CVPR 2026 Papers

Cognitive science suggests that spatial ability develops progressively--from perception to reasoning and interaction. Yet in multimodal LLMs (MLLMs), this hierarchy remains poorly understood, as most studies focus on a narrow set of tasks. We introduce SpatialTree, a cognitive-science-inspired hierarchy that organizes spatial abilities into four levels: low-level perception (L1), mental mapping (L2), simulation (L3), and agentic competence (L4). Based on this taxonomy, we construct the first capability-centric hierarchical benchmark, thoroughly evaluating mainstream MLLMs across 27 sub-abilities. The evaluation results reveal a clear structure: L1 skills are largely orthogonal, whereas higher-level skills are strongly correlated, indicating increasing interdependency. Through targeted supervised fine-tuning, we uncover a surprising transfer dynamic--negative transfer within L1, but strong cross-level transfer from low- to high-level abilities with notable synergy. Finally, we explore how to improve the entire hierarchy. We find that naive RL that encourages extensive "thinking" is unreliable: it helps complex reasoning but hurts intuitive perception. We propose a simple auto-think strategy that suppresses unnecessary deliberation, enabling RL to consistently improve performance across all levels. By building SpatialTree, we provide a proof-of-concept framework for understanding and systematically scaling spatial abilities in MLLMs.

Long video understanding is challenging due to dense visual redundancy, long-range temporal dependencies, and the tendency of chain-of-thought and retrieval-based agents to accumulate semantic drift and correlation-driven errors. We argue that long-video reasoning should begin not with reactive retrieval, but with deliberate task formulation: the model must first articulate what must be true in the video for each candidate answer to hold. This thinking-before-finding principle motivates VideoHV-Agent, a framework that reformulates video question answering as a structured hypothesis-verification process.Based on video summaries, a Thinker rewrites answer candidates into testable hypotheses, a Judge derives a discriminative clue specifying what evidence must be checked, a Verifier grounds and tests the clue using localized, fine-grained video content, and an Answer agent integrates validated evidence to produce the final answer.Experiments on three long-video understanding benchmarks show that VideoHV-Agent achieves state-of-the-art accuracy while providing enhanced interpretability, improved logical soundness, and lower computational cost.

Realistic 3D city generation is fundamental to a wide range of applications, including virtual reality and digital twins. However, most existing methods rely on training a single diffusion model, which limits their ability to generate personalized and boundless city-scale scenes. In this paper, we present Yo'City, a novel agentic framework that enables user-customized and infinitely expandable 3D city generation by leveraging the reasoning and compositional capabilities of off-the-shelf large models. Specifically, Yo'City first conceptualizes the city through a top-down planning strategy that defines a hierarchical "City-District-Grid" structure. The Global Planner determines the overall layout and potential functional districts, while the Local Designer further refines each district with detailed grid-level descriptions. Subsequently, the grid-level 3D generation is achieved through a produce-refine-evaluate isometric image synthesis loop, followed by image-to-3D generation. To simulate continuous city evolution, Yo'City further introduces a user-interactive, relationship-guided expansion mechanism, which performs scene graph-based distance- and semantics-aware layout optimization, ensuring spatially coherent city growth. To comprehensively evaluate our method, we construct a diverse benchmark dataset and design six multi-dimensional metrics that assess generation quality from the perspectives of semantics, geometry, texture, and layout. Extensive experiments demonstrate that Yo'City consistently outperforms existing state-of-the-art methods across all evaluation aspects.

Most of the multi-agent video understanding frameworks adopt static and non-learnable tool invocation mechanisms, which limit the discovery of diverse clues essential for robust perception and reasoning regarding temporally or spatially complex videos. To address this challenge, we propose a novel Multi-agent system for video understanding, namely VideoChat-M1. Instead of using a single or fixed policy, we adopt a distinct Collaborative Policy Planning (CPP) paradigm with multiple policy agents, which comprises three key processes. (1)Policy Generation: Each agent generates its unique tool invocation policy tailored to the user's query. (2) Policy Execution: Each agent sequentially invokes relevant tools to execute its policy and explore the video content. (3) Policy Communication: During the intermediate stages of policy execution, agents interact with one another to update their respective policies. Through this collaborative framework, all agents work in tandem, dynamically refining their preferred policies based on contextual insights from peers. Moreover, we equip our CPP paradigm with Multi-Agent Reinforcement Learning (MARL). Consequently, policy agents can be jointly optimized to enhance the performance, guided by both the final answer reward and intermediate collaborative process feedback. Extensive experiments demonstrate that VideoChat-M1 achieves SOTA performance across eight benchmarks on four tasks. Notably, on LongVideoBench, our method outperforms Gemini 2.5 pro by 3.6% and GPT-4o by 15.6%.

