Mingfei Han

Yuetian Weng, Mingfei Han, Haoyu He et al.

Video Semantic Segmentation (VSS) involves assigning a semantic label to each pixel in a video sequence. Prior work in this field has demonstrated promising results by extending image semantic segmentation models to exploit temporal relationships across video frames; however, these approaches often incur significant computational costs. In this paper, we propose an efficient mask propagation framework for VSS, called MPVSS. Our approach first employs a strong query-based image segmentor on sparse key frames to generate accurate binary masks and class predictions. We then design a flow estimation module utilizing the learned queries to generate a set of segment-aware flow maps, each associated with a mask prediction from the key frame. Finally, the mask-flow pairs are warped to serve as the mask predictions for the non-key frames. By reusing predictions from key frames, we circumvent the need to process a large volume of video frames individually with resource-intensive segmentors, alleviating temporal redundancy and significantly reducing computational costs. Extensive experiments on VSPW and Cityscapes demonstrate that our mask propagation framework achieves SOTA accuracy and efficiency trade-offs. For instance, our best model with Swin-L backbone outperforms the SOTA MRCFA using MiT-B5 by 4.0% mIoU, requiring only 26% FLOPs on the VSPW dataset. Moreover, our framework reduces up to 4x FLOPs compared to the per-frame Mask2Former baseline with only up to 2% mIoU degradation on the Cityscapes validation set. Code is available at https://github.com/ziplab/MPVSS.

Mingfei Han, David Junhao Zhang, Yali Wang et al.

Learning spatial-temporal relation among multiple actors is crucial for group activity recognition. Different group activities often show the diversified interactions between actors in the video. Hence, it is often difficult to model complex group activities from a single view of spatial-temporal actor evolution. To tackle this problem, we propose a distinct Dual-path Actor Interaction (DualAI) framework, which flexibly arranges spatial and temporal transformers in two complementary orders, enhancing actor relations by integrating merits from different spatiotemporal paths. Moreover, we introduce a novel Multi-scale Actor Contrastive Loss (MAC-Loss) between two interactive paths of Dual-AI. Via self-supervised actor consistency in both frame and video levels, MAC-Loss can effectively distinguish individual actor representations to reduce action confusion among different actors. Consequently, our Dual-AI can boost group activity recognition by fusing such discriminative features of different actors. To evaluate the proposed approach, we conduct extensive experiments on the widely used benchmarks, including Volleyball, Collective Activity, and NBA datasets. The proposed Dual-AI achieves state-of-the-art performance on all these datasets. It is worth noting the proposed Dual-AI with 50% training data outperforms a number of recent approaches with 100% training data. This confirms the generalization power of Dual-AI for group activity recognition, even under the challenging scenarios of limited supervision.

Yuetian Weng, Zizheng Pan, Mingfei Han et al.

The task of action detection aims at deducing both the action category and localization of the start and end moment for each action instance in a long, untrimmed video. While vision Transformers have driven the recent advances in video understanding, it is non-trivial to design an efficient architecture for action detection due to the prohibitively expensive self-attentions over a long sequence of video clips. To this end, we present an efficient hierarchical Spatio-Temporal Pyramid Transformer (STPT) for action detection, building upon the fact that the early self-attention layers in Transformers still focus on local patterns. Specifically, we propose to use local window attention to encode rich local spatio-temporal representations in the early stages while applying global attention modules to capture long-term space-time dependencies in the later stages. In this way, our STPT can encode both locality and dependency with largely reduced redundancy, delivering a promising trade-off between accuracy and efficiency. For example, with only RGB input, the proposed STPT achieves 53.6% mAP on THUMOS14, surpassing I3D+AFSD RGB model by over 10% and performing favorably against state-of-the-art AFSD that uses additional flow features with 31% fewer GFLOPs, which serves as an effective and efficient end-to-end Transformer-based framework for action detection.

Mingjie Li, Rui Liu, Guangsi Shi et al.

Long-term time-series forecasting (LTSF) is fundamental to various real-world applications, where Transformer-based models have become the dominant framework due to their ability to capture long-range dependencies. However, these models often experience overfitting due to data redundancy in rolling forecasting settings, limiting their generalization ability particularly evident in longer sequences with highly similar adjacent data. In this work, we introduce CLMFormer, a novel framework that mitigates redundancy through curriculum learning and a memory-driven decoder. Specifically, we progressively introduce Bernoulli noise to the training samples, which effectively breaks the high similarity between adjacent data points. This curriculum-driven noise introduction aids the memory-driven decoder by supplying more diverse and representative training data, enhancing the decoder's ability to model seasonal tendencies and dependencies in the time-series data. To further enhance forecasting accuracy, we introduce a memory-driven decoder. This component enables the model to capture seasonal tendencies and dependencies in the time-series data and leverages temporal relationships to facilitate the forecasting process. Extensive experiments on six real-world LTSF benchmarks show that CLMFormer consistently improves Transformer-based models by up to 30%, demonstrating its effectiveness in long-horizon forecasting.

