Yiming Zhao

Yiming Zhao

We present a post-training quantization (PTQ) approach for Wan2.1-T2V-14B, a 14-billion-parameter text-to-video diffusion transformer, targeting the W8A8 HiFloat8 (HiF8) format on Ascend 910B NPUs. A central challenge in quantizing video DiT models is the heterogeneous activation distribution across transformer blocks: boundary blocks (the first and last few blocks) exhibit fundamentally different statistical properties from middle blocks, making uniform quantization ineffective. We conduct a systematic per-block activation analysis across all 40 WanAttentionBlocks and use the findings to motivate a boundary-protection strategy that retains the first two and last three blocks in BF16 while quantizing the remaining 35 blocks with W8A8 HiF8. The proposed PTQ method matches or marginally exceeds the BF16 baseline on all five VBench dimensions evaluated, indicating no measurable accuracy loss within the 5-prompt evaluation set. An ablation study over four protection configurations confirms that full boundary protection yields the highest average VBench score, validating the data-driven block selection. We additionally investigate quantization-aware training (QAT) as a complementary fine-tuning stage and analyze the conditions under which it fails to outperform plain PTQ on single-card hardware.

Xinyu Zou, Zhi Hu, Yiming Zhao et al.

Multi-scenario learning (MSL) enables a service provider to cater for users' fine-grained demands by separating services for different user sectors, e.g., by user's geographical region. Under each scenario there is a need to optimize multiple task-specific targets e.g., click through rate and conversion rate, known as multi-task learning (MTL). Recent solutions for MSL and MTL are mostly based on the multi-gate mixture-of-experts (MMoE) architecture. MMoE structure is typically static and its design requires domain-specific knowledge, making it less effective in handling both MSL and MTL. In this paper, we propose a novel Automatic Expert Selection framework for Multi-scenario and Multi-task search, named AESM^{2}. AESM^{2} integrates both MSL and MTL into a unified framework with an automatic structure learning. Specifically, AESM^{2} stacks multi-task layers over multi-scenario layers. This hierarchical design enables us to flexibly establish intrinsic connections between different scenarios, and at the same time also supports high-level feature extraction for different tasks. At each multi-scenario/multi-task layer, a novel expert selection algorithm is proposed to automatically identify scenario-/task-specific and shared experts for each input. Experiments over two real-world large-scale datasets demonstrate the effectiveness of AESM^{2} over a battery of strong baselines. Online A/B test also shows substantial performance gain on multiple metrics. Currently, AESM^{2} has been deployed online for serving major traffic.

Yiming Zhao, Denys Rozumnyi, Jie Song et al.

We propose a method to estimate 3D human poses from substantially blurred images. The key idea is to tackle the inverse problem of image deblurring by modeling the forward problem with a 3D human model, a texture map, and a sequence of poses to describe human motion. The blurring process is then modeled by a temporal image aggregation step. Using a differentiable renderer, we can solve the inverse problem by backpropagating the pixel-wise reprojection error to recover the best human motion representation that explains a single or multiple input images. Since the image reconstruction loss alone is insufficient, we present additional regularization terms. To the best of our knowledge, we present the first method to tackle this problem. Our method consistently outperforms other methods on significantly blurry inputs since they lack one or multiple key functionalities that our method unifies, i.e. image deblurring with sub-frame accuracy and explicit 3D modeling of non-rigid human motion.

Yu Zeng, Wenxuan Huang, Zhen Fang et al.

Multimodal Large Language Models (MLLMs) have advanced VQA and now support Vision-DeepResearch systems that use search engines for complex visual-textual fact-finding. However, evaluating these visual and textual search abilities is still difficult, and existing benchmarks have two major limitations. First, existing benchmarks are not visual search-centric: answers that should require visual search are often leaked through cross-textual cues in the text questions or can be inferred from the prior world knowledge in current MLLMs. Second, overly idealized evaluation scenario: On the image-search side, the required information can often be obtained via near-exact matching against the full image, while the text-search side is overly direct and insufficiently challenging. To address these issues, we construct the Vision-DeepResearch benchmark (VDR-Bench) comprising 2,000 VQA instances. All questions are created via a careful, multi-stage curation pipeline and rigorous expert review, designed to assess the behavior of Vision-DeepResearch systems under realistic real-world conditions. Moreover, to address the insufficient visual retrieval capabilities of current MLLMs, we propose a simple multi-round cropped-search workflow. This strategy is shown to effectively improve model performance in realistic visual retrieval scenarios. Overall, our results provide practical guidance for the design of future multimodal deep-research systems. The code will be released in https://github.com/Osilly/Vision-DeepResearch.

Qingnan Ren, Shun Zou, Shiting Huang et al.

As autonomous coding agents become capable of handling increasingly long-horizon tasks, they have gradually demonstrated the potential to complete end-to-end software development. Although existing benchmarks have recently evolved from localized code editing to from-scratch project generation, they remain confined to structurally simplified, single-stack applications. Consequently, they fail to capture the heterogeneous environments, full-stack orchestration, and system-level complexity of real enterprise Software as a Service (SaaS) systems, leaving a critical gap in assessing agents under realistic engineering constraints. To fill this gap, we introduce SaaSBench, the first benchmark designed to explore the boundaries of AI agents in enterprise SaaS engineering. Spanning 30 complex tasks across 6 SaaS domains with 5,370 validation nodes, it incorporates 8 programming languages, 6 databases, and 13 frameworks to meticulously mirror real-world software heterogeneity. Furthermore, we design a dependency-aware hybrid evaluation paradigm tailored for complex systems with long horizons and multi-component coupling, enabling fine-grained, reproducible assessment. Crucially, our extensive experiments reveal a striking insight: the primary bottleneck for state-of-the-art agents is not generating isolated code logic, but successfully configuring and integrating a multi-component system. Over 95\% of task failures occur before agents even reach deep business logic, with models often falling victim to overconfidence and prematurely halting during foundational system setup, or getting trapped in ineffective debugging loops. We hope SaaSBench serves as a practical and challenging testbed to drive the evolution of reliable, system-level coding agents. The code is available at \url{https://github.com/ShadeCloak/SaaSbench}.

