Preprint reports targeted evidence agent improves document VQA

A preprint introduces Q-Guide, an agent that selects targeted tools to retrieve missing evidence from documents before answering visual questions. The authors report 65.0% accuracy on DocVQA2026 versus 40.0% for direct prompting, and 32.4% versus 24.4% on Manga109. They say gains appeared within two or three deliberate rounds and came from directing perception rather than adding planners, routers, or multiple agents.