Ensemble Watermarks for Large Language Models
Georg Niess, Roman Kern
Abstract
As large language models (LLMs) reach human-like fluency, reliably distinguishing AI-generated text from human authorship becomes increasingly difficult. While watermarks already exist for LLMs, they often lack flexibility and struggle with attacks such as paraphrasing. To address these issues, we propose a multi-feature method for generating watermarks that combines multiple distinct watermark features into an ensemble watermark. Concretely, we combine acrostica and sensorimotor norms with the established red-green watermark to achieve a 98% detection rate. After a paraphrasing attack, the performance remains high with 95% detection rate. In comparison, the red-green feature alone as a baseline achieves a detection rate of 49% after paraphrasing. The evaluation of all feature combinations reveals that the ensemble of all three consistently has the highest detection rate across several LLMs and watermark strength settings. Due to the flexibility of combining features in the ensemble, various requirements and trade-offs can be addressed. Additionally, the same detection function can be used without adaptations for all ensemble configurations. This method is particularly of interest to facilitate accountability and prevent societal harm.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
2 top-tier papers have cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- Stability-Aware Feature Design for Robust Watermark Detection in Machine-Generated Text cites A Watermark for Large Language Models
- DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature cites Authorship Attribution for Neural Text Generation
- DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature cites A Watermark for Large Language Models
- Provable Robust Watermarking for AI-Generated Text cites DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature
- IPMark: A Sentence-Level Watermark for LLMs with Hierarchical Personalization and Efficient Detection cites Provable Robust Watermarking for AI-Generated Text
- IPMark: A Sentence-Level Watermark for LLMs with Hierarchical Personalization and Efficient Detection cites A Watermark for Large Language Models
Cited by top-tier papers2
Builds on6
- DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability CurvatureEric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning et al.ICML 2023 · 988 citations
- A Watermark for Large Language ModelsJohn Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz et al.ICML 2023 · 854 citations
- Provable Robust Watermarking for AI-Generated TextXuandong Zhao, Prabhanjan Vijendra Ananth, Lei Li, Yu-Xiang WangICLR 2024 · 312 citations
- Large Language Models are Human-Level Prompt EngineersYongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster et al.ICLR 2023 · 297 citations
- Authorship Attribution for Neural Text GenerationAdaku Uchendu, Thai Le, Kai Shu, Dongwon LeeEMNLP 2020 · 110 citations
Related papers
- An Ensemble Framework for Unbiased Language Model WatermarkingYihan Wu, Ruibo Chen, Georgios Milis, Heng HuangICLR 2026 · 9 citations
- From Trade-off to Synergy: A Versatile Symbiotic Watermarking Framework for Large Language ModelsYidan Wang, Yubing Ren, Yanan Cao, Binxing FangACL 2025 · 4 citations
- Adaptive Text Watermark for Large Language ModelsYepeng Liu, Yuheng BuICML 2024 · 63 citations
- On the Reliability of Watermarks for Large Language ModelsJohn Kirchenbauer, Jonas Geiping, Yuxin Wen, Manli Shu et al.ICLR 2024 · 202 citations
- Topic-Based Watermarks for Large Language ModelsAlexander Nemecek, Yuzhou Jiang, Erman AydayACL Findings 2026
