CVPR 2026 论文列表

Reinforcement learning (RL) has recently achieved remarkable success in eliciting visual reasoning within Multimodal Large Language Models (MLLMs). However, existing approaches typically train separate models for different tasks and treat image and video reasoning as disjoint domains. This results in limited scalability toward a multimodal reasoning generalist, which restricts practical versatility and hinders potential knowledge sharing across tasks and modalities.To this end, we propose OneThinker, an all-in-one reasoning model that unifies image and video understanding across diverse fundamental visual tasks, including question answering, captioning, spatial and temporal grounding, tracking, and segmentation. To achieve this, we construct the OneThinker-600k training corpus covering all these tasks and employ commercial models for CoT annotation, resulting in OneThinker-SFT-340k for SFT cold start.Moreover, we propose EMA-GRPO to handle reward heterogeneity in multi-task RL by tracking task-wise moving averages of reward standard deviations for balanced optimization. Extensive experiments on diverse visual benchmarks show that OneThinker delivers strong performance on 31 benchmarks, across 10 fundamental visual understanding tasks.Moreover, it exhibits effective knowledge transfer between certain tasks and preliminary zero-shot generalization ability, marking a step toward a unified multimodal reasoning generalist. All code, model, and data are released.

Large multimodal models (LMMs) are increasingly adopted as judges in multimodal evaluation systems due to their strong instruction following and consistency with human preferences. However, their ability to follow diverse, fine-grained evaluation criteria remains underexplored. We develop Multi-Crit, a benchmark for evaluating multimodal judges on their capacity to follow pluralistic criteria and produce reliable criterion-level judgments. Covering both open-ended generation and verifiable reasoning tasks, Multi-Crit is built through a rigorous data curation pipeline that gathers challenging response pairs with multi-criterion human annotations. It further introduces three novel metrics for systematically assessing pluralistic adherence, criterion-switching flexibility, and the ability to recognize criterion-level preference conflicts. Comprehensive analysis of 25 LMMs reveals that 1) proprietary models still struggle to maintain consistent adherence to pluralistic criteria--especially in open-ended evaluation; 2) open-source models lag further behind in flexibly following diverse criteria; and 3) critic fine-tuning with holistic judgment signals enhances visual grounding but fails to generalize to pluralistic criterion-level judgment. Additional analyses on reasoning fine-tuning, test-time scaling, and boundary consistency between open-source and proprietary models further probe the limits of current multimodal judges. As a pioneering study, Multi-Crit lays the foundation for building reliable and steerable multimodal AI evaluation.

Multimodal Large Language Models (MLLMs) often generate hallucinations, where the output deviates from the visual content. Given that these hallucinations can take diverse forms, detecting hallucinations at a fine-grained level is essential for comprehensive evaluation and analysis. To this end, we propose a novel task of multimodal fine-grained hallucination detection and editing for MLLMs. Moreover, we propose ZINA, a novel method that identifies hallucinated spans at a fine-grained level, classifies their error types into six categories, and suggests appropriate refinements. To train and evaluate models for this task, we constructed VisionHall, a dataset comprising 6.9k outputs from twelve MLLMs manually annotated by 211 annotators, and 20k synthetic samples generated using a graph-based method that captures dependencies among error types. We demonstrated that ZINA outperformed existing methods, including GPT-4o and Llama-3.2, in both detection and editing tasks.

Image memorability, i.e., how likely an image is to be remembered, has traditionally been studied in computer vision either as a passive prediction task, with models regressing a scalar score, or with generative methods altering the visual input to boost the image likelihood of being remembered. Yet, none of these paradigms supports users at capture time, when the crucial question is how to improve a photo memorability. We introduce the task of **Mem**orability **Feed**back (**MemFeed**), where an automated model should provide actionable, human-interpretable guidance to users with the goal to enhance an image future recall. We also present **MemCoach**, the first approach designed to provide concrete suggestions in natural language for memorability improvement (e.g., "emphasize facial expression," "bring the subject forward"). Our method, based on Multimodal Large Language Models (MLLMs), is training-free and employs a teacher-student steering strategy, aligning the model internal activations toward more memorable patterns learned from a teacher model progressing along least-to-most memorable samples. To enable systematic evaluation on this novel task, we further introduce **MemBench**, a new benchmark featuring sequence-aligned photoshoots with annotated memorability scores. Our experiments, considering multiple MLLMs, demonstrate the effectiveness of MemCoach, showing consistently improved performance over several zero-shot models. The results indicate that memorability can not only be predicted but also taught and instructed, shifting the focus from mere prediction to actionable feedback for human creators. Dataset and code will be publicly released upon publication.

