List your company
ResearchFormerly WizWand

Class-Incremental Learning on VTAB

96.7Avg AccuracySoTU
51.647263.343675.0486.7364Jul 5, 2023Dec 26, 2023Jun 18, 2024Dec 9, 2024Jun 2, 2025Nov 23, 2025May 17, 2026
Updated 5mo ago

Evaluation results

MethodLinks
SoTUBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=102024.1196.7——93.4
RanPACBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=102024.1195.3——92.2
StaR-MoEBackbone=ViT-B/16-IN1K2026.0594.4393.79——
DERBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=10, Examples per class=202024.1193.1——89.8
RanPACalgorithm=Algorithm 12023.0792.6———
CoOp + OursBackbone=ViT-B/162024.0392.5191.28——
Joint full fine-tuningstrategy=full fine-tuning, backbone=ViT2023.0792———
SEMABackbone=ViT-B/16-IN21K2024.0391.99——90.86
ADAM (SSF)Backbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=10, variant=SSF2024.1191.6——84.3
OursBackbone=ViT-B/162024.0391.3789.67——
SEMA2024.0391.2689.64——
SEMABackbone=ViT-B/16-IN1K2024.0391.26——89.64
MaPLe + OursBackbone=ViT-B/162024.0390.9788.83——
SEMABackbone=ViT-B/16-IN1K2026.0590.9388.12——
SimpleCILBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=102024.1190.9——83.6
ADAM (Ada)Backbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=10, variant=Ada2024.1190.9——83.6
Ours w/o VIBackbone=ViT-B/162024.0390.7488.64——
SD-LoRABackbone=ViT-B/16-IN1K2026.059087.61——
InfLORABackbone=ViT-B/16-IN21K2024.0389.61——86.52
InfLoRABackbone=ViT-B/16-IN1K2024.0388.9——87.63
CODA-PromptBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=102024.1188.6——89.3
InfLoRABackbone=ViT-B/16-IN1K2026.0588.4388.54——
DualPromptBackbone=ViT-B/16-IN21K2024.0388.11——77.58
FOSTERBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=10, Examples per class=202024.1187.9——80.2
CODA-PBackbone=ViT-B/162024.0387.581.2——
CoOpBackbone=ViT-B/162024.0387.0681.25——
APER + SSFBackbone=ViT-B/16-IN1K2026.0586.7584.51——
DualPromptBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=102024.1186.2——73.6
CODA-PromptBackbone=ViT-B/16-IN1K2026.0586.1686.34——
SimpleCILBackbone=ViT-B/16-IN21K2024.0385.99——84.38
ADAMBackbone=ViT-B/16-IN21K2024.0385.95——84.35
ESNBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=102024.1185.9——85.2
SimpleCILBackbone=ViT-B/16-IN1K2026.0585.4283.61——
APER + AdapterBackbone=ViT-B/16-IN1K2026.0585.4283.59——
SimpleCILBackbone=ViT-B/16-IN1K2024.0385.29——83.61
ADAMBackbone=ViT-B/16-IN1K2024.0385.29——83.61
CODA-PBackbone=ViT-B/16-IN21K2024.0385.13——85.85
HiDe-PromptBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=102024.1184.8——84.6
MaPLeBackbone=ViT-B/162024.0383.9181.81——
L2PBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=102024.1183.4——61.7
DualPromptBackbone=ViT-B/16-IN1K2024.0382.89——79.79
DualPromptBackbone=ViT-B/16-IN1K2026.0582.8979.79——
MEMOBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=10, Examples per class=202024.1181.5——86.9
L2PBackbone=ViT-B/16-IN1K2024.0381.19——80.83
L2PBackbone=ViT-B/16-IN1K2026.0581.1980.83——
CODA-PBackbone=ViT-B/16-IN1K2024.0379.88——81.58
L2PBackbone=ViT-B/16-IN21K2024.0379.17——63.56
AttriCLIP + OursBackbone=ViT-B/162024.0374.8471.12——
iCaRLBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=10, Examples per class=202024.1174——85.9
AttriCLIPBackbone=ViT-B/162024.0371.8464.09——
Continual-CLIPBackbone=ViT-B/162024.0368.560.97——
CLIP-AdapterBackbone=ViT-B/162024.0368.2361.02——
VPTBackbone=ViT-B/162024.0367.257.26——
FT AdapterBackbone=ViT-B/16-IN1K2024.0359.98——43.5
iCaRLBackbone=ViT-B/162024.0353.3841.6——
Adam-AdaptBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——83.6—
Adam-AdaptBackbone=ViT-B/16, Pre-training=ImageNet-21K, Task Setting=2 init classes, 1 class per task2024.02——84.4—
Adam-FTBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——83.8—
Adam-FTBackbone=ViT-B/16, Pre-training=ImageNet-21K, Task Setting=2 init classes, 1 class per task2024.02——84.4—
Adam-SSFBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——84.2—
Adam-SSFBackbone=ViT-B/16, Pre-training=ImageNet-21K, Task Setting=2 init classes, 1 class per task2024.02——84.9—
Adam-VPTBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——86.9—
Adam-VPTBackbone=ViT-B/16, Pre-training=ImageNet-21K, Task Setting=2 init classes, 1 class per task2024.02——81.1—
CODA-PromptBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——6.3—
DualPromptBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——4.2—
FinetuneBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——2.1—
FinetuneBackbone=ViT-B/16, Pre-training=ImageNet-21K, Task Setting=2 init classes, 1 class per task2024.02——0.5—
Joint full tuningBackbone=ViT-B/16, Pre-training=ImageNet21K, Tasks=102024.11———92
L2PBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——1.3—
NCMBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——88.2—
NCMBackbone=ViT-B/16, Pre-training=ImageNet-21K, Task Setting=2 init classes, 1 class per task2024.02——88.5—
RanDumbBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——93.6—
RanDumbBackbone=ViT-B/16, Pre-training=ImageNet-21K, Task Setting=2 init classes, 1 class per task2024.02——92.4—
RanPACBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——93—
RanPACBackbone=ViT-B/16, Pre-training=ImageNet-21K, Task Setting=2 init classes, 1 class per task2024.02——88.2—
RanPAC-impBackbone=ViT-B/16, Pre-training=ImageNet-1K, Task Setting=2 init classes, 1 class per task2024.02——93—
RanPAC-impBackbone=ViT-B/16, Pre-training=ImageNet-21K, Task Setting=2 init classes, 1 class per task2024.02——91.9—