DependencyRAG: Dependency-Guided RAG for Multi-hop Question Answering

Multi-hop question answering (MHQA) requires integrating information from multiple evidence sources through a sequence of dependent reasoning steps. While large language models (LLMs) exhibit strong reasoning capabilities, their reliance on static internal knowledge often leads to hallucinations in knowledge-intensive scenarios. Retrieval-Augmented Generation (RAG) alleviates this issue by grounding generation in external evidence, yet existing RAG-based iterative frameworks fail to model the dependency structure among sub-questions, causing ineffective context propagation and suboptimal retrieval. In light of this, we propose DependencyRAG, a dependency-guided RAG framework for MHQA. The framework first employs an LLM-based dependency analysis module to decompose the original question and identify prerequisite relations among sub-questions, producing a topologically ordered reasoning plan. It then executes an iterative “Rewrite–Judge–Retrieve–Generate” procedure that rewrites each sub-question using only required prerequisite answers and selectively triggers retrieval based on dependency cues. This design reduces retrieval noise and ensures coherent multi-step reasoning. Experiments on three MHQA benchmarks show that DependencyRAG consistently outperforms state-of-the-art baselines, with an average improvement of 2.65 points in accuracy and 4.20 points in F1-score, demonstrating the effectiveness of dependency modeling for improving retrieval accuracy and reasoning quality in complex multi-hop QA.