DualG-MRAG: Decoupling Macro-Reasoning and Micro-Matching for Multimodal Retrieval-Augmented Generation
The paper’s core move is to split multimodal retrieval into one graph for routing and another for local evidence checks.
DualG-MRAG is aimed at MM-RAG systems that lose accuracy on complex multi-hop questions. The authors argue that graph-based retrieval either gets noisy when it includes fine visual detail or misses evidence when it stays coarse. Their framework uses a Macro Graph for global structure, a Micro Graph for fine-grained verification, and a GNN retriever to pass relevance across evidence sources. It also extracts explicit reasoning paths for the generator instead of feeding isolated chunks. ArXiv · AI/CL/LG's note
DualG-MRAG is aimed at MM-RAG systems that lose accuracy on complex multi-hop questions. The authors argue that graph-based retrieval either gets noisy when it includes fine visual detail or misses evidence when it stays coarse. Their framework uses a Macro Graph for global structure, a Micro Graph for fine-grained verification, and a GNN retriever to pass relevance across evidence sources. It also extracts explicit reasoning paths for the generator instead of feeding isolated chunks. ArXiv · AI/CL/LG's note
score 4