CVPR 2026 论文列表

Vision-language models like CLIP excel at zero-shot recognition but struggle with continual learning due to two critical issues: (1) severe distribution gap between pretraining captions and post-training class names, and (2) performance mismatch between vision-only and dual-encoder approaches--vision-only methods achieve 20% higher accuracy on fine-grained tasks while CLIP dominates on natural images. We propose Learning from Itself (LfI), which mines CLIP's internal knowledge to address both challenges. First, we generate pseudo-captions by optimizing learnable tokens to minimize CLIP's contrastive loss, creating auxiliary training signals that bridge the pretraining-finetuning distribution gap without external models. Second, we introduce adaptive mutual distillation that dynamically weights knowledge transfer between CLIP's text encoder and a temporary vision classifier based on their instantaneous performance--stronger branches teach more, weaker ones learn more. At inference, only the original CLIP architecture is used, having absorbed discriminative knowledge from both branches. LfI achieves state-of-the-art results across multiple continual learning benchmarks, demonstrating that CLIP can effectively teach itself to continually learn new tasks. Code is available at: https://github.com/jordangong/continual-learning.

We introduce GaussianZoom, a generative zoom-in 3D reconstruction system with an iterative progressive framework that combines geometry-consistent scene modeling and multi-scale semantic reasoning to enable high-fidelity extreme zoom-in rendering from low-resolution inputs. To achieve this, we develop a novel multi-view consistent super-resolution module with depth-based feature warping and VLM-driven detail synthesis, ensuring accurate multi-view correspondence while enriching fine-scale appearance beyond the observed resolution. To support zooming across large magnification ranges, we further introduce a new expandable continuous Level-of-Detail hierarchy that dynamically modulates Gaussian visibility for smooth, alias-free cross-scale rendering. Experiments on Mip-NeRF360 and Tanks&Temples demonstrate that GaussianZoom achieves superior perceptual quality, multi-view consistency, and robustness under extreme magnification, establishing a strong baseline for generative zoom-in 3D scene reconstruction.

Vision Transformers (ViTs) exhibit robust performance across diverse visual scenarios. However, their efficiency is constrained by excessive token counts. Token merging offers a viable solution for achieving efficient ViTs. Existing methods merge tokens based solely on specific characteristics within the attention mechanism, which changes significantly across layers. In this paper, we propose a novel training-free SAliency-Driven Token Merging (SAD-TM) approach by leveraging not only the semantic relevance in the attention space but also the latent visual saliency of input patches. Our SAD-TM is inspired by the discovery that saliency-based statistics can directly capture the causal relationship between model input and output, regardless of the layers. Based on the observation, we develop a method that is mathematically formulated to merge tokens with high saliency outliers. The principle behind our merging is that tokens with high saliency outliers usually imply inconsistencies with the global gradient direction, and thus can be merged safely. Besides, our systematic analysis indicates that class attention shows considerable variation across early blocks, so a deferred merging strategy is introduced to optimize the selection of merging rates. In a training-free manner, SAD-TM demonstrates superior performance across various ViT architectures. Especially, with a FLOPs compression of 23.08% on DeiT-Tiny, SAD-TM achieves a Top-1 Accuracy comparable to the pretrained baseline on ImageNet dataset. The code will be available soon.

As large language models (LLMs) continue to advance, there is increasing interest in their ability to infer human mental states and demonstrate a human-like Theory of Mind (ToM). Most existing ToM evaluations, however, are centered on text-based inputs, while scenarios relying solely on visual information receive far less attention. This leaves a gap, since real-world human-AI interaction typically requires multimodal understanding. In addition, many current methods regard the model as a black box and rarely probe how its internal attention behaves in multiple-choice question answering (QA). The impact of LLM hallucinations on such tasks is also underexplored from an interpretability perspective. To address these issues, we introduce VisionToM, a vision-oriented intervention framework designed to strengthen task-aware reasoning. The core idea is to compute intervention vectors that align visual representations with the correct semantic targets, thereby steering the model's attention through different layers of visual features. This guidance reduces the model's reliance on spurious linguistic priors, leading to more reliable multimodal language model (MLLM) outputs and better QA performance. Experiments on the EgoToM benchmark--an egocentric, real-world video dataset for ToM with three multiple-choice QA settings--demonstrate that our method substantially improves the ToM abilities of MLLMs. Furthermore, results on an additional open-ended generation task show that VisionToM enables MLLMs to produce free-form explanations that more accurately capture agents' mental states, pushing machine-human collaboration toward greater alignment.

