Agentic and Non-Agentic Multi-Hop Systems for Medical Question Answering

Abstract

This paper presents two systems developed for the MedHopQA Shared Task on multi-hop biomedical question answering. Our first system, Agentic-Qwen-Wikipedia, uses a lightweight agentic framework using SmolAgents to iteratively retrieve and reason over Wikipedia content via sub-query decomposition. Our second system, LLM-Qwen-Wikipedia-PubMed, offers a non-agentic, explicitly controlled pipeline that decomposes questions, retrieves evidence from both Wikipedia and PubMed, and synthesizes answers through a structured multi-hop process. The systems employ Qwen2.5-Coder-32B-Instruct-GPTQ-Int4 and Qwen-3-8B-AWQ, respectively. They are deployed efficiently using vLLM. The non-agentic system achieves higher performance on the shared task.

This article is part of the Proceedings of the BioCreative IX Challenge and Workshop (BC9): Large Language Models for Clinical and Biomedical NLP at the International Joint Conference on Artificial Intelligence (IJCAI).