CVPR 2026 论文列表

Multimodal Large Language Models (MLLMs) have shown strong potential in remote sensing (RS) through multi-task reasoning and cross-modal generalization.However, existing RS-MLLMs mainly rely on a single shared expert for all tasks, making it hard to produce reliable results. Meanwhile, the intrinsic redundancy and homogeneity of RS images bring substantial difficulties for both training and inference. These challenges directly conflict with the demands of remote sensing, which values task precision and trustworthy reasoning.To address these limitations, we propose GeoCoT, a manifold-driven mixture-of-experts (MoE) system with Chain-of-Thought (CoT) reasoning. GeoCoT introduces Mani-MoE, a sparse expert architecture grounded in local manifold mapping. It projects high-dimensional tokens onto low-rank subspaces adaptively to eliminate redundancy and uncover intrinsic structure, and then routes them through a sparse expert pathway, where gating decisions are guided by the manifold structure of the input.To optimize this architecture, we adopt a CoT-driven multi-stage training strategy. It leverages a cold-start phase for domain adaptation, followed by our RS Vision Group Relative Policy Optimization (RSV-GRPO) to systematically strengthen structured reasoning from global to objectives. Furthermore, we innovatively build *RS-CoT-20k* dataset for task-specific supervision.Extensive experiments on multi-task datasets demonstrate that GeoCoT outperforms prior approaches, achieving 5.27 \% higher average accuracy than the state-of-the-art method. Our code will be available.

Diffusion models have achieved remarkable progress in video generation, but their controllability remains a major limitation. Key scene factors such as layout, lighting, and camera trajectory are often entangled or only weakly modeled, restricting their applicability in domains like filmmaking and virtual production where explicit scene control is essential. We present LiVER, a diffusion-based framework for scene-controllable video generation. To achieve this, we introduce a novel framework that conditions video synthesis on explicit 3D scene properties, supported by a new large-scale dataset with dense annotations of object layout, lighting, and camera parameters. Our method disentangles these properties by rendering control signals from a unified 3D representation. We propose a lightweight conditioning module and a progressive training strategy to integrate these signals into a foundational video diffusion model, ensuring stable convergence and high fidelity. Our framework enables a wide range of applications, including image-to-video and video-to-video synthesis where the underlying 3D scene is fully editable. To further enhance usability, we develop a scene agent that automatically translates high-level user instructions into the required 3D control signals. Experiments show that LiVER achieves state-of-the-art photorealism and temporal consistency while enabling precise, disentangled control over scene factors, setting a new standard for controllable video generation.

Radar is a critical perception modality in autonomous driving systems due to its all-weather characteristics and ability to measure range and Doppler velocity. However, the sheer volume of high-dimensional raw radar data saturates the communication link to the computing engine (e.g., an NPU), which is often a low-bandwidth interface with data rate provisioned only for a few low-resolution range-Doppler frames. A generalized codec for utilizing high-dimensional radar data is notably absent, while existing image-domain approaches are unsuitable, as they typically operate at fixed compression ratios and fail to adapt to varying or adversarial conditions. In light of this, we propose radar data compression with adaptive feedback. It dynamically adjusts the compression ratio by performing gradient descent from the proxy gradient of detection confidence with respect to the compression rate. We employ a zeroth-order gradient approximation as it enables gradient computation even with non-differentiable core operations--pruning and quantization. This also avoids transmitting the gradient tensors over the band-limited link, which, if estimated, would be as large as the original radar data. In addition, we have found that radar feature maps are heavily concentrated on a few frequency components. Thus, we apply the discrete cosine transform to the radar data cubes and selectively prune out the coefficients effectively. We preserve the dynamic range of each radar patch through scaled quantization. Combining those techniques, our proposed online adaptive compression scheme achieves over 100x feature size reduction at minimal performance drop ( 1%p). We validate our results on the RADIal, CARRADA, and Radatron datasets.

