Paper bounds how much edited watermarked text a corpus holds
Original title: Prevalence Bounds for Edited Watermarked Text: Power Envelopes and Calibration Sensitivity
On OPT-1.3B and Qwen3-1.7B text, a share of 0.2 yields an exact-null bound of about [0.14, 0.38].
A cs.CR paper derives prevalence bounds for the share of a corpus produced under Gumbel-max and green-list (KGW) watermarks when documents may have been edited by an unknown amount. Without edit restrictions the upper end of the identified set is 1. With retention, entropy and substitution budgets, the authors give a sharp power envelope, finite-sample bounds and confidence sequences. On text from OPT-1.3B and Qwen3-1.7B, 12,000 fresh human documents under three keys showed no robust departure from the exact null. Coverage breaks at 100,000 documents under a simulated null error of 0.01, while a null calibrated on an independent human sample still covers.
Unconfirmed: The authors say LLM-paraphrased KGW pools fall below the envelope, so those bounds are fragile.
- Unrestricted edits make the identified upper share 1.
- At a true share of 0.2 the exact-null bound is about [0.14, 0.38].
- 12,000 human documents scored under three keys stayed near the exact null.
- A 0.01 null error breaks exact-null coverage at 100,000 documents.
- LLM-paraphrased KGW pools fall below the envelope, yet bounds covered at 2,000 documents.
From the original article
1 min read at arxiv.orgWatermark detectors decide whether one document came from a watermarked language model. Auditors often need a different number: the share of a corpus that did. We study this share for Gumbel-max and green-list (KGW) watermarks when every document may have been edited by an unknown amount.
By Kuancheng Ye · 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.