CVPR 2026 Papers

Recent deep unfolding networks (DUNs) have advanced Compressive Sensing (CS) by effectively integrating iterative optimization with deep learning architectures. However, most CS approaches predominantly confine their inference to a single solution space, neglecting the inherent ill-posedness of CS problems that intrinsically permits multiple plausible candidate hypotheses. In this paper, a novel Multi-Hypothesis Collaborative Deep Unfolding CS Network (MHC-DUN) is proposed, which explicitly models and leverages multiple hypotheses by jointly optimizing across diverse solution spaces. Specifically, following the Proximal Gradient Descent algorithm, MHC-DUN jointly performs gradient descent and proximal mapping within this multi-hypothesis paradigm. i) For gradient descent, a well-designed AlphaNet is introduced to dynamically predict spatially varying step sizes for all hypotheses, enabling collaborative gradient updates across multiple solutions. ii) For proximal operator, a sophisticated multi-hypothesis collaborative proximal mapping module is designed, which leverages both intra-hypothesis and inter-hypothesis correlation priors to jointly refine multiple solutions. To enable end-to-end training, a novel composite loss function is designed, which balances measurement fidelity, hypothesis diversity, and reconstruction accuracy, encouraging exploration of complementary solutions while maintaining reconstruction fidelity. Experimental results reveal that the proposed CS method outperforms existing CS networks.

Unified Multimodal Models (UMMs) are redefining the landscape of artificial intelligence by coupling perception and generation across language, vision, and structured reasoning. Yet, despite their growing sophistication, a critical gap persists in evaluation: existing benchmarks largely measure discriminative understanding or unconstrained generation in isolation, overlooking the integrated generative reasoning required for genuine multimodal intelligence. To address this, we introduce GGBench, the benchmark explicitly designed to evaluate geometric generative reasoning--the ability of a model to understand, reason about, and construct a solution within a unified framework. Each instance in GGBench contains precisely aligned natural-language instructions, executable GeoGebra code, and rendered diagrams, enabling deterministic and interpretable verification of a model's reasoning and constructive fidelity. The benchmark comprises 1,411 rigorously curated problems covering eight categories and multiple difficulty levels, resulting in over 7,000 aligned visualizations. We propose a comprehensive tri-modal evaluation protocol that jointly assesses textual planning quality, code executability, and geometric accuracy of generated diagrams through both automated and human-in-the-loop judging. Extensive experiments on both state-of-the-art UMMs and general Large Language Models (LLMs) reveal a large performance gap between end-to-end generation and reasoning-grounded construction. GGBench establishes a new standard for testing multimodal systems that must not only understand but also build, marking a crucial step toward grounded, verifiable generative intelligence.

Recent advances in Remote Sensing Foundation Models (RSFMs) have demonstrated considerable potential for Earth Observation (EO) tasks. While adopting natural image foundation models (e.g., DINO) provides a data-efficient strategy for building RSFMs, their strong generalization capability does not fully transfer to complex remote sensing (RS) scenarios due to severe background interference, notably in perceiving challenging targets like low-contrast objects. To this end, we propose ORSATR-X, a novel RSFM that effectively integrates the generalizable representations of DINOv3 with a dedicated mechanism for exciting local contrast information. ORSATR-X comprises two core components: (1) a DINOv3 encoder, which provides rich feature representation under limited RS pre-training data, and (2) a carefully designed side network incorporating a Weber Local Adapter (WLA) and a Multi-scale Aggregation Module (MSAM). The WLA enhances discriminability of low-contrast boundaries in complex scenes through center-surround contrast and directional gradient information enhancement, while the MSAM handles inherent object scale variations in RS imagery by adaptive aggregation of features across multiple scales. Furthermore, we pre-train the side network using an efficient self-supervised distillation strategy. Extensive experiments on scene classification, object detection, and semantic segmentation demonstrate that ORSATR-X achieves state-of-the-art performance among existing RSFMs, demonstrating the effectiveness of our design.

