CVPR 2026 论文列表

Zero-Shot Temporal Action Localization (ZS-TAL) aims to classify and localize actions in untrimmed videos that are unseen during training. Existing training-based ZS-TAL methods typically rely on fine-tuning models on large-scale annotated training data. This can be impractical in real-world applications and damage its generalization. As a result, Training-Free ZS-TAL has gained attention, which directly leveraging Vision-Language Models (VLM) enables action localization without any additional training. However, current techniques perform test-time adaptation independently on each video, neglecting the potential benefit of accumulating knowledge from historical test videos. To address this, we propose a learnable lookup table (LLT) framework. During testing, we continuously update lookup table by incorporating high-confidence, diverse lookup candidates to construct action-positive lookup item. Additionally, we introduce a learnable residual module to adapt the corresponding lookup item to the current video context features. Finally, we employ refined activation scores to select accurate video frames and further adjust the text prototypes. This simple yet effective text-visual collaboration enables training-free ZS-TAL to harness historical videos. Extensive experiments show our method significantly outperforms state-of-the-art zero-shot VLM baselines, validating the effectiveness of our framework.

Open-vocabulary 3D semantic segmentation aims to segment arbitrary categories beyond the training set. Existing methods predominantly rely on distilling knowledge from 2D open-vocabulary models. However, aligning 3D features to the 2D representation space restricts intrinsic 3D geometric learning and inherits errors from 2D predictions. To address these limitations, we propose GeoGuide, a novel framework that leverages pretrained 3D models to integrate hierarchical geometry-semantic consistency for open-vocabulary 3D segmentation. Specifically, we introduce an Uncertainty-based Superpoint Distillation module to fuse geometric and semantic features for estimating per-point uncertainty, adaptively weighting 2D features within superpoints to suppress noise while preserving discriminative information to enhance local semantic consistency. Furthermore, our Instance-level Mask Reconstruction module leverages geometric priors to enforce semantic consistency within instances by reconstructing complete instance masks. Additionally, our Inter-Instance Relation Consistency module aligns geometric and semantic similarity matrices to calibrate cross-instance consistency for same-category objects, mitigating viewpoint-induced semantic drift. Extensive experiments on ScanNet v2, Matterport3D, and nuScenes demonstrate the superior performance of GeoGuide.

The effective integration and utilization of multimodal data acquired from image cameras and LiDAR is of paramount importance for perception systems. This paper proposes **I**mage-to-**P**oint Cloud **F**eature Back-**P**rojection (**IPFP**), a novel method for training multimodal fusion networks that back-projects aggregated image-feature centers (from non-projection-aligned image pixels) into the point-cloud feature set via the estimated depth map. Consequently, image features and point cloud features reside within the same three-dimensional space, enabling the natural enrichment of image information into the point cloud during the network forward pass. This process can be selectively enabled when desired -- for instance, at training time -- and turned off in the absence of multimodal data -- for example, at testing time if only LiDAR sensors are available. Experimental results demonstrate that **IPFP** can consistently improve state-of-the-art 3D semantic segmentation models, while retaining the ability to process LiDAR-only data at inference.

Unified image generation and editing models suffer from severe task interference in dense diffusion transformers architectures, where a shared parameter space must compromise between conflicting objectives (e.g., local editing v.s. subject-driven generation). While the sparse Mixture-of-Experts (MoE) paradigm is a promising solution, its gating networks remain task-agnostic, operating based on local features, unaware of global task intent. This task-agnostic nature prevents meaningful specialization and fails to resolve the underlying task interference.In this paper, we propose a novel framework to inject semantic intent into MoE routing. We introduce a Hierarchical Task Semantic Annotation scheme to create structured task descriptors (e.g., scope, type, preservation). We then design Predictive Alignment Regularization to align internal routing decisions with the task's high-level semantics. This regularization evolves the gating network from a task-agnostic executor to a dispatch center. Our model effectively mitigates task interference, outperforming dense baselines in fidelity and quality, and our analysis shows that experts naturally develop clear and semantically correlated specializations.

