Heng Zhou

Yuhao Wu, Maojia Song, Yihuai Lan et al.

Understanding the physical structure is essential for real-world applications such as embodied agents, interactive design, and long-horizon manipulation. Yet, prevailing Vision-Language Model (VLM) evaluations still center on structure-agnostic, single-turn setups (e.g., VQA), which fail to assess agents' ability to reason about how geometry, contact, and support relations jointly constrain what actions are possible in a dynamic environment. To address this gap, we introduce the Causal Hierarchy of Actions and Interactions (CHAIN) benchmark, an interactive 3D, physics-driven testbed designed to evaluate whether models can understand, plan, and execute structured action sequences grounded in physical constraints. CHAIN shifts evaluation from passive perception to active problem solving, spanning tasks such as interlocking mechanical puzzles and 3D stacking and packing. We conduct a comprehensive study of state-of-the-art VLMs and diffusion-based models under unified interactive settings. Our results show that top-performing models still struggle to internalize physical structure and causal constraints, often failing to produce reliable long-horizon plans and cannot robustly translate perceived structure into effective actions. The project is available at https://social-ai-studio.github.io/CHAIN/.

Heng Zhou, Chunna Tian, Zhenxi Zhang et al.

RGB-Thermal salient object detection (SOD) combines two spectra to segment visually conspicuous regions in images. Most existing methods use boundary maps to learn the sharp boundary. These methods ignore the interactions between isolated boundary pixels and other confident pixels, leading to sub-optimal performance. To address this problem,we propose a position-aware relation learning network (PRLNet) for RGB-T SOD based on swin transformer. PRLNet explores the distance and direction relationships between pixels to strengthen intra-class compactness and inter-class separation, generating salient object masks with clear boundaries and homogeneous regions. Specifically, we develop a novel signed distance map auxiliary module (SDMAM) to improve encoder feature representation, which takes into account the distance relation of different pixels in boundary neighborhoods. Then, we design a feature refinement approach with directional field (FRDF), which rectifies features of boundary neighborhood by exploiting the features inside salient objects. FRDF utilizes the directional information between object pixels to effectively enhance the intra-class compactness of salient regions. In addition, we constitute a pure transformer encoder-decoder network to enhance multispectral feature representation for RGB-T SOD. Finally, we conduct quantitative and qualitative experiments on three public benchmark datasets.The results demonstrate that our proposed method outperforms the state-of-the-art methods.

Heng Zhou, Li Kang, Yiran Qin et al.

Understanding the world from distributed, partial viewpoints is a fundamental challenge for embodied multi-agent systems. Each agent perceives the environment through an ego-centric view that is often limited by occlusion and ambiguity. To study this problem, we introduce the Ego-to-World (E2W) benchmark, which evaluates a vision-language model's ability to fuse heterogeneous viewpoints across three tasks: (i) global counting, (ii) relational location reasoning, and (iii) action-oriented grasping that requires predicting view-specific image coordinates. To address this setting, we propose CoRL, a two-stage framework that combines Chain-of-Thought supervised fine-tuning with reinforcement learning using Group-Relative Policy Optimization. Its core component, the Cross-View Spatial Reward (CVSR), provides dense task-aligned feedback by linking reasoning steps to visual evidence, ensuring coherent cross-view entity resolution, and guiding the model toward correct final predictions. Experiments on E2W show that CoRL consistently surpasses strong proprietary and open-source baselines on both reasoning and perception-grounding metrics, while ablations further confirm the necessity of each CVSR component. Beyond that, CoRL generalizes to external spatial reasoning benchmarks and enables effective real-world multi-robot manipulation with calibrated multi-camera rigs, demonstrating cross-view localization and successful grasp-and-place execution. Together, E2W and CoRL provide a principled foundation for learning world-centric scene understanding from distributed, ego-centric observations, advancing collaborative embodied AI.

Heng Zhou, Xiaoxiong Liu, Zhenxi Zhang et al.

