CVPR 2026 Papers

We introduce EMMA, a physics-informed multimodal framework that recovers all identifiable dynamical parameters of a system directly from raw video, audio, and image-based time-series observations. Unlike prior video-only approaches that struggle with occluded states, hidden actuation inputs, or assumptions about known initial conditions and coordinate frames, EMMA performs joint inference of explicit parameters, implicit dynamical components, and calibration invariants within a unified continuous-time model. EMMA leverages a Liquid Time-Constant (LTC) network to learn latent dynamics from heterogeneous modalities while a physics-constrained loss enforces consistency with the governing differential equations. A unified feature pipeline enables consistent alignment across video trajectories, acoustic signatures, and chart-derived measurements, allowing EMMA to estimate parameters under forced, implicit, and multivariate dynamics without requiring segmentation masks, differentiable rendering, or specialized sensors. Across 100+ scenarios including five standard dynamical benchmarks (75 Delfys videos), real-world rover and quadrotor systems with hidden inputs, and simulation-chart case studies spanning biological and chaotic systems, EMMA delivers robust multi-parameter recovery and significantly outperforms existing single-modality and equation-discovery baselines. Our results establish EMMA as a general, scalable solution for physics-consistent model extraction from opportunistic multimodal data.

Zero-Shot Composed Image Retrieval (ZS-CIR) aims to retrieve target images given a multimodal query (comprising a reference image and a modification text), without training on annotated triplets. Existing methods typically convert the multimodal query into a single modality--either as an edited caption for Text-to-Image retrieval (T2I) or as an edited image for Image-to-Image retrieval (I2I). However, T2I often loses fine-grained visual details, while I2I struggles with complex semantic modifications. To effectively leverage their complementary strengths under diverse query intents, we propose WISER, a training-free framework that unifies T2I and I2I via a "retrieve-verify-refine" pipeline, explicitly modeling intent awareness and uncertainty awareness. Specifically, WISER first performs Wider Search by generating both edited captions and images for parallel retrieval to broaden the candidate pool. Then, it conducts Adaptive Fusion with a verifier to assess retrieval confidence, triggering refinement for uncertain retrievals, and dynamically fusing the dual-path for reliable ones. For uncertain retrievals, WISER generates refinement suggestions through structured self-reflection to guide the next retrieval round toward Deeper Thinking. Extensive experiments show the superior performance of WISER across multiple benchmarks, achieving relative improvements of 45% on CIRCO (mAP@5) and 57% on CIRR (Recall@1) over existing training-free methods. Code is released at https://github.com/Physicsmile/WISER.

The significance of cross-view 3D geometric modeling capabilities for autonomous driving is self-evident, yet existing Vision-Language Models (VLMs) inherently lack this capability, resulting in their mediocre performance. While some promising approaches attempt to mitigate this by constructing Q&A data for auxiliary training, they still fail to fundamentally equip VLMs with the ability to comprehensively handle diverse evaluation protocols. We thus chart a new course, advocating for the infusion of VLMs with the cross-view geometric grounding of mature 3D foundation models, closing this critical capability gap in autonomous driving. In this spirit, we propose a novel architecture, VGGDrive, which empowers Vision-language models with cross-view Geometric Grounding for autonomous Driving. Concretely, to bridge the cross-view 3D geometric features from the frozen visual 3D model with the VLM's 2D visual features, we introduce a plug-and-play Cross-View 3D Geometric Enabler (CVGE). The CVGE decouples the base VLM architecture and effectively empowers the VLM with 3D features through a hierarchical adaptive injection mechanism. Extensive experiments show that VGGDrive enhances base VLM performance across five autonomous driving benchmarks, including tasks like cross-view risk perception, motion prediction, and trajectory planning. It's our belief that mature 3D foundation models can empower autonomous driving tasks through effective integration, and we hope our initial exploration demonstrates the potential of this paradigm to the autonomous driving community.

