<?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>Moe on Njx&#39;Log</title>
    <link>https://njx-njx.github.io/tags/moe/</link>
    <description>Recent content in Moe on Njx&#39;Log</description>
    <image>
      <title>Njx&#39;Log</title>
      <url>https://njx-njx.github.io/blog-cover.jpg</url>
      <link>https://njx-njx.github.io/blog-cover.jpg</link>
    </image>
    <generator>Hugo -- 0.161.1</generator>
    <language>zh</language>
    <lastBuildDate>Wed, 26 Aug 2026 17:40:00 +0800</lastBuildDate>
    <atom:link href="https://njx-njx.github.io/tags/moe/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>比率用错了单位：GSPO 把重要性采样提到序列级</title>
      <link>https://njx-njx.github.io/posts/gspo-sequence-level-importance-sampling/</link>
      <pubDate>Wed, 26 Aug 2026 17:40:00 +0800</pubDate>
      <guid>https://njx-njx.github.io/posts/gspo-sequence-level-importance-sampling/</guid>
      <description>GRPO 沿用 PPO 的 token 级重要性比率，Qwen 团队认定这是对重要性采样的根本误用，会在 MoE 长回答训练中导致不可逆崩溃。GSPO 把比率、裁剪、优化全部提到序列级，被裁掉的 token 多两个数量级，训练反而更高效。本文拆解这条论证链，以及论文没点破的前提与代价。</description>
    </item>
  </channel>
</rss>
