CVPR 2026 Papers

Vision-Language Models (VLMs) have demonstrated remarkable capabilities in instruction following and 2D visual understanding. However, state-of-the-art VLMs, including GPT-5, still struggle with 3D perception, particularly in tasks such as monocular 3D visual grounding. While specialized vision-only models excel in this domain, they often lack the rich semantic understanding inherent to VLMs. To bridge this gap, we propose \texttt MonoVLM , a novel triple-stage training framework that enables VLMs to perform accurate monocular 3D grounding. The core of our method is a progressive training process that uses Group Relative Policy Optimization (GRPO) to gradually teach the model to first localize the described object, then understand its 3D structure, and finally estimate the full 3D bounding box accurately. Comprehensive experiments show that \texttt MonoVLM models significantly outperform existing VLMs and even surpass the performance of specialized vision-only models. We validate our design via extensive comparisons and ablation studies.

Autonomous driving has made substantial progress recently, achieving reliable performance in most real-world environments. However, existing algorithms still depend heavily on high-definition maps, making them ineffective in mapless scenarios such as indoor parking lots. These limitations hinder seamless point-to-point navigation and restrict the broader deployment of the autonomous driving system.To address this challenge, we propose DriveVLN, a new task that extends Vision-and-Language Navigation (VLN) to autonomous driving. DriveVLN employs visual and linguistic priors to guide vehicles toward destinations based solely on concise natural-language descriptions, without access to predefined maps or routes. Unlike conventional VLN, which relies on detailed step-wise instructions in indoor environments, DriveVLN requires models to produce navigation information based on diverse visual cues and history, including signs, landmarks, and textual indicators.We further develop a CARLA-based simulation engine comprising over 200 realistic scenes reconstructed from real road scans, enabling large-scale training and closed-loop evaluation. A baseline model is established through supervised fine-tuning on real data, followed by reinforcement learning in simulation.Comprehensive experiments show that DriveVLN effectively bridges map-based and mapless driving, providing a new foundation for unified, language-driven autonomous navigation in complex real-world environments.

Vision-language models (VLMs) have demonstrated strong potential for adapting to downstream biomedical tasks with limited training samples. However, their generalization to unseen classes within the same dataset remains limited, as the image-text alignment semantics often rely on spurious cues present in seen classes that do not transfer. To tackle this, we propose BiomedCCPL (Causal Conditional Prompt Learning), a framework that uses VGAP (Visual Grounder with Adaptive Prototype) to generate image-conditional prompts from multi-scale adaptive prototypes and employs SCD (Synergistic Causal Disentanglement) to regularize the generation of image-conditional prompts. Guided by insights from a causal analysis of generalization to unseen classes, SCD leverages multiple synergistic learning objectives to perform front-door adjustment, ensuring that the dynamically generated image-conditional prompts focus on underlying diagnostic image features shared across seen and unseen classes. Experiments on 11 datasets across 9 modalities demonstrate that BiomedCCPL effectively enhances the model's data efficiency and generalization ability. In particular, on the Base-to-Novel task, BiomedCCPL achieves an average HM of 79.98%, surpassing the previous state-of-the-art by 6.45%. Code is available at https://github.com/burgers0708/BiomedCCPL.

Multi-model routing has evolved from an engineering technique into essential infrastructure, yet existing work lacks a systematic, reproducible benchmark for evaluating vision-language models (VLMs). We present VL-RouterBench to assess the overall capability of VLM routing systems systematically. The benchmark is grounded in raw inference and scoring logs from VLMs and constructs quality and cost matrices over sample-model pairs. In scale, VL-RouterBench covers 14 datasets across 3 task groups, totaling 30,540 samples, and includes 15 open-source models and 2 API models, yielding 519,180 sample-model pairs and a total input-output token volume of 34,494,977. The evaluation protocol jointly measures average accuracy, average cost, and throughput, and builds a ranking score from the harmonic mean of normalized cost and accuracy to enable comparison across router configurations and cost budgets. On this benchmark, we evaluate 10 routing methods and baselines and observe a significant routability gain, while the best current routers still show a clear gap to the ideal Oracle, indicating considerable room for improvement in router architecture through finer visual cues and modeling of textual structure. We open-source the complete data construction and evaluation toolchain to promote comparability, reproducibility, and practical deployment in multimodal routing research.

