CVPR 2026 Papers

Chest X-ray radiography (CXR) is an essential medical imaging technique for disease diagnosis. However, as 2D projectional images, CXRs are limited by structural superposition and hence fail to capture 3D anatomies. This limitation makes representation learning and disease diagnosis challenging. To address this challenge, we propose a novel CXR world model named X-WIN, which distills volumetric knowledge from chest computed tomography (CT) by learning to predict its 2D projections in latent space. The core idea is that a world model with internalized knowledge of 3D anatomical structure can predict CXRs under various transformations in 3D space. During projection prediction, we introduce an affinity-guided contrastive alignment loss that leverages mutual similarities to capture rich, correlated information across projections from the same volume. To improve model adaptability, we incorporate real CXRs into training through masked image modeling and employ a domain classifier to encourage statistically similar representations for real and simulated CXRs. Comprehensive experiments show that X-WIN outperforms existing foundation models on diverse downstream tasks using linear probing and few-shot fine-tuning. X-WIN also demonstrates the ability to render 2D projections for reconstructing a 3D CT volume.

Vision-Language-Action (VLA) models for autonomous driving often hit a performance plateau during Reinforcement Learning (RL) optimization. This stagnation arises from exploration capabilities constrained by previous Supervised Fine-Tuning (SFT), leading to "persistent failures" in long-tail scenarios. In these critical situations, all explored actions yield a zero-value driving score. This information-sparse reward signals a failure, yet fails to identify its root cause--whether it is due to incorrect planning, flawed reasoning, or poor trajectory execution. To address this limitation, we propose **VLA** with **E**xplicit **L**earning from **F**ailures (**ELF-VLA**), a framework that augments RL with structured diagnostic feedback. Instead of relying on a vague scalar reward, our method produces detailed, interpretable reports that identify the specific failure mode. The VLA policy then leverages this explicit feedback to generate a **Feedback-Guided Refinement**. By injecting these corrected, high-reward samples back into the RL training batch, our approach provides a targeted gradient, which enables the policy to solve critical scenarios that unguided exploration cannot. Extensive experiments demonstrate that our method unlocks the latent capabilities of VLA models, achieving state-of-the-art (SOTA) performance on the public Navsim benchmark for overall PDMS, EPDMS score and high-level planning accuracy.

Open-set fine-grained retrieval (OSFR) is a challenging task where models must generalize to unseen subcategories. Existing methods often fail this, as they embed category-specific semantics from closed-set training labels. Recently, diffusion transformers (DiT) have shown promise by encoding attribute-centric, generative curriculum knowledge that is agnostic to these labels. However, the vanilla DiT is not optimized for fine-grained visual discrepancies and its massive size makes deployment infeasible. To solve this, we propose DiT-Distill, a framework to first refine and then distill this knowledge. We introduce a conditional discrepancy refinement strategy to fine-tune the DiT, forcing it to focus on discrepancy-aware, attribute-centric details rather than holistic context. Subsequently, a generative curriculum distillation mechanism transfers this refined, hierarchical knowledge from multiple diffusion timesteps of the DiT into a lightweight backbone using a generative infusion module and a curriculum alignment loss. This process results in an efficient retrieval model that enables DiT-free inference. Extensive experiments show DiT-Distill achieves state-of-the-art performance on open-set fine-grained datasets.

We introduce Diff4Splat, a feed-forward framework for dynamic scene generation from a single image. Our method synergizes the powerful generative priors of video diffusion models with geometric and motion constraints learned from a large-scale 4D dataset. Given a single image, a camera trajectory, and an optional text prompt, our model directly predicts a dynamic scene represented by a deformable 3D Gaussian field. This approach captures appearance, geometry, and motion in a single pass, eliminating the need for test-time optimization or post-hoc processing. At the core of our framework is a video latent transformer that enhances existing video diffusion models, enabling them to jointly model spatio-temporal dependencies and predict 3D Gaussian Primitives over time. Supervised by objectives targeting appearance fidelity, geometric accuracy, and motion consistency, Diff4Splat generates high-fidelity dynamic scenes within 30 seconds. We demonstrate the effectiveness of Diff4Splat across video generation, novel view synthesis, and geometry extraction, where it matches or surpasses optimization-based methods for dynamic scene synthesis while being significantly more efficient.

