BioHarness: Substrate-Aware Evidence Assembly for Biomedical Question Answering across Literature, Knowledge Bases,...

Biomedical question answering often requires evidence beyond topically retrieved literature, including gene alias resolution, database identifier normalization, and atlas-derived biological measurements. However, existing retrieval-augmented generation (RAG) systems typically follow a fixed workflow and lack an explicit mechanism for deciding when retrieved text is sufficient, when biomedical entity mapping is needed, or when executable evidence assembly over structured measurements should be invoked. This motivates a substrate-aware large language model (LLM) harness that selectively assembles sufficient evidence by combining literature retrieval, biomedical entity mapping, and access to atlas-derived measurements.

We introduce BioHarness, an LLM harness for staged biomedical evidence assembly across literature retrieval, biomedical entity-mapping resources, and atlas-derived structured measurements. Across 22,123 biomedical QA items spanning seven answer formats, BioHarness improves the pooled score from 62.2 to 67.2 over the strongest baseline. Ablations, case studies, and backbone-scaling analyses show that these gains arise from repairing evidence-substrate mismatches through reranking, entity grounding, and structured measurement access, rather than from indiscriminately invoking more reasoning steps, retrieving additional literature, or relying on a particular answer-model scale. Comparisons with the frontier models and harness show that BioHarness, despite using an open-weight backbone with only 3B active parameters, approaches frontier-level performance on selected literature QA tasks and outperforms frontier harnesses without web access on biomedical identifier and locus lookup.