Vision-language models (VLMs) with dynamic resolution vision encoders achieve strong performance, but face significant efficiency challenges due to long input sequences. A common approach is to assess the importance of tokens and prune those that are less informative. Recent methods utilizing a small VLM to provide the importance map of visual tokens have outperformed existing rule-based and similarity-driven pruning approaches, particularly under high pruning ratios. However, directly using the small VLM remains unreliable, as it utilizes the aggregated visual attention weights as importance score, which can lead to noisy guidance if the generated tokens are incorrect.To address this, we invert the approach by having it detect non-informative visual tokens according to the user's input query. By adding a variational information bottleneck in the small VLM, we can approximate the entropy of each visual token as pruning guidance. Such a posteriori-guided pruning method allows the large VLM to retain its reasoning capacity with improved efficiency. Extensive experiments on eight benchmarks demonstrate the effectiveness of our approach. With only 5% of visual tokens retained, the large VLM preserves 95% of its original performance, outperforming the state of the art by 8%.

Robotic manipulation in complex 3D environments requires unifying spatial reasoning with intuitive visual perception, which is a capability that current Vision-Language-Action paradigms address separately. While 3D VLAs excel in geometric and physical reasoning, they lack intuitive, image-level understanding and dense visual semantics; conversely, 2D VLAs (even with depth image) provide rich visual intuition and semantic continuity but miss explicit spatial global grounding. We introduce DiffRender-VLA, a differentiable rendering-based framework that bridges 3D and 2D Vision-Language-Action models through gradient-consistent visual mediation. It generates differentiable images by localizing the next end-effector target with a world-aligned cube marker, differentiably structuring surrounding geometry whose color encodes spatial relations to the marker, and rendering adaptive viewpoints optimized to reveal the target-environment spatial relationships. These differentiable images serve as visual bridges, embedding spatial semantics while allowing gradients from 2D VLAs to backpropagate into 3D representations, thereby coupling geometric reasoning with visual perception. This closed differentiable loop unifies reasoning and perception, substantially improving performance under occlusion, clutter, and complex spatial manipulation tasks, achieving average improvements of +12.1% over state-of-the-art methods. Codes are available at https://github.com/zyl123456aB/DIFFVLA.

Dataset distillation compresses large training sets into compact synthetic datasets while preserving downstream performance. As modern systems increasingly operate on paired vision-language inputs, multimodal distillation must preserve representation quality and cross-modal alignment under tight compute and memory budgets, yet prior methods often require heavy computes and overlook their correlations. To address this, we present Multimodal Distribution Matching (MDM), a geometry-aware framework for efficient and generalizable multimodal distillation. Specifically, MDM integrates complementary components at the data, model, and loss levels. At the data level, it initializes synthetic image-text pairs by sampling from clusters in the joint embedding space. At the model level, it forms a mixed teacher by interpolating independently fine-tuned models in weight space according to their angular deviation from the pretrained anchor. At the loss level, it matches joint distributions on the unit hypersphere using a geometry-aware matching objective that exploits the joint features in the cross-modal alignment and discrepancy directions along with symmetric contrastive learning. Across image-text retrieval benchmarks with cross-architecture evaluation, MDM yields compact synthetic sets that preserve multimodal semantics, substantially reduce distillation cost, and remain robust across architectures.

Vision-Language Models (VLMs) have emerged as a promising paradigm in autonomous driving (AD), providing a unified framework for perception and decisionmaking. However, their real-world deployment is hindered by significant computational overhead when processing high-resolution, multi-view images. This complexity stems from the massive number of visual tokens, which increases inference latency and memory consumption due to the quadratic complexity of self-attention. To address these challenges, we propose Prune2Drive, a plug-and-play visual token pruning framework for multi-view VLMs in AD. Prune2Drive introduces two core innovations: (i) a diversity-aware token selection mechanism that prioritizes semantic and spatial coverage across views, and (ii) a view-adaptive pruning controller that automatically learns optimal pruning ratios based on camera importance to downstream tasks. Unlike prior methods, Prune2Drive requires no model retraining or access to attention maps, ensuring compatibility with modern efficient attention implementations. Extensive experiments on the DriveLM and DriveLMM-o1 benchmarks demonstrate that Prune2Drive achieves significant speedups and memory savings with minimal performance impact. When retaining only 10% of visual tokens, our method achieves a 6.40x speedup in the prefilling phase and consumes only 13.4% of the original FLOPs, with a mere 3% average performance drop on the DriveLM benchmark. Code is available at: https://github.com/MinhaoXiong/Prune2Drive.git

Fine-tuning large-scale Vision-Language Models (VLMs) is crucial for specialized tasks, but their performance is often undermined by the label noise prevalent in real-world datasets. Traditional approaches to learning with noisy labels typically rely on a self-referential loop, using a model's own predictions to correct errors. While recent VLM-specific methods have begun to leverage cross-modal information to aid in noise detection, we explore an alternative direction: using the text modality not just to identify noise, but to establish a source of ground truth that is fully independent of the training data's potentially corrupt labels. To this end, we propose Text-ANchored Guided Optimization (TANGO), a framework centered on "semantic anchors"--a set of pure, immutable reference points generated from diverse text descriptions. Building upon these anchors, our approach reframes two key aspects of learning with noisy labels. First, we replace the conventional linear classifier with a parameter-free Text-Anchored Classifier, making predictions a direct, weighted consensus of the clean anchors. Second, we introduce an Anchor-Guided Refinement mechanism that validates each sample's given label against this external ground truth, providing a more robust signal for sample selection and label correction. Extensive experiments demonstrate that this approach achieves competitive and often state-of-the-art performance. Our code will be publicly available.

In this paper, we tackle the Aerial Vision-and-Dialog Navigation (AVDN) task in the training-free setting for resource-efficient high-altitude UAV navigation.Naively applying MLLMs leads to unreliable navigation due to weak directional grounding and the lack of explicit spatial memory.To address these issues, we propose PSC-AVDN, a training-free framework that tightly couples a three-stage Parsing-Search-Confirmation reasoning pipeline with a Structured Spatial Memory (SSM).The parsing stage uses an LLM to convert ambiguous dialogue instructions into stable geometric directional and destination cues.A Search Chain-of-Thought (S-CoT) then performs stepwise target exploration under high-altitude observations, and a Confirmation Chain-of-Thought (C-CoT) conducts fine-grained verification around candidate regions to resolve visual ambiguity.Meanwhile, SSM integrates three complementary sources of spatial cues, including multi-scale visual observation, spatial visual memory, and structured geometric memory to provide global spatial context and long-horizon consistency.Extensive experiments on ANDH and ANDH-Full show that PSC-AVDN establishes new state-of-the-art performance in the training-free setting, matching or surpassing several finetuned methods.

This paper reveals that many open-source large language models (LLMs) lack hierarchical knowledge about our visual world, unaware of even well-established biology taxonomies. This shortcoming makes LLMs a bottleneck for vision LLMs' hierarchical visual recognition (e.g., recognizing Anemone Fish but not Vertebrate). We arrive at these findings using about one million four-choice visual question answering (VQA) tasks constructed from six taxonomies and four image datasets. Interestingly, finetuning a vision LLM using our VQA tasks reaffirms LLMs' bottleneck effect because the VQA tasks improve the LLMs' hierarchical consistency in text-only tasks more than the vision LLMs'. We believe that one cannot make vision LLMs understand our visual world hierarchically until LLMs possess corresponding taxonomy knowledge.

Network pruning is an effective technique for enabling lightweight Large Vision-Language Models (LVLMs), which primarily incorporates both weights and activations into the importance metric. However, existing efforts typically process calibration data from different modalities in a unified manner, overlooking modality-specific behaviors. This raises a critical challenge: how to address the divergent behaviors of textual and visual tokens for accurate pruning of LVLMs. To this end, we systematically investigate the sensitivity of visual and textual tokens to the pruning operation by decoupling their corresponding weights, revealing that: (i) the textual pathway should be calibrated via text tokens, since it exhibits higher sensitivity than the visual pathway; (ii) the visual pathway exhibits high redundancy, permitting even 50% sparsity. Motivated by these insights, we propose a simple yet effective Asymmetric Text-Visual Weight Pruning method for LVLMs, dubbed ATV-Pruning, which establishes the importance metric for accurate weight pruning by selecting the informative tokens from both textual and visual pathways. Specifically, ATV-Pruning integrates two primary innovations: first, a calibration pool is adaptively constructed by drawing on all textual tokens and a subset of visual tokens; second, we devise a layer-adaptive selection strategy to yield important visual tokens. Finally, extensive experiments across standard multimodal benchmarks verify the superiority of our ATV-Pruning over state-of-the-art methods.

