MirrorMark: A Distortion-Free Multi-Bit Watermark for Large Language Models

2026-01-29

AI Summary

Existing watermarking methods for large language models struggle to simultaneously achieve multi-bit embedding, high text quality, and robust detection. This work proposes a distortion-free multi-bit watermarking mechanism that embeds watermarks via measure-preserving random mirroring without altering the token distribution. To enhance robustness against insertion and deletion attacks, a context-aware scheduler is designed to evenly allocate watermark bits across the generated text. The proposed method uniquely unifies high-quality generation with strong detectability, embedding 54 bits of watermark within 300 tokens while preserving text quality comparable to original generation. It improves detection accuracy by 8–12% and achieves up to an 11% gain in identification rate at a 1% false positive rate. Furthermore, the study provides a theoretical analysis framework based on the equal error rate.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Generation
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

As large language models (LLMs) become integral to applications such as question answering and content creation, reliable content attribution has become increasingly important. Watermarking is a promising approach, but existing methods either provide only binary signals or distort the sampling distribution, degrading text quality; distortion-free approaches, in turn, often suffer from weak detectability or robustness. We propose MirrorMark, a multi-bit and distortion-free watermark for LLMs. By mirroring sampling randomness in a measure-preserving manner, MirrorMark embeds multi-bit messages without altering the token probability distribution, preserving text quality by design. To improve robustness, we introduce a context-based scheduler that balances token assignments across message positions while remaining resilient to insertions and deletions. We further provide a theoretical analysis of the equal error rate to interpret empirical performance. Experiments show that MirrorMark matches the text quality of non-watermarked generation while achieving substantially stronger detectability: with 54 bits embedded in 300 tokens, it improves bit accuracy by 8-12% and correctly identifies up to 11% more watermarked texts at 1% false positive rate.

Problem

• watermarking
• large language models
• distortion-free
• multi-bit
• content attribution

Innovation

• distortion-free watermarking
• multi-bit watermark
• measure-preserving sampling
• LLM watermarking
• context-based scheduler