CVPR 2026 Papers

Autonomous driving requires generating safe and reliable trajectories from complex multimodal inputs. Traditional modular pipelines separate perception, prediction, and planning, while recent end-to-end (E2E) systems learn them jointly. Vision-language models (VLMs) further enrich this paradigm by introducing cross-modal priors and commonsense reasoning, yet current VLM-based planners face three key challenges: (i) a mismatch between discrete text reasoning and continuous control, (ii) high latency from autoregressive chain-of-thought decoding, and (iii) inefficient or non-causal planners that limit real-time deployment. We propose ColaVLA, a unified vision-language-action framework that transfers reasoning from text to a unified latent space and couples it with a hierarchical, parallel trajectory decoder. The Cognitive Latent Reasoner compresses scene understanding into compact, decision-oriented meta-action embeddings through ego-adaptive selection and only two VLM forward passes. The Hierarchical Parallel Planner then generates multi-scale, causality-consistent trajectories in a single forward pass. Together, these components preserve the generalization and interpretability of VLMs while enabling efficient, accurate and safe trajectory generation. Experiments on the nuScenes benchmark show that ColaVLA achieves state-of-the-art performance in both open-loop and closed-loop settings with favorable efficiency and robustness.

Large multimodal models (LMMs) have achieved impressive performance on multimodal reasoning, becoming crucial technology for the advancement of intelligent question-answering systems. In real-world educational scenarios, effective teaching extends far beyond providing answers. Experienced teachers analyze students' incorrect answers to trace underlying errors and provide corrective feedback, termed educational diagnostic reasoning, a capability that remains under-explored in existing LMMs. To bridge this research gap, we introduce Edudiag benchmark, requiring LMMs to reconstruct erroneous reasoning chains from incorrect answers and generate corrective feedback. Through an AI-assisted annotation pipeline with rigorous human verification, we create 8K erroneous reasoning chains and corresponding feedback, spanning three representative educational domains: commonsense, science, and mathematics. Extensive evaluation across 28 leading LMMs highlights Edudiag as a challenging testbed, where even leading proprietary LMMs struggle on it and supervised fine-tuning (SFT) on open-source LMMs achieves marginal performance gains. Moreover, we conduct analysis experiments and identify three critical insights for educational diagnostic reasoning: (i) Effective error tracing remains the primary bottleneck, while SFT models still fail to reversely identify errors that commonly occur. (ii) Group relative policy optimization (GRPO) mitigates this bottleneck and boosts performance. (iii) LMMs optimized with GRPO can generate plausible yet challenging distractors for multiple-choice questions based on their self-constructed erroneous reasoning chains. We believe Edudiag provides a new direction for evaluating the advanced LMMs.

3D Large Vision-Language Models (3D LVLMs) built upon Large Language Models (LLMs) have achieved remarkable progress across various multimodal tasks. However, their inherited position-dependent modeling mechanism, Rotary Position Embedding (RoPE), remains suboptimal for 3D multimodal understanding. The vanilla RoPE formulation fails to preserve essential three-dimensional spatial structures when encoding 3D tokens, and its relative distance computation overlooks angular dependencies hindering the model's ability to capture directional variations in visual representations. To overcome these limitations, we introduce Spherical Coordinate-based Positional Embedding (SoPE). Our method maps point-cloud token indices into a 3D spherical coordinate space, enabling unified modeling of spatial locations and directional angles. This formulation preserves the inherent geometric structure of point-cloud data, enhances spatial awareness, and yields more consistent and expressive geometric representations for multimodal learning. In addition, we introduce a multi-scale frequency mixing strategy to fuse feature information across different frequency domains. Experimental results on multiple 3D scene benchmarks validate the effectiveness of our approach, while real-world deployment experiments further demonstrate its strong generalization capability.