Yuetian Weng, Mingfei Han, Haoyu He et al.

Empowered by Large Language Models (LLMs), recent advancements in Video-based LLMs (VideoLLMs) have driven progress in various video understanding tasks. These models encode video representations through pooling or query aggregation over a vast number of visual tokens, making computational and memory costs affordable. Despite successfully providing an overall comprehension of video content, existing VideoLLMs still face challenges in achieving detailed understanding due to overlooking local information in long-term videos. To tackle this challenge, we introduce LongVLM, a simple yet powerful VideoLLM for long video understanding, building upon the observation that long videos often consist of sequential key events, complex actions, and camera movements. Our approach proposes to decompose long videos into multiple short-term segments and encode local features for each segment via a hierarchical token merging module. These features are concatenated in temporal order to maintain the storyline across sequential short-term segments. Additionally, we propose to integrate global semantics into each local feature to enhance context understanding. In this way, we encode video representations that incorporate both local and global information, enabling the LLM to generate comprehensive responses for long-term videos. Experimental results on the VideoChatGPT benchmark and zero-shot video question-answering datasets demonstrate the superior capabilities of our model over the previous state-of-the-art methods. Qualitative examples show that our model produces more precise responses for long video understanding. Code is available at https://github.com/ziplab/LongVLM.

Kecheng Zhang, Zongxin Yang, Mingfei Han et al.

Visual agents operating in the wild must respond to queries precisely when sufficient evidence first appears in a video stream, a critical capability that is overlooked by conventional video LLMs evaluated in offline settings. The shift to an online, streaming paradigm introduces significant challenges: a lack of decision transparency, the difficulty of aligning response timing with visual evidence, and the need to maintain a global, causally consistent understanding under tight computational budgets. To address these issues, we propose a novel framework that decouples reasoning control from memory integration. We introduce \textbf{\model{}}, an instantiation of this framework with two core components. First, the \emph{Active Thinking Decision Maker (ATDM)} is a transparent reasoning controller that externalizes its decision process using observable progress ($\boldsymbolρ$) and confidence ($\boldsymbol{c}$) metrics. This allows it to precisely time its response $t_r$ to match the first-sufficient-evidence timestamp $t^\star$ while streaming its reasoning to the user. Second, the \emph{Hierarchical Progressive Semantic Integration (HPSI)} module acts as an efficient memory system. It employs a set of learnable, multi-level aggregation tokens that are propagated across clips to build a rich, global cognitive state without exceeding token budgets. %Our approach sets a new standard on key online video understanding benchmarks, achieving strong performance of \textbf{71.6\%} on StreamingBench and \textbf{46.9\%} on OVOBench, demonstrating a robust solution for evidence-aligned and transparent online video analysis. Extensive experiments demonstrate the effectiveness of ATDM and HPSI, e.g., Thinking-QwenVL improves the accuracy of the previous state-of-the-art from 67.63\% to 71.60\% on the StreamingBench benchmark.

Tingjun Dai, Mingfei Han, Tingwen Du et al.

Measurement of task progress through explicit, actionable milestones is critical for robust robotic manipulation. This progress awareness enables a model to ground its current task status, anticipate verifiable intermediate states, and detect and recover from failures when progress stalls. To embody this capability, we introduce See, Plan, Rewind (SPR), a progress-aware vision-language-action framework that dynamically grounds language instructions into a sequence of spatial subgoals. SPR operates through a continuous core cycle, Seeing the current state and upcoming milestone, Planning a trajectory towards the next 2D waypoint, and Rewinding to a recoverable state upon failure by monitoring progress against the expected sequence. This closed-loop approach enables robust error correction without requiring additional training data or auxiliary models. Extensive experiments demonstrate the framework's effectiveness, generalization and robustness: SPR outperforms the MolmoAct baseline by 5\% on the LIBERO benchmark. On the challenging LIBERO-Plus benchmark with unseen instructions and initial states, SPR achieves state-of-the-art robustness with the smallest performance drop, surpassing OpenVLA-OFT and UniVLA, demonstrating superior out-of-distribution robustness.

Yongxin Wang, Meng Cao, Haokun Lin et al.

Multimodal large language models (MLLMs) have achieved remarkable progress on various visual question answering and reasoning tasks leveraging instruction fine-tuning specific datasets. They can also learn from preference data annotated by human to enhance their reasoning ability and mitigate hallucinations. Most of preference data is generated from the model itself. However, existing methods require high-quality critical labels, which are costly and rely on human or proprietary models like GPT-4V. In this work, we propose Enhancing Alignment in MLLMs via Critical Observation (EACO), which aligns MLLMs by self-generated preference data using only 5k images economically. Our approach begins with collecting and refining a Scoring Evaluation Instruction-tuning dataset to train a critical evaluation model, termed the Critic. This Critic observes model responses across multiple dimensions, selecting preferred and non-preferred outputs for refined Direct Preference Optimization (DPO) tuning. To further enhance model performance, we employ an additional supervised fine-tuning stage after preference tuning. EACO reduces the overall hallucinations by 65.6% on HallusionBench and improves the reasoning ability by 21.8% on MME-Cognition. EACO achieves an 8.5% improvement over LLaVA-v1.6-Mistral-7B across multiple benchmarks. Remarkably, EACO also shows the potential critical ability in open-source MLLMs, demonstrating that EACO is a viable path to boost the competence of MLLMs.

Mingfei Han, Haihong Hao, Liang Ma et al.

Vision-and-Language Navigation (VLN) has long been constrained by the limited diversity and scalability of simulator-curated datasets, which fail to capture the complexity of real-world environments. To overcome this limitation, we introduce a large-scale video-instruction framework derived from web-based room tour videos, enabling agents to learn from natural human walking demonstrations in diverse, realistic indoor settings. Unlike existing datasets, our framework integrates both open-ended description-enriched trajectories and action-enriched trajectories reconstructed in 3D, providing richer spatial and semantic supervision. A key extension in this work is the incorporation of implicit geometry representations, which extract spatial cues directly from RGB frames without requiring fragile 3D reconstruction. This approach substantially improves data utilization, alleviates reconstruction failures, and unlocks large portions of previously unusable video data. Comprehensive experiments across multiple VLN benchmarks (CVDN, SOON, R2R, and REVERIE) demonstrate that our method not only sets new state-of-the-art performance but also enables the development of robust zero-shot navigation agents. By bridging large-scale web videos with implicit spatial reasoning, this work advances embodied navigation towards more scalable, generalizable, and real-world applicable solutions.

Meng Cao, Haoran Tang, Haoze Zhao et al.

Understanding the physical world, including object dynamics, material properties, and causal interactions, remains a core challenge in artificial intelligence. Although recent multi-modal large language models (MLLMs) have demonstrated impressive general reasoning capabilities, they still fall short of achieving human-level understanding of physical principles. Existing datasets for physical reasoning either rely on real-world videos, which incur high annotation costs, or on synthetic simulations, which suffer from limited realism and diversity. In this paper, we propose a novel paradigm that leverages glitches in gameplay videos, referring to visual anomalies that violate predefined physical laws, as a rich and scalable supervision source for physical world understanding. We introduce PhysGame, an meta information guided instruction-tuning dataset containing 140,057 glitch-centric question-answer pairs across five physical domains and sixteen fine-grained categories. To ensure data accuracy, we design a prompting strategy that utilizes gameplay metadata such as titles and descriptions to guide high-quality QA generation. Complementing PhysGame, we construct GameBench, an expert-annotated benchmark with 880 glitch-identified gameplay videos designed to evaluate physical reasoning capabilities. Extensive experiments show that PhysGame significantly enhances both Game2Real transferability, improving the real world physical reasoning performance of Qwen2.5VL by 2.5% on PhysBench, and Game2General transferability, yielding a 1.9% gain on the MVBench benchmark. Moreover, PhysGame-tuned models achieve a 3.7% absolute improvement on GameBench, demonstrating enhanced robustness in detecting physical implausibilities. These results indicate that learning from gameplay anomalies offers a scalable and effective pathway toward advancing physical world understanding in multimodal intelligence.

Jinxing Zhou, Yanghao Zhou, Mingfei Han et al.

Referring Audio-Visual Segmentation (Ref-AVS) aims to segment target objects in audible videos based on given reference expressions. Prior works typically rely on learning latent embeddings via multimodal fusion to prompt a tunable SAM/SAM2 decoder for segmentation, which requires strong pixel-level supervision and lacks interpretability. From a novel perspective of explicit reference understanding, we propose TGS-Agent, which decomposes the task into a Think-Ground-Segment process, mimicking the human reasoning procedure by first identifying the referred object through multimodal analysis, followed by coarse-grained grounding and precise segmentation. To this end, we first propose Ref-Thinker, a multimodal language model capable of reasoning over textual, visual, and auditory cues. We construct an instruction-tuning dataset with explicit object-aware think-answer chains for Ref-Thinker fine-tuning. The object description inferred by Ref-Thinker is used as an explicit prompt for Grounding-DINO and SAM2, which perform grounding and segmentation without relying on pixel-level supervision. Additionally, we introduce R\textsuperscript{2}-AVSBench, a new benchmark with linguistically diverse and reasoning-intensive references for better evaluating model generalization. Our approach achieves state-of-the-art results on both standard Ref-AVSBench and proposed R\textsuperscript{2}-AVSBench. Code will be available at https://github.com/jasongief/TGS-Agent.

Haihong Hao, Lei Chen, Mingfei Han et al.

Existing vision-and-language navigation (VLN) models primarily reason over past and current visual observations, while largely ignoring the future visual dynamics induced by actions. As a result, they often lack an effective understanding of the causal relationship between actions and how the visual world changes, limiting robust decision-making. Humans, in contrast, can imagine the near future by leveraging action-dynamics causality, which improves both environmental understanding and navigation choices. Inspired by this capability, we propose LatentPilot, a new paradigm that exploits future observations during training as a valuable data source to learn action-conditioned visual dynamics, while requiring no access to future frames at inference. Concretely, we propose a flywheel-style training mechanism that iteratively collects on-policy trajectories and retrains the model to better match the agent's behavior distribution, with an expert takeover triggered when the agent deviates excessively. LatentPilot further learns visual latent tokens without explicit supervision; these latent tokens attend globally in a continuous latent space and are carried across steps, serving as both the current output and the next input, thereby enabling the agent to dream ahead and reason about how actions will affect subsequent observations. Experiments on R2R-CE, RxR-CE, and R2R-PE benchmarks achieve new SOTA results, and real-robot tests across diverse environments demonstrate LatentPilot's superior understanding of environment-action dynamics in scene. Project page:https://abdd.top/latentpilot/

Rocktim Jyoti Das, Harsh Singh, Diana Turmakhan et al.

Scaling data and models has played a pivotal role in the remarkable progress of computer vision and language. Inspired by these domains, recent efforts in robotics have similarly focused on scaling both data and model size to develop more generalizable and robust policies. However, unlike vision and language, robotics lacks access to internet-scale demonstrations across diverse robotic tasks and environments. As a result, the scale of existing datasets typically suffers from the need for manual data collection and curation. To address this problem, here we propose BLAZER, a framework that learns manipulation policies from automatically generated training data. We build on the zero-shot capabilities of LLM planners and automatically generate demonstrations for diverse manipulation tasks in simulation. Successful examples are then used to finetune an LLM and to improve its planning capabilities without human supervision. Notably, while BLAZER training requires access to the simulator's state, we demonstrate direct transfer of acquired skills to sensor-based manipulation. Through extensive experiments, we show BLAZER to significantly improve zero-shot manipulation in both simulated and real environments. Moreover, BLAZER improves on tasks outside of its training pool and enables downscaling of LLM models. Our code and data will be made publicly available on the project page.

Longtao Jiang, Jie Huang, Mingfei Han et al.

Text-guided image inpainting aims to inpaint masked image regions based on a textual prompt while preserving the background. Although diffusion-based methods have become dominant, their property of modeling the entire image in latent space makes it challenging for the results to align well with prompt details and maintain a consistent background. To address these issues, we explore Mask AutoRegressive (MAR) models for this task. MAR naturally supports image inpainting by generating latent tokens corresponding to mask regions, enabling better local controllability without altering the background. However, directly applying MAR to this task makes the inpainting content either ignore the prompts or be disharmonious with the background context. Through analysis of the attention maps from the inpainting images, we identify the impact of background tokens on text tokens during the MAR generation, and leverage this to design \textbf{Token Painter}, a training-free text-guided image inpainting method based on MAR. Our approach introduces two key components: (1) Dual-Stream Encoder Information Fusion (DEIF), which fuses the semantic and context information from text and background in frequency domain to produce novel guidance tokens, allowing MAR to generate text-faithful inpainting content while keeping harmonious with background context. (2) Adaptive Decoder Attention Score Enhancing (ADAE), which adaptively enhances attention scores on guidance tokens and inpainting tokens to further enhance the alignment of prompt details and the content visual quality. Extensive experiments demonstrate that our training-free method outperforms prior state-of-the-art methods across almost all metrics. Codes: https://github.com/longtaojiang/Token-Painter.

Bingqian Lin, Yunshuang Nie, Khun Loun Zai et al.

