Automatic creation of problems that combine multiple modalities (text, images, diagrams) with aligned annotations and ground-truth answers.