Author name cluster

Zijian Wu

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

8 papers

2 author rows

AAMAS Conference 2026 Conference Paper

HLSMAC: A New StarCraft Multi-Agent Challenge for High-Level Strategic Decision-Making

Xingxing Hong
Yungong Wang
Dexin Jin
Ye Yuan
Ximing Huang
Zijian Wu
Yirui Rao
Wenxin Li

Benchmarks are crucial for assessing multi-agent reinforcement learning (MARL) algorithms. While StarCraft II-related environments have driven significant advances in MARL, existing benchmarks like SMAC focus primarily on micromanagement, limiting comprehensiveevaluationofhigh-levelstrategicintelligence. Toaddress this, we introduce HLSMAC, a new cooperative MARL benchmark with 12 carefully designed StarCraft II scenarios based on classical stratagems from the Thirty-Six Stratagems. Each scenario corresponds to a specific stratagem and is designed to challenge agents with diverse strategic elements, including tactical maneuvering, timing coordination, and deception, thereby opening up avenues for evaluating high-level strategic decision-making capabilities. We also propose novel metrics across multiple dimensions beyond conventional win rate, such as ability utilization and advancement efficiency, to assess agents’ overall performance within the HLSMAC environment. We conduct a large-scale evaluation of 21 state-of-the-art MARL algorithms and LLM-based agents, with additional multi-seed analysis for relatively better-performing methods. The results demonstrate that HLSMAC serves as a robust testbed for advancing multi-agent strategic decision-making.

PDF

AAAI Conference 2025 Conference Paper

DepMGNN: Matrixial Graph Neural Network for Video-based Automatic Depression Assessment

Zijian Wu
Leijing Zhou
Shuanglin Li
Changzeng Fu
Jun Lu
Jing Han
Yi Zhang
Zhuang Zhao

Depression can be reflected by long-term human spatio-temporal facial behaviours. While human face videos recorded in real-world usually have long and variable lengths, existing video-based depression assessment approaches frequently re-sample/down-sample such videos to short and equal-length videos, or split each video into several equal-length segments, where segment-level spatio-temporal facial behaviours are suppressed as a vector-style representations for RNN-based long-term (video-level) modelling. Both strategies lead to crucial information loss and distortion. In this paper, we propose a novel graph-style data structure called Matrixial Graph and an effective Matrixial Graph Neural Network (MGNN) for face video-based depression assessment, which can directly and end-to-end model long-term depression-specific spatio-temporal facial cues from variable-length videos without resampling/splitting videos or suppressing video segments to vectors. Importantly, the nodes in our matrixial graph are capable of including matrices of different shapes, and thus nodes of a matrix graph can directly represent all frame-level 2D facial feature maps (or images themselves) of an entire video regardless of its length. Then, our MGNN is the first GNN that can jointly process matrixial graphs containing varying numbers of nodes, which further learns matrix-style edge features, thereby facilitating to explicit model video-level multi-scale spatio-temporal facial behaviours among matrixial graph nodes for depression assessment. Experiments show that the explicit spatio-temporal modeling on 2D facial feature maps, facilitated by our matrixial graph/MGNN, provided significant benefits, leading our approach to achieve new state-of-the-art performances on AVEC2013 and AVEC2014 datasets with large advantages.

PDF Details DOI

NeurIPS Conference 2025 Conference Paper

NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation

Xiangyan Liu
Jinjie Ni
Zijian Wu
Chao Du
Longxu Dou
Haonan Wang
Tianyu Pang
Michael Shieh

Recent advances in reinforcement learning (RL) have strengthened the reasoning capabilities of vision-language models (VLMs). However, enhancing policy exploration to better scale test-time compute remains largely underexplored. In addition, VLMs continue to struggle with imperfect visual perception, which in turn affects the subsequent reasoning process. To this end, we propose **NoisyRollout**, a simple yet effective data augmentation method that mixes trajectories from both clean and moderately distorted images during RL training. By injecting targeted diversity in visual perception and the resulting reasoning patterns, NoisyRollout promotes better policy exploration through vision-oriented inductive biases, ultimately leading to more robust reasoning behaviors. We further adopt a noise annealing schedule that gradually reduces distortion strength over training, leveraging noisy signals early on while ensuring training stability in later stages. Crucially, our method is easy-to-adopt—**requiring no additional training cost and no modifications to the RL objective**. Extensive experiments on $2$ distinct training datasets demonstrate that NoisyRollout achieves state-of-the-art performance among open-source RL-tuned models across $5$ out-of-domain reasoning and perception benchmarks. Furthermore, we validate the effectiveness of NoisyRollout across model sizes ($7$B and $32$B) and data scales (from $1$K to $6$K), highlighting its generalizability and scalability.

PDF Details

NeurIPS Conference 2025 Conference Paper

On Reasoning Strength Planning in Large Reasoning Models

Leheng Sheng
An Zhang
Zijian Wu
Weixiang Zhao
Changshuo Shen
Zhang Yi
Xiang Wang
Tat-Seng Chua