Recent studies have demonstrated that Reinforcement Learning (RL), notably Group Relative Policy Optimization (GRPO), can intrinsically elicit and enhance the reasoning capabilities of Vision-Language Models (VLMs). However, despite the promise, the underlying mechanisms that drive the effectiveness of RL models as well as their limitations remain underexplored. In this paper, we highlight a fundamental behavioral distinction between RL and base models, where the former engages in deeper yet narrow reasoning, while base models, despite less refined along individual path, exhibit broader and more diverse thinking patterns. Through further analysis of training dynamics, we show that GRPO is prone to diversity collapse, causing models to prematurely converge to a limited subset of reasoning strategies while discarding the majority of potential alternatives, leading to local optima and poor scalability. To address this, we propose Multi-Group Policy Optimization (MUPO), a simple yet effective approach designed to incentivize divergent thinking across multiple solutions, and demonstrate its effectiveness on established benchmarks.

Multimodal Large Language Models (MLLMs) perform strong vision-language reasoning under standard conditions but fail in extreme illumination, where RGB inputs lose irrevocable structure and semantics. We propose Event-MLLM, an event-enhanced model that performs all-light visual reasoning by dynamically fusing event streams with RGB frames. Two key components drive our approach: an Illumination Indicator -- a learnable signal derived from a DINOv2 branch that represents exposure degradation and adaptively modulates event-RGB fusion -- and an Illumination Correction Loss that aligns fused features with non-degraded (normal-light) semantics in the latent space, compensating for information lost in extreme lighting. We curate the first multi-illumination event-instruction corpus for MLLMs, with 2,241 event-RGB samples (around 6 QA pairs each) across diverse scenes and 17 brightness rates (0.05x - 20x), plus an instruct-following benchmark for reasoning, counting, and fine-grained recognition under extreme lighting. Experiments show that Event-MLLM markedly outperforms general-purpose, illumination-adaptive, and event-only baselines, setting a new state of the art in robust multimodal perception and reasoning under challenging illumination.

Recent Vision-Language Models (VLMs) have demonstrated remarkable multimodal understanding capabilities, yet the redundant visual tokens incur prohibitive computational overhead and degrade inference efficiency. Prior studies typically relies on [CLS] attention or text-vision cross-attention to identify and discard redundant visual tokens. Despite promising results, such solutions are prone to introduce positional bias and, more critically, are incompatible with efficient attention kernels such as FlashAttention, limiting their practical deployment for VLM acceleration. In this paper, we step away from attention dependencies and revisit visual token compression from an information-theoretic perspective, aiming to maximally preserve visual information without any attention involvement. We present ApET, an Approximation-Error Token compression framework. ApET first reconstructs the original visual tokens with a small set of basis tokens via linear approximation, then leverages the approximation error to identify and drop the least informative tokens. Extensive experiments across multiple VLMs and benchmarks demonstrate that ApET retains 95.2% of the original performance on image-understanding tasks and even attains 100.4% on video-understanding tasks, while compressing the token budgets by 88.9% and 87.5%, respectively. Thanks to its attention-free design, ApET seamlessly integrates with FlashAttention, enabling further inference acceleration and making VLM deployment more practical.

Recent Multimodal Large Language Models (MLLMs) have significantly advanced e-commerce product understanding. However, they still face three challenges: (i) the modality imbalance induced by modality mixed training; (ii) underutilization of the intrinsic alignment relationships among visual and textual information within a product; and (iii) limited handling of noise in e-commerce multimodal data. To address these, we propose MOON2.0, a dynamic modality-balanced MultimOdal representation learning framework for e-commerce prOduct uNderstanding. It comprises: (1) a Modality-driven Mixture-of-Experts (MoE) that adaptively processes input samples by their modality composition, enabling Multimodal Joint Learning to mitigate the modality imbalance; (2) a Dual-level Alignment method to better leverage semantic alignment properties inside individual products; and (3) an MLLM-based Image-text Co-augmentation strategy that integrates textual enrichment with visual expansion, coupled with Dynamic Sample Filtering to improve training data quality. We further release MBE2.0, a co-augmented Multimodal representation Benchmark for E-commerce representation learning and evaluation at https://huggingface.co/datasets/ZHNie/MBE2.0. Experiments show that MOON2.0 delivers state-of-the-art zero-shot performance on MBE2.0 and multiple public datasets. Furthermore, attention-based heatmap visualization provides qualitative evidence of improved multimodal alignment of MOON2.0.

