Member of Technical Staff at OpenAI

Email: leo [dot] dingliyu [at] gmail [dot] com

[Google Scholar]

 

About

I am a researcher at OpenAI, working on large language model research and development, with a focus on synthetic data generation.

Previously, I worked as a researcher at Microsoft where I participated in the development of Phi-4. I completed my Ph.D. of Computer Science at Princeton University, supervised by Sanjeev Arora. We worked at the intersects of deep learning theory and practices, including optimization and evaluation of large models.

Before that, I was a Yao Class student studying Computer Science at Institute for Interdisciplinary Information Science, Tsinghua University.

I participated in algorithmic programming competitions during high school and college, and won gold medals at the IOI and the ACM-ICPC World Finals.

Publications 

News

• I joined OpenAI in Feb. 2025.
• Three papers accepted by NeurIPS 2024!
• Depth-µP and Skill-Mix are accepted by ICLR 2024!
• We release Skill-Mix, a new type of evaluation of LLMs on their capability to combine basic skills. (See the demo here.) Skill-Mix resists data contaminations and can detect “cramming for leaderboard“! We also find GPT-4 is beyond “stochastic parrot” behavior based on its good performance on Skill-Mix!
• Tensor Program VI is here! Our new parametrization, Depth-µP, can scale up networks to infinite depth! You also get hyperparameter transfer for free, meaning using Depth-µP, optimal hyperparameters in a shallow network also work for a deep network.