Current research in autonomous driving heavily relies on large-scale driving datasets for model fitting or trial-and-error learning strategies in simulation environments. However, these approaches suffer from limited behavioral diversity and fail to cover complex edge-case interactions. To address these limitations, we model the driving environment as an Active Markov Game (AMG) and introduce a multi-agent co-evolutionary training framework for more realistic interactive learning. The AMG formulation extends traditional Markov games by explicitly making state transitions and rewards dependent on the evolving strategies of the agents, thus capturing the interactive dynamics and strategic coupling between the ego vehicle and surrounding agents. Building on this, our multi-agent co-evolutionary training mechanism jointly optimizes the ego vehicle's policy and a diverse pool of opponent strategies, allowing all agents to adapt to each other's behaviors during training. This game-theoretic approach produces a robust ego agent capable of handling diverse, non-stationary driving strategies, overcoming the "non-responsive opponent" limitation found in prior methods. In CARLA simulations of unsignalized intersections and long-tail scenarios, our method performs exceptionally well, achieving near-perfect success rates (98%), and significantly outperforming state-of-the-art baselines such as PPO, DDPG, and IPPO in terms of generalization, safety margins, and control smoothness. These results demonstrate that our approach substantially enhances the robustness, safety, and strategic adaptability of autonomous driving in complex multi-agent environments.

Embodied Multi-Agent Systems have proven highly effective in addressing complex tasks through coordinated collaboration among heterogeneous agents. However, real-world environments and task specifications are inherently dynamic, exhibiting frequent changes, uncertainty, and variability. Despite these characteristics, most existing frameworks employ static architectures with fixed agent capabilities and rigid task allocation strategies, which substantially constrain their adaptability to evolving conditions. This inflexibility presents significant challenges to maintaining robust and efficient multi-agent cooperation in dynamic and unpredictable settings.To address these limitations, we propose DRAMA, short for Dynamic Orchestration for Resilient Multi-Agent Ecosystems, tailored for rapidly changing environments. DRAMA adopts a multilayer architecture that incorporates three principal mechanisms: adaptive scheduling through an affinity-driven mechanism, fault-tolerant continuity via hierarchical trust-chain task takeover, and collective spatial intelligence that consolidates distributed observations for predictive reasoning. Together, these components enable event-triggered rescheduling and decentralized fault recovery, ensuring uninterrupted task execution amid agent arrivals, dropouts, or recoveries. Extensive experiments in the embodied VirtualHome-Social environment demonstrate that DRAMA achieves a 7% improvement in runtime efficiency and a 10% increase in throughput compared with state-of-the-art baselines, while maintaining superior stability and robustness under dynamic agent populations.

Despite rapid advances in text-to-video synthesis, generated video quality remains critically dependent on precise user prompts. Existing test-time optimization methods, successful in other domains, struggle with the multi-faceted nature of video. In this work, we introduce VISTA (Video Iterative Self-improvemenT Agent), a novel multi-agent system that autonomously improves video generation through refining prompts in an iterative loop. VISTA first decomposes a user's idea into a structured temporal plan. After generation, the best video is identified through a robust pairwise tournament. This winning video is then critiqued by a trio of specialized agents focusing on visual, audio, and contextual fidelity. Finally, a reasoning agent synthesizes this feedback to introspectively rewrite and enhance the prompt for the next generation cycle. Experiments on single- and multi-scene video generation scenarios show that while prior methods yield inconsistent gains, VISTA consistently improves video quality and alignment with user intent, achieving up to 60% pairwise win rate against state-of-the-art baselines. Human evaluators concur, preferring VISTA's outputs in 66.4% of comparisons.

