CliniCARE-Bench - Over-Abstention Rate: leaderboard
CliniCARE-Bench (arXiv 2608.07796v1, Table 3): agent systems audit de-identified EHR records against a clinical policy corpus across 25 clinician-built scenarios (750 cases, one run each) and return Yes, No, or Indeterminate (lack of data or medically ambiguous). Share of cases with a definitive reference verdict on which the system abstains, in percent; lower is better. Rows are harness plus model systems.
Metric: Abstentions on cases with a definitive reference verdict (%; lower is better). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.6 Flash | 7.3 | #57 |
| 2 | Gemini 3.5 Flash | 7.5 | #66 |
| 3 | GLM-5.2 | 7.5 | #80 |
| 4 | GPT-5.5 | 7.7 | #33 |
| 5 | DeepSeek V4 Flash | 7.8 | #155 |
| 6 | DeepSeek V4 Pro | 9.7 | #112 |
| 7 | MiniMax-M3 | 10.5 | #151 |
| 8 | Claude Opus 5 | 10.5 | #16 |
| 9 | Kimi K2.7 Code | 10.5 | #95 |
| 10 | Claude Sonnet 5 | 10.8 | #65 |
| 11 | Gemini 3.1 Pro (Preview) | 11.2 | #64 |
| 12 | GPT-5.4 | 11.2 | #83 |
| 13 | GPT-5.6 Sol | 12 | #22 |
| 14 | GPT-5.4 Mini | 12.2 | #228 |
| 15 | Qwen 3.7 Plus | 14 | #98 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=clinicare-bench-over-abstention-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.