CVPR 2026 Papers

Recent advances in vision-language foundation models have enabled text-driven evaluation of image aesthetics and visual quality. However, existing models are typically optimized for fixed prompts or specific datasets, limiting their adaptability to diverse evaluation criteria. This paper presents Probabilistic Prompt Adaptation (PPA), a unified probabilistic framework that flexibly predicts aesthetic and quality scores conditioned on arbitrary text prompts. PPA formulates score prediction as a mixture over prompts, dynamically estimating prompt suitability based on both image content and task context. By marginalizing over prompts pre-sampled from a large language model (LLM), it enables annotation-free training using only triplets of task, image, and score. Experiments across multiple IAA and IQA benchmarks demonstrate that PPA achieves consistent and perceptually aligned prompt-based scoring, allowing fine-grained control over evaluation semantics.

Object hallucination in Large Vision-Language Models (LVLMs) significantly hinders their reliable deployment. Existing methods struggle to balance efficiency and accuracy: they often require expensive reference models and multiple forward passes, or apply static edits that risk suppressing genuine visual evidence. To address this, we introduce HulluEdit, a single-pass, reference-free intervention framework. Our core innovation is orthogonal subspace editing: we decompose the hidden states of the model into orthogonal subspaces--visual evidence, conflicting priors, and residual uncertainty--enabling selective suppression of hallucinatory patterns without interfering with visual grounding. This approach mathematically guarantees that edits applied to the prior subspace leave the visual component entirely unaffected. Extensive experiments show that HulluEdit achieves state-of-the-art hallucination reduction on benchmarks including POPE and CHAIR across diverse architectures, while preserving general capabilities on MME and maintaining efficient inference. Our method consistently outperforms contrastive decoding and static subspace editing baselines, offering a new pathway toward more trustworthy LVLMs. Code is released at https://github.com/VioAgnes/HulluEdit.

Current visual grounding models are either based on a Multimodal Large Language Model (MLLM) that performs auto-regressive decoding, which is slow and risks hallucinations, or on re-aligning an LLM with vision features to learn new special or object tokens for grounding, which may undermine the LLM's pretrained reasoning ability. In contrast, we propose **VGent**, a modular encoder-decoder architecture that explicitly disentangles high-level reasoning and low-level bounding box prediction. Specifically, a frozen MLLM serves as the encoder to provide untouched powerful reasoning capabilities, while a decoder takes high-quality boxes proposed by detectors as queries and selects target box(es) via cross-attending on encoder's hidden states. This design fully leverages advances in both object detection and MLLM, avoids the pitfalls of auto-regressive decoding, and enables fast inference. Moreover, it supports modular upgrades of both the encoder and decoder to benefit the whole system: we introduce (i) **QuadThinker**, an RL-based training paradigm for enhancing multi-target reasoning ability of the encoder; (ii) **mask-aware label** for resolving detection-segmentation ambiguity; and (iii) **global target recognition** to improve the recognition of all the targets which benefits the selection among augmented proposals. Experiments on multi-target visual grounding benchmarks show that VGent achieves a new state-of-the-art with **+20.6%** F1 improvement over prior methods, and further boosts gIoU by **+8.2%** and cIoU by **+5.8%** under visual reference challenges, while maintaining constant, fast inference latency.

Textual adapter-based tuning methods have shown significant potential in transferring knowledge from pre-trained Vision-Language Models (VLMs) to downstream tasks. Existing works generally employ the deterministic textual feature adapter to refine each category textual representation. However, due to inherent factors such as different attributes and contexts, there exists significant diversity in textual descriptions for each category. Such description diversity offers rich discriminative semantic knowledge that can benefit downstream visual learning tasks. Obviously, the traditional deterministic adapter model cannot adequately capture this varied semantic information. Also, it is desirable to exploit the inter-class relationships in VLM adapter. To address these issues, we propose to exploit random graph model into VLM adapter and develop a novel Vertex Random Graph Adapter (VRGAdapter). VRGAdapter first models the inherent diverse descriptions of each category and inter-class relationships of different categories simultaneously by leveraging a Vertex Random Knowledge Graph (VRKG) model. Then, it employs probabilistic message propagation on VRKG to learn context-aware distribution representation for each class node. Finally, it adopts a reparameterized sampling function to achieve textual adapter learning. Note that, VRGAdapter provides a more general adapter solution that encompasses traditional graph-based adapter as a special case. In addition, to enable more robust performance for downstream tasks, we also introduce a new Uncertainty-guided Multi-branch Fusion (UMF) scheme that dynamically integrates multiple pre-trained models for ensemble prediction. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of our approach.

We introduce VoxTell, a vision-language model for text-prompted volumetric medical image segmentation. It maps free-form descriptions, from single words to full clinical sentences, to 3D masks. Trained on 62K+ CT, MRI, and PET volumes spanning 1K+ anatomical and pathological classes, VoxTell uses multi-stage vision-language fusion across decoder layers to align textual and visual features at multiple scales. It achieves state-of-the-art zero-shot performance across modalities on unseen datasets, excelling on familiar concepts while generalizing to related unseen classes. Extensive experiments further demonstrate strong cross-modality transfer, robustness to linguistic variations and clinical language, as well as accurate instance-specific segmentation from real-world text. Code is available at: https://github.com/MIC-DKFZ/VoxTell

Event-based video reconstruction seeks to recover high-speed, high-dynamic-range videos from event streams. While existing approaches rely exclusively on motion-triggered events, these events are inherently sparse and primarily capture dynamic regions. Therefore, they often suffer from error accumulation and degraded quality in regions with few events. In this work, we introduce aperture-modulation-triggered events as a complementary mechanism to enrich the captured scene information. Specifically, we periodically modulate the aperture to actively generate dense event signals, thereby encoding intensity cues even in static or low-motion regions. Building upon this idea, we design an AE2VID framework that jointly leverages aperture-modulation-triggered and motion-triggered events to enhance the fidelity of predictions. The proposed framework consists of two subnetworks for the dedicated processing of both event types. We further collect a real dataset and validate the effectiveness of our method. Extensive experiments show our superiority over state-of-the-art methods. Code and data will be available at https://github.com/a1henu/AE2VID/.

Recent advances in snapshot multispectral (MS) imaging have enabled compact, low-cost spectral sensors for consumer and mobile devices. By capturing richer spectral information than conventional RGB sensors, these systems can enhance key imaging tasks, including color correction. However, most existing methods treat the color correction pipeline in separate stages, often discarding MS data early in the process. We propose a unified, learning-based framework that performs end-to-end color correction and jointly leverages data from a high-resolution RGB sensor and an auxiliary low-resolution MS sensor. Our approach integrates the full pipeline within a single model, producing coherent and color-accurate outputs. We demonstrate the flexibility and generality of our framework by refactoring two different state-of-the-art image-to-image architectures. To support training and evaluation, we construct a dedicated dataset by aggregating and repurposing publicly available spectral datasets, rendering under multiple RGB camera sensitivities. Extensive experiments show that our approach improves color accuracy and stability, reducing error by up to 50% compared to RGB-only and MS-driven baselines. Code, models and dataset available at: https://lucacogo.github.io/Mobile-Spectral-CC/.

We study how different Chain-of-Thought (CoT) designs affect the acquisition of the generalizable visual reasoning ability in vision-language models (VLMs). While CoT data, especially long or visual CoT such as "think with image", has been widely used to supervise intermediate reasoning, it remains unclear why specific CoT designs help and which ones truly support generalizable reasoning. To systematically evaluate this, we focus on a controlled maze-solving task where reasoning rules are fully visual, difficulty can be tuned by grid size, and all the intermediate steps can be automatically generated. Using Qwen2.5-VL-7B under a standard SFT-then-RL pipeline, we compare three representative CoT formats: Language CoT, Grounding CoT (with spatial coordinate trajectories), and Visual CoT (with image manipulations). Our experiments reveal that visual and longer CoT mainly accelerate convergence but do not lift the final performance ceiling; concise CoT containing only essential grounding steps outperforms longer traces; and, strikingly, CoT retaining only the minimal grounding results generalizes best across different maze sizes. We further validate these insights on other vision-centric tasks. These findings highlight a "short is long" effect and provide practical guidance for constructing more generalizable SFT datasets for visual reasoning. Our code and data will be publicly released at https://github.com/RUCAIBox/Revisiting-Visual-CoT.

