A new highly technical paper from researchers at Inria, Imatag and Meta AI proposes methods to make watermarking of large language models more robust and reliable.