Learning to Correct for QA Reasoning with Black-box LLMs