<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>RLHF on C.CUI's Log</title><link>https://cuicaihao.github.io/zh/tags/rlhf/</link><description>Recent content in RLHF on C.CUI's Log</description><generator>Hugo</generator><language>zh-Hans</language><lastBuildDate>Sun, 15 Nov 2026 10:00:00 +1100</lastBuildDate><atom:link href="https://cuicaihao.github.io/zh/tags/rlhf/index.xml" rel="self" type="application/rss+xml"/><item><title>高分不等于好回答：大模型对齐的理论上限与走偏机制</title><link>https://cuicaihao.github.io/zh/posts/2026-11-15-how-far-can-llm-alignment-go-and-where-can-it-go-wrong/</link><pubDate>Sun, 15 Nov 2026 10:00:00 +1100</pubDate><guid>https://cuicaihao.github.io/zh/posts/2026-11-15-how-far-can-llm-alignment-go-and-where-can-it-go-wrong/</guid><description>大模型对齐并不是把价值观直接写进模型，而是借助奖励信号重新分配回答概率。论文给出了奖励与策略偏离之间可估计的理论前沿，也说明当代理奖励偏离真实目标时，激进优化可能放大误差。本文进一步讨论这一结论对模型训练与使用的启示。</description></item></channel></rss>