Towards Robust Mathematical Reasoning