MAGE-RAG: Multigranular Adaptive Graph Evidence for Agentic Multimodal RAG in Long-Document QA

MAGE-RAG introduces a multigranular adaptive graph evidence framework designed for long-document multimodal question answering, addressing the limitations of existing retrieval-augmented generation (RAG) methods. The framework constructs an evidence graph that captures various relationships among text, images, and layout elements, allowing for dynamic evidence selection during query processing. Achieving 52.75% accuracy on LongDocURL and 53.26% on MMLongBench-Doc, MAGE-RAG's approach enhances evidence relevance while managing context noise, which is critical for practitioners developing efficient long-document QA systems.