Remote sensing images (RSIs) are frequently degraded by haze, fog, and thin clouds, which obscure surface reflectance and hinder downstream applications. This study presents the first systematic and unified survey of RSIs dehazing, integrating methodological evolution, benchmark assessment, and physical consistency analysis. We categorize existing approaches into a three-stage progression: from handcrafted physical priors, to data-driven deep restoration, and finally to hybrid physical-intelligent generation, and summarize more than 30 representative methods across CNNs, GANs, Transformers, and diffusion models. To provide a reliable empirical reference, we conduct large-scale quantitative experiments on five public datasets using 12 metrics, including PSNR, SSIM, CIEDE, LPIPS, FID, SAM, ERGAS, UIQI, QNR, NIQE, and HIST. Cross-domain comparison reveals that recent Transformer- and diffusion-based models improve SSIM by 12%~18% and reduce perceptual errors by 20%~35% on average, while hybrid physics-guided designs achieve higher radiometric stability. A dedicated physical radiometric consistency experiment further demonstrates that models with explicit transmission or airlight constraints reduce color bias by up to 27%. Based on these findings, we summarize open challenges: dynamic atmospheric modeling, multimodal fusion, lightweight deployment, data scarcity, and joint degradations, and outline promising research directions for future development of trustworthy, controllable, and efficient (TCE) dehazing systems. All reviewed resources, including source code, benchmark datasets, evaluation metrics, and reproduction configurations are publicly available at https://github.com/VisionVerse/RemoteSensing-Restoration-Survey.

Heng Zhou, Chunna Tian, Zhenxi Zhang et al.

A key challenge of infrared small target segmentation (ISTS) is to balance false negative pixels (FNs) and false positive pixels (FPs). Traditional methods combine FNs and FPs into a single objective by weighted sum, and the optimization process is decided by one actor. Minimizing FNs and FPs with the same strategy leads to antagonistic decisions. To address this problem, we propose a competitive game framework (pixelGame) from a novel perspective for ISTS. In pixelGame, FNs and FPs are controlled by different player whose goal is to minimize their own utility function. FNs-player and FPs-player are designed with different strategies: One is to minimize FNs and the other is to minimize FPs. The utility function drives the evolution of the two participants in competition. We consider the Nash equilibrium of pixelGame as the optimal solution. In addition, we propose maximum information modulation (MIM) to highlight the tar-get information. MIM effectively focuses on the salient region including small targets. Extensive experiments on two standard public datasets prove the effectiveness of our method. Compared with other state-of-the-art methods, our method achieves better performance in terms of F1-measure (F1) and the intersection of union (IoU).

Li Kang, Heng Zhou, Xiufeng Song et al.

Recent advancements in multimodal large language models and vision-languageaction models have significantly driven progress in Embodied AI. As the field transitions toward more complex task scenarios, multi-agent system frameworks are becoming essential for achieving scalable, efficient, and collaborative solutions. This shift is fueled by three primary factors: increasing agent capabilities, enhancing system efficiency through task delegation, and enabling advanced human-agent interactions. To address the challenges posed by multi-agent collaboration, we propose the Multi-Agent Robotic System (MARS) Challenge, held at the NeurIPS 2025 Workshop on SpaVLE. The competition focuses on two critical areas: planning and control, where participants explore multi-agent embodied planning using vision-language models (VLMs) to coordinate tasks and policy execution to perform robotic manipulation in dynamic environments. By evaluating solutions submitted by participants, the challenge provides valuable insights into the design and coordination of embodied multi-agent systems, contributing to the future development of advanced collaborative AI systems.

Xiaofang Yang, Lijun Li, Heng Zhou et al.

Recent years have witnessed increasing interest in extending large language models into agentic systems. While the effectiveness of agents has continued to improve, efficiency, which is crucial for real-world deployment, has often been overlooked. This paper therefore investigates efficiency from three core components of agents: memory, tool learning, and planning, considering costs such as latency, tokens, steps, etc. Aimed at conducting comprehensive research addressing the efficiency of the agentic system itself, we review a broad range of recent approaches that differ in implementation yet frequently converge on shared high-level principles including but not limited to bounding context via compression and management, designing reinforcement learning rewards to minimize tool invocation, and employing controlled search mechanisms to enhance efficiency, which we discuss in detail. Accordingly, we characterize efficiency in two complementary ways: comparing effectiveness under a fixed cost budget, and comparing cost at a comparable level of effectiveness. This trade-off can also be viewed through the Pareto frontier between effectiveness and cost. From this perspective, we also examine efficiency oriented benchmarks by summarizing evaluation protocols for these components and consolidating commonly reported efficiency metrics from both benchmark and methodological studies. Moreover, we discuss the key challenges and future directions, with the goal of providing promising insights.

