KL-Regularized Reward Maximization

A classical objective formulation in reinforcement learning and sequence generation that maximizes expected reward while constraining the KL divergence between the policy and a reference policy, preventing the policy from deviating too far from the prior.