Towards Efficient Second Order Thinking