Weipu Zhang (张维璞)
Ph.D. student since 2025, Beijing Institute of Technology,
National Key Lab of Autonomous Intelligent Unmanned System, Supervisor: Prof. Gang Wang
北京理工大学自动化学院，自主智能无人系统全国重点实验室，导师：王钢 教授
Joint Ph.D. research track since 2026, Zhongguancun Academy (ZGCA)
AI & Game Research group, Supervisor: Prof. Jian Zhao
北京中关村学院，AI+游戏研究组，导师：赵鉴 教授
Research interests
My research interests center on building game agents that can learn, understand, and interact with games at a level of efficiency comparable to humans. To move toward this goal, I am particularly interested in reinforcement learning, computer vision, world models, and game generation, especially in settings that require strong generalization, long-horizon decision making, and efficient use of data and supervision.
Highlights
ICLR26: OC-STORM on Hollow Knight. Among the game-agent results I have worked on so far, this is the one that feels the most amazing to me: seeing an agent act coherently in a rich, difficult game world still feels remarkable.
Education
-
Joint Ph.D. research track, Zhongguancun Academy
Supervisor: Prof. Jian Zhao -
Ph.D. student, Beijing Institute of Technology
Supervisor: Prof. Gang Wang -
MSc in Cognitive Science, The University of Edinburgh
Graduated with Distinction -
B.Eng. in Automation, Beijing Institute of Technology
Awards
-
Cognitive Science MSc Dissertation Prize, The University of Edinburgh
Awarded to one student in the programme -
Poster Competition Prize, The University of Edinburgh
Awarded to one student in the School of Informatics -
Facebook Image Similarity Challenge 2021, NeurIPS 2021 Workshop
Global first place in Track 1 and third place in Track 2 Link
Publications
-
Object-Centric World Models from Few-Shot Annotations for Sample-Efficient Reinforcement Learning
ICLR, 2026.
-
STORM: Efficient Stochastic Transformer based World Models for Reinforcement Learning
NeurIPS, 2023.
-
Mixture-of-World Models: Scaling Multi-Task Reinforcement Learning with Modular Latent Dynamics
ICLR, 2026.
-
DyMoDreamer: World Modeling with Dynamic Modulation
NeurIPS, 2025.
-
Results and findings of the 2021 Image Similarity Challenge
NeurIPS Competition Track, PMLR, 2022.