Guibin Zhang, Hejia Geng, Xiaohang Yu et al.

The emergence of agentic reinforcement learning (Agentic RL) marks a paradigm shift from conventional reinforcement learning applied to large language models (LLM RL), reframing LLMs from passive sequence generators into autonomous, decision-making agents embedded in complex, dynamic worlds. This survey formalizes this conceptual shift by contrasting the degenerate single-step Markov Decision Processes (MDPs) of LLM-RL with the temporally extended, partially observable Markov decision processes (POMDPs) that define Agentic RL. Building on this foundation, we propose a comprehensive twofold taxonomy: one organized around core agentic capabilities, including planning, tool use, memory, reasoning, self-improvement, and perception, and the other around their applications across diverse task domains. Central to our thesis is that reinforcement learning serves as the critical mechanism for transforming these capabilities from static, heuristic modules into adaptive, robust agentic behavior. To support and accelerate future research, we consolidate the landscape of open-source environments, benchmarks, and frameworks into a practical compendium. By synthesizing over five hundred recent works, this survey charts the contours of this rapidly evolving field and highlights the opportunities and challenges that will shape the development of scalable, general-purpose AI agents.

Yawen Xiang, Heng Zhou, Chengyang Li et al.

Motion deblurring is one of the fundamental problems of computer vision and has received continuous attention. The variability in blur, both within and across images, imposes limitations on non-blind deblurring techniques that rely on estimating the blur kernel. As a response, blind motion deblurring has emerged, aiming to restore clear and detailed images without prior knowledge of the blur type, fueled by the advancements in deep learning methodologies. Despite strides in this field, a comprehensive synthesis of recent progress in deep learning-based blind motion deblurring is notably absent. This paper fills that gap by providing an exhaustive overview of the role of deep learning in blind motion deblurring, encompassing datasets, evaluation metrics, and methods developed over the last six years. Specifically, we first introduce the types of motion blur and the fundamental principles of deblurring. Next, we outline the shortcomings of traditional non-blind deblurring algorithms, emphasizing the advantages of employing deep learning techniques for deblurring tasks. Following this, we categorize and summarize existing blind motion deblurring methods based on different backbone networks, including convolutional neural networks, generative adversarial networks, recurrent neural networks, and Transformer networks. Subsequently, we elaborate not only on the fundamental principles of these different categories but also provide a comprehensive summary and comparison of their advantages and limitations. Qualitative and quantitative experimental results conducted on four widely used datasets further compare the performance of SOTA methods. Finally, an analysis of present challenges and future pathways. All collected models, benchmark datasets, source code links, and codes for evaluation have been made publicly available at https://github.com/VisionVerse/Blind-Motion-Deblurring-Survey

Heng Zhou, Zelin Tan, Zhemeng Zhang et al.

When an LLM-based agent improves on a task, is the gain from the model itself or from the reasoning paradigm wrapped around it? We study this question by comparing six inference-time paradigms, namely Direct, CoT, ReAct, Plan-Execute, Reflection, and ReCode, across four frontier LLMs and ten benchmarks, yielding roughly 18,000 runs. We find that reasoning structure helps dramatically on some tasks but hurts on others: ReAct improves over Direct by 44pp on GAIA, while CoT degrades performance by 15pp on HumanEval. No single paradigm dominates, and oracle per-task selection beats the best fixed paradigm by 17.1pp on average. Motivated by this complementarity, we propose a select-then-solve approach: before answering each task, a lightweight embedding-based router selects the most suitable paradigm. Across four models, the router improves average accuracy from 47.6% to 53.1%, outperforming the best fixed paradigm at 50.3% by 2.8pp and recovering up to 37% of the oracle gap. In contrast, zero-shot self-routing only works for GPT-5 at 67.1% and fails for weaker models, all trailing the learned router. Our results argue that reasoning paradigm selection should be a per-task decision made by a learned router, not a fixed architectural choice.

Yiran Qin, Jiahua Ma, Li Kang et al.

