In reinforcement learning, a policy is the function that maps an agent's observation to an action — it can be a neural network, a lookup table, or any decision rule. Policy gradient methods directly optimize the policy by estimating how changes affect expected reward. Actor-critic architectures combine a policy (actor) with a value function (critic) for more stable training. In the context of LLMs, RLHF fine-tunes the language model itself as a policy that generates text actions to maximize a reward model's score.