Watermarking LLMs*

*This project is a work in progress.

We engineer and analyze random walk attacks on state-of-the-art watermarking schemes in different entropy settings. To support our claims, we present experimental evidence obtained by watermarking open-source models such as Mixtral.