Reproduce “Adaptive Testing for Segmenting Watermarked Texts From Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xingchi Li, Xiaochi Liu, Guanxun Li
The authors adapt watermark-detection statistics to locate watermarked spans in mixed text. Their method tests overlapping windows with token-weighted statistics, then detects changes in the resulting p-value sequence; an empty-prompt estimate supplies next-token probabilities. In experiments with Llama 3 8B and Mistral 7B-v0.1, adaptive methods outperform baseline, while empty estimates match or sometimes exceed optimized estimates in tested settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.