---
id: 20260806-T0-03
title: "扩散语言模型解码提速新方案：草稿-精炼，双向修正更高效"
title_en: "Speculative Correction: Draft-then-Refine Decoding Boosts Diffusion LM Efficiency"
url: https://ai.daily.yangsir.net/daily/20260806-T0-03
issue_date: 2026-08-06
publish_date: 2026-08-05T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2608.02625
---

# 扩散语言模型解码提速新方案：草稿-精炼，双向修正更高效

一项新研究提出 'Speculative Correction' 推理方法，应用于扩散语言模型（DLM）。DLM 本身支持双向 token 修正，但标准解码常将其限制为从左到右的逐块生成。该方法采用两阶段流程：先快速生成草稿，再进行精炼修正。作为一种即插即用的推理模式，它有望在不改变模型结构的前提下提升解码效率。

## English Version

**Speculative Correction: Draft-then-Refine Decoding Boosts Diffusion LM Efficiency**

A new study presents 'Speculative Correction,' a plug-and-play inference pattern for diffusion language models (DLMs). DLMs allow bidirectional token revision, but standard decoding adapts them to left-to-right generation. This approach first generates a draft, then refines it, potentially improving efficiency without model changes.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2608.02625)

**详情页**：https://ai.daily.yangsir.net/daily/20260806-T0-03

---

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