CVPR 2026 论文列表

Multimodal LLM agents operating in complex game environments must continually reuse past experience to solve new tasks efficiently. In this work, we propose Echo, a transfer-oriented memory framework that enables agents to derive actionable knowledge from prior interactions rather than treating memory as a passive repository of static records. To make transfer explicit, Echo decomposes reusable knowledge into five dimensions: structure, attribute, process, function, and interaction. This formulation allows the agent to identify recurring patterns shared across different tasks and infer what prior experience remains applicable in new situations. Building on this formulation, Echo leverages In-Context Analogy Learning (ICAL) to retrieve relevant experiences and adapt them to unseen tasks through contextual examples. Experiments in Minecraft show that, under a from-scratch learning setting, Echo achieves a 1.3x-1.7x speed-up on object-unlocking tasks. Moreover, Echo exhibits a burst-like chain-unlocking phenomenon, rapidly unlocking multiple similar items within a short time interval after acquiring transferable experience. These results suggest that experience transfer is a promising direction for improving the efficiency and adaptability of multimodal LLM agents in complex interactive environments.

Capturing scenes with both high dynamic range (HDR) and high-speed motion remains challenging for conventional cameras. Existing alternating-exposure approaches exacerbate temporal resolution loss, making them unsuitable for high-speed scenes. Consequently, current solutions typically compromise either spatial resolution through fixed spatial-varying attenuation levels or employ multi-sensor configurations to maintain temporal resolution. In this paper, we leverage an ultra-high speed spike camera to enable spatial and temporal attenuation of incident light, thereby reconstructing high frame rate (HFR) and HDR video with a single sensor. We achieve this by placing a rapidly rotating spoke-pattern neutral density (SpokeND) filter in front of the spike camera, enabling each pixel to periodically capture multi-attenuated spikes while maintaining full spatial resolution. Building on these multi-attenuated spikes, we propose ReST-Net, which comprises the ReGain and ReFine modules. The ReGain module reconstructs spatially consistent frames by learning to recover relative gain from the multi-attenuated spikes, and the ReFine module removes temporal fluctuations to produce temporally consistent HDR videos. Extensive experiments on synthetic and real-world data demonstrate that our method can reconstruct HDR video at up to 2000 FPS.

Recent advances in multimodal large language models (MLLMs) have substantially expanded the capabilities of multimodal retrieval, enabling systems to align and retrieve information across visual and textual modalities. Yet, existing benchmarks largely focus on coarse-grained or single-condition alignment, overlooking real-world scenarios where user queries specify multiple interdependent constraints across modalities. To bridge this gap, we introduce MCMR (Multi-Conditional Multimodal Retrieval): a large-scale benchmark designed to evaluate fine-grained, multi-condition cross-modal retrieval under natural-language queries. MCMR spans five product domains: upper and bottom clothing, jewelry, shoes, and furniture. It also preserves rich long-form metadata essential for compositional matching. Each query integrates complementary visual and textual attributes, requiring models to jointly satisfy all specified conditions for relevance. We benchmark a diverse suite of MLLM-based multimodal retrievers and vision-language rerankers to assess their condition-aware reasoning abilities. Experimental results reveal: (i) distinct modality asymmetries across models; (ii) visual cues dominate early-rank precision, while textual metadata stabilizes long-tail ordering; and (iii) MLLM-based pointwise rerankers markedly improve fine-grained matching by explicitly verifying query-candidate consistency. Overall, MCMR establishes a challenging and diagnostic benchmark for advancing multimodal retrieval toward compositional, constraint-aware, and interpretable understanding. Our code and dataset is available at https://github.com/EIT-NLP/MCMR

Learning joint representations across multiple modalities remains a central challenge in multimodal machine learning. Prevailing approaches predominantly operate in pairwise settings, aligning two modalities at a time. While some recent methods aim to capture higher-order interactions among multiple modalities, they often overlook or insufficiently preserve pairwise relationships, limiting their effectiveness on single-modality tasks. In this work, we introduce Contrastive Fusion (ConFu), a framework that jointly embeds both individual modalities and their fused combinations into a unified representation space, where modalities and their fused counterparts are aligned. ConFu extends traditional pairwise contrastive objectives with an additional fused-modality contrastive term, encouraging the joint embedding of modality pairs with a third modality. This formulation enables ConFu to capture higher-order dependencies, such as XOR-like relationships, that cannot be recovered through pairwise alignment alone, while still maintaining strong pairwise correspondence. We evaluate ConFu on synthetic and real-world multimodal benchmarks, assessing its ability to exploit cross-modal complementarity, capture higher-order dependencies, and scale with increasing multimodal complexity. Across these settings, ConFu demonstrates competitive performance on retrieval and classification tasks, while supporting unified one-to-one and two-to-one retrieval within a single contrastive framework. We release our code and dataset at https://github.com/estafons/confu.