No-reference image quality assessment (NR IQA) has recently benefited from deep and multimodal models, yet many SOTA systems still violate at least one basic requirement: they either discard critical quality cues via aggressive resizing, fail to generalize across resolutions, cannot be jointly trained on heterogeneous IQA datasets with mismatched MOS scales, or require prohibitive computation. We present ReLIQS, a model for Resolution-agnostic Learning for Image Quality with Saliency, which is resolution-agnostic, preserves original-resolution quality cues, learns from multiple subjective studies, and remains computationally efficient and budget-adaptive. ReLIQS is a CLIP-based multiscale patch-driven architecture that learns both where to look and how to judge quality. Fixed-size patches are sampled across multiple resolutions, including the original resolution, and encoded with a CLIP vision backbone. A lightweight Perceptual Importance Estimator then predicts IQA-specific importance maps to select a small set of informative patches, and a Quality Aspect Module aggregates their embeddings into a single image-level score. Across authentic, synthetic, and AIGC benchmarks spanning diverse resolutions and distortions, ReLIQS generalizes better than strong CNN-, CLIP-, and MLLM-based baselines with matching or reduced computational cost.

The pursuit of autonomous agents with predictive cognitive world models is hindered by a fundamental flaw in current vision-language models (VLMs): they lack cognitive inertia. Operating on isolated snapshots, these models cannot form a temporally coherent world view, leading to erratic decision jitter and a failure to execute complex, multi-step maneuvers. To remedy this, we introduce CogDriver, a framework designed to build a coherent world model by instilling this crucial cognitive property. Our work makes two key contributions: (1) We present CogDriver-Data, a large-scale vision-language-action dataset whose narrative annotations provide the supervisory signal for learning the temporal dynamics of a world model. (2) We develop the CogDriver-Agent, an architecture featuring a sparse temporalmemory to maintain a stable internal state, the foundation of a world model. This is enabled by a spatiotemporal knowledge distillation approach that explicitly teaches decision coherence. Comprehensive experiments validate our paradigm: CogDriver-Agent achieves a 22% increase in the closed-loop Driving Score on Bench2Drive and a 21% reduction in mean L2 error on nuScenes, establishing a new state-of-the-art. These significant gains in both long-term decision-making and imitation accuracy provide strong evidence that our agent is developing a more stable internal world model.

Long-horizon navigation in complex urban environments relies heavily on continuous human operation, which leads to fatigue, reduced efficiency, and safety concerns. Shared autonomy, where a Vision-Language AI agent and a human operator collaborate on maneuvering the mobile machine, presents a promising solution to address these issues. However, existing shared autonomy methods often require humans and AI to operate within the same action space, leading to high cognitive overhead. We present Assistive Urban Robot Autonomy (AURA), a new multi-modal framework that decomposes urban navigation into high-level human instruction and low-level AI control. AURA incorporates a Spatial-Aware Instruction Encoder to align various human instructions with visual and spatial context. To facilitate training, we construct MM-CoS, a large-scale dataset comprising teleoperation and vision-language descriptions. Experiments in simulation and the real world demonstrate that AURA effectively follows human instructions, reduces manual operation effort, and improves navigation stability, while enabling online adaptation. Moreover, under similar takeover conditions, our shared autonomy framework reduces the frequency of takeovers by more than 44%. Demo video and more detail are provided in the project page.

Video Large Language Models (VLLMs) encounter significant computational challenges due to the large volume of visual tokens generated from multiple frames. Existing visual token pruning methods fail to account for the uneven spatiotemporal information density, thus squandering scarce token budgets on regions with low information density. In this paper, we propose a training-free Metadata-guided Token Merging framework (MeToM) that leverages intrinsic video metadata to adaptively allocate budgets and merge visual tokens based on content complexity. Specifically, MeToM exploits residual data from codec metadata as spatial information density cues. It merges less informative regions during tokenization, avoiding redundant encoding and improving the efficiency of the visual encoder. Additionally, MeToM captures temporal variations in information density by utilizing the average Group of Pictures (GoP) packet size to represent scene complexity. This mechanism enables dynamic per-frame token allocation across time, assigning more tokens to content-complex frames and fewer tokens to information-sparse ones. Finally, we merge low-contribution visual tokens via multi-layer attention to reduce prefill FLOPs and the visual KV-cache footprint inside the LLM. Extensive experimental results demonstrate that MeToM outperforms prior state-of-the-art methods. Notably, it achieves a 2.65xinference speedup over the baseline VLLM without sacrificing accuracy.

