The paper explores the use of code in enhancing the reasoning capability of large language models (LLMs) for math problems. They propose a prompting method called explicit code-based self-verification (CSV) that significantly improves accuracy on math datasets.