, but conventional approaches primarily preserve in-domain accuracy while neglecting out-of-domain generalization, which is essential under distribution shifts. This limitation becomes more severe with the emergence of vision foundation models (VFMs): although VFMs exhibit strong robustness on unseen data, distilling them with conventional KD often compromises this ability. We propose Generalizable Knowledge Distillation (GKD), a multi-stage framework that explicitly enhances generalization. GKD decouples representation learning from task learning. In the first stage, the student acquires domain-agnostic representations through selective feature distillation, and in the second stage, these representations are frozen for task adaptation, thereby mitigating overfitting to visible domains. To further support transfer, we introduce a query-based soft distillation mechanism, where student features act as queries to teacher representations to selectively retrieve transferable spatial knowledge from VFMs. Extensive experiments on five domain generalization benchmarks demonstrate that GKD consistently outperforms existing KD methods, achieving average gains of +1.9% in foundation-to-foundation (F2F) and +10.6% in foundation-to-local (F2L) distillation. The code will be available at https://github.com/Younger-hua/GKD.

Cause-and-effect reasoning in video is a significant challenge for Vision-Language Models (VLMs), as it requires going beyond surface-level perception to a deeper understanding of causal mechanisms. However, existing benchmarks rarely provide the fine-grained, grounded evidence needed to rigorously evaluate this capability. To address this gap, we introduce CaST-Bench, a benchmark for Causal Chain-Grounded Spatio-Temporal Video Reasoning. CaST-Bench presents complex causal questions that require models to identify and localize a chain of multiple spatio-temporal evidences. Through a human-AI collaborative pipeline, we construct a high-quality dataset of 2,066 questions over 1,015 videos, with causal chains annotated by temporal segments and bounding-box tracks. Furthermore, we design a comprehensive evaluation suite with novel metrics that assess not only answer correctness but also the capability for visual evidence grounded reasoning. This grounding is crucial for improving accuracy by mitigating spurious correlations and for enhancing user trust by making models more transparent. Our experiments show that current VLMs struggle with causal questions, largely due to their limited ability to construct precise and grounded causal chains. This highlights an important direction for improving future VLMs.

Polynomial systems in multiview geometry are often over-constrained, and naively removing equations can lead to unstable or inconsistent solutions. We revisit this issue through constraint relaxation---selectively removing equations to obtain a finite set of isolated (and, ideally, well-conditioned) solutions. Focusing on the three-view Kruppa equations for autocalibration, we introduce a minimal relaxation framework for identifying subsets of constraints leading to geometrically valid solutions. Using both symbolic computation and numerical homotopy continuation methods, we enumerate and classify all possible relaxation patterns arising from our formulation of Kruppa's equations, isolating all patterns that yield finitely-many solutions. In experiments with synthetic and real images, we demonstrate empirically that specific relaxations are well-conditioned and consistently outperform multiple baselines, including a more classical Kruppa formulation. Overall, our findings establish the minimal relaxation framework as a practical tool in multiview geometry computation.

Zero-shot image restoration provides a flexible way to handle diverse degradations without task-specific training. However, existing methods typically rely on stacked layers or pre-trained features to enhance degradation expression, while overlooking physically consistent priors. The insufficient degradation prompts impose the heavy training burden and high sampling costs during zero-shot diffusion. Moreover, the fixed inference trajectory often collapses to suboptimal solutions under complex corruptions. We observe that heterogeneous degradations can be reparameterized into a minimal set of physically coherent parameters for compact representation. Based on this insight, we first propose a unified physical zero-shot image restoration (UP-ZeroIR) framework that explicitly models heterogeneous degradations into a homogeneous all-in-one distribution. The distribution can be optimized directly in the latent space, enabling principled solution exploration and effective prompt adaptation. Besides, we introduce a dynamic quality-refinement strategy that adaptively adjusts the diffusion trajectory for robust globally optimal convergence. Extensive experiments demonstrate that our method achieves state-of-the-art performance across both single and mixed degradations. Our code is available at https://github.com/yangjinglyy/UP-ZeroIR.

