<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Alignment on Nalar</title>
    <link>https://nalar.dev/tags/alignment/</link>
    <description>Recent content in Alignment on Nalar</description>
    <generator>Hugo</generator>
    <language>en-us</language>
    <lastBuildDate>Sat, 12 Sep 2026 00:00:00 +0700</lastBuildDate>
    <atom:link href="https://nalar.dev/tags/alignment/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Align LLMs with Direct Preference Optimization</title>
      <link>https://nalar.dev/align-llms-with-direct-preference-optimization/</link>
      <pubDate>Sat, 12 Sep 2026 00:00:00 +0700</pubDate>
      <guid>https://nalar.dev/align-llms-with-direct-preference-optimization/</guid>
      <description>&lt;h1 id=&#34;align-llms-with-direct-preference-optimization&#34;&gt;Align LLMs with Direct Preference Optimization&lt;/h1&gt;&#xA;&lt;p&gt;Supervised fine-tuning works well when you can provide a target response for each prompt. It becomes less natural when the signal is comparative: one answer is preferred over another, but neither is a perfect target to copy. &lt;strong&gt;Direct Preference Optimization (DPO)&lt;/strong&gt; turns those preference pairs into a training objective for a language model without requiring a separately trained reward model or an online reinforcement step.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
