Microsoft publishes MAI-DxO diagnostic results in Science

Summary

Microsoft AI published results in Science on October 22 for MAI-DxO, an orchestrator that runs multiple models as a virtual panel of physicians, proposing diagnoses and selecting tests step by step. On 304 difficult NEJM cases it reached 83% diagnostic accuracy, versus 22% for physicians without reference resources and 36% with them; on MIMIC-IV-ED emergency cases it scored 80% against 47% (all company-reported). A cost-optimized configuration held 81% accuracy, and the company says gains persisted when repeated with a newer model, GPT-5.5. The evaluation used simulated written cases rather than real patients, and the benchmark itself was built by Microsoft, which is open-sourcing part of it.

Why it matters for our work

It is a concrete sign that orchestrating multiple AI agents, rather than taking a single model's immediate answer, can sharply improve expert knowledge work such as research and diagnosis. The numbers come from Microsoft's own benchmark, however, and external replication and real-patient validation are still pending.

Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.