Accurate robot segmentation is a fundamental capability for robotic perception. It enables precise visual servoing for VLA systems, scalable robot-centric data augmentation, accurate real-to-sim transfer, and reliable safety monitoring in dynamic human-robot environments. Despite the strong capabilities of modern segmentation models, surprisingly it remains challenging to segment robots. This is due to robot embodiment diversity, appearance ambiguity, structural complexity, and rapid shape changes. Embracing these challenges, we introduce RobotSeg, a foundation model for robot segmentation in image and video. RobotSeg is built upon the versatile SAM 2 foundation model but addresses its three limitations for robot segmentation, namely the lack of adaptation to articulated robots, reliance on manual prompts, and the need for per-frame training mask annotations, by introducing a structure-enhanced memory associator, a robot prompt generator, and a label-efficient training strategy. These innovations collectively enable a structure-aware, automatic, and label-efficient solution. We further construct the video robot segmentation (VRS) dataset comprising over 2.8k videos (138k frames) with diverse robot embodiments and environments. Extensive experiments demonstrate that RobotSeg achieves state-of-the-art performance on both images and videos, establishing a strong foundation for future advances in robot perception.

In computational pathology, few-shot whole slide image classification is primarily driven by the extreme scarcity of expert-labeled slides. Recent vision-language methods incorporate textual semantics generated by large language models, but treat these descriptions as static class-level priors that are shared across all samples and lack sample-wise refinement. This limits both the diversity and precision of visual-semantic alignment, hindering generalization under limited supervision. To overcome this, we propose the stochastic MUlti-view Semantic Enhancement (MUSE), a framework that first refines semantic precision via sample-wise adaptation and then enhances semantic richness through retrieval-augmented multi-view generation. Specifically, MUSE introduces Sample-wise Fine-grained Semantic Enhancement (SFSE), which yields a fine-grained semantic prior for each sample through MoE-based adaptive visual-semantic interaction. Guided by this prior, Stochastic Multi-view Model Optimization (SMMO) constructs an LLM-generated knowledge base of diverse pathological descriptions per class, then retrieves and stochastically integrates multiple matched textual views during training. These dynamically selected texts serve as enriched semantic supervisions to stochastically optimize the vision-language model, promoting robustness and mitigating overfitting. Experiments on three benchmark WSI datasets show that MUSE consistently outperforms existing vision-language baselines in few-shot settings, demonstrating that effective few-shot pathology learning requires not only richer semantic sources but also their active and sample-aware semantic optimization.

The advent of one-step text-to-image (T2I) models offers unprecedented synthesis speed. However, their application to text-guided image editing remains severely hampered, as forcing existing training-free editors into a single inference step fails. This failure manifests as severe object distortion and a critical loss of consistency in non-edited regions, resulting from the high-energy, erratic trajectories produced by naive vector arithmetic on the models' structured fields. To address this problem, we introduce ChordEdit, a model agnostic, training-free, and inversion-free method that facilitates high-fidelity one-step editing. We recast editing as a transport problem between the source and target distributions defined by the source and target text prompts. Leveraging dynamic optimal transport theory, we derive a principled, low-energy control strategy. This strategy yields a smoothed, variance-reduced editing field that is inherently stable, facilitating the field to be traversed in a single, large integration step. A theoretically grounded and experimentally validated approach allows ChordEdit to deliver fast, lightweight and precise edits, finally achieving true real-time editing on these challenging models.

Composed Image Retrieval (CIR) aims to retrieve target images based on a hybrid query comprising a reference image and a modification text. Early dual-tower Vision-Language Models (VLMs) struggle with cross-modality compositional reasoning required for this task. While adapting generative Multimodal Large Language Models (MLLMs) for retrieval offers a promising direction, we identify that this strategy overlooks a fundamental issue: compressing a generative MLLM into a single-embedding discriminative retriever triggers a paradigm conflict, which leads to Capability Degradation--the deterioration of native fine-grained reasoning after retrieval adaptation. To address this challenge, we propose ReCALL, a model-agnostic framework that follows a diagnose-generate-refine pipeline: First, we diagnose cognitive blind spots of the retriever via self-guided informative instance mining. Next, we generate corrective instructions and triplets by prompting the foundation MLLM and conduct quality control with VQA-based consistency filtering. Finally, we refine the retriever through continual training on these triplets with a grouped contrastive scheme, thereby internalizing fine-grained visual-semantic distinctions and realigning the discriminative embedding space of retriever with intrinsic compositional reasoning within the MLLM. Extensive experiments on CIRR and FashionIQ show that ReCALL consistently recalibrates degraded capabilities and achieves state-of-the-art performance. Code is available at https://github.com/RemRico/Recall.

