---
title: "Optimal Watermark Localization in Mixed-Source Large Language Model Texts"
type: paper-summary
source_paper: "2026-08-14_21-41-50Z_OptimalWatermarkLocalizationinMixed_SourceLargeLan.md"
---
# Summary: Optimal Watermark Localization in Mixed-Source Large Language Model Texts

**Original paper:** [arXiv: 2608.14906](http://arxiv.org/abs/2608.14906v1)

## Summary
The paper treats token-level watermark localization in mixed-source LLM text as a multiple-testing problem. It shows that discovering the watermarked spans is harder than global detection and derives optimal detection boundaries; adaptive thresholding estimates the surviving watermark fraction without prior knowledge of the relevant exponents.

## Why it matters
Watermarking is not only a yes/no attribution problem. Real documents mix sources, so practical provenance systems must localize evidence under sparse and changing signals.