# Summary: 2026-08-14_21-41-50Z_OptimalWatermarkLocalizationinMixed_SourceLargeLan.md
Saved: 2026-08-17 21:57
Source: 2026-08-14_21-41-50Z_OptimalWatermarkLocalizationinMixed_SourceLargeLan.md
Model: None

---

## Summary  
The paper tackles the problem of locating watermark signals in text generated by large language models when the output is mixed‑source due to rewriting, insertion, deletion, or paraphrasing. It treats each token position as a test where a latent indicator records whether the watermark dependence survives, framing the task as a token‑level multiple‑testing problem based on pivotal statistics. Theoretical analysis reveals sharp phase transitions for global detection, discovery, and classification under varying sparsity exponents. An adaptive thresholding method uses a data‑driven estimate of the surviving watermark fraction to achieve optimal discovery performance without requiring explicit knowledge of these exponents.

## Key Contributions  
- [Finding 1] Formulation of watermark localization as a token‑level multiple‑testing problem with pivotal statistics and latent indicators.  
- [Finding 2] Derivation of sharp boundaries and phase transitions for global detection, discovery, and classification within the class of coordinatewise pivot‑based localization rules.  
- [Finding 3] Development of an adaptive thresholding method that employs a data‑driven estimate of the surviving watermark fraction to attain optimal discovery power.

## Methodology  
The authors model each token as a test whose outcome is governed by whether the watermark signal persists after editing operations. They analyze three asymptotic regimes indexed by exponents for signal sparsity, next‑token concentration, and effective‑vocabulary growth. Using these exponents they compute critical thresholds that separate detection from discovery and classification. The adaptive method gathers local estimates of the surviving fraction from observed data and applies a threshold to produce localized watermark signals, thereby avoiding reliance on pre‑specified exponent values.

## Results  
Theoretical results show that discovery is strictly harder than detection and that consistent classification cannot be achieved across the parameter regime within the pivot‑based class. Simulations confirm the predicted phase transitions, while experiments on model‑generated texts demonstrate practical localization performance under common edit mechanisms such as insertion, deletion, and paraphrasing.

## Significance  
The work provides optimal theoretical bounds for watermark localization in mixed‑source LLM outputs, enabling reliable authentication and traceability. The adaptive thresholding approach offers a practical solution that does not require prior knowledge of the underlying exponents, making it suitable for real‑world deployment where such parameters are unknown.

## Related Concepts  
Watermarking, large language models, token‑level multiple testing, pivotal statistics, asymptotic regimes, phase transitions, coordinatewise pivot‑based rules, discovery vs detection, classification, adaptive thresholding, effective vocabulary growth.