Evidence / original sources

A claim should lead somewhere.

Follow the references behind Health Benchmark. Each entry explains its role, links to the original publication, and identifies the guides that use it.

Content updated September 28, 2026. Inclusion does not imply endorsement.

Rahul K. Arora and colleagues · OpenAI

HealthBench: Evaluating Large Language Models Towards Improved Human Health ↗

Original benchmark construction, coverage, scoring and historical evaluation variability.

Version: arXiv v1 · 2025-05-13

Evidence locator: Sections 2–5 and 8; Tables 2, 3 and 7; Appendix D

arxiv.org

Used inHealthBench dossier →How HealthBench scoring works, step by step →Seven themes, five axes: two different HealthBench denominators →What HealthBench grader validation does and does not establish →

OpenAI

Introducing HealthBench ↗

Original release and explanation of physician-authored evaluation rubrics.

Version: Published 2025-05-12

Evidence locator: Benchmark release

openai.com

Used inHealthBench dossier →Seven themes, five axes: two different HealthBench denominators →

OpenAI

HealthBench reference evaluator ↗

Pinned reference scoring implementation, including positive-point normalization.

Version: Commit 652c89d0ca9df547706735883097e9537d40dc47

Evidence locator: calculate_score; calculate_length_adjusted_score; aggregate results

github.com

Used inHealthBench dossier →How HealthBench scoring works, step by step →What HealthBench grader validation does and does not establish →

OpenAI

HealthBench official dataset card ↗

Official dataset access and MIT license metadata.

Version: Accessed 2026-09-28

Evidence locator: Dataset card

huggingface.co

Used inHealthBench dossier →Seven themes, five axes: two different HealthBench denominators →What HealthBench grader validation does and does not establish →