New EAGR framework enhances multi-hop reasoning in RAG systems

Researchers have developed a new framework called Evidence-Guided Adaptive Graph Retrieval (EAGR) for multi-hop reasoning in retrieval-augmented generation (RAG) systems. EAGR operates without a dedicated graph database or explicit graph materialization, instead using an implicit semantic graph derived from passage metadata stored in a VectorDB. The system iteratively assesses accumulated evidence to guide the exploration of semantic directions, expanding the retrieval frontier through metadata lookups. This approach achieved higher LLM-based answer accuracy on benchmarks like HotpotQA, 2WikiMultiHopQA, and MuSiQue, outperforming vanilla RAG by 17% and a graph-aware baseline by 4% on average. AI

IMPACT Enhances multi-hop reasoning capabilities in RAG systems, potentially improving performance on complex question-answering tasks.

RANK_REASON The item describes a new research framework and its performance on benchmarks, published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.IR (Information Retrieval) →

• 2WikiMultiHopQA
• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• HotpotQA
• Hugging Face
• Litmaps
• MuSiQue
• Retrieval-Augmented Generation
• ScienceCast
• scite Smart Citations
• VectorDB

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →