Problem
Research questions and friction points this paper is trying to address.
online watermark detection
large language models
anytime-valid inference
streaming text generation
AI-generated text detection
Efficient Online LLM Watermark Detection via Rao-Blackwellized E-Processes
🤖 AI Summary
This work addresses the limitations of existing large language model (LLM) watermark detection methods, which typically rely on fixed time windows and thus cannot support early stopping in streaming generation. The paper proposes the first online detection framework based on Rao-Blackwellized e-processes, integrating the Gumbel-max watermarking mechanism with a pivot-induced sequential test to enable recursive evidence updating at any time without storing the full generation history. By introducing e-processes into LLM watermark detection for the first time, the method achieves strict anytime validity, consistency, and strong theoretical guarantees. Empirical evaluations on both synthetic and real-world text demonstrate its computational efficiency and practical effectiveness.
📝 Abstract
As large language models (LLMs) are increasingly deployed, reliable and efficient mechanisms for distinguishing AI-generated text from human-written content have become essential. Statistical watermarking has emerged as a promising solution, yet most existing methods are typically fixed-horizon procedures, precluding valid early stopping in streaming generation. In this paper, we develop an efficient online watermark detection framework with anytime-valid inference based on Rao-Blackwellized e-processes, enabling recursive token-level evidence updates without storing the full history. In particular, we instantiate the framework for the Gumbel-max watermark and reduce the original token-level dependence testing problem to a pivot-induced sequential testing problem with an explicit null distribution. Theoretically, we prove anytime-valid Type I error control under arbitrary optional stopping and establish positive asymptotic log-growth under watermarking, implying consistency of the proposed stopping rules. Simulations and experiments on real LLM-generated text demonstrate efficient online detection with rigorous anytime-valid guarantees.
Innovation
Methods, ideas, or system contributions that make the work stand out.
online watermark detection
Rao-Blackwellization
e-processes
anytime-valid inference
sequential testing
🔎 Similar Papers
No similar papers found.