Autoregressive video models are promising for world modeling via next-frame prediction, but they suffer from exposure bias: a mismatch between training on clean contexts and inference on self-generated frames, causing errors to compound and quality to drift over time. We introduce Backwards Aggregation (BAgger), a self-supervised scheme that constructs corrective trajectories from the model's own rollouts, teaching it to recover from its mistakes. Unlike prior approaches that rely on few-step distillation and distribution-matching losses --which can hurt quality and diversity -- BAgger trains with standard score or flow matching objectives, avoiding large teachers and long-chain backpropagation through time. We instantiate BAgger on causal diffusion transformers and evaluate on text-to-video, video extension, and multi-prompt generation, observing more stable long-horizon motion and better visual consistency with reduced drift.

Humans anticipate, from a glance and a contemplated action of their bodies, how the 3D world will respond, a capability that is equally vital for robotic manipulation. We introduce PointWorld, a large pre-trained 3D world model that unifies state and action in a shared 3D space as 3D point flows: given one or few RGB-D images and a sequence of low-level robot action commands, PointWorld forecasts per-pixel displacements in 3D that respond to the given actions. By representing actions as 3D point flows instead of embodiment-specific action spaces (e.g., joint positions), this formulation directly conditions on physical geometries of robots while seamlessly integrating learning across embodiments. To train our 3D world model, we curate a large-scale dataset spanning real and simulated robotic manipulation in open-world environments, enabled by recent advances in 3D vision and simulated environments, totaling about 2M trajectories and 500 hours across a single-arm Franka and a bimanual humanoid. Through rigorous, large-scale empirical studies of backbones, action representations, learning objectives, partial observability, data mixtures, domain transfers, and scaling, we distill design principles for large-scale 3D world modeling. With a real-time (0.1s) inference speed, PointWorld can be efficiently integrated in the model-predictive control (MPC) framework for manipulation. We demonstrate that a single pre-trained checkpoint enables a real-world Franka robot to perform rigid-body pushing, deformable and articulated object manipulation, and tool use, without requiring any demonstrations or post-training and all from a single image captured in-the-wild.

Federated Prototype Learning (FedCL) has emerged as an effective strategy for handling data heterogeneity in Federated Learning (FL). In FedCL, clients collaboratively construct a set of global feature centers (prototypes), and let local features align with these prototypes to mitigate the effects of data heterogeneity. The performance of FedCL highly depends on the quality of prototypes. Existing methods assume that larger inter-class distances among prototypes yield better performance, and thus design different methods to increase these distances. However, we observe that while these methods increase prototype distances to enhance class discrimination, they inevitably disrupt essential semantic relationships among classes, which are crucial for model generalization. This raises an important question: how to construct prototypes that inherently preserve semantic relationships among classes? Directly learning these relationships from limited and heterogeneous client data can be problematic in FL. Recently, the success of pre-trained language models (PLMs) demonstrates their ability to capture semantic relationships from vast textual corpora. Motivated by this, we propose FedTSP, a novel method that leverages PLMs to construct semantically enriched prototypes from the textual modality, enabling more effective collaboration in heterogeneous data settings. We first use a large language model (LLM) to generate fine-grained textual descriptions for each class, which are then processed by a PLM on the server to form textual prototypes. To address the modality gap between client image models and the PLM, we introduce trainable prompts, allowing prototypes to adapt better to client tasks. Extensive experiments demonstrate that FedTSP mitigates data heterogeneity while significantly accelerating convergence.

