---
id: 20260627-T0-07
title: "Dynamic-dLLM：无需训练的扩散模型加速方案"
title_en: "Dynamic-dLLM: Training-Free Acceleration for Diffusion LLMs"
url: https://ai.daily.yangsir.net/daily/20260627-T0-07
issue_date: 2026-06-27
publish_date: 2026-06-26T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2606.26120
---

# Dynamic-dLLM：无需训练的扩散模型加速方案

arXiv:2606.26120v1 发布。扩散大语言模型（dLLM）在文本生成上表现优异，但计算复杂度高达L的立方级（O(L^3)）。新提出的Dynamic-dLLM方法通过动态缓存预算（Dynamic Cache-Budget）和自适应并行解码（Adaptive Parallel Decoding）技术，在不重新训练模型的前提下实现了加速生成。

## English Version

**Dynamic-dLLM: Training-Free Acceleration for Diffusion LLMs**

arXiv:2606.26120v1 released. Diffusion LLMs excel in text generation but suffer from O(L^3) complexity. The proposed Dynamic-dLLM introduces dynamic cache-budget allocation and adaptive parallel decoding, achieving training-free acceleration for diffusion-based text generation.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2606.26120)

**详情页**：https://ai.daily.yangsir.net/daily/20260627-T0-07

---

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