Guo Pu, Yiming Zhao, Zhouhui Lian

Recent single-view 3D generative methods have made significant advancements by leveraging knowledge distilled from extensive 3D object datasets. However, challenges persist in the synthesis of 3D scenes from a single view, primarily due to the complexity of real-world environments and the limited availability of high-quality prior resources. In this paper, we introduce a novel approach called Pano2Room, designed to automatically reconstruct high-quality 3D indoor scenes from a single panoramic image. These panoramic images can be easily generated using a panoramic RGBD inpainter from captures at a single location with any camera. The key idea is to initially construct a preliminary mesh from the input panorama, and iteratively refine this mesh using a panoramic RGBD inpainter while collecting photo-realistic 3D-consistent pseudo novel views. Finally, the refined mesh is converted into a 3D Gaussian Splatting field and trained with the collected pseudo novel views. This pipeline enables the reconstruction of real-world 3D scenes, even in the presence of large occlusions, and facilitates the synthesis of photo-realistic novel views with detailed geometry. Extensive qualitative and quantitative experiments have been conducted to validate the superiority of our method in single-panorama indoor novel synthesis compared to the state-of-the-art. Our code and data are available at \url{https://github.com/TrickyGo/Pano2Room}.

Yuyuan Yang, Junkun Hong, Hongrong Wang et al.

Embodied task planning demands vision-language models to generate action sequences that are both visually grounded and causally coherent over time. However, existing training paradigms face a critical trade-off: joint end-to-end training often leads to premature temporal binding, while standard reinforcement learning methods suffer from optimization instability. To bridge this gap, we present Staged Vision-Language Learning (SVLL), a unified three-stage framework for robust, physically-grounded embodied planning. In the first two stages, SVLL decouples spatial grounding from temporal reasoning, establishing robust visual dependency before introducing sequential action history. In the final stage, we identify a key limitation of standard Direct Preference Optimization (DPO), its purely relative nature -- optimizing only the preference gap between winning and losing trajectories while neglecting absolute likelihood constraints on optimal path, often yields unsafe or hallucinated behaviors. To address this, we further introduce Bias-DPO, a novel alignment objective that injects an inductive bias toward expert trajectories by explicitly maximizing likelihood on ground-truth actions while penalizing overconfident hallucinations. By anchoring the policy to the expert manifold and mitigating causal misalignment, SVLL, powered by Bias-DPO, ensures strict adherence to environmental affordances and effectively suppresses physically impossible shortcuts. Finally, extensive experiments on the interactive AI2-THOR benchmark and real-world robotic deployments demonstrate that SVLL outperforms both state-of-the-art open-source (e.g., Qwen2.5-VL-7B) and closed-source models (e.g., GPT-4o, Gemini-2.0-flash) in task success rate, while significantly reducing physical constraint violations.

Huanyu Bian, Zhilong Jia, Menghan Dou et al.

With the rapid development of classical and quantum machine learning, a large number of machine learning frameworks have been proposed. However, existing machine learning frameworks usually only focus on classical or quantum, rather than both. Therefore, based on VQNet 1.0, we further propose VQNet 2.0, a new generation of unified classical and quantum machine learning framework that supports hybrid optimization. The core library of the framework is implemented in C++, and the user level is implemented in Python, and it supports deployment on quantum and classical hardware. In this article, we analyze the development trend of the new generation machine learning framework and introduce the design principles of VQNet 2.0 in detail: unity, practicality, efficiency, and compatibility, as well as full particulars of implementation. We illustrate the functions of VQNet 2.0 through several basic applications, including classical convolutional neural networks, quantum autoencoders, hybrid classical-quantum networks, etc. After that, through extensive experiments, we demonstrate that the operation speed of VQNet 2.0 is higher than the comparison method. Finally, through extensive experiments, we demonstrate that VQNet 2.0 can deploy on different hardware platforms, the overall calculation speed is faster than the comparison method. It also can be mixed and optimized with quantum circuits composed of multiple quantum computing libraries.

Lin Bai, Yiming Zhao, Xinming Huang

Point cloud semantic segmentation has attracted attentions due to its robustness to light condition. This makes it an ideal semantic solution for autonomous driving. However, considering the large computation burden and bandwidth demanding of neural networks, putting all the computing into vehicle Electronic Control Unit (ECU) is not efficient or practical. In this paper, we proposed a light weighted point cloud semantic segmentation network based on range view. Due to its simple pre-processing and standard convolution, it is efficient when running on deep learning accelerator like DPU. Furthermore, a near sensor computing system is built for autonomous vehicles. In this system, a FPGA-based deep learning accelerator core (DPU) is placed next to the LiDAR sensor, to perform point cloud pre-processing and segmentation neural network. By leaving only the post-processing step to ECU, this solution heavily alleviate the computation burden of ECU and consequently shortens the decision making and vehicles reaction latency. Our semantic segmentation network achieved 10 frame per second (fps) on Xilinx DPU with computation efficiency 42.5 GOP/W.

Yiming Zhao, Taein Kwon, Paul Streli et al.

Touch contact and pressure are essential for understanding how humans interact with and manipulate objects, insights which can significantly benefit applications in mixed reality and robotics. However, estimating these interactions from an egocentric camera perspective is challenging, largely due to the lack of comprehensive datasets that provide both accurate hand poses on contacting surfaces and detailed annotations of pressure information. In this paper, we introduce EgoPressure, a novel egocentric dataset that captures detailed touch contact and pressure interactions. EgoPressure provides high-resolution pressure intensity annotations for each contact point and includes accurate hand pose meshes obtained through our proposed multi-view, sequence-based optimization method processing data from an 8-camera capture rig. Our dataset comprises 5 hours of recorded interactions from 21 participants captured simultaneously by one head-mounted and seven stationary Kinect cameras, which acquire RGB images and depth maps at 30 Hz. To support future research and benchmarking, we present several baseline models for estimating applied pressure on external surfaces from RGB images, with and without hand pose information. We further explore the joint estimation of the hand mesh and applied pressure. Our experiments demonstrate that pressure and hand pose are complementary for understanding hand-object interactions. ng of hand-object interactions in AR/VR and robotics research. Project page: \url{https://yiming-zhao.github.io/EgoPressure/}.

Yiming Zhao, Yuanpeng Gao, Yuxuan Luo et al.

