Statistical Foundations of LLMs
An up-to-date list is available on Google Scholar.
Research Philosophy
My research aims to establish the statistical and algorithmic foundations for modern AI. Large language models (LLMs) and other powerful architectures have achieved remarkable success, yet in many domains we still need theoretical guarantees of efficiency, reliability, and safety. I believe that statistical principles and algorithmic designs together offer a rigorous way to make AI systems interpretable, reliable, and generalizable—statistics by providing principled tools to reason about data, dependence structure, and uncertainty, and algorithms by turning these principles into methods that make learning robust and efficient.
Research Blueprint
The following figure represents my current understanding of the statistical foundations of LLMs. It outlines a broad conceptual landscape rather than the full scope of my ongoing work—so far, my research has focused on only a few of these components. I view this structure as an evolving framework that helps conceptualize how statistical principles, empirical behaviors, and algorithmic mechanisms interact in modern AI systems.
Recent Research
My current research focuses on LLM watermarking and LLM evaluation. Please see the Publications page for more details.
- For watermarking, I presented a short course on it at ICSA 2025, summarizing recent advances and my ongoing research on its statistical foundations. Lecture slides are available here.