Vision-language models (VLM) excel at general understanding yet remain weak at dynamic spatial reasoning (DSR), i.e., reasoning about the evolvement of object geometry and relationship in 3D space over time, largely due to the scarcity of scalable 4D-aware training resources. To bridge this gap across aspects of dataset, benchmark and model, we introduce DSR Suite. First, we propose an automated pipeline that generates multiple-choice question-answer pairs from in-the-wild videos for DSR. By leveraging modern vision foundation models, the pipeline extracts rich geometric and motion information, including camera poses, local point clouds, object masks, orientations, and 3D trajectories. These geometric cues enable the construction of DSR-Train for learning and further human-refined DSR-Bench for evaluation. Compared with previous works, our data emphasize (i) in-the-wild video sources, (ii) object- and scene-level 3D requirements, (iii) viewpoint transformations, (iv) multi-object interactions, and (v) fine-grained, procedural answers. Beyond data, we propose a lightweight Geometry Selection Module (GSM) to seamlessly integrate geometric priors into VLMs, which condenses question semantics and extracts question-relevant knowledge from pretrained 4D reconstruction priors into a compact set of geometry tokens. This targeted extraction avoids overwhelming the model with irrelevant knowledge. Experiments show that integrating DSR-Train and GSM into Qwen2.5-VL-7B significantly enhances its dynamic spatial reasoning capability, while maintaining accuracy on general video understanding benchmarks.

People can view the same image differently: they focus on different regions, objects, and details in varying orders and describe them in distinct linguistic styles. This leads to substantial variability in image descriptions. However, existing models for personalized image description focus on linguistic style alone, with no prior work leveraging individual viewing patterns. We address this gap by explicitly modeling personalized viewing behavior as a core factor in description generation. Our method, DEPER (DEscription-PERception persona encoder), learns a subject embedding that captures both linguistic style and viewing behavior, guided by an auxiliary attention-prediction task. A lightweight adapter aligns these embeddings with a frozen vision-language model, enabling few-shot personalization without retraining. Across four datasets spanning diverse viewing tasks and both short and detailed descriptions, DEPER achieves a 24% average improvement, showing that modeling personalized attention produces more human-aligned and high-quality descriptions. We posit that understanding how people see helps predict what they say; modeling human diversity in perception can improve both performance and human alignment in multi-modal systems. Code is available at: https://github.com/cvlab-stonybrook/Personalized-Image-Description

Referring video object segmentation (RVOS) aims to identify, track and segment the objects in a video based on language descriptions, which has received great attention in recent years. However, existing datasets remain focus on short video clips within several seconds, with salient objects visible in most frames. To advance the task towards more practical scenarios, we introduce Long-RVOS, a large-scale benchmark for long-term referring video object segmentation. Long-RVOS contains 2,000+ videos of an average duration exceeding 60 seconds, covering a variety of objects that undergo occlusion, disappearance-reappearance and shot changing. The objects are manually annotated with three different types of descriptions to individually evaluate the understanding of static attributes, motion patterns and spatiotemporal relationships. Moreover, we introduce two new metrics to assess the temporal and spatiotemporal consistency. We benchmark 7 state-of-the-art methods on Long-RVOS to show that current approaches struggle severely with the long-video challenges. We further propose ReferMo, a promising baseline method that integrates motion information to expand the temporal receptive field, and employs a local-to-global architecture to capture both short-term dynamics and long-term dependencies. We hope that Long-RVOS and our baseline can drive future RVOS research towards more realistic and long-form videos. Our dataset and code is available at https://isee-laboratory.github.io/Long-RVOS.

