---
id: 20260626-T0-14
title: "iLLaDA：8B参数的掩码扩散语言模型挑战自回归范式"
title_en: "iLLaDA: An 8B Masked Diffusion Language Model Challenges Autoregressive Training"
url: https://ai.daily.yangsir.net/daily/20260626-T0-14
issue_date: 2026-06-26
publish_date: 2026-06-25T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2606.25331
---

# iLLaDA：8B参数的掩码扩散语言模型挑战自回归范式

研究人员提出了iLLaDA，一个拥有8B参数的掩码扩散语言模型。该模型从头开始训练，采用完全双向注意力机制，挑战了现代LLM主流的自回归因子分解和因果注意力训练范式。

## English Version

**iLLaDA: An 8B Masked Diffusion Language Model Challenges Autoregressive Training**

Researchers presented iLLaDA, an 8B parameter masked diffusion language model trained from scratch. Using fully bidirectional attention, the model challenges the predominant autoregressive factorization and causal attention found in modern LLMs.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2606.25331)

**详情页**：https://ai.daily.yangsir.net/daily/20260626-T0-14

---

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