Diffusion- and flow-based models usually allocate compute uniformly across space, updating all patches with the same timestep and number of function evaluations. While convenient, this ignores the heterogeneity of natural images: some regions are easy to denoise, whereas others benefit from more refinement or additional context. Motivated by this, we explore patch-level noise scales for image synthesis. We find that naively varying timesteps across image tokens performs poorly, as it exposes the model to overly informative training states that do not occur at inference. We therefore introduce a timestep sampler that explicitly controls the maximum patch-level information available during training, and show that moving from global to patch-level timesteps already improves image generation over standard baselines. By further augmenting the model with a lightweight per-patch difficulty head, we enable adaptive samplers that allocate compute dynamically where it is most needed. Combined with noise levels varying over both space and diffusion time, this yields Patch Forcing (PF), a framework that advances easier regions earlier so they can provide context for harder ones. PF achieves superior results on class-conditional ImageNet, remains orthogonal to representation alignment and guidance methods, and scales to text-to-image synthesis. Our results suggest that patch-level denoising schedules provide a promising foundation for adaptive image generation.

Vision-Language Models (VLMs) like CLIP exhibit extraordinary out-of-distribution (OOD) generalization, while the theoretical foundations underlying this robustness remain largely unexplored. This work establishes a connection between CLIP and Invariant Risk Minimization (IRM), the principled paradigm to overcome OOD problems, through token-level causal representation learning. Our key insight is that CLIP's contrastive objective, when optimally trained, recovers modality-invariant causal factors at the word-and-phrase granularity. By decomposing text prompts into class-specific tokens (causal factors) and class-agnostic context tokens (environmental factors), we prove that a vocabulary-constrained InfoNCE objective becomes formally equivalent to IRM's invariance criterion. Grounded in this equivalence, we propose a mid-training paradigm aiming to inject invariant learning signals into pre-trained CLIP without architectural modification, yielding CLIP-IRM with superior OOD performance. We further extend this causal alignment to multimodal reasoning via using CLIP-IRM's invariant alignment scores as process-level rewards in reinforcement learning, effectively transplanting IRM's guarantees to robust sequential decision-making in Multimodal Large Language Models. Extensive experiments validate our theoretical framework and present substantial improvements in both multimodal OOD understanding and reasoning tasks.

Current text-to-image models struggle to provide precise camera control using natural language alone. In this work, we present a framework for precise camera control with global scene understanding in text-to-image generation by learning parametric camera tokens. We fine-tune image generation models for viewpoint-conditioned text-to-image generation on a curated dataset that combines 3D-rendered images for geometric supervision and photorealistic augmentations for appearance and background diversity. Qualitative and quantitative experiments demonstrate that our method achieves state-of-the-art accuracy while preserving image quality and prompt fidelity. Unlike prior methods that overfit to object-specific appearance correlations, our viewpoint tokens learn factorized geometric representations that transfer to unseen object categories. Our work shows that text-vision latent spaces can be endowed with explicit 3D camera structure, offering a pathway toward geometrically-aware prompts for text-to-image generation. Project page: https://randdl.github.io/viewtoken_control/

We introduce Complet4R, a novel end-to-end framework for Geometric Complete 4D Reconstruction, which aims to recover temporally coherent and geometrically complete reconstruction for dynamic scenes. Our method formalizes the task of Geometric Complete 4D Reconstruction as a unified framework of reconstruction and completion, by directly accumulating full contexts onto each frame. Unlike previous approaches that rely on pairwise reconstruction or local motion estimation, Complet4R utilizes a decoder-only transformer to operate all context globally directly from sequential video input, reconstructing a complete geometry for every single timestamp, including occluded regions visible in other frames. Our method demonstrates the state-of-the-art performance on our proposed benchmark for Geometric Complete 4D Reconstruction and the 3D Point Tracking task. Code will be released to support future research.