As an emerging multi-granularity clustering paradigm, granular-ball computing (GBC) hierarchically represents samples through granular-balls (GBs) to capture compact, multi-scale features. Nevertheless, its effective application to clustering-based segmentation methods (CSMs) remains challenging due to two key issues: representing intrinsic uncertainties and defining a justifiable, semantics-aware quality criterion. To address them, the first segmentation framework based on GBC (SegGBC) is proposed to alleviate the single-granularity limitation of existing CSMs. Concretely, we leverage intuitionistic fuzzy sets (IFS) to explicitly quantify image uncertainty: membership and non-membership encode evidence, and the IFS hesitation degree models residual ambiguity. In addition, a semantic compactness metric criterion (SCM_GB) is designed to characterize semantic information by considering the "stable region" in conjunction with the overall density of GBs. The proposal of "stable region" ensures robust semantics concurrently with high computational efficiency. Extensive experiments demonstrate that the proposed SegGBC achieves promising performance for segmentation. The proposed segmentation GB representation is a plug-and-play front-end, significantly boosting the performance of CSMs by >+3.25% SA and >+3.92% mIoU on standard image and COCO benchmarks. Code is available at supplementary material.

Zero-shot anomaly detection (ZSAD) aims to identify unseen anomalies without abnormal supervision, which is essential for open-world scenarios. Recent vision-language models such as CLIP enable anomaly reasoning through shared visual-textual embeddings, but existing methods often rely on coarse prompt fusion, leading to unstable alignment and imprecise localization under domain shifts. To address this issue, we propose the Semantic Graviton Network (SGNet), a physics-inspired framework that models multimodal alignment as an adaptive potential field. We introduce semantic gravitons, learnable dynamic mediators that bridge visual and textual modalities by establishing localized semantic equilibria through attraction and equilibrium forces. A graviton interaction network alternates text-to-graviton and vision-to-graviton coupling to progressively refine multimodal correspondence, while an energy-based potential regularization further stabilizes the interaction process. Extensive experiments on ten industrial and medical benchmarks show that SGNet achieves state-of-the-art performance for zero-shot anomaly detection.

Geo-spatial analysis of our world benefits from a multimodal approach, as every single geographic location can be described in numerous ways (images from various viewpoints, textual descriptions, geographic coordinates, etc.). Current benchmarks have limited coverage across modalities, leading to specialized models that perform well in their respective domains, but do not fully take advantage of other geo-spatial modalities. We introduce the Multi-Modal Landmark dataset (MMLandmarks), a benchmark composed of four modalities: 197k high-resolution aerial images, 329k ground-view images, textual information, and geographic coordinates for 18,557 distinct landmarks in the United States. The MMLandmarks dataset has a one-to-one landmark level correspondence across every modality, which enables training and benchmarking models for various geo-spatial tasks, including cross-view Ground-to-Satellite retrieval, ground and satellite geolocalization, Text-to-Image, and Text-to-GPS retrieval. We show that current specialized and off-the-shelf foundation models cannot be trivially used to solve this variety of geo-spatial tasks, illustrating a gap where multimodal datasets lead to broader geo-spatial understanding. We employ a simple CLIP-inspired baseline that reflects versatility and broad generalization when trained with MMLandmarks.

Despite progress, Vision-Language-Action models (VLAs) are limited by a scarcity of large-scale, diverse robot data. While human manipulation videos offer a rich alternative, existing methods are forced to choose between small, precisely-labeled datasets and vast in-the-wild footage with unreliable hand tracking labels. We present JALA, a pretraining framework that learns Jointly-Aligned Latent Actions. JALA bypasses full visual dynamic reconstruction, instead learns a predictive action embedding aligned with both inverse dynamics and real actions. This yields a transition-aware, behavior-centric latent space for learning from heterogeneous human data. We scale this approach with UniHand-Mix, a 7.5M video corpus (>2,000 hours) blending laboratory and in-the-wild footage. Experiments demonstrate that JALA generates more realistic hand motions in both controlled and unconstrained scenarios, significantly improving downstream robot manipulation performance in both simulation and real-world tasks. These results indicate that jointly-aligned latent actions offer a scalable pathway for VLA pretraining from human data.

Long streaming video QA remains challenging due to growing visual tokens and limited reasoning length of large language models (LLMs). KV-caching stores the Key-Value (KV) of the historical tokens via LLM prefill and enables more efficient streaming QA. However, existing methods cache every one or two frames, causing redundant memory usage and losing fine-grained spatial details within frame or temporal contexts across frames. This paper proposes MuKV, a method that features a multi-grained KV cache compression module and a semi-hierarchical retrieval approach to improve both efficiency and accuracy for long streaming VideoQA. For the offline KV cache, MuKV extracts visual representations at patch-, frame-, and segment-levels. The multiple levels of granularity preserve both local cues and global temporal context, while maintaining efficiency with a dual signal token compression mechanism guided by self-attention and frequency. For online QA, MuKV designs a semi-hierarchical retrieval method to retrieve relevant KV caches for answer generation. Experiments on long-streaming VideoQA benchmarks show that MuKV significantly improves answer accuracy, without sacrificing memory and online QA efficiency. Moreover, our compression mechanism alone brings consistent benefits across answer accuracy, memory, and QA efficiency over baselines, showcasing highly effective contribution.