Object counting in remote sensing imagery becomes challenging when visual cues are obscured by clouds, fog, shadows, or low-light conditions. Yet earth observation inherently provides complementary geo-modalities, including land use and map, which offer stable structural and contextual priors that remain available when appearance cues fail. In this paper, we introduce GROC, the first large-scale dataset Geo-guided Reasoning in Object Counting under adverse earth observation conditions. GROC contains 1.2 million point annotations over 14K images, each aligned with 3 modalities that preserve original geospatial information. We also provide a data engine to collect a large-scale object counting dataset with multiple geo-modalities, realistic degradations, and reliable annotations. We further present an counting agent that adaptively leverages geo-modalities to produce reliable estimates. Extensive experiments show that existing models struggle to "see" through adverse conditions, whereas geo-modalities improve robustness. GROC establishes the first benchmark that explicitly challenges models to see what they cannot see, charting a new direction for geo-guided amodal reasoning in earth observation.

Implicit Neural Representations (INRs) parameterize continuous signals via multilayer perceptrons (MLPs), enabling compact, resolution-independent modeling for tasks like image, audio, and 3D reconstruction. However, fitting high-resolution signals demands optimizing over millions of coordinates, incurring prohibitive computational costs. To address it, we propose NTK-Guided Implicit Neural Teaching (NINT), which accelerates training by dynamically selecting coordinates that maximize global functional updates. Leveraging the Neural Tangent Kernel (NTK), NINT scores examples by the norm of their NTK-augmented loss gradients, capturing both fitting errors and heterogeneous leverage (self-influence and cross-coordinate coupling). This dual consideration enables faster convergence compared to existing methods. Through extensive experiments, we demonstrate that NINT significantly reduces training time by nearly half while maintaining or improving representation quality, establishing state-of-the-art acceleration among recent sampling-based strategies.

Deploying pretrained visual models in real-world environments often suffers from significant performance degradation due to the diversity of testing scenarios. Continuous adaptation of learning models on edge devices via unlabeled data collected from the target domain is highly effective for boosting generalization capability. However, gradient-backpropagation-based optimization of the massive parameters in deep neural networks is vastly more time-consuming than forward inference, rendering online learning infeasible on low-power edge devices. To address this critical challenge, we propose a lightweight gradient-free forward-memorizing framework, namely MemFlow, which leverages a frozen backbone and enables efficient fine-tuning of the mapping between features and predictions. Specifically, MemFlow employs randomly connected neurons to memorize feature-label associations; within the network, spiking signals are propagated, and predictions are generated by associating neuron-stored memories according to their confidence levels. More notably, MemFlow supports reinforced memorization of feature mappings using unlabeled data, thereby enabling rapid adaptation to new domains. Extensive experiments on four real-world cross-domain datasets demonstrate that MemFlow achieves performance improvements of up to 10% while consuming less than 1% of the computational time required by traditional domain adaptation methods.

One-shot Federated Learning (OSFL) has emerged as a promising paradigm to mitigate the high communication overhead of traditional federated learning. However, its effectiveness is often hindered by data heterogeneity across clients. While recent methods leverage pre-trained diffusion models to generate data for OSFL, they often struggle with some practical limitations, including a lack of semantic fidelity in capturing the fine-grained characteristics of local data, and insufficient diversity in the generated data, which collectively degrade the performance of the global model. To address these challenges, we propose Espresso, a novel framework that enhances both the fidelity and diversity of synthetic data in OSFL. Espresso consists of two main components: (1) Fine-Grained Condition Learning, which learns fine-grained conditional embeddings to improve semantic fidelity and diversity by modeling intra-category patterns, and (2) Semantics-Preserving Sampling, which diversifies the generated data by modeling the distribution of latent noises and applying a self-reflection sampling strategy. Extensive experiments on benchmark datasets demonstrate that Espresso can improve the semantic fidelity and diversity of the synthetic data, leading to an enhancement in the performance of the global model compared to state-of-the-art OSFL methods under data heterogeneity.