The rise of AI-generated images (AIGIs) poses growing challenges for digital authenticity, prompting the need for efficient, generalizable image forgery detection systems. Existing methods, whether non-LLM-based or LLM-based, exhibit distinct advantages and limitations. While non-LLM-based models offer efficient low-level artifact detection, they often lack semantic understanding. Conversely, LLM-based methods provide strong semantic reasoning and explainability but are computationally intensive and less sensitive to subtle visual artifacts. Moreover, the true contribution of explanatory reasoning texts to forgery detection performance remains unclear. In this work, we investigate the intrinsic value and potential of LLM-generated reasoning texts, considering it a source of generalization and semantic-error sensitivity. Based on these findings, we propose ReAlign, a novel framework that distills high-quality reasoning texts generated by a GRPO-optimized LLM into a lightweight AIGI detector via contrastive learning. ReAlign effectively inherits the generalization ability and semantic sensitivity capability of reasoning textual representations, while remaining efficient and lightweight for deployment. Moreover, ReAlign adopts a tailored joint optimization strategy that integrates contrastive loss for image-text alignment and classification loss for accurate forgery discrimination. Experimental results on AIGCDetectBenchmark, AIGI-Holmes, and our newly constructed UltraSynth-10k demonstrate that ReAlign consistently outperforms existing state-of-the-art detectors in both accuracy and generalization, particularly when facing complex, high-fidelity forgeries from modern generative models.

Vision-Language Models (VLMs) have achieved remarkable success in visual question answering tasks, but their reliance on large numbers of visual tokens introduces significant computational overhead. While existing efficient VLM approaches reduce visual tokens through fixed-ratio compression, they operate passively and lack the ability to adapt to varying task requirements. This motivates a fundamental question: Can VLMs autonomously determine the minimum number of visual tokens required for each sample? Inspired by human active vision mechanisms, we introduce AdaptVision, an efficient VLM paradigm that enables adaptive visual token acquisition through a coarse-to-fine approach. Our model initially processes compressed visual tokens from low-resolution images and selectively acquires additional visual information by invoking a bounding box tool to crop key regions when necessary. We train AdaptVision using a reinforcement learning framework that carefully balances accuracy and efficiency. Central to our approach is Decoupled Turn Policy Optimization (DTPO), which decouples the learning objective into two components: (1) tool learning, which optimizes correct tool utilization, and (2) accuracy improvement, which refines the generated responses to improve answer correctness. Based on this formulation, we further decouple advantage estimation by computing separate advantages for tokens associated with each objective. This formulation enables more effective optimization for AdaptVision compared to vanilla GRPO. Comprehensive experiments across multiple VQA benchmarks demonstrate that AdaptVision achieves superior performance while consuming substantially fewer visual tokens than state-of-the-art efficient VLM methods.

Brain imaging analysis is crucial for diagnosing and treating brain disorders, and multimodal large language models (MLLMs) are increasingly supporting it. However, current brain imaging visual question-answering (VQA) benchmarks either cover a limited number of imaging modalities or are restricted to coarse-grained pathological descriptions, hindering a comprehensive assessment of MLLMs across the full clinical continuum. To address these, we introduce OmniBrainBench, the first comprehensive multimodal VQA benchmark specifically designed to assess the multimodal comprehension capabilities of MLLMs in brain imaging analysis with closed- and open-ended evaluations. OmniBrainBench comprises 15 distinct brain imaging modalities collected from 30 verified medical sources, yielding 9,527 validated VQA pairs and 31,706 images. It simulates clinical workflows and encompasses 15 multi-stage clinical tasks rigorously validated by a professional radiologist. Evaluations of 24 state-of-the-art models, including open-source general-purpose, medical, and proprietary MLLMs, highlight the substantial challenges posed by OmniBrainBench. Experiments reveal that proprietary MLLMs like Gemini-2.5-Pro (66.58%) outperform open-source and medical MLLMs yet lag far behind physicians (91.35%), while medical MLLMs show wide variance in closed- and open-ended VQA. Open-source general-purpose MLLMs generally trail but excel in specific tasks, and all MLLMs fall short in complex preoperative reasoning, revealing a critical visual-to-clinical gap. OmniBrainBench establishes a new standard to assess MLLMs in brain imaging analysis, highlighting the gaps against physicians. We publicly release our benchmark and code at link.

