Sequential diagnosis with language models.

Diagnostic reasoning unfolds over time as clinicians gather information and revise hypotheses. This model evaluation examines frontier language models on consecutive diagnostic tasks, introducing a benchmark designed to examine performance across evolving clinical information rather than isolated questions. The work offers a more process-oriented approach to AI evaluation and suggests benchmark performance alone cannot improve patient outcomes or safety in clinical practice. Prospective comparisons with clinicians and real-world workflow testing are noted as important next steps