ObGynLongBench - History-Level EHR - L1 Single Evidence: leaderboard
ObGynLongBench (arXiv 2609.07601v1, Table 2): 17 LLMs answer 1,500 rule-grounded obstetric multiple-choice decision points from real longitudinal EHRs of 976 pregnant patients; open-source and medical models run with thinking disabled. Full-history input, cases whose rule needs recognition of a single piece of evidence; accuracy percent, higher is better.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around July 2028. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Flash | 66.5 | #155 |
| 2 | Gemini 3 Flash | 66 | #103 |
| 3 | Gemma 4 26B A4B (Non-reasoning) | 65 | #360 (Gemma 4 26B A4B) |
| 4 | Qwen 3.5 4B (Non-reasoning) | 63.2 | #514 (Qwen 3.5 4B) |
| 5 | Claude Haiku 4.5 | 63 | #287 |
| 6 | Qwen 3.5 9B (Non-reasoning) | 62.7 | #391 (Qwen 3.5 9B) |
| 7 | GPT-5.4 Mini | 61.8 | #228 |
| 8 | Qwen 3.5 35B A3B (Non-reasoning) | 58 | #268 (Qwen 3.5 35B A3B) |
| 9 | Phi-4 Mini Instruct | 56.2 | #1232 |
| 10 | Gemma 4 E4B (Non-reasoning) | 56 | #762 (Gemma 4 E4B) |
| 11 | MedGemma 1.5 4B | 45.8 | #788 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=obgynlongbench-history-level-ehr-l1-single-evidence · How It Works · Data refreshed daily, snapshot 2026-11-05.