# DPO

Optimizes a model directly on preferred versus rejected answers, without training a separate reward model or running reinforcement learning.

## What it is used for

A simpler, more stable alternative to RLHF for preference tuning.