Hyper-M2RAG: High-Order Hypergraph Learning for Multimodal Document Retrieval

A new framework, Hyper-M2RAG, addresses limitations in Multimodal Retrieval-Augmented Generation (M-RAG) systems by using high-order hypergraph representation learning. Traditional simple graphs fail to capture complex N-ary relationships among heterogeneous entities like visual charts, textual descriptions, and numerical data. The proposed method formalizes document structure as a Multimodal Hypergraph, where hyperedges serve as unified semantic containers for multi-way associations. This approach reduces computational redundancy and contextual noise compared to exhaustive full-page reconstruction. The paper is available on arXiv (2608.16628) and introduces incremental refinement strategies for long-form document processing.