Overview
How do agentic and non-agentic multi-hop pipelines differ in medical QA accuracy, evidence use, and reliability?
System design
- Developed three multi-hop systems combining open-source LLMs with Wikipedia and PubMed retrieval
- Implemented an agentic workflow with smolagents and Qwen2.5-Coder-32B
- Built controlled non-agentic pipelines for evidence retrieval, reranking, and answer extraction
- Compared short-answer accuracy, evidence use, and reliability under the same evaluation setup
Results
- Published the results at BioCreative IX / IJCAI 2025
- Identified the best-performing retrieval configuration for the development set
- Produced a side-by-side analysis framework for agentic vs non-agentic system behavior