Recent advancements in foundational models, such as large language models and world models, have greatly enhanced the capabilities of robotics, enabling robots to autonomously perform complex tasks. However, acquiring large-scale, high-quality training data for robotics remains a challenge, as it often requires substantial manual effort and is limited in its coverage of diverse real-world environments. To address this, we propose a novel hybrid approach called Compositional Simulation, which combines classical simulation and neural simulation to generate accurate action-video pairs while maintaining real-world consistency. Our approach utilizes a closed-loop real-sim-real data augmentation pipeline, leveraging a small amount of real-world data to generate diverse, large-scale training datasets that cover a broader spectrum of real-world scenarios. We train a neural simulator to transform classical simulation videos into real-world representations, improving the accuracy of policy models trained in real-world environments. Through extensive experiments, we demonstrate that our method significantly reduces the sim2real domain gap, resulting in higher success rates in real-world policy model training. Our approach offers a scalable solution for generating robust training data and bridging the gap between simulated and real-world robotics.

Ao Sun, Hongtao Zhang, Heng Zhou et al.

Linear Attention Large Language Models (LLMs) offer a compelling recurrent formulation that compresses context into a fixed-size state matrix, enabling constant-time inference. However, the internal dynamics of this compressed state remain largely opaque. In this work, we present a comprehensive study on the runtime state dynamics of state-of-the-art Linear Attention models. We uncover a fundamental phenomenon termed State Rank Stratification, characterized by a distinct spectral bifurcation among linear attention heads: while one group maintains an effective rank oscillating near zero, the other exhibits rapid growth that converges to an upper bound. Extensive experiments across diverse inference contexts reveal that these dynamics remain strikingly consistent, indicating that the identity of a head,whether low-rank or high-rank,is an intrinsic structural property acquired during pre-training, rather than a transient state dependent on the input data. Furthermore, our diagnostic probes reveal a surprising functional divergence: low-rank heads are indispensable for model reasoning, whereas high-rank heads exhibit significant redundancy. Leveraging this insight, we propose Joint Rank-Norm Pruning, a zero-shot strategy that achieves a 38.9\% reduction in KV-cache overhead while largely maintaining model accuracy.

Yuchen Dong, Heng Zhou, Chengyang Li et al.

Camouflage object detection (COD) poses a significant challenge due to the high resemblance between camouflaged objects and their surroundings. Although current deep learning methods have made significant progress in detecting camouflaged objects, many of them heavily rely on additional prior information. However, acquiring such additional prior information is both expensive and impractical in real-world scenarios. Therefore, there is a need to develop a network for camouflage object detection that does not depend on additional priors. In this paper, we propose a novel adaptive feature aggregation method that effectively combines multi-layer feature information to generate guidance information. In contrast to previous approaches that rely on edge or ranking priors, our method directly leverages information extracted from image features to guide model training. Through extensive experimental results, we demonstrate that our proposed method achieves comparable or superior performance when compared to state-of-the-art approaches.

Jiuqing Dong, Yongbin Gao, Heng Zhou et al.

Out-of-distribution (OOD) detection is critical for ensuring the reliability of open-world intelligent systems. Despite the notable advancements in existing OOD detection methodologies, our study identifies a significant performance drop under the scarcity of training samples. In this context, we introduce a novel few-shot OOD detection benchmark, carefully constructed to address this gap. Our empirical analysis reveals the superiority of ParameterEfficient Fine-Tuning (PEFT) strategies, such as visual prompt tuning and visual adapter tuning, over conventional techniques, including fully fine-tuning and linear probing tuning in the few-shot OOD detection task. Recognizing some crucial information from the pre-trained model, which is pivotal for OOD detection, may be lost during the fine-tuning process, we propose a method termed DomainSpecific and General Knowledge Fusion (DSGF). This approach is designed to be compatible with diverse fine-tuning frameworks. Our experiments show that the integration of DSGF significantly enhances the few-shot OOD detection capabilities across various methods and fine-tuning methodologies, including fully fine-tuning, visual adapter tuning, and visual prompt tuning. The code will be released.

Heng Zhou, Ao Yu, Yuchen Fan et al.