Vision-Language-Action (VLA) models are promising for embodied intelligence, yet they often overlook the predictive and temporal-causal structure underlying visual dynamics. World-model VLAs address this by predicting future frames, but waste capacity reconstructing redundant backgrounds. To overcome these limitations, we introduce CoWVLA, a new "Chain of World" paradigm that unifies world-model temporal reasoning with a disentangled latent motion representation. First, a pretrained video VAE serves as a latent motion extractor, explicitly factorizing video segments into structure and motion latents. Then, during pre-training, the VLA learns from an instruction and an initial frame to infer a continuous latent motion chain and predict the segment's terminal frame. Finally, during co-fine-tuning, this latent dynamic is aligned with discrete action prediction by jointly modeling sparse keyframes and action sequences in a unified autoregressive decoder. This design preserves the world-model benefits of temporal reasoning and world knowledge while retaining the compactness and interpretability of latent actions, enabling efficient visuomotor learning. Extensive experiments on robotic simulation benchmarks show that CoWVLA outperforms existing world-model and latent-action approaches and achieves moderate computational efficiency, highlighting its potential as a more effective VLA pretraining paradigm. Project website: https://fx-hit.github.io/cowvla-io.

Recent advancements in Unified Multimodal Models (UMMs) have enabled remarkable image understanding and generation capabilities. However, while models like Gemini-2.5-Flash-Image show emerging abilities to reason over multiple related images, existing benchmarks rarely address the challenges of multi-image context generation, focusing mainly on text-to-image or single-image editing tasks. In this work, we introduce MICON-Bench, a comprehensive benchmark covering six tasks that evaluate cross-image composition, contextual reasoning, and identity preservation. We further propose an MLLM-driven Evaluation-by-Checkpoint framework for automatic verification of semantic and visual consistency, where multimodal large language model (MLLM) serves as a verifier. Additionally, we present Dynamic Attention Rebalancing (DAR), a training-free, plug-and-play mechanism that dynamically adjusts attention during inference to enhance coherence and reduce hallucinations. Extensive experiments on various state-of-the-art open-source models demonstrate both the rigor of MICON-Bench in exposing multi-image reasoning challenges and the efficacy of DAR in improving generation quality and cross-image coherence.

Despite significant progress in text-to-video generation, current models still suffer from unrealistic dynamics, temporal inconsistency, and unstable semantic alignment. Existing preference alignment approaches rely on costly and often ambiguous human or VLM-based video preference annotation, which has become a major bottleneck for scaling data. To address this challenge, we propose an annotation-free preference alignment method that constructs accurate preference pairs through video continuation.We extend a pretrained video generation model into a continuation model and apply continuation with different amounts of reference frames while keeping the total video length fixed. As generated segments are inferior to ground-truth frames and and fixed-length continuations conditioned on more reference frames contain less generated content, they exhibit higher fidelity than those with fewer references, naturally inducing a preference order.We further introduce Asymmetrical DPO, which computes preference loss on all continuation regions except the shared prefix conditioning frames and normalizes it by their length, preventing spurious preference signals from leaking into the conditioned portion.Experiments across multiple benchmarks show that our method delivers significant improvements in dynamics realism, temporal coherence, and semantic alignment over existing DPO-based approaches, while fully eliminating the need for human preference labeling or auxiliary reward models.

Event-based Action Recognition (EAR) provides a promising pathway for understanding dynamic behaviors under challenging conditions. Recent progress in vision-language models has introduced a cross-modal learning paradigm into EAR, enabling models to associate event streams with textual semantics for enhancing conceptual understanding. However, existing methods typically overlook the intrinsic polarity-driven motion cues that are fundamental to event data, leading to suboptimal spatiotemporal representations. To address this limitation, we propose a POlarity Knowledge Enhanced framework (POKER), which explicitly incorporates event polarity-aware motion knowledge across visual and textual modalities.POKER consists of two synergistic components: Polarity Motion Capturer (PMC) and Polarity Motion Reasoner (PMR). Specifically, PMC decouples positive and negative polarities to capture polarity-sensitive motion cues, while PMR semantically analyzes polarity-induced motion dynamics via large language models. Through the polarity alignment, POKER couples semantic reasoning with visual dynamics, achieving more discriminative representations. Extensive experiments on multiple benchmarks demonstrate that POKER enhances performance across diverse event representations.

