---
id: 20260416-T0-10
title: "LoSA：扩散模型长文本生成速度提升40%"
title_en: "LoSA Speeds Up Diffusion Models by 40%"
url: https://ai.daily.yangsir.net/daily/20260416-T0-10
issue_date: 2026-04-16
publish_date: 2026-04-15T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2604.12056
---

# LoSA：扩散模型长文本生成速度提升40%

MIT团队推出LoSA方法，解决扩散语言模型(DLMs)长文本生成效率问题。该方法通过局部感知稀疏注意力机制，将长场景下的生成速度提升40%，同时保持与自回归模型相当的生成质量。DLMs能够以任意顺序生成多个token，是传统生成方式的潜在替代方案。

## English Version

**LoSA Speeds Up Diffusion Models by 40%**

MIT researchers introduce LoSA, improving diffusion language models (DLMs) efficiency for long text generation. The method uses locality-aware sparse attention to boost speed by 40% while maintaining generation quality comparable to autoregressive models. DLMs generate tokens in any order, offering a promising alternative to traditional generation.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2604.12056)

**详情页**：https://ai.daily.yangsir.net/daily/20260416-T0-10

---

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