Haotian Ma

Explaining AI from learned features to abstract concepts.

Hi, I’m Haotian, a Ph.D. candidate in Computer Science at the University of Wisconsin–Madison, advised by Prof. Daifeng Wang. I have also had the pleasure of working with Prof. Aws Albarghouthi and Prof. Frederic Sala. Previously, I was a research intern at the John Hopcroft Center, where I worked with Prof. Quanshi Zhang.

My goal is to explain AI behavior from the bottom up, starting with learned features and building toward abstract concepts. To pursue this goal, I develop methods to understand how learned representations shape model behavior across interventions, combinations, and changing contexts. My research combines theory, method development, controlled interventions, open-source research infrastructure, and large-scale evaluation across mechanistic interpretability, LLM reasoning, AI safety, and AI for scientific discovery.

Recently, I founded OpenSAE to build broad research infrastructure for LLM evaluation. I am seeking Research Scientist and postdoctoral research opportunities in mechanistic interpretability, AI safety, and frontier model evaluation.