Classification on CIFAR-100 (Accuracy)
92.5AccuracyCBM-Suite
Evaluation Results
| Method | Links | |
|---|---|---|
| CBM-SuiteBottleneck type=CBMs with soft concepts, Backbone=ViT2026.07 | 92.5 | |
| HybridCBMBottleneck type=CBMs with soft concepts, Backbone=ViT2026.07 | 86.22 | |
| CaBMBottleneck type=Strictly text bottleneck, Protocol=text-only classifier2026.07 | 81.14 | |
| Linear ProbingLoss=SCL, Passes=T × N, Backbone=ResNet182026.05 | 73.9 | |
| Fixed PrototypesLoss=SCL, Passes=N, Backbone=ResNet182026.05 | 73.9 | |
| Normalized Linear ProbingLoss=SCL, Passes=T × N, Backbone=ResNet182026.05 | 73.6 | |
| PS-CBMBottleneck type=CBMs with soft concepts, Backbone=CNN2026.07 | 72.1 | |
| LF-CBMBottleneck type=CBMs with soft concepts, Backbone=CNN2026.07 | 68.8 | |
| VLG-CBMBottleneck type=CBMs with soft concepts, Backbone=CNN2026.07 | 68.3 | |
| LaBoBottleneck type=CBMs with soft concepts, Backbone=CNN2026.07 | 68.1 | |
| FLQMIAAlignment=frozen pretrained DINOv2 and GTE-1.5, Setting=Zero-shot2026.05 | 53.31 | |
| PCBMBottleneck type=CBMs with soft concepts, Backbone=CNN2026.07 | 52 | |
| FLVMIAAlignment=frozen pretrained DINOv2 and GTE-1.5, Setting=Zero-shot2026.05 | 51.69 | |
| SAIL w/ SigLIPAlignment=frozen pretrained DINOv2 and GTE-1.5, Loss=SigLIP, Setting=Zero-shot2026.05 | 50.5 | |
| SAIL w/ CLIPAlignment=frozen pretrained DINOv2 and GTE-1.5, Loss=InfoNCE, Setting=Zero-shot2026.05 | 45.22 | |
| CSAAlignment=frozen pretrained DINOv2 and GTE-1.5, Setting=Zero-shot2026.05 | 44.18 | |
| Multilinear-CovJac2026.05 | 28.91 | |
| M-CovJac#p/n=4, L=6, k=256k, iterations=100k, τ=1.02026.05 | 28.37 | |
| Soft-Mix2026.05 | 28.21 | |
| Gumbel-ST2026.05 | 28.08 | |
| Soft-Mix#p/n=16, L=6, k=256k, iterations=100k, τ=1.02026.05 | 27.92 | |
| Gumbel-ST#p/n=16, L=6, k=256k, iterations=100k, τ=1.02026.05 | 27.72 | |
| Multilinear-STE2026.05 | 25.02 | |
| M-STE#p/n=4, L=6, k=256k, iterations=100k, τ=1.02026.05 | 24.02 |