Optimal Watermark Localization in Mixed-Source Large Language Model Texts
The research tackles locating watermarks in LLM text after editing by treating it as a token-level multiple testing problem using pivotal statistics. It proves that while global detection is possible, finding specific watermark positions (discovery) is strictly harder than detection and consistent classification is impossible for simple rules. A new adaptive method, SPOT, achieves near-optimal discovery power without needing prior knowledge of complex model parameters.