Retrieval-augmented large language models (LLMs) have shown strong performance in open-domain question answering but often struggle with multi-hop reasoning and noisy evidence aggregation. Unlike recent methods that rely on LLMs excessively at every stage, we present ExpandFuse, a lightweight and modular framework that strategically integrates LLMs to enhance complex reasoning. ExpandFuse performs LLM-driven query expansion, hybrid sparse-dense retrieval, and topic-aware reranking, followed by a weighted fusion of semantic and topical relevance scores. The top-ranked evidence supports step-by-step answer generation via Chain-of-Thought prompting. Extensive experiments on HotpotQA, 2WikiMultihopQA, and StrategyQA demonstrate that ExpandFuse achieves competitive performance compared to Iter-RetGen and BlendFilter while maintaining efficiency. Ablation results confirm the effectiveness of query expansion and topic-guided fusion, highlighting ExpandFuse as a scalable, domain-agnostic solution for multi-hop question answering.
Murugaraj et al. (Mon,) studied this question.