Evaluating large language models (LLMs) on question answering often relies on static benchmarks that reward memorization and understate the role of retrieval, failing to capture the dynamic nature of world knowledge. We present LiveSearchBench, an automated pipeline for constructing retrieval-dependent benchmarks from recent knowledge updates. Our method computes deltas between successive Wikidata snapshots, filters candidate triples for quality, and synthesizes natural-language questions at three levels of reasoning difficulty, each guaranteed to admit a unique, verifiable answer through SPARQL validation. The pipeline is fully automated, scalable across time, and minimizes human intervention, enabling continual regeneration of temporally grounded benchmarks. Experiments show a pronounced performance drop when models confront facts that post-date pretraining, with the gap most salient on multi-hop queries. Retrieval augmented methods and larger, instruction-tuned models provide partial gains but fail to close this recency gap. By design, LiveSearchBench shifts evaluation from static memorization toward tasks that require up-to-date retrieval and reasoning, offering a foundation for systematic, long-term assessment of LLMs under evolving knowledge.

Yinuo Liu, Zi Qian, Heng Zhou et al.

Interleaved text-and-image generation represents a significant frontier for Multimodal Large Language Models (MLLMs), offering a more intuitive way to convey complex information. Current paradigms rely on either image generation or retrieval augmentation, yet they typically treat the two as mutually exclusive paths, failing to unify factuality with creativity. We argue that the next milestone in this field is Agentic Tool Planning, where the model serves as a central controller that autonomously determines when, where, and which tools to invoke to produce interleaved responses for visual-critical queries. To systematically evaluate this paradigm, we introduce ATP-Bench, a novel benchmark comprising 7,702 QA pairs (including 1,592 VQA pairs) across eight categories and 25 visual-critical intents, featuring human-verified queries and ground truths. Furthermore, to evaluate agentic planning independent of end-to-end execution and changing tool backends, we propose a Multi-Agent MLLM-as-a-Judge (MAM) system. MAM evaluates tool-call precision, identifies missed opportunities for tool use, and assesses overall response quality without requiring ground-truth references. Our extensive experiments on 10 state-of-the-art MLLMs reveal that models struggle with coherent interleaved planning and exhibit significant variations in tool-use behavior, highlighting substantial room for improvement and providing actionable guidance for advancing interleaved generation. Dataset and code are available at https://github.com/Qwen-Applications/ATP-Bench.

Li Kang, Yutao Fan, Rui Li et al.

Multi-agent embodied systems hold promise for complex collaborative manipulation, yet face critical challenges in spatial coordination, temporal reasoning, and shared workspace awareness. Inspired by human collaboration where cognitive planning occurs separately from physical execution, we introduce the concept of compositional environment -- a synergistic integration of real-world and simulation components that enables multiple robotic agents to perceive intentions and operate within a unified decision-making space. Building on this concept, we present CoEnv, a framework that leverages simulation for safe strategy exploration while ensuring reliable real-world deployment. CoEnv operates through three stages: real-to-sim scene reconstruction that digitizes physical workspaces, VLM-driven action synthesis supporting both real-time planning with high-level interfaces and iterative planning with code-based trajectory generation, and validated sim-to-real transfer with collision detection for safe deployment. Extensive experiments on challenging multi-arm manipulation benchmarks demonstrate CoEnv's effectiveness in achieving high task success rates and execution efficiency, establishing a new paradigm for multi-agent embodied AI.

Mingrui Li, Shuhong Liu, Heng Zhou et al.

We present SGS-SLAM, the first semantic visual SLAM system based on Gaussian Splatting. It incorporates appearance, geometry, and semantic features through multi-channel optimization, addressing the oversmoothing limitations of neural implicit SLAM systems in high-quality rendering, scene understanding, and object-level geometry. We introduce a unique semantic feature loss that effectively compensates for the shortcomings of traditional depth and color losses in object optimization. Through a semantic-guided keyframe selection strategy, we prevent erroneous reconstructions caused by cumulative errors. Extensive experiments demonstrate that SGS-SLAM delivers state-of-the-art performance in camera pose estimation, map reconstruction, precise semantic segmentation, and object-level geometric accuracy, while ensuring real-time rendering capabilities.

Heng Zhou, Zhetao Guo, Shuhong Liu et al.

