InstructGPT — RLHF at Scale (2022)
研究论文
全部解读文章
How raw language models became assistants people can use.
您正在查看一条研究论文记录：一篇里程碑式的论文，以及它为该领域带来了什么。
OpenAI showed that a 1.3B model tuned with human feedback was preferred over the raw 175B GPT-3 — alignment as capability. The three-stage recipe (demonstrations, reward model, RL) became the industry standard and made ChatGPT possible months later.
它如何与整个技术栈相连
本记录的关联链接一览——点击分组可直接跳转。
推出于
基础来源
来源
记录审阅截至 2026-01-15 — 查看 方法论.