Few-shot multi-class anomaly detection is crucial in real industrial settings, where only a few normal samples are available while numerous object types must be inspected. This setting is challenging as defect patterns vary widely across categories while normal samples remain scarce. Existing vision-language model-based approaches typically depend on class-specific anomaly descriptions or auxiliary modules, limiting both scalability and computational efficiency. In this work, we propose AnoPLe, a lightweight multimodal prompt learning framework that removes reliance on anomaly-type textual descriptions and avoids any external modules. AnoPLe employs bidirectional interactions between textual and visual prompts, allowing class semantics and instance-level cues to refine one another and form class-conditioned representations that capture shared normal patterns across categories. To enhance localization, we design a scale-aware prefix trained on both global and local views, enabling the prompts to capture both global context and fine-grained details. In addition, alignment loss propagates local anomaly evidence to global features, strengthening the consistency between pixel- and image-level predictions. Despite its simplicity, AnoPLe achieves strong performance on MVTec-AD, VisA, and Real-IAD under the few-shot multi-class setting, surpassing prior approaches while remaining efficient and free from expert-crafted anomaly descriptions. Moreover, AnoPLe generalizes well to unseen anomalies and extends effectively to the medical domain.

Diffusion large language models (dLLMs) are emerging as promising alternatives to autoregressive (AR) LLMs. Recently, this paradigm has been extended to multimodal tasks, leading to the development of diffusion multimodal large language models (dMLLMs). These models are expected to retain the reasoning capabilities of LLMs while enabling faster inference through parallel generation. However, when combined with Chain-of-Thought (CoT) reasoning, dMLLMs exhibit two critical issues.First, we observe that dMLLMs often generate the final answer token at a very early timestep. This trend indicates that the model determines the answer before sufficient reasoning, leading to degraded reasoning performance. Second, during the initial timesteps, dMLLMs show minimal dependency on visual prompts, exhibiting a fundamentally different pattern of visual information utilization compared to AR vision-language models. In summary, these findings indicate that dMLLMs tend to generate premature final answers without sufficiently grounding on visual inputs.To address these limitations, we propose Position & Step Penalty (PSP) and Visual Reasoning Guidance (VRG). PSP penalizes tokens in later positions during early timesteps, delaying premature answer generation and encouraging progressive reasoning across timesteps. VRG, inspired by the classifier-free guidance, amplifies visual grounding signals to enhance the model's alignment with visual evidence.Extensive experiments across various dMLLMs demonstrate that our method achieves up to 7.5% higher accuracy while delivering more than 3x speedup compared to reasoning with four times more diffusion steps. Our code will be released after publication.

Test-time prompt tuning (TPT) has emerged as a powerful technique for adapting pre-trained vision-language models (VLMs) to diverse downstream tasks, including image classification and visual reasoning. With the rise of text-driven object detectors, we extend TPT to object detection, unlocking new capabilities for cross-domain adaptation. However, a critical challenge in TPT is the inherent miscalibration caused by entropy minimization: domain shifts often lead to incorrect predictions, and enforcing high confidence exacerbates miscalibration, ultimately degrading performance. To tackle this, we introduce InsCal, a novel framework designed to enhance cross-domain object detection through three key innovations: (1) extending TPT to a multi-source paradigm, enabling knowledge aggregation across diverse domains; (2) reducing domain gaps via a novel text-driven style transfer strategy that aligns features to the source domain without requiring reference images; and (3) refining the entropy minimization objective with instance-specific calibration, ensuring robust and well-calibrated adaptation. Our approach not only mitigates miscalibration but also significantly improves cross-domain object detection performance, setting a new benchmark for test-time adaptation in VLMs.

Recent unified models integrate understanding experts (e.g., LLMs) with generative experts (e.g., diffusion models), achieving strong multimodal performance. However, recent advanced methods such as BAGEL and LMFusion follow the Mixture-of-Transformers (MoT) paradigm, adopting a symmetric design that mirrors one expert to another for convenient initialization and fusion, which remains suboptimal due to inherent modality discrepancies. In this work, we propose HBridge, an asymmetric H-shaped architecture that enables heterogeneous experts to optimally leverage pretrained priors from their respective Recent unified models integrate understanding experts (e.g., LLMs) with generative experts (e.g., diffusion models), achieving strong multimodal performance. However, recent advanced methods such as BAGEL and LMFusion follow the Mixture-of-Transformers (MoT) paradigm, adopting a symmetric design that mirrors one expert to another for convenient initialization and fusion, which remains suboptimal due to inherent modality discrepancies. modality domains. Unlike prior dense fusion strategies that straightforwardly connect all layers between experts via shared attention, HBridge selectively bridges intermediate layers, reducing over 40% attention sharing, which improves efficiency and enhances generation quality. Shallow and deep layers, which capture modality-specific representations, are decoupled, while mid-layer bridging promotes semantic alignment. To further strengthen cross-modal coherence, we introduce semantic reconstruction tokens that explicitly guide the generative expert to reconstruct visual semantic tokens of the target image. Extensive experiments across multiple benchmarks demonstrate the effectiveness and superior performance of HBridge, establishing a new paradigm for unified multimodal generation.