Vision-Language Models (VLMs) typically assume a uniform spatial fidelity across the entire field of view of visual inputs, dedicating equal precision to even the uninformative regions. By contrast, human vision is neither uniform nor static. It is adaptive, selective, and resource-efficient. In light of this, we present the first systematic analysis of bio-inspired visual representation methods, providing insights for more efficient and adaptive VLMs. We propose LLMind (Looking Like the Mind), a novel training-free framework that mimics foveated encoding and cortical magnification in human vision to achieve adaptive, efficient representations for VLMs under tight pixel budgets. Our key idea is to explore a Bio-inspired Adaptive Sampling Strategy (BASS). This empowers us to design a Mobius-parameterized module that performs non-uniform sampling while preserving global scene structure. On top of BASS, we introduce closed-loop semantic feedback (CSF) via test-time adaptation to align the perceptual saliency with the textual information from the frozen VLM. We evaluate LLMind against uniform and other sampling baselines across diverse scene-level and region-guided visual question answering (VQA) benchmarks. The results show that ours achieves dramatic gains, with average improvements by +20% on VQAv2, +38% on Seed-Bench, and +37% on A-OKVQA compared to uniform sampling under tight pixel budgets. More surprisingly, results reveal that LLMind can retain up to 82%, 92% and 97% of the full-resolution performance with only 1%, 3% and 5% of the pixels, respectively. Moreover, LLMind is lightweight, plug-and-play, and compatible with existing VLMs without requiring architectural changes.

We present ReMoT, a unified training paradigm to systematically address the fundamental shortcomings of VLMs in spatio-temporal consistency--a critical failure point in navigation, robotics, and autonomous driving. ReMoT integrates two core components: (i) A rule-based automatic framework that generates ReMoT-16K, a large-scale (16.5K triplets) motion-contrast dataset derived from video meta-annotations, surpassing costly manual or model-based generation. (ii) Group Relative Policy Optimization, which we empirically validate, yields optimal performance and data efficiency for learning this contrastive reasoning, far exceeding standard Supervised Fine-Tuning. We also construct the first benchmark for fine-grained motion contrast triplets to measure a VLM's discrimination of subtle motion attributes (e.g., opposing directions). The resulting model achieves SOTA performance on our new benchmark and multiple standard VLM benchmarks, culminating in a remarkable 25.1 performance leap on spatio-temporal reasoning tasks.

Current multimodal image retrieval benchmarks focus on relatively simple queries where target images are either described directly or by simple composition with an input image. When retrieval requires complex reasoning to determine the target image, the task becomes significantly more challenging, yet standardized benchmarks for this setting do not exist. To fill this gap, we introduce RMIR, a benchmark dataset of 1634 queries requiring reasoning across three categories: functional (object affordances), temporal (time-based relationships), and causal (cause-effect reasoning). Each query combines visual and textual inputs that demand robust visual understanding together with logical inference, beyond surface-level matching, to identify correct target images. In addition to the dataset itself, we present a pipeline to generate the dataset, which can be used to generate additional reasoning-intensive retrieval data at scale. Evaluation of state-of-the-art models on RMIR reveals significant performance gaps, with the best model achieving only 46.53% recall@20 averaged across reasoning categories. Our evaluation also shows that generative embedding models with explicit reasoning substantially outperform discriminative approaches, with reasoning-aware training proving more impactful than model scale. Our systematic analysis exposes fundamental limitations in current multimodal retrieval systems and establishes RMIR as a challenging testbed for developing multimodal, reasoning-capable retrieval models. Our dataset and code are available at https://github.com/amazon-science/rmir

Chain-of-Thought (CoT) has recently shown encouraging progress in the vision language model. However, the pure-vision CoT (i.e., chain-of-vision) has been underexplored in visual in-context learning. In this paper, we introduce Diffusion Guided Chain-of-Vision, which integrates an explicit chain-of-thought process into autoregressive vision models through vision prior from pre-trained diffusion models. Concretely, we find that pre-trained diffusion models induce a reliable probability flow in image space, where intermediate images sampled along this flow exhibit visual coherence and serve as task-free, chain-of-vision supervision for pure-vision autoregressive models. Extensive experiments on diverse vision tasks and multi-scale models validate the effectiveness of our proposed method for visual in-context learning. Code and dataset will be publicly available.

The application of Large Multimodal Models (LMMs) to long-form video understanding is constrained by limited context lengths and the computationally prohibitive cost of processing dense video tokens. Consequently, recent research has focused on query-aware frame selection, methods that often incur significant computational overhead. This paper challenges the assumption that such complex search mechanisms are universally necessary. We first identify and validate a query typology distinguishing between global query and localized query. We demonstrate that while uniform sampling is both effective and efficient for global queries, localized queries indeed necessitate query-aware selection for optimal performance. Building on this insight, we propose DIG, a training-free frame selection framework that adapts its strategy based on the query type. Specifically, DIG employs efficient uniform sampling for global queries while activating a specialized pipeline to extract query-relevant frames for localized queries. Experiments on three long-form video understanding benchmarks demonstrate that DIG consistently outperforms existing baselines and robustly improves LMM performance, even when scaling the input frame count to 256.

