Image Classification on MNIST (test) (Standard Classification Report)
99.7AccuracyIndividuals
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| IndividualsBackbone=ViT-B/32, #Params=x8, Merging Strategy=N/A2026.05 | 99.7 | — | — | — | |
| Fine-tuned (STL)Backbone=CLIP-ViT-B/32, Evaluation Protocol=Fine-tuned (STL)2026.05 | 99.6 | — | — | — | |
| WEMoEBackbone=ViT-B/32, #Params=x6.27, Merging Strategy=Dynamic Merging2026.05 | 99.6 | — | — | — | |
| SMILEBackbone=ViT-B/32, #Params=x3.