All terms
Training
Preference Learning
Training a model to align with which of several candidate outputs people prefer.
Definition
Preference learning trains a model to align with human or AI preferences by showing it pairs or sets of candidate outputs labeled by which is better. A reward model or implicit reward signal then guides the model toward higher-quality, safer, and more helpful responses. It underlies RLHF, DPO, and GRPO, and avoids needing an exact correct answer for every query.