Tag
This paper systematically evaluates input representations (images, text, or both) for multimodal document QA, finding that images improve accuracy but increase latency, and proposes a TF-IDF router to optimize performance.