AI-assisted graphic design has emerged as a powerful tool for automating the creation and editing of design elements such as posters, banners, and advertisements. While diffusion-based text-to-image models have demonstrated strong capabilities in visual content generation, their text rendering performance, particularly for small-scale typography and non-Latin scripts, remains limited. In this paper, we propose UTDesign, a unified framework for high-precision stylized text editing and conditional text generation in design images, supporting both English and Chinese scripts. Our framework introduces a novel DiT-based text style transfer model trained from scratch on a synthetic dataset, capable of generating transparent RGBA text foregrounds that preserve the style of reference glyphs. We further extend this model into a conditional text generation framework by training a multi-modal condition encoder on a curated dataset with detailed text annotations, enabling accurate, style-consistent text synthesis conditioned on background images, prompts, and layout specifications. Finally, we integrate our approach into a fully automated text-to-design (T2D) pipeline by incorporating pre-trained text-to-image (T2I) models and an MLLM-based layout planner. Extensive experiments demonstrate that UTDesign achieves state-of-the-art performance among open-source methods in terms of stylistic consistency and text accuracy, and also exhibits unique advantages compared to proprietary commercial approaches. Code and data for this paper are available at https://github.com/ZYM-PKU/UTDesign.

Qisheng Su, Zhen Fang, Shiting Huang et al.

Recent development of agents has renewed demand for long-context reasoning capacity of LLMs. However, training LLMs for this capacity requires costly long-document curation or heuristic context synthesis. We observe that agents produce massive trajectories when solving problems, invoking tools and receiving environment observations across many turns. The evidence needed to answer the original question is thus scattered throughout these turns, requiring integration of distant context segments. Nevertheless, standard agent SFT masks tool responses and only trains turn-level tool selection, creating a supervision blind spot where these scattered signals go unused. We propose Agent Context Compilation (ACC), which converts trajectories from search, software engineering, and database querying agents into long-context QA pairs that combine the original question with tool responses and environment observations gathered across multiple turns, training the model to answer directly without tool use. This makes the dependencies between the question and the evidence explicit, enabling direct supervision of long-context reasoning over distant segments without additional annotation. ACC is a simple but effective approach that can be combined with any existing long-context extension or training method, providing scalable supervised fine-tuning data. We validate ACC on long-range dependency modeling tasks through MRCR and GraphWalks, challenging benchmarks requiring cross-turn coreference resolution and graph traversal over extended contexts. Training Qwen3-30B-A3B with ACC achieves 68.3 on MRCR (+18.1) and 77.5 on GraphWalks (+7.6), results comparable to Qwen3-235B-A22B, while preserving general capabilities on GPQA, MMLU-Pro, AIME, and IFEval. Further mechanism analysis reveals that the ACC-trained model exhibits task-adaptive attention restructuring and expert specialization.

Yiming Zhao, Zhouhui Lian

Text-to-Image (T2I) generation methods based on diffusion model have garnered significant attention in the last few years. Although these image synthesis methods produce visually appealing results, they frequently exhibit spelling errors when rendering text within the generated images. Such errors manifest as missing, incorrect or extraneous characters, thereby severely constraining the performance of text image generation based on diffusion models. To address the aforementioned issue, this paper proposes a novel approach for text image generation, utilizing a pre-trained diffusion model (i.e., Stable Diffusion [27]). Our approach involves the design and training of a light-weight character-level text encoder, which replaces the original CLIP encoder and provides more robust text embeddings as conditional guidance. Then, we fine-tune the diffusion model using a large-scale dataset, incorporating local attention control under the supervision of character-level segmentation maps. Finally, by employing an inference stage refinement process, we achieve a notably high sequence accuracy when synthesizing text in arbitrarily given images. Both qualitative and quantitative results demonstrate the superiority of our method to the state of the art. Furthermore, we showcase several potential applications of the proposed UDiffText, including text-centric image synthesis, scene text editing, etc. Code and model will be available at https://github.com/ZYM-PKU/UDiffText .

Ziao Zhang, Kou Shi, Shiting Huang et al.

As the capability frontier of autonomous agents continues to expand, they are increasingly able to complete specialized tasks through plug-and-play external skills. Yet current benchmarks mostly test whether models can use provided skills, leaving open whether they can discover skills from experience, repair them after failure, and maintain a coherent library over time. We introduce SkillFlow, a benchmark of 166 tasks across 20 families in which task construction within each family follows a Domain-Agnostic Execution Flow (DAEF) that defines an agent workflow framework, allowing these tasks to share a consistent workflow. Agents are evaluated under an Agentic Lifelong Learning protocol in which they begin without skills, solve tasks sequentially within each family, externalize lessons through trajectory- and rubric-driven skill patches, and carry the updated library forward. Experiments reveal a substantial capability gap. For Claude Opus 4.6, lifelong skill evolution improves task success from 62.65% to 71.08% (+8.43 points). However, high skill usage does not necessarily imply high utility: Kimi K2.5 gains only +0.60 points despite 66.87% skill usage, while Qwen-Coder-Next reaches only a 44.58% task completion rate and still regresses relative to the vanilla setting. SkillFlow contributes a structured testbed for this direction and an in-depth empirical analysis of skill discovery, patching, transfer, and their failure modes under lifelong evaluation.

Yiming Zhao, Dewen Guo, Zhouhui Lian et al.

To bridge the gap between artists and non-specialists, we present a unified framework, Neural-Polyptych, to facilitate the creation of expansive, high-resolution paintings by seamlessly incorporating interactive hand-drawn sketches with fragments from original paintings. We have designed a multi-scale GAN-based architecture to decompose the generation process into two parts, each responsible for identifying global and local features. To enhance the fidelity of semantic details generated from users' sketched outlines, we introduce a Correspondence Attention module utilizing our Reference Bank strategy. This ensures the creation of high-quality, intricately detailed elements within the artwork. The final result is achieved by carefully blending these local elements while preserving coherent global consistency. Consequently, this methodology enables the production of digital paintings at megapixel scale, accommodating diverse artistic expressions and enabling users to recreate content in a controlled manner. We validate our approach to diverse genres of both Eastern and Western paintings. Applications such as large painting extension, texture shuffling, genre switching, mural art restoration, and recomposition can be successfully based on our framework.

Jiwei Tang, Jin Xu, Tingwei Lu et al.

