loss = -final_score + β * KL(new_policy || reference_policy)