Watermarks for Language Models via Probabilistic Automata

Yangkun Wang, Jingbo Shang

A recent watermarking scheme for language models achieves distortion-free embedding and robustness to edit-distance attacks. However, it suffers from limited generation diversity a…