Constitution-Guided Watermarking
Constitution-Guided Watermarking selects per-request watermark trade-offs from natural-language provider rules.
Constitution-Guided Watermarking lets model providers express competing watermark goals as natural-language principles rather than one shared configuration. Offline, a pretrained reasoning agent refines rule-specific settings using empirical feedback; at deployment, a separate monitor applies the matching policy, including exemptions, without changing the serving model. In a proof of concept with KGW and a five-rule constitution, robustness-prioritized requests improved post-paraphrase detection by up to 14 percentage points over fixed configurations, while matching or exceeding baselines on aggregate quality and clean detection at a nominal 0.1% false-positive rate.
- Offline agent refines rule-specific watermark settings from provider principles.
- Deployment monitor selects a policy without modifying the serving model.
- Proof of concept uses KGW and a five-rule constitution.
- Post-paraphrase detection rises by up to 14 points on robustness requests.
From the original article
1 min read at arxiv.orgWatermarking enables language model providers to identify text generated by their models. However, its desired properties can conflict (\ie~stronger watermark signals can degrade text quality), while designs that resist editing may also facilitate forgery. Providers address these trade-offs by choosing configurations that balance competing objectives or prioritize particular properties.
By Toluwani Aremu · 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.