Med-PaLM 2

Announced via preprint in May 2023 and published in Nature that August ('Large language models encode clinical knowledge'), Med-PaLM 2 is Google Research's PaLM 2-based model for medical question answering. It reached 86.5% on USMLE-style MedQA questions — nearly 20 points above the original Med-PaLM — and in physician evaluations its long-form answers were preferred over physician-written answers on several axes including clinical accuracy. It was piloted at Mayo Clinic for non-clinical administrative tasks.

Med-PaLM 2 was never publicly released: no weights, no API, and no path for independent replication. Google's own researchers stress it remains inferior to clinicians, with residual hallucination rates of a few percent on consumer health questions and worse on rare diseases. Treat it as the landmark result that launched medical LLM evaluation methodology — and as a model that should never be used for real clinical decisions.