Near-Optimal Detection for Gumbel Watermarking Scheme in LLM Output
arXiv·medium signal
Proposes a provably near-optimal detection mechanism for the Gumbel watermarking scheme (Aaronson 2022), a leading approach for watermarking LLM-generated text. Provides formal guarantees on detection power, advancing the practical viability of AI-generated text detection. Relevant to anyone building or evaluating content authenticity systems as watermarking mandates gain regulatory momentum.