CVPR 2026 Papers

Understanding the physical world is essential for generalist AI agents. However, it remains unclear whether state-of-the-art vision perception models (e.g., large VLMs) can perform quantitative physical reasoning tasks. Existing evaluations are predominantly VQA-based and qualitative, offering limited insight into whether these models can infer the kinematic quantities of moving objects from videos. To address this, we present QuantiPhy, the first benchmark designed to quantitatively measure a VLM's physical reasoning ability. Comprising more than 3.3K video-text instances with numerical ground truth, QuantiPhy evaluates a VLM's performance on estimating an object's size, velocity, and acceleration at a given timestamp, using one of these properties as an input prior. The benchmark standardizes prompts and scoring to assess numerical accuracy, enabling fair comparisons across models. Our experiments on state-of-the-art VLMs reveal a consistent gap between their qualitative plausibility and actual numerical correctness. We further provide an in-depth analysis of key factors like background noise, counterfactual priors, and strategic prompting and find that state-of-the-art VLMs lean heavily on pre-trained world knowledge rather than faithfully using the provided visual and textual inputs when reasoning about objects' kinematic properties. QuantiPhy offers the first rigorous, scalable testbed to move VLMs beyond mere verbal plausibility toward a numerically grounded physical understanding.

Recent advances in text-to-image (T2I) generation via reinforcement learning (RL) have benefited from reward models that assess semantic alignment and visual quality. However, most existing reward models pay limited attention to fine-grained spatial relationships, often producing images that appear plausible overall yet contain inaccuracies in object positioning. In this work, we present SpatialReward, a verifiable reward model explicitly designed to evaluate spatial layouts in generated images. SpatialReward adopts a multi-stage pipeline: a Prompt Decomposer extracts entities, attributes, and spatial metadata from free-form prompts; expert detectors provide accurate visual grounding of object positions and attributes; and a vision-language model applies chain-of-thought reasoning over grounded observations to assess complex spatial relations that are challenging for rule-based methods. To more comprehensively evaluate spatial relationships in generated images, we introduce SpatRelBench, a benchmark covering object attributes, orientation, inter-object relations, and rendered text placement. Experiments on Stable Diffusion and FLUX show that incorporating SpatialReward into RL training consistently improves spatial consistency and overall generation quality, with results aligned more closely to human judgments. These findings indicate that verifiable reward models hold considerable potential for enabling more accurate and controllable optimization in text-to-image generation models.

Spatial intelligence in vision-language models (VLMs) attracts research interest with the practical demand to reason in the 3D world. Despite promising results, most existing methods follow the conventional 2D pipeline in VLMs and use pixel-aligned representations for the vision modality. However, correspondence-based models with implicit 3D scene understanding often fail to achieve spatial consistency, and representation-based models with 3D geometric priors lack efficiency in vision sequence serialization. To address this, we propose a Proxy3D method with compact yet comprehensive 3D proxy representations for the vision modality. Given only video frames as input, we employ semantic and geometric encoders to extract scene features and then perform their semantic-aware clustering to obtain a set of proxies in the 3D space. For representation alignment, we further curate the SpaceSpan dataset and apply multi-stage training to adopt the proposed 3D proxy representations with the VLM. When using shorter sequences for vision information, our method achieves competitive or state-of-the-art performance in 3D visual question answering, visual grounding and general spatial intelligence benchmarks.

Detecting out-of-distribution (OOD) objects is indispensable for safely deploying object detectors in the wild. Current approaches enable the unknown-aware ability by regularizing the instance-level feature space, such as outlier synthesis. Despite the general efficacy, it is challenging to truly learn the concept of 'unknown' under the absence of real unknown data. In this paper, we propose UNO-Adapter, a simple yet highly effective framework tailored for OOD object detection. Our key insight is that in object detection, where in-distribution (ID) and OOD objects may coexist within the same context, we need global abstraction and reasoning to help the detector learn their differences, i.e., unknown injection. UNO-Adapter consists of two key steps: unsupervised concept discovery and neural concept binder. The former introduces an object-centric learning paradigm to abstract and model the holistic image, including both ID and OOD, obtaining sparse and compressed slot-based representations with relational constraints. The latter dynamically combines slots with object candidates extracted by the detector, binding the concept of unknown to the de facto detector. During inference, we introduce an image-guided OOD object score to reinforce the distinction between ID and OOD. Experiments on standard benchmarks demonstrate the superiority of the proposed method. In particular, UNO-Adapter reduces the FPR95 by up to 11.96% compared to the previous best OOD object detection method.

