Policy

In RL, a policy maps states to a distribution over actions. In LLM alignment, the policy is the language model itself — sampling tokens as actions.

RLHF optimizes this policy against a reward model with a KL penalty to a reference policy.

← Back to AI Glossary