Learnable In-Context Vector for Visual Question Answering