Monocular SLAM has received a lot of attention due to its simple RGB inputs and the lifting of complex sensor constraints. However, existing monocular SLAM systems are designed for bounded scenes, restricting the applicability of SLAM systems. To address this limitation, we propose MoD-SLAM, the first monocular NeRF-based dense mapping method that allows 3D reconstruction in real-time in unbounded scenes. Specifically, we introduce a Gaussian-based unbounded scene representation approach to solve the challenge of mapping scenes without boundaries. This strategy is essential to extend the SLAM application. Moreover, a depth estimation module in the front-end is designed to extract accurate priori depth values to supervise mapping and tracking processes. By introducing a robust depth loss term into the tracking process, our SLAM system achieves more precise pose estimation in large-scale scenes. Our experiments on two standard datasets show that MoD-SLAM achieves competitive performance, improving the accuracy of the 3D reconstruction and localization by up to 30% and 15% respectively compared with existing state-of-the-art monocular SLAM systems.

Zhenxi Zhang, Heng Zhou, Xiaoran Shi et al.

Semi-supervised segmentation presents a promising approach for large-scale medical image analysis, effectively reducing annotation burdens while achieving comparable performance. This methodology holds substantial potential for streamlining the segmentation process and enhancing its feasibility within clinical settings for translational investigations. While cross-supervised training, based on distinct co-training sub-networks, has become a prevalent paradigm for this task, addressing critical issues such as predication disagreement and label-noise suppression requires further attention and progress in cross-supervised training. In this paper, we introduce an Evidential Tri-Branch Consistency learning framework (ETC-Net) for semi-supervised medical image segmentation. ETC-Net employs three branches: an evidential conservative branch, an evidential progressive branch, and an evidential fusion branch. The first two branches exhibit complementary characteristics, allowing them to address prediction diversity and enhance training stability. We also integrate uncertainty estimation from the evidential learning into cross-supervised training, mitigating the negative impact of erroneous supervision signals. Additionally, the evidential fusion branch capitalizes on the complementary attributes of the first two branches and leverages an evidence-based Dempster-Shafer fusion strategy, supervised by more reliable and accurate pseudo-labels of unlabeled data. Extensive experiments conducted on LA, Pancreas-CT, and ACDC datasets demonstrate that ETC-Net surpasses other state-of-the-art methods for semi-supervised segmentation. The code will be made available in the near future at https://github.com/Medsemiseg.

Li Kang, Xiufeng Song, Heng Zhou et al.

Coordinating multiple embodied agents in dynamic environments remains a core challenge in artificial intelligence, requiring both perception-driven reasoning and scalable cooperation strategies. While recent works have leveraged large language models (LLMs) for multi-agent planning, a few have begun to explore vision-language models (VLMs) for visual reasoning. However, these VLM-based approaches remain limited in their support for diverse embodiment types. In this work, we introduce VIKI-Bench, the first hierarchical benchmark tailored for embodied multi-agent cooperation, featuring three structured levels: agent activation, task planning, and trajectory perception. VIKI-Bench includes diverse robot embodiments, multi-view visual observations, and structured supervision signals to evaluate reasoning grounded in visual inputs. To demonstrate the utility of VIKI-Bench, we propose VIKI-R, a two-stage framework that fine-tunes a pretrained vision-language model (VLM) using Chain-of-Thought annotated demonstrations, followed by reinforcement learning under multi-level reward signals. Our extensive experiments show that VIKI-R significantly outperforms baselines method across all task levels. Furthermore, we show that reinforcement learning enables the emergence of compositional cooperation patterns among heterogeneous agents. Together, VIKI-Bench and VIKI-R offer a unified testbed and method for advancing multi-agent, visual-driven cooperation in embodied AI systems.

Yuchen Fan, Kaiyan Zhang, Heng Zhou et al. · pku, tsinghua

We investigate the potential of large language models (LLMs) to serve as efficient simulators for agentic search tasks in reinforcement learning (RL), thereby reducing dependence on costly interactions with external search engines. To this end, we first quantify the intrinsic search capability of LLMs via structured prompting and repeated sampling, which we term Self-Search. Our results reveal that LLMs exhibit strong scaling behavior with respect to the inference budget, achieving high pass@k on question-answering benchmarks, including the challenging BrowseComp task. Building on these observations, we introduce Self-Search RL (SSRL), which enhances LLMs' Self-Search capability through format-based and rule-based rewards. SSRL enables models to iteratively refine their knowledge utilization internally, without requiring access to external tools. Empirical evaluations demonstrate that SSRL-trained policy models provide a cost-effective and stable environment for search-driven RL training, reducing reliance on external search engines and facilitating robust sim-to-real transfer. We draw the following conclusions: 1) LLMs possess world knowledge that can be effectively elicited to achieve high performance; 2) SSRL demonstrates the potential of leveraging internal knowledge to reduce hallucination; 3) SSRL-trained models integrate seamlessly with external search engines without additional effort. Our findings highlight the potential of LLMs to support more scalable RL agent training.

