Debiasing Watermarks for Large Language Models via Maximal Coupling

DOI10.1080/01621459.2025.2520455MaRDI QIDQ6910923zbMATH OpenFDO

Authors Yangxinyu Xie, Weijie Su, Ruixun Zhang

Publication date 28 October 2025

Published in Journal of the American Statistical Association (Search for Journal in Brave)

Full work available at URL https://doi.org/10.1080/01621459.2025.2520455

Mathematics Subject Classification ID

Nonparametric hypothesis testing (62G10)

Cites work

• Coupling, stationarity, and regeneration
• Distribution-free tests for sparse heterogeneous mixtures
• Higher criticism for detecting sparse heterogeneous mixtures.
• Higher criticism for large-scale inference, especially for rare and weak effects
• Higher criticism: \(p\)-values and criticism
• Optimal detection of heterogeneous and heteroscedastic mixtures
• Optimal Detection of Sparse Mixtures Against a Given Null Distribution
• The intermediates take it all: asymptotics of higher criticism statistics and a powerful alternative based on equal local levels

Cited in

(2)

• Adaptive testing for segmenting watermarked texts from language models
• Robust detection of watermarks for large language models under human edits

This page was built for publication: Debiasing Watermarks for Large Language Models via Maximal Coupling

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6910923)