Medical image segmentation is challenging due to limited annotated data, high labeling costs, and substantial image heterogeneity. Although large-scale vision foundation models (e.g., SAM) have shown great potential in this field, existing SAM-based methods typically rely on expert-defined geometric prompts or complex clinical text prompts, which limits their generalizability across diverse medical image segmentation tasks. To overcome these challenges, we propose Simple-ViLMedSAM, a CLIP-SAM integration framework that enables high-accuracy segmentation in zero-shot and few-shot settings using only simple text queries, that is, using only basic anatomical or disease-related text labels. At its core is an Implicit Pos-Prompter (IPP), which generates attribution maps containing implicit positional cues to replace traditional geometric prompts. IPP incorporates a multi-modal information bottleneck and an affinity-based refinement strategy to ensure high-quality guidance from CLIP-SAM interactions. To further enhance segmentation, we introduce a Bidirectional Interaction Decoder (BID) that employs bidirectional cross-attention to align IPP's positional maps with SAM's pixel-level features. By jointly modeling global semantics and local details, BID significantly improves segmentation accuracy. Extensive experiments on four public datasets demonstrate that Simple-ViLMedSAM consistently outperforms existing methods in both zero-shot and few-shot medical image segmentation tasks, using only simple text queries. The code will be publicly available upon acceptance.

Real-world human action understanding remains challenging due to long-tailed label distributions, compositional motion patterns, and viewpoint variations. Existing skeleton-based methods often lack a structured and transferable representation of motion, and task-specific models for generation, classification, and detection are usually trained independently, resulting in fragmented pipelines and limited cross-task generalization. We present PRISM, a PRImitive-centric Skeleton Modeling framework that learns a shared motion representation from a motion generation objective and transfers it to perception tasks. PRISM represents each action sequence as a trajectory in a primitive coefficient space, which captures how a set of learned atomic motion primitives contribute to the observed motion. A structured decomposition module learns this representation in a physically grounded and view-invariant manner via motion generation. Instead of enforcing joint or unified training across tasks, PRISM provides a single primitive-centric representation that can be sequentially transferred to downstream classification and frame-wise detection through lightweight task heads. This representation introduces structure, compositionality, and improved generalization across distinct supervisions. PRISM consistently improves performance on long-tailed and multi-label datasets and enables interpretable reasoning over compositional and rare actions. Extensive experimental results show that the structured primitive space serves as a transferable and robust foundation for diverse action understanding tasks in real-world datasets.

In large-scale industrial documents with scanned images, complex layouts, and multiple pages, the effectiveness of retrieval-augmented generation (RAG) is highly dependent on chunking quality. However, existing text-centric chunkers overlook the visual and structural cues present in real-world documents, leading to redundant or ambiguous chunks that impair retrieval and answer accuracy. To address this problem, we propose \ours which integrates (i) SharedDet for normalizing document parsing and OCR outputs into a document-level frame, (ii) Multi-modal block embeddings with boundary-aware SoftROI, (iii) global document-tree reconstruction via biaffine scoring, and (iv) structure-aware dependency chunking that preserves boundaries and reduces redundancy. \ours achieves consistent gains across both Document Hierarchical Parsing (DHP) and corpus-level RAG evaluations, improving STEDS by +28.5--39.6%, retrieval nDCG by +1.1--15.3%, and QA ANLS by +4.5--15.3%. These results demonstrate that modeling document-level dependencies with Multi-modal, structure-aware chunking improves RAG performance on long, multi-page industrial documents.

