Can We Understand How Large Language Models Reason?