SELF-[IN]CORRECT: LLMs Struggle with Refining Self-Generated Responses