---
id: 20260926-T0-04
title: "ELF-REG让连续扩散语言模型首次胜任推理任务"
title_en: "ELF-REG Enables Continuous Diffusion Language Models to Handle Reasoning Tasks"
url: https://ai.daily.yangsir.net/daily/20260926-T0-04
issue_date: 2026-09-26
publish_date: 2026-09-25T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2609.29102
---

# ELF-REG让连续扩散语言模型首次胜任推理任务

连续扩散语言模型一直在推理任务上表现不佳，原因是缺乏有效的正则化手段。研究者提出ELF-REG方法，通过在训练中对连续表示施加正则化约束，使模型在GSM8K等推理基准上的准确率大幅提升。该方法的核心思路是在去噪过程中引导模型学习更稳定的连续表示，从而在最终并行解码阶段生成更准确的答案。这意味着扩散语言模型有望在推理场景中与自回归模型竞争，为并行生成文本开辟新的可能性。

## English Version

**ELF-REG Enables Continuous Diffusion Language Models to Handle Reasoning Tasks**

Continuous diffusion language models (dLMs) have lagged behind autoregressive models on reasoning tasks due to unstable continuous representations during denoising. ELF-REG introduces a regularization method that constrains the continuous latent space during training, significantly boosting accuracy on reasoning benchmarks like GSM8K. The approach stabilizes the denoising process so that parallel decoding at the final step produces more accurate answers. This suggests diffusion language models can now compete with autoregressive models on reasoning, opening a path toward faster parallel text generation.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2609.29102)

**详情页**：https://ai.daily.yangsir.net/daily/20260926-T0-04

---

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