MedHopper: an agentic RAG-LLM system for multi-hop biomedical QA

This multi-hop biomedical QA system combines dense retrieval with a bounded agentic workflow for the MedHopQA (BioCreative IX) task, which requires short, entity-centric answers from a Wikipedia-derived corpus evaluated via exact match with normalization. Our approach replaces prior prompt decomposition with a controlled state machine that routes questions into four strategies (direct, definition, intersection, and multi-hop). The pipeline performs dense retrieval with cross-encoder reranking, executes up to three hops of intermediate entity extraction when needed, and applies answer validation with a bounded query-repair loop. On the MedHopQA test set (N = 1000), MedHopper achieves 0.55 Exact Match (MedHopQA evaluation metric) under official CodaBench evaluation. Ablation confirms capped multi-hop execution, reranking, and query repair as primary contributors; validation modules provide smaller consistent gains. Analysis reveals persistent challenges under exact-match scoring, including answer-type inconsistency (chromosome vs cytoband; yes/no vs symptom) and surface variation across plausible lexicalizations. All experiments were run on a single consumer GPU (RTX 5080) with no API dependencies, producing deterministic outputs at zero marginal cost per query.