InfoResearchPreprintLLM-specific
Fundamental Limits of Statistical Watermarking for LLMs with Fixed and Optimizable Common Randomness
- Published
- Record updated
Summary
This paper develops a unified theoretical framework for statistical watermarking of LLMs, treating it as an independence test between generated text and common randomness while jointly designing the embedder and detector. It characterizes the fundamental limits under both fixed and optimizable common randomness and finds that optimizing the common randomness does not improve the best achievable performance, though it allows a fixed, deterministic, model-agnostic detector to reach optimal performance.