Continual unlearning poses the challenge of enabling large vision-language models to selectively refuse specific image-instruction pairs in response to sequential deletion requests, while preserving general utility. However, sequential unlearning updates distort shared representations, creating spurious associations between vision-language pairs and refusal behaviors that hinder precise identification of refusal targets, resulting in inappropriate refusals. To address this challenge, we propose a novel continual unlearning framework that grounds refusal behavior in fine-grained descriptions of visual and textual concepts decomposed from deletion targets. We first identify which visual-linguistic concept combinations characterize each forget category through a concept modulator, then determine how to generate appropriate refusal responses via a mixture of refusal experts, termed refusers, each specialized for concept-aligned refusal generation. To generate concept-specific refusal responses across sequential tasks, we introduce a multimodal, concept-driven routing scheme that reuses refusers for tasks sharing similar concepts and adapts underutilized ones for novel concepts. Extensive experiments on vision-language benchmarks demonstrate that the proposed framework outperforms existing methods by generating concept-grounded refusal responses and preserving the general utility across unlearning sequences.

Referring video object segmentation (RVOS) aims to segment objects within a video according to natural language expressions. Unlike earlier works focusing on static single-object scenarios, recent studies address more complex motion scenes. Previous methods typically adopt a query-based, logically multi-stage pipeline to handle these scenarios. However, this paradigm learns trajectory consistency modeling and multimodal fusion from scratch, which often leads to trajectory inconsistencies and insufficient multimodal understanding. To address these limitations, we propose DeRVOS, a framework that decouples RVOS into two key branches: consistent trajectory generation and multimodal understanding. We extract temporally consistent object representations using a powerful pretrained instance trajectory generation model and perform cross-modal alignment via a unified multimodal encoder, enabling upstream modeling of trajectory consistency and vision-language understanding. This design reduces RVOS to the task of modeling the relationship between referring expressions and instance trajectories. To connect the two branches and enable efficient motion-aware semantic understanding, we introduce the Trajectory Alignment and Implicit Selection (TAIS) module, which progressively performs cross-frame multimodal alignment and motion-guided implicit trajectory selection. Extensive experiments demonstrate that DeRVOS achieves state-of-the-art results on both traditional RVOS benchmarks and the challenging MeViS dataset, surpassing LVLM-based methods by 4.7%.

With the advancement of multi-modal Large Language Models (LLMs), Video LLMs have been further developed to perform on holistic and specialized video understanding. However, existing works are limited to specialized video understanding tasks, failing to achieve a comprehensive and multi-grained video perception. To bridge this gap, we introduce UFVideo, the first Video LLM with unified multi-grained cooperative understanding capabilities. Specifically, we design unified visual-language guided alignment to flexibly handle video understanding across global, pixel and temporal scales within a single model. UFVideo dynamically encodes the visual and text inputs of different tasks and generates the textual response, temporal localization, or grounded mask. Additionally, to evaluate challenging multi-grained video understanding tasks, we construct the UFVideo-Bench consisting of three distinct collaborative tasks within the scales, which demonstrates UFVideo's flexibility and advantages over GPT-4o. Furthermore, we validate the effectiveness of our model across 9 public benchmarks covering various common video understanding tasks, providing valuable insights for future Video LLMs.

Dataset distillation (DD) aims to synthesize compact training sets that enable models to achieve high accuracy with significantly fewer samples. Recent diffusion-based DD methods commonly introduce semantic guidance through late-stage cross-attention, where textual prompts tend to dominate the generative process. Although this strategy enforces label relevance, it diminishes the contribution of visual latents, resulting in over-corrected samples that mirror prompt patterns rather than reflecting intrinsic visual features. To solve this problem, we introduce an Early Vision-Language Fusion (EVLF) method that aligns textual and visual embeddings at the transition between the encoder and the generative backbone. By incorporating a lightweight cross-attention module at this transition, the early representations simultaneously encode local textures and global semantic directions across the denoising process. Importantly, EVLF is plug-and-play and can be easily integrated into any diffusion-based dataset distillation pipeline with an encoder. It works across different denoiser architectures and sampling schedules without any task-specific modifications. Extensive experiments demonstrate that EVLF generates semantically faithful and visually coherent synthetic data, yielding consistent improvements in downstream classification accuracy across varied settings. Source code is available at https://github.com/wenqi-cai297/earlyfusion-for-dd .