Fine-grained bird species identification in the wild is frequently unanswerable from a single image: key cues may be non-visual (e.g. vocalization), or obscured due to occlusion, camera angle, or low resolution. Yet today's multimodal systems are typically judged on answerable, in-schema cases, encouraging confident guesses rather than principled abstention. We propose the RealBirdID benchmark: given an image of a bird, a system should either answer with a species or abstain with a concrete, evidence-based rationale: "requires vocalization," "low quality image," or "view obstructed". For each genus, the dataset includes a validation split composed of curated unanswerable examples with labeled rationales, paired with a companion set of clearly answerable instances. We find that (1) the species identification on the answerable set is challenging for a variety of open-source and proprietary models (\leq13% accuracy for MLLMs including GPT-5 and Gemini-2.5 Pro), (2) models with greater classification ability are not necessarily more calibrated to abstain from unanswerable examples, and (3) that MLLMs generally fail at providing correct reasons even when they do abstain. RealBirdID establishes a focused target for abstention-aware fine-grained recognition and a recipe for measuring progress.

The integration of vision-language models (VLMs) is driving a new generation of embodied agents capable of operating in human-centered environments. However, as deployment expands, these systems face growing safety risks, particularly when executing hazardous instructions. Current safety evaluation benchmarks remain limited: they cover only narrow scopes of hazards and focus primarily on final outcomes, neglecting the agent's full perception-planning-execution process and thereby obscuring critical failure modes. Therefore, we present SAFE, a benchmark for systematically assessing the safety of embodied VLM agents on hazardous instructions. SAFE comprises three components: SAFE-THOR, an extensible adversarial simulation sandbox with a universal adapter that maps high-level VLM outputs to low-level embodied controls, supporting diverse agent workflow integration; SAFE-VERSE, a risk-aware task suite inspired by Asimov's Three Laws of Robotics, comprising 45 adversarial scenarios, 1,350 hazardous tasks, and 9,900 instructions that span risks to humans, environments, and agents; and SAFE-DIAGNOSE, a multi-level and fine-grained evaluation protocol measuring agent performance across perception, planning, and execution. Applying SAFE to nine state-of-the-art VLMs and two embodied agent workflows, we uncover systematic failures in translating hazard recognition into safe planning and execution. Our findings reveal fundamental limitations in current safety alignment and demonstrate the necessity of a comprehensive, multi-stage evaluation for developing safer embodied intelligence.

While recent vision-language models (VLMs) demonstrate strong image understanding, their ability to "think with images," i.e., to reason through multi-step visual interactions, remains limited. We introduce VISTA-Gym, a scalable training environment for incentivizing tool-integrated visual reasoning capabilities in VLMs. VISTA-Gym unifies diverse real-world multimodal reasoning tasks (7 tasks from 13 datasets in total) with a standardized interface for visual tools (e.g., grounding, parsing), executable interaction loops, verifiable feedback signals, and efficient trajectory logging, enabling visual agentic reinforcement learning at scale. While recent VLMs exhibit strong text-only reasoning, both proprietary and open-source models still struggle with tool selection, invocation, and coordination. With VISTA-Gym, we train VISTA-R1 to interleave tool-use with agentic reasoning via multi-turn trajectory sampling and end-to-end reinforcement learning. Extensive experiments across 11 public reasoning-intensive VQA benchmarks show that VISTA-R1-8B outperforms state-of-the-art baselines with similar sizes by 9.51%-18.72%, demonstrating VISTA-Gym as an effective training ground to unlock the tool-integrated reasoning capabilities for VLMs.

Accurate and fast localization is vital for safe autonomous navigation in GPS-denied areas. Fine-Grained Cross-View Geolocalization (FG-CVG) aims to estimate the precise 2-Degree-of-Freedom (2-DoF) location of a ground image relative to a satellite image. However, current methods force a difficult trade-off, with high-accuracy models being slow for real-time use. In this paper, we introduce GeoFlow, a new approach that offers a lightweight and highly efficient framework that breaks this accuracy-speed trade-off. Our technique learns a direct probabilistic mapping, predicting the displacement (in distance and direction) required to correct any given location hypothesis. This is complemented by our novel inference algorithm, Iterative Refinement Sampling (IRS). Instead of trusting a single prediction, IRS refines a population of hypotheses, allowing them to iteratively 'flow' from random starting points to a robust, converged consensus. Even its iterative nature, this approach offers flexible inference-time scaling, allowing a direct trade-off between performance and computation without any re-training. Experiments on the KITTI and VIGOR datasets show that GeoFlow achieves state-of-the-art efficiency, running at real-time speeds of 29 FPS while maintaining competitive localization accuracy. This work opens a new path for the development of practical real-time geolocalization systems.

