Improved Unbiased Watermark for Large Language Models

About

As artificial intelligence surpasses human capabilities in text generation, the necessity to authenticate the origins of AI-generated content has become paramount. Unbiased watermarks offer a powerful solution by embedding statistical signals into language model-generated text without distorting the quality. In this paper, we introduce MCmark, a family of unbiased, Multi-Channel-based watermarks. MCmark works by partitioning the model's vocabulary into segments and promoting token probabilities within a selected segment based on a watermark key. We demonstrate that MCmark not only preserves the original distribution of the language model but also offers significant improvements in detectability and robustness over existing unbiased watermarks. Our experiments with widely-used language models demonstrate an improvement in detectability of over 10% using MCmark, compared to existing state-of-the-art unbiased watermarks. This advancement underscores MCmark's potential in enhancing the practical application of watermarking in AI-generated texts.

Ruibo Chen, Yihan Wu, Junfeng Guo, Heng Huang · 2025

Related benchmarks

Task	Dataset	Result
Watermark Detection	C4 subset	Accuracy—	24
Text Generation	C4	TPR @ FPR=1%98.96	15
Watermark Detection	LLaMA-2 Token Replacement Attack epsilon=0.05 (1,000 generated sequences)	TPR@FPR=0.1%97.11	7
Watermark Detection	LLaMA-2 Token Replacement Attack epsilon=0.1 (1,000 generated sequences)	TPR@FPR=0.1%96.07	7
Watermark Detection	LLaMA-2 Token Replacement Attack, epsilon=0.2 (1,000 generated sequences)	TPR @ FPR=0.1%88.79	7
Watermark Robustness under GPT rephrasing attack	LLaMA-2 generated sequences (1,000)	TPR@FPR=5%62.5	7