Recent studies empirically reveal that large reasoning models (LRMs) can automatically allocate more reasoning strengths (\ie the number of reasoning tokens) for harder problems, exhibiting difficulty-awareness for better task performance. While this automatic reasoning strength allocation phenomenon has been widely observed, its underlying mechanism remains largely unexplored. To this end, we provide explanations for this phenomenon from the perspective of model activations. \textbf{We find evidence that LRMs pre-plan the reasoning strengths in their activations even before generation, with this reasoning strength causally controlled by the magnitude of a pre-allocated directional vector. } Specifically, we show that the number of reasoning tokens is predictable solely based on the question activations using linear probes, indicating that LRMs estimate the required reasoning strength in advance. We then uncover that LRMs encode this reasoning strength through a pre-allocated directional vector embedded in the activations of the model, where the vector’s magnitude modulates the reasoning strength. Subtracting this vector can lead to reduced reasoning token number and performance, while adding this vector can lead to increased reasoning token number and even improved performance. We further reveal that this direction vector consistently yields positive reasoning length prediction, and it modifies the logits of end-of-reasoning token \texttt{} to affect the reasoning length. Finally, we demonstrate two potential applications of our findings: overthinking behavior detection and enabling efficient reasoning on simple problems. Our work provides new insights into the internal mechanisms of reasoning in LRMs and offers practical tools for controlling their reasoning behaviors. Our code is available at \url{https: //anonymous. 4open. science/r/LRM-plans-CoT-7E04}.

PDF Details

ICML Conference 2024 Conference Paper

Evolving Subnetwork Training for Large Language Models

Hanqi Li
Lu Chen 0002
Da Ma
Zijian Wu
Su Zhu
Kai Yu 0004

Large language models have ushered in a new era of artificial intelligence research. However, their substantial training costs hinder further development and widespread adoption. In this paper, inspired by the redundancy in the parameters of large language models, we propose a novel training paradigm: Evolving Subnetwork Training (EST). EST samples subnetworks from the layers of the large language model and from commonly used modules within each layer, Multi-Head Attention (MHA) and Multi-Layer Perceptron (MLP). By gradually increasing the size of the subnetworks during the training process, EST can save the cost of training. We apply EST to train GPT2 model and TinyLlama model, resulting in 26. 7% FLOPs saving for GPT2 and 25. 0% for TinyLlama without an increase in loss on the pre-training dataset. Moreover, EST leads to performance improvements in downstream tasks, indicating that it benefits generalization. Additionally, we provide intuitive theoretical studies based on training dynamics and Dropout theory to ensure the feasibility of EST.

Details

NeurIPS Conference 2024 Conference Paper

Lean Workbook: A large-scale Lean problem set formalized from natural language math problems

Huaiyuan Ying
Zijian Wu
Yihan Geng
Jiayu Wang
Dahua Lin
Kai Chen

Large language models have demonstrated impressive capabilities across various natural language processing tasks, especially in solving mathematical problems. However, large language models are not good at math theorem proving using formal languages like Lean. A significant challenge in this area is the scarcity of training data available in these formal languages. To address this issue, we propose a novel pipeline that iteratively generates and filters synthetic data to translate natural language mathematical problems into Lean 4 statements, and vice versa. Our results indicate that the synthetic data pipeline can provide useful training data and improve the performance of LLMs in translating and understanding complex mathematical problems and proofs. Our final dataset contains about 57K formal-informal question pairs along with searched proof from the math contest forum and 21 new IMO questions. We open-source our code at \url{https: //github. com/InternLM/InternLM-Math} and our data at \url{https: //huggingface. co/datasets/InternLM/Lean-Workbook}.

PDF Details DOI

EAAI Journal 2023 Journal Article

Real-time segmentation network for accurate weld detection in large weldments

Zijian Wu
Peng Gao
Jing Han
Lianfa Bai
Jun Lu
Zhuang Zhao

Aiming at the defects of inaccurate weld extraction and high matching error rate in automatic welding system of large weldments currently. We propose a multi task detection model based on CNN architecture, which integrates the semantic segmentation technology required for weldment merging as well as the edge detection technology needed for weld matching. In particular, for the purpose of predicting smoother edges and welds, we carefully construct a new segment head, which adopts the sub-pixel convolution technology for up-sampling. Furthermore, a joint optimization loss function is explored to alleviate the imbalance of category distribution in large-scale weldment datasets. To verify the effectiveness of the model, abundant groups of data are collected for training and testing. The experimental results indicate that the proposed method has achieved the optimal trade-off between detection accuracy (83. 35% mIoU, 95. 15% F-score of welds and edges) as well as speed (74FPS) on a 2080Ti GPU compared with other state-of-the-arts, which greatly improves the robustness of the automatic welding system for large weldments.

Details DOI

JBHI Journal 2020 Journal Article

Data-Driven Analysis of Radiologists’ Behavior for Diagnosing Thyroid Nodules

Leilei Chang
Chao Fu
Zijian Wu
Weiyong Liu
Shanlin Yang

Thyroid nodule has been a common and serious threaten to human health. With the identification and diagnosis of thyroid nodules in the general population, large volumes of examination reports in clinical practice have been accumulated. They provide data basics of analyzing radiologists’ behavior of diagnosing thyroid nodules. To conduct data-driven analysis of radiologists’ behavior, an experimental framework is designed based on belief rule base, which is essentially a white box for knowledge representation and uncertain reasoning. Under the framework, with 2744 examination reports of thyroid nodules in the period from January 2012 to February 2019 that have been collected from a tertiary hospital located in Hefei, Anhui, China, experimental results are obtained from conducting missing validation, self-validation, and mutual validation. Three principles are then concluded from the results and corresponding analysis. The first is that missing features on some criteria are considered as benign ones by default, the second is that there is generally inconsistency between the recorded features on criteria and the overall diagnosis, and the third is that different radiologists have different diagnostic preferences. These three principles reflect three diagnostic behavioral characteristics of radiologists, namely reliability, inconsistency, and independence. Based on the three principles and radiologists’ behavioral characteristics, managerial insights in a general case are concluded to make the findings in this study available in other situations.

Details DOI