CVPR 2026 论文列表

Multimodal large language models (MLLMs) project visual tokens into the embedding space of language models, yet the internal structuring and processing of visual semantics remain poorly understood. In this work, we introduce a two-fold analytical framework featuring a novel probing tool, EmbedLens, to conduct a fine-grained analysis. We uncover a pronounced semantic sparsity at the input level: visual tokens consistently partition into sink, dead, and alive categories. Remarkably, only the alive tokens, comprising ~60% of the total input, carry image-specific meaning. Furthermore, using a targeted patch-compression benchmark, we demonstrate that these alive tokens already encode rich, fine-grained cues (e.g., objects, colors, and OCR) prior to entering the LLM. Internal visual computations (such as visual attention and feed-forward networks) are redundant for most standard tasks. For the small subset of highly vision-centric tasks that actually benefit from internal processing, we reveal that alive tokens naturally align with intermediate LLM layers rather than the initial embedding space, indicating that shallow-layer processing is unnecessary and that direct mid-layer injection is both sufficient. Ultimately, our findings provide a unified mechanistic view of visual token processing, paving the way for more efficient and interpretable MLLM architectures through selective token pruning, minimized visual computation, and mid-layer injection.

Incremental open-vocabulary 3D instance-semantic mapping is essential for autonomous agents operating in complex everyday environments. However, it remains challenging due to the need for robust instance segmentation, real-time processing, and flexible open-set reasoning. Existing methods often rely on the closed-set assumption or dense per-pixel language fusion, which limits scalability and temporal consistency. We introduce OVI-MAP that decouples instance reconstruction from semantic inference. We propose to build a class-agnostic 3D instance map that is incrementally constructed from RGB-D input, while semantic features are extracted only from a small set of automatically selected views using vision-language models. This design enables stable instance tracking and zero-shot semantic labeling throughout online exploration. Our system operates in real time and outperforms state-of-the-art open-vocabulary mapping baselines on standard benchmarks.

Video Temporal Grounding (VTG) aims to localize the video segment that corresponds to a natural language query, which requires a comprehensive understanding of complex temporal dynamics. Existing Vision-LMMs typically perceive temporal dynamics via positional encoding, text-based timestamps, or visual frame numbering. However, these approaches exhibit notable limitations: assigning each frame a text-based timestamp token introduces additional computational overhead and leads to sparsity in visual attention, positional encoding struggles to capture absolute temporal information, and visual frame numbering often compromises spatial detail. To address these issues, we propose Temporal to Spatial Gridification (T2SGrid), a novel framework that reformulates video temporal understanding as a spatial understanding task. The core idea of T2SGrid is to process video content in clips rather than individual frames. we employ a overlapping sliding windows mechanism to segment the video into temporal clips. Within each window, frames are arranged chronologically in a row-major order into a composite grid image, effectively transforming temporal sequences into structured 2D layouts. The gridification not only encodes temporal information but also enhances local attention within each grid. Furthermore, T2SGrid enables the use of composite text timestamps to establish global temporal awareness. Experiments on standard VTG benchmarks demonstrate that T2SGrid achieves superior performance.

Vision-language models (VLMs) such as CLIP exhibit strong Out-of-distribution (OOD) detection capabilities by aligning visual and textual representations. Recent CLIP-based test-time adaptation methods further improve detection performance by incorporating external OOD labels. However, such labels are finite and fixed, while the real OOD semantic space is inherently open-ended. Consequently, fixed labels fail to represent the diverse and evolving OOD semantics encountered in test streams.To address this limitation, we introduce Test-time Textual Learning (TTL), a framework that dynamically learns OOD textual semantics from unlabeled test streams, without relying on external OOD labels.TTL updates learnable prompts using pseudo-labeled test samples to capture emerging OOD knowledge. To suppress noise introduced by pseudo-labels, we introduce an OOD knowledge purification strategy that selects reliable OOD samples for adaptation while suppressing noise. In addition, TTL maintains an OOD Textual Knowledge Bank that stores high-quality textual features, providing stable score calibration across batches. Extensive experiments on two standard benchmarks with nine OOD datasets demonstrate that TTL consistently achieves state-of-the-art performance, highlighting the value of textual adaptation for robust test-time OOD detection. Our code is available at https://github.com/figec/TTL.