Large multimodal language models have made rapid progress on vision-language tasks, yet their potential for zero-/few-shot object detection (ZSOD/FSOD) under a closed set of target classes has yet to be fully explored.. ZSOD/FSOD is hampered by data scarcity and catastrophic forgetting. Although vision-language models (VLMs) demonstrated excellent performance on multiple benchmarks, they typically rely on large-scale visual pre-training, which is inconsistent with the goal of FSOD to test generalization to new categories under limited supervision. We introduce AgentDet, a shared-blackboard multi-agent framework that unifies ZSOD and FSOD via pseudo-incremental learning. AgentDet decouples detection into four cooperating roles--Agent-Scout, Agent-Pinner, Agent-Curator, and Agent-Judge--which collaboratively maintain a Shared Blackboard and a Knowledge Base. For efficiency, we only train Agent-Judge by updating its image encoder and LLM-based detection head, which is a lightweight recipe that encourages generalization to previously unseen categories. On PASCAL VOC and MS COCO ZSOD/FSOD datasets, AgentDet achieves strongly competitive performance with state-of-the-art results in several settings.

End-to-end autonomous driving (E2E-AD) demands effective processing of multi-view sensor data and robust handling of diverse and complex driving scenarios, particularly rare maneuvers such as aggressive turns. The recent success of the Mixture-of-Experts (MoE) architecture in Large Language Models (LLMs) demonstrates that expert specialization enables strong scalability. In this work, we propose DriveMoE, a novel MoE-based E2E-AD framework, with a Scene-Specialized Vision MoE and a Skill-Specialized Action MoE. First, we introduce Drive-p0, a Vision-Language-Action (VLA) baseline adapted from Embodied AI for autonomous driving, which serves as the foundation model for DriveMoE. Building on this, we strengthen perception through a carefully designed Vision MoE, where a router adaptively selects context-relevant camera views. This mechanism is inspired by human driving cognition, in which attention is directed to key visual cues rather than to all sensory inputs simultaneously. Beyond perception, we introduce an Action MoE that augments the framework by training a router to activate specialized expert modules tailored to distinct driving behaviors. Within the Action MoE, we implement two distinct styles(Token-level Router and Trajectory-level Router) and extensively explore their applicability in autonomous driving. In Bench2Drive closed-loop evaluations, DriveMoE demonstrates robust performance across diverse driving scenarios, alleviates the mode-averaging effect that limits existing models, and achieves state-of-the-art results with significant improvements over Drive-p0. We will release our code and models of DriveMoE and Drive-p0.

Video Question Answering (VideoQA) has made significant strides by leveraging multimodal learning to align visual and textual modalities. However, current benchmarks overwhelmingly focus on questions answerable through explicit visual content - actions, objects, and events - directly observable within individual frames or short clips. To truly understand videos as humans do, models must go beyond what is directly shown, inferring hidden relationships and contextual cues that are only implied across frames. Current benchmarks fail to capture this essential aspect of video understanding. To address this gap, we introduce VRR-QA, a benchmark for Visual Relational Reasoning Beyond Explicit Cues. We curate our benchmark from creative and cinematic videos such as movies, that deliberately employ storytelling techniques which omit direct depictions of certain events or relations, requiring viewers to infer them. VRR-QA comprises 1K meticulously expert-annotated QA pairs drawn from 1K creative video clips covering 15 genres across 7 decades of content, from both live-action and animated titles. Our extensive evaluations on 14 leading VideoQA models reveals consistent and significant performance degradation, underscoring their reliance on surface-level visual cues and highlighting the difficulty of implicit reasoning. Even the best model substantially underperforms human baselines with only 64% accuracy. Performance variations across models further illustrate the complexity and diversity of the challenges presented by VRR-QA. By releasing both dataset and data collection framework, VRR-QA establishes a rigorous, diverse, and reproducible testbed for advancing VideoQA.