We propose Tea-Adapter, a plug-and-play adapter designed to efficiently integrate conditional knowledge from a smaller teacher model into a larger student video diffusion model. Existing controllable video DiT methods face critical challenges: full fine-tuning of billion-parameter models is extremely expensive. At the same time, ControlNets introduce substantial parameter overhead and exhibit limited flexibility for novel multi-condition compositions. To overcome these issues, Tea-Adapter introduces a novel reverse distillation method that enables large video diffusion models to inherit precise control capabilities from smaller, efficiently tuned teacher diffusion models, eliminating the need for full fine-tuning. Moreover, recognizing the intrinsic relationships between different conditions, we replace the cascaded ControlNet design with a Mixture of Condition Experts (MCE) layer. This structure dynamically routes diverse conditional inputs within a unified architecture, supporting both single-condition control and multiple condition combinations without additional training cost. To achieve cross-scale knowledge transfer, we further develop a Feature Propagation Module to ensure efficient and temporally consistent feature propagation across video frames. Experiments demonstrate that Tea-Adapter enables high-fidelity, multi-condition video synthesis, making advanced, controllable video generation feasible on low-resource hardware.

Autoregressive (AR) models offer stable and efficient training, but standard next-token prediction is not well aligned with the temporal structure required for text-conditioned motion generation. We introduce MoScale, a next-scale AR framework that generates motion hierarchically from coarse to fine temporal resolutions. By providing global semantics at the coarsest scale and refining them progressively, MoScale establishes a causal hierarchy better suited for long-range motion structure. To improve robustness under limited text-motion data, we further incorporate cross-scale hierarchical refinement for improving per-scale initial predictions and in-scale temporal refinement for selective bidirectional re-prediction. MoScale achieves SOTA text-to-motion performance with high training efficiency, scales effectively with model size, and generalizes zero-shot to diverse motion generation and editing tasks. Code and additional materials are available on the webpage.

Most existing 3D referring expression segmentation (3DRES) methods rely on dense, high-quality point clouds, while real-world agents such as robots and mobile phones operate with only a few sparse RGB views and strict latency constraints. We introduce Multi-view 3D Referring Expression Segmentation (MV-3DRES), where the model must recover scene structure and segment the referred object directly from sparse multi-view images. Traditional two-stage pipelines, which first reconstruct a point cloud and then perform segmentation, often yield low-quality geometry, produce coarse or degraded target regions, and run slowly. We propose the Multimodal Visual Geometry Grounded Transformer (MVGGT), an efficient end-to-end framework that integrates language information into sparse-view geometric reasoning through a dual-branch design. Training in this setting exposes a critical optimization barrier, termed Foreground Gradient Dilution (FGD), where sparse 3D signals lead to weak supervision. To resolve this, we introduce Per-view No-target Suppression Optimization (PVSO), which provides stronger and more balanced gradients across views, enabling stable and efficient learning. To support consistent evaluation, we build MVRefer, a benchmark that defines standardized settings and metrics for MV-3DRES. Experiments show that MVGGT establishes the first strong baseline and achieves both high accuracy and fast inference, outperforming existing alternatives. The code is available at https://mvggt.github.io/.

Flow-GRPO successfully applies reinforcement learning to flow models, but uses uniform credit assignment across all steps. This ignores the temporal structure of diffusion generation: early steps determine composition and content (low-frequency structure), while late steps resolve details and textures (high-frequency details). Moreover, assigning uniform credit based solely on the final image can inadvertently reward suboptimal intermediate steps, especially when errors are corrected later in the diffusion trajectory. We propose Stepwise-Flow-GRPO, which assigns credit based on each step's reward improvement. By leveraging Tweedie's formula to obtain intermediate reward estimates and introducing gain-based advantages, our method achieves superior sample efficiency and faster convergence. We also introduce a DDIM-inspired SDE that improves reward quality while preserving stochasticity for policy gradients.

