Medical Logical Reasoning (Rule 3: is_mel_on_back) on HAM 5000 samples (test)
99.8F1 (NS-CL)GroundTruth (Oracle)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GroundTruth (Oracle)Supervision (%)=Oracle/GT2026.05 | 99.8 | 100 | |
| GlobalVAESupervision (%)=75%2026.05 | 90.8 | 88.4 | |
| SlotVAE (k=1)Supervision (%)=75%2026.05 | 60.6 | 40 | |
| DINOv2Supervision (%)=75%2026.05 | 0.8 | 18.4 |