Recent large vision-language models (LVLMs) have been applied to diverse VQA tasks. However, achieving practical performance typically requires task-specific fine-tuning with large numbers of image-text pairs, which are costly to collect. In this work, we study text-centric training, a setting where only textual descriptions are available and no real images are provided, as a paradigm for low-cost data scaling. Unlike images, whose collection is often restricted by privacy constraints and scarcity in niche domains, text is widely available. Moreover, text is easily editable, enabling automatic diversification and expansion with LLMs at minimal human effort. While this offers clear advantages over image collection in terms of scalability and cost, training on raw text without images still yields limited gains on VQA tasks because of the image-text modality gap. To address this issue, we propose a Text-Printed Image (TPI), which generates synthetic images by directly rendering the given textual description on a plain white canvas. This simple rendering projects text into the image modality and can be integrated into many existing LVLM training pipelines at low cost. Moreover, TPI preserves the semantics of the text, whereas text-to-image models often fail to do. Across four models and seven benchmarks, our systematic experiments show that TPI enables more effective text-centric training than synthetic images generated by a diffusion model. We further explore TPI as a low-cost data-augmentation strategy and demonstrate its practical utility. Overall, our findings highlight the significant potential of text-centric training and, more broadly, chart a path toward fully automated data generation for LVLMs.

Domain Generalization in Semantic Segmentation (DG-SS) aims to enable segmentation models to perform robustly in unseen environments. However, conventional DG-SS methods are restricted to a fixed set of known categories, limiting their applicability in open-world scenarios. Recent progress in Vision-Language Models (VLMs) has advanced Open-Vocabulary Semantic Segmentation (OV-SS) by enabling models to recognize a broader range of concepts. Yet, these models remain sensitive to domain shifts and struggle to maintain robustness when deployed in unseen environments, a challenge that is particularly severe in urban-driving scenarios. To bridge this gap, we introduce Open-Vocabulary Domain Generalization in Semantic Segmentation (OVDG-SS), a new setting that jointly addresses unseen domains and unseen categories. We introduce the first benchmark for OVDG-SS in autonomous driving, addressing a previously unexplored problem and covering both synthetic-to-real and real-to-real generalization across diverse unseen domains and unseen categories. In OVDG-SS, we observe that domain shifts often distort text-image correlations in pre-trained VLMs, which hinders the performance of OV-SS models. To tackle this challenge, we propose S^2-Corr, a state-space-driven text-image correlation refinement mechanism that can mitigate domain-induced distortions and produce a more consistent text-image correlation under distribution changes. Extensive experiments on our constructed benchmark demonstrate that the proposed method achieves superior cross-domain performance and efficiency compared to existing OV-SS approaches.

Vision-language models (VLMs) have achieved remarkable success across diverse tasks. However, concerns about their trustworthiness persist, particularly regarding tendencies to lean more on textual cues than visual evidence and the risk of producing ungrounded or fabricated responses. To address these issues, we propose Saliency-R1, a framework for improving the interpretability and faithfulness of VLMs reasoning. Specifically, we introduce a novel saliency map technique that efficiently highlights critical image regions contributing to generated tokens without additional computational overhead. This can further be extended to trace how visual information flows through the reasoning process to the final answers, revealing the alignment between the thinking process and the visual context. We use the overlap between the saliency maps and human-annotated bounding boxes as the reward function, and apply Group Relative Policy Optimization (GRPO) to align the salient parts and critical regions, encouraging models to focus on relevant areas when conduct reasoning. Experiments show Saliency-R1 improves reasoning faithfulness, interpretability, and overall task performance.

