<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>中文 on Xiaocan (Bruce) Li&#39;s Blog</title>
    <link>https://hsiaotsan.github.io/blog/tags/%E4%B8%AD%E6%96%87/</link>
    <description>Recent content in 中文 on Xiaocan (Bruce) Li&#39;s Blog</description>
    <generator>Hugo -- 0.164.0</generator>
    <language>en-us</language>
    <lastBuildDate>Wed, 27 May 2026 10:30:00 -0500</lastBuildDate>
    <atom:link href="https://hsiaotsan.github.io/blog/tags/%E4%B8%AD%E6%96%87/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>分解 MXFP4 量化误差：为大模型强化学习对症下药</title>
      <link>https://hsiaotsan.github.io/blog/posts/mxfp4-quantization-error-decomposition-zh/</link>
      <pubDate>Wed, 27 May 2026 10:30:00 -0500</pubDate>
      <guid>https://hsiaotsan.github.io/blog/posts/mxfp4-quantization-error-decomposition-zh/</guid>
      <description>MXFP4 的量化误差不是单一噪声，而是由三种性质完全不同的成分叠加而成；每一种主导一种 RL 训练失效模式，因此需要对症下药的修正策略。</description>
    </item>
  </channel>
</rss>
