Glossary

Preference model — Glossary

A model trained to score which responses humans prefer, used in RLHF-like pipelines.

Also known as: reward model

Reviewed:

Full definition

Preference or reward models approximate human rankings. If the preference data are narrow, optimized policies may overfit those tastes.

Related encyclopedia topic

RLHF — Reinforcement Learning from Human Feedback (RLHF) shapes model behavior using preference data and reward models rather than only static supervised targets.

← Back to glossary · Learning hub