Large language models (LLMs) demonstrate exceptional capabilities in various scenarios. However, they suffer from much redundant information and are sensitive to the position of key information in long context scenarios. To address these challenges, we present Perception Compressor, a training-free prompt compression framework. It includes a perception retriever that leverages guiding questions and instruction to retrieve the most relevant demonstrations, a dual-slope ratio allocator to dynamically allocate compression ratios and open-book ratios, and a semi-guided iterative compression that retains key information at the token level while removing tokens that distract the LLM. We conduct extensive experiments on long context benchmarks, i.e., NaturalQuestions, LongBench, and MuSiQue. Experiment results show that Perception Compressor outperforms existing methods by a large margin, achieving state-of-the-art performance.

Yiming Zhao, Yu Zeng, Wenxuan Huang et al.

Large Vision-Language Models (LVLMs) have shown significant progress in video understanding, yet they face substantial challenges in tasks requiring precise spatiotemporal localization at the instance level. Existing methods primarily rely on text prompts for human-model interaction, but these prompts struggle to provide precise spatial and temporal references, resulting in poor user experience. Furthermore, current approaches typically decouple visual perception from language reasoning, centering reasoning around language rather than visual content, which limits the model's ability to proactively perceive fine-grained visual evidence. To address these challenges, we propose VideoSeeker, a novel paradigm for instance-level video understanding through visual prompts. VideoSeeker seamlessly integrates agentic reasoning with instance-level video understanding tasks, enabling the model to proactively perceive and retrieve relevant video segments on demand. We construct a four-stage fully automated data synthesis pipeline to efficiently generate large-scale, high-quality instance-level video data. We internalize tool-calling and proactive perception capabilities into the model via cold-start supervision and RL training, building a powerful video understanding model. Experiments demonstrate that our model achieves an average improvement of +13.7% over baselines on instance-level video understanding tasks, surpassing powerful closed-source models such as GPT-4o and Gemini-2.5-Pro, while also showing effective transferability on general video understanding benchmarks. The relevant datasets and code will be released publicly.

Chengsi Yao, Ge Wang, Kai Kang et al.

Generative Control Policies (GCPs) show immense promise in robotic manipulation but struggle to simultaneously model stable global motions and high-frequency local corrections. While modern architectures extract multi-scale spatial features, their underlying Probability Flow ODEs apply a uniform temporal integration schedule. Compressed to a single step for real-time Receding Horizon Control (RHC), uniform ODE solvers mathematically smooth over sparse, high-frequency transients entangled within low-frequency steady states. To decouple these dynamics without accumulating pipelined errors, we introduce KoopmanFlow, a parameter-efficient generative policy guided by a Koopman-inspired structural inductive bias. Operating in a unified multimodal latent space with visual context, KoopmanFlow bifurcates generation at the terminal stage. Because visual conditioning occurs before spectral decomposition, both branches are visually guided yet temporally specialized. A macroscopic branch anchors slow-varying trajectories via single-step Consistency Training, while a transient branch uses Flow Matching to isolate high-frequency residuals stimulated by sudden visual cues (e.g., contacts or occlusions). Guided by an explicit spectral prior and optimized via a novel asymmetric consistency objective, KoopmanFlow establishes a fused co-training mechanism. This allows the variant branch to absorb localized dynamics without multi-stage error accumulation. Extensive experiments show KoopmanFlow significantly outperforms state-of-the-art baselines in contact-rich tasks requiring agile disturbance rejection. By trading a surplus latency buffer for a richer structural prior, KoopmanFlow achieves superior control fidelity and parameter efficiency within real-time deployment limits.

Wanxin Tian, Shijie Zhang, Kevin Zhang et al.

Self-evolution, the ability of agents to autonomously improve their reasoning and behavior, is essential for the embodied domain with long-horizon, real-world tasks. Despite current advancements in reinforcement fine-tuning (RFT) showing strong performance in enhancing reasoning in LLMs, its potential to enable self-evolving embodied intelligence with multi-modal interactions remains largely unexplored. Specifically, reinforcement fine-tuning faces two fundamental obstacles in embodied settings: (i) the lack of accessible intermediate rewards in multi-step reasoning tasks limits effective learning signals, and (ii) reliance on hand-crafted reward functions restricts generalization to novel tasks and environments. To address these challenges, we present Self-Evolving Embodied Agents-R1, SEEA-R1, the first RFT framework designed for enabling the self-evolving capabilities of embodied agents. Specifically, to convert sparse delayed rewards into denser intermediate signals that improve multi-step reasoning, we propose Tree-based group relative policy optimization (Tree-GRPO) integrates Monte Carlo Tree Search into GRPO. To generalize reward estimation across tasks and scenes, supporting autonomous adaptation and reward-driven self-evolution, we further introduce Multi-modal Generative Reward Model (MGRM). To holistically evaluate the effectiveness of SEEA-R1, we evaluate on the ALFWorld benchmark, surpassing state-of-the-art methods with scores of 85.07% (textual) and 46.27% (multi-modal), outperforming prior models including GPT-4o. SEEA-R1 also achieves scores of 80.3% (textual) and 44.03% (multi-modal) without ground truth reward, surpassing all open-source baselines and highlighting its scalability as a self-evolving embodied agent. Additional experiments and qualitative analysis further support the potential of SEEA-R1 for future research in scalable embodied intelligence.

Yiming Zhao, Yu Zeng, Yukun Qi et al.

Large Vision-Language Models (LVLMs) have made significant progress in the field of video understanding recently. However, current benchmarks uniformly lean on text prompts for evaluation, which often necessitate complex referential language and fail to provide precise spatial and temporal references. This limitation diminishes the experience and efficiency of human-model interaction. To address this limitation, we propose the Video Visual Prompt Benchmark(V2P-Bench), a comprehensive benchmark specifically designed to evaluate LVLMs' video understanding capabilities in multimodal human-model interaction scenarios. V2P-Bench includes 980 unique videos and 1,172 QA pairs, covering 5 main tasks and 12 dimensions, facilitating instance-level fine-grained understanding aligned with human cognition. Benchmarking results reveal that even the most powerful models perform poorly on V2P-Bench (65.4% for GPT-4o and 67.9% for Gemini-1.5-Pro), significantly lower than the human experts' 88.3%, highlighting the current shortcomings of LVLMs in understanding video visual prompts. We hope V2P-Bench will serve as a foundation for advancing multimodal human-model interaction and video understanding evaluation. Project page: https://github.com/gaotiexinqu/V2P-Bench.

Mingcong Lei, Ge Wang, Yiming Zhao et al.

