RLHF (Reinforcement Learning from Human Feedback)
Alignment technique tuning LLM outputs using reward models trained on human pairwise preference rankings.
Mechanism & Definition
An alignment methodology that uses human preferences to optimize model behavior. A reward model is trained on human pairwise rankings of candidate model responses, followed by PPO (Proximal Policy Optimization) or DPO (Direct Preference Optimization) to align responses to human intent.