Unified models aim to support both understanding and generation by encoding images into discrete tokens and processing them alongside text within a single autoregressive framework. This unified design offers architectural simplicity and cross-modal synergy, which facilitates shared parameterization, consistent training objectives, and seamless transfer between modalities. However, the large number of visual tokens required by such models introduces substantial computation and memory overhead, and this inefficiency directly hinders deployment in resource constrained scenarios such as embodied AI systems. In this work, we propose a unified token compression algorithm UniCompress that significantly reduces visual token count while preserving performance on both image understanding and generation tasks. Our method introduces a plug-in compression and decompression mechanism guided with learnable global meta tokens. The framework is lightweight and modular, enabling efficient integration into existing models without full retraining. Experimental results show that our approach reduces image tokens by up to (4x), achieves substantial gains in inference latency and training cost, and incurs only minimal performance degradation, which demonstrates the promise of token-efficient unified modeling for real world multimodal applications.

3D medical image classification is essential to modern clinical workflows. Medical foundation models (FMs) have emerged as a promising approach for scaling to new tasks, yet current research suffers from three critical pitfalls: data-regime bias, suboptimal adaptation, and insufficient task coverage. In this paper, we address these pitfalls and introduce AnyMC3D, a scalable 3D classifier adapted from 2D FMs. It allows efficient scaling to new tasks by adding only lightweight plugins ( 1M parameters per task) to a single frozen backbone. Besides, this versatile framework also supports multi-view inputs, auxiliary pixel-level supervision, and interpretable heatmap generation. We establish a comprehensive benchmark of 12 tasks covering diverse pathologies, anatomies, and modalities and systematically evaluate state-of-the-art 3D classification techniques. Our analysis reveals several key insights: (1) effective adaptation is critical to unlock FM potential, (2) general-purpose FMs can match medical-specific FMs if properly adapted, and (3) 2D-based methods surpass 3D architectures for 3D classification. For the first time, we demonstrate the feasibility of achieving state-of-the-art performance across diverse applications using a single scalable framework (e.g., 1st place in the *** challenge), eliminating the need for separate task-specific 3D models.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable proficiency in multimodal tasks.Despite their impressive performance, MLLMs suffer from the modality imbalance issue, where visual information is often underutilized compared to textual representations in deeper layers, leading to degraded visual performance or hallucinations.This issue stems from the predominant reliance on next-text-token-prediction during training, which fails to provide direct visual supervisory signals, resulting in progressive homogenization of visual representations throughout the layers.To this end, we propose Latent Visual Reconstruction (LaVer), a novel training framework that facilitates MLLMs in learning more discriminative visual representations via masked image modeling in the joint latent semantic space of LLM.Our method offers direct visual activation to MLLMs, which exhibit increased visual attention allocation, indicating enhanced utilization of visual information.Extensive experiments across diverse benchmarks prove the superiority of our approach in various scenarios, especially those requiring dense visual capabilities.Code will be publicly available upon publication.

Multimodal large language models (MLLMs) achieve remarkable progress in cross-modal perception and reasoning, yet a fundamental question remains unresolved: should the vision encoder be fine-tuned or frozen? Despite the success of models such as LLaVA and Qwen-VL, inconsistent design choices and heterogeneous training setups hinder a unified understanding of visual fine-tuning (VFT) in MLLMs. Through a configuration-aligned benchmark, we find that existing VFT methods fail to consistently outperform the frozen baseline across multimodal tasks. Our analysis suggests that this instability arises from visual preference conflicts, where the context-agnostic nature of vision encoders induces divergent parameter updates under diverse multimodal context. To address this issue, we propose the Context-aware Visual Fine-tuning (CoVFT) framework, which explicitly incorporates multimodal context into visual adaptation. By integrating a Context Vector Extraction (CVE) and a Contextual Mixture-of-Experts (CoMoE) module, CoVFT decomposes conflicting optimization signals and enables stable, context-sensitive visual updates. Extensive experiments on 12 multimodal benchmarks demonstrate that CoVFT achieves state-of-the-art performance with superior stability. Notably, fine-tuning a 7B MLLM with CoVFT surpasses the average performance of its 13B counterpart, revealing substantial untapped potential in visual encoder optimization within MLLMs. Code is available at https://github.com/weeknan/CoVFT.