Deploying embodied agents that can answer questions about their surroundings in realistic real-world settings remains difficult, partly due to the scarcity of benchmarks for episodic memory Embodied Question Answering (EQA). Inspired by the challenges of infrastructure inspections, we propose Inspection EQA as a compelling problem class for advancing episodic memory EQA, as it demands multi-scale reasoning and long-range spatial understanding, while offering standardized evaluation, professional inspection reports as grounding, and egocentric imagery. We introduce BridgeEQA, a benchmark of 2,200 open-vocabulary question-answer pairs (in the style of OpenEQA) grounded in professional inspection reports across 200 real-world bridge scenes with 47.93 images on average per scene. We further propose a new EQA metric Image Citation Relevance to evaluate the ability of a model to cite relevant images. Evaluations of state-of-the-art vision-language models reveal substantial performance gaps. To address this, we propose Embodied Memory Visual Reasoning (EMVR), which formulates the inspection EQA task as a Markov decision process. EMVR shows strong performance over the baselines. Code and dataset available at: https://drags99.github.io/bridge-eqa/

Multimodal learning often grapples with the challenge of low-quality data, which predominantly manifests as two facets: modality imbalance and noisy corruption. While these issues are often studied in isolation, we argue that they share a common root in the predictive uncertainty towards the reliability of individual modalities and instances during learning. In this paper, we propose a unified framework, termed Conformal Predictive Self-Calibration (CPSC), which leverages conformal prediction to equip the model with the ability to perform self-guided calibration on-the-fly. The core of our proposed CPSC lies in a novel self-calibrating training loop that seamlessly integrates two key modules: (1) Representation Self-Calibration, which decomposes unimodal features into components, selectively fuses the most robust ones identified by a conformal predictor to enhance feature resilience. (2) Gradient Self-Calibration, which recalibrates the gradient flow during backpropagation based on instance-wise reliability scores, steering the optimization towards more trustworthy directions. Furthermore, we also devise a self-update strategy for the conformal predictor to ensure the entire system co-evolves consistently throughout the training process. Extensive experiments on six benchmark datasets under both imbalanced and noisy settings demonstrate that our CPSC framework consistently outperforms existing state-of-the-art methods. Our code is available at https://github.com/XunCHN/CPSC.

Fine-grained visual understanding is shifting from static classification to knowledge-augmented reasoning, where models must justify as well as recognise. Existing approaches remain limited by closed-set taxonomies and single-label prediction, leading to significant degradation under open-set or context-dependent conditions. We present the Knowledge-Augmented Fine-Grained Reasoning Agent (KFRA), a unified framework that transforms fine-grained perception into evidence-driven reasoning. KFRA operates through a three-stage closed reasoning loop that emulates expert analysis. It first performs open-vocabulary detection and web-scale retrieval to generate category hypotheses. It then conducts discriminative regions localisation by aligning textual knowledge with visual evidence through a global-to-local focusing mechanism. Finally, it integrates all multimodal evidence within a large multimodal model to perform interpretable reasoning. Unlike existing agents that treat retrieval and reasoning as independent processes, KFRA establishes a retrieval-grounding coupling that converts retrieved knowledge into spatially grounded evidence for verification. This design enables factual, interpretable, and task-agnostic reasoning across diverse fine-grained scenarios. To evaluate this capability, we construct FGExpertBench, a benchmark designed to assess reasoning depth and cross-task generalisation across six knowledge dimensions. Extensive experiments demonstrate that KFRA consistently surpasses both standalone large multimodal models and current agent frameworks, achieving up to 19 percent improvement in reasoning accuracy and delivering evidence-grounded interpretability in open-set fine-grained visual understanding.

