Event Status
Scheduled
Abstract:
LLM watermarking has emerged as a promising tool for attributing AI-generated text, but its practical deployment remains limited. In this talk, I will first present a unified theoretical framework that jointly optimizes watermark design and detection, characterizing the fundamental trade-off between detectability, false positive control, and text distortion. While this formulation yields sharp optimality insights, it typically requires the LLM provider to implement the watermark during generation, which has limited its deployment in practice. I will then discuss the reasons behind this gap, emphasizing stakeholder incentives and real-world deployment constraints. Motivated by these challenges, I will present in-context watermarking, a model-agnostic approach that embeds detectable signals through prompts rather than decoding-time access. This perspective suggests that effective watermarking requires not only strong algorithms, but also incentive-aligned designs tailored to practical domains such as peer review and education.
Bio:
Yuheng Bu is an Assistant Professor in the Department of Computer Science at the University of California, Santa Barbara. His research lies at the intersection of information theory, machine learning, and trustworthy AI. He received the NSF CAREER Award in 2026. Before joining UC Santa Barbara, he was an Assistant Professor in the Department of Electrical and Computer Engineering at the University of Florida. Prior to that, he was a Postdoctoral Research Associate at the Research Laboratory of Electronics and the Institute for Data, Systems, and Society at the Massachusetts Institute of Technology. He received his Ph.D. in 2019 from the Coordinated Science Laboratory and the Department of Electrical and Computer Engineering at the University of Illinois Urbana-Champaign. He earned his B.S. degree with honors in Electronic Engineering from Tsinghua University in 2014.