Multimodal Large Language Models (MLLMs) show strong potential for interpreting and interacting with complex, pixel-rich Graphical User Interface (GUI) environments. However, building agents that are both efficient for high-level tasks and precise for fine-grained interactions remains challenging. GUI agents must perform routine actions efficiently while also handling tasks that demand exact visual grounding, yet existing approaches struggle when accuracy depends on identifying specific interface elements. These MLLMs also remain large and cannot adapt their reasoning depth to the task at hand. In this work, we introduce iSHIFT: Implicit Slow-fast Hybrid Inference with Flexible Tokens, a lightweight agent that integrates latent thinking (implicit chain-of-thought) with a perception control module. iSHIFT enables an MLLM to switch between a slow mode, which leverages detailed visual grounding for high precision and a fast mode that uses global cues for efficiency. Special perception tokens guide attention to relevant screen regions, allowing the model to decide both how to reason and where to focus. Despite its compact 2.5B size, iSHIFT matches state-of-the-art performance on multiple benchmark datasets.

Large vision-language models (LVLMs) have achieved impressive performance across a variety of multimodal tasks, yet remain vulnerable to targeted adversarial attacks, particularly in black-box settings. In this paper, we propose VCP-Attack, a transferable targeted attack framework that combines structured contrastive supervision with subspace-guided perturbation optimization. Specifically, we employ a dynamic PCA-based projection to constrain perturbations within semantically meaningful low-dimensional subspaces, and design a multi-sample contrastive loss to align adversarial features with target semantics while pushing them away from the source semantics. Extensive experiments on seven open-source and three proprietary LVLMs--including GPT-4o, Claude, and Gemini--show that VCP-Attack achieves state-of-the-art performance in black-box targeted attacks. Under a fixed perturbation budget (epsilon = 16/255), our method achieves an average attack success rate (ASR) of 94.2% on open-source models and 83.1% on proprietary models, surpassing the strongest baselines by 23.3% and 16.8%, respectively. Notably, VCP-Attack achieves a 95.6% ASR on GPT-4o. Comprehensive ablation studies and visualizations further validate the effectiveness of the dynamic subspace projection and semantic contrastive supervision. While evaluated on image captioning, our approach is model-agnostic and exhibits strong potential for broader applications to black-box adversarial settings in vision-language tasks.

When video reasoning requires external knowledge, many systems with large multimodal models (LMMs) adopt retrieval augmentation to supply the missing context. Appending textual or multi-clip evidence, however, forces heterogeneous signals into a single attention space. We observe diluted attention and higher cognitive load even on non-long videos. The bottleneck is not only what to retrieve but how to represent and fuse external knowledge with the video backbone.We present **Graph-to-Frame RAG (G2F-RAG)**, a training free and auditable paradigm that delivers knowledge in the visual space. On the offline stage, an agent builds a problem-agnostic video knowledge graph that integrates entities, events, spatial relations, and linked world knowledge. On the online stage, a hierarchical multi-agent controller decides whether external knowledge is needed, retrieves a minimal sufficient subgraph, and renders it as a single reasoning frame appended to the video. LMMs then perform joint reasoning in a unified visual domain. This design reduces cognitive load and leaves an explicit, inspectable evidence trail.G2F-RAG is plug-and-play across backbones and scales. It yields consistent gains on diverse public benchmarks, with larger improvements in knowledge-intensive settings. Ablations further confirm that knowledge representation and delivery matter. G2F-RAG reframes retrieval as visual space knowledge fusion for robust and interpretable video reasoning.

Current referring video segmentation methods typically operate in an offline manner, where sparse frames are first selected for image-level referring segmentation, and the resulting masks are then propagated across the video. Although video sampling captures global context, its isolated processing steps not only complicate optimization but also restrict applicability to real-world streaming scenarios. In this paper, we propose a simple but efficient MLLM-based framework StreamingRVOS, which can extend image-level segmentation to video-level via a streaming pipeline without introducing extra parameters. Specifically, we employ a Semantic Embedding Recycling (SER) method to propagate temporal context across frames, enabling the model to perceive semantic representation in the video. Then, we propose an Online Mask Consistency Perception (OMCP) strategy to adaptively invoke the MLLM to re-perceive the current scene and regenerate the semantic embedding. We conduct extensive experiments on multiple downstream datasets to prove the effectiveness of StreamingRVOS. Compared to previous methods, our method achieves excellent performance in referring video segmentation (1B variant improves upon Sa2VA by 19.2 on the MeViS dataset), while operating at an average speed of 7 FPS under streaming inference on 1 xA800 GPU.

