Three Bricks to Consolidate Watermarks for Large Language Models
This work addresses the challenge of attributing generated text to specific models for users in AI and NLP, but it appears incremental as it consolidates existing watermarking techniques rather than proposing a fundamentally new approach.
The research tackled the problem of distinguishing generated from natural text by consolidating watermarks for large language models, introducing new statistical tests with robust theoretical guarantees at false-positive rates below 10^-6, comparing effectiveness on NLP benchmarks, and developing advanced detection schemes including multi-bit watermarking.
The task of discerning between generated and natural texts is increasingly challenging. In this context, watermarking emerges as a promising technique for ascribing generated text to a specific model. It alters the sampling generation process so as to leave an invisible trace in the generated output, facilitating later detection. This research consolidates watermarks for large language models based on three theoretical and empirical considerations. First, we introduce new statistical tests that offer robust theoretical guarantees which remain valid even at low false-positive rates (less than 10$^{\text{-6}}$). Second, we compare the effectiveness of watermarks using classical benchmarks in the field of natural language processing, gaining insights into their real-world applicability. Third, we develop advanced detection schemes for scenarios where access to the LLM is available, as well as multi-bit watermarking.