MedSNIP-Bench: leaderboard
MedSNIP-Bench (Iqbal et al., 2026) tests medical fact verification on human-annotated snippets, clause groups that keep clinical dependencies together, drawn from answers to consumer-health and clinical-vignette questions. The verifier labels each snippet true or false from its own knowledge, and the score is F1 on the false class in percent; higher is better.
Metric: False-class F1 (%; F1 on false snippets among the 2,524 human-annotated snippets of 276 consumer-health and clinical-vignette answers, 13.6% false; each snippet verified on its own from parametric knowledge (claim-only mode); same prompt for every verifier). Source: arxiv.org. Saturation forecast: Around 2030. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (High) | 43.1 | #121 (GPT-5.4) |
| 2 | GPT-4o | 40.1 | #741 |
| 3 | GPT-OSS-20B | 40.1 | #1150 |
| 4 | Gemma 4 31B (IT) | 38.9 | #343 |
| 5 | Llama 3.3 70B Instruct | 33.1 | #1353 |
| 6 | Llama 3.1 8B Instruct | 27.8 | #4277 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=medsnip-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.