The rapid advancement of Vision-Language Models (VLMs) has brought their safety vulnerabilities into sharp focus. However, existing red teaming methods are fundamentally constrained by an inherent linear exploration paradigm, confining them to optimizing within a predefined strategy set and preventing the discovery of novel, diverse exploits. To transcend this limitation, we introduce TreeTeaming, an automated red teaming framework that reframes strategy exploration from static testing to a dynamic, evolutionary discovery process. At its core lies a strategic Orchestrator, powered by a Large Language Model (LLM), which autonomously decides whether to evolve promising attack paths or explore diverse strategic branches, thereby dynamically constructing and expanding a strategy tree. A multimodal actuator is then tasked with executing these complex strategies. In the experiments across 12 prominent VLMs, TreeTeaming achieves state-of-the-art attack success rates on 11 models, outperforming existing methods and reaching up to 87.60% on GPT-4o. The framework also demonstrates superior strategic diversity over the union of previously public jailbreak strategies. Furthermore, the generated attacks exhibit an average toxicity reduction of 23.09%, showcasing their stealth and subtlety. Our work introduces a new paradigm for automated vulnerability discovery, underscoring the necessity of proactive exploration beyond static heuristics to secure frontier AI models. The code and data are available at: https://github.com/ChunXiaostudy/TreeTeaming. Warning: This paper contains examples of harmful texts and images, and reader discretion is recommended.

Zero-shot anomaly detection aims to detect and localise abnormal regions in the image without access to any in-domain training images. While recent approaches leverage vision-language models (VLMs), such as CLIP, to transfer high-level concept knowledge, methods based on purely vision foundation models (VFMs), like DINOv2, have lagged behind in performance. We argue that this gap stems from two practical issues: (i) limited diversity in existing auxiliary anomaly detection datasets and (ii) overly shallow VFM adaptation strategies. To address both challenges, we propose AnomalyVFM, a general and effective framework that turns any pretrained VFM into a strong zero-shot anomaly detector. Our approach combines a robust three-stage synthetic dataset generation scheme with a parameter-efficient adaptation mechanism, utilising low-rank feature adapters and a confidence-weighted pixel loss. Together, these components enable modern VFMs to substantially outperform current state-of-the-art methods. More specifically, with RADIO as a backbone, AnomalyVFM achieves an average image-level AUROC of 94.1% across 9 diverse datasets, surpassing previous methods by significant 3.3 percentage points. Code: \href https://maticfuc.github.io/anomaly_vfm/ Project Page

Infrared and Visible Image Fusion (IVIF) aims to combine complementary information from infrared and visible images to overcome the limitations of a single modality. While existing methods typically employ fixed or sample-adaptive fusion paradigms where fusion weights are static or derived from global pixel distributions, they often overlook spatial inconsistencies in pixel distribution within images, leading to suboptimal performance. To address this issue, we propose RegionFuse, a Region-Adaptive Pixel Distribution Learning Network for IVIF, which dynamically generates fusion weights based on local pixel distributions to construct a region-wise adaptive fusion paradigm. RegionFuse introduces a Mixture of Region Attention (MoRA) mechanism, which assigns each region to several specialized experts, enabling region-level feature interaction tailored to specific local distributions. Furthermore, we design a Region Feature Compression Module (RFCM) and place it after each MoRA to enhance informative regions and suppress redundant ones. Extensive experiments on various benchmarks demonstrate the superiority and robustness of RegionFuse, especially in handling non-uniform pixel distributions. Evaluations on NIR-VIS and downstream tasks further confirm its generalizability and practical utility. The code is available at https://github.com/DarkIceField/RegionFuse.

