BLADE improves multi-mode attribution of LLM watermarks
Original title: Token-Level Attribution of Multi-Mode LLM Watermarks: Optimal Rule and Belief Weighting
On OPT-1.3B and other open-weight models, BLADE beat compared detectors across modes and substitutions.
The paper studies token-level attribution when a watermark verifier holds many secret modes and a document mixes those modes with unmarked text. Picking the largest pivotal statistic inflates Type I error, while restoring the nominal level cuts power as modes are added. The authors derive an optimal prior-weighted rule and propose BLADE, which estimates the prior and alternative density and keeps exact Type I error under a conditional product model.
- BLADE reports the mode with the largest prior-weighted alternative density.
- It estimates the mode prior from neighboring pivots.
- Power stayed higher across candidate counts, substitutions, and temperatures.
- The rule approaches optimality as prior and density estimates improve.
- Evaluations used OPT-1.3B and other open-weight models.
From the original article
1 min read at arxiv.orgA watermark verifier may hold many modes, i. e. , secret keys, and a document may mix spans of different modes with unwatermarked text.
By Rui Zhou · 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.