Authors
Rezarta Islamaj, Robert Leaman, Joey Chan, Nicholas Wan, Qiao Jin, Natalie Xie, John Wilbur, Shubo Tian, Lana Yeganova, Po-Ting Lai, Chih-Hsuan Wei, Yifan Yang, Yao Ge, Qingqing Zhu, Zhizheng Wang, Zhiyong Lu
Published in
Database : the journal of biological databases and curation. Volume 2026. Jan 15, 2026.
Abstract
Evaluating large language models (LLMs) in the biomedical domain requires benchmarks that can distinguish genuine reasoning from pattern matching and that remain discriminative as model capabilities improve. Existing biomedical question answering (QA) benchmarks are limited in this respect: multiple-choice formats allow models to succeed by answer elimination rather than inference, and widely circulated exam-style datasets are subject to performance saturation and training data contamination. Multi-hop reasoning, i.e. the ability to integrate information across multiple sources to derive an answer, is central to clinically meaningful tasks such as diagnosis support, literature-based discovery, and hypothesis generation, yet remains underrepresented in current biomedical QA benchmarks. We present MedHopQA, a disease-centred multi-hop reasoning benchmark of 1000 expert-curated question-answer pairs introduced as a shared task at BioCreative IX. Each question requires synthesis of information across two distinct Wikipedia articles, and answers are provided in open-ended free-text format rather than as multiple-choice selections. Gold annotations are augmented with ontology-grounded synonym sets (MONDO, NCBI Gene, and NCBI Taxonomy) to support both lexical and concept-level evaluation. The dataset was constructed through a multi-stage human-AI pipeline combining structured human annotation, triage, iterative verification, and LLM-as-a-judge validation. To reduce leaderboard gaming and contamination risk, the 1000 scored questions are embedded within a publicly downloadable set of 10 000 questions, with answers withheld, on a CodaBench leaderboard. Evaluation of four frontier LLMs under a zero-shot setting (GPT-5.1, Gemini 2.5 Pro, Claude Sonnet 4.5, and GPT-4o) reveals performance variation across answer types, with overall accuracy ranging from 66.3% to 83.4%. Performance is strongest on chemical and anatomical questions and most variable on disease and gene/protein categories, where fine-grained semantic discrimination is required. MedHopQA provides both a benchmark and a reusable framework for constructing future biomedical QA datasets that prioritize compositional reasoning, saturation resistance, and contamination mitigation as design constraints.
PMID:
42742456
Bibliographic data and abstract were imported from PubMed on 15 Sep 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 1
- Comments 0