Behavioral audit framework evaluates evidence-based reasoning in medical LLMs

A new audit method decomposes medical cases into evidence units to determine if LLM diagnoses rely on appropriate clinical facts. Testing five open-weight models on DDXPlus and MedCase shows that many evidence interactions are clinically plausible, though distinguishing between differential diagnosis and reasoning errors remains critical.