Yubin Hu, Kairui Wen, Heng Zhou et al.

Reconstructing accurate 3D surfaces for street-view scenarios is crucial for applications such as digital entertainment and autonomous driving simulation. However, existing street-view datasets, including KITTI, Waymo, and nuScenes, only offer noisy LiDAR points as ground-truth data for geometric evaluation of reconstructed surfaces. These geometric ground-truths often lack the necessary precision to evaluate surface positions and do not provide data for assessing surface normals. To overcome these challenges, we introduce the SS3DM dataset, comprising precise \textbf{S}ynthetic \textbf{S}treet-view \textbf{3D} \textbf{M}esh models exported from the CARLA simulator. These mesh models facilitate accurate position evaluation and include normal vectors for evaluating surface normal. To simulate the input data in realistic driving scenarios for 3D reconstruction, we virtually drive a vehicle equipped with six RGB cameras and five LiDAR sensors in diverse outdoor scenes. Leveraging this dataset, we establish a benchmark for state-of-the-art surface reconstruction methods, providing a comprehensive evaluation of the associated challenges. For more information, visit our homepage at https://ss3dm.top.

Zelin Tan, Hejia Geng, Mulei Zhang et al.

While scaling laws for large language models (LLMs) during pre-training have been extensively studied, their behavior under reinforcement learning (RL) post-training remains largely unexplored. This paper presents a systematic empirical investigation of scaling behaviors in RL-based post-training, with a particular focus on mathematical reasoning. Based on 54 experiments across diverse model sizes and training settings, we characterize how model scale, data volume, and computational budget interact to shape performance. Our analysis leads to four key findings: (1). Under a fixed computational budget, larger models trained for fewer steps consistently outperform smaller models trained for more steps. (2). Given a fixed amount of training data, larger models achieve superior sample efficiency, yielding lower loss. (3). In data-constrained regimes, repeated reuse of high-quality data proves highly effective, as final performance is primarily governed by the total number of optimization steps rather than the uniqueness of samples. (4). These scaling behaviors are robust across both base and instruction-tuned models, which share similar learning dynamics (e.g., larger models show faster convergence) even while differing in absolute accuracy. Collectively, these results provide a principled foundation and practical guidelines for efficiently scaling the reasoning capabilities of LLMs through RL post-training.

Ying Zhu, Heng Zhou, Rui Su et al.

Recently, many approaches, such as Chain-of-Thought (CoT) prompting and Multi-Agent Debate (MAD), have been proposed to further enrich Large Language Models' (LLMs) complex problem-solving capacities in reasoning scenarios. However, these methods may fail to solve complex problems due to the lack of ability to find optimal solutions. Swarm Intelligence has been serving as a powerful tool for finding optima in the field of traditional optimization problems. To this end, we propose integrating swarm intelligence into the reasoning process by introducing a novel Agent-based Swarm Intelligence (ASI) paradigm. In this paradigm, we formulate LLM reasoning as an optimization problem and use a swarm intelligence scheme to guide a group of LLM-based agents in collaboratively searching for optimal solutions. To avoid swarm intelligence getting trapped in local optima, we further develop a Swarm Intelligence Enhancing Reasoning (SIER) framework, which develops a density-driven strategy to enhance the reasoning ability. To be specific, we propose to perform kernel density estimation and non-dominated sorting to optimize both solution quality and diversity simultaneously. In this case, SIER efficiently enhances solution space exploration through expanding the diversity of the reasoning path. Besides, a step-level quality evaluation is used to help agents improve solution quality by correcting low-quality intermediate steps. Then, we use quality thresholds to dynamically control the termination of exploration and the selection of candidate steps, enabling a more flexible and efficient reasoning process. Extensive experiments are ...

Yawen Xiang, Heng Zhou, Chengyang Li et al.