Despite rapid developments and widespread applications of MLLM agents, they still struggle with long-form video understanding (LVU) tasks, which are characterized by high information density and extended temporal spans. Recent research on LVU agents demonstrates that simple task decomposition and collaboration mechanisms are insufficient for long-chain reasoning tasks. Moreover, directly reducing the time context through embedding-based retrieval may lose key information of complex problems. In this paper, we propose Symphony, a multi-agent system, to alleviate these limitations. By emulating human cognition patterns, Symphony decomposes LVU into fine-grained subtasks and incorporates a deep reasoning collaboration mechanism enhanced by reflection, effectively improving the reasoning capability. Additionally, Symphony provides a VLM-based grounding approach to analyze LVU tasks and assess the relevance of video segments, which significantly enhances the ability to locate complex problems with implicit intentions and large temporal spans. Experimental results show that Symphony achieves state-of-the-art performance on LVBench, LongVideoBench, VideoMME, and MLVU, with a 5.0% improvement over the prior state-of-the-art method on LVBench. Code is available at https://github.com/Haiyang0226/Symphony.

Vision-Language Models (VLMs) have demonstrated strong capability in a wide range of tasks such as visual recognition, document parsing, and visual grounding. Nevertheless, recent work shows that while VLMs often manage to capture the correct image region corresponding to the question, they do not necessarily produce the correct answers. In this work, we demonstrate that this misalignment could be attributed to suboptimal information flow within VLMs, where text tokens distribute too much attention to irrelevant visual tokens, leading to incorrect answers. Based on the observation, we show that modulating the information flow during inference can improve the perception capability of VLMs. The idea is that text tokens should only be associated with important visual tokens during decoding, eliminating the interference of irrelevant regions. To achieve this, we propose a token dynamics-based method to determine the importance of visual tokens, where visual tokens that exhibit distinct activation patterns during different decoding stages are viewed as important. We apply our approach to representative open-source VLMs and evaluate on various datasets, including visual question answering, visual grounding and counting, optical character recognition, and object hallucination. The results show that our approach significantly improves the performance of baselines. Project page: https://cxliu0.github.io/AIF/.

Recent advances in Vision-Language Models (VLMs) have enabled video-instructed robotic programming, allowing agents to interpret video demonstrations and generate executable control code. We formulate video-instructed robotic programming as a cross-domain adaptation problem, where perceptual and physical differences between demonstration and deployment induce procedural mismatches. However, current VLMs lack the procedural understanding needed to reformulate causal dependencies and achieve task-compatible behavior under such domain shifts. We introduce NeSyCR, a neurosymbolic counterfactual reasoning framework that enables verifiable adaptation of task procedures, providing a reliable synthesis of code policies. NeSyCR abstracts video demonstrations into symbolic trajectories that capture the underlying task procedure. Given deployment observations, it derives counterfactual states that reveal cross-domain incompatibilities. By exploring the symbolic state space with verifiable checks, NeSyCR proposes procedural revisions that restore compatibility with the demonstrated procedure. NeSyCR achieves a 31.14% improvement in task success over the strongest baseline Statler, showing robust cross-domain adaptation across both simulated and real-world manipulation tasks.

Test-time prompt tuning (TPT) has emerged as a promising technique for enhancing the adaptability of vision-language models by optimizing textual prompts using unlabeled test data. However, prior studies have observed that TPT often produces poorly calibrated models, raising concerns about the reliability of their predictions. Recent works address this issue by incorporating additional regularization terms that constrain model outputs, which improve calibration but often degrade performance. In this work, we reveal that these regularization strategies implicitly encourage optimization toward flatter minima, and that the sharpness of the loss landscape around adapted prompts is a key factor governing calibration quality. Motivated by this observation, we introduce Flatness-aware Prompt Pretraining (FPP), a simple yet effective pretraining framework for TPT that initializes prompts within flatter regions of the loss landscape prior to adaptation. We show that simply replacing the initialization in existing TPT pipelines--without modifying any other components--is sufficient to improve both calibration and performance. Notably, FPP requires no labeled data and incurs no additional computational costs during test-time tuning, making it highly practical for real-world deployment. The code is available at: https://github.com/YonseiML/fpp.

