Model Graders in Health AI Evaluations

Why a model does the grading

HealthBench holds 48,562 criteria across 5,000 conversations. Scoring one model means a decision on every criterion of every case it is given, and each rerun repeats all of it. Physician grading at that volume is not affordable for a board that has to refresh whenever a new model ships. So the benchmark sends the grader the conversation, the response, and one criterion, and takes back a met or not-met decision. The weights are applied afterward, which is covered in rubric grading.

This makes the grader part of the instrument rather than a detail of the harness. HealthBench Hard reuses the original physician-written rubrics on the 1,000 conversations where frontier models did worst, and it keeps GPT-4.1 as its default grader for the same reason a lab keeps the assay fixed when it changes the sample.

Grading the grader

To justify a grader, you evaluate it against physicians. OpenAI compared candidate models with physician grades on the same criteria and reported macro-F1: GPT-4.1 at 0.709, o4-mini at 0.692, o3 at 0.681, GPT-4.1-mini at 0.661, and GPT-4.1-nano at 0.580. GPT-4.1 was selected on that result.

Two things are worth taking from those numbers. The stronger reasoning model did not win: o3 graded worse than GPT-4.1 here, so grading is a separate skill from answering and cannot be assumed to track capability. And the chosen grader's agreement exceeded the average physician's agreement in 5 of the 7 themes, which sets the realistic bar. The target is not a perfect key; it is how well one qualified human matches other qualified humans.

Where graders go wrong

Three failure modes come up repeatedly. Self-preference: a grader from the same model family as the system under test may reward its own style. Verbosity: a longer answer touches more criteria by covering more ground, which is why length-adjusted variants exist, penalizing about 2.99 points per 500 characters beyond 2,000 on HealthBench and about 1.47 on HealthBench Professional. Configuration: the grader's own settings count, and HealthBench Professional specifies not just GPT-5.4 but GPT-5.4 at low reasoning effort.

Designs exist that push back. Health Optimization Bench has each task authored by one model family and scored blind by a panel of three independent families, bars the authoring family from grading its own task, and publishes 95 percent bootstrap confidence intervals with every score. CHI-Bench splits the job: deterministic unit tests decide what can be checked mechanically, and an LLM judge handles only evidence grounding, consent, and cross-stage consistency.

What changing the grader does to scores

A score is a model plus a configuration, and the grader is part of the configuration. Two runs over the same tasks with different graders are two measurements, not one measurement repeated. HealthBench shows this in public: Anthropic reports a raw score under its own protocol, OpenAI leads with length-adjusted numbers at maximum reasoning effort and reports its production settings separately, and Baichuan ran its competitors itself. Those rows sit under one benchmark name and are not directly comparable.

Before you compare two numbers, check four things: the grader model, the grader's effort setting, whether length adjustment was applied, and the version of the task set. If any of them differ, part of the gap between the numbers is a gap between instruments. Reading a leaderboard covers the rest of the checks, and current standings with their configurations are at clinicalbenchmarks.ai.

how accurate is an LLM grader compared with a physician?

HealthBench chose its default grader by meta-evaluation against physician grades on the same criteria, reporting macro-F1 of 0.709 for GPT-4.1, ahead of o4-mini at 0.692, o3 at 0.681, GPT-4.1-mini at 0.661, and GPT-4.1-nano at 0.580. The grader's agreement exceeded the average physician's agreement in 5 of the 7 themes. That is agreement with how physicians grade, measured on a fixed criterion set, and it is not evidence the grader would be right about a case no physician wrote a rubric for.

does changing the grader model change benchmark scores?

Yes. The grader is part of the measuring instrument, so the same task set scored by a different grader, or by the same grader at a different reasoning effort, yields numbers that are not directly comparable. HealthBench uses GPT-4.1 as its default grader while HealthBench Professional uses GPT-5.4 at low reasoning effort with a length adjustment. Before comparing two scores, confirm the grader model, its effort setting, the length adjustment, and the task set version all match.