Roadside perception datasets are typically constructed via cooperative labeling between synchronized vehicle and roadside frame pairs, but real deployment is often limited roadside-only data due to hardware and privacy constraints. The observation that even human experts struggle to produce accurate labels without vehicle-side data reveals a fundamental learnability problem: many roadside-only scenes contain distant, blurred, or occluded objects whose 3D properties are ambiguous from a single view and can only be reliably annotated by cross-checking paired vehicle--roadside frames. We refer to such cases as inherently ambiguous samples. In this work, we develop an active learning framework for roadside monocular 3D object detection and propose a learnability-driven framework that selects scenes which are both informative and reliably labelable, suppressing inherently ambiguous samples while ensuring coverage. Experiments demonstrate that our method significantly outperforms uncertainty-based baselines, which suggests that learnability, not uncertainty, matters for roadside 3D perception.

Inverse rendering is an ill-posed problem, but priors such as illumination priors can help simplify it. Existing work either disregards the spherical and rotation-equivariant nature of illumination environments or does not provide a well-behaved latent space. We propose a rotation-equivariant variational autoencoder that models natural illumination on the sphere without relying on 2D projections. To preserve the SO(2)-equivariance of environment maps, we use a novel Vector Neuron Vision Transformer (VN-ViT) as encoder and a rotation-equivariant conditional neural field as decoder. In the encoder, we reduce the equivariance from SO(3) to SO(2) using a novel SO(2)-equivariant fully connected layer, an extension of Vector Neurons. We show that our SO(2)-equivariant fully connected layer outperforms standard Vector Neurons when used in our SO(2)-equivariant model. Compared to previous methods, our variational autoencoder enables smoother interpolation in latent space and offers a more well-behaved latent space.

Recent advances in vision-language models (VLMs) have made impressive strides in understanding spatio-temporal relationships in videos. However, when spatial information is obscured, these models struggle to capture purely temporal patterns. We introduce SpookyBench, a benchmark where information is encoded solely in temporal sequences of noise-like frames, mirroring natural phenomena from biological signaling to covert communication. Interestingly, while humans can recognize shapes, text, and patterns in these sequences with over 98% accuracy, state-of-the-art VLMs achieve 0% accuracy. This highlights a critical limitation: an over-reliance on frame-level spatial features and an inability to extract meaning from temporal cues. Overcoming this limitation will require novel architectures or training paradigms that decouple spatial dependencies from temporal processing. Our systematic analysis shows that this issue persists across model scales and architectures. We release SpookyBench to catalyze research in temporal pattern recognition and bridge the gap between human and machine video understanding. Dataset and code are available on the project website.

MLLMs exhibit strong reasoning on isolated queries, yet they operate de novo--solving each problem independently and often repeating the same mistakes. Existing memory-augmented agents mainly store past trajectories for reuse. However, trajectory-based memory suffers from brevity bias, gradually losing essential domain knowledge. More critically, even in truly multimodal problem-solving settings, it records only a **single-modality** trace of past behavior, failing to preserve how visual attention and logical reasoning jointly contributed to the solution. This is fundamentally misaligned with human cognition: semantic memory is both **multimodal and integrated**, preserving visual and abstract knowledge through coordinated but distinct representational streams. We thus introduce **ViLoMem**, a dual-stream memory framework that constructs compact, schema-based memory. It separately encodes visual distraction patterns and logical reasoning errors, enabling MLLMs to learn from their successful and failed experiences. Following a grow-and-refine principle, the system incrementally accumulates and updates multimodal semantic knowledge--preserving stable, generalizable strategies while avoiding catastrophic forgetting. Across nine multimodal benchmarks, **ViLoMem** consistently improves pass@1 accuracy and substantially reduces repeated visual and logical errors. Ablations confirm the necessity of dual-stream memory with explicit distraction-hallucination separation, demonstrating the value of error-aware multimodal memory for lifelong and cross-domain agentic learning.

