RLHF

Reinforcement learning from human feedback: people rank answers, a reward model learns their taste, and the assistant is tuned to score well by it. It's how models pick up hard-to-specify qualities like being helpful, kind, and safe.