AutoMark: Enabling Autoresearch to Discover Better LLM Watermarks
AutoMark uses frontier-model agents to autonomously discover distortion-free LLM watermarks that beat prior schemes.
AutoMark defines reliability criteria, statistical tests, and a suite ranking LLM watermarks by detectability, quality, and robustness so agents can search for new distortion-free schemes. Using GPT-6 Astra, Opus 5, and Gemini-3.8 Flash, the authors report more than 50 schemes, including several that outperform prior work on all three dimensions. A manual study decomposes the schemes and identifies new ideas, such as aligning watermark scores with a random per-request direction. Code is released at eth-sri/automark.
- Agents discovered more than 50 distortion-free watermarking schemes.
- Several schemes beat prior work on detectability, quality, and robustness.
- New idea aligns watermark scores with a random per-request direction.
- Three frontier models drove the search and the code is public.
- Organizations
- ETH Zurich
- AI models
- GPT-6 AstraOpus 5Gemini-3.8 Flash
From the original article
1 min read at arxiv.orgWith LLM watermarking being deployed commercially and now required by regulations, improving its reliability and effectiveness has become crucial. Yet, recent progress in the field of LLM watermarking has increasingly been driven by improving details of existing methods, an effort fundamentally limited by the pace of human researchers.
By Thibaud Gloaguen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.