Medical Visual Question Answering on OmniMedVQA (Accuracy, AUROC, ECE, Brier Score)
5ECESteerConf
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SteerConfModel Architecture=MedGemma-4B-IT, Training Pathway=OmniMedVQA-trained2026.06 | 5 | 19.6 | 69.3 | 69.6 | |
| ConfTunerModel Architecture=MedGemma-4B-IT, Training Pathway=OmniMedVQA-trained2026.06 | 7.4 | 19.4 | 70.8 | 69.3 | |
| Base ModelModel Architecture=MedGemma-4B-IT, Training Pathway=OmniMedVQA-trained2026.06 | 14.6 | 21.9 | 62.8 | 70.5 | |
| OursModel Architecture=MedGemma-4B-IT, Training Pathway=OmniMedVQA-trained2026.06 | 18.4 | 11.1 | 92.6 | 80.5 | |
| Top-K SamplingModel Architecture=MedGemma-4B-IT, Training Pathway=OmniMedVQA-trained2026.06 | 40.3 | 40.6 | 55.3 | 45.9 |