What it measures
A named HealthBench evaluation slice appears in model safety and health reports.
How to read the result
Preserve the exact subset and grading definition instead of treating every HealthBench result as the same test.
Limits and comparison traps
Use the source’s rubric and denominator. Agreement-focused evaluation does not remove the need to inspect task coverage and clinical context.
Published evaluation
Overall source score on the predefined consensus criteria, converted from a fraction to percentage units. This is distinct from the original HealthBench and the Hard subset.
Where this fits in an evaluation suite
Read accuracy together with error and abstention rates. Domain averages can hide language gaps, specialist failures, and poorly calibrated confidence.
Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.
A reproducible comparison
Overall source score on the predefined consensus criteria, converted from a fraction to percentage units. This is distinct from the original HealthBench and the Hard subset.
Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.
PRIMARY EVIDENCE & FURTHER READING
- HealthBench: Evaluating Large Language Models Towards Improved Human Health — published result table ↗
All 12 authors
Arora, Rahul K., Wei, Jason, Hicks, Rebecca Soskin, Bowman, Preston, Quiñonero-Candela, Joaquin, Tsimpourlas, Foivos, Sharman, Michael, Shah, Meghan, Vallone, Andrea, Beutel, Alex, Heidecke, Johannes, Singhal, Karan
Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.