Large Language Models (LLMs) exhibit remarkable capabilities in the hierarchical decomposition of complex tasks through semantic reasoning. However, their application in embodied systems faces challenges in ensuring reliable execution of subtask sequences and achieving one-shot success in long-term task completion. To address these limitations in dynamic environments, we propose Closed-Loop Embodied Agent (CLEA) -- a novel architecture incorporating four specialized open-source LLMs with functional decoupling for closed-loop task management. The framework features two core innovations: (1) Interactive task planner that dynamically generates executable subtasks based on the environmental memory, and (2) Multimodal execution critic employing an evaluation framework to conduct a probabilistic assessment of action feasibility, triggering hierarchical re-planning mechanisms when environmental perturbations exceed preset thresholds. To validate CLEA's effectiveness, we conduct experiments in a real environment with manipulable objects, using two heterogeneous robots for object search, manipulation, and search-manipulation integration tasks. Across 12 task trials, CLEA outperforms the baseline model, achieving a 67.3% improvement in success rate and a 52.8% increase in task completion rate. These results demonstrate that CLEA significantly enhances the robustness of task planning and execution in dynamic environments.

Rui Ge, Yichao Fu, Yuyang Qian et al.

Large language models are increasingly deployed as autonomous agents that must plan, act, and recover from mistakes through long-horizon interaction with environments that provide rich feedback. However, prevailing outcome-driven post-training methods (e.g., RL with verifiable rewards) primarily optimize final success signals, leaving rich environment feedback underutilized. Consequently, they often lead to distribution sharpening: the policy becomes better at reproducing a narrow set of already-successful behaviors, while failing to improve the feedback-grounded agency needed to expand problem-solving capacity (e.g., Pass@k) in long-horizon settings. To address this, we propose LEAFE (Learning Feedback-Grounded Agency from Reflective Experience), a framework that internalizes recovery agency from reflective experience. Specifically, during exploration, the agent summarizes environment feedback into actionable experience, backtracks to earlier decision points, and explores alternative branches with revised actions. We then distill these experience-guided corrections into the model through supervised fine-tuning, enabling the policy to recover more effectively in future interactions. Across a diverse set of interactive coding and agentic tasks under fixed interaction budgets, LEAFE consistently improves Pass@1 over the base model and achieves higher Pass@k than outcome-driven baselines (GRPO) and experience-based methods such as Early Experience, with gains of up to 14% on Pass@128.

Zhen Fang, Wenxuan Huang, Yu Zeng et al.

Existing Flow Matching (FM) text-to-image models suffer from two critical bottlenecks under multi-task alignment: the reward sparsity induced by scalar-valued rewards, and the gradient interference arising from jointly optimizing heterogeneous objectives, which together give rise to a 'seesaw effect' of competing metrics and pervasive reward hacking. Inspired by the success of On-Policy Distillation (OPD) in the large language model community, we propose Flow-OPD, the first unified post-training framework that integrates on-policy distillation into Flow Matching models. Flow-OPD adopts a two-stage alignment strategy: it first cultivates domain-specialized teacher models via single-reward GRPO fine-tuning, allowing each expert to reach its performance ceiling in isolation; it then establishes a robust initial policy through a Flow-based Cold-Start scheme and seamlessly consolidates heterogeneous expertise into a single student via a three-step orchestration of on-policy sampling, task-routing labeling, and dense trajectory-level supervision. We further introduce Manifold Anchor Regularization (MAR), which leverages a task-agnostic teacher to provide full-data supervision that anchors generation to a high-quality manifold, effectively mitigating the aesthetic degradation commonly observed in purely RL-driven alignment. Built upon Stable Diffusion 3.5 Medium, Flow-OPD raises the GenEval score from 63 to 92 and the OCR accuracy from 59 to 94, yielding an overall improvement of roughly 10 points over vanilla GRPO, while preserving image fidelity and human-preference alignment and exhibiting an emergent 'teacher-surpassing' effect. These results establish Flow-OPD as a scalable alignment paradigm for building generalist text-to-image models.

Yu Zeng, Wenxuan Huang, Shiting Huang et al.

Although current large Vision-Language Models (VLMs) have advanced in multimodal understanding and reasoning, their fundamental perceptual and reasoning abilities remain limited. Specifically, even on simple jigsaw tasks, existing VLMs perform near randomly, revealing deficiencies in core perception and reasoning capabilities. While high-quality vision-language data can enhance these capabilities, its scarcity and limited scalability impose significant constraints. To address this, we propose AGILE, an Agentic jiGsaw Interaction Learning for Enhancing visual perception and reasoning in VLMs. AGILE formulates jigsaw solving as an interactive process, enabling the model to progressively engage with the environment. At each step, the model generates executable code to perform an action based on the current state, while the environment provides fine-grained visual feedback to guide task completion. Through this iterative cycle of observation and interaction, the model incrementally improves its perceptual and reasoning capabilities via exploration and feedback. Experimental results show that AGILE not only substantially boosts performance on jigsaw tasks of varying complexity (e.g., increasing accuracy from 9.5% to 82.8% under the 2 $\times$ 2 setting) but also demonstrates strong generalization across 9 general vision tasks, achieving an average improvement of 3.1%. These results indicate notable enhancements in both perceptual and reasoning abilities. This work opens a new avenue for advancing reasoning and generalization in multimodal models and provides an efficient, scalable solution to the scarcity of multimodal reinforcement learning data. The code and datasets is available at https://github.com/yuzeng0-0/AGILE .

Tianfei Ren, Zhipeng Yan, Yiming Zhao et al.

While text-to-image models have made strong progress in visual fidelity, faithfully realizing complex visual intents remains challenging because many requirements must be tracked across grounding, generation, and verification. We refer to these requirements as semantic commitments and formalize their lifecycle discontinuity as the Conceptual Rift, where commitments may be locally resolved or checked but fail to remain identifiable as the same operational units throughout the generation lifecycle. To address this, we propose SCOPE, a specification-guided skill orchestration framework that maintains semantic commitments in an evolving structured specification and conditionally invokes retrieval, reasoning, and repair skills around unresolved or violated commitments. To evaluate commitment-level intent realization, we introduce Gen-Arena, a human-annotated benchmark with entity- and constraint-level specifications, together with Entity-Gated Intent Pass Rate (EGIP), a strict entity-first pass criterion. SCOPE substantially outperforms all evaluated baselines on Gen-Arena, achieving 0.60 EGIP, and further achieves strong results on WISE-V (0.907) and MindBench (0.61), demonstrating the effectiveness of persistent commitment tracking for complex image generation.

