Watermarking large language model output typically embeds a secret signal that a verifier can later detect. But a verifier may hold many modes—that is, secret keys—and a document can contain spans generated under different keys as well as unwatermarked text. The challenge is to attribute each span to the correct key without making too many false claims.

The simplest approach, reporting the mode whose pivotal statistic is largest, turns out to be too permissive. According to the paper, this directly inflates the Type I error rate, meaning it will too often declare a match when there is none. The authors analyze this problem and propose an optimal rule for attribution, along with a belief-weighting method that accounts for uncertainty across modes.

The result is relevant for forensic analysis of LLM output, where false attribution could have real consequences. The paper frames the rule as optimal in a formal sense, though the abstract does not lay out the full derivation. As a preprint, the method has not yet undergone peer review.