Trajectory prediction in autonomous driving has traditionally been studied from a model-centric perspective. However, existing datasets exhibit a strong long-tail distribution in scenario density, where common low-density cases dominate and safety-critical high-density cases are severely underrepresented. This imbalance limits model robustness and hides failure modes when standard evaluations average errors across all scenarios. We revisit trajectory prediction from a data-centric perspective and present Den-TP, a framework for density-aware dataset curation and evaluation. Den-TP first partitions data into density-conditioned regions using agent count as a dataset-agnostic proxy for interaction complexity. It then applies a gradient-based submodular selection objective to choose representative samples within each region while explicitly rebalancing across densities. The resulting subset reduces the dataset size by 50% yet preserves overall performance and significantly improves robustness in high-density scenarios. We further introduce density-conditioned evaluation protocols that reveal long-tail failure modes overlooked by conventional metrics. Experiments on Argoverse 1 and 2 with state-of-the-art models show that robust trajectory prediction depends not only on data scale, but also on balancing scenario density.

Single-view RGB model-based object pose estimation methods achieve strong generalization but are fundamentally limited by depth ambiguity, clutter, and occlusions. Multi-view pose estimation methods have the potential to solve these issues, but existing works rely on precise single-view pose estimates or lack generalization to unseen objects. We address these challenges via the following three contributions. First, we introduce AlignPose, a 6D object pose estimation method that aggregates information from multiple extrinsically calibrated RGB views and does not require any object-specific training or symmetry annotation. Second, the key component of this approach is a new multi-view feature-metric refinement specifically designed for object pose. It optimizes a single, consistent world-frame object pose by minimizing the feature discrepancy between on-the-fly rendered object features and observed image features across all views simultaneously. Third, we report extensive experiments on six datasets (YCB-V, T-LESS, HouseCat6D, ITODD-MV, IPD, XYZ-IBD) using the BOP benchmark evaluation and show that AlignPose outperforms other published methods, especially on challenging industrial datasets where multiple views are readily available in practice.

Efficient streaming video generation is critical for simulating interactive and dynamic worlds. Existing methods distill few-step video diffusion models with sliding window attention, using initial frames as sink tokens to maintain attention performance and reduce error accumulation. However, video frames become overly dependent on these static tokens, resulting in copied initial frames and diminished motion dynamics. To address this, we introduce Reward Forcing, a novel framework with two key designs. First, we propose EMA-Sink, which maintains fixed-size tokens initialized from initial frames and continuously updated by fusing evicted tokens via exponential moving average as they exit the sliding window. Without additional computation cost, EMA-Sink tokens capture both long-term context and recent dynamics, preventing initial frame copying while maintaining long-horizon consistency. Second, to better distill motion dynamics from teacher models, we propose a novel Rewarded Distribution Matching Distillation (Re-DMD). Vanilla distribution matching treats every training sample equally, limiting the model's ability to prioritize dynamic content. Instead, Re-DMD biases the model's output distribution toward high-reward regions by prioritizing samples with greater dynamics rated by a vision-language model. Re-DMD significantly enhances motion quality while preserving data fidelity. We include both quantitative and qualitative experiments to show that Reward Forcing achieves state-of-the-art performance on standard benchmarks while enabling high-quality streaming video generation at 23.1 FPS on a single H100 GPU.

We present PaNDaS, a novel deep learning framework for Partial Non-Rigid Deformations and interpolations of Surfaces (PaNDaS). PaNDaS learns a per-face feature field on the source mesh and fuses it with a global encoding of the target. A deformation generator predicts a Jacobian field and recovers a smooth displacement, enabling precise regional control, pose mixing, and transferable local edits. Unlike previous approaches, our method can restrict the deformations to specific parts of the shape in a versatile way. Across various human body part datasets, PaNDaS achieves state-of-the-art interpolation accuracy and stronger locality than methods based on global shape codes or handles, while remaining robust to remeshing. We demonstrate several localized shape manipulation tasks and show that our method can generate new shapes by combining different input deformations.

