Statistical Foundations of Generative AI

Large language models (LLMs) and other powerful architectures have achieved remarkable success, yet in many domains there remains a need for theoretical guarantees of efficiency, reliability, and safety. I believe that statistical principles and algorithmic design together provide a rigorous foundation for making AI systems interpretable, reliable, and generalizable. Statistics offers principled tools for reasoning about data, dependence structures, and uncertainty, while algorithms translate these principles into practical methods that enable robust and efficient learning.

My current work focuses on statistical watermarking, which provides provable methods for detecting and tracing AI-generated content, as well as on model evaluation and uncertainty quantification for large language models. Beyond these topics, I am also actively investigating related statistical problems, including synthetic data, reasoning, and fundamental limits.

LLM Watermarking

• A Statistical Framework of Watermarks for Large Language Models: Pivot, Detection Efficiency, and Optimal Rules
  X. Li, F. Ruan, H. Wang, Q. Long, and W. J. Su. The Annals of Statistics, 2025.
• Optimal Detection for Language Watermarks with Pseudorandom Collision
  T. T. Cai, X. Li, Q. Long, W. J. Su, and G. G. Wen (Alphabetical). arXiv preprint arXiv:2510.22007, 2025.
• Robust Detection of Watermarks in Large Language Models under Human Edits
  X. Li, F. Ruan, H. Wang, Q. Long, and W. J. Su. Journal of the Royal Statistical Society: Series B, 2025.
• On the Empirical Power of Goodness-of-Fit Tests in Watermark Detection
  W. He*, X. Li*, T. Shang, L. Shen, W. J. Su, and Q. Long. NeurIPS, 2025 (Spotlight).
• Debiasing Watermarks for Large Language Models via Maximal Coupling
  Y. Xie, X. Li, T. Mallick, W. J. Su, and R. Zhang. Journal of the American Statistical Association, 2025.
• Optimal Estimation of Watermark Proportions in Hybrid AI-Human Texts
  X. Li, G. G. Wen, W. He, J. Wu, Q. Long, and W. J. Su. arXiv preprint arXiv:2506.22343, 2025.
• Improving the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models
  W. He*, X. Li*, L. Shen, W. J. Su, and Q. Long. ICLR, 2026.
• Selective Disclosure Watermarking for Large Language Models
  X. Chen, X. Li, Y. Xie, and Q. Long. ICML, 2026.

AI Content Detection

• Steer-to-Detect: Probing Hidden Representations for Detection of LLM-Generated Texts
  L. Liang and X. Li. arXiv preprint arXiv:2605.12890, 2026.
• Robust Spectral Watermark for Synthetic Tabular Data
  Y. Zhao, X. Li, P. Song, Q. Long, and W. J. Su. Statistical Learning and Data Science, 2026.

LLM Evaluation

• Evaluating the Unseen Capabilities: How Many Theorems Do LLMs Know?
  X. Li, J. Xin, Q. Long, and W. J. Su. arXiv preprint arXiv:2506.02058, 2025.
• UCS: Estimating Unseen Coverage for Improved In-Context Learning
  J. Xin, X. Li, E. Qiang, W. He, T. Shang, W. J. Su, and Q. Long. Findings of ACL, 2026.