Glossary

DPO (direct preference optimization)

A fine-tuning method that trains a model directly on preferred-versus-rejected response pairs, without building a separate reward model or running reinforcement learning.

DPO consumes the same preference pairs as RLHF but replaces the reward-model-plus-RL loop with a single classification-style training objective. That makes preference tuning simpler, cheaper, and more stable to run, and it has become a standard alignment recipe, especially for open-weight models. The human preference pairs remain the input that must be sourced.

Why it matters

DPO lowered the engineering cost of using preference data, which broadened demand for it.

See also

← All terms