Major state-of-the-art unified tokenizers predominantly adopt pixel reconstruction and feature alignment as pretext tasks, they leave key domains largely unexplored such as architecture, supervised objectives and tasks interaction, potentially resulting in limited performance. We systematically investigate the critical factors of a unified visual tokenizer and propose a novel framework that strengthens synergy between understanding and generation in various aspects. Our initial analysis focus on properties of frontier vision models, confirming inherent conflict in contrastive learning style models for unifying generation and understanding, and demonstrate distinct convergence behavior of codebooks. To address the above bottleneck, we hierarchically decouple the conflicting proxy tasks, enriching the diversity of semantic features supervision to enhance thesemantic and low-level capabilities. Subsequently, we further introduce attention-prioritized mapping strategy, which guides fine-grained generation with powerful semantic prior. Our method achieves rFID of 0.33 and zero-shot accuracy of 80.9% on ImageNet at 256x256 resolution, surpassing VILA-U by 7.6% and outperforms continuous embedding of SigLIP. When applied to discrete unified MLLMs, our 7B model exceeds TokenFlow-13B by 3.1% in understanding and achieve SOTA performance in GenAI-Bench and MJHQ-30K.

Vision-Language-Action (VLA) models have emerged as a promising paradigm where pretrained Vision-Language Models (VLMs) serve as System 2 for high-level reasoning, connected to action experts as System 1 for low-level motor control.However, current works fail to genuinely leverage VLM capabilities: VLMs produce latent embeddings that lack semantic interpretability, providing ambiguous and unstable guidance to downstream policies, while solely action supervision further causes VLMs to degenerate into parameter-heavy fusion encoders that memorize action patterns rather than perform generalized reasoning.To bridge this gap, we introduce SemanticVLA, which leverages VLM reasoning through synergistic dual-path design. Explicit trace reasoning generates interpretable spatial waypoints as textual coordinate sequences through the VLM's native language interface, directly reusing its pretrained spatial grounding to provide a thinking process for task planning. Latent action tokens complement trace reasoning by learning compact visuomotor primitives grounded in visual observations, providing more fine-grained action representations beyond pure coordinate prediction. This synergy enables trace reasoning to leverage VLM's multimodal understanding for refining latent token prediction, while latent tokens provide stable and grounded guidance that compensates for trace's numerical sensitivity.SemanticVLA achieves 97.0% average success rate on LIBERO and 65.1% on SimplerEnv WidowX, substantially outperforming strong baselines. More importantly, SemanticVLA maintains significantly more stable performance under instruction rephrasing in both simulation suites, and demonstrates strong advantages on real-world long-horizon and reasoning-intensive tasks.By bridging VLM reasoning and action expert through semantically explicit trace and visually grounded latent action tokens, our approach enables genuine reasoning rather than action memorization.

Existing methods achieve high-quality facial albedo capture under controllable lighting, which increases capture cost and limits usability. We propose WildCap, a novel method for high-quality facial albedo capture from a smartphone video recorded in the wild. To disentangle high-quality albedo from complex lighting effects in in-the-wild captures, we propose a novel hybrid inverse rendering framework. We first apply a data-driven method, i.e., SwitchLight, to convert the captured images into more constrained conditions and then adopt model-based inverse rendering. However, unavoidable local artifacts in network predictions, such as shadow-baking, are non-physical and thus hinder accurate inverse rendering of lighting and material. To address this, we propose a novel texel grid lighting model to explain non-physical effects as clean albedo illuminated by local physical lighting. During optimization, we jointly sample a diffusion prior for the albedo map and optimize the lighting, effectively resolving scale ambiguity between local lights and albedo. Other reflectance maps are then predicted from the albedo. Our method achieves significantly better results than prior arts in the same capture setup, closing the quality gap between in-the-wild and controllable recordings by a large margin.

Humans perceive and reason about their surroundings in four dimensions - three spatial and one temporal axis - by building persistent, structured internal representations that encode semantic meaning, spatial layout, and temporal dynamics. These multimodal memories enable them to recall past events, infer unobserved states, and integrate new information into context-dependent reasoning. Inspired by this capability, we introduce R4, a training-free framework for retrieval-augmented reasoning in 4D spatio-temporal space that equips vision-language models (VLMs) with structured, lifelong memory. R4 continuously constructs a 4D knowledge database by anchoring object-level semantic descriptions in metric space and time, yielding a persistent world model that can be shared across agents. At inference, natural language queries are decomposed into semantic, spatial, and temporal keys to retrieve relevant observations, which are integrated into the VLM's reasoning through an iterative retrieval-reasoning loop. Unlike classical retrieval-augmented generation methods, retrieval in R4 operates directly in structured 4D space, enabling episodic and collaborative reasoning without training. Experiments on embodied question answering and navigation benchmarks demonstrate that R4 substantially improves retrieval and reasoning over spatio-temporal information compared to baselines, advancing a new paradigm for embodied 4D reasoning in dynamic environments.

