A HealthBench Score Measures Three Things, Not One

On HealthBench, which model looked best depended on the prompt. Under baseline instructions one model led. With prompts optimized per model, two others passed it. No weights changed in between. A benchmark table built on one fixed prompt is partly ranking how well each model suits that prompt, and a reader has no way to pull that apart from medical skill.

HealthBench invites the problem by being open-ended. Rather than multiple-choice items, it scores free-text answers to around 5,000 realistic healthcare conversations against rubrics written by physicians, so more than one answer can legitimately earn credit. The realism is the point, and it is also where the degrees of freedom come from. A system instruction changes structure, hedging and safety behavior. A judge then has to decide whether a half-satisfied rubric item counts.

We compared five hand-written strategies against Cost-Aware Prompt Optimization, an automated search that mutates and recombines prompts under a budget. Safety-first beat baseline everywhere, and forcing brevity was reliably the worst. Read the rubrics and that stops being mysterious, since they credit disclaimers, escalation advice and explanation, so an instruction to be concise mostly deletes things the model would have been paid for. The finding I liked most is not in the scores at all. All three optimized prompts, searched separately per model from different seeds, landed on safety-first language about patient safety, epistemic limits and referral. Independent searches converging on the same behavior is evidence that the search is finding rubric alignment rather than a lucky phrasing.

The judge experiments generalize furthest. We changed one thing at a time: fix the judge model and vary the grading instruction, then fix the instruction and vary the model. Across prompt variants the pairwise correlations held between roughly 0.71 and 0.80. Across models they dropped to somewhere between 0.12 and 0.60. Who grades mattered considerably more than how the grading was worded. We looked at correlation rather than agreement in means, since the question for model selection is whether the relative ordering survives, not whether two evaluators happen to average out the same. My guess is that judges diverge most on the borderline, partially satisfied items, which in open-ended medical answers is most of them.

The weakness in our own headline number is worth saying out loud. During the search, candidate prompts were graded by the model that had produced the answers, which keeps the budget manageable and introduces self-preference, while the manual strategies were graded externally. Those two sets of scores are not points on one scale and should not be read as if they were. Rerunning the optimized prompts under the external judge is the obvious next experiment.

There is also a claim this work cannot support, and it matters more here than in most domains. Prompt optimization does not give a model medical knowledge. It changes how existing capability comes out, measured against a rubric that stands in for clinical quality rather than being it. The prompt with the best HealthBench score is not therefore the safest one to deploy, and none of this says anything about patient outcomes.

The piece I would rebuild first is the dull one. Our pipeline keeps generation and scoring apart, stores responses as JSONL, resumes after interruptions and runs local judges through vLLM alongside API models. Because the answers sit on disk independent of any evaluator, re-scoring them under a different judge costs a single scoring pass. That is the difference between robustness analysis you recommend in a discussion section and robustness analysis you actually run. The harder version is what I would like to work on: designing open-ended evaluations, including multimodal ones, whose conclusions do not depend on which judge you happened to pick.