Hello! ☃

I am an Assistant Professor at Stony Brook University, with appointments in Data Science, the Department of Applied Mathematics & Statistics, and the Department of Computer Science. I do research in natural language processing (NLP) and machine learning (ML).

Previously I was a Research Assistant Professor at the Toyota Technological Institute at Chicago (TTIC), situated on the University of Chicago campus. I obtained my Ph.D. from the School of Engineering and Applied Sciences (SEAS) at Harvard University, affiliated with the Natural Language Processing (NLP) group at Harvard/Cornell. I am fortunate to be advised by Professor Alexander (Sasha) Rush, currently at Cornell University. I have also worked with Professor Minlan Yu at Harvard on ML applications in networks and cybersecurity. I obtained an M.A. degree in Statistics from Harvard University. Prior to Harvard, I received my B.S. degree in EE from Tsinghua University.

Research

My previous webpage is here for reference of past research and projects.

My primary research interests lie in the areas of natural language processing (NLP) and machine learning (ML). The ultimate goal is to build general intelligent machines that can understand, interact, and help human beings on a wide variety of tasks with trustworthiness and efficiency. I am glad to see LLMs are bringing us one step closer 🤔

I am broadly interested in understanding and improving state-of-the-art deep learning models such as (large) language models (LLMs), on a variety of aspects such as efficiency, knowledge represention and manipulation, memorization, factualness, AI safety, fair evaluation, reasoning and planning, as well as multimodality. Check Research for more details. Feel free to reach out to collaborate on any interesting problems.

Here is a general talk I gave in Nov 2024 at Stony Brook for the CS department research seminar to all students and faculty - Generating from Generative ⸤Language⸣ Models (link)

Recent Interests

LLM + RL/reasoning: I have been working (and thinking a lot) on LLM + RL lately, where RL may unlock new capabilities of LLMs or new ways of generation. There is still much to explore - I maintain an online project of the fast progress in the space after DeepSeek R1 (with my personal bias).