<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>语义向量 on AI内参</title>
    <link>https://neican.ai/tags/%E8%AF%AD%E4%B9%89%E5%90%91%E9%87%8F/</link>
    <description>Recent content in 语义向量 on AI内参</description>
    <generator>Hugo</generator>
    <language>zh-cn</language>
    <lastBuildDate>Mon, 11 May 2026 09:10:03 +0800</lastBuildDate>
    <atom:link href="https://neican.ai/tags/%E8%AF%AD%E4%B9%89%E5%90%91%E9%87%8F/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>从“稳稳接住”到“失语”：AI生成能力背后的长尾幻觉与工程边界</title>
      <link>https://neican.ai/insights/article-20260511091003884-3/</link>
      <pubDate>Mon, 11 May 2026 09:10:03 +0800</pubDate>
      <guid>https://neican.ai/insights/article-20260511091003884-3/</guid>
      <description>AI模型出现的“过度讨好”与“关键词失语”现象，揭示了后训练（SFT）阶段中奖励机制偏差与长尾知识分布稀释的工程挑战。未来的大模型竞争将转向精细化的行为调控与底层空间的可观测治理，而非单纯的参数扩张。</description>
    </item>
  </channel>
</rss>
