Can Large Language Models Perform Retrieval-Augmented Generation as Multi-Hop Reasoning Over Knowledge Graphs?

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by grounding their outputs in external knowledge. However, conventional chunk-based retrieval methods are limited in handling complex multi-hop questions due to their lack of structural reasoning and semantic alignment. In this work, we propose a novel pipeline where an LLM performs multi-hop subgraph retrieval over a knowledge graph (KG), followed by hallucination detection using triplet-level graph alignment. The final answer is generated by conditioning the LLM on both the user query and the verified subgraph. To construct our benchmark, we automatically generate 500 diverse multi-hop questions from 10 real-world websites and construct their corresponding KGs using GPT-4o. Experimental results show that our method significantly outperforms traditional chunking-based RAG in answer accuracy (84.29% vs. 68.77%). Furthermore, we introduce a fine-tuning procedure using Group Relative Policy Optimization (GRPO) guided by FactAlign rewards, which further improves answer accuracy to 87.62% while reducing hallucination by over 8 percentage points. Our approach demonstrates a principled and interpretable alternative to retrieval that leverages the structure and semantics of knowledge graphs, offering stronger factual consistency and reasoning capabilities.