Existing video frame interpolation (VFI) methods often adopt a frame-centric approach, processing videos as independent short segments (e.g., triplets), which leads to temporal inconsistencies and motion artifacts. To overcome this, we propose a holistic, video-centric paradigm named Local Diffusion Forcing for Video Frame Interpolation (LDF-VFI). Our framework is built upon an auto-regressive diffusion transformer that models the entire video sequence to ensure long-range temporal coherence. To mitigate error accumulation inherent in auto-regressive generation, we introduce a novel skip-concatenate sampling strategy that effectively maintains temporal stability. Furthermore, LDF-VFI incorporates sparse, local attention and tiled VAE encoding, a combination that not only enables efficient processing of long sequences but also allows generalization to arbitrary spatial resolutions (e.g., 4K) at inference without retraining. An enhanced conditional VAE decoder, which leverages multi-scale features from the input video, further improves reconstruction fidelity. Empirically, LDF-VFI achieves state-of-the-art performance on challenging VFI benchmarks, demonstrating superior per-frame quality and temporal consistency, especially in scenes with large motion.

Vision language models (VLMs) have achieved remarkable success in semantic understanding tasks under language guidance, yet their potential for geometric perception remains largely underexplored. This paper introduces the first VLM-based depth completion framework. With almost no architectural modifications, we propose a sparse depth injection mechanism that extends the capability of VLM toward 3D perception through three key aspects: visual tokenization, textual prompt, and textual supervision. At the visual input side, sparse depth is tokenized to provide absolute scale and accurate geometric cues, alleviating the scale and camera ambiguities of RGB-only inputs. At the textual input side, a binary mask derived from sparse depth serves as a prompt, instructing the model where to complete and where to preserve. At the supervision side, the model is fine-tuned using text labels generated from sparse depth, requiring no ground-truth depth. Benefiting from the strong semantic priors and cross-modal expressiveness of VLM, our framework achieves superior zero-shot performance across diverse sensors, sparsity levels, and scenes.

Clinical video diagnosis, in which physicians assess dynamic tissue responses across procedural stages, is critical for detecting diseases such as cervical and colorectal cancers. Recent spatiotemporal models map visual progressions directly to diagnostic outputs, yet overlook two hallmarks of expert reasoning: clinically grounded diagnostic principles and hypothesis-driven counterfactual thinking. This gap leads existing methods to conflate causal pathological cues with non-pathological variations, thereby limiting their reliability in data-scarce settings. Here we show that explicitly modeling counterfactual tissue evolution, guided by clinical rules that encode expert diagnostic principles, substantially improves diagnostic robustness-emulating the hypothesis-driven reasoning clinicians employ in practice. We introduce MEDVCR, comprising a Counterfactual Generator (CG) that synthesizes hypothetical tissue transitions via diffusion modeling, a Counterfactual Representation Learning (CRL) module that enforces temporal consistency, pathological separability, and counterfactual alignment, and a Dual Diagnostic Prediction (DDP) strategy that combines video-level context with frame-level counterfactual contrast. On two representative tasks, MEDVCR achieves 93.0% Recall@1 on colposcopy (+10.2%) and 94.8% Average Precision (AP) on colonoscopy (+2.6%), outperforming all state-of-the-art (SOTA) baselines. We anticipate that this counterfactual reasoning paradigm will open new avenues for vision-based clinical diagnostic tasks toward more transparent and interpretable decision support.

