Abstract
Large Language Models (LLMs) often face challenges in performing reliable multi-hop reasoning due to issues such as incomplete evidence chains and hallucinations. Incorporating knowledge graphs (KGs) can mitigate these problems, but existing approaches either suffer from suboptimal accuracy or are computationally expensive. To address these issues, we propose Reasoning Path Retrieval for RAG (RPR-RAG), a novel graph-based retrieval method that incrementally builds a subgraph from the knowledge graph, extracts explicit reasoning paths, and provides them as structured external evidence to downstream LLMs. The experimental results on both WebQSP and CWQ datasets clearly demonstrate that RPR-RAG outperforms existing methods in multi-hop reasoning tasks, providing stronger accuracy and more efficient performance. Meanwhile, RPR-RAG significantly outperforms other KG-RAG method under zero-shot conditions on the MetaQA dataset, showing its superior capability in multi-hop reasoning without requiring task-specific training. RPR-RAG is designed to be lightweight and can be trained and executed on a single consumer-grade GPU (e.g., RTX 3060, 6 GB). Ablation studies reveal that the path validity evaluation and stopping criterion are essential for improving both coverage and efficiency. RPR-RAG is adaptable to various backbone LLMs, from smaller 7B models to larger models like ChatGPT, providing a reliable, cost-effective, explainable, and scalable solution for KG-grounded reasoning tasks. The source code is available at
https://github.com/threeOldFarmers/rpr-rag.