Common image editing tasks typically adopt powerful generative diffusion models as the leading paradigm for real-world content editing. Meanwhile, although reinforcement learning (RL) methods such as Diffusion-DPO and Flow-GRPO have further improved generation quality, efficiently applying Reinforcement Learning from Human Feedback (RLHF) to diffusion-based editing remains largely unexplored, due to a lack of scalable human-preference datasets and frameworks tailored to diverse editing needs. To fill this gap, we propose HP-Edit, a post-training framework for Human Preference-aligned Editing, and introduce RealPref-50K, a real-world dataset across eight common tasks and balancing common object editing. Specifically, HP-Edit leverages a small amount of human-preference scoring data and a pretrained visual large language model (VLM) to develop HP-Scorer--an automatic, human preference-aligned evaluator. We then use HP-Scorer both to efficiently build a scalable preference dataset and to serve as the reward function for post-training the editing model. We also introduce RealPref-Bench, a benchmark for evaluating real-world editing performance. Extensive experiments demonstrate that our approach significantly enhances models such as Qwen-Image-Edit-2509, aligning their outputs more closely with human preference.

Ultrasound imaging is widely used in clinical diagnostics due to its real-time capability and radiation-free nature. However, existing vision-language pre-training models, such as CLIP, are primarily designed for other modalities, and are difficult to directly apply to ultrasound data, which exhibit heterogeneous anatomical structures and diverse diagnostic attributes. To bridge this gap, we construct US-365K, a large-scale ultrasound image-text dataset containing 365k paired samples across 52 anatomical categories. We establish Ultrasonographic Diagnostic Taxonomy (UDT) containing two hierarchical knowledge frameworks. Ultrasonographic Hierarchical Anatomical Taxonomy standardizes anatomical organization, and Ultrasonographic Diagnostic Attribute Framework formalizes nine diagnostic dimensions, including body system, organ, diagnosis, shape, margins, echogenicity, internal characteristics, posterior acoustic phenomena, and vascularity. Building upon these foundations, we propose Ultrasound-CLIP, a semantic-aware contrastive learning framework that introduces semantic soft labels and semantic loss to refine sample discrimination. Moreover, we construct a heterogeneous graph modality derived from UDAF's textual representations, enabling structured reasoning over lesion-attribute relations. Extensive experiments with patient-level data splitting demonstrate that our approach achieves state-of-the-art performance on classification and retrieval benchmarks, while also delivering strong generalization to zero-shot, linear probing, and fine-tuning tasks.

What data should a vision-language model be trained on? To answer this question, many data curation efforts center on the quality of a dataset. However, most of these existing methods are (i) offline, i.e. they produce a static dataset from a set of predetermined filtering criteria, and (ii) concept-agnostic, i.e. they use model-based filters which induce additional dataset bias. In this work, we go beyond such offline, concept-agnostic methods and advocate for more flexible, task-adaptive online concept-based curation. Our first contribution is DataConcept, a collection of 128M web-crawled image-text pairs annotated with fine-grained details about their concept composition. Building on DataConcept, we introduce CABS (Concept-Aware Batch Sampling), a simple yet effective batch-sampling framework that flexibly constructs batches on-the-fly based on specific target distributions. We propose two variants: (i) Diversity Maximization, to curate batches with the broadest coverage of available concepts, and (ii) CABS-FM (Frequency Maximization), to curate batches with maximal object multiplicity. Through extensive evaluations with four visual backbones and a suite of 28 benchmarks, we demonstrate that CABS significantly benefits Language-Image Pretraining (LIP) and yields highly performant models on long-tailed evaluations. Overall, CABS represents a strong open-source alternative to proprietary online curation algorithms, enabling practitioners to define custom concept distributions that optimize for specific downstream tasks. Both DataConcept and the source code for CABS will be made public.

