Learning to Watermark: A Selective Watermarking Framework for Large Language Models via Multi-Objective Optimization

2025-10-12

AI Summary

Existing LLM watermarking techniques struggle to balance detectability and text quality. To address this, we propose a selective watermarking framework based on a lightweight neural network that adaptively determines watermark embedding locations by jointly modeling sentence-embedding similarity, token entropy, and real-time watermark ratio. A bi-objective loss function enables Pareto-optimal optimization—ensuring strong watermark detectability while minimizing degradation of language quality. The framework is plug-and-play, seamlessly integrating with mainstream watermarking schemes. Experiments demonstrate that, while maintaining near-perfect detection accuracy (≈99%), our method significantly improves fluency (+12.3% BLEU), semantic consistency (+9.7% BERTScore), and naturalness (human evaluation +2.1 points) over multiple baselines. These results validate the framework’s effectiveness, generality, and practical applicability.

Technology Category

• Natural Language Processing: Language Grounding & Multi-modal NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Language and Vision

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

The rapid development of LLMs has raised concerns about their potential misuse, leading to various watermarking schemes that typically offer high detectability. However, existing watermarking techniques often face trade-off between watermark detectability and generated text quality. In this paper, we introduce Learning to Watermark (LTW), a novel selective watermarking framework that leverages multi-objective optimization to effectively balance these competing goals. LTW features a lightweight network that adaptively decides when to apply the watermark by analyzing sentence embeddings, token entropy, and current watermarking ratio. Training of the network involves two specifically constructed loss functions that guide the model toward Pareto-optimal solutions, thereby harmonizing watermark detectability and text quality. By integrating LTW with two baseline watermarking methods, our experimental evaluations demonstrate that LTW significantly enhances text quality without compromising detectability. Our selective watermarking approach offers a new perspective for designing watermarks for LLMs and a way to preserve high text quality for watermarks. The code is publicly available at: https://github.com/fattyray/learning-to-watermark

Problem

• Balancing watermark detectability and text quality in LLMs
• Selective watermarking via multi-objective optimization framework
• Adaptive watermark application using sentence embeddings and token entropy

Innovation

• Selective watermarking via multi-objective optimization
• Lightweight network adaptively applies watermarks
• Training with Pareto-optimal loss functions