Recent works on vision-language-action (VLA) models have made great progress in exploring action tokenizers that convert continuous control signals into discrete tokens to align with LLM/VLM training paradigms.These approaches typically train a single tokenizer over entire manipulation trajectories, which often comprise multiple distinct skills and thus pose a challenging optimization trade-off.To address this issue, we introduce MoEActok, a novel action tokenizer that employs a mixture-of-experts (MoE) quantizer to produce skill-aware discrete representations for VLA models. MoEActok utilizes a clustering-driven MoE VQ-VAE mechanism in which each expert specializes in a particular skill.The key components are: (a) an action-skill decoupling strategy that uses k-means clustering to group action chunks, aligning clusters having similar skills; (b) a skill-aware training paradigm that augments VLA models with skill-conditioned context, improving skill grounding; and (c) an adapter that projects shared encoder representations into skill-specific latent spaces for specialized quantization, and subsequently harmonize the heterogeneous quantized representations back into a unified space for coherent reconstruction by the shared decoder.We evaluate MoEActok-based VLA models against multiple prior action tokenizer baselines in the RoboTwin and Simpler-Env simulators, and further assess zero-shot transfer on three real-world tasks. Across both simulated and real-world settings, MoEActok-based VLA substantially outperforms existing discrete tokenization methods.

Metric depth estimation aims to recover depth maps with absolute scale, high resolution, and cross-scene consistency from visual observations. Existing approaches either rely on large-scale models or costly sensors to preserve metric accuracy and generalization, both ill-suited to resource-constrained deployment. In this paper, we propose LiteSense, a lightweight RGB-ToF fusion framework that leverages compact normalized histogram (CNH) signals together with RGB cues to achieve efficient and reliable metric depth estimation. Specifically, LiteSense leverages a U-Net-style encoder-decoder that forms an RGB-D input by concatenating RGB with upsampled ToF depth, providing explicit metric priors. To address resolution disparity and recover fine details, we introduce the **Patch-wise CNH Spatial Injection (PCSI)** module, which leverages zone-wise histogram measurements via cross-attention to guide high-level feature fusion. Extensively evaluated on NYUv2 and SUN RGB-D, LiteSense consistently outperforms monocular baselines and DELTAR with substantially lower computational cost, and demonstrates promising zero-shot generalization. We further introduce THDR3K, the first indoor RGB-ToF-CNH dataset, where LiteSense achieves real-world accuracy comparable to--and in challenging cases surpassing--Intel RealSense. All the relevant source codes and the collected dataset are available at GitHub.

Image aesthetic assessment (IAA) has extensive applications in content creation, album management, and recommendation systems, etc. In such applications, it is commonly needed to pick out the most aesthetically pleasing image from a series of images with subtle aesthetic variations, a topic we refer to as fine-grained IAA. Unfortunately, state-of-the-art IAA models are typically designed for coarse-grained evaluation, where images with notable aesthetic differences are evaluated independently on an absolute scale. These models are inherently limited in discriminating fine-grained aesthetic differences. To address the dilemma, we contribute FGAesthetics, a fine-grained IAA database with 32,217 images organized into 10,028 series, which are sourced from diverse categories including Natural, AIGC, and Cropping. Annotations are collected via pairwise comparisons within each series. We also devise Series Refinement and Rank Calibration to ensure the reliability of data and labels. Based on FGAesthetics, we further propose FGAesQ, a novel IAA framework that learns discriminative aesthetic scores from relative ranks through Difference-preserved Tokenization (DiffToken), Comparative Text-assisted Alignment (CTAlign), and Rank-aware Regression (RankReg). FGAesQ enables accurate aesthetic assessment in fine-grained scenarios while still maintains competitive performance in coarse-grained evaluation. Extensive experiments and comparisons demonstrate the superiority of the proposed method.

Semantic route planning involves generating itineraries that align with user intent while respecting real-world spatial constraints. However, text-only large language models (LLMs) often hallucinate geographically implausible routes due to poor spatial grounding. Inspired by how humans use maps for route planning, we propose the SMAP, which is the first multimodal framework combining user queries, POI metadata, and map tiles to produce spatially coherent, preference-aware routes. To enhance the spatial consistency, the SMAP features a two-stage anti-hallucination mechanism: (1) a map-grounded self-editing pipeline where a multimodal LLM (MLLM) drafts routes and a second MLLM verifies and refines them using geographic evidence; and (2) hallucination-penalized Direct Preference Optimization (HDPO) that steers the route generator toward spatially plausible routes by using verified routes as accepted responses and hallucinated drafts as rejected ones. Additionally, we introduce MM-Route, the first multimodal dataset for semantic route planning, with 3,000 diverse queries annotated with POI metadata and map tiles, covering a broad spectrum of geographic granularities and user intents. Experimental results demonstrate that SMAP significantly reduces geographical hallucinations and outperforms strong baselines in spatial plausibility and user alignment. Our project is at https://amap-mobility-intelligence.github.io/SMAP/.