The past three years have seen significant interest in applying language models to the task of visual document understanding – integrating spatial, textual, and visual signals to make sense of PDFs and scanned documents.