A Watermark for Large Language Models
The green-list scheme that most later watermark work builds on: the model is nudged toward a secret subset of tokens, and a detector counts how often they appear.
What it is
This is the paper that made text watermarking practical. The idea is simple: before each token, a hash of recent tokens splits the vocabulary into a green list and a red list, and the model is nudged toward green. A detector that knows the hash counts green tokens and computes how unlikely that count would be by chance. The scheme needs no access to the model at detection time and works on a few hundred tokens of ordinary prose. Nearly every later watermark, including SynthID-Text, is a refinement of this design.
When to use this entry
Cite it for the origin of the green-list scheme and the z-score style of detection.
Questions
Who wrote the first LLM watermark paper?
John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers and Tom Goldstein at the University of Maryland, 2023.