Recent progress in deep learning has been driven by increasingly large-scale models, but the resulting computational cost has become a critical bottleneck. Sparse Mixture of Experts (MoE) offers an effective solution by activating only a small subset of experts for each input, achieving high scalability without sacrificing inference speed. Although effective, sparse MoE training exhibits characteristic optimization difficulties. Because the router receives informative gradients only through the experts selected in the forward pass, it suffers from gradient blocking and obtains little information from unselected routes. This limited, highly localized feedback makes it difficult for the router to learn appropriate expert-selection scores and often leads to unstable routing dynamics, such as fluctuating expert assignments during training. To address this issue, we propose TGR-MoE: Teacher-Guided Routing for Sparse Vision Mixture-of-Experts, a simple yet effective method that stabilizes router learning using supervision derived from a pretrained dense teacher model. TGR-MoE constructs a teacher router from the teacher's intermediate representations and uses its routing outputs as pseudo-supervision for the student router, suppressing frequent routing fluctuations during training and enabling knowledge-guided expert selection from the early stages of training. Extensive experiments on ImageNet-1K and CIFAR-100 demonstrate that TGR consistently improves both accuracy and routing consistency, while maintaining stable training even under highly sparse configurations.

In the progress of industrial anomaly detection, general anomaly detection (GAD) is an emerging trend and also the ultimate goal. Unlike the conventional single- and multi-class AD, general AD aims to train a general AD model that can directly detect anomalies in diverse novel classes without any retraining or fine-tuning on the target data. Recently, Multimodal Large Language Models (MLLMs) have shown great promise in achieving general anomaly detection due to their revolutionary visual understanding and language reasoning capabilities. However, MLLM's general AD ability remains underexplored due to: (1) MLLMs are pretrained on amounts of data sourced from the Web, these data still have significant gaps with the data in AD scenarios. Moreover, the image-text pairs during pretraining are also not specifically for AD tasks. (2) The current mainstream AD datasets are image-based and not yet suitable for post-training MLLMs. To facilitate MLLM-based general AD research, we present MMR-AD, which is a comprehensive benchmark for both training and evaluating MLLM-based AD models. With MMR-AD, we reveal that the AD performance of current SOTA generalist MLLMs still falls far behind the industrial requirements. Based on MMR-AD, we also propose a baseline model, Anomaly-R1, which is a reasoning-based AD model that learns from the CoT data in MMR-AD and is further enhanced by reinforcement learning. Extensive experiments show that our Anomaly-R1 achieves remarkable improvements over generalist MLLMs in both anomaly detection and localization.

Multimodal Large Language Models (MLLMs) suffer from cross-modal hallucinations, where one modality inappropriately influences generation about another, leading to fabricated output. This exposes a more fundamental deficiency in modality-interaction control. To address this, we propose Modality-Adaptive Decoding (MAD), a training-free method that adaptively weights modality-specific decoding branches based on task requirements. MAD leverages the model's inherent ability to self-assess modality relevance by querying which modalities are needed for each task. The extracted modality probabilities are then used to adaptively weight contrastive decoding branches, enabling the model to focus on relevant information while suppressing cross-modal interference. Extensive experiments on CMM and AVHBench demonstrate that MAD significantly reduces cross-modal hallucinations across multiple audio-visual language models (7.8% and 2.0% improvements for VideoLLaMA2-AV, 8.7% and 4.7% improvements for Qwen2.5-Omni). Our approach demonstrates that explicit modality awareness through self-assessment is crucial for robust multimodal reasoning, offering a principled extension to existing contrastive decoding methods.

With the rapid development of Vision-Language Models (VLMs), there is a growing demand for automatic analysis of structured visual data. Charts and tables carry quantitative information through regular layouts, explicit numbers, and chart-specific reading patterns, yet current VLMs still underuse these properties, often causing value errors and unreliable analysis. To overcome these limitations, we propose Twin-T, a two-stage expert VLM for comprehensive chart-table tasks across Image, LaTeX, and Python. In stage 1, we propose a dual-head image encoder that can separate structural cues and fine details from input images. In stage 2, we propose MINT, a preference learning method that emphasizes numerical and keyword fidelity, as well as vision-text matching. Furthermore, we introduce a comprehensive TwintVQA benchmark with 17 chart types, 11 task types, 3 data formats, and short / medium / long QA settings. Our model narrows the gap between open-source and closed-source models on mainstream chart-table benchmarks, outperforming open-source models while even remaining competitive with GPT-4o and Gemini-2.5-Pro.