HOI detection has long been dominated by task-specific models, sometimes with early vision-language backbones such as CLIP. With the rise of large generative VLMs, a key question is whether standalone VLMs can perform HOI detection competitively against specialized HOI methods. Existing benchmarks such as HICO-DET require exact label matching under incomplete annotations, so any unmatched prediction is marked wrong. This unfairly penalizes valid outputs, especially from less constrained VLMs, and makes cross-paradigm comparison unreliable. To address this limitation, we introduce CrossHOI-Bench, a multiple-choice HOI benchmark with explicit positives and curated negatives, enabling unified and reliable evaluation of both VLMs and HOI-specific models. We further focus on challenging scenarios, such as multi-person scenes and fine-grained interaction distinctions, which are crucial for revealing real differences between the two paradigms. Experiments show that large VLMs achieve competitive, sometimes superior, zero-shot performance, yet they struggle with multiple concurrent actions and with correctly assigning interactions to the target person. Conversely, HOI-specific methods remain weaker in general HOI reasoning but demonstrate stronger multi-action recognition and more reliable identification of which person performs which action. These findings expose complementary strengths and weaknesses of VLMs and HOI-specific methods, which existing benchmarks fail to reveal due to incorrect penalization.

Video instance segmentation (VIS) for low-light content remains highly challenging for both humans and machines alike, due to noise, blur and other adverse conditions. The lack of large-scale annotated datasets and the limitations of current synthetic pipelines, particularly in modeling temporal degradations, further hinder progress. Moreover, existing VIS methods are not robust to the degradations found in low-light videos and, consequently, perform poorly even after finetuning. In this paper, we introduce ELVIS (Enhance Low-Light for Video Instance Segmentation), a framework that enables domain adaptation of state-of-the-art VIS models to low-light scenarios. ELVIS is comprised of an unsupervised synthetic low-light video pipeline that models both spatial and temporal degradations, a calibration-free degradation profile estimation network (VDP-Net) and an enhancement decoder head that disentangles degradations from content features. ELVIS improves performances by up to +3.7AP on the synthetic low-light YouTube-VIS 2019 dataset and beats two-stage baselines by at least +2.8AP on real low-light videos. Code and dataset available at: https://joannelin168.github.io/research/ELVIS

Research on the intelligent interpretation of all-weather, all-time Synthetic Aperture Radar (SAR) is crucial for advancing remote sensing applications. In recent years, although Visual Language Models (VLMs) have demonstrated strong open-world understanding capabilities on RGB images, their performance is severely limited when directly applied to the SAR field due to the complexity of the imaging mechanism, sensitivity to scattering features, and the scarcity of high-quality text corpora. To systematically address this issue, we constructed the inaugural SAR Image-Text-AlphaEarth feature triplet dataset and developed FUSAR-GPT, a VLM specifically for SAR. FUSAR-GPT innovatively introduces a geospatial baseline model as a 'world knowledge' prior and embeds multi-source remote-sensing temporal features into the model's visual backbone via 'spatiotemporal anchors', enabling dynamic compensation for the sparse representation of targets in SAR images. Furthermore, we designed a two-stage SFT strategy to decouple the knowledge injection and task execution of large models. The spatiotemporal feature embedding and the two-stage decoupling paradigm enable FUSAR-GPT to achieve state-of-the-art performance across several typical remote sensing visual-language benchmark tests, significantly outperforming mainstream baseline models by over 10%.