Recent studies have revealed the potential of training open-source Large Language Models (LLMs) to unleash LLMs' reasoning ability for enhancing vision-language navigation (VLN) performance, and simultaneously mitigate the domain gap between LLMs' training corpus and the VLN task. However, these approaches predominantly adopt straightforward input-output mapping paradigms, causing the mapping learning difficult and the navigational decisions unexplainable. Chain-of-Thought (CoT) training is a promising way to improve both navigational decision accuracy and interpretability, while the complexity of the navigation task makes the perfect CoT labels unavailable and may lead to overfitting through pure CoT supervised fine-tuning. To address these issues, we propose EvolveNav, a novel sElf-improving embodied reasoning paradigm that realizes adaptable and generalizable navigational reasoning for boosting LLM-based vision-language Navigation. Specifically, EvolveNav involves a two-stage training process: (1) Formalized CoT Supervised Fine-Tuning, where we train the model with curated formalized CoT labels to first activate the model's navigational reasoning capabilities, and simultaneously increase the reasoning speed; (2) Self-Reflective Post-Training, where the model is iteratively trained with its own reasoning outputs as self-enriched CoT labels to enhance the supervision diversity. A self-reflective auxiliary task is also designed to encourage the model to learn correct reasoning patterns by contrasting with wrong ones. Experimental results under both task-specific and cross-task training paradigms demonstrate the consistent superiority of EvolveNav over previous LLM-based VLN approaches on various popular benchmarks, including R2R, REVERIE, CVDN, and SOON. Code is available at https://github.com/expectorlin/EvolveNav.

Mingfei Han, Linjie Yang, Xiaojun Chang et al.

A short clip of video may contain progression of multiple events and an interesting story line. A human need to capture both the event in every shot and associate them together to understand the story behind it. In this work, we present a new multi-shot video understanding benchmark Shot2Story with detailed shot-level captions, comprehensive video summaries and question-answering pairs. To facilitate better semantic understanding of videos, we provide captions for both visual signals and human narrations. We design several distinct tasks including single-shot video captioning, multi-shot video summarization, and multi-shot video question answering. Preliminary experiments show some challenges to generate a long and comprehensive video summary for multi-shot videos. Nevertheless, the generated imperfect summaries can already achieve competitive performance on existing video understanding tasks such as video question-answering, promoting an under-explored setting of video understanding with detailed summaries.

Panwen Hu, Jin Jiang, Jianqi Chen et al.

The advent of AI-Generated Content (AIGC) has spurred research into automated video generation to streamline conventional processes. However, automating storytelling video production, particularly for customized narratives, remains challenging due to the complexity of maintaining subject consistency across shots. While existing approaches like Mora and AesopAgent integrate multiple agents for Story-to-Video (S2V) generation, they fall short in preserving protagonist consistency and supporting Customized Storytelling Video Generation (CSVG). To address these limitations, we propose StoryAgent, a multi-agent framework designed for CSVG. StoryAgent decomposes CSVG into distinct subtasks assigned to specialized agents, mirroring the professional production process. Notably, our framework includes agents for story design, storyboard generation, video creation, agent coordination, and result evaluation. Leveraging the strengths of different models, StoryAgent enhances control over the generation process, significantly improving character consistency. Specifically, we introduce a customized Image-to-Video (I2V) method, LoRA-BE, to enhance intra-shot temporal consistency, while a novel storyboard generation pipeline is proposed to maintain subject consistency across shots. Extensive experiments demonstrate the effectiveness of our approach in synthesizing highly consistent storytelling videos, outperforming state-of-the-art methods. Our contributions include the introduction of StoryAgent, a versatile framework for video generation tasks, and novel techniques for preserving protagonist consistency.

Chengyou Jia, Minnan Luo, Xiaojun Chang et al.

Exploring open-vocabulary video action recognition is a promising venture, which aims to recognize previously unseen actions within any arbitrary set of categories. Existing methods typically adapt pretrained image-text models to the video domain, capitalizing on their inherent strengths in generalization. A common thread among such methods is the augmentation of visual embeddings with temporal information to improve the recognition of seen actions. Yet, they compromise with standard less-informative action descriptions, thus faltering when confronted with novel actions. Drawing inspiration from human cognitive processes, we argue that augmenting text embeddings with human prior knowledge is pivotal for open-vocabulary video action recognition. To realize this, we innovatively blend video models with Large Language Models (LLMs) to devise Action-conditioned Prompts. Specifically, we harness the knowledge in LLMs to produce a set of descriptive sentences that contain distinctive features for identifying given actions. Building upon this foundation, we further introduce a multi-modal action knowledge alignment mechanism to align concepts in video and textual knowledge encapsulated within the prompts. Extensive experiments on various video benchmarks, including zero-shot, few-shot, and base-to-novel generalization settings, demonstrate that our method not only sets new SOTA performance but also possesses excellent interpretability.

Harsh Singh, Rocktim Jyoti Das, Mingfei Han et al.

Large Language Models (LLMs) have demonstrated remarkable planning abilities across various domains, including robotics manipulation and navigation. While recent efforts in robotics have leveraged LLMs both for high-level and low-level planning, these approaches often face significant challenges, such as hallucinations in long-horizon tasks and limited adaptability due to the generation of plans in a single pass without real-time feedback. To address these limitations, we propose a novel multi-agent LLM framework, Multi-Agent Large Language Model for Manipulation (MALMM) that distributes high-level planning and low-level control code generation across specialized LLM agents, supervised by an additional agent that dynamically manages transitions. By incorporating observations from the environment after each step, our framework effectively handles intermediate failures and enables adaptive re-planning. Unlike existing methods, our approach does not rely on pre-trained skill policies or in-context learning examples and generalizes to a variety of new tasks. We evaluate our approach on nine RLBench tasks, including long-horizon tasks, and demonstrate its ability to solve robotics manipulation in a zero-shot setting, thereby overcoming key limitations of existing LLM-based manipulation methods.

Mingfei Han, Liang Ma, Kamila Zhumakhanova et al.

Vision-and-Language Navigation (VLN) suffers from the limited diversity and scale of training data, primarily constrained by the manual curation of existing simulators. To address this, we introduce RoomTour3D, a video-instruction dataset derived from web-based room tour videos that capture real-world indoor spaces and human walking demonstrations. Unlike existing VLN datasets, RoomTour3D leverages the scale and diversity of online videos to generate open-ended human walking trajectories and open-world navigable instructions. To compensate for the lack of navigation data in online videos, we perform 3D reconstruction and obtain 3D trajectories of walking paths augmented with additional information on the room types, object locations and 3D shape of surrounding scenes. Our dataset includes $\sim$100K open-ended description-enriched trajectories with $\sim$200K instructions, and 17K action-enriched trajectories from 1847 room tour environments. We demonstrate experimentally that RoomTour3D enables significant improvements across multiple VLN tasks including CVDN, SOON, R2R, and REVERIE. Moreover, RoomTour3D facilitates the development of trainable zero-shot VLN agents, showcasing the potential and challenges of advancing towards open-world navigation.

Ruiheng Liu, Haihong Hao, Mingfei Han et al.

Advancing towards artificial superintelligence requires rich and intelligent perceptual capabilities. A critical frontier in this pursuit is overcoming the limited spatial understanding of Multimodal Large Language Models (MLLMs), where geometry information is essential. Existing methods often address this by rigidly injecting geometric signals into every input, while ignoring their necessity and adding computation overhead. Contrary to this paradigm, our framework endows the model with an awareness of perceptual insufficiency, empowering it to autonomously engage geometric features in reasoning when 2D cues are deemed insufficient. To achieve this, we first introduce an independent geometry input channel to the model architecture and conduct alignment training, enabling the effective utilization of geometric features. Subsequently, to endow the model with perceptual awareness, we curate a dedicated spatial-aware supervised fine-tuning dataset. This serves to activate the model's latent internal cues, empowering it to autonomously determine the necessity of geometric information. Experiments across multiple spatial reasoning benchmarks validate this approach, demonstrating significant spatial gains without compromising 2D visual reasoning capabilities, offering a path toward more robust, efficient and self-aware multi-modal intelligence.

Mingfei Han, Linjie Yang, Xiaojie Jin et al.

The creation of new datasets often presents new challenges for video recognition and can inspire novel ideas while addressing these challenges. While existing datasets mainly comprise landscape mode videos, our paper seeks to introduce portrait mode videos to the research community and highlight the unique challenges associated with this video format. With the growing popularity of smartphones and social media applications, recognizing portrait mode videos is becoming increasingly important. To this end, we have developed the first dataset dedicated to portrait mode video recognition, namely PortraitMode-400. The taxonomy of PortraitMode-400 was constructed in a data-driven manner, comprising 400 fine-grained categories, and rigorous quality assurance was implemented to ensure the accuracy of human annotations. In addition to the new dataset, we conducted a comprehensive analysis of the impact of video format (portrait mode versus landscape mode) on recognition accuracy and spatial bias due to the different formats. Furthermore, we designed extensive experiments to explore key aspects of portrait mode video recognition, including the choice of data augmentation, evaluation procedure, the importance of temporal information, and the role of audio modality. Building on the insights from our experimental results and the introduction of PortraitMode-400, our paper aims to inspire further research efforts in this emerging research area.

Liang Ma, Jiajun Wen, Min Lin et al.

