[CVPR 2025] Docopilot: Improving Multimodal Models for Document-Level Understanding - View it on GitHub