---
id: 20261117-T0-09
title: "See Every Block Once：降低掩码扩散语言模型后训练方差"
title_en: "See Every Block Once: Variance-Reduced Post-Training for Masked Diffusion LMs"
url: https://ai.daily.yangsir.net/daily/20261117-T0-09
issue_date: 2026-11-17
publish_date: 2026-11-16T05:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2611.17754
---

# See Every Block Once：降低掩码扩散语言模型后训练方差

新论文针对掩码扩散语言模型（MDLM）的后训练难题提出方法。策略梯度方法常用于让LLM输出对齐人类偏好，但其所需的对数似然在MDLM中无法直接计算，导致后训练依赖近似手段、方差偏高。论文提出"每个块只看一次"的方差缩减思路，用于改进MDLM的后训练效果。对扩散类语言模型的研究者而言，这关系到这类模型能否用上与自回归模型相同的偏好对齐流程。

## English Version

**See Every Block Once: Variance-Reduced Post-Training for Masked Diffusion LMs**

A new paper addresses post-training for masked diffusion language models (MDLMs). Policy gradient methods are widely used to align LLM outputs with human preferences, but the log-likelihoods they require cannot be computed directly in MDLMs, so post-training relies on approximations with high variance. The paper proposes a variance-reduction approach that sees every block once to improve MDLM post-training. For diffusion language model researchers, this bears on whether such models can use the same preference-alignment pipelines as autoregressive models.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2611.17754)

**详情页**：https://ai.daily.yangsir.net/daily/20261117-T0-09

---

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