$G^2$-Reader: Dual Evolving Graphs for Multimodal Document QA

Retrieval-augmented generation is a practical paradigm for question answering over long documents, but it remains brittle for multimodal reading where text, tables, and figures are interleaved across many pages. First, flat chunking breaks document-native structure and cross-modal alignment, yielding semantic fragments that are hard to interpret in isolation. Second, even iterative retrieval can fail in long contexts by looping on partial evidence or drifting into irrelevant sections as noise accumulates, since each step is guided only by the current snippet without a persistent global search state. We introduce $G^2$-Reader, a dual-graph system, to address both issues. It evolves a Content Graph to preserve document-native structure and cross-modal semantics, and maintains a Planning Graph, an agentic directed acyclic graph of sub-questions, to track intermediate findings and guide stepwise navigation for evidence completion. On VisDoMBench across five multimodal domains, $G^2$-Reader with Qwen3-VL-32B-Instruct reaches 66.21\% average accuracy, outperforming strong baselines and a standalone GPT-5 (53.08\%). Code is available: https://github.com/DorothyDUUU/G2_Reader.

Many important documents, such as research papers, slide decks, reports, and web pages, mix text with tables, figures, and captions across many pages. Current AI reading systems often break these documents into small pieces before searching for an answer, but this can separate related information and make the system miss how different parts of the document connect. We introduce G2-Reader, a system designed to help AI read long, visually rich documents more reliably. It builds one graph to organize the document itself, linking related paragraphs, tables, figures, and captions, and another graph to organize the reasoning process by breaking a question into smaller sub-questions. As the system gathers evidence, it checks whether the information is sufficient and updates its plan when something is missing. This makes the system better at answering complex questions that require evidence from multiple pages or modalities. On a benchmark covering slides, papers, tables, web pages, and scientific figures, G2-Reader outperforms strong existing systems, showing that better organization of evidence and reasoning can make AI document understanding more accurate and trustworthy.