New LLM watermark detection works in real-time with early stopping

A team of researchers led by Lu Luo has introduced an efficient online watermark detection framework for large language models (LLMs) that solves the fixed-horizon problem of prior methods. Their approach, detailed in a preprint on arXiv, uses Rao-Blackwellized e-processes to enable anytime-valid inference — meaning detection can stop early as soon as evidence is sufficient, without needing to pre-specify the number of tokens. The framework recursively updates token-level evidence without storing the entire generation history, a major practical advantage for streaming text applications.

The method is instantiated specifically for the Gumbel-max watermark, reducing the original token-level dependence test to a pivot-based sequential test with an explicit null distribution. Theoretically, the authors prove strict Type I error control under arbitrary optional stopping and establish positive asymptotic log-growth under watermarking, ensuring consistency. Experiments on real LLM-generated text confirm efficient online detection with rigorous guarantees. This work advances the deployment of watermarking for real-time AI content monitoring, a critical need as LLMs generate vast streams of text in chat, code, and translation services.