BIOBRIEFPAPER
A Nature Machine Intelligence paper describes a deep research agent for biomedical literature. The authors' benchmark win is real, but it was measured on benchmarks they designed themselves.
DeepEvidence, published this week in Nature Machine Intelligence, is a deep-research agent built for one job: turning the flood of biomedical papers into a traceable evidence map a researcher can audit.
The system does not just retrieve studies. It runs coordinated agents that combine breadth-first exploration across heterogeneous sources with depth-first drilling into specific findings, then incrementally builds an evidence graph: a structured map of entities, observations, and the citations that support each one, the authors write. Reference code is released as part of BioDSA on GitHub.
The authors report DeepEvidence "substantially outperforms generic AI agents" across four open benchmarks, and introduce seven new tasks spanning drug discovery, preclinical work, clinical trial development, and evidence-based medicine, with a benchmark dataset on Hugging Face and a preprint on arXiv.
The wins are author-reported on benchmarks the authors themselves designed and released; the paper does not document deployment in a working systematic review, drug-discovery pipeline, or clinical trial design workflow. The seven tasks, the dataset, and the reference code are all public, so outside teams can rerun the comparison on their own review sets before the benchmark claim settles.