SPIQA: A Dataset for Multimodal Question Answering on Scientific Papers