---
id: 20260513-T0-17
title: "熵门控连续比特流扩散模型缩小语言模型自回归差距"
title_en: "Entropy-Gated Bitstream Diffusion Closes Autoregressive Gap in Language Models"
url: https://ai.daily.yangsir.net/daily/20260513-T0-17
issue_date: 2026-05-13
publish_date: 2026-05-12T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2605.07013
---

# 熵门控连续比特流扩散模型缩小语言模型自回归差距

最新研究提出熵门控连续比特流扩散模型，解决扩散语言模型在标准基准上落后于自回归模型的问题。该方法通过熵门控机制和连续比特流扩散技术，实现了并行且顺序无关的文本生成，在样本质量和多样性上取得了突破性进展。

## English Version

**Entropy-Gated Bitstream Diffusion Closes Autoregressive Gap in Language Models**

New entropy-gated continuous bitstream diffusion model closes the performance gap between diffusion and autoregressive language models. The approach enables parallel, order-agnostic text generation through entropy gating and continuous flow techniques, achieving breakthroughs in sample quality and diversity.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2605.07013)

**详情页**：https://ai.daily.yangsir.net/daily/20260513-T0-17

---

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