<?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>Normalization on Njx&#39;Log</title>
    <link>https://njx-njx.github.io/tags/normalization/</link>
    <description>Recent content in Normalization 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>Sat, 29 Aug 2026 12:05:00 +0800</lastBuildDate>
    <atom:link href="https://njx-njx.github.io/tags/normalization/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>“答错就写长一点”：GRPO 归一化项在扭曲什么</title>
      <link>https://njx-njx.github.io/posts/dr-grpo-length-and-difficulty-bias/</link>
      <pubDate>Sat, 29 Aug 2026 12:05:00 +0800</pubDate>
      <guid>https://njx-njx.github.io/posts/dr-grpo-length-and-difficulty-bias/</guid>
      <description>R1-Zero 的‘长度与奖励齐涨’被当成了推理涌现的标志性曲线，Sea AI Lab 的 Dr.GRPO 论文指出其中一半涨幅可能是优化器自己的偏好：1/|y| 让模型在答错时倾向于写长，除以 std 让极端难度的题拿走更多权重。两个归一化项、两种偏差的机制拆解，外加一个藏在 masked_mean 里的实现层偏差。</description>
    </item>
  </channel>
</rss>