High-quality 3D avatar modeling faces a critical trade-off between fidelity and generalization. On the one hand, multi-view studio data enables high-fidelity modeling of humans with precise control over expressions and poses, but it struggles to generalize to real-world data due to limited scale and the domain gap between the studio environment and the real world. On the other hand, recent large-scale avatar models trained on millions of in-the-wild samples show promise for generalization across a wide range of identities, yet the resulting avatars are often of low-quality due to inherent 3D ambiguities. To address this, we present Large-Scale Codec Avatars (LCA), a high-fidelity, full-body 3D avatar model that generalizes to world-scale populations in a feedforward manner, enabling efficient inference. Inspired by the success of large language models and vision foundation models, we present, for the first time, a pre/post-training paradigm for 3D avatar modeling at scale: we pretrain on 1M in-the-wild videos to learn broad priors over appearance and geometry, then post-train on high-quality curated data to enhance expressivity and fidelity. LCA generalizes across hair styles, clothing, and demographics while providing precise, fine-grained facial expressions and finger-level articulation control, with strong identity preservation. Notably, we observe emergent generalization to relightability and loose garment support to unconstrained inputs, and zero-shot robustness to stylized imagery, despite the absence of direct supervision.

Generative models can now produce photorealistic imagery, yet they still struggle with the long, multi-goal prompts that professional designers issue. To expose this gap and better evaluate models' performance in real-world, we introduce Long Goal Bench(LGBench), a 2000-task suite (1000 T2I, 1000 I2I) whose average instruction contains 18---22 tightly coupled goals spanning global layout, local object placement, typography, and logo fidelity. We find even state-of-the-art commercial APIs satisfy fewer than 72% of the goals and routinely miss localized edits, confirming the brittleness of current pipelines. To address this, we present VisionDirector, a training-free, vision-language supervisor that (i) extracts structured goals from long instructions, (ii) dynamically decides between one-shot generation and staged edits, (iii) runs micro-grid sampling plus semantic verification/rollback after every edit, and (iv) logs goal-level rewards. We further fine-tune the planner with Group Relative Policy Optimization, yielding shorter edit trajectories (3.1 vs. 4.2 steps) and stronger alignment. VisionDirector achieves new state of the art on GenEval (+7% overall), and ImgEdit (+0.07 absolute) while producing consistent qualitative improvements on typography, multi-object scenes, and pose editing. The code, benchmark, and evaluation scripts will be released.

Model merging combines independently fine-tuned checkpoints without joint multi-task training. In the era of foundation-model, fine-tuning with Low-Rank Adaptation (LoRA) is prevalent, making LoRA merging a promising target. Existing approaches can work in homogeneous settings where all target tasks are classification but often fail when tasks span classification and regression. Approaches using entropy-based surrogates do not apply to regression and are costly for large language models due to long token sequences. We introduce Null-Space Compression (NSC) Merging, a label-free, output-agnostic method that sets merge weights from adapter geometry. Our key observation is that during LoRA finetuning the down-projection factor (A) in (\Delta W = BA) compresses its null space, and the compression correlates with performance. NSC uses this as an optimization signal for merging that can generalize across classification, regression, and sequence generation. NSC achieves state-of-the-art performance across twenty heterogeneous vision tasks with balanced gains where prior methods overfit subsets of tasks. It also outperforms baselines on six NLI benchmarks and on vision-language evaluations for VQA and image captioning, demonstrating scalability and effectiveness. Our code is available at https://github.com/wonyoung01/nsc_merging.

Vision-Language Models (VLMs) have achieved impressive performance on spatial reasoning benchmarks, yet these evaluations mask critical weaknesses in understanding object interactions. Current benchmarks test high-level relationships ("left of," "behind", etc.) but ignore fine-grained spatial understanding needed for real-world applications: precise 3D localization, physical compatibility between objects, object affordances and multi-step spatial planning. In this work, we present BOP-ASK, a novel large-scale dataset for object-interaction reasoning for both training and benchmarking. Our data generation pipeline leverages 6D object poses from the Benchmark for Object Pose Estimation (BOP) datasets from which we derive fine-grained annotations such as grasp poses, referred object poses, path planning trajectories, relative spatial and depth relationships, and object-to-object relationships. BOP-ASK comprises over 150k images and 33M question-answer pairs spanning six tasks (four novel), providing a rich resource for training and evaluating VLMs. We evaluate proprietary and open-sourced VLMs, and conduct human evaluations on BOP-ASK-core, a contributed test benchmark. We also release BOP-ASK-lab, an out-of-distribution benchmark with images not sourced from BOP, enabling testing of generalization. Our experiments demonstrate that models trained on BOP-ASK outperform baselines and exhibit emergent capabilities such as precise object and grasp pose estimation, trajectory planning, and fine-grained object-centric spatial reasoning in cluttered environments. Project website: https://bop-ask.github.io/

