Xuheng Li

Xuheng Li

PhD Candidate · Department of Computer Science · University of California, Los Angeles

Email: [FIRST] [dot] [LAST] [at] cs [dot] ucla [dot] edu

CV Google Scholar X GitHub

I am a Ph.D. student of the AGI Lab in the Department of Computer Science at the University of California, Los Angeles, advised by Prof. Quanquan Gu. I received my B.Sc. at the School of Mathematical Sciences at Peking University.

My research focuses on optimization and RL applied to the pre-training and post-training of LLMs. I am also interested in sampling based-methods, including the diffusion models and their applications. I am fascinated with in how the dynamics of high-dimensional models are shaped by the low-dimensional structure of the data and training algorithms.

News

Publications

* denotes equal contribution.

Loading publications…

Research

Optimization in High Dimensions

Modern machine learning models are trained on high-dimensional loss landscapes whose behavior is far from well understood. I study how stochastic optimization algorithms interact with the intrinsic low-dimensional structure of data.

Sampling and Diffusion Models

Score-based generative models and Markov chain Monte Carlo samplers share a deep connection through stochastic differential equations. I work on the theoretical foundations of sampling algorithms, and on applying diffusion models to structured domains such as mixed-type electronic health records.

RL in Post-Training and Reasoning of LLMs

Reinforcement learning from human feedback and inference-time scaling are central to aligning and eliciting reasoning in large language models. I develop principled algorithms and statistical frameworks for contextual bandits and inference strategies.

A Little More About Me

Beyond research, hiking and stargazing are two of my favorite activities in life. Trying to make the most of a finite life in the vastness of nature and the universe.

"Look again at that dot. That's here. That's home. That's us."

Carl Sagan