Jiayuan Du, Yuebing Song, Yiming Zhao et al.

End-to-End autonomous driving (E2E-AD) systems face challenges in lifelong learning, including catastrophic forgetting, difficulty in knowledge transfer across diverse scenarios, and spurious correlations between unobservable confounders and true driving intents. To address these issues, we propose DeLL, a Deconfounded Lifelong Learning framework that integrates a Dirichlet process mixture model (DPMM) with the front-door adjustment mechanism from causal inference. The DPMM is employed to construct two dynamic knowledge spaces: a trajectory knowledge space for clustering explicit driving behaviors and an implicit feature knowledge space for discovering latent driving abilities. Leveraging the non-parametric Bayesian nature of DPMM, our framework enables adaptive expansion and incremental updating of knowledge without predefining the number of clusters, thereby mitigating catastrophic forgetting. Meanwhile, the front-door adjustment mechanism utilizes the DPMM-derived knowledge as valid mediators to deconfound spurious correlations, such as those induced by sensor noise or environmental changes, and enhances the causal expressiveness of the learned representations. Additionally, we introduce an evolutionary trajectory decoder that enables non-autoregressive planning. To evaluate the lifelong learning performance of E2E-AD, we propose new evaluation protocols and metrics based on Bench2Drive. Extensive evaluations in the closed-loop CARLA simulator demonstrate that our framework significantly improves adaptability to new driving scenarios and overall driving performance, while effectively retaining previous acquired knowledge.

Niclas Pokel, Yiming Zhao, Pehuén Moure et al.

Personalizing Automatic Speech Recognition (ASR) for non-normative speech remains challenging because data collection is labor-intensive and model training is technically complex. To address these limitations, we propose Adapt4Me, a web-based decentralized environment that operationalizes Bayesian active learning to enable end-to-end personalization without expert supervision. The app exposes data selection, adaptation, and validation to lay users through a three-stage human-in-the-loop workflow: (1) rapid profiling via greedy phoneme sampling to capture speaker-specific acoustics; (2) backend personalization using Variational Inference Low-Rank Adaptation (VI-LoRA) to enable fast, incremental updates; and (3) continuous improvement, where users guide model refinement by resolving visualized model uncertainty via low-friction top-k corrections. By making epistemic uncertainty explicit, Adapt4Me reframes data efficiency as an interactive design feature rather than a purely algorithmic concern. We show how this enables users to personalize robust ASR models, transforming them from passive data sources into active authors of their own assistive technology.

Yukun Qi, Yiming Zhao, Yu Zeng et al.

The advancement of Chain-of-Thought (CoT) reasoning has significantly enhanced the capabilities of large language models (LLMs) and large vision-language models (LVLMs). However, a rigorous evaluation framework for video CoT reasoning remains absent. Current video benchmarks fail to adequately assess the reasoning process and expose whether failures stem from deficiencies in perception or reasoning capabilities. Therefore, we introduce VCR-Bench, a novel benchmark designed to comprehensively evaluate LVLMs' Video Chain-of-Thought Reasoning capabilities. VCR-Bench comprises 859 videos spanning a variety of video content and durations, along with 1,034 high-quality question-answer pairs. Each pair is manually annotated with a stepwise CoT rationale, where every step is tagged to indicate its association with the perception or reasoning capabilities. Furthermore, we design seven distinct task dimensions and propose the CoT score to assess the entire CoT process based on the stepwise tagged CoT rationals. Extensive experiments on VCR-Bench highlight substantial limitations in current LVLMs. Even the top-performing model, o1, only achieves a 62.8% CoT score and an 56.7% accuracy, while most models score below 40%. Experiments show most models score lower on perception than reasoning steps, revealing LVLMs' key bottleneck in temporal-spatial information processing for complex video reasoning. A robust positive correlation between the CoT score and accuracy confirms the validity of our evaluation framework and underscores the critical role of CoT reasoning in solving complex video reasoning tasks. We hope VCR-Bench to serve as a standardized evaluation framework and expose the actual drawbacks in complex video reasoning task.

Yifan Pu, Yiming Zhao, Zhicong Tang et al.

Multi-layer image generation is a fundamental task that enables users to isolate, select, and edit specific image layers, thereby revolutionizing interactions with generative models. In this paper, we introduce the Anonymous Region Transformer (ART), which facilitates the direct generation of variable multi-layer transparent images based on a global text prompt and an anonymous region layout. Inspired by Schema theory suggests that knowledge is organized in frameworks (schemas) that enable people to interpret and learn from new information by linking it to prior knowledge.}, this anonymous region layout allows the generative model to autonomously determine which set of visual tokens should align with which text tokens, which is in contrast to the previously dominant semantic layout for the image generation task. In addition, the layer-wise region crop mechanism, which only selects the visual tokens belonging to each anonymous region, significantly reduces attention computation costs and enables the efficient generation of images with numerous distinct layers (e.g., 50+). When compared to the full attention approach, our method is over 12 times faster and exhibits fewer layer conflicts. Furthermore, we propose a high-quality multi-layer transparent image autoencoder that supports the direct encoding and decoding of the transparency of variable multi-layer images in a joint manner. By enabling precise control and scalable layer generation, ART establishes a new paradigm for interactive content creation.

Mingcong Lei, Yiming Zhao, Ge Wang et al.

A key objective of embodied intelligence is enabling agents to perform long-horizon tasks in dynamic environments while maintaining robust decision-making and adaptability. To achieve this goal, we propose the Spatio-Temporal Memory Agent (STMA), a novel framework designed to enhance task planning and execution by integrating spatio-temporal memory. STMA is built upon three critical components: (1) a spatio-temporal memory module that captures historical and environmental changes in real time, (2) a dynamic knowledge graph that facilitates adaptive spatial reasoning, and (3) a planner-critic mechanism that iteratively refines task strategies. We evaluate STMA in the TextWorld environment on 32 tasks, involving multi-step planning and exploration under varying levels of complexity. Experimental results demonstrate that STMA achieves a 31.25% improvement in success rate and a 24.7% increase in average score compared to the state-of-the-art model. The results highlight the effectiveness of spatio-temporal memory in advancing the memory capabilities of embodied agents.