Diffusion Transformer (DiT) models have achieved unprecedented quality in image and video generation, yet their iterative sampling process remains computationally prohibitive. To accelerate inference, feature caching methods have emerged by reusing or forecasting intermediate representations across timesteps. However, existing caching approaches treat all feature components uniformly. We reveal that DiT feature spaces contain distinct principal and residual subspaces with divergent temporal behavior: the principal subspace evolves smoothly and predictably, while the residual subspace exhibits volatile, low-energy oscillations that resist accurate prediction. Building on this insight, we propose SVD-Cache, a subspace-aware caching framework that decomposes diffusion features via Singular Value Decomposition (SVD), applies exponential moving average (EMA) prediction to the dominant low-rank components, and directly reuses the residual subspace. Extensive experiments demonstrate that SVD-Cache achieves near-lossless across diverse models and methods, including 5.55xspeedup on FLUX and HunyuanVideo, and compatibility with model acceleration techniques including distillation, quantization and sparse attention. Our code is available at \href https://github.com/BlackMaple1203/SVDCache https://github.com/BlackMaple1203/SVDCache .

For video-text retrieval, the use of CLIP has been a de facto standard. However, as CLIP provides only image and text encoders, this consensus has led to a biased paradigm that entirely ignores the sound track of videos. While several attempts have been made to reintroduce audio -- typically by incorporating an audio encoder and fusing its output with visual features -- these methods face two challenges: ineffective representation of speech content and suboptimal vision-audio fusion. To address these issues jointly, we propose SAVE, a Speech Aware Video rEpresentation learning method. SAVE improves upon AVIGATE, a SOTA audiovisual method, with a dedicated speech branch for more effective speech embedding. Furthermore, we introduce soft-ALBEF for early vision-audio alignment that facilitates fusion. Extensive experiments on five benchmarks show that SAVE compares favorably against the SOTA, outperforming AVIGATE by +4.1% on MSRVTT-1k, +1.9% on MSRVTT-3k, +2.5% on VATEX, +9.8% on Charades, and +2.1% on LSMDC, in light of the SumR metric.

Multimodal Large Language Models (MLLMs) are rapidly becoming the intelligence brain of end-to-end autonomous driving systems. A key challenge is to assess whether MLLMs can truly understand and follow complex real-world traffic rules. However, existing benchmarks mainly focus on single-rule scenarios like traffic sign recognition, neglecting the complexity of multi-rule concurrency and conflicts in real driving. Consequently, models perform well on simple tasks but often fail or violate rules in real world complex situations. To bridge this gap, we propose DriveCombo, a text and vision-based benchmark for compositional traffic rule reasoning. Inspired by human drivers' cognitive development, we propose a systematic Five-Level Cognitive Ladder that evaluates reasoning from single-rule understanding to multi-rule integration and conflict resolution, enabling quantitative assessment across cognitive stages. We further propose a Rule2Scene Agent that maps language-based traffic rules to dynamic driving scenes through rule crafting and scene generation, enabling scene-level traffic rule visual reasoning. Evaluations of 14 mainstream MLLMs reveal a pronounced decline in performance as task difficulty increases, especially in the rule conflict resolution task. After splitting the dataset and fine-tuning on the training set, we further observe substantial improvements in both traffic rule reasoning and downstream planning capabilities. These results highlight the effectiveness of DriveCombo in advancing compliant and intelligent autonomous driving systems.

Recent advances in machine unlearning have focused on developing algorithms to remove specific training samples from a trained model. In contrast, we observe that not all models are equally easy to unlearn. Hence, we introduce a family of deep semi-parametric models (SPMs) that exhibit non-parametric behavior during unlearning. SPMs use a fusion module that aggregates information from each training sample, enabling explicit test-time deletion of selected samples without altering model parameters. Empirically, we demonstrate that SPMs achieve competitive task performance to parametric models in image classification and generation, while being significantly more efficient for unlearning. Notably, on ImageNet classification, SPMs reduce the prediction gap relative to a retrained (oracle) baseline by 11% and achieve over 10xfaster unlearning compared to existing approaches on parametric models.