![[2602.17608] Towards Anytime-Valid Statistical Watermarking](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.17608] Towards Anytime-Valid Statistical Watermarking
Summary
The paper presents a novel framework for statistical watermarking in machine-generated content, addressing limitations of existing methods and enhancing sample efficiency.
Why It Matters
As Large Language Models (LLMs) proliferate, distinguishing machine-generated content from human text becomes crucial. This research introduces a principled approach to watermarking that allows for valid early stopping, improving detection efficiency and reliability, which is vital for maintaining content integrity in AI applications.
Key Takeaways
- Introduces Anchored E-Watermarking, a new watermarking framework.
- Addresses limitations of existing statistical watermarking methods.
- Enables anytime-valid inference, improving detection reliability.
- Demonstrates a 13-15% reduction in token budget for detection.
- Theoretical claims supported by simulations and benchmark evaluations.
Computer Science > Machine Learning arXiv:2602.17608 (cs) [Submitted on 19 Feb 2026] Title:Towards Anytime-Valid Statistical Watermarking Authors:Baihe Huang, Eric Xu, Kannan Ramchandran, Jiantao Jiao, Michael I. Jordan View a PDF of the paper titled Towards Anytime-Valid Statistical Watermarking, by Baihe Huang and Eric Xu and Kannan Ramchandran and Jiantao Jiao and Michael I. Jordan View PDF Abstract:The proliferation of Large Language Models (LLMs) necessitates efficient mechanisms to distinguish machine-generated content from human text. While statistical watermarking has emerged as a promising solution, existing methods suffer from two critical limitations: the lack of a principled approach for selecting sampling distributions and the reliance on fixed-horizon hypothesis testing, which precludes valid early stopping. In this paper, we bridge this gap by developing the first e-value-based watermarking framework, Anchored E-Watermarking, that unifies optimal sampling with anytime-valid inference. Unlike traditional approaches where optional stopping invalidates Type-I error guarantees, our framework enables valid, anytime-inference by constructing a test supermartingale for the detection process. By leveraging an anchor distribution to approximate the target model, we characterize the optimal e-value with respect to the worst-case log-growth rate and derive the optimal expected stopping time. Our theoretical claims are substantiated by simulations and evaluations on est...
