How it works

Researchers tested large reasoning models on clinical cases, comparing error types and detectability across models of varying capability.