Search

Type to search the site.

RLHF (reinforcement learning from human feedback)

Fine-tuning LLMs from human preference comparisons, often followed by reinforcement learning, widely used for alignment after pretraining.

RLHF (reinforcement learning from human feedback) refines LLMs using human preference comparisons, often ranking two model outputs, then optimizes with reinforcement learning (commonly PPO). It is a main practical lever for alignment after pretraining, distinct from supervised learning and fine-tuning on fixed input–output pairs.