Have there been any experiments using Qwen2.5-VL? What is the reason for choosing InternVL instead? Thanks for the awsome work.