Image deblurring is an essential image preprocessing technique, aiming to recover clear and detailed images form blurry ones. However, existing algorithms often fail to effectively integrate multi-scale feature extraction with frequency enhancement, limiting their ability to reconstruct fine textures. Additionally, non-uniform blur in images also restricts the effectiveness of image restoration. To address these issues, we propose a multi-scale frequency enhancement network (MFENet) for blind image deblurring. To capture the multi-scale spatial and channel information of blurred images, we introduce a multi-scale feature extraction module (MS-FE) based on depthwise separable convolutions, which provides rich target features for deblurring. We propose a frequency enhanced blur perception module (FEBP) that employs wavelet transforms to extract high-frequency details and utilizes multi-strip pooling to perceive non-uniform blur, combining multi-scale information with frequency enhancement to improve the restoration of image texture details. Experimental results on the GoPro and HIDE datasets demonstrate that the proposed method achieves superior deblurring performance in both visual quality and objective evaluation metrics. Furthermore, in downstream object detection tasks, the proposed blind image deblurring algorithm significantly improves detection accuracy, further validating its effectiveness androbustness in the field of image deblurring.

Zhenxi Zhang, Ran Ran, Chunna Tian et al.

Semi-supervised medical image segmentation offers a promising solution for large-scale medical image analysis by significantly reducing the annotation burden while achieving comparable performance. Employing this method exhibits a high degree of potential for optimizing the segmentation process and increasing its feasibility in clinical settings during translational investigations. Recently, cross-supervised training based on different co-training sub-networks has become a standard paradigm for this task. Still, the critical issues of sub-network disagreement and label-noise suppression require further attention and progress in cross-supervised training. This paper proposes a cross-supervised learning framework based on dual classifiers (DC-Net), including an evidential classifier and a vanilla classifier. The two classifiers exhibit complementary characteristics, enabling them to handle disagreement effectively and generate more robust and accurate pseudo-labels for unlabeled data. We also incorporate the uncertainty estimation from the evidential classifier into cross-supervised training to alleviate the negative effect of the error supervision signal. The extensive experiments on LA and Pancreas-CT dataset illustrate that DC-Net outperforms other state-of-the-art methods for semi-supervised segmentation. The code will be released soon.

Zhenxi Zhang, Ran Ran, Chunna Tian et al.

Consistency learning plays a crucial role in semi-supervised medical image segmentation as it enables the effective utilization of limited annotated data while leveraging the abundance of unannotated data. The effectiveness and efficiency of consistency learning are challenged by prediction diversity and training stability, which are often overlooked by existing studies. Meanwhile, the limited quantity of labeled data for training often proves inadequate for formulating intra-class compactness and inter-class discrepancy of pseudo labels. To address these issues, we propose a self-aware and cross-sample prototypical learning method (SCP-Net) to enhance the diversity of prediction in consistency learning by utilizing a broader range of semantic information derived from multiple inputs. Furthermore, we introduce a self-aware consistency learning method that exploits unlabeled data to improve the compactness of pseudo labels within each class. Moreover, a dual loss re-weighting method is integrated into the cross-sample prototypical consistency learning method to improve the reliability and stability of our model. Extensive experiments on ACDC dataset and PROMISE12 dataset validate that SCP-Net outperforms other state-of-the-art semi-supervised segmentation methods and achieves significant performance gains compared to the limited supervised training. Our code will come soon.

Heng Zhou, Zhiqiang Xu

The $p$-set, which is in a simple analytic form, is well distributed in unit cubes. The well-known Weil's exponential sum theorem presents an upper bound of the exponential sum over the $p$-set. Based on the result, one shows that the $p$-set performs well in numerical integration, in compressed sensing as well as in UQ. However, $p$-set is somewhat rigid since the cardinality of the $p$-set is a prime $p$ and the set only depends on the prime number $p$. The purpose of this paper is to present generalizations of $p$-sets, say $\mathcal{P}_{d,p}^{{\mathbf a},ε}$, which is more flexible. Particularly, when a prime number $p$ is given, we have many different choices of the new $p$-sets. Under the assumption that Goldbach conjecture holds, for any even number $m$, we present a point set, say ${\mathcal L}_{p,q}$, with cardinality $m-1$ by combining two different new $p$-sets, which overcomes a major bottleneck of the $p$-set. We also present the upper bounds of the exponential sums over $\mathcal{P}_{d,p}^{{\mathbf a},ε}$ and ${\mathcal L}_{p,q}$, which imply these sets have many potential applications.