We introduce NitroGen, a vision-action foundation model for generalist gaming agents that is trained on 40,000 hours of gameplay videos across more than 1,000 games. We scale embodied agents through three key ingredients: 1) an internet-scale video-action dataset constructed by automatically extracting player actions from publicly available gameplay videos, 2) a multi-game benchmark environment that can measure cross-game generalization, and 3) a unified vision-action model trained with large-scale behavior cloning. NitroGen exhibits competence across diverse domains, including combat encounters in 3D action games, high-precision control in 2D platformers, and exploration in procedurally generated worlds. When fine-tuned, pre-training transfers effectively to unseen games, achieving up to 52% relative improvement in task success rates over models trained from scratch. We release the dataset, evaluation suite, and model weights to advance research on generalist embodied agents.

Recent advancements in multimodal large language models (MLLMs) and video agent systems have significantly improved general video understanding. However, when applied to scientific video understanding and educating--a domain that demands external professional knowledge integration and rigorous step-wise reasoning--existing approaches often struggle. To bridge this gap, we propose SciEducator, an iterative self-evolving multi-agent system for scientific video comprehension and education. Rooted in the classical Deming Cycle from management science, our design reformulates its Plan-Do-Study-Act philosophy into a self-evolving reasoning and feedback mechanism, which facilitates the interpretation of intricate scientific activities in videos. Moreover, SciEducator can produce multimodal educational content tailored to specific scientific processes, including textual instructions, visual guides, audio narrations, and interactive references. To support evaluation, we construct SciVBench, a benchmark consisting of 500 expert-verified and literature-grounded science QA pairs across five categories, covering physical, chemical, and everyday phenomena. Extensive experiments demonstrate that SciEducator substantially outperforms leading closed-source MLLMs (e.g., Gemini, GPT-4o) and state-of-the-art video agents on the benchmark, establishing a new paradigm for the community.

Adversarial patches have emerged as a popular privacy-preserving approach for resisting AI-driven surveillance systems. However, their conspicuous appearance makes them difficult to deploy in real-world scenarios. In this paper, we propose a thermally activated adversarial wearable designed to ensure adaptability and effectiveness in complex real-world environments. The system integrates thermochromic dyes with flexible heating units to induce visually dynamic adversarial patterns on clothing surfaces. In its default state, the clothing appears as an ordinary black T-shirt. Upon heating via an embedded thermal unit, hidden adversarial patterns on the fabric are activated, allowing the wearer to effectively evade detection across both visible and infrared modalities. Physical experiments demonstrate that the adversarial wearable achieves rapid texture activation within 50 seconds and maintains an adversarial success rate above 80% across diverse real-world surveillance environments. This work demonstrates a new pathway toward physically grounded, user-controllable anti-AI systems, highlighting the growing importance of proactive adversarial techniques for privacy protection in the age of ubiquitous AI surveillance.

Video agentic models have advanced challenging video-language tasks. However, most agentic approaches still heavily rely on greedy parsing over densely sampled video frames, resulting in high computational cost. We present VideoSeek, a long-horizon video agent that leverages video logic flow to actively seek answer-critical evidence instead of exhaustively parsing the full video. This insight allows the model to use far fewer frames while maintaining, or even improving, its video understanding capability. VideoSeek operates in a think-act-observe loop with a well-designed toolkit for collecting multi-granular video observations. This design enables query-aware exploration over accumulated observations and supports practical video understanding and reasoning. Experiments on four challenging video understanding and reasoning benchmarks demonstrate that VideoSeek achieves strong accuracy while using far fewer frames than prior video agents and standalone LMMs. Notably, VideoSeek achieves a 10.2 absolute points improvement on LVBench over its base model, GPT-5, while using 93% fewer frames. Further analysis highlights the significance of leveraging video logic flow, strong reasoning capability, and the complementary roles of toolkit design.

The rapid advancement of Multimodal Large Language Models (MLLMs) has given rise to sophisticated autonomous agents capable of performing complex, human-like tasks across the web. However, this also introduces significant security risks, particularly from stealthy MLLM agents that can evade conventional detection mechanisms by mimicking human behavior. In this paper, we propose DualMirage, a novel CAPTCHA framework that proactively counters and identifies stealthy agents by exploiting fundamental disparities between human and machine perception. DualMirage employs a dual-pronged strategy: (1) Contour Illusions, which utilize cognitive principles to generate illusory contours that humans perceive effortlessly yet pose interpretation challenges for MLLMs; and (2) Adversarial Illusions, which embed human-imperceptible perturbations optimized to mislead the visual encoders of target MLLMs and thereby elicit characteristic, identifiable model responses. Evaluations on five state-of-the-art MLLMs demonstrate that DualMirage achieves an average 95.8% human success rate while blocking MLLM agents (up to 100% agent blocking rate), outperforming existing CAPTCHAs. Furthermore, DualMirage induces models to expose identities actively, achieving 58.8% white-box and 21.9% black-box attack success rates, proving effective against stealthy multimodal agents.

