Agentic Graph-RAG: A Multi-Agent Framework for Robust, Decomposed Multi-Hop Reasoning

作者：Jiayu Wang, Hao Shen, Bangpeng Xie, Yun Chen, Wenkai Zhao, Xiaohui Wang, Yiqi Hong, Chaoran Fu, Zhijun Pan, Lupeng Sun, Yanan Li · 年份：2025 · DOI：10.1109/iccc68654.2025.11437910 · 被引用次数：1 · 研究领域：Constraint Satisfaction and Optimization、Bayesian Modeling and Causal Inference、Advanced Graph Neural Networks

Large Language Models (LLMs) excel at fluency but fail in complex, multi-hop reasoning, which requires synthesizing multiple, verifiable facts. Existing solutions are fragmented; standard Retrieval-Augmented Generation (RAG) is vulnerable to retrieval noise, and Multi-Agent Systems (MAS) suffer from error propagation due to ambiguous communication. This paper introduces Agentic Graph-RAG, a multi-agent framework where agents collaborate via a structured Common Factual Ground rather than direct dialogue. A Planner Agent decomposes the query, a Dual-Retrieval Module gathers hybrid evidence from Knowledge Graphs (KGs) and text, and a Synthesizer-Verifier Agent performs cross-modal consistency checks to generate verified, cited conclusions for each step. Empirical evaluation shows the framework establishes a new state-of-the-art. It achieves 75.8% accuracy on the long-tail PopQA dataset and 94.2% faithfulness in a manual evaluation. The framework provides a robust solution by transforming reasoning into a transparent, verifiable process that mitigates error propagation.