Reproduce “Doc-V*:Coarse-to-Fine Interactive Visual Reasoning for Multi-Page Document VQA” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuanlei Zheng, Pei Fu, Hang Li, Ziyang Wang, Yuyi Zhang, Wenyu Ruan, Xiaojin Zhang, Zhongyu Wei and 4 more
Doc-V* reframes multi-page document VQA as an interactive search process: a model first inspects a low-resolution page overview, then retrieves or directly fetches selected pages while maintaining a structured evidence summary. Built on Qwen2.5-VL-7B-Instruct, the approach improves over fixed Top-5 retrieval across five benchmarks, while using less memory than processing every page at high resolution.
Key insights
- Global thumbnails and detailed page analysis serve complementary roles: the overview guides document navigation, while high-resolution pages support precise grounding and verification.
- Adaptive interaction is more effective than selecting a fixed number of pages in advance, especially when relevant evidence is discontinuous or requires multi-hop reasoning.
- Working memory, retrieval, and direct fetching each contribute to performance; removing any of these components reduces accuracy or increases the number of processed pages.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.