RLHF
Concept
Training & Inference
Reinforcement learning from human feedback — the technique behind ChatGPT's manners.
You're looking at a concept explainer: what this term means in plain English, and the records it ties together.
RLHF trains a reward model from human preference comparisons, then optimizes the language model against it. Introduced at scale by InstructGPT (2022), it is why assistants follow instructions and refuse harm — and its quirks (sycophancy, reward hacking) drive research into alternatives like constitutional AI and direct preference optimization.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Introduced
Based on
Sources
Record reviewed as of 2026-01-15 — see Methodology.