Mixture-of-Experts (MoE) has emerged as an effective approach to reduce the computational overhead of Transformer architectures by sparsely activating a subset of parameters for each token while preserving high model capacity. This paradigm has recently been extended to Vision-Language Models (VLMs), enabling scalable multi-modal understanding with reduced computational cost. However, the widely adopted deterministic top-K routing mechanism may overlook more optimal expert combinations and lead to expert overfitting. To address this limitation and improve the diversity of expert selection, we propose MoE-GRPO, a reinforcement learning (RL)-based framework for optimizing expert routing in MoE-based VLMs. Specifically, we formulate expert selection as a sequential decision-making problem and optimize it using Group Relative Policy Optimization (GRPO), allowing the model to learn adaptive expert routing policies through exploration and reward-based feedback. Furthermore, we introduce a modality-aware router guidance that enhances training stability and efficiency by discouraging the router from exploring experts that are infrequently activated for a given modality. Extensive experiments on multi-modal image and video benchmarks show that MoE-GRPO consistently outperforms standard top-K routing and its variants by promoting more diverse expert selection, thereby mitigating expert overfitting and enabling a task-level expert specialization.

The expansion of Transformers and the collection of high-quality multimodal datasets have propelled deep neural networks to achieve unprecedented performance in vision and language tasks. However, applying these advances is non-trivial in real-world applications. The extensive number of parameters complicates model updates, and real-world data often features a long-tailed distribution along with noisy labels. To address the above issues, we propose to explore the internal structure of the neural network for learning with sample relationships, rather than just increasing the number of model parameters. Specifically, we introduce RetFormer, a model enhanced with a multimodal knowledge base for storing world knowledge, and a retrieval cross-fusion module designed to establish robust multimodal sample relationships by leveraging content from the knowledge base. RetFormer establishes a robust relationship between image and text modalities by integrating information from external knowledge bases into the model's decision-making process, thus overcoming the limitations of traditional approaches on model size and datasets. Our experiments demonstrate the benefits of integrating large-scale image-text datasets into vision tasks and exemplify the importance of modeling the relationship between image and text modalities. We have evaluated our approach on the task of long-tailed recognition and learning with noisy labels and have shown that it achieves state-of-the-art accuracies.

Diffeomorphic image registration (DIR) seeks topology-preserving transformations and is fundamental in medical imaging. Existing DIR methods rely on integration schemes (e.g., scaling-and-squaring) and multiple regularizers to enforce invertibility. We introduce **SGDIR**, a continuous-time registration framework, parameterized by known time-embedded backbones, that models diffeomorphisms using only a single semigroup-based regularization, eliminating explicit integration and auxiliary constraints. We mathematically prove that this formulation directly learns the flow of an underlying ODE, inherently enforcing inverse and cycle consistencies. We evaluate on eight 2D and 3D MR and CT datasets. Under strict semigroup enforcement, our model achieves near-perfect diffeomorphism and significantly outperforms existing diffeomorphic methods, while remaining competitive with leading non-diffeomorphic deformable models. When the regularization is relaxed, the same architecture functions as a deformable method and substantially surpasses state-of-the-art non-diffeomorphic approaches in registration accuracy. These results demonstrate that continuous-time deformation modeling, guided solely by our semigroup-based regularization, yields a unified framework capable of both rigorously diffeomorphic mapping and state-of-the-art deformable registration.

Vision-Language Model (VLM) is now an important component to enable robust robot manipulation. Yet, using it to translate human instructions into an action-resolvable intermediate representation often needs a tradeoff between VLM-comprehensibility and generalizability. Inspired by context-free grammar structure, we design the Semantic Assembly representation named SEAM, by decomposing the intermediate representation into vocabulary and grammar. Doing so leads us to a concise vocabulary of semantically-rich operations and a VLM-friendly grammar for handling diverse unseen tasks. Also, we design a novel open-vocabulary segmentation paradigm with an in-context learning strategy to precisely localize fine-grained object parts for manipulation (e.g., cup handle, teapot opening) effectively with the shortest inference time over all state-of-the-art parallel works. We then formulate new metrics for action-generalizability and VLM-comprehensibility to evaluate mainstream representations, demonstrating the strong performance of SEAM on both aspects. Extensive real-world experiments further manifest the SOTA performance of SEAM under varying settings and tasks.