The development of 3D Vision-Language Models (VLMs), crucial for applications in robotics, autonomous driving, and augmented reality, is severely constrained by the scarcity of paired 3D-text data. Existing methods rely solely on next-token prediction loss, using only language tokens for supervision. This results in inefficient utilization of limited 3D data and leads to a significant degradation and loss of valuable geometric information in intermediate representations.To address these limitations, we propose \mname , a novel feature-level alignment regularization method. \mname explicitly supervises intermediate point cloud representations to preserve fine-grained 3D geometric-semantic information throughout the language modeling process. Specifically, we constrain the intermediate point cloud tokens within the LLM to align with visual input tokens via a consistency loss. By training only a lightweight alignment projector and LoRA adapters, \mname achieves explicit feature-level supervision with minimal computational overhead, effectively preventing geometric degradation.Extensive experiments on ModelNet40 and Objaverse datasets demonstrate that our method achieves 2.08 pp improvement on average for classification tasks, with a substantial 7.50 pp gain on the challenging open-vocabulary Objaverse classification task and 4.88 pp improvement on 3D object captioning evaluated by Qwen2-72B-Instruct, validating the effectiveness of \mname .

Sinusoidal neural networks (SIRENs) are powerful implicit neural representations (INRs) for low-dimensional signals in vision and graphics. By encoding input coordinates with sinusoidal functions, they enable high-frequency image and surface reconstruction. However, training SIRENs is often unstable and highly sensitive to frequency initialization: small frequencies produce overly smooth reconstructions in detailed regions, whereas large ones introduce spurious high-frequency components that manifest as noise in smooth areas such as image backgrounds. To address these challenges, we propose SASNet, a Spatially-Adaptive Sinusoidal Network that couples a frozen frequency embedding layer, which explicitly fixes the network's frequency support, with jointly learned spatial masks that localize neuron influence across the domain. This pairing stabilizes optimization, sharpens edges, and suppresses noise in smooth areas. Experiments on 2D image and 3D volumetric data fitting as well as signed distance field (SDF) reconstruction benchmarks demonstrate that SASNet achieves faster convergence, superior reconstruction quality, and robust frequency localization--assigning low- and high-frequency neurons to smooth and detailed regions respectively--while maintaining parameter efficiency.

Automatic radiology report generation (RRG) aims to translate medical images into diagnostic text, reducing radiologists' workload and standardizing clinical documentation. Nonetheless, existing approaches mainly focus on single-time point analysis and fail to capture temporal disease evolution across longitudinal examinations. While recent longitudinal RRG (LRRG) approaches incorporate historical data, they often combine images from different time points within a single representation space, leading to blurred semantics and inconsistent temporal reasoning. In this work, we propose a Temporal Decoupling with Iterative MutualRefinement Model (TIM), a two-stage framework that explicitly decouples spatial pathology from temporal progression and iteratively refines reports through mutual feedback. Stage I performs temporal-decoupled representation learning, separating temporal evolution patterns from disease-specific features and generating radiology reports for both prior and current studies. Stage II introduces a mutual report refinement mechanism that identifies diagnostic inconsistencies within prior reports and iteratively rectifies both prior and current reports through error-sensitive feedback. Experiments on the Longitudinal-MIMIC dataset demonstrate that TIM surpasses existing single-image and longitudinal baselines, achieving new state-of-the-art performance across both language and clinical metrics. Code is available at https://github.com/yihengd/TIM.

Automatic X-ray prohibited items detection is vital for security inspection and has been widely studied. Traditional methods rely on visual modal, often struggling with complex threats. While recent studies incorporate language to guide single-view images, human inspectors typically use dual-view images in practice. This raises the question: can the second view provide constraints similar to a language modality? In this work, we introduce DualXrayBench, the first comprehensive benchmark for X-ray inspection that includes multiple views and modalities. It supports eight tasks designed to test cross-view reasoning. In DualXrayBench, we introduce a dual-view caption corpus consisting of 45,613 dual-view images across 12 categories with corresponding captions. Building upon these data, we propose the Geometric (cross-view)-Semantic (cross-modality) Reasoner (GSR), a multimodal model that jointly learns correspondences between cross-view geometry and cross-modal semantics, treating the second-view images as a "language-like modality". To enale this, we construct the GSXray dataset, with structured Chain-of-Thought sequences: top , side, conclusion. Comprehensive evaluations on DualXrayBench demonstrate that GSR achieves significant improvements across all X-ray tasks, offering a new sight for real-world X-ray inspection.

