<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title>Long Context - Tag - lilfry's library</title><link>https://lilfry09.github.io/en/tags/long-context/</link><description>Long Context - Tag - lilfry's library</description><generator>Hugo -- gohugo.io</generator><language>en</language><managingEditor>lilfry@sjtu.edu.cn (lilfry)</managingEditor><webMaster>lilfry@sjtu.edu.cn (lilfry)</webMaster><copyright>All rights reserved.</copyright><lastBuildDate>Tue, 07 Apr 2026 21:12:00 +0800</lastBuildDate><atom:link href="https://lilfry09.github.io/en/tags/long-context/" rel="self" type="application/rss+xml"/><item><title>Read YaRN: Long context expansion, the core is not to pull the window hard, but not to break the RoPE</title><link>https://lilfry09.github.io/en/ai/yarn-context-window-extension/</link><pubDate>Tue, 07 Apr 2026 21:12:00 +0800</pubDate><author>fry</author><guid>https://lilfry09.github.io/en/ai/yarn-context-window-extension/</guid><description><![CDATA[<p>It would be a pity if we only regard <code>YaRN</code> as &ldquo;another trick to pull the context from <code>4k</code> to <code>64k/128k</code>&rdquo;.</p>
<p>What it really encounters is a more essential geometric problem than &ldquo;window expansion&rdquo;:</p>
<p>**The model is already accustomed to a position and phase system. Now that you forcefully increase the length, how can you make this system continue to work at a longer range while trying not to lose the positional resolution at close range? **</p>
<p>Once this question is clarified, <code>YaRN</code>’s design will appear very natural.</p>]]></description></item></channel></rss>