EPAG - HPI Guideline Score: leaderboard
EPAG pre-consultation (Seo et al., 2026): the model acts as a doctor agent asking five turns of multiple-choice history questions of a GPT-4o-mini patient agent, in Korean, over 520 patient profiles covering 26 diseases. The collected history of present illness is split into information units and matched against clinician-curated diagnostic guidelines for the patient's disease. Mean number of guideline-matched units per dialogue (an unbounded count); higher is better.
Metric: Guideline-Matched History Units per Dialogue (unweighted). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 | 4.82 | #492 |
| 2 | Claude 3.5 Sonnet | 4.62 | #747 |
| 3 | Claude 3.7 Sonnet | 4.59 | #455 |
| 4 | Claude 3.5 Haiku | 4.58 | #1563 |
| 5 | GPT-4.1 Mini | 4.46 | #775 |
| 6 | GPT-4o Mini | 4.46 | #1627 |
| 7 | GPT-4o | 4.39 | #741 |
| 8 | MedGemma-4B | 4.19 | #2460 |
| 9 | Phi-3.5-mini-instruct | 3.91 | #5757 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=epag-hpi-guideline-score · How It Works · Data refreshed daily, snapshot 2026-11-12.