PSF-Med - Frontier Hardest 500 - MIMIC-CXR: leaderboard
Tests whether medical image interpretations change when a question is paraphrased. Lower paraphrase flip rate indicates more consistent answers, not greater diagnostic accuracy. Open medical models use the full suite; frontier models use the hardest 500 cases. Dataset scores omit cells the publisher marks as degenerate always-yes behavior. This board reports PSF-Med - Frontier Hardest 500 - MIMIC-CXR using Paraphrase Flip Rate (%).
Metric: Paraphrase Flip Rate (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 5.21 | #287 |
| 2 | Kimi K3 (Low) | 6.85 | #35 (Kimi K3) |
| 3 | Claude Opus 5 (Thinking) | 7.88 | #16 (Claude Opus 5) |
| 4 | GPT-5.6 Sol (Low) | 8.93 | #22 (GPT-5.6 Sol) |
| 5 | GPT-5 Mini (Low) | 11.18 | #193 (GPT-5 Mini) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=psf-med-frontier-hardest-500-mimic-cxr · How It Works · Data refreshed daily, snapshot 2026-11-05.