Multimodal misinformation poses an escalating challenge that often evades traditional detectors, which are opaque black boxes and fragile against new manipulation tactics. We present Probabilistic Concept Graph Reasoning (PCGR), an interpretable, modular, and evolvable framework that reframes multimodal misinformation detection (MMD) as structured, concept-based reasoning. PCGR follows a build-then-infer paradigm, which first constructs a graph of human-understandable concept nodes, including novel high-level concepts automatically discovered and validated by multimodal large language models (MLLMs), and then applies hierarchical attention over this concept graph to infer claim veracity. This design produces interpretable reasoning chains linking evidence to conclusions. Experiments demonstrate that PCGR achieves state-of-the-art MMD accuracy and robustness to emerging manipulation types, outperforming prior methods in both coarse detection and fine-grained manipulation recognition.

Modern vision-language models (VLMs) are expected to have abilities of spatial reasoning with diverse scene complexities, but evaluating such abilities is difficult due to the lack of benchmarks that are not only diverse and scalable but also fully customizable. Existing benchmarks offer limited customizability over the scene complexity and are incapable of isolating and analyzing specific VLM failure modes under distinct spatial conditions. To address this gap, instead of individually presenting benchmarks for different scene complexities, in this paper we present InfiniBench, a fully automated, customizable and user-friendly benchmark generator that can synthesize a theoretically infinite variety of 3D scenes with parameterized control on scene complexity. InfiniBench uniquely translates scene descriptions in natural language into photo-realistic videos with complex and physically plausible 3D layouts. This is achieved through three key innovations: 1) a LLM-based agentic framework that iteratively refines procedural scene constraints from scene descriptions; 2) a flexible cluster-based layout optimizer that generates dense and cluttered scenes previously intractable for procedural methods; and 3) a task-aware camera trajectory optimization method that renders scenes into videos with full object coverage as VLM input. Experiments demonstrate that InfiniBench outperforms state-of-the-art procedural and LLM-based 3D generation methods in prompt fidelity and physical plausibility, especially in high-complexity scenarios. We further showcased the usefulness of InfiniBench, by generating benchmarks for representative spatial reasoning tasks including measurement, perspective-taking and spatiotemporal tracking. The source codes of InfiniBench are available at https://github.com/pittisl/infinibench, and some sample benchmarks being generated by InfiniBench are available at https://huggingface.co/datasets/Haoming645/infinibench.

Earth observation data presents a unique challenge: it is spatial like images, sequential like video or text, and highly multimodal. We present Helios: a multimodal, spatio-temporal foundation model that employs a novel self-supervised learning formulation, masking strategy, and loss all designed for the Earth observation domain. Helios achieves state-of-the-art performance compared to 12 other foundation models across a variety of research benchmarks and real-world tasks from external partners. When evaluating embeddings Helios achieves the best performance on 15 out of 24 tasks, and with full fine-tuning it is the best on 19 of 29 tasks. We deploy Helios as the backbone of an end-to-end platform for data collection, labeling, training, and inference of Earth observation models. The Helios platform puts frontier foundation models and powerful data management tools into the hands of non-profits and NGOs working to solve the world's biggest problems. Helios source code, training data, and pre-trained weights are available at REDACTED.

Despite recent advances in personalized image generation, existing models consistently fail to produce reliable multi-human scenes, often merging or losing facial identity. We present Ar2Can, a novel two-stage framework that disentangles spatial planning from identity rendering for multi-human generation. The Architect predicts structured layouts, specifying where each person should appear. The Artist then synthesizes photorealistic images, guided by a spatially-grounded face matching reward that combines Hungarian spatial alignment with identity similarity. This approach ensures faces are rendered at correct locations and faithfully preserve reference identities. We develop two Architect variants, seamlessly integrated with our diffusion-based Artist model. This is optimized via Group Relative Policy Optimization (GRPO) using compositional rewards for count accuracy, image quality, and identity matching. Evaluated on the MultiHuman-Testbench, Ar2Can achieves substantial improvements in both count accuracy and identity preservation, while maintaining high perceptual quality. Notably, our method achieves these results using primarily synthetic data, without requiring real multi-human images. Project page: https://qualcomm-ai-research.github.io/ar2can/