Recent text-to-video generation models have made remarkable progress in visual realism, motion fidelity, and text-video alignment, yet they still struggle to produce socially coherent behavior. Unlike humans, who readily infer intentions, beliefs, emotions, and social norms from brief visual cues, current models often generate literal scenes without capturing the underlying causal and psychological dynamics. To systematically assess this limitation, we introduce the first benchmark for social reasoning in video generation. Grounded in developmental and social psychology, the benchmark covers thirty classic social cognition paradigms spanning seven core dimensions: mental-state inference, goal-directed action, joint attention, social coordination, prosocial behavior, social norms, and multi-agent strategy. To operationalize these paradigms, we build a fully training-free agent-based pipeline that distills the reasoning structure of each paradigm, synthesizes diverse video-ready scenarios, enforces conceptual neutrality and difficulty control through cue-based critique, and evaluates generated videos with a high-capacity VLM judge along five interpretable dimensions of social reasoning. Using this framework, we conduct the first large-scale evaluation of seven state-of-the-art video generation systems. Results show a clear gap between surface-level plausibility and deeper social reasoning, suggesting that current models remain limited in their ability to generate socially grounded behavior. https://github.com/Gloria2tt/SVBench-Evaluation

Significant advancements made in reconstructing hands from images have delivered accurate single-frame estimates, yet they often lack physics consistency and provide no notion of how confidently the motion satisfies physics. In this paper, we propose a novel physics-aware conditional diffusion framework that refines noisy pose sequences into physically plausible hand motion while estimating the physics variance in motion estimates. Building on a MeshCNN-Transformer backbone, we formulate Euler-Lagrange dynamics for articulated hands. Unlike prior works that enforce zero residuals, we treat the resulting dynamic residuals as virtual observables to more effectively integrate physics. Through a last-layer Laplace approximation, our method produces per-joint, per-time variances that measure physics consistency and offers interpretable variance maps indicating where physical consistency weakens. Experiments on two well-known hand datasets show consistent gains over strong image-based initializations and competitive video-based methods. Qualitative results confirm that our variance estimations are aligned with the physical plausibility of the motion in image-based estimates.

While existing equivariant methods enhance data efficiency, they suffer from high computational intensity, reliance on single-modality inputs, and instability when combined with fast-sampling methods. In this work, we propose E3Flow, a novel framework that addresses the critical limitations of equivariant diffusion policies. E3Flow overcomes these challenges, successfully unifying efficient rectified flow with stable, multi-modal equivariant learning for the first time. Our framework is built upon spherical harmonic representations to ensure rigorous SO(3) equivariance. We introduce a novel invariant Feature Enhancement Module (FEM) that dynamically fuses hybrid visual modalities (point clouds and images), injecting rich visual cues into the spherical harmonic features. We evaluate E3Flow on 8 manipulation tasks from the MimicGen and further conduct 4 real-world experiments to validate its effectiveness in physical environments. Simulation results show that E3Flow achieves a 3.12% improvement in average success rate over the state-of-the-art Spherical Diffusion Policy (SDP) while simultaneously delivering a 7x inference speedup. E3Flow thus demonstrates a new and highly effective trade-off between performance, efficiency, and data efficiency for robotic policy learning. Code: https://github.com/zql-kk/E3Flow.

Vision-Language Models (VLMs), such as CLIP, have demonstrated remarkable zero-shot capabilities in various computer vision tasks. However, their application to medical imaging remains challenging due to the high variability and complexity of medical data. Specifically, medical images often exhibit significant domain shifts caused by various confounders, including equipment differences, procedure artifacts, and imaging modes, which can lead to poor generalization when models are applied to unseen domains. To address this limitation, we propose Multimodal Causal-Driven Representation Learning (MCDRL), a novel framework that integrates causal inference with the VLM to tackle domain generalization in medical image segmentation. MCDRL is implemented in two steps: first, it leverages CLIP's cross-modal capabilities to identify candidate lesion regions and construct a confounder dictionary through text prompts, specifically designed to represent domain-specific variations; second, it trains a causal intervention network that utilizes this dictionary to identify and eliminate the influence of these domain-specific variations while preserving the anatomical structural information critical for segmentation tasks. Extensive experiments demonstrate that MCDRL consistently outperforms competing methods, yielding superior segmentation accuracy and exhibiting robust generalizability.