Yiming Zhao, Jiwei Tang, Shimin Di et al.

Recommending event schedules is a key issue in Event-based Social Networks (EBSNs) in order to maintain user activity. An effective recommendation is required to maximize the user's preference, subjecting to both time and geographical constraints. Existing methods face an inherent trade-off among efficiency, effectiveness, and generalization, due to the NP-hard nature of the problem. This paper proposes the Chain-of-Scheduling (CoS) framework, which activates the event scheduling capability of Large Language Models (LLMs) through a guided, efficient scheduling process. CoS enhances LLM by formulating the schedule task into three atomic stages, i.e., exploration, verification and integration. Then we enable the LLMs to generate CoS autonomously via Knowledge Distillation (KD). Experimental results show that CoS achieves near-theoretical optimal effectiveness with high efficiency on three real-world datasets in a interpretable manner. Moreover, it demonstrates strong zero-shot learning ability on out-of-domain data.

Jiayuan Du, Yiming Zhao, Zhenglong Guo et al.

This paper introduces a novel architecture for trajectory-conditioned forecasting of future 3D scene occupancy. In contrast to methods that rely on variational autoencoders (VAEs) to generate discrete occupancy tokens, which inherently limit representational capacity, our approach predicts multi-frame future occupancy in an end-to-end manner directly from raw image features. Inspired by the success of attention-based transformer architectures in foundational vision and language models such as GPT and VGGT, we employ a sparse occupancy representation that bypasses the intermediate bird's eye view (BEV) projection and its explicit geometric priors. This design allows the transformer to capture spatiotemporal dependencies more effectively. By avoiding both the finite-capacity constraint of discrete tokenization and the structural limitations of BEV representations, our method achieves state-of-the-art performance on the nuScenes benchmark for 1-3 second occupancy forecasting, outperforming existing approaches by a significant margin. Furthermore, it demonstrates robust scene dynamics understanding, consistently delivering high accuracy under arbitrary future trajectory conditioning.

Mingcong Lei, Honghao Cai, Zezhou Cui et al.

Embodied agents face persistent challenges in real-world environments, including partial observability, limited spatial reasoning, and high-latency multi-memory integration. We present RoboMemory, a brain-inspired framework that unifies Spatial, Temporal, Episodic, and Semantic memory under a parallelized architecture for efficient long-horizon planning and interactive environmental learning. A dynamic spatial knowledge graph (KG) ensures scalable and consistent memory updates, while a closed-loop planner with a critic module supports adaptive decision-making in dynamic settings. Experiments on EmbodiedBench show that RoboMemory, built on Qwen2.5-VL-72B-Ins, improves average success rates by 25% over its baseline and exceeds the closed-source state-of-the-art (SOTA) Gemini-1.5-Pro by 3%. Real-world trials further confirm its capacity for cumulative learning, with performance improving across repeated tasks. These results highlight RoboMemory as a scalable foundation for memory-augmented embodied intelligence, bridging the gap between cognitive neuroscience and robotic autonomy.

Yiming Zhao, Guorong Li, Laiyun Qing et al.

Open-world object counting leverages the robust text-image alignment of pre-trained vision-language models (VLMs) to enable counting of arbitrary categories in images specified by textual queries. However, widely adopted naive fine-tuning strategies concentrate exclusively on text-image consistency for categories contained in training, which leads to limited generalizability for unseen categories. In this work, we propose a plug-and-play Semantic-Driven Visual Prompt Tuning framework (SDVPT) that transfers knowledge from the training set to unseen categories with minimal overhead in parameters and inference time. First, we introduce a two-stage visual prompt learning strategy composed of Category-Specific Prompt Initialization (CSPI) and Topology-Guided Prompt Refinement (TGPR). The CSPI generates category-specific visual prompts, and then TGPR distills latent structural patterns from the VLM's text encoder to refine these prompts. During inference, we dynamically synthesize the visual prompts for unseen categories based on the semantic correlation between unseen and training categories, facilitating robust text-image alignment for unseen categories. Extensive experiments integrating SDVPT with all available open-world object counting models demonstrate its effectiveness and adaptability across three widely used datasets: FSC-147, CARPK, and PUCPR+.

Zeyu Liu, Weicong Liang, Yiming Zhao et al.

Recently, Glyph-ByT5 has achieved highly accurate visual text rendering performance in graphic design images. However, it still focuses solely on English and performs relatively poorly in terms of visual appeal. In this work, we address these two fundamental limitations by presenting Glyph-ByT5-v2 and Glyph-SDXL-v2, which not only support accurate visual text rendering for 10 different languages but also achieve much better aesthetic quality. To achieve this, we make the following contributions: (i) creating a high-quality multilingual glyph-text and graphic design dataset consisting of more than 1 million glyph-text pairs and 10 million graphic design image-text pairs covering nine other languages, (ii) building a multilingual visual paragraph benchmark consisting of 1,000 prompts, with 100 for each language, to assess multilingual visual spelling accuracy, and (iii) leveraging the latest step-aware preference learning approach to enhance the visual aesthetic quality. With the combination of these techniques, we deliver a powerful customized multilingual text encoder, Glyph-ByT5-v2, and a strong aesthetic graphic generation model, Glyph-SDXL-v2, that can support accurate spelling in 10 different languages. We perceive our work as a significant advancement, considering that the latest DALL-E3 and Ideogram 1.0 still struggle with the multilingual visual text rendering task.

Yiming Zhao, Xiao Zhang, Xinming Huang

Clustering objects from the LiDAR point cloud is an important research problem with many applications such as autonomous driving. To meet the real-time requirement, existing research proposed to apply the connected-component-labeling (CCL) technique on LiDAR spherical range image with a heuristic condition to check if two neighbor points are connected. However, LiDAR range image is different from a binary image which has a deterministic condition to tell if two pixels belong to the same component. The heuristic condition used on the LiDAR range image only works empirically, which suggests the LiDAR clustering algorithm should be robust to potential failures of the empirical heuristic condition. To overcome this challenge, this paper proposes a divide-and-merge LiDAR clustering algorithm. This algorithm firstly conducts clustering in each evenly divided local region, then merges the local clustered small components by voting on edge point pairs. Assuming there are $N$ LiDAR points of objects in total with $m$ divided local regions, the time complexity of the proposed algorithm is $O(N)+O(m^2)$. A smaller $m$ means the voting will involve more neighbor points, but the time complexity will become larger. So the $m$ controls the trade-off between the time complexity and the clustering accuracy. A proper $m$ helps the proposed algorithm work in real-time as well as maintain good performance. We evaluate the divide-and-merge clustering algorithm on the SemanticKITTI panoptic segmentation benchmark by cascading it with a state-of-the-art semantic segmentation model. The final performance evaluated through the leaderboard achieves the best among all published methods. The proposed algorithm is implemented with C++ and wrapped as a python function. It can be easily used with the modern deep learning framework in python.

