---
id: 20260318-T0-14
title: "扩散语言模型在推理任务上表现落后，自回归规划可改进"
title_en: "Diffusion Models Underperform on Reasoning, AR Plans Help"
url: https://ai.daily.yangsir.net/daily/20260318-T0-14
issue_date: 2026-03-18
publish_date: 2026-03-17T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2603.13243
---

# 扩散语言模型在推理任务上表现落后，自回归规划可改进

新研究指出，扩散语言模型(dLLMs)通过迭代去噪生成文本，但在多步推理任务上持续落后自回归模型。研究表明，这种差距源于协调问题：自回归模型逐词构建连贯性，而扩散模型缺乏全局规划。研究者提出通过自回归规划条件化改进扩散模型推理能力。

## English Version

**Diffusion Models Underperform on Reasoning, AR Plans Help**

Diffusion language models underperform on multi-step reasoning due to coordination issues. Researchers propose autoregressive plan conditioning to improve dLLM reasoning.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2603.13243)

**详情页**：https://ai.daily.yangsir.net/daily/20260318-T0-14

---

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