MedHopper: an agentic RAG-LLM system for multi-hop biomedical QA
Rustam Ruslanovich Taktashov1, Nadezhda Yurievna Biziukova1, Alexander Viktorovich Dmitriev1
1Institute of Biomedical Chemistry, 10 bld. 8, Pogodinskaya str., 119121 Moscow, Russia.
Database : the Journal of Biological Databases and Curation
|August 13, 2026
Summary
This study introduces MedHopper, a biomedical question-answering system for the MedHopQA task. It achieves 0.55 Exact Match using dense retrieval and an agentic workflow, improving biomedical information retrieval.
Area of Science:
- Biomedical Informatics
- Natural Language Processing
- Artificial Intelligence
Background:
- Biomedical question answering (QA) systems are crucial for accessing complex health information.
- Existing systems often struggle with multi-hop reasoning and entity-centric answers required in specialized biomedical datasets.
- The MedHopQA dataset presents a challenge due to its Wikipedia-derived corpus and exact match evaluation.
Purpose of the Study:
- To develop and evaluate a novel multi-hop biomedical QA system for the MedHopQA task.
- To improve the accuracy and efficiency of retrieving short, entity-centric answers from a biomedical corpus.
- To address limitations in current QA approaches for complex biomedical queries.
Main Methods:
- A multi-hop biomedical QA system, MedHopper, combining dense retrieval with a bounded agentic workflow.
- Utilized a controlled state machine for question routing into four strategies: direct, definition, intersection, and multi-hop.
- Implemented dense retrieval with cross-encoder reranking, up to three hops of intermediate entity extraction, and a bounded query-repair loop for answer validation.
Main Results:
- Achieved 0.55 Exact Match (EM) on the MedHopQA test set (N=1000) under official CodaBench evaluation.
- Ablation studies identified capped multi-hop execution, reranking, and query repair as key contributors to performance.
- Demonstrated efficient, deterministic, and zero-marginal-cost query processing on a single consumer GPU.
Conclusions:
- The MedHopper system effectively addresses the MedHopQA task by integrating dense retrieval and an agentic workflow.
- The proposed state machine and query repair loop enhance the system's ability to handle complex biomedical questions.
- Challenges remain in exact-match scoring, particularly with answer-type inconsistencies and lexical variations, highlighting areas for future research.
