level: research
large language models are being used more in healthcare for tasks like diagnosis and patient communication. this survey looks at recent work on medical llms, focusing on reasoning. it uses a two-part view: one from clinical practice and one from computation. the clinical side uses miller's pyramid to define five levels of skill, from basic knowledge recall to managing changing cases. the computational side connects three types of reasoning—deductive, inductive, and abductive—to common medical goals.
the authors built a benchmark dataset covering all five clinical reasoning levels. they tested 18 current models on it. results show that models trained specifically on medical data do better on diagnosis tasks. general-purpose models are stronger in decision support and dialogue. this split suggests that no single model type is best for all clinical uses. the benchmark helps show where each kind of model fits.
the survey points out that linking clinical needs to ai methods is still a challenge. current models can handle some reasoning steps but struggle with complex, real-world case management. the five-level framework and benchmark aim to guide future work. by making the clinical requirements clear, researchers can build models that better match what doctors actually need. the paper calls for more work on aligning ai capabilities with practical medical workflows.
why it matters: it gives a clear map of where medical ai models succeed and fail, helping developers focus on real clinical gaps.