QID: Efficient Query-Informed ViTs in Data-Scarce Regimes for OCR-free Visual Document Understanding | Read Paper on Bytez