Yu Huang
Hi! I’m Yu, a final-year Ph.D. student in Statistics at the Wharton School, University of Pennsylvania, where I am fortunate to be co-advised by Professors Yuxin Chen and Yuting Wei. I’m currently interning on the data team at the Institute of Foundation Models (IFM), MBZUAI, which develops the fully open K2 models.
Previously, I received my M.S. in Computer Science and B.S. in Mathematics from Tsinghua University. I also interned at Amazon AGI Foundations and Alibaba DAMO Academy.
My research is on the mathematical foundations of modern AI. I have concentrated on understanding the reasoning abilities of large language models and, more recently, on improving them through the science of data, including synthetic data generation and pretraining data mixtures for mathematical reasoning. My long-term goal is twofold: to make these foundations more predictive, so that theory can better inform how models are built; and to build AI systems that can themselves advance these foundations at scale.
Research Highlights
Reasoning in Language Models
How transformers learn in-context and chain-of-thought reasoning through training, when this reasoning generalizes to longer problems, and when test-time computation provably helps.
The Science of Training Data
How the difficulty mix of RL data induces an implicit curriculum; pretraining data mixtures and synthetic ML theory data for mathematical reasoning.
Representation Learning and Generative Models
When multimodal learning succeeds and when it fails, what self-supervised learning teaches vision transformers, and how to provably accelerate diffusion sampling.
Full list on the Publications page.