Watermarkable Multi-Draft Speculative Sampling via Poisson Processes
The paper claims a speculative decoding scheme can keep watermark strength without giving up acceptance efficiency.
Liu, Wan and Gao build a multi-draft sampler around Poisson processes, aimed at the hard pairing of faster LLM inference and output provenance. The abstract says the method supports an unbiased watermark “without degrading speculative acceptance.” Its coupling scheme is described as drafter-invariant, which the authors say helps both sampling and watermarking. They call it the first multi-draft, drafter-invariant speculative sampler to maintain both properties, with experiments verifying performance on both fronts. ArXiv · AI/CL/LG's note
Liu, Wan and Gao build a multi-draft sampler around Poisson processes, aimed at the hard pairing of faster LLM inference and output provenance. The abstract says the method supports an unbiased watermark “without degrading speculative acceptance.” Its coupling scheme is described as drafter-invariant, which the authors say helps both sampling and watermarking. They call it the first multi-draft, drafter-invariant speculative sampler to maintain both properties, with experiments verifying performance on both fronts. ArXiv · AI/CL/LG's note
score 6