Watermarkable Multi-Draft Speculative Sampling via Poisson Processes
Poisson-process multi-draft speculative sampling stays efficient while embedding an unbiased LLM watermark.
The paper introduces a multi-draft speculative sampling algorithm based on Poisson processes that addresses the tension between faster LLM inference and output watermarking. The scheme embeds an unbiased watermark without reducing speculative acceptance and uses exact list coupling without communication, which yields drafter invariance. The authors present it as the first multi-draft, drafter-invariant speculative sampler that preserves both watermark strength and sampling efficiency, with experiments supporting both claims.
- Poisson-process multi-draft sampler improves speculative decoding efficiency.
- Unbiased watermark embeds without lowering speculative acceptance rates.
- List coupling without communication makes the scheme drafter-invariant.
- Reported as the first design preserving watermark strength and efficiency.
Full article151 words · extracted from arxiv.org · click to collapse
Large language models (LLMs) have achieved state-of-the-art performance across a wide range of tasks, motivating two important aspects of deployment: inference efficiency and output provenance, which can be tackled by speculative sampling and watermarking, respectively. However, recent works have shown that combining these two goals is highly nontrivial and can be potentially impossible. In this work, we develop a novel multi-draft speculative sampling algorithm based on Poisson processes that improves the frontier of this fundamental trade-off. The proposed algorithm has strong sampling efficiency on its own and, more interestingly, is naturally watermarkable: we can embed an unbiased watermark without degrading speculative acceptance. Moreover, our algorithm is based on an exact list-coupling-without-communication scheme, which yields a drafter invariance property that benefits both sampling and watermarking. It is the first multi-draft, drafter-invariant speculative sampling scheme that maintains both watermark strength and sampling efficiency, and we experimentally verify its strong performance in both aspects.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.21858