Characterization of the watermarking–efficiency trade-off

Characterize the fundamental trade-off among target–draft communication, watermark strength, and speculative sampling efficiency for watermarkable speculative sampling schemes.

Background

The paper develops a drafter-invariant, multi-draft speculative sampling algorithm based on Poisson processes that embeds an unbiased watermark while preserving the target distribution and achieving strong sampling efficiency. Its theoretical and experimental results improve the known trade-off between watermark detectability and speculative sampling efficiency, but they do not provide a general characterization of how this trade-off depends jointly on communication between the target and draft models, watermark strength, and sampling efficiency.

The authors explicitly identify this missing characterization as unresolved, leaving open the task of determining the fundamental limits or achievable region governing these three quantities in watermarkable speculative sampling.