In bandwidth-limited online video streaming, videos are usually downsampled and compressed. Although recent online video super-resolution (online VSR) approaches achieve promising results, they are still compute-intensive and fall short of real-time processing at higher resolutions, due to complex motion estimation for alignment and redundant processing of consecutive frames. To address these issues, we propose a compressed-domain-aware network (CDA-VSR) for online VSR, which utilizes compressed-domain information, including motion vectors, residual maps, and frame types to balance quality and efficiency. Specifically, we propose a motion-vector-guided deformable alignment module that uses motion vectors for coarse warping and learns only local residual offsets for fine-tuned adjustments, thereby maintaining accuracy while reducing computation. Then, we utilize a residual map gated fusion module to derive spatial weights from residual maps, suppressing mismatched regions and emphasizing reliable details. Further, we design a frame-type-aware reconstruction module for adaptive compute allocation across frame types, balancing accuracy and efficiency. On the REDS4 dataset, our CDA-VSR surpasses the state-of-the-art method TMP, with a maximum PSNR improvement of 0.13 dB while delivering more than double the inference speed. The code will be released at https://github.com/sspBIT/CDA-VSR.

Autoregressive (AR) vision-language models (VLMs) have long dominated multimodal understanding, reasoning, and graphical user interface (GUI) grounding. Recently, discrete diffusion vision-language models (DVLMs) have shown strong performance in multimodal reasoning, offering bidirectional attention, parallel token generation, and iterative refinement. However, their potential for GUI grounding remains unexplored. In this work, we evaluate whether discrete DVLMs can serve as a viable alternative to AR models for GUI grounding. We adapt LLaDA-V for single-turn action and bounding-box prediction, framing the task as text generation from multimodal input. To better capture the hierarchical structure of bounding-box geometry, we propose a hybrid masking schedule that combines linear and deterministic masking, improving grounding accuracy by up to 6.1 points in Step Success Rate (SSR) over the GUI-adapted LLaDA-V trained with linear masking. Evaluations on four datasets spanning web, desktop, and mobile interfaces show that the adapted diffusion model with hybrid masking consistently outperforms the linear-masked variant and performs competitively with autoregressive counterparts despite limited pretraining. Systematic ablations reveal that increasing diffusion steps, generation length, and block length improves accuracy but also increases latency, with accuracy plateauing beyond a certain number of diffusion steps. Expanding the training data with diverse GUI domains further reduces latency by about 1.3 seconds and improves grounding accuracy by an average of 20 points across benchmarks. These results demonstrate that discrete DVLMs are a promising modeling framework for GUI grounding and represent an important step toward diffusion-based GUI agents.

GUI grounding, which translates natural language instructions into precise pixel coordinates, is essential for developing practical GUI agents. However, we observe that existing grounding models exhibit significant coordinate prediction instability--minor visual perturbations (e.g., cropping a few pixels) can drastically alter predictions, flipping results between correct and incorrect. This instability severely undermines model performance, especially for samples with high-resolution and small UI elements. To address this issue, we propose Multi-View Prediction (MVP), a training-free framework that enhances grounding performance through multi-view inference. Our key insight is that while single-view predictions may be unstable, aggregating predictions from multiple carefully cropped views can effectively distinguish correct coordinates from outliers. MVP comprises two components: (1) Attention-Guided View Proposal, which derives diverse views guided by instruction-to-image attention scores, and (2) Multi-Coordinates Clustering, which ensembles predictions by selecting the centroid of the densest spatial cluster. Extensive experiments demonstrate MVP's effectiveness across various models and benchmarks. Notably, on ScreenSpot-Pro, MVP boosts UI-TARS-1.5-7B to 56.1%, GTA1-7B to 61.7%, Qwen3VL-8B-Instruct to 65.3%, and Qwen3VL-32B-Instruct to 74.0%.

