Watermarking is a common way to prove that text came from an LLM, but real-world text is often edited, paraphrased, or mixed with human writing. That leaves watermark evidence at only some token positions. This paper asks a more precise question than "is this text watermarked?": can we locate which tokens still carry the signal?

The authors model localization as a token-level multiple-testing problem with a latent indicator for each position. They derive sharp asymptotic boundaries for detection, discovery, and classification under a regime indexed by signal sparsity, next-token concentration, and vocabulary growth. Two results stand out: discovering any surviving positions is strictly harder than merely detecting a watermark somewhere, and no coordinatewise pivot-based rule can consistently classify all surviving positions across the parameter regime.

The proposed remedy is an adaptive thresholding method that estimates the surviving watermark fraction from the data, without knowing the exponents or time-varying token distributions. It reaches the optimal discovery boundary and near-optimal power relative to homogeneous pivot-based rules. Simulations confirm the phase transitions, and experiments on model-generated text show practical localization under common edits such as rewriting, insertion, deletion, and paraphrasing.