Jiaxun Li 李佳讯University of Michigan · Statistics
Learning theory, generation,
and reliable statistical inference.
I am a Statistics PhD student at the University of Michigan, advised by Ambuj Tewari and Saptarshi Chakraborty. I study the theoretical foundations of generative learning and the statistical principles behind reliable machine learning evaluation.
My work connects mathematical characterizations of generation with benchmark design, Python experiments, and finite-sample statistical inference.
I am looking for 2027 summer internship opportunities.
We study generation in statistical learning theory and characterize when a hypothesis class supports uniform, non-uniform, or prompted generation. These characterizations use a new combinatorial measure, the Closure dimension, and show that a class can be generatable without being predictable, or predictable without being generatable.
Characterizes generation in metric spaces using a scale-sensitive closure dimension, showing how geometry governs the stability of generation across novelty scales.
We introduce no-regret generation, where a generator makes a sublinear number of mistakes uniformly over all demonstration sequences and compatible target languages, and extend this framework to the agnostic setting. We establish necessary and sufficient conditions for no-regret generation and show that finite Littlestone dimension guarantees no-regret generation in both the realizable and agnostic settings.
Studies reliable detection under text contamination. I independently designed the benchmark and implemented the models, experimental pipelines, and Python evaluations, including AUROC and low-false-positive-rate detection.
We develop three methods for finite-sample inference on the average treatment effect in stratified randomized experiments with binary outcomes. Simulations and a case study compare their statistical and computational efficiency, and algorithmic improvements reduce the computational cost of the most statistically efficient method.
Earlier work: Analysis of a stochastic SIR model with media effects · Jiaxun Li and Yanni Xiao · Preprint, 2023.
* denotes equal contribution.
University of Michigan
PhD in Statistics · 2024–2029 (expected)
Advisors: Ambuj Tewari & Saptarshi Chakraborty
University of California, Berkeley
Visiting Student · 2023
GPA: 4.0/4.0
Xi’an Jiaotong University
BS in Mathematics, Honors Program · 2024
GPA: 4.21/4.30
Generation through the lens of learning theory
Contributed presentation · Conference on Learning Theory (COLT 2025)
Slides ↗
Graduate Student Instructor · University of Michigan
Selected courses:
Outstanding Graduate Student Instructor Team Award
STATS 485: Capstone Seminar · Winter 2026
For research collaborations and Summer 2027 internship opportunities:
jasonli@umich.edu ↗Outside research, I enjoy indie games, tennis, photography, movies, and progressive rock. Personal blog ↗