We revisit dataset distillation from an outcome-centric perspective. Rather than aligning process surrogates (per-step gradients or training trajectories), Influence Matching (Inf-Match) aligns the final outcome of training: it learns a compact synthetic set whose effect on the converged parameters matches that of the full dataset. Concretely, we introduce a fully differentiable, sample-level influence estimator that quantifies parameter shifts from adding or removing data, without time-consuming inverse-Hessian products or convexity assumptions. The estimator runs in linear time by unrolling the optimization dynamics and applying a first-order Taylor approximation. We then learn the synthetic set by minimizing the mismatch between its influence and that of the real dataset, yielding outcome alignment rather than heuristic process imitation. Inf-Match delivers the best accuracy across standard classification benchmarks. For instance, on Tiny-ImageNet (IPC=10), Inf-Match attains 31.5%, a +4.7% improvement over NCFM. Beyond classification, Inf-Match scales to vision-language distillation on Flickr30K, outperforming strong process-matching baselines. For instance, with 200 to 1000 synthetic samples, our method achieved a leading impressive average on image/text retrieval tasks, higher than NCFM by 2.5%. The code will be released via https://github.com/hrtan/infmatch.

Long video understanding is essential for human-like intelligence, enabling coherent perception and reasoning over extended temporal contexts. While the emerging thinking-with-frames paradigm--which alternates between global temporal reasoning and local frame examination--has advanced the reasoning capabilities of video multi-modal large language models (MLLMs), it suffers from a significant efficiency bottleneck due to the progressively growing and redundant multi-modal context. To address this, we propose SpecTemp, a reinforcement learning-based Speculative Temporal reasoning framework that decouples temporal perception from reasoning via a cooperative dual-model design. In SpecTemp, a lightweight draft MLLM rapidly explores and proposes salient frames from densely sampled temporal regions, while a powerful target MLLM focuses on temporal reasoning and verifies the draft's proposals, iteratively refining its attention until convergence. This design mirrors the collaborative pathways of the human brain, balancing efficiency with accuracy. To support training, we construct the SpecTemp-80K dataset, featuring synchronized dual-level annotations for coarse evidence spans and fine-grained frame-level evidence. Experiments across multiple video understanding benchmarks demonstrate that SpecTemp not only maintains competitive accuracy but also significantly accelerates inference compared with existing thinking-with-frames methods.

Vision language models (VLMs) have achieved remarkable success in broad visual understanding, yet they remain challenged by object-centric reasoning on rare objects due to the scarcity of such instances in pretraining data. While prior efforts alleviate this issue by retrieving additional data or introducing stronger vision encoders, these methods are still computationally intensive during finetuning VLMs and don't fully exploit the original training data. In this paper, we introduce an efficient plug-and-play module that substantially improves VLMs' reasoning over rare objects by refining visual tokens and enriching input text prompts, without VLMs finetuning. Specifically, we propose to learn multi-modal class embeddings for rare objects by leveraging prior knowledge from vision foundation models and synonym-augmented text descriptions, compensating for limited training examples. These embeddings refine the visual tokens in VLMs through a lightweight attention-based enhancement module that improves fine-grained object details. In addition, we use the learned embeddings as object-aware detectors to generate informative hints, which are injected into the text prompts to help guide the VLM's attention toward relevant image regions. Experiments on two benchmarks show consistent and substantial gains for pretrained VLMs in rare object recognition and reasoning. Further analysis reveals how our method strengthens the VLM's ability to focus on and reason about rare objects. Code is available at https://github.com/XinHu98/seeing

Generalization remains the central challenge for interactive 3D scene generation. Existing learning-based approaches ground spatial understanding in limited scene dataset, restricting generalization to new layouts. We instead reprogram a pre-trained 3D instance generator to act as a scene-level learner via, replacing dataset-bounded supervision with model-centric spatial supervision. This reprogramming unlocks the generator's transferable spatial knowledge, enabling generalization to unseen layouts and novel object compositions. Remarkably, spatial reasoning still emerges even when the training scenes are randomly composed objects. This demonstrates that the generator's transferable scene prior provides a rich learning signal for inferring proximity, support, and symmetry from purely geometric cues. Replacing widely used canonical space, we instantiate this insight with a view-centric formulation of the scene space, yielding a fully feed-forward, generalizable scene generator that learns spatial relations directly from the instance model. Quantitative and qualitative results show that a 3D instance generator is an implicit spatial learner and reasoner, pointing toward foundation models for interactive 3D scene understanding and generation. The code is accessible at \href https://luling06.github.io/I-Scene-project/ the project page .