Many audio-visual learning methods have focused on aligning audio and visual information, either through semantic or temporal correspondence. However, most of these works have utilized monaural audio, which does not contain information about the spatial location of the sound source. In contrast, humans and other animals utilize binaural hearing to perceive this spatial information. Combining spatial sound and visual perception enables powerful high-level reasoning: for example, a person looking for their phone may hear the ringing sound coming from a backpack sitting on a table, and quickly infer that the missing phone is inside the backpack. In this paper, we investigate the problem of Audio-Visual Spatial Reasoning. We design a spatial audio-visual question answering dataset to cover scenarios where semantic correspondence between audio and visual signals is absent but spatial alignment exists, as well as cases with multiple audio-visual semantic correspondences that require spatial reasoning to disambiguate. We propose a model that learns spatial comprehension across the audio and vision modalities by connecting them with a large language model and experimentally demonstrate that spatial sound perception is an essential part of our task.

In-Context Learning (ICL) is a significant paradigm for Large Multimodal Models (LMMs), using a few in-context demonstrations (ICDs) for new task adaptation. However, its performance is sensitive to demonstration configurations and computationally expensive. Mathematically, the influence of these demonstrations can be decomposed into a dynamic mixture of the standard attention output and the context values. Current approximation methods simplify this process by learning a "shift vector". Inspired by the exact decomposition, we introduce High-Fidelity In-Context Learning (HiFICL) to more faithfully model the ICL mechanism. HiFICL consists of three key components: 1) a set of "virtual key-value pairs" to act as a learnable context, 2) a low-rank factorization for stable and regularized training, and 3) a simple end-to-end training objective. From another perspective, this mechanism constitutes a form of context-aware Parameter-Efficient Fine-Tuning (PEFT). Extensive experiments show that HiFICL consistently outperforms existing approximation methods on several multimodal benchmarks.

The rapid advancement of educational applications, artistic creation, and AI-generated content (AIGC) technologies has substantially increased practical requirements for comprehensive Image Aesthetics Assessment (IAA), particularly demanding methods capable of delivering both quantitative scoring and professional understanding. Multimodal Large Language Model (MLLM)-based IAA methods demonstrate stronger perceptual and generalization capabilities compared to traditional approaches, yet they suffer from modality bias (score-only or text-only) and lack fine-grained attribute decomposition, thereby failing to support further aesthetic assessment. In this paper, we present: (1) ArtiMuse, an innovative MLLM-based IAA model with Joint Scoring and Expert-Level Understanding capabilities; (2) ArtiMuse-10K, the first expert-curated IAA dataset comprising 10,000 images spanning 5 main categories and 15 subcategories, each annotated by professional experts with 8-dimensional attributes analysis and a holistic score. Both the model and dataset will be made public.

This paper studies the problem of universal test-time prompt learning for vision-language models (VLMs) which aims to enhance prompt learning for a pre-trained VLM via unlabeled target data containing out-of-distribution (OOD) samples. However, existing test-time adaptation approaches often overlook class-specific diversity in the target domain and rely on unreliable pseudo-labels due to inadequate uncertainty estimation, which may result in additional adaptation bias during test time. Towards this end, we propose a novel framework named Separability-aware Conjugate Optimization with Prototypical Retrieval (STAR) for universal test-time prompt learning of VLMs. The core of our STAR is to incorporate a separability-aware gating mechanism into conjugate optimization for reliable pseudo-learning with OOD samples. In particular, we first compute the Fisher score to quantify the separability between in-distribution (ID) and OOD samples, which guides our soft gating mechanism for divided training. Then, we employ conjugate optimization to derive reliable pseudo-labels of unlabeled data for test-time adaptation. To further mitigate biases in OOD detection, we maintain a dynamic memory bank which stores high-confidence samples to build class-wise prototypes, which would serve as queries for prototypical retrieval to calibrate OOD detection. Extensive experiments on multiple benchmarks demonstrate that STAR consistently outperforms competing baseline methods.

