Optimizing watermarks for large language models

TL;DR

This paper presents a systematic multi-objective optimization approach to designing watermarks for large language models, balancing their detectability with minimal impact on text quality, and identifies Pareto optimal solutions that outperform existing methods.

Contribution

It introduces a novel multi-objective optimization framework for watermark design in LLMs, providing Pareto optimal solutions that improve upon current watermarking techniques.

Findings

Pareto optimal watermark solutions outperform default methods

Trade-off between watermark detectability and text quality is effectively modeled

Systematic approach enhances robustness and efficiency of watermarks

Abstract

With the rise of large language models (LLMs) and concerns about potential misuse, watermarks for generative LLMs have recently attracted much attention. An important aspect of such watermarks is the trade-off between their identifiability and their impact on the quality of the generated text. This paper introduces a systematic approach to this trade-off in terms of a multi-objective optimization problem. For a large class of robust, efficient watermarks, the associated Pareto optimal solutions are identified and shown to outperform the currently default watermark.