Zhenghao XuI am a fourth-year PhD candidate at Georgia Tech in the Machine Learning program, advised by Prof. Tuo Zhao and Prof. Molei Tao (co-advise). Before joining Georgia Tech, I received my bachelor's degree in Computer Science and Technology from Zhejiang University, where I was a member of the Qiushi Science Class (Computer Science). Email / GitHub / Google Scholar / LinkedIn / X (Twitter) |
|
ResearchMy current research focuses on reinforcement learning (RL) and its application in LLM post-training. I am also interested in continuous optimization and deep learning theory. Selected Papers |
Approximation of Log-Partition Function in Policy Mirror Descent Induces Implicit Regularization for LLM Post-TrainingZhenghao Xu, Qin Lu, Changlong Yu, Tuo Zhao arXiv, 2026 paper / code / blog (Revisiting (Kimi's) Policy Mirror Descent) Policy mirror descent for LLM post-training with an implicit regularization perspective. |
Ask a Strong LLM Judge when Your Reward Model is UncertainZhenghao Xu, Qin Lu, Qingru Zhang, Liang Qiu, Ilgee Hong, Changlong Yu, Wenlin Yao, Yao Liu, Haoming Jiang, Lihong Li, Hyokun Yun, Tuo Zhao NeurIPS, 2025 paper / code Uncertainty-based routing between reward models and strong LLM judges for efficient pairwise RLHF. |
Sample Complexity of Neural Policy Mirror Descent for Policy Optimization on Low-Dimensional ManifoldsZhenghao Xu, Xiang Ji, Minshuo Chen, Mengdi Wang, Tuo Zhao JMLR, 2024 paper / code Theory for policy optimization with neural function approximation under low-dimensional structure. |
Provable Acceleration of Nesterov's Accelerated Gradient for Rectangular Matrix Factorization and Linear Neural NetworksZhenghao Xu, Yuqing Wang, Tuo Zhao, Rachel Ward, Molei Tao NeurIPS, 2024 paper Acceleration and low-rank adaptivity guarantees for Nesterov's method in matrix factorization and linear networks. |
Good Regularity Creates Large Learning Rate Implicit Biases: Edge of Stability, Balancing, and CatapultYuqing Wang, Zhenghao Xu, Tuo Zhao, Molei Tao JMLR, accepted (to appear) paper / short version (M3L@NeurIPS 2023) Large learning rate dynamics in nonconvex optimization, including edge of stability, balancing, and catapult. |
Selected ProjectsUse this section for non-paper projects, demos, internships, or open-source work. |
|
Design and source code from Jon Barron's website |