Chengchun Shi

(i), I am an associate professor of data science at London School of Economics and Political Science (LSE).

(ii) Pior to (i), I was an assistant professor of data science at LSE.

(iii) Prior to (ii), I was a Ph.D. student in Statistics at North Carolina State University (NCSU). I work with Dr. Wenbin Lu and Dr. Rui Song.

(iv) Prior to (iii), I obtained a B.S. in Statistics from Zhejiang University in July 2014.

(v) Prior to (iv), I graduated from Ningbo Xiaoshi High School in July 2010.

I am looking for students interested in LLMs (see our short course slides), reinforcement learning (see my lecture slides) and statistics.

Mar 2026:
- New paper on demystifying GRPO through U-statistics theory
Feb 2026:
- One paper on detecting AI-generated images accepted to CVPR.
- Talk on Stats-powered AI at QMUL.
Jan 2026:
- Our proposal Lowering the Computational Barrier: Resource- and Energy-Efficient Reinforcement Learning for LLM Reasoning has been awarded!
- Two papers on LLM and A/B testing accepted to ICLR.
- Talks on Stats-powered AI at Renmin University, ECNU and Sanya Workshop on Causality and ML.
Dec 2025:
- Our proposal Developing Reinforcement Learning Methods for Building Reliable Interventional Digital Health Systems has been awarded!
- One paper on RL accepted to JASA, the other to Brain.
Nov 2025:
- Talks on Stats-powered AI at Manchester, PolyU and Yunnan University.
Oct 2025:
- Talks on Stats-powered AI at UCL, University of Birmingham.
- One paper accepted to STAT.
Sep 2025:
- Four papers on LLM and RL accepted to NeurIPS 2025. Congratulations to all co-authors!
- Two of my AOS papers on RL were selected to present @ NeurIPS 2025.
- Two papers accepted to JASA. Congratulations to all co-authors!
August 2025:
- One paper on RL accepted to TMLR.
- Talk at BIRS, JSM and NKU on Doubly Robust LLM Fine-Tuning.

My research is motivated from the following applications:

Large language models:
- Reasoning: Demystify GRPO; Fin-R1
- RLHF: DRPO (slides, talk); VRPO (talk); Dual active learning
- AI-content detection: AdadetectGPT (talk); L2D; StatDetectLLM (have a try); DID
Ridesharing (see our AAAI tutorial (Youtube, Bilibili) and my talk; see also some simulated environments for Order Dispatch and Spatio-temporal Policy Evaluation);
Video-sharing (see our KDD paper successfully deployed in a widely used mobile app with millions of daily active users)
Mobile health (some simulated environments for Diabetes and Intern Health);
Psychology (see our paper on using RL for modelling human decision making).

Some of my recent talks and slides on AI, RL, statistical inference, causal inference and experimental design:

Stats-powered AI Youtube, Bilibili;
StatRL Bilibili, Youtube; Chinese versions: Bilibili, Youtube; slides; the accompanying paper and its Chinese version
CausalRL Youtube, slides; the accompanying paper
ARMAdesign 2-hour slides, 1-hour slides, 30-minute slides; the accompanying paper
OPE my slides and our review paper.

A summary of my past research: Statistical Methods in Reinforcement Learning (partly supported by EPSRC)