We present **ReViSe** (_**Re**asoning with **Vi**deo **S**parsity_), a framework that combines multi-round reasoning with adaptive frame selection for video question answering (VQA). Existing vision-language models (VLMs) uniformly sample video frames, which introduces redundancy or irrelevancy. In contrast, ReViSe*interactively selects informative frames through multi-round reasoning. To achieve this, ReViSe includes three modules: a multi-round conversation module that retains frame selection history as memory; a reasoning tracer that maintains a chain-of-thought across rounds; and a self-correction mechanism that enforces structural and behavioral validity. ReViSe integrates seamlessly with both proprietary and open-source VLMs. It supports proprietary models in a "plug-and-play" manner and enables reinforcement fine-tuning for open-source models. Experiments on multiple VQA benchmarks show that **ReViSe** improves the video understanding ability of VLMs by improving accuracy while reducing the number of frames used.

Videos are continuous 2D projections of 3D worlds. After training on large video data, will global 3D understanding naturally emerge? We study this by quantifying the 3D understanding of existing Video Foundation Models (VidFMs) pretrained on vast video data. We propose the first model-agnostic framework that measures the 3D awareness of various VidFMs by estimating multiple 3D properties from their features via shallow read-outs. Our study presents meaningful findings regarding the 3D awareness of VidFMs on multiple axes. In particular, we show that state-of-the-art video generation models exhibit a strong understanding of 3D objects and scenes, despite not being trained on any 3D data. Such understanding can even surpass that of large expert models specifically trained for 3D tasks. Our findings, together with the 3D benchmarking of major VidFMs, provide valuable observations for building scalable 3D models.

Continual learning for video--language understanding is increasingly important as models face non-stationary data, domains, and query styles, yet prevailing solutions blur what should stay stable versus what should adapt, rely on static routing/capacity, or require replaying past videos. We aim to explicitly specify where stability lives and where plasticity should be focused under realistic memory and privacy constraints. We introduce Affordance-First Decomposition (AFD): videos are mapped to slowly varying affordance tokens that form a shared, time-aligned substrate, while a lightweight, query-routed, conflict-aware scheduler concentrates adaptation and grows capacity only when needed. The substrate is stabilized via weak alignment and teacher consistency, and training uses question-only replay. AFD achieves state-of-the-art across protocols: 51.6% average accuracy with -1.8% forgetting on domain-incremental VideoQA, ViLCo R@1@0.5 of 29.6% (MQ) and 20.7% (NLQ) with 18.4% stAP@0.25 (VQ), and 39.5% accuracy with -1.6% forgetting on time-incremental iVQA. Overall, AFD offers an explicit, interpretable split between a stable interaction-centered substrate and targeted adaptation.

Instruction-based image editing through natural language has emerged as a powerful paradigm for intuitive visual manipulation. While recent models achieve impressive results on single edits, they suffer from severe quality degradation under multi-turn editing. Through systematic analysis, we identify progressive loss of high-frequency information as the primary cause of this quality degradation. We present FreqEdit, a training-free framework that enables stable editing across 10+ consecutive iterations. Our approach comprises three synergistic components: (1) high-frequency feature injection from reference velocity fields to preserve fine-grained details, (2) an adaptive injection strategy that spatially modulates injection strength for precise region-specific control, and (3) a path compensation mechanism that periodically recalibrates the editing trajectory to prevent over-constraint. Extensive experiments demonstrate that FreqEdit achieves superior performance in both identity preservation and instruction following compared to seven state-of-the-art baselines.

