Abstract: In Visual Document Understanding (VDU) tasks, finetuning a pre-trained Vision-Language Model (VLM) with new datasets often falls short in optimizing the vision encoder to identify ...