Technique improves the reasoning capabilities of large language models