**Document QA** requires not only accurate answers but also identifying where each answer is grounded on the page. Most models treat the task as text-only generation, while existing answer grounding methods generate coarse bounding boxes that fail to capture curved text. We introduce **M3Grounder, a hybrid vision-language and segmentation architecture that formulates document grounding as pixel-level segmentation. It produces fine-grained evidence masks** refined by a bleed-suppression loss to prevent spillover. M3Grounder autoregressively generates answer text interleaved with [GROUND] tokens that link individual answer spans to their corresponding evidence regions. Also, **M3Grounder grounds evidence hierarchically across phrase, line, and block levels** using an enclosure loss that enforces spatial containment. We release **GroundingDocQA dataset (200K documents, 2M multi-span and multi-granular QA pairs with pixel-level grounding masks)**, built through a data engine that handles complex layouts, curved-text, and graphics-rich documents. We also release **GroundingDocQA-Bench, a diverse and challenging human-verified benchmark**. M3Grounder sets **a new state of the art in grounded DocVQA, advancing from coarse boxes to hierarchical, fine-grained and contextually grounded evidence.**

Human action recognition (HAR) in low-light environments remains challenging due to degraded visibility, illumination variance, and loss of appearance cues. We introduce DarkAct, a large-scale and high-quality RGB-thermal video dataset purpose-built for multimodal action recognition under low illumination. DarkAct contains 12,778 paired RGB-thermal videos covering 27 human actions across diverse viewpoints and scenes, offering a novel and comprehensive benchmark for understanding human actions in dark environments. We conduct extensive experiments on DarkAct, systematically benchmarking unimodal HAR models, multimodal fusion frameworks, and vision-language foundation models. Their limited performance on DarkAct underscores the urgent need for more robust perception systems under adverse illumination. To address this, we propose DarkAct-Net, an RGB-thermal fusion framework that enhances human-centric representation and achieves adaptive cross-modal fusion, enabling robust and fine-grained action recognition across diverse lighting conditions. Dataset and code are available at https://github.com/darkact-creator/DarkAct.git.

Recent advancements in reinforcement learning with verifiable rewards (RLVR) have significantly improved the complex reasoning ability of vision-language models (VLMs). However, its outcome-level supervision is too coarse to diagnose and correct errors within the reasoning chain. To this end, we propose Perceval, a process reward model (PRM) that enables token-level error grounding, which can extract image-related claims from the response and compare them one by one with the visual evidence in the image, ultimately returning claims that contain perceptual errors. Perceval is trained with perception-intensive supervised training data. We then integrate Perceval into the RL training process to train the policy models. Specifically, compared to traditional GRPO, which applies sequence-level advantages, we apply token-level advantages by targeting penalties on hallucinated spans identified by Perceval, thus enabling fine-grained supervision signals. In addition to augmenting the training process, Perceval can also assist VLMs during the inference stage. Using Perceval, we can truncate the erroneous portions of the model's response, and then either have the model regenerate the response directly or induce the model to reflect on its previous output. This process can be repeated multiple times to achieve test-time scaling. Experiments show significant improvements on benchmarks from various domains across multiple reasoning VLMs trained with RL. For test-time scaling, it also demonstrates consistent performance gains over other strategies, such as major voting. Our code and data will be publicly released.

Vision-Language Models (VLMs) have achieved remarkable progress in multimodal understanding, yet their positional encoding mechanisms remain suboptimal. Existing approaches uniformly assign positional indices to all tokens, overlooking variations in information density within and across modalities, which leads to inefficient attention allocation where redundant visual regions dominate while informative content is underrepresented. We identify positional granularity as an implicit resource and propose MODIX (Multimodal Information-Driven Positional IndeX Scaling), a training-free framework that dynamically adapts positional strides based on modality-specific contributions. MODIX jointly models intra-modal density via covariance-based entropy and inter-modal interaction via cross-modal alignment to derive unified scores, which rescale positional indices to allocate finer granularity to informative modalities while compressing redundant ones, without requiring any modification to model parameters or architecture. Experiments across diverse architectures and benchmarks demonstrate that MODIX consistently improves multimodal reasoning and adaptively reallocates attention according to task-dependent information distributions, suggesting that positional encoding should be treated as an adaptive resource in Transformers for multimodal sequence modeling.