DualG-MRAG: Decoupling Macro-Reasoning and Micro-Matching for Multimodal Retrieval-Augmented Generation

2026-07-30

AI Summary

Existing multimodal retrieval-augmented generation methods struggle to effectively model explicit cross-modal and cross-document relationships in complex multi-hop reasoning: fine-grained visual features often lead to graph structure explosion and noise, while coarse-grained representations lose critical local evidence. To address this, this work proposes a two-level decoupled framework that constructs a macro reasoning graph for global path planning and a micro matching graph for local evidence verification. A query-driven graph neural network retriever is introduced to propagate relevance across heterogeneous evidence, and a dynamic programming decoder based on GNN forward propagation directly generates coherent reasoning paths. This approach significantly improves evidence recall and multi-hop question answering accuracy, achieving the first effective decoupling of macro-structure reasoning and micro-evidence matching.

Abstract

While Multimodal Retrieval-Augmented Generation (MM-RAG) has shown promising results, it still struggles with complex multi-hop reasoning tasks. Existing methods primarily focus on independent instance-level matching, which often fails to capture explicit relationships across modalities and documents. Although Graph-enhanced methods introduce structural modeling, they face a fundamental challenge in multimodal scenarios: incorporating fine-grained visual features leads to rapid graph expansion and retrieval noise, whereas coarse-grained representations cause the discarding of critical local evidence. To address this dilemma, we propose DualG-MRAG, a Dual-tier framework that introduces a decoupled architecture comprising Macro-reasoning and Micro-matching Graphs for Multimodal RAG. Specifically, to suppress retrieval noise by isolating global structural reasoning from fine-grained evidence matching, we construct a Macro Graph for global topological routing and a Micro Graph for precise local verification. Subsequently, to enable dynamic relevance propagation across heterogeneous evidence sources, we formulate retrieval as a query-driven message passing process via a GNN Retriever. Furthermore, to provide the generative model with coherent structural guidance, we introduce a dynamic programming decoding mechanism that extracts explicit reasoning paths directly from the GNN's forward pass, replacing the standard input of isolated document chunks. Extensive experiments demonstrate that DualG-MRAG outperforms baselines in both evidence recall and complex QA accuracy.

Problem

• Multimodal Retrieval-Augmented Generation
• multi-hop reasoning
• graph expansion
• retrieval noise
• fine-grained visual features

Innovation

• Decoupled Architecture
• Macro-Reasoning
• Micro-Matching
• Graph Neural Network Retriever
• Reasoning Path Extraction