The utilization of temporal information has always been an open topic in the tracking community. However, existing trackers tend to employ more and more inputs or parameters for temporal learning, hindering their deployment in resource-constrained unmanned aerial vehicles (UAVs). More importantly, this raises ambiguity whether the performance gains come from the temporal learning itself, or come from the increased inputs and parameters. In this study, we advocate designing temporal learning components from a more balanced perspective that jointly considers performance gains and computational costs. To achieve this goal, we introduce a new evaluation metric, i.e., precision per FLOP (PPF). The PPF is introduced to quantify the tracking precision gains achieved by temporal learning components per unit of FLOP, thus enabling fair and efficiency-aware comparisons among these components and driving them toward more efficient designs. Based on this metric, we propose a low-cost yet effective temporal learning (LETL) approach to model contextual relationships. This approach continuously propagates and merges representative appearance tokens in video streams, allowing the tracker to efficiently capture the changing patterns of targets with relatively low costs. We integrate the LETL approach into existing one-stream frameworks, thereby building a simple yet effective tracker, namely LETrack, for robust UAV tracking. Extensive experimental demonstrate the superiority of our LETrack, and show that the proposed LETL approach achieves higher PPF scores. The Code is available at https://github.com/GXNU-ZhongLab/LETrack.

We introduce ProM3E, a probabilistic masked multimodal embedding model for any-to-any generation of multimodal representations for ecology. ProM3E is based on masked modality reconstruction in the embedding space, learning to infer missing modalities given a few context modalities. By design, our model supports modality inversion in the embedding space. The probabilistic nature of our model allows us to analyze the feasibility of fusing various modalities for given downstream tasks, essentially learning what to fuse. Using these features of our model, we propose a novel cross-modal retrieval approach that mixes inter-modal and intra-modal similarities to achieve superior performance across all retrieval tasks. We further leverage the hidden representation from our model to perform linear probing tasks and demonstrate the superior representation learning capability of our model. All our code, datasets and model will be released at https://vishu26.github.io/prom3e.

Current diffusion-based portrait animation models predominantly focus on enhancing visual quality and expression realism, while overlooking generation latency and real-time performance, which restricts their application range in the live streaming scenario. We propose PersonaLive, a novel diffusion-based framework towards streaming real-time portrait animation with multi-stage training recipes. Specifically, we first adopt hybrid implicit signals, namely implicit facial representations and 3D implicit keypoints, to achieve expressive image-level motion control. Then, a fewer-step appearance distillation strategy is proposed to eliminate appearance redundancy in the denoising process, greatly improving inference efficiency. Finally, we introduce an autoregressive micro-chunk streaming generation paradigm equipped with a sliding training strategy and a historical keyframe mechanism to enable low-latency and stable long-term video generation. Extensive experiments demonstrate that PersonaLive achieves state-of-the-art performance with up to **7-22**xspeedup over prior diffusion-based portrait animation models. The code will be publicly available.

We propose SLARM, a feed-forward model that unifies dynamic scene reconstruction, semantic understanding, and real-time streaming inference. SLARM captures complex, non-uniform motion through higher-order motion modeling, trained solely on differentiable renderings without any flow supervision. Besides, SLARM distills semantic features from LSeg to obtain language-aligned representations. This design enables semantic querying via natural language, and the tight coupling between semantics and geometry further enhances the accuracy and robustness of dynamic reconstruction. Moreover, SLARM processes image sequences using window-based causal attention, achieving stable, low-latency streaming inference without accumulating memory cost. Within this unified framework, SLARM achieves state-of-the-art results in dynamic estimation, rendering quality, and scene parsing, improving motion accuracy by 21%, reconstruction PSNR by 1.6 dB, and segmentation mIoU by 20% over existing methods.