Analyzing Chain-of-thought Prompting in Black-Box Large Language Models via Estimated V-information.