A CMU research blog post argues that the well-documented gap between LLM benchmark performance and real-world deployment in healthcare (e.g., a 61 percentage point drop in one study) stems not from poorly designed benchmarks but from implicit assumptions embedded in evaluation protocols. The authors classify these assumptions into two types: task assumptions (how conversations are structured) and outcome assumptions (how human behavior mediates results). They propose BenchmarkCards — structured documentation for benchmark designers to make assumptions explicit — and a staged evaluation process to systematically test which assumptions hold before deployment. The framework is illustrated with a healthcare LLM-as-medical-assistant case where the gap decomposes into 12 points from query distribution, 19 from interaction type, and 30 from decision mediation.
