Three Bricks to Consolidate Watermarks for Large Language Models

TL;DR

This paper introduces three foundational approaches to improve watermarking techniques for large language models, enhancing detection robustness, real-world applicability, and multi-bit encoding capabilities.

Contribution

It consolidates watermarking methods through new statistical tests, benchmark evaluations, and advanced detection schemes for large language models.

Findings

Robust statistical tests with guarantees at very low false-positive rates

Benchmark evaluations demonstrate watermark effectiveness in real-world NLP tasks

Development of advanced detection schemes and multi-bit watermarking methods

Abstract

The task of discerning between generated and natural texts is increasingly challenging. In this context, watermarking emerges as a promising technique for ascribing generated text to a specific model. It alters the sampling generation process so as to leave an invisible trace in the generated output, facilitating later detection. This research consolidates watermarks for large language models based on three theoretical and empirical considerations. First, we introduce new statistical tests that offer robust theoretical guarantees which remain valid even at low false-positive rates (less than 10$^{\text{-6}}$). Second, we compare the effectiveness of watermarks using classical benchmarks in the field of natural language processing, gaining insights into their real-world applicability. Third, we develop advanced detection schemes for scenarios where access to the LLM is available, as well as multi-bit watermarking.