Estimating camera pose in dynamic environments is a critical challenge, as most visual SLAM and SfM methods assume inputs from static environments. While recent dynamic-aware methods exist, they are often not unified: semantic-based approaches are brittle, per-sequence optimization methods fail on short sequences, and other learned models sometimes perform badly on static-only scenes. We present Wildpose, a unified monocular pose estimation framework that is robust in dynamic environments while maintaining state-of-the-art performance on static and low-ego-motion datasets. Our key insight is to connect the two powerful paradigms in modern 3D vision: the rich perceptual frontend of feed-forward models and the end-to-end optimization of differentiable bundle adjustment (BA). We achieve this by enhancing the differentiable BA pipeline in two ways. First, we introduce a new 3D-aware update operator by integrating a frozen, pre-trained MASt3R feature backbone and training the operator's subsequent layers on a diverse curriculum of static and dynamic data. Second, we propose a high-capacity motion mask detector that leverages rich, multi-level 3D-aware features from the same frozen backbone. Extensive experiments show Wildpose consistently outperforms prior methods across a wide variety of benchmarks, including dynamic (Wild-SLAM, Bonn), static (TUM, 7-Scenes), and low-ego-motion (Sintel) datasets.

Autonomous Vehicles (AVs) collect and pseudo-label terabytes of multi-modal data localized to HD maps during normal fleet testing. However, identifying interesting and safety-critical scenarios from uncurated driving logs remains a significant challenge. Traditional scenario mining techniques are error-prone and prohibitively time-consuming, often relying on hand-crafted structured queries. In this work, we revisit spatio-temporal scenario mining through the lens of recent vision-language models (VLMs) to detect whether a described scenario occurs in a driving log and, if so, precisely localize it in both time and space. To address this problem, we introduce RefAV, a large-scale dataset of 10,000 diverse natural language queries that describe complex multi-agent interactions relevant to motion planning derived from 1000 driving logs in the Argoverse 2 Sensor dataset. We evaluate several referential multi-object trackers and present an empirical analysis of our baselines. Notably, we find that naively repurposing off-the-shelf VLMs yields poor performance, suggesting that scenario mining presents unique challenges. Lastly, we discuss our recently held competition and share insights from the community. Our code and dataset are available on GitHub and Argoverse.

Object counting in remote sensing imagery becomes challenging when visual cues are obscured by clouds, fog, shadows, or low-light conditions. Yet earth observation inherently provides complementary geo-modalities, including land use and map, which offer stable structural and contextual priors that remain available when appearance cues fail. In this paper, we introduce GROC, the first large-scale dataset Geo-guided Reasoning in Object Counting under adverse earth observation conditions. GROC contains 1.2 million point annotations over 14K images, each aligned with 3 modalities that preserve original geospatial information. We also provide a data engine to collect a large-scale object counting dataset with multiple geo-modalities, realistic degradations, and reliable annotations. We further present an counting agent that adaptively leverages geo-modalities to produce reliable estimates. Extensive experiments show that existing models struggle to "see" through adverse conditions, whereas geo-modalities improve robustness. GROC establishes the first benchmark that explicitly challenges models to see what they cannot see, charting a new direction for geo-guided amodal reasoning in earth observation.

Large vision-language models (LVLMs) have achieved impressive performance across a variety of multimodal tasks, yet remain vulnerable to targeted adversarial attacks, particularly in black-box settings. In this paper, we propose VCP-Attack, a transferable targeted attack framework that combines structured contrastive supervision with subspace-guided perturbation optimization. Specifically, we employ a dynamic PCA-based projection to constrain perturbations within semantically meaningful low-dimensional subspaces, and design a multi-sample contrastive loss to align adversarial features with target semantics while pushing them away from the source semantics. Extensive experiments on seven open-source and three proprietary LVLMs--including GPT-4o, Claude, and Gemini--show that VCP-Attack achieves state-of-the-art performance in black-box targeted attacks. Under a fixed perturbation budget (epsilon = 16/255), our method achieves an average attack success rate (ASR) of 94.2% on open-source models and 83.1% on proprietary models, surpassing the strongest baselines by 23.3% and 16.8%, respectively. Notably, VCP-Attack achieves a 95.6% ASR on GPT-4o. Comprehensive ablation studies and visualizations further validate the effectiveness of the dynamic subspace projection and semantic contrastive supervision. While evaluated on image captioning, our approach is model-agnostic and exhibits strong potential for broader applications to black-box adversarial settings in vision-language tasks.