Yunzhe Zhang

Ziyue Yang, Da Ma, Hanqi Li et al.

As scientific literature grows rapidly, automated survey generation has become a key capability for AI scientists and human researchers. However, existing systems suffer from limited analytical depth due to reliance on abstracts and isolated paper processing, and unreliable citations from imprecise retrieval and post-hoc grounding, producing superficial surveys and may mislead researchers. We present DeepSurvey, an agentic system that addresses both. To enhance depth, DeepSurvey extracts structured keynotes from full-text papers, models cross-paper relationships through clustering and comparative analysis, and integrates code-repository analysis to recover implementation-level details. To fortify reliability, it combines citation-graph expansion with hybrid filtering for topic-focussed retrieval, enforces evidence-constrained citation assignment, and deploys multi-granularity agentic refinement to validate citation-claim alignment. Experiments show that DeepSurvey achieves the highest content score (8.644/10) and citation quality (12.3% and 9.3% recall and precision gains over the strongest baseline), generalizes more robustly across domains (0.14 vs 0.22 to 0.69 CS-to-non-CS drop), and is preferred over human-written surveys by domain experts (83.3% overall quality, 100% content depth).

Haotian Xu, Shengjie Wang, Zhaolei Wang et al.

Reinforcement learning (RL) has achieved promising results on most robotic control tasks. Safety of learning-based controllers is an essential notion of ensuring the effectiveness of the controllers. Current methods adopt whole consistency constraints during the training, thus resulting in inefficient exploration in the early stage. In this paper, we propose an algorithm named Constrained Policy Optimization with Extra Safety Budget (ESB-CPO) to strike a balance between the exploration efficiency and the constraints satisfaction. In the early stage, our method loosens the practical constraints of unsafe transitions (adding extra safety budget) with the aid of a new metric we propose. With the training process, the constraints in our optimization problem become tighter. Meanwhile, theoretical analysis and practical experiments demonstrate that our method gradually meets the cost limit's demand in the final training stage. When evaluated on Safety-Gym and Bullet-Safety-Gym benchmarks, our method has shown its advantages over baseline algorithms in terms of safety and optimality. Remarkably, our method gains remarkable performance improvement under the same cost limit compared with baselines.

Yunzhe Zhang, Hongfu Liu, Pengyu Hong

Text-to-image diffusion models can synthesize high-quality images, yet the outcome is notoriously sensitive to the random seed: different initial seeds often yield large variations in image quality and prompt-image alignment. We revisit this "seed effect" and show that attention dynamics over prompt core tokens, the content-bearing words, measured during the first few denoising steps, strongly predict final generation quality. Building on this observation, we introduce Attention-Based Seed Selection (ABSS), a training-free, plug-and-play method that ranks seeds for a given prompt by leveraging cross-attention to core tokens during the denoising process. ABSS requires no finetuning and does not alter the initial noise; it scores and ranks all candidate seeds, keeps only the top-k for full generation, and discards the rest, without relying on a fixed accept/reject threshold. Operating purely at inference time, ABSS can serve as a lightweight pre-selection add-on for existing seed-optimization pipelines, enabling additional gains. Across three benchmarks, extensive experiments show that ABSS enables consistent improvements in text-image alignment and visual quality for Stable Diffusion variants, as corroborated by human preference and alignment metrics.

Fengbo Lan, Shengjie Wang, Yunzhe Zhang et al.

Achieving human-like dexterous manipulation remains a crucial area of research in robotics. Current research focuses on improving the success rate of pick-and-place tasks. Compared with pick-and-place, throwing-catching behavior has the potential to increase the speed of transporting objects to their destination. However, dynamic dexterous manipulation poses a major challenge for stable control due to a large number of dynamic contacts. In this paper, we propose a Learning-based framework for Throwing-Catching tasks using dexterous hands (LTC). Our method, LTC, achieves a 73\% success rate across 45 scenarios (diverse hand poses and objects), and the learned policies demonstrate strong zero-shot transfer performance on unseen objects. Additionally, in tasks where the object in hand faces sideways, an extremely unstable scenario due to the lack of support from the palm, all baselines fail, while our method still achieves a success rate of over 60\%.

Yaru Niu, Yunzhe Zhang, Mingyang Yu et al.

Quadrupedal robots have demonstrated impressive locomotion capabilities in complex environments, but equipping them with autonomous versatile manipulation skills in a scalable way remains a significant challenge. In this work, we introduce a cross-embodiment imitation learning system for quadrupedal manipulation, leveraging data collected from both humans and LocoMan, a quadruped equipped with multiple manipulation modes. Specifically, we develop a teleoperation and data collection pipeline, which unifies and modularizes the observation and action spaces of the human and the robot. To effectively leverage the collected data, we propose an efficient modularized architecture that supports co-training and pretraining on structured modality-aligned data across different embodiments. Additionally, we construct the first manipulation dataset for the LocoMan robot, covering various household tasks in both unimanual and bimanual modes, supplemented by a corresponding human dataset. We validate our system on six real-world manipulation tasks, where it achieves an average success rate improvement of 41.9% overall and 79.7% under out-of-distribution (OOD) settings compared to the baseline. Pretraining with human data contributes a 38.6% success rate improvement overall and 82.7% under OOD settings, enabling consistently better performance with only half the amount of robot data. Our code, hardware, and data are open-sourced at: https://human2bots.github.io.

Wanshu Fan, Yue Wang, Cong Wang et al.

Single-Image Super-Resolution (SISR) plays a pivotal role in enhancing the accuracy and reliability of measurement systems, which are integral to various vision-based instrumentation and measurement applications. These systems often require clear and detailed images for precise object detection and recognition. However, images captured by visual measurement tools frequently suffer from degradation, including blurring and loss of detail, which can impede measurement accuracy.As a potential remedy, we in this paper propose a Semantic-Guided Global-Local Collaborative Network (SGGLC-Net) for lightweight SISR. Our SGGLC-Net leverages semantic priors extracted from a pre-trained model to guide the super-resolution process, enhancing image detail quality effectively. Specifically,we propose a Semantic Guidance Module that seamlessly integrates the semantic priors into the super-resolution network, enabling the network to more adeptly capture and utilize semantic priors, thereby enhancing image details. To further explore both local and non-local interactions for improved detail rendition,we propose a Global-Local Collaborative Module, which features three Global and Local Detail Enhancement Modules, as well as a Hybrid Attention Mechanism to work together to efficiently learn more useful features. Our extensive experiments show that SGGLC-Net achieves competitive PSNR and SSIM values across multiple benchmark datasets, demonstrating higher performance with the multi-adds reduction of 12.81G compared to state-of-the-art lightweight super-resolution approaches. These improvements underscore the potential of our approach to enhance the precision and effectiveness of visual measurement systems. Codes are at https://github.com/fanamber831/SGGLC-Net.

Yunzhe Zhang, Yao Lu, Qi Xuan

Contrastive learning, a dominant self-supervised technique, emphasizes similarity in representations between augmentations of the same input and dissimilarity for different ones. Although low contrastive loss often correlates with high classification accuracy, recent studies challenge this direct relationship, spotlighting the crucial role of inductive biases. We delve into these biases from a clustering viewpoint, noting that contrastive learning creates locally dense clusters, contrasting the globally dense clusters from supervised learning. To capture this discrepancy, we introduce the "RLD (Relative Local Density)" metric. While this cluster property can hinder linear classification accuracy, leveraging a Graph Convolutional Network (GCN) based classifier mitigates this, boosting accuracy and reducing parameter requirements. The code is available \href{https://github.com/xsgxlz/How-does-Contrastive-Learning-Organize-Images/tree/main}{here}.

Yao Lu, Wen Yang, Yunzhe Zhang et al.

There are good arguments to support the claim that deep neural networks (DNNs) capture better feature representations than the previous hand-crafted feature engineering, which leads to a significant performance improvement. In this paper, we move a tiny step towards understanding the dynamics of feature representations over layers. Specifically, we model the process of class separation of intermediate representations in pre-trained DNNs as the evolution of communities in dynamic graphs. Then, we introduce modularity, a generic metric in graph theory, to quantify the evolution of communities. In the preliminary experiment, we find that modularity roughly tends to increase as the layer goes deeper and the degradation and plateau arise when the model complexity is great relative to the dataset. Through an asymptotic analysis, we prove that modularity can be broadly used for different applications. For example, modularity provides new insights to quantify the difference between feature representations. More crucially, we demonstrate that the degradation and plateau in modularity curves represent redundant layers in DNNs and can be pruned with minimal impact on performance, which provides theoretical guidance for layer pruning. Our code is available at https://github.com/yaolu-zjut/Dynamic-Graphs-Construction.

Ansh Kumar Sharma, Yixiang Sun, Ninghao Lu et al.

Robot learning from interacting with the physical world is fundamentally bottlenecked by the cost of physical interaction. The two alternatives, supervised finetuning (SFT) from expert demonstrations and reinforcement learning (RL) in a software-based simulator, are limited by the amount of expert data available and the sim-to-real gap for manipulation. With the recent emergence of world models learned from real-world video-action data, we ask the question of whether training a policy in a world model can be more effective than supervised learning or software simulation in achieving better real-robot performance. We propose World-Gymnast, which performs RL finetuning of a vision-language-action (VLA) policy by rolling out the policy in an action-conditioned video world model and rewarding the rollouts with a vision-language model (VLM). On the Bridge robot setup, World-Gymnast outperforms SFT by as much as 18x and outperforms software simulator by as much as 2x. More importantly, World-Gymnast demonstrates intriguing capabilities of RL with a world model, including training on diverse language instructions and novel scenes from the world model, test-time training in a novel scene, and online iterative world model and policy improvement. Our results suggest learning a world model and training robot policies in the cloud could be the key to bridging the gap between robots that work in demonstrations and robots that can work in anyone's household.

Yunzhe Zhang, Yifei Wang, Khanh Vinh Nguyen et al.

Current SMILES-based diffusion models for molecule generation typically support only unimodal constraint. They inject conditioning signals at the start of the training process and require retraining a new model from scratch whenever the constraint changes. However, real-world applications often involve multiple constraints across different modalities, and additional constraints may emerge over the course of a study. This raises a challenge: how to extend a pre-trained diffusion model not only to support cross-modality constraints but also to incorporate new ones without retraining. To tackle this problem, we propose the Cross-Modality Controlled Molecule Generation with Diffusion Language Model (CMCM-DLM), demonstrated by two distinct cross modalities: molecular structure and chemical properties. Our approach builds upon a pre-trained diffusion model, incorporating two trainable modules, the Structure Control Module (SCM) and the Property Control Module (PCM), and operates in two distinct phases during the generation process. In Phase I, we employs the SCM to inject structural constraints during the early diffusion steps, effectively anchoring the molecular backbone. Phase II builds on this by further introducing PCM to guide the later stages of inference to refine the generated molecules, ensuring their chemical properties match the specified targets. Experimental results on multiple datasets demonstrate the efficiency and adaptability of our approach, highlighting CMCM-DLM's significant advancement in molecular generation for drug discovery applications.