Fast segmentation of watermarked texts from large language models through an epidemic change-point framework

Soham Bonnerjee introduced WISER, a watermark segmentation algorithm that localizes watermarked segments in machine-generated text using epidemic change-point methods. The approach establishes finite-sample error bounds and consistency for detecting multiple watermarked segments within a single text. Experiments across benchmark datasets with diverse watermarking schemes show WISER outperforms existing methods in both computational speed and accuracy, demonstrating how classical statistical ideas can address the modern problem of watermark localization with theoretical guarantees.