While vision-language models (VLMs) have demonstrated promising capabilities in reasoning and planning for embodied agents, their ability to comprehend physical phenomena, particularly within structured 3D environments, remains severely limited. To close this gap, we introduce PhyBlock, a progressive benchmark designed to assess VLMs on physical understanding and planning through robotic 3D block assembly tasks. PhyBlock integrates a novel four-level cognitive hierarchy assembly task alongside targeted Visual Question Answering (VQA) samples, collectively aimed at evaluating progressive spatial reasoning and fundamental physical comprehension, including object properties, spatial relationships, and holistic scene understanding. PhyBlock includes 2600 block tasks (400 assembly tasks, 2200 VQA tasks) and evaluates models across three key dimensions: partial completion, failure diagnosis, and planning robustness. We benchmark 21 state-of-the-art VLMs, highlighting their strengths and limitations in physically grounded, multi-step planning. Our empirical findings indicate that the performance of VLMs exhibits pronounced limitations in high-level planning and reasoning capabilities, leading to a notable decline in performance for the growing complexity of the tasks. Error analysis reveals persistent difficulties in spatial orientation and dependency reasoning. Surprisingly, chain-of-thought prompting offers minimal improvements, suggesting spatial tasks heavily rely on intuitive model comprehension. We position PhyBlock as a unified testbed to advance embodied reasoning, bridging vision-language understanding and real-world physical problem-solving.

Jinxing Zhou, Zhihui Li, Yongqiang Yu et al.

We present \textbf{Met}a-\textbf{T}oken \textbf{Le}arning (Mettle), a simple and memory-efficient method for adapting large-scale pretrained transformer models to downstream audio-visual tasks. Instead of sequentially modifying the output feature distribution of the transformer backbone, Mettle utilizes a lightweight \textit{Layer-Centric Distillation (LCD)} module to distill in parallel the intact audio or visual features embedded by each transformer layer into compact meta-tokens. This distillation process considers both pretrained knowledge preservation and task-specific adaptation. The obtained meta-tokens can be directly applied to classification tasks, such as audio-visual event localization and audio-visual video parsing. To further support fine-grained segmentation tasks, such as audio-visual segmentation, we introduce a \textit{Meta-Token Injection (MTI)} module, which utilizes the audio and visual meta-tokens distilled from the top transformer layer to guide feature adaptation in earlier layers. Extensive experiments on multiple audiovisual benchmarks demonstrate that our method significantly reduces memory usage and training time while maintaining parameter efficiency and competitive accuracy.

Zhiheng Liu, Xueqing Deng, Shoufa Chen et al.

Generative video modeling has made significant strides, yet ensuring structural and temporal consistency over long sequences remains a challenge. Current methods predominantly rely on RGB signals, leading to accumulated errors in object structure and motion over extended durations. To address these issues, we introduce WorldWeaver, a robust framework for long video generation that jointly models RGB frames and perceptual conditions within a unified long-horizon modeling scheme. Our training framework offers three key advantages. First, by jointly predicting perceptual conditions and color information from a unified representation, it significantly enhances temporal consistency and motion dynamics. Second, by leveraging depth cues, which we observe to be more resistant to drift than RGB, we construct a memory bank that preserves clearer contextual information, improving quality in long-horizon video generation. Third, we employ segmented noise scheduling for training prediction groups, which further mitigates drift and reduces computational cost. Extensive experiments on both diffusion- and rectified flow-based models demonstrate the effectiveness of WorldWeaver in reducing temporal drift and improving the fidelity of generated videos.

Minghao Jin, Mozheng Liao, Mingfei Han et al.

Recent world-model-based Vision-Language-Action (VLA) architectures have improved robotic manipulation through predictive visual foresight. However, dense future prediction introduces visual redundancy and accumulates errors, causing long-horizon plan drift. Meanwhile, recent sparse methods typically represent visual foresight using high-level semantic subtasks or implicit latent states. These representations often lack explicit kinematic grounding, weakening the alignment between planning and low-level execution. To address this, we propose StructVLA, which reformulates a generative world model into an explicit structured planner for reliable control. Instead of dense rollouts or semantic goals, StructVLA predicts sparse, physically meaningful structured frames. Derived from intrinsic kinematic cues (e.g., gripper transitions and kinematic turning points), these frames capture spatiotemporal milestones closely aligned with task progress. We implement this approach through a two-stage training paradigm with a unified discrete token vocabulary: the world model is first trained to predict structured frames and subsequently optimized to map the structured foresight into low-level actions. This approach provides clear physical guidance and bridges visual planning and motion control. In our experiments, StructVLA achieves strong average success rates of 75.0% on SimplerEnv-WidowX and 94.8% on LIBERO. Real-world deployments further demonstrate reliable task completion and robust generalization across both basic pick-and-place and complex long-horizon tasks.

Haihong Hao, Mingfei Han, Changlin Li et al.

