RLAC: Reinforcement Learning with Adversarial Critic for Free-Form Generation Tasks