Saved in:
Bibliographic Details
Main Authors: Hu, Zhiqing, Zhao, Chenxu, Lu, Jiazhong, Liu, Xiaolei
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2512.19378
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918259172835328
author Hu, Zhiqing
Zhao, Chenxu
Lu, Jiazhong
Liu, Xiaolei
author_facet Hu, Zhiqing
Zhao, Chenxu
Lu, Jiazhong
Liu, Xiaolei
contents Misuse of LLM-generated text can be curbed by watermarking techniques that embed implicit signals into the output. We propose a watermark that partitions the vocabulary at each decoding step into three sets (Green/Yellow/Red) with fixed ratios and restricts sampling to the Green and Yellow sets. At detection time, we replay the same partitions, compute Green-enrichment and Red-depletion statistics, convert them to one-sided z-scores, and aggregate their p-values via Fisher's method to decide whether a passage is watermarked. We implement generation, detection, and testing on Llama 2 7B, and evaluate true-positive rate, false-positive rate, and text quality. Results show that the triple-partition scheme achieves high detection accuracy at fixed FPR while preserving readability.
format Preprint
id arxiv_https___arxiv_org_abs_2512_19378
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle HATS: High-Accuracy Triple-Set Watermarking for Large Language Models
Hu, Zhiqing
Zhao, Chenxu
Lu, Jiazhong
Liu, Xiaolei
Computation and Language
Misuse of LLM-generated text can be curbed by watermarking techniques that embed implicit signals into the output. We propose a watermark that partitions the vocabulary at each decoding step into three sets (Green/Yellow/Red) with fixed ratios and restricts sampling to the Green and Yellow sets. At detection time, we replay the same partitions, compute Green-enrichment and Red-depletion statistics, convert them to one-sided z-scores, and aggregate their p-values via Fisher's method to decide whether a passage is watermarked. We implement generation, detection, and testing on Llama 2 7B, and evaluate true-positive rate, false-positive rate, and text quality. Results show that the triple-partition scheme achieves high detection accuracy at fixed FPR while preserving readability.
title HATS: High-Accuracy Triple-Set Watermarking for Large Language Models
topic Computation and Language
url https://arxiv.org/abs/2512.19378