Education & Work

(I've recently moved my homepage to this location, so please bear with me as I work out any bugs or issues that may arise.)


Zeyuan awarded the World Champion Prize
by Rob Kostald, the head coach of USACO
@ Merida, Mexico, 2006
MIT

Tsinghua University

Personal Information

Research Interests

My current research focuses on investigating the physics of language models and AI in a broader sense. This involves designing experiments to elucidate the underlying universal principles governing how LLMs learn to accomplish diverse AI tasks. By probing into the neurons, one can uncover intricate (and sometimes surprising!) mechanisms behind how these AI models function. The ultimate goal is to provide theoretical guidance and practical suggestions on how we can ultimately achieve AGI. This line of work is featured at ICML2024 tutorial.

Before that, I work on the mathematics of deep learning. That involves developing rigorous theoretical proofs towards the learnability of neural networks, in ideal and theory-friendly settings, to explain certain mysterious phenomena observed in deep learning. In this area, our paper on ensemble / knowledge distillation received some award from ICLR'23; although I am most proud of our COLT'23 result that provably shows why deep learning is actually deep –– better than shallow learners such as layer-wise training, kernel methods, etc.

In my past life, I have also worked in machine learning, optimization theory, and theoretical computer science.

Conferences


Journals


Email

Service

Some Awards

In algorithm competitions, I was fortunate to win a few awards in my past life, including two IOI gold medals, a USACO world champion, an ACM/ICPC world-final gold medal, a Google Codejam world runner-up, and a USA MCM Top Prize.

In research, I used to be supported by a Microsoft Young Fellow Award, a Simons Student Award and a Microsoft Azure Research Award.

For a full list, click here.