Glossary entry

RLHF

Reinforcement learning from human feedback used to align model behavior with human preferences.

Definition

What it means

RLHF typically gathers human preferences, learns a reward signal from them, and then optimizes a model to produce outputs that receive higher predicted reward.

Why it matters

Role in the system

It connects qualitative human judgment to an optimization process that can improve model behavior at scale.

Example

In practice

Reviewers compare assistant answers, and those rankings help train a reward model.