HealthBench Consensus: Medical LLM Benchmark

• HealthBench Consensus is a streamlined evaluation tool that focuses on 34 high-impact criteria to measure LLM safety and reliability in medical dialogues.
• It employs a rigorous multi-rater physician consensus to validate criteria and assess improvements in emergency referrals, context seeking, and expert communication.
• Benchmark results demonstrate significant error rate reductions across successive LLM versions, underscoring the efficacy of a consensus-driven approach.

HealthBench Consensus is a publicly released variant of the HealthBench evaluation suite for LLMs in healthcare, specifically designed to provide high-precision, physician-validated measurement of critical model behaviors in medical dialogues. It distills the comprehensive HealthBench rubric—spanning 48,562 criteria—down to 34 consensus criteria that have been validated for high impact and broad applicability by multiple independent clinicians. This consensus-driven approach aims to facilitate precise benchmarking of safety-critical and clinically relevant LLM competencies, informing both model development cycles and deployment decisions in healthcare settings (Arora et al., 13 May 2025).

1. Design Rationale and Definition

HealthBench Consensus was developed to address key limitations of broad, rubric-based evaluation frameworks: namely, noise from less critical criteria and the risk of penalizing models for failing dimensions irrelevant to clinical safety. In contrast to the base HealthBench suite, which prioritizes breadth by evaluating model performance over tens of thousands of rubric criteria, HealthBench Consensus narrows the focus to the most essential behaviors, as endorsed by clinicians through a multi-rater consensus process. This allows for rapid, noise-minimized measurement of models’ performance on safety-critical dimensions such as emergency triage, clarity of referral, and context-seeking behavior.

Key facts:

• Includes only the 34 high-impact consensus dimensions, each validated to be broadly applicable and clinically consequential.
• Applied to the subset of 3,671 examples for which at least one consensus criterion is relevant.
• Serves both as a tool for rapid tracking of core clinical safety behaviors and as a complement to the full HealthBench suite for surfacing narrow failure modes with minimal confounding signal (Arora et al., 13 May 2025).

2. Consensus Criterion Selection and Validation

The construction of the 34 consensus criteria was a multi-stage process:

1. Rubric Authoring: 262 physicians authored 48,562 conversation-specific criteria, spanning seven clinical themes and five behavioral axes.
2. Clinical Categorization: For each theme, examples were categorized (e.g., “emergent,” “non-emergent”).
3. Criterion Proposal: Physician advisors pre-wrote a candidate set of criteria per category (e.g., “Does the response include a clear emergency referral?”).
4. Multi-Rater Review: Each candidate criterion was independently reviewed by at least two physicians. If >50% of raters agreed on a criterion’s relevance for an example, it was retained as a consensus dimension for that example.
5. Meta-Evaluation: Multiple physicians graded whether model responses met each consensus criterion; this produced approximately 60,896 “meta-examples” for transparent downstream analysis of grading trustworthiness.

No formal Delphi process was used; a simple >50% majority sufficed to establish consensus. This pragmatic approach enabled efficient, scalable validation of critical evaluation dimensions (Arora et al., 13 May 2025).

3. Scoring, Aggregation, and Evaluation Metrics

The consensus scoring pipeline is grounded in explicit, mathematically defined aggregation formulas:

• Per-Example Consensus Score: For example $i$ with assigned consensus criteria $C_i$:

$$s^{(cons)}_i = \frac{\sum_{j\in C_i} r_{ij}p_{ij}}{\sum_{j\in C_i} p_{ij}}$$

where $r_{ij} = 1$ if criterion $j$ is met, $0$ otherwise, and $p_{ij} \in (0,10]$ is the positive, safety-weighted point value.

• Clipping: Scores are clipped to $[0,1]$, avoiding negative marking.
• Aggregate HealthBench Consensus Score:

$S^{(cons)} = \frac{1}{M}\sum_i s^{(cons)}_i$

where the sum is over $M$ examples with at least one relevant consensus criterion. Per-dimension aggregations (e.g., by clinical theme) are supported via restriction of $C_i$ to the target subset.

Three principal evaluation metrics are implemented:

1. Error Rate: Defined as $1 - S^{(cons)}$.
2. Meta-Evaluation MF1: Agreement between model-based graders and physicians, computed as macro-averaged F1 over the binary “met”/“not met” class for each criterion.
3. Worst-at-k Reliability: Measures expected minimum score among $k$ independent model outputs to estimate reliability under multiple sampling; this is optional but informative for deployment robustness (Arora et al., 13 May 2025).

4. Empirical Benchmark Results and Comparative Performance

HealthBench Consensus enables precise tracking of LLM safety advances:

• From GPT-3.5 Turbo to GPT-4.1, the overall error rate dropped from ≈25% to ≈6%; GPT-4o (Aug 2024) scored ≈12%, and o3 (Apr 2025) ≈7%.
• Dimension-specific trends:
  • Emergency Referrals (“Emergency behavior” criterion): error reduced from ≈22% (GPT-3.5 Turbo) to ≈8% (GPT-4.1), further to ≈6% (o3).
  • Context-Seeking remains challenging (GPT-4.1 ≈81% success).
  • Expertise-Tailored Communication: accuracy/completeness for GPT-4.1 and o3 exceeds 99%.

A plausible implication is that focusing model development and evaluation on consensus dimensions yields rapid progress on safety-critical behaviors, with error rates reduced more than four-fold on these axes (Arora et al., 13 May 2025).

5. Methodological Innovations and Meta-Evaluation

HealthBench Consensus institutionalizes rigorous meta-evaluation practices:

• Each criterion-example pair (“meta-example”) is graded by multiple physicians, enabling robust calculation of agreement rates and alignment metrics for both human and model-based graders.
• GPT-4.1 used as an automated grader achieved MF1 scores in the 70th–88th percentile relative to average physician performance across themes, indicating that machine evaluation is becoming at least as reliable as human rater consensus on these dimensions.
• The dataset structure supports systematic auditing and further calibration of automated grading, establishing a transparent, reproducible measurement ecosystem (Arora et al., 13 May 2025).

6. Context, Applications, and Future Directions

HealthBench Consensus is positioned both as a practical tool for rapid iteration on core safety dimensions and as a foundation for ongoing benchmark stewardship:

• It complements full-scale HealthBench, enabling detection of narrow but clinically impactful model failure modes with minimal noise.
• In development cycles, Consensus can be used to triage model changes, prioritize high-impact improvements, and validate model-based graders before extensive deployment.
• The consensus-based selection and meta-evaluation process is extendable: new clinical domains can be incorporated by repeating the physician-majority voting and meta-evaluation sequence with relevant experts.
• Regular monitoring and reporting of worst-at-k reliability on Consensus axes is recommended to guard against rare, catastrophic failure scenarios in clinical deployments (Arora et al., 13 May 2025).

Consensus-building implications from broader benchmark governance frameworks reinforce the need for rigorous, multi-axis evaluation and dynamic lifecycle management. Although not specific to HealthBench Consensus, the Benchmark Health Index (BHI) suggests that high-precision, consensus-validated benchmarks should be integrated into continual auditing frameworks that track discrimination, anti-saturation, and impact over time (Zhu et al., 12 Feb 2026). This suggests that regular reassessment and dynamic updating of consensus criteria may be beneficial as model capabilities and evaluation needs evolve.