<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title>Preference Learning - 标签 - lilfry's library</title><link>https://lilfry09.github.io/tags/preference-learning/</link><description>Preference Learning - 标签 - lilfry's library</description><generator>Hugo -- gohugo.io</generator><language>zh-CN</language><managingEditor>lilfry@sjtu.edu.cn (lilfry)</managingEditor><webMaster>lilfry@sjtu.edu.cn (lilfry)</webMaster><copyright>All rights reserved.</copyright><lastBuildDate>Fri, 30 Oct 2026 00:22:00 +0800</lastBuildDate><atom:link href="https://lilfry09.github.io/tags/preference-learning/" rel="self" type="application/rss+xml"/><item><title>读 Selective DPO：偏好数据为什么不是越难越好</title><link>https://lilfry09.github.io/ai/selective-dpo-data-difficulty/</link><pubDate>Fri, 30 Oct 2026 00:22:00 +0800</pubDate><author>fry</author><guid>https://lilfry09.github.io/ai/selective-dpo-data-difficulty/</guid><description><![CDATA[<p>偏好对齐里有一个很自然的直觉：只要 chosen 和 rejected 没有标错，数据越多越好；难样本甚至应该更有训练价值。</p>
<p>Selective DPO 挑战了这个直觉。它提出，偏好数据是否有用，不只取决于标签质量，也取决于样本难度是否落在当前模型的能力范围内。</p>
<p>论文的核心主张是：</p>
<blockquote>
<p>偏好样本具有稳定的难度差异；过难样本会因为超过模型容量而伤害对齐，而且“过难”的阈值随模型能力变化。</p></blockquote>]]></description></item></channel></rss>