---
id: 20261022-T0-12
title: "扩散草稿模型提速推测解码：分解引导训练兼顾吞吐与接受率"
title_en: "Factorization-Guided Training Boosts Diffusion-Drafted Speculative Decoding"
url: https://ai.daily.yangsir.net/daily/20261022-T0-12
issue_date: 2026-10-22
publish_date: 2026-10-21T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.23935
---

# 扩散草稿模型提速推测解码：分解引导训练兼顾吞吐与接受率

arXiv新论文针对扩散草稿模型在推测解码中的权衡问题提出改进：扩散草稿器能进一步并行生成候选token，但提高候选容量或平均接受token数（MAT）时会出现性能折衷。论文用分解引导的训练与推理方法处理这一矛盾。推测解码是当前LLM推理加速的主流路线之一，这类改进直接关系到单位算力的输出吞吐。

## English Version

**Factorization-Guided Training Boosts Diffusion-Drafted Speculative Decoding**

A new arXiv paper tackles a tradeoff in diffusion-drafted speculative decoding: diffusion drafters parallelize proposal generation further, but scaling proposal capacity or mean accepted tokens (MAT) introduces performance compromises. The authors propose factorization-guided training and inference to address it. Speculative decoding is a mainstream path for LLM inference acceleration, so such improvements translate directly into output throughput per unit of compute.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.23935)

**详情页**：https://ai.daily.yangsir.net/daily/20261022-T0-12

---

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