Sequential diagnosis with language models.

An AI system simulating sequential clinical reasoning with language models achieved 83.2% diagnostic accuracy on NEJM clinicopathological conference cases and 80.3% on MIMIC-IV-ED admissions.

• Why it matters: Sequential diagnosis mirrors iterative clinical reasoning, where clinicians trade off diagnostic accuracy against workup cost, but it is hard to benchmark and evaluate reliably.
• What they did: The authors built an interactive benchmark and an AI system that continually revises differentials and selects cost-effective tests, simulating a panel of physicians.
• The result: In the benchmark, the system outperformed generalist physicians (36.1 and 46.8%, respectively), suggesting language models may support more accurate, lower-cost diagnoses.