Ddog: Optimizing Multi-Hop Inference via Dual-Driven Retrieval and Reasoning Path

Abstract

Retrieval-Augmented Generation (RAG) mitigates issues such as hallucinations, limited knowledge coverage, and outdated information in large language models by incorporating external knowledge. In the real world, structured data like knowledge graphs are widely used. However, RAG struggles when processing structured data, particularly in scenarios requiring efficient handling of complex reasoning tasks. While Graph Retrieval-Augmented Generation (GraphRAG) improves multi-hop reasoning by leveraging graph structures, its performance is limited by incomplete coverage of static knowledge graphs, rigid path searches, and semantic gaps in cross-document reasoning. To address these challenges, this paper proposes Dual-Driven Optimization on GraphRAG (DDOG), which enhances the integrity of multi-hop reasoning paths and the reliability of answer generation through a dual-driven collaborative mechanism that combines static graph retrieval and dynamic knowledge expansion. Experimental results demonstrate that DDOG improves answer accuracy by an average of 18.6% on the HotPotQA, WikiHop, and CWQ datasets compared to directly using all retrieved documents, and achieves an average 9.5% improvement in EM over the baseline GraphRAG. DDOG exhibits significantly enhanced robustness in scenarios involving multi-hop inference path breaks. This study offers an efficient and scalable solution for complex reasoning tasks and introduces a new methodological perspective for dynamic knowledge integration and path optimization.