Hanxuan Li 李瀚轩

First-year PhD student

Biography

I am a first-year PhD student at SphereLab, the Chinese University of Hong Kong, advised by Prof. Weiyang Liu. I received my B.S. (2022.9 – 2026.6) from Turing Class, Chu Kochen Honors College, Zhejiang University. During my undergraduate studies, I had the privilege of working with Prof. Yulun Zhang and Prof. Huan Wang.

My research vision is to understand the principles underlying large language model pretraining.

Optimization. Pretraining algorithms and learning dynamics.
Architecture. Inductive biases in large-scale model design, with emphasis on neural memory design.
Scaling laws. How performance and behavior scale with width, depth, and data.

Selected papers

(* equal contribution, † corresponding author)

Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation
Kexuan Shi*, Hanxuan Li*, Zeju Qiu, Yandong Wen, Simon Buchholz, Weiyang Liu
arXiv preprint
tl;dr: A spectrum-preserving optimizer that updates each weight matrix via left and right orthogonal transformations, keeping singular values fixed while modulating geometry; competitive for LLM pretraining and finetuning.

For the full list, see Publications.