面向最优多草稿投机解码

AI 摘要

该论文研究多草稿投机解码（MDSD）的最优接受率问题。作者利用最优传输问题的对偶形式，将最优接受率计算转化为子集选择问题，首次在词表规模达数千的真实场景下计算了MDSD效率的理论上界，并量化了现有验证算法与该上界的差距。研究发现草稿采样方法显著影响最优接受率，条件泊松采样（CPS）优于有放回采样，且现有验证算法在有放回采样下未达到理论最优。

正文节选

Towards Optimal Multi-draft Speculative Decoding Abstract Large Language Models (LLMs) have become an indispensable part of natural language processing tasks. However, autoregressive sampling has become an efficiency bottleneck. Multi-Draft Speculative Decoding (MDSD) is a recent approach where, when generating each token, a small draft model generates multiple drafts, and the target LLM verifies them in parallel, ensuring that the final output conforms to the target model distribution. The two

发布时间：2026-11-04 13:00

抓取时间：2026-11-04 13:06

来源机构：arXiv