Critic based reward model

critique is easier than generate

• : Reward modeling → generate that maximize reward like PPO

• : AI critic