In-Context Probing: Toward Building Robust Classifiers via Probing Large Language Models