Medical image segmentation is vital for clinical diagnosis and quantitative analysis, yet remains challenging due to the heterogeneity of imaging modalities and the high cost of pixel-level annotations. Although general interactive segmentation models like SAM have achieved remarkable progress, their transfer to medical imaging still faces two key bottlenecks: (i) the lack of adaptive mechanisms for modality- and anatomy-specific tasks, which limits generalization in out-of-distribution medical scenarios; and (ii) current medical adaptation methods fine-tune on large, heterogeneous datasets without selection, leading to noisy supervision, higher cost, and negative transfer. To address these issues, we propose SegMoTE, an efficient and adaptive framework for medical image segmentation. SegMoTE preserves SAM's original prompt interface, efficient inference, and zero-shot generalization while introducing only a small number of learnable parameters to dynamically adapt across modalities and tasks. In addition, we design a progressive prompt tokenization mechanism that enables fully automatic segmentation, significantly reducing annotation dependence. Trained on MedSeg-HQ, a curated dataset less than 1% of existing large-scale datasets, SegMoTE achieves SOTA performance across diverse imaging modalities and anatomical tasks. It represents the first efficient, robust, and scalable adaptation of general segmentation models to the medical domain under extremely low annotation cost, advancing the practical deployment of foundation vision models in clinical applications.

Radiology report generation (RRG) aims to automatically describe medical images via free-text reports. In clinical practice, comparing current and prior chest X-rays is essential for assessing disease progression, motivating the development of longitudinal RRG methods. However, most existing approaches often struggle to capture fine-grained temporal changes, as they often rely on unidirectional alignments or static reasoning pipelines, overlooking the bidirectional and asymmetric nature of disease evolution. To tackle these challenges, we propose BiOTPrompt, a novel framework for disease evolution-aware radiology report generation, which introduces a Bidirectional Optimal Transport (BiOT) mechanism to explicitly model progression dynamics between historical and current chest X-rays. By analyzing the asymmetry between bidirectional transport plans, BiOTPrompt can identify newly emerged and resolved regions, which are then used to construct dynamic prompts that guide large language models (LLMs) in generating clinically relevant diagnostic reports. Furthermore, we incorporate a vision-language consistency constraint to ensure alignment between visual evidence and textual descriptions, mitigating hallucinations and enhancing factual correctness. Extensive experiments on the Longitudinal-MIMIC dataset demonstrate that BiOTPrompt achieves state-of-the-art performance in both language metrics and clinical relevance, setting a new standard for longitudinal radiology report generation.

Cross-degradation generalization remains a critical challenge for RGB-infrared multimodal object detection, especially when training data covers limited degradation types. This paper presents a distribution alignment framework with a key insight: aligning fused features to the pretrained distribution where the frozen detector performs optimally, rather than adapting to training-specific degradations. By freezing the pretrained detector and training only a lightweight fusion module, our approach leverages complementary infrared information to reduce distribution shift while maintaining computational efficiency. The method achieves state-of-the-art results on three benchmarks with 4x faster training. Critically, we demonstrate that aligning to the pretrained distribution substantially outperforms aligning to training degradations when generalizing to unseen scenarios.

Vision-Language Models (VLMs) are increasingly adopted for medical applications, but their clinical utility is limited by a core weakness in quantitative reasoning. This limitation affects tasks ranging from regression of lesion sizes to prediction of bounding-box coordinates and stems from the discrete tokenization schemes underlying Large Language Models (LLMs). To address this, we propose Rotary Number Encoding and Decoding (RNED), a principled method for embedding continuous numerical values directly in the representation space of a VLM. Analogous to rotary position encoding, RNED represents a scalar by applying a number-specific rotation matrix to a dedicated numeric token embedding. This norm-preserving transformation maintains ordinal structure over a wide numerical range and integrates seamlessly with pretrained model weights. For decoding, we introduce a robust score-matching-based scheme to recover continuous values from hidden states in the presence of stochastic noise. We evaluate RNED on two quantitative tasks: radiological measurement estimation and medical visual grounding. On both internal and public benchmarks, RNED consistently outperforms existing VLM baselines. Together, these results show that RNED offers a robust, generalizable solution for numerical reasoning in medical VLMs, enabling models that are both quantitatively reliable and clinically applicable.