Distributed Image Compression (DIC) is crucial for multi-view transmission, especially when operating at extremely low bitrates (< 0.1 bpp). Its core challenge is effectively utilizing side information to achieve high-quality reconstruction under strict bitrate budgets. However, existing DIC approaches struggle to exploit global context and object-level details from side information, leading to local blurring and the loss of fine details in the reconstruction. To address these limitations, we propose a Multimodal DIC framework (MDIC), which, for the first time, leverages side information in a multimodal manner into the DIC paradigm, effectively preserving fine-grained local details and enhancing global perceptual quality in reconstructed images. Specifically, we introduce a text-to-image diffusion-based decoder conditioned on textual side information extracted from correlated images to capture shared global semantics. Moreover, we design a feature-mask generator, supervised by a multimodal fine-grained alignment task, to strengthen the exploitation of visual side information. The generated mask serves two purposes: first, it guides the extraction of fine-grained details from losslessly transmitted side information to preserve the semantic consistency of reconstructed details; second, it regulates the extraction of clustered feature representations from the quantized VQ-VAE embeddings, compensating for category information lost under the extreme compression of the primary image. Extensive experiments on the widely used KITTI Stereo and Cityscapes datasets demonstrate that MDIC achieves state-of-the-art perceptual quality at extremely low bitrates.

Vision-Language-Action (VLA) models trained via imitation learning suffer from significant performance degradation in data-scarce scenarios due to their reliance on large-scale demonstration datasets. Although reinforcement learning (RL)-based post-training has proven effective in addressing data scarcity, its application to VLA models is hindered by the non-resettable nature of real-world environments. This limitation is particularly critical in high-risk domains such as industrial automation, where interactions often induce state changes that are costly or infeasible to revert. Furthermore, existing VLA approaches lack a reliable mechanism for detecting task completion, leading to redundant actions that reduce overall task success rates. To address these challenges, we propose RehearseVLA, an RL-based post-training framework that replaces physical interaction with a low-cost world model-based virtual simulator. RehearseVLA consists of two key components: (1) a physically-consistent world simulator that generates temporally consistent future visual observations, and (2) a vision-language model (VLM)-guided instant reflector that provides continuous reward signals and predicts action termination. This simulated environment enables VLA models to safely explore and generalize beyond their initial imitation learning distribution. Our method achieves notable performance gains with as few as five expert demonstrations per task. Experiments on complex robotic manipulation tasks demonstrate that RehearseVLA effectively overcomes the data inefficiency, safety constraints, and inefficient execution of conventional VLA models that rely on real-world interaction, offering a practical and scalable solution for post-training in resource-constrained settings. Our code is available at https://github.com/iSEE-Laboratory/RehearseVLA.

High-quality data has become a primary driver of progress under scale laws, with curated datasets often outperforming much larger unfiltered ones at lower cost. Online data curation extends this idea by dynamically selecting training samples based on the model's evolving state. While effective in classification and multimodal learning, existing online sampling strategies rarely extend to object detection because of its structural complexity and domain gaps. We introduce DetGain, an online data curation method specifically for object detection that estimates the marginal perturbation of each image to dataset-level Average Precision (AP) based on its prediction quality. By modeling global score distributions, DetGain efficiently estimates the global AP change and computes teacher-student contribution gaps to select informative samples at each iteration. The method is architecture-agnostic and minimally intrusive, enabling straightforward integration into diverse object detection architectures. Experiments on the COCO dataset with multiple representative detectors show consistent improvements in accuracy. DetGain also demonstrates strong robustness under low-quality data and can be effectively combined with knowledge distillation techniques to further enhance performance, highlighting its potential as a general and complementary strategy for data-efficient object detection.

Theory of Mind (ToM) refers to the ability to infer others' mental states, such as beliefs, desires, and intentions. Current vision-language embodied agents lack ToM-based decision-making, and existing benchmarks focus solely on human mental states while ignoring the agent's own perspective, hindering coherent decision and action generation. To address this, we propose MindPower, a Robot-Centric framework integrating Perception, Mental Reasoning, Decision Making and Action. Given multimodal inputs, MindPower first perceives the environment and human states, then performs ToM Reasoning to model both self and others, and finally generates decisions and actions guided by inferred mental states. Furthermore, we introduce Mind-Reward, a novel optimization objective that encourages VLMs to produce consistent ToM Reasoning and behavior. Our model outperforms GPT-4o by 12.77% in decision making and 12.49% in action generation. Project page: https://zhangdaxia22.github.io/MindPower/.