<?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>RLHF on GopherDing's Blog</title><link>/tags/rlhf/</link><description>Recent content in RLHF on GopherDing's Blog</description><generator>Hugo -- gohugo.io</generator><language>zh</language><managingEditor>GopherDing</managingEditor><webMaster>GopherDing</webMaster><lastBuildDate>Sat, 21 Nov 2026 11:30:00 +0800</lastBuildDate><atom:link href="/tags/rlhf/index.xml" rel="self" type="application/rss+xml"/><item><title>Finetuning with Sampling 深度解读：不动损失函数，改数据分布</title><link>/posts/tech/finetuning-with-sampling-%E6%B7%B1%E5%BA%A6%E8%A7%A3%E8%AF%BB%E4%B8%8D%E5%8A%A8%E6%8D%9F%E5%A4%B1%E5%87%BD%E6%95%B0%E6%94%B9%E6%95%B0%E6%8D%AE%E5%88%86%E5%B8%83/</link><pubDate>Sat, 21 Nov 2026 11:30:00 +0800</pubDate><author>GopherDing</author><guid>/posts/tech/finetuning-with-sampling-%E6%B7%B1%E5%BA%A6%E8%A7%A3%E8%AF%BB%E4%B8%8D%E5%8A%A8%E6%8D%9F%E5%A4%B1%E5%87%BD%E6%95%B0%E6%94%B9%E6%95%B0%E6%8D%AE%E5%88%86%E5%B8%83/</guid><description>解读 Karan、Chen 与 Du 的论文 Finetuning with Sampling：用 MCMC 投影采样把离线专家数据改造得更接近基座模型分布，让 SFT 泛化更好、遗忘更少，正面挑战 RL 后训练的主流叙事。</description></item></channel></rss>