Understanding the global structure of neural network loss landscapes is important for gaining insight into model merging, hyperparameter selection, generalization, and the relationships between distinct solutions. Visualizing the global structure of loss landscapes is very challenging because of the high dimensionality of the parameter space of neural networks. Prior work has primarily focused on visualizing the loss landscape around one single basin, missing how different minima or solutions relate to each other. We introduce Globscope, a framework that provides a global view of the loss landscape across multiple solutions or basins. Globscope learns a low-dimensional non-linear manifold of model parameters using an autoencoder, enabling both latent-space visualization and reconstruction of full model weights. Then it summarizes the relations and connections among minima on this manifold through topological data analysis. Our framework produces continuous, interpretable visualizations that reveal global connectivity patterns in the landscape. We compare Globscope with kernel-based methods and demonstrate its superior performance in preserving the global structure across diverse solutions. We further show how Globscope can be used to analyze two applications: revealing global low-loss solution pathways between distinct solutions using mode connectivity algorithms, and visualizing permutation symmetries of different solutions using re-basin approaches.

Although industrial inspection systems should be capable of recognizing unprecedented defects, most existing approaches operate under a closed-set assumption, which prevents them from detecting novel anomalies. While visual prompting offers a scalable alternative for industrial inspection, existing methods often suffer from prompt embedding collapse due to high intra-class variance and subtle inter-class differences. To resolve this, we propose UniSpector, which shifts the focus from naive prompt-to-region matching to the principled design of a semantically structured and transferable prompt topology. UniSpector employs the Spatial-Spectral Prompt Encoder to extract orientation-invariant, fine-grained representations; these serve as a solid basis for the Contrastive Prompt Encoder to explicitly regularize the prompt space into a semantically organized angular manifold. Additionally, Prompt-guided Query Selection generates adaptive object queries aligned with the prompt. We introduce Inspect Anything, the first benchmark for visual-prompt-based open-set defect localization, where UniSpector significantly outperforms baselines by at least 19.7% and 15.8% in AP50b and AP50m, respectively. These results show that our method enable a scalable, retraining-free inspection paradigm for continuously evolving industrial environments, while offering critical insights into the design of generic visual prompting.

Always-on sensing is essential for next-generation edge/wearable AI systems, yet continuous high-fidelity RGB video capture remains prohibitively expensive for resource-constrained mobile and edge platforms. We present a new paradigm for efficient streaming video understanding: grayscale-always, color-on-demand. Through preliminary studies, we discover that color is not always necessary. Sparse RGB frames suffice for comparable performance when temporal structure is preserved via continuous grayscale streams. Building on this insight, we propose ColorTrigger, an online training-free trigger that selectively activates color capture based on windowed grayscale affinity analysis. Designed for real-time edge deployment, ColorTrigger uses lightweight quadratic programming to detect chromatic redundancy causally, coupled with credit-budgeted control and dynamic token routing to jointly reduce sensing and inference costs. On streaming video understanding benchmarks, ColorTrigger achieves 91.6% of full-color baseline performance while using only 8.1% RGB frames, demonstrating substantial color redundancy in natural videos and enabling practical always-on video sensing on resource-constrained devices.

Post-training alignment of video generation models with human preferences is a critical goal. Developing effective Reward Models (RMs) for this process faces significant methodological hurdles. Current data collection paradigms, reliant on in-prompt pairwise annotations, suffer from labeling noise. Concurrently, the architectural design of VLM-based RMs, particularly their output mechanisms, remains underexplored. Furthermore, RM is susceptible to reward hacking in post-training. To mitigate these limitations, we propose SoliReward, a systematic framework for video RM training. Our framework first sources high-quality, cost-efficient data via single-item binary annotations, then constructs preference pairs using a cross-prompt pairing strategy. Architecturally, we employ a Hierarchical Progressive Query Attention mechanism to enhance feature aggregation. Finally, we introduce a modified BT loss that explicitly accommodates win-tie scenarios. This approach regularizes the RM's score distribution for positive samples, providing more nuanced preference signals to alleviate over-focus on a small number of top-scoring samples. Our approach is validated on benchmarks evaluating physical plausibility, subject deformity, and semantic alignment, demonstrating improvements in direct RM evaluation metrics and in the efficacy of post-training on video generation models. Code and benchmark are available at https://github.com/lian700/SoliReward.