Definition
Plain language
A benchmark of medical questions that hide the answer inside very long clinical documents.
As stated in the literature
A long-context medical question-answering benchmark used to evaluate reasoning over extended clinical text.
Why it matters: It tests whether a model can find and reason over key details scattered across long documents, a demand common in real medical work.
For example, it asks a model a clinical question whose answer is buried deep inside a patient's lengthy medical record.
Heard on the show
“On a long-context medical benchmark called LongHealth, it beats Deepseek-V3.”Episode 079 — An Old Idea From Cognitive Psychology Reshapes How We Reward Reasoning Models