Abstract: Clinical large language models are usually judged by whether the final answer is right. Yet a model can reach a fluent, plausible output through invalid reasoning that static benchmarks do not detect, and confident hallucination, causally invalid inference, and spurious demographic shortcuts can pass conventional evaluation. This critical review
[…]
Content was trimmed to protect the source. Please visit the original article for the full text.
Read the original article:
