<?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>Logit-Lens on Xueping's Blog</title><link>https://hellogxp.github.io/tags/logit-lens/</link><description>Recent content in Logit-Lens on Xueping's Blog</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Sat, 06 Jun 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://hellogxp.github.io/tags/logit-lens/index.xml" rel="self" type="application/rss+xml"/><item><title>Late Crystallization in LLMs: How Facts Emerge in the Final Layers</title><link>https://hellogxp.github.io/posts/late-crystallization/</link><pubDate>Sat, 06 Jun 2026 00:00:00 +0000</pubDate><guid>https://hellogxp.github.io/posts/late-crystallization/</guid><description>85.9% of correct answers never appear in any intermediate layer prediction. We discovered that factual knowledge in LLMs undergoes a sudden phase transition—crystallizing only at the final layer. This phenomenon explains why some interventions work and others don&amp;rsquo;t.</description></item></channel></rss>