Medical Logical Reasoning (Rule 1: is_malignant) on HAM 5000 samples (test)
100F1-score (NS-CL)GroundTruth (Oracle)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GroundTruth (Oracle)Supervision (%)=Oracle/GT2026.05 | 100 | 100 | |
| GlobalVAESupervision (%)=75%2026.05 | 92.9 | 92.8 | |
| SlotVAE (k=1)Supervision (%)=75%2026.05 | 75.9 | 75.1 | |
| DINOv2Supervision (%)=1%2026.05 | 37 | 29.8 | |
| ResNet+SA (k=2)Supervision (%)=1%2026.05 | 29 | 18.8 | |
| GlobalVAESupervision (%)=1%2026.05 | 16.6 | 13.7 | |
| SlotVAE (k=1)Supervision (%)=1%2026.05 | 13.6 | 12.5 | |
| SlotVAE (k=2)Supervision (%)=1%2026.05 | 8.3 | 15.7 | |
| SlotVAE (k=10)Supervision (%)=1%2026.05 | 0.1 | 1.7 | |
| SlotVAE (k=5)Supervision (%)=1%2026.05 | 0 | 11.7 |