---
id: 20260807-T0-09
title: "OPTD新方法：用策略蒸馏将扩散语言模型推理步骤压缩至几步"
title_en: "OPTD: On-Policy Distillation Cuts Diffusion LM Steps to a Few"
url: https://ai.daily.yangsir.net/daily/20260807-T0-09
issue_date: 2026-08-07
publish_date: 2026-08-06T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2608.02942
---

# OPTD新方法：用策略蒸馏将扩散语言模型推理步骤压缩至几步

研究团队提出OPTD（On-Policy Transition Distillation），一种面向少步扩散语言模型的蒸馏加速方案。扩散语言模型虽然能并行预测多个token，但准确的生成仍需要多次迭代去噪。OPTD通过一致性引导的自适应压缩，将教师模型的多个步骤合并为学生的单步转换，在不牺牲生成质量的前提下大幅减少推理时间。该方案为扩散模型在实际应用中的部署提供了实用路径。

## English Version

**OPTD: On-Policy Distillation Cuts Diffusion LM Steps to a Few**

Researchers proposed OPTD (On-Policy Transition Distillation), a distillation method for few-step diffusion language models. While diffusion LMs can predict multiple tokens in parallel, accurate generation still requires many iterative denoising steps. OPTD compresses multiple teacher steps into a single student transition via consistency-guided adaptive compression, significantly reducing inference time without sacrificing quality. The approach offers a practical path for deploying diffusion models in real applications.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2608.02942)

**详情页**：https://ai.daily.yangsir.net/daily/20260807-T0-09

---

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