March 19, 2026 Webinar
Online LLM watermark detection via e-processes
Ruodu Wang, PhD
Abstract:
E-values and e-processes are an alternative to classic p-values, with key advantages including sequential validity, post-hoc decision validity, robustness to dependence, and deep connections to martingales. We first review the theory of e-values and e-processes and their advantages. Then we discuss the applications of e-values in watermarking for large language models (LLMs). Watermarking has is an effective tool for distinguishing AI-generated text from human-written content. Statistically, watermark schemes induce dependence between generated tokens and a pseudo-random sequence, reducing watermark detection to a hypothesis testing problem on independence. We develop a unified framework for LLM watermark detection based on e-processes, providing anytime-valid guarantees for online testing. Some experiments demonstrate that the proposed framework achieves competitive performance compared to existing watermark detection methods.
Short Bio:
Dr. Ruodu Wang is a Tier 1 Canada Research Chair in Quantitative Risk Management and a Professor of Actuarial Science and Quantitative Finance at the University of Waterloo. He received his PhD in Mathematics (2012) from the Georgia Institute of Technology after bachelor's (2006) and master's (2009) degrees from Peking University. He holds editorial positions at eight academic journals across different areas, and he is a Fellow of the Institute of Mathematical Statistics (2022).
Hope to see you at the webinar!