Intrinsic image decomposition (IID) of outdoor scenes is crucial for relighting, editing, and understanding large-scale environments, but progress has been limited by the lack of real-world datasets with reliable albedo and shading supervision. We introduce Olbedo, a large-scale aerial dataset for outdoor albedo--shading decomposition in the wild. Olbedo contains 5,664 UAV images captured across four landscape types, multiple years, and diverse illumination conditions. Each view is accompanied by multi-view consistent albedo and shading maps, metric depth, surface normals, sun and sky shading components, camera poses, and, for recent flights, measured HDR sky domes. These annotations are derived from an inverse-rendering refinement pipeline over multi-view stereo reconstructions and calibrated sky illumination, together with per-pixel confidence masks. We demonstrate that Olbedo enables state-of-the-art IID models, including both CNN-based and diffusion-based architectures originally trained on synthetic indoor data, to generalize to real outdoor imagery: fine-tuning on Olbedo significantly improves single-view outdoor albedo prediction on the MatrixCity benchmark. We further illustrate applications of Olbedo-trained models to multi-view consistent relighting of 3D assets, material editing, and scene change analysis for urban digital twins. We release the dataset, baseline models, and an evaluation protocol to support future research in outdoor intrinsic decomposition and illumination-aware aerial vision.

In this paper, we explore an important yet underexplored task in robot manipulation: cycle-based manipulation, where robots need to perform cyclic or repetitive actions with an expected terminal time. These tasks are crucial in daily life, such as shaking a bottle or knocking a nail. However, few prior works have explored this task, leading to two main challenges: 1) the imitation methods often fail to complete these tasks within the expected terminal time due to the ineffective utilization of history; 2) the absence of a benchmark with sufficient data and automatic evaluation tools hinders development of effective solutions in this area. To address these challenges, we first propose the CycleManip framework to achieve cycle-based task manipulation in an end-to-end imitation manner without requiring any extra models, hierarchical structure or significant computational overhead. The core insight is to enhance effective history perception by a cost-aware sampling strategy and to improve historical understanding by multi-task learning. Second, we introduce a cycle-based task manipulation benchmark, which provides diverse cycle-based tasks, and an automatic evaluation method. Extensive experiments conducted in both simulation and real-world settings demonstrate that our method achieves high success rates in cycle-based task manipulation. The results further show strong adaptability performance in general manipulation, and the plug-and-play ability on imitation policies such as Vision-Language-Action (VLA) models. Moreover, the results show that our approach can be applied across diverse robotic platforms, including bi-arm grippers, dexterous hands, and humanoid robots.

Density ratio estimation is a core concept in statistical machine learning because it provides a unified mechanism for tasks such as importance weighting, divergence estimation, and likelihood-free inference, but its potential in vision and language models has not been fully explored. Modern vision-language encoders such as CLIP and SigLIP are trained with contrastive objectives that implicitly optimize log density ratios between joint and marginal image-text distributions, which implicitly learn similarity scores proportional to log density ratios. However, prior work has largely focused on their embedding utility, and the density-ratio structure induced by contrastive learning has not been systematically examined or exploited in multimodal applications. To address this gap, we reinterpret CLIP-style models as pretrained and general-purpose density ratio estimators and show that this perspective enables new algorithmic capabilities. We present a unified explanation of how contrastive objectives estimate density ratios and propose two practical applications: Importance Weight Learning and KL divergence estimation. Our Importance Weight Learning method requires only a single additional prompt and improves F1 scores by up to 7 points. We further show that CLIP-based density ratios support estimation of KL divergences that quantify how conditioning on an image or text alters the distribution of the other modality. Through qualitative examples and an N-gram analysis of captions, we find that these divergences capture semantic diversity and mode structure in multimodal data. Leveraging this property, we introduce a simple KL-guided data curation method that achieves performance competitive with LAION2B filtering. Our code is available at https://github.com/fumiyauchiyama/CLIP_Density_Ratio .

Medical image segmentation remains challenging due to limited annotations for training, ambiguous anatomical features, and domain shifts. While vision-language models such as CLIP offer strong cross-modal representations, their potential for dense, text-guided medical image segmentation remains underexplored. We present MedCLIPSeg, a novel framework that adapts CLIP for robust, data-efficient, and uncertainty-aware medical image segmentation. Our approach leverages patch-level CLIP embeddings through probabilistic cross-modal attention, enabling bidirectional interaction between image and text tokens and explicit modeling of predictive uncertainty. Together with a soft patch-level contrastive loss that encourages more nuanced semantic learning across diverse textual prompts, MedCLIPSeg effectively improves data efficiency and domain generalizability. Extensive experiments across 16 datasets spanning five imaging modalities and six organs demonstrate that MedCLIPSeg outperforms prior methods in accuracy, efficiency, and robustness, while providing interpretable uncertainty maps that highlight local reliability of segmentation results. This work demonstrates the potential of probabilistic vision-language modeling for text-driven medical image segmentation.