New LLM watermark detection uses anytime-valid inference

Researchers have developed a new online detection framework for watermarks in large language models (LLMs) that allows for anytime-valid inference. This method, based on Rao-Blackwellized e-processes, enables recursive token-level evidence updates without needing to store the entire text history. The framework has been instantiated for the Gumbel-max watermark, transforming the detection problem into a sequential testing problem. Theoretical analysis confirms anytime-valid Type I error control and positive asymptotic log-growth, indicating consistency. Experiments on real LLM-generated text show efficient online detection with guaranteed validity. AI

IMPACT Enables more efficient and reliable detection of AI-generated text in real-time applications.

RANK_REASON The cluster contains an academic paper detailing a new method for LLM watermark detection. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• arXiv
• Gumbel-max watermark
• large language models
• Rao-Blackwellized e-processes

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →