Decomposing Label Space, Format and Discrimination: Rethinking How LLMs Respond and Solve Tasks via In-Context Learning