Multi-Task Learning for Routing Problem with Cross-Problem Zero-Shot Generalization