Bird's-Eye-View (BEV) detection has become a dominant paradigm for 3D object detection in autonomous driving, due to its strong perception capability. However, most existing methods mainly focus on constructing high-quality BEV feature representations, while neglecting the design of task-specific detection heads. In practice, they directly adopt the center-based head originally developed for 2D detection, without any specific optimization. This leads to three inherent limitations: (i) a geometric mismatch between the Gaussian kernel used for classification and the real BEV object, (ii) degraded end-to-end performance without Non-Maximum Suppression (NMS), and (iii) sparse supervisory signals. To address these issues, we propose Spe-BEVHead, a detection head specifically tailored for BEV 3D object detection. Spe-BEVHead introduces three BEV-specific adaptations: (1) a Rotated Box Kernel that generates geometry-aligned classification weights, (2) a Local Response Refinement Module (LRRM) that suppresses non-peak responses and improves end-to-end performance, and (3) a dual-branch architecture that provides richer supervisory signals to promote more robust learning while inherently preserving the performance for end-to-end inference. Extensive experiments show that Spe-BEVHead can be seamlessly integrated into existing BEV backbones, delivering direct performance gains while retaining competitive performance under the challenging end-to-end setting.

Safe highway autonomy for heavy trucks remains an open and unsolved challenge: due to long braking distances, scene understanding of hundreds of meters is required for anticipatory planning and to allow safe braking margins. However, existing driving datasets primarily cover urban scenes, with perception effectively limited to short ranges of only up to 100 meters. To address this gap, we introduce TruckDrive, a highway-scale multimodal driving dataset, captured with a sensor suite purpose-built for long range sensing: seven long-range FMCW LiDARs measuring range and radial velocity, three high-resolution short-range LiDARs, eleven 8MP surround cameras with varying focal lengths and ten 4D FMCW radars. The dataset offers 475 thousands samples with 165 thousands densely annotated frames for driving perception benchmarking up to 1,000 meters for 2D detection and 400 meters for 3D detection, depth estimation, tracking, planning and end to end driving over 20 seconds sequences at highway speeds. We find that state-of-the-art autonomous driving models do not generalize to ranges beyond 150 meters, with drops between 31% and 99% in 3D perception tasks, exposing a systematic long-range gap that current architectures and training signals cannot close.

Reading measurement instruments is effortless for humans and requires relatively little domain expertise, yet it remains surprisingly challenging for current vision-language models (VLMs) as we find in preliminary evaluation. In this work, we introduce MeasureBench, a benchmark on visual measurement reading covering both real-world and synthesized images of various types of measurements, along with an extensible pipeline for data synthesis. Our pipeline procedurally generates a specified type of gauge with controllable visual appearance, enabling scalable variation in key details such as pointers, scales, fonts, lighting, and clutter. Evaluation on popular proprietary and open-weight VLMs shows that even the strongest frontier VLMs struggle with measurement reading in general. We have also conducted preliminary experiments with reinforcement finetuning (RFT) over synthetic data, and find a significant improvement on both in-domain synthetic subset and real-world images. Our analysis highlights a fundamental limitation of current VLMs in fine-grained spatial grounding. We hope this resource and our code releases can help future advances on visually grounded numeracy and precise spatial perception of VLMs, bridging the gap between recognizing numbers and measuring the world.

Few-shot object detection (FSOD) is challenging due to unstable optimization and limited generalization arising from the scarcity of training samples. To address these issues, we propose a hybrid ensemble decoder that enhances generalization during fine-tuning. Inspired by ensemble learning, the decoder comprises a shared hierarchical layer followed by multiple parallel decoder branches, where each branch employs denoising queries either inherited from the shared layer or newly initialized to encourage prediction diversity. This design fully exploits pretrained weights without introducing additional parameters, and the resulting diverse predictions can be effectively ensembled to improve generalization. We further leverage a unified progressive fine-tuning framework with a plateau-aware learning rate schedule, which stabilizes optimization and achieves strong few-shot adaptation without complex data augmentations or extensive hyperparameter tuning. Extensive experiments on CD-FSOD, ODinW-13, and RF100-VL validate the effectiveness of our approach. Notably, on RF100-VL, which includes 100 datasets across diverse domains, our method achieves an average performance of 41.9 in the 10-shot setting, significantly outperforming the recent approach SAM3, which obtains 35.7. We further construct a mixed-domain test set from CD-FSOD to evaluate robustness to out-of-distribution (OOD) samples, showing that our proposed modules lead to clear improvement gains. These results highlight the effectiveness, generalization, and robustness of the proposed method. Code is available at: https://github.com/Intellindust-AI-Lab/FT-FSOD.