Modular arithmetic prediction on MNIST modular arithmetic Unseen operations (test)
99.46Zero-Shot Op AccuracyResNet w/ multi-frequency rotation (MFR)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ResNet w/ multi-frequency rotation (MFR)Backbone=ResNet, Training Objective=BRo-JEPA, Rotation type (SFR/MFR)=MFR, Embedding type=block rotation2026.05 | 99.46 | 99.46 | 70.9 | |
| ResNet18 w/ compositional consistency lossBackbone=ResNet18, Training Objective=JEPA, Compositional consistency loss=true, Embedding type=additive2026.05 | 99.31 | 99.31 | 78.39 | |
| ResNet w/ single-frequency rotation (SFR)Backbone=ResNet, Training Objective=BRo-JEPA, Rotation type (SFR/MFR)=SFR, Embedding type=block rotation2026.05 | 98.93 | 98.93 | 69.22 | |
| MLP w/ compositional consistency lossBackbone=MLP, Training Objective=JEPA, Compositional consistency loss=true, Embedding type=additive2026.05 | 97.77 | 97.86 | 97.81 | |
| MLP w/ multi-frequency rotation (MFR)Backbone=MLP, Training Objective=BRo-JEPA, Rotation type (SFR/MFR)=MFR, Embedding type=block rotation2026.05 | 97.24 | 97.21 | 97.23 | |
| MLP w/ single-frequency rotation (SFR)Backbone=MLP, Training Objective=BRo-JEPA, Rotation type (SFR/MFR)=SFR, Embedding type=block rotation2026.05 | 94.26 | 94.23 | 94.28 | |
| MLP w/ multi-frequency rotation (MFR)Backbone=MLP, Training Objective=Supervised, Rotation type (SFR/MFR)=MFR, Embedding type=block rotation2026.05 | 56.5 | — | 53.59 | |
| ResNet18 w/ multi-frequency rotation (MFR)Backbone=ResNet18, Training Objective=Supervised, Rotation type (SFR/MFR)=MFR, Embedding type=block rotation2026.05 | 54.92 | — | 50.41 | |
| MLP w/ single-frequency rotation (SFR)Backbone=MLP, Training Objective=Supervised, Rotation type (SFR/MFR)=SFR, Embedding type=block rotation2026.05 | 13.92 | — | 14.22 | |
| ResNet18 w/ single-frequency rotation (SFR)Backbone=ResNet18, Training Objective=Supervised, Rotation type (SFR/MFR)=SFR, Embedding type=block rotation2026.05 | 10.1 | — | 12.78 | |
| MLPBackbone=MLP, Training Objective=JEPA, Embedding type=additive2026.05 | 9.95 | 97.93 | 10.44 | |
| ResNet18Backbone=ResNet18, Training Objective=JEPA, Embedding type=additive2026.05 | 9.74 | 99.47 | 9.03 | |
| MLPBackbone=MLP, Training Objective=Supervised, Embedding type=additive2026.05 | 5.93 | — | 1.8 | |
| ResNet18Backbone=ResNet18, Training Objective=Supervised, Embedding type=additive2026.05 | 2.42 | — | 0.11 |