Differentiable Information Enhanced Model-Based Reinforcement Learning