Yiming Zhao, Lin Bai, Xinming Huang

Projecting the point cloud on the 2D spherical range image transforms the LiDAR semantic segmentation to a 2D segmentation task on the range image. However, the LiDAR range image is still naturally different from the regular 2D RGB image; for example, each position on the range image encodes the unique geometry information. In this paper, we propose a new projection-based LiDAR semantic segmentation pipeline that consists of a novel network structure and an efficient post-processing step. In our network structure, we design a FID (fully interpolation decoding) module that directly upsamples the multi-resolution feature maps using bilinear interpolation. Inspired by the 3D distance interpolation used in PointNet++, we argue this FID module is a 2D version distance interpolation on $(θ, φ)$ space. As a parameter-free decoding module, the FID largely reduces the model complexity by maintaining good performance. Besides the network structure, we empirically find that our model predictions have clear boundaries between different semantic classes. This makes us rethink whether the widely used K-nearest-neighbor post-processing is still necessary for our pipeline. Then, we realize the many-to-one mapping causes the blurring effect that some points are mapped into the same pixel and share the same label. Therefore, we propose to process those occluded points by assigning the nearest predicted label to them. This NLA (nearest label assignment) post-processing step shows a better performance than KNN with faster inference speed in the ablation study. On the SemanticKITTI dataset, our pipeline achieves the best performance among all projection-based methods with $64 \times 2048$ resolution and all point-wise solutions. With a ResNet-34 as the backbone, both the training and testing of our model can be finished on a single RTX 2080 Ti with 11G memory. The code is released.

Yiming Zhao, Xiao Zhang, Xinming Huang

LiDAR panoptic segmentation is a newly proposed technical task for autonomous driving. In contrast to popular end-to-end deep learning solutions, we propose a hybrid method with an existing semantic segmentation network to extract semantic information and a traditional LiDAR point cloud cluster algorithm to split each instance object. We argue geometry-based traditional clustering algorithms are worth being considered by showing a state-of-the-art performance among all published end-to-end deep learning solutions on the panoptic segmentation leaderboard of the SemanticKITTI dataset. To our best knowledge, we are the first to attempt the point cloud panoptic segmentation with clustering algorithms. Therefore, instead of working on new models, we give a comprehensive technical survey in this paper by implementing four typical cluster methods and report their performances on the benchmark. Those four cluster methods are the most representative ones with real-time running speed. They are implemented with C++ in this paper and then wrapped as a python function for seamless integration with the existing deep learning frameworks. We release our code for peer researchers who might be interested in this problem.

Lin Bai, Yiming Zhao, Xinming Huang

Light Detection And Ranging (LiDAR) has been widely used in autonomous vehicles for perception and localization. However, the cost of a high-resolution LiDAR is still prohibitively expensive, while its low-resolution counterpart is much more affordable. Therefore, using low-resolution LiDAR for autonomous driving is an economically viable solution, but the point cloud sparsity makes it extremely challenging. In this paper, we propose a two-stage neural network framework that enables 3D object detection using a low-resolution LiDAR. Taking input from a low-resolution LiDAR point cloud and a monocular camera image, a depth completion network is employed to produce dense point cloud that is subsequently processed by a voxel-based network for 3D object detection. Evaluated with KITTI dataset for 3D object detection in Bird-Eye View (BEV), the experimental result shows that the proposed approach performs significantly better than directly applying the 16-line LiDAR point cloud for object detection. For both easy and moderate cases, our 3D vehicle detection results are close to those using 64-line high-resolution LiDARs.

Yiming Zhao, Xinming Huang, Ziming Zhang

Estimating homography to align image pairs captured by different sensors or image pairs with large appearance changes is an important and general challenge for many computer vision applications. In contrast to others, we propose a generic solution to pixel-wise align multimodal image pairs by extending the traditional Lucas-Kanade algorithm with networks. The key contribution in our method is how we construct feature maps, named as deep Lucas-Kanade feature map (DLKFM). The learned DLKFM can spontaneously recognize invariant features under various appearance-changing conditions. It also has two nice properties for the Lucas-Kanade algorithm: (1) The template feature map keeps brightness consistency with the input feature map, thus the color difference is very small while they are well-aligned. (2) The Lucas-Kanade objective function built on DLKFM has a smooth landscape around ground truth homography parameters, so the iterative solution of the Lucas-Kanade can easily converge to the ground truth. With those properties, directly updating the Lucas-Kanade algorithm on our feature maps will precisely align image pairs with large appearance changes. We share the datasets, code, and demo video online.

Yiming Zhao, Lin Bai, Ziming Zhang et al.

LiDAR depth completion is a task that predicts depth values for every pixel on the corresponding camera frame, although only sparse LiDAR points are available. Most of the existing state-of-the-art solutions are based on deep neural networks, which need a large amount of data and heavy computations for training the models. In this letter, a novel non-learning depth completion method is proposed by exploiting the local surface geometry that is enhanced by an outlier removal algorithm. The proposed surface geometry model is inspired by the observation that most pixels with unknown depth have a nearby LiDAR point. Therefore, it is assumed those pixels share the same surface with the nearest LiDAR point, and their respective depth can be estimated as the nearest LiDAR depth value plus a residual error. The residual error is calculated by using a derived equation with several physical parameters as input, including the known camera intrinsic parameters, estimated normal vector, and offset distance on the image plane. The proposed method is further enhanced by an outlier removal algorithm that is designed to remove incorrectly mapped LiDAR points from occluded regions. On KITTI dataset, the proposed solution achieves the best error performance among all existing non-learning methods and is comparable to the best self-supervised learning method and some supervised learning methods. Moreover, since outlier points from occluded regions is a commonly existing problem, the proposed outlier removal algorithm is a general preprocessing step that is applicable to many robotic systems with both camera and LiDAR sensors.