Skip to content
InfoResearchPreprintLLM-specific

PIVOT: Distribution-Preserving Multi-BitWatermarking with Minimax-Optimal Scoring

Published
Record updated
View JSON

Summary

PIVOT is a multi-bit watermark for LLM output that embeds a recoverable message without changing the token distribution. It uses keyed pseudorandom scalars and a secret vocabulary permutation, with a message bit selecting forward or reverse rank-interval layouts. The authors frame detection and message recovery as multiple-hypothesis testing, derive a minimax-optimal score, and report experiments on two LLMs.