A small learned component that guides a frozen policy's behavior by modifying its input noise without changing the policy weights.