Home › Research Papers › InstructGPT — RLHF at Scale (2022)
InstructGPT — RLHF at Scale (2022)
Research Paper
Alignment
How raw language models became assistants people can use.
You're looking at a research-paper record: a landmark paper and what it introduced to the field.
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.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Introduced in
Basis for
Sources
Record reviewed as of 2026-01-15 — see Methodology.