|
Xinheng "Asher" Ding | 丁欣恒 Hi, I am currently a Ph.D. student in Computer Science at the University of Minnesota Twin Cities, fortunate to work with my awesome advisor, Dr. Zirui "Ray" Liu🤩. From August 2024 to July 2025, I spent a year as a Computer Engineering Ph.D. student at Indiana University Bloomington, where I was supervised by Dr. Dingwen Tao and Dr. Fengguang Song. Prior to that, I received my Bachelor’s degree in Engineering from the Department of Automation at the University of Science and Technology of China (USTC). 🎯 I am striving to grow into a professional, active, and influential researcher. If you have any suggestions, comments, critiques, encouragement, or collaboration opportunities, feel free to share them with me—I would greatly appreciate it.(^_^) |
|
Research InterestsI am broadly interested in the fields related to high-performance computing (HPC), artificial intelligence, and their intersection.
|
News
|
Blog |
|
π(A | S; θ)
π(A | S; θold)
|
RL for LLMs: From What?! to Why.
Feb 9, 2026 A comprehensive overview of the principles and procedures of classical reinforcement learning algorithms, covering value-based methods such as SARSA as well as policy-based approaches, represented by PPO and GRPO, which are widely used in large language models. |
|
Ut = Rt + γUt+1
|
Mathematical Foundations of RL
Feb 9, 2026 Instead of focusing on specific RL algorithms, this blog explains in as much detail as possible all the mathematical foundations involved in reinforcement learning: from random variables to the Bellman equations, in a way that even high school students can understand. |
|
O =
c
∑
i=1
XiWi
|
Enabling Large-Scale True-on-Policy RL by Bringing Tensor-Parallelism to Order
Nov 25, 2025 A detailed analysis of why, in modern LLM systems, different tensor-parallel sizes (TP size) can lead to non-deterministic behavior—i.e., the same input producing different outputs. It also explains how using different parallelization strategies in the training engine and the inference engine during RL training can cause RL to shift from on-policy to off-policy. |
Selected PublicationsFor a full list of publications, please refer to my Google Scholar Profile. |
|
Deterministic Inference across Tensor Parallel Sizes That Eliminates Training-Inference Mismatch
Ziyang Zhang*, Xinheng Ding*, Jiayi Yuan, Rixin Liu, Huizi Mao, Jiarong Xing, Zirui Liu Preprint, 2025, Used in SGLang Paper / Blog |
|
Understanding and Mitigating Numerical Sources of Nondeterminism in LLM Inference
Jiayi Yuan*, Hao Li*, Xinheng Ding, Wenya Xie, Yu-Jhe Li, Wentian Zhao, Kun Wan, Jing Shi, Xia Hu, Zirui Liu NeurIPS, 2025, Oral Presentation (Acceptance Rate: 0.03%) Paper / Talk |
|
GFormer: Accelerating Large Language Models with Optimized Transformers on Gaudi Processors
Chengming Zhang, Xinheng Ding, Baixi Sun, Xiaodong Yu, Weijian Zheng, Zhen Xie, Dingwen Tao Preprint, 2024 Paper |
Selected Awards and Honors
|
Services
|