Referring Video Object Segmentation (RVOS) aims to segment objects in videos based on textual queries. Current methods mainly rely on large-scale supervised fine-tuning (SFT) of Multi-modal Large Language Models (MLLMs). However, this paradigm suffers from heavy data dependence and limited scalability against the rapid evolution of MLLMs. Although recent zero-shot approaches offer a flexible alternative, their performance remains significantly behind SFT-based methods, due to the straightforward workflow designs. To address these limitations, we propose Refer-Agent, a collaborative multi-agent system with alternating reasoning-reflection mechanisms. This system decomposes RVOS into step-by-step reasoning process. During reasoning, we introduce a Coarse-to-Fine frame selection strategy to ensure the frame diversity and textual relevance, along with a Dynamic Focus Layout that adaptively adjusts the agent's visual focus.Furthermore, we propose a Chain-of-Reflection mechanism, which employs a Questioner-Responder pair to generate a self-reflection chain, enabling the system to verify intermediate results and generates feedback for next-round reasoning refinement. Extensive experiments on five challenging benchmarks demonstrate that Refer-Agent significantly outperforms state-of-the-art methods, including both SFT-based models and zero-shot approaches.Moreover, Refer-Agent is flexible and enables fast integration of new MLLMs without any additional fine-tuning costs. Code will be released.

Large multimodal models (LMMs) have shown great potential for video reasoning with textual Chain-of-Thought. However, they remain vulnerable to hallucinations, especially when processing long-form videos where evidence is sparse and temporally dispersed. Inspired by how humans comprehend long videos--by first skimming globally and then examining relevant clips for details--we introduce LongVT, an end-to-end agentic framework that enables "Thinking with Long Videos" via interleaved Multimodal Chain-of-Tool-Thought. Specifically, we exploit LMMs' inherent temporal grounding ability as a native video cropping tool to zoom in on a specific video clip and resample finer-grained video frames. This global-to-local reasoning loop continues until answers are grounded in retrieved visual evidence. Given the scarcity of fine-grained question-answering (QA) data for the long video reasoning task, we curate and release a data suite named VideoSIAH to facilitate both training and evaluation. Specifically, our training dataset consists of 247.9K samples for tool-integrated cold-start supervised fine-tuning, 1.6K samples for agentic reinforcement learning, and 15.4K samples for agentic reinforcement fine-tuning. Our evaluation benchmark consists of 652 QA pairs that are carefully curated through a semi-automatic data pipeline with human-in-the-loop validation. With a meticulously designed three-stage training strategy and extensive empirical validation, LongVT consistently outperforms existing strong baselines across four challenging long-video understanding and reasoning benchmarks. Our code, data, and model weights are publicly available at https://github.com/EvolvingLMMs-Lab/LongVT.

Text-guided 3D editing aims to modify existing 3D assets using natural-language instructions. Current methods struggle to jointly understand complex prompts, automatically localize edits in 3D, and preserve unedited content. We introduce Vinedresser3D, an agentic framework for high-quality text-guided 3D editing that operates directly in the latent space of a native 3D generative model. Given a 3D asset and an editing prompt, Vinedresser3D uses a multimodal large language model to infer rich descriptions of the original asset, identify the parts to edit and the edit type (addition, modification, deletion), and generate decomposed structural and appearance-level text guidance. The agent then selects an informative view and applies an image editing model to obtain visual guidance. Finally, an inversion-based rectified-flow inpainting pipeline with an interleaved sampling module performs editing in the 3D latent space, enforcing prompt alignment while maintaining 3D coherence and unedited regions. Experiments on diverse 3D edits demonstrate that Vinedresser3D outperforms prior baselines in both automatic metrics and human preference studies, while enabling precise, coherent, and mask-free 3D editing.

