Unsolvable Problem Detection: Evaluating Trustworthiness of Vision Language Models