Multimodal Summarization (MMS) aims to generate concise textual summaries by understanding and integrating information across videos, transcripts, and images. However, existing approaches still suffer from three main challenges: (1) reliance on domain-specific supervision, (2) implicit fusion with weak cross-modal grounding, and (3) flat temporal modeling without event transitions. To address these issues, we introduce **CoE**, a training-free MMS framework that performs structured reasoning through a **Chain-of-Events** guided by a Hierarchical Event Graph (HEG). The HEG encodes textual semantics into an explicit event hierarchy that scaffolds cross-modal grounding and temporal reasoning. Guided by this structure, **CoE** localizes key visual cues, models event evolution and causal transitions, and refines outputs via lightweight style adaptation for domain alignment. Extensive experiments on eight diverse datasets demonstrate that **CoE** consistently outperforms state-of-the-art video CoT baselines, achieving average gains of **+3.04 ROUGE**, **+9.51 CIDEr**, and **+1.88 BERTScore**, highlighting its robustness, interpretability, and cross-domain generalization. Our code is available at https://github.com/youxiaoxing/CoE.

This paper provides a simplification on OpenVision's architecture and loss design for enhancing its training efficiency. Following the prior vision-language pretraining works CapPa and AIMv2, as well as modern multimodal designs like LLaVA, our changes are straightforward: we remove the text encoder (and therefore the contrastive loss), retaining only the captioning loss as a purely generative training signal. We name this new version OpenVision 2. The initial results are promising: despite this simplification, OpenVision 2 competitively matches the original model's performance on a broad set of multimodal benchmarks while substantially cutting both training time and memory consumption. For example, with ViT-L/14, it reduces training time by about 1.5x (from 83h to 57h), and memory usage by about 1.8x (from 24.5GB to 13.8GB, equivalently allowing the maximum batch size to grow from 2k to 8k). This superior training efficiency also allows us to scale far beyond the largest vision encoder used in OpenVision, reaching more than 1 billion parameters. We hold a strong belief that this lightweight, generative-only paradigm is compelling for future vision encoder development in multimodal foundation models.

The performance of Vision-Language Transformers drops sharply when an input modality (e.g., image) is missing, because the model is forced to make predictions using incomplete information. Existing missing-aware prompt methods help reduce this degradation, but they still rely on conventional prediction heads (e.g., a Fully-Connected layer) that compute class scores in the same way regardless of which modality is present or absent. We introduce Decoupled Prototype Learning (DPL), a new prediction head architecture that explicitly adjusts its decision process to the observed input modalities. For each class, DPL selects a set of prototypes specific to the current missing-modality cases (image-missing, text-missing, or mixed-missing). Each prototype is then decomposed into image-specific and text-specific components, enabling the head to make decisions that depend on the information actually present. This adaptive design allows DPL to handle inputs with missing modalities more effectively while remaining fully compatible with existing prompt-based frameworks. Extensive experiments on MM-IMDb, UPMC Food-101, and Hateful Memes demonstrate that DPL outperforms state-of-the-art approaches across all widely used multimodal image-text datasets and various missing cases.

Brain tumor MRI segmentation is essential for clinical diagnosis and treatment planning, enabling accurate lesion detection and radiotherapy target delineation. However, tumor lesions occupy only a small fraction of the volumetric space, resulting in severe spatial sparsity, while existing segmentation networks often overlook clinically observed spatial priors of tumor occurrence, leading to redundant feature computation over extensive background regions. To address this issue, we propose PGR-Net (Prior-Guided ROI Reasoning Network) - an explicit ROI-aware framework that incorporates a data-driven spatial prior set to capture the distribution and scale characteristics of tumor lesions, providing global guidance for more stable segmentation. Leveraging these priors, PGR-Net introduces a hierarchical Top-K ROI decision mechanism that progressively selects the most confident lesion candidate regions across encoder layers to improve localization precision. We further develop the WinGS-ROI (Windowed Gaussian-Spatial Decay ROI) module, which uses multi-window Gaussian templates with a spatial decay function to produce center-enhanced guidance maps, thus directing feature learning throughout the network. With these ROI features, a windowed RetNet backbone is adopted to enhance localization reliability. Experiments on BraTS 2019/2023 and MSD Task01 show that PGR-Net consistently outperforms existing approaches while using only 8.64M Params, achieving Dice scores of 89.02%, 91.82%, and 89.67% on the Whole Tumor region. Code is available at https://github.com/CNU-MedAI-Lab/PGR-Net.