Are LLMs Ready to Assist Physicians? PhysAssistBench for Interactive Doctor-Patient-EHR Assistance

In three linesPhysAssistBench is an interactive medical assistance benchmark with 1,296 physician-validated turns built from real MIMIC-IV cases. It evaluates LLMs' ability to coordinate clinical knowledge, patient communication, and EHR system interaction within single dialogues. Experiments show current models remain unreliable in this setting.

Summary generated by Claude — human-verified