Cross-modal biomedical signals such as pathology and genomics can provide richer and more robust semantic guidance for medical image representation learning. However, the availability of such guidance remains limited, as privacy constraints and acquisition costs severely restrict access to medical images paired with other biomedical data. A further challenge lies in modality discrepancy, which introduces intra-modal statistical bias and cross-modal noise, thereby degrading the quality of medical image representations. To address these challenges, we propose KAMP, a large language model (LLM)-driven multimodal pretraining framework for medical image representation learning. KAMP leverages textual priors as the semantic anchor to enhance medical image representations and align them with multimodal biomedical representations, enabling the learning of rich and robust features even when paired data are scarce. KAMP operates in three stages. First, the LLM generates personalized diagnostic knowledge from patient clinical text and imaging metadata. This knowledge is injected as a prior to enrich medical image representations and serves as a semantic anchor to reduce the representation gap between medical images and other biomedical modalities. Second, the LLM is optimized using Group Relative Policy Optimization (GRPO), with the cross-modal aligner pretrained in the first stage serving as the reward model. Third, the refined knowledge is used to retrain the cross-modal aligner, yielding more robust medical image representations while mitigating bias and noise introduced by other modalities. Comprehensive evaluations on brain, bladder, and liver cancer datasets demonstrate that KAMP outperforms existing methods in most downstream few-shot classification tasks.

While Video Large Language Models (Video-LLMs) have shown significant potential in multimodal understanding and reasoning tasks, how to efficiently select the most informative frames from videos remains a critical challenge. Existing methods attempt to optimize frame sampling by reducing inter-frame redundancy or employing unsupervised event localization. However, these approaches often fall short in handling complex instruction-following tasks and scenarios that demand precise temporal modeling, resulting in limited performance in both semantic alignment and temporal reasoning. To address the above challenges, we introduce Instructed Temporal Grounding for Videos (VideoITG), a framework aiming to adaptively customize frame sampling strategies based on user instructions. Specifically, we design the VidThinker pipeline, which automates annotation by generating instruction-conditioned captions, retrieving relevant video segments, and selecting key frames to enable efficient supervision. Using VidThinker, we build the VideoITG-40K dataset with 40K videos and 500K temporal grounding annotations. Our plug-and-play VideoITG model leverages Video-LLMs' visual-language alignment and reasoning for discriminative frame selection. VideoITG consistently boosts the performance on multiple multimodal video understanding benchmarks, demonstrating its effectiveness and potential.

Reward models (RMs) are essential for training large language models (LLMs), but remain underexplored for omni models that handle interleaved image and text sequences. We introduce Multimodal RewardBench 2 (MMRB2), the first comprehensive benchmark for reward models on multimodal understanding and (interleaved) generation. MMRB2 spans four tasks: text-to-image, image editing, interleaved generation, and multimodal reasoning ("thinking-with-images"), providing 1,000 expert-annotated preference pairs per task from 23 models and agents across 21 source tasks. MMRB2 is designed with: (1) practical but challenging prompts; (2) responses from state-of-the-art models, and agents; and (3) preference pairs with strong human-expert consensus, curated via an ensemble filtering strategy. Using MMRB2, we study existing judges for each subtask, including multimodal LLM-as-a-judge and models trained with human preferences. Top judges like GPT-5 and Gemini-2.5-Pro reach 66-75% accuracy, compared to >90% for humans, and outperform the commonly used GPT-4o (59% accuracy). The best performing open-source model Qwen3-VL-32B achieves similar accuracies as Gemini-2.5-Flash (64%). We also show that MMRB2 performance strongly correlates with downstream task success, and conduct an in-depth analysis that shows key areas to improve the reward models going forward.

In real-world robotic manipulation, states typically admit a neighborhood of near-equivalent actions. That is, for each state, there exist a feasible action neighborhood (FAN) rather than a single correct action, within which motions yield indistinguishable progress. However, prevalent VLA training methodologies are directly inherited from linguistic settings and do not exploit the FAN property, thus leading to poor generalization and low sample efficiency. To address this limitation, we introduce a FAN-guided regularizer that shapes the model's output distribution to align with the geometry of FAN. Concretely, we introduce a Gaussian prior that promotes locally smooth and unimodal predictions around the preferred direction and magnitude. In extensive experiments across both reinforced finetuning (RFT) and supervised finetuning (SFT), our method achieves significant improvement in sample efficiency, and success rate in both in-distribution and out-of-distribution (OOD) scenarios. By aligning with the intrinsic action tolerance of physical manipulation, FAN-guided regularization provides a principled and practical method for sample-efficient, and generalizable VLA adaptation. The source code is publicly available at: https://github.com/Haochen-Niu/FAN-VLA.