---
id: 20261031-T0-03
title: "多草稿策略蒸馏：让扩散草稿模型推测解码更省算力"
title_en: "Multi-Draft On-Policy Distillation Improves Diffusion Drafter Efficiency"
url: https://ai.daily.yangsir.net/daily/20261031-T0-03
issue_date: 2026-10-31
publish_date: 2026-10-30T04:00:00.000Z
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2610.37778
---

# 多草稿策略蒸馏：让扩散草稿模型推测解码更省算力

这篇arXiv论文研究推测解码的推理成本问题。一步扩散草稿模型通过并行提出多个token来降低推测解码的串行开销，而在线策略蒸馏用于训练这类草稿模型。论文提出多草稿在线策略蒸馏方法，目标是提升扩散草稿模型的质量与效率。具体实验数据和加速比在摘要中未给出。对开发者而言，若方法有效，可降低LLM推理延迟，尤其适合需要低延迟的在线服务场景。

## English Version

**Multi-Draft On-Policy Distillation Improves Diffusion Drafter Efficiency**

This arXiv paper addresses inference cost in speculative decoding. One-step diffusion drafters propose multiple tokens in parallel to reduce the sequential overhead of speculative decoding, and on-policy distillation is used to train these drafters. The work introduces multi-draft on-policy distillation for diffusion drafters, aiming to improve drafter quality and efficiency. Concrete benchmark numbers and speedup figures are not provided in the abstract. If effective, the method could lower LLM inference latency, particularly for latency-sensitive online services.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2610.37778)

**详情页**：https://ai.daily.yangsir.net/daily/20261031-T0-03

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*