BioinformaticsJClub
BioHarness: Substrate-Aware Evidence Assembly for Biomedical Question Answering across Literature, Knowledge Bases, and Biological Atlases.
Bioinformatics · · Journal Article · Open access
Xiao, Qin + more
Abstract ↗AI summary
The abstract is read at the publisher; the summary is JClub's.
BioHarness improves biomedical question answering by raising pooled score from 62.2 to 67.2 across 22,123 QA items spanning seven answer formats using substrate-aware evidence assembly.
- Why it matters: Biomedical QA needs evidence beyond fixed literature retrieval, including gene alias resolution, identifier normalization, and atlas-derived measurements, but existing RAG workflows lack substrate-aware decision mechanisms.
- What they did: BioHarness stages evidence assembly by combining literature retrieval, biomedical entity mapping resources, and access to atlas-derived structured measurements, with ablations and scaling analyses to attribute gains.
- The result: Gains come from repairing evidence-substrate mismatches via reranking, entity grounding, and structured measurement access, and it approaches frontier-level performance on selected literature QA tasks while outperforming frontier harnesses without web access on identifier and locus lookup.
The findingWhy it mattersWhat they didThe result
- Open access