Glossary entry

Reward model

A learned scorer that predicts how desirable a model output is.

Definition

What it means

A reward model converts examples of preferred behavior into a numerical signal. An optimization process can then use that signal to improve another model.

Why it matters

Role in the system

It makes subjective feedback usable in reinforcement-learning pipelines.

Example

In practice

A scorer estimates whether a coding answer is correct, useful, and well explained.