Driving World Models (DWMs) have been developing rapidly with the advances of generative models. However, existing DWMs lack 3D scene understanding capabilities and can only generate content conditioned on input data, without the ability to interpret or reason about the driving environment. Moreover, current approaches represent 3D spatial information with point cloud or BEV features do not accurately align textual information with the underlying 3D scene. To address these limitations, we propose a novel unified DWM framework based on 3D Gaussian scene representation, which enables both 3D scene understanding and multi-modal scene generation, while also enabling contextual enrichment for understanding and generation tasks. Our approach directly aligns textual information with the 3D scene by embedding rich linguistic features into each Gaussian primitive, thereby achieving early modality alignment. In addition, we design a novel task-aware language-guided sampling strategy that removes redundant 3D Gaussians and injects accurate and compact 3D tokens into LLM. Furthermore, we design a dual-condition multi-modal generation model, where the information captured by our vision-language model is leveraged as a high-level language condition in combination with a low-level image condition, jointly guiding the multi-modal generation process. We conduct comprehensive studies on the nuScenes, and NuInteract datasets to validate the effectiveness of our framework. Our method achieves state-of-the-art performance. We will release the code publicly on GitHub.

Accurate depth estimation in wide field is highly desired in applications of autonomous driving, robot vision and drone controls. Biological compound eyes inspire wide Field of View (FOV) depth estimation, yet their artificial implementations face the challenge of modality misalignment. Specifically, the spherical imaging data do not align well with planar neural networks, diminishing the learning efficiency. Herein, we propose SCE-Depth, a bio-inspired framework for spherical compound eye depth estimation, which processes spherical images natively on a HEALPix grid using a spherical neural network. This approach achieves a unified 180deg FOV while avoiding the errors typically introduced by modality conversion. Additionally, we identify a depth-sensitive gradient feature from the overlapping FOVs of adjacent ommatidia. To exploit it, we introduce a spherical Sobel operator called the Spherical Gradient Feature Extractor (SGFE) and a corresponding Spherical Gradient Loss (SGL), which jointly extract gradient features on the HEALPix grid, enabling gradient-aware depth prediction. Extensive benchmark experiments demonstrate that these strategies enable SCE-Depth to substantially reduce depth estimation error compared to fisheye-based baselines, with particularly large improvements in peripheral accuracy. We also demonstrate the generalization capability of SCE-Depth to other wide FOV data modalities, such as fisheye and panoramic imagery. Code is available at https://github.com/ZhuYi2000/SCE-Depth.

Pre-trained vision encoders like DINOv2 have demonstrated exceptional performance on unimodal tasks. However, we observe that their features are poorly aligned across different modalities. For instance, the feature embedding for an RGB image and its corresponding depth map of the same scene exhibit a cosine similarity that is nearly identical to that of two random, unrelated images. To address this, we propose the Omnivorous Vision Encoder, a novel framework that learns a modality-agnostic feature space. We train the encoder with a dual objective: first, to maximize the feature alignment between different modalities of the same scene; and second, a distillation objective that anchors the learned representations to a fully frozen teacher such as DINOv2. The resulting student encoder becomes "omnivorous" by producing a consistent, unified embedding for a given scene, regardless of the input modality (RGB, Depth, Segmentation, etc.). This approach enables robust cross-modal understanding while retaining the discriminative semantics of the original foundation model.

Recent advances in 3D vision-language models (VLMs) highlight a strong potential for 3D scene understanding and reasoning.However, effectively tokenizing 3D scenes into holistic scene tokens, and leveraging these tokens across diverse 3D understanding tasks, remain highly challenging. We present NDTokenizer3D, a generalist 3D VLM that performs a wide range of 3D scene understanding tasks while naturally supporting human interactions, thereby bridging language-level reasoning with 3D spatial understanding. The core of our approach is a novel three-stage scene tokenization pipeline built upon a Multi-Scale Normal Distributions Transform (NDT) representation, paired with a Multi-Scale NDT Decoder (MSDec).Specifically, NDTokenizer3D first constructs a multi-scale NDT representation from raw high-resolution point clouds, preserving both global context and fine-grained geometric details. Next, the MSDec progressively fuses cross-scale NDT features, producing holistic scene tokens consumable by LLM endpoints. Beyond tokenization, MSDec is repurposed as a general interface for human-interactive prompting (points, boxes, masks) and segmentation-mask decoding, unifying diverse 3D scene understanding tasks within a single architecture. With this compact and unified design, NDTokenizer3D offers a fine-grained, general-purpose 3D VLM, achieving remarkable improvements in 3D Referring Segmentation, 3D Visual Question Answering, and 3D Dense Captioning.