MC-CXR - Misleading Prior Image: leaderboard
Accuracy on MC-CXR (Seoul National University, 2026) when the chest X-ray comes with a cross-patient prior chest X-ray whose labels include a plausible wrong finding; the model must still name the image finding among 12 CheXpert labels. Percent, higher is better.
Metric: Accuracy under misleading context (%; the image plus a cross-patient prior chest X-ray whose labels include the wrong finding; 240 radiologist-reviewed MIMIC-CXR cases, one current chest X-ray and a fixed target finding each; zero-shot constrained single-letter answer from 12 CheXpert labels (chance about 8.3%), invalid answers scored wrong; greedy or temperature-0 decoding). Source: arxiv.org. Saturation forecast: Around 2031. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 30.8 | #803 |
| 2 | InternVL3-8B | 22.9 | #1688 |
| 3 | Gemini 3.5 Flash (Minimal) | 22.9 | #101 (Gemini 3.5 Flash) |
| 4 | Claude Opus 4.7 | 16.7 | #79 |
| 5 | MedGemma 1.5 4B | 16.2 | #2562 |
| 6 | Gemma 3 12B (IT) | 10 | #1837 |
| 7 | Phi-4 Multimodal Instruct | 9.2 | #3882 |
| 8 | Llama 3.2 11B Instruct | 8.8 | #4991 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mc-cxr-misleading-prior-image · How It Works · Data refreshed daily, snapshot 2026-11-13.