A Watermark for Large Language Models

Researchers at the University of Maryland and Facebook AI Research proposed a watermarking framework for proprietary large language models that embeds invisible signals into generated text, detectable by an efficient open-source algorithm without API or parameter access, with negligible impact on text quality. The method, tested on a multi-billion parameter model from the Open Pretrained Transformer (OPT) family, uses a randomized set of 'green' tokens promoted during sampling and a statistical test with interpretable p-values, as detailed in the paper 'A Watermark for Large Language Models' submitted to arXiv on January 24, 2023.

Computer Science Machine Learning Submitted on 24 Jan 2023 v1 https://arxiv.org/abs/2301.10226v1 , last revised 1 May 2024 this version, v4 Title:A Watermark for Large Language Models View PDF /pdf/2301.10226 HTML experimental https://arxiv.org/html/2301.10226v4 Abstract:Potential harms of large language models can be mitigated by watermarking model output, i.e., embedding signals into generated text that are invisible to humans but algorithmically detectable from a short span of tokens. We propose a watermarking framework for proprietary language models. The watermark can be embedded with negligible impact on text quality, and can be detected using an efficient open-source algorithm without access to the language model API or parameters. The watermark works by selecting a randomized set of "green" tokens before a word is generated, and then softly promoting use of green tokens during sampling. We propose a statistical test for detecting the watermark with interpretable p-values, and derive an information-theoretic framework for analyzing the sensitivity of the watermark. We test the watermark using a multi-billion parameter model from the Open Pretrained Transformer OPT family, and discuss robustness and security. Submission history From: John Kirchenbauer view email /show-email/48291975/2301.10226 Tue, 24 Jan 2023 18:52:59 UTC 3,550 KB v1 /abs/2301.10226v1 Fri, 27 Jan 2023 18:54:34 UTC 3,620 KB v2 /abs/2301.10226v2 Tue, 6 Jun 2023 17:50:01 UTC 3,618 KB v3 /abs/2301.10226v3 v4 Wed, 1 May 2024 22:04:31 UTC 3,825 KB Current browse context: cs.LG References & Citations Loading... Bibliographic and Citation Tools Bibliographic Explorer What is the Explorer? https://info.arxiv.org/labs/showcase.html arxiv-bibliographic-explorer Connected Papers What is Connected Papers? https://www.connectedpapers.com/about Litmaps What is Litmaps? https://www.litmaps.co/ scite Smart Citations What are Smart Citations? https://www.scite.ai/ Code, Data and Media Associated with this Article alphaXiv What is alphaXiv? https://alphaxiv.org/ CatalyzeX Code Finder for Papers What is CatalyzeX? https://www.catalyzex.com DagsHub What is DagsHub? https://dagshub.com/ Gotit.pub What is GotitPub? http://gotit.pub/faq Hugging Face What is Huggingface? https://huggingface.co/huggingface ScienceCast What is ScienceCast? https://sciencecast.org/welcome Demos Recommenders and Search Tools Influence Flower What are Influence Flowers? https://influencemap.cmlab.dev/ CORE Recommender What is CORE? https://core.ac.uk/services/recommender IArxiv Recommender What is IArxiv? https://iarxiv.org/about arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs https://info.arxiv.org/labs/index.html .