Models and training

Direct preference optimization (DPO)

stable definition

Definition

A preference-training method that directly adjusts a model toward preferred responses and away from rejected ones without first training a separate reward model in the classic RLHF pipeline.

Distinguish it from nearby terms

DPO is one model-training technique, not a general term for evaluating preferences.

Check your understanding

It changes model weights and should not be confused with runtime ranking of candidate outputs.

Also called

DPO

Public evidence