Embodied navigation demands comprehensive scene understanding and precise spatial reasoning. While image-text models excel at interpreting pixel-level color and lighting cues, 3D-text models capture volumetric structure and spatial relationships. However, unified fusion approaches that jointly fuse 2D images, 3D point clouds, and textual instructions face challenges in limited availability of triple-modality data and difficulty resolving conflicting beliefs among modalities. In this work, we introduce CoNav, a collaborative cross-modal reasoning framework where a pretrained 3D-text model explicitly guides an image-text navigation agent by providing structured spatial-semantic knowledge to resolve ambiguities during navigation. Specifically, we introduce Cross-Modal Belief Alignment, which operationalizes this cross-modal guidance by simply sharing textual hypotheses from the 3D-text model to the navigation agent. Through lightweight fine-tuning on a small 2D-3D-text corpus, the navigation agent learns to integrate visual cues with spatial-semantic knowledge derived from the 3D-text model, enabling effective reasoning in embodied navigation. CoNav achieves significant improvements on four standard embodied navigation benchmarks (R2R, CVDN, REVERIE, SOON) and two spatial reasoning benchmarks (ScanQA, SQA3D). Moreover, under close navigation Success Rate, CoNav often generates shorter paths compared to other methods (as measured by SPL), showcasing the potential and challenges of fusing data from different modalities in embodied navigation. Project Page: https://oceanhao.github.io/CoNav/

Yongxin Wang, Zhicheng Yang, Meng Cao et al.

Group-relative reinforcement learning with verifiable rewards (RLVR) often wastes the most informative data it already has the failures. When all rollouts are wrong, gradients stall; when one happens to be correct, the update usually ignores why the others are close-but-wrong, and credit can be misassigned to spurious chains. We present CARE (Contrastive Anchored REflection), a failure-centric post-training framework for multimodal reasoning that turns errors into supervision. CARE combines: (i) an anchored-contrastive objective that forms a compact subgroup around the best rollout and a set of semantically proximate hard negatives, performs within-subgroup z-score normalization with negative-only scaling, and includes an all-negative rescue to prevent zero-signal batches; and (ii) Reflection-Guided Resampling (RGR), a one-shot structured self-repair that rewrites a representative failure and re-scores it with the same verifier, converting near-misses into usable positives without any test-time reflection. CARE improves accuracy and training smoothness while explicitly increasing the share of learning signal that comes from failures. On Qwen2.5-VL-7B, CARE lifts macro-averaged accuracy by 4.6 points over GRPO across six verifiable visual-reasoning benchmarks; with Qwen3-VL-8B it reaches competitive or state-of-the-art results on MathVista and MMMU-Pro under an identical evaluation protocol.

Zirui Song, Yuan Huang, Junchang Liu et al.

Social deduction games like Werewolf combine language, reasoning, and strategy, providing a testbed for studying natural language and social intelligence. However, most studies reduce the game to LLM-based self-play, yielding templated utterances and anecdotal cases that overlook the richness of social gameplay. Evaluation further relies on coarse metrics such as survival time or subjective scoring due to the lack of quality reference data. To address these gaps, we curate a high-quality, human-verified multimodal Werewolf dataset containing over 100 hours of video, 32.4M utterance tokens, and 15 rule variants. Based on this dataset, we propose a novel strategy-alignment evaluation that leverages the winning faction's strategies as ground truth in two stages: 1) Speech evaluation, formulated as multiple-choice-style tasks that assess whether the model can adopt appropriate stances across five dimensions of social ability; and 2) Decision evaluation, which assesses the model's voting choices and opponent-role inferences. This framework enables a fine-grained evaluation of models' linguistic and reasoning capabilities, while capturing their ability to generate strategically coherent gameplay. Our experiments show that state-of-the-art LLMs show diverse performance, with roughly half remain below 0.50, revealing clear gaps in deception and counterfactual reasoning. We hope our dataset further inspires research on language, reasoning, and strategy in multi-agent interaction.

Mingfei Han, Haihong Hao, Jinxing Zhou et al.

Vision-language models often hallucinate details, generating non-existent objects or inaccurate attributes that compromise output reliability. Existing methods typically address these issues via extensive human annotations or external supervision from more powerful models. In this work, we present a novel framework that leverages the model's self-consistency between long responses and short answers to generate preference pairs for training. We observe that short binary questions tend to yield highly reliable responses, which can be used to query the target model to evaluate and rank its generated responses. Specifically, we design a self-reflection pipeline where detailed model responses are compared against concise binary answers, and inconsistency signals are utilized to automatically curate high-quality training data without human annotations or external model-based supervision. By relying solely on self-consistency rather than external supervision, our method offers a scalable and efficient solution that effectively reduces hallucinations using unlabeled data. Extensive experiments on multiple benchmarks, i.e., AMBER, MultiObject-Hal (ROPE), Object HalBench, and MMHal-Bench, demonstrate significant improvements in factual grounding and reliability. Moreover, our approach maintains robust instruction-following ability, as evidenced by enhanced performance on LLaVA-Bench and MMBench.