Large Language Models are increasingly integrated as the cognitive core of 3D embodied agents to enable complex environmental reasoning. However, these agents tend to inherit the critical flaw of hallucination, often failing to ground their responses to their 3D view. While Visual Contrastive Decoding (VCD) is a powerful training-free method for mitigating hallucinations in 2D image-based models, it has not been adapted to the complex 3D embodied environment. In this paper, we embarked on the ambitious goal of being the very first to bridge this gap by introducing a VCD framework for 3D embodied agents. Our method operates at inference time by generating a "negative" 3D context, not by blurring an image, but by applying novel distortions directly to a 3D scene graph, such as swapping object category labels or noising positional coordinates. We evaluate our approach on standard evaluation benchmarks and find that it consistently outperforms existing models. For example, in the random category of 3D-POPE, our 3D-VCD method reduced Yes-rate from 99.9% to 75.1% while simultaneously increasing precision from 50.0% to 62.2%. These results demonstrate that our training-free approach effectively curbs hallucination, yielding 3D agents that are significantly more reliable and grounded.

Text-to-image (T2I) diffusion models such as SDXL and FLUX have achieved impressive photorealism, yet small-scale distortions remain pervasive in limbs, face, text and so on. Existing refinement approaches either perform costly iterative re-generation or rely on vision-language models (VLMs) with weak spatial grounding, leading to semantic drift and unreliable local edits. To close this gap, we propose **Agentic Retoucher**, a hierarchical decision-driven framework that reformulates post-generation correction as a human-like *perception-reasoning-action* loop.Specifically, we design (1) a **perception agent** that learns contextual saliency for fine-grained distortion localization under text-image consistency cues, (2) a **reasoning agent** that performs human-aligned inferential diagnosis via progressive preference alignment, and (3) an **action agent** that adaptively plans localized inpainting guided by user preference. This design integrates perceptual evidence, linguistic reasoning, and controllable correction into a unified, self-corrective decision process. To enable fine-grained supervision and quantitative evaluation, we further construct **GenBlemish-27K**, a dataset of 6K T2I images with 27K annotated artifact regions across 12 categories.Extensive experiments demonstrate that Agentic Retoucher consistently outperforms state-of-the-art methods in perceptual quality, distortion localization and human preference alignment, establishing a new paradigm for self-corrective and perceptually reliable T2I generation.

Document understanding is a long-standing practical task. Vision-Language Models (VLMs) have gradually become a primary approach in this domain, demonstrating effective performance on single-page tasks. However, their effectiveness diminishes when handling long documents. In such scenarios, clues are often scattered across multiple pages and modalities, and redundancy from lengthy inputs can impair the model's judgment. While retrieval-augmented generation mitigates this issue by filtering for question-relevant content, the retrieved results still contain substantial redundancy. To address these limitations, we propose SLEUTH, a multi-agent framework. Concretely, SLEUTH orchestrates a retriever and four collaborative agents in a coarse-to-fine process. The framework identifies key textual and visual clues within the retrieved pages, filters for salient visual evidence such as tables and charts, and analyzes the query to devise a reasoning strategy. It ultimately synthesizes a distilled, evidence-dense multimodal context to generate the final prediction. SLEUTH is model-agnostic and scalable. When paired with advanced VLM backbones, it consistently improves performance on multiple long-document benchmarks, achieving SOTA results. Ablation studies verify each module's effectiveness and confirm the benefits of our hierarchical refinement paradigm.

Large Vision-Language Models (VLMs) are increasingly deployed across diverse applications, such as AI agents, yet remain vulnerable to targeted adversarial attacks. However, the practical robustness of such attacks often remains unclear with limited evaluation under defenses. Diffusion-based purification (DBP), a widely adopted black-box defense, effectively blocks current attacks by removing adversarial perturbations via generative diffusion. Prior DBP evasion attacks target white-box image classifiers and are ill-suited to VLMs, incurring high computational costs and gradient instability when adapted. In this paper, we present PureProof, a black-box targeted attack on VLMs resilient to DBP. It consists of three core components. Stochastic Reverse Alignment (SRA) guides adversarial optimization via single-step reverse prediction, avoiding costly full-trajectory backpropagation. Adaptive Re-noising Augmentation (ARA) mitigates diffusion stochasticity through timestep-adaptive re-noising. Self-Consistency Regularization (SCR) stabilizes optimization by promoting local temporal coherence. Extensive experiments on open-source and commercial VLMs show that PureProof consistently outperforms prior attacks against DBP and achieves strong noise resilience, revealing critical vulnerabilities in VLMs and offering broader safety implications for real-world VLM deployments, including emerging agentic settings.