MEGRAG: Answer-Aware Multi-Granular Evidence Graphs for Multi-Hop RAG

A novel framework named MEGRAG (Multi-Granular Evidence Graphs for Answer-Aware Multi-Hop RAG) has been introduced to overcome challenges in multi-hop question answering related to retrieval-augmented generation (RAG). This framework, outlined in arXiv paper 2608.02195, addresses two primary concerns: the dependence on single-granularity evidence in iterative RAG (iRAG) methods, which often fails to effectively balance information density and contextual noise, and the issues of redundant evidence and intermediate retrieval errors that can compromise final answers. MEGRAG conceptualizes multi-hop reasoning as a path-structured multi-granular evidence graph. It connects passages to sentences and extracted triples via a cross-granularity index offline, while online, it retrieves passages and builds a graph to facilitate answer-aware reasoning. The paper is categorized as 'new' on arXiv and can be found at https://arxiv.org/abs/2608.02195. The authors suggest this framework as a means to enhance the accuracy and efficiency of multi-hop RAG systems.