Edge-based representations are fundamental cues for visual understanding, a principle rooted in early vision research and still central today. We extend this principle to vision-language alignment, showing that isolating and aligning structural cues across modalities can greatly benefit fine-tuning on long, detail-rich captions, with a specific focus on improving cross modal retrieval. We introduce StructXLIP, a fine-tuning alignment paradigm that extracts edge maps (e.g., Canny), treating them as proxies for the visual structure of an image, and filters the corresponding captions to emphasize structural cues, making them "structure-centric". Fine-tuning augments the standard alignment loss with three structure-centric losses: (i) aligning edge maps with structural text, (ii) matching local edge regions to textual chunks, and (iii) connecting edge maps to color images to prevent representation drift. From a theoretical standpoint, while standard CLIP maximizes the mutual information between visual and textual embeddings, StructXLIP additionally maximizes the mutual information between multimodal structural representations. This auxiliary optimization is intrinsically harder, guiding the model toward more robust and semantically stable minima, enhancing vision-language alignment. Beyond outperforming current competitors on cross-modal retrieval on both general and specialized domains, our method serves as a general boosting recipe that can be integrated into future approaches in a plug-and-play manner. Code and pretrained models are available at: https://github.com/intelligolabs/StructXLIP.

Recent Vision-Language-Action (VLA) models reformulate vision-language models by tuning them with millions of robotic demonstrations. While they perform well when fine-tuned for a single embodiment or task family, extending them to multi-skill settings remains challenging: directly merging VLA experts trained on different tasks results in near-zero success rates. This raises a fundamental question: what prevents VLAs from mastering multiple skills within one model? With an empirical decomposition of learnable parameters during VLA fine-tuning, we identify two key sources of non-mergeability:(1) Finetuning drives LoRA adapters in the VLM backbone toward divergent, task-specific directions beyond the capacity of existing merging methods to unify.(2) Action experts develop inter-block dependencies through self-attention feedback, causing task information to spread across layers and preventing modular recombination.To address these challenges, we present MergeVLA, a merging-oriented VLA architecture that preserves mergeability by design.MergeVLA introduces sparsely activated LoRA adapters via task masks to retain consistent parameters and reduce irreconcilable conflicts in the VLM.Its action expert replaces self-attention with cross-attention-only blocks to keep specialization localized and composable.When the task is unknown, it uses a test-time task router to adaptively select the appropriate task mask and expert head from the initial observation, enabling unsupervised task inference.Across LIBERO, LIBERO-Plus, RoboTwin, and multi-task experiments on the real SO101 robotic arm, MergeVLA achieves performance comparable to or even exceeding individually finetuned experts, demonstrating robust generalization across tasks, embodiments, and environments. Project page: https://mergevla.github.io/

The Abstraction and Reasoning Corpus (ARC) is designed to promote research on abstract reasoning, a fundamental aspect of human intelligence. Common approaches to ARC treat it as a language-oriented problem, addressed by large language models (LLMs) or recurrent reasoning models. However, although the puzzle-like tasks in ARC are inherently visual, existing research has rarely approached the problem from a vision-centric perspective. In this work, we formulate ARC within a vision paradigm, framing it as an image-to-image translation problem. To incorporate visual priors, we represent the inputs on a "canvas" that can be processed like natural images. It is then straightforward for us to apply standard vision architectures, such as a vanilla Vision Transformer (ViT), to perform image-to-image mapping. Our model is trained from scratch solely on ARC data and generalizes to unseen tasks through test-time training. Our framework, termed Vision ARC (VARC), achieves 60.4% accuracy on the ARC-1 benchmark, substantially outperforming existing methods that are also trained from scratch. Our results are competitive with those of leading LLMs and close the gap to average human performance.

We present Scalable Pixel-anchored End-to-end Diffusion (SpeeDiff), a latent diffusion method that jointly trains the VAE and the diffusion model from scratch. In principle, joint training allows the diffusion loss gradient to directly guide the VAE encoder, encouraging the formation of a generation-friendly latent space and potentially yielding faster convergence than the conventional two-stage approach with a pretrained frozen VAE. However, a naive end-to-end implementation severely degrades performance, as unrestricted backpropagation of the diffusion loss leads to latent space collapse. Our main technical contribution is a simple yet effective Tweedie Pixel Reconstruction (TPR) loss, which provides additional pixel-level feedback by decoding a predicted clean latent from an intermediate noisy state using Tweedie's formula, thereby alleviating collapse. Furthermore, our method enables jointly scaling a fully transformer-based architecture and enhances representation alignment within the end-to-end framework. Our SpeeDiff-XL model achieves over 140x and 61x faster training compared to Vanilla SiT and REPA, respectively, while attaining an FID of 1.50 without guidance on ImageNet 256x256 generation. With a more efficient 32x compressed VAE, our model further reaches an FID of 1.53 without guidance on ImageNet 512x512 generation.