Reproduce “Watermarking Language Models through Language Models” on OpenResearch

About this paper

Agnibh Dasgupta, Abdullah Tanvir, Xin Zhong

Researchers at the University of Nebraska Omaha developed a prompt-guided framework for watermarking language model outputs, enabling the embedding and detection of origin signals without requiring access to model internals. This method demonstrates high detection accuracies, such as 95% with GPT-4o and 88.79% with Mistral-7B, while exhibiting robustness against model fine-tuning, distillation, and prompt-based adversarial attacks.

Key insights

• Imperceptible and detectable watermarks can be dynamically embedded in LLM outputs purely through input-level prompt guidance, enabling black-box compatibility.
• Instruction-tuned LLMs can act as 'soft controllers' to implicitly encode and instantiate complex watermarking objectives, leveraging their natural language understanding for subtle output modulation.
• Watermark detection is effective by learning the statistical regularities induced by the prompt-based modulation, rather than relying on explicit scoring functions or token manipulation.

Summary by alphaXiv.