Watermarkable Multi-Draft Speculative Sampling via Poisson Processes

TL;DR - This paper introduces a Poisson-process-based multi-draft speculative sampling algorithm that supports unbiased watermarking without reducing speculative acceptance. It advances the trade-off between efficient LLM inference and reliable output provenance.

• Uses an exact list-coupling-without-communication scheme to enable multi-draft speculative sampling.
• Provides drafter invariance, benefiting both sampling efficiency and watermark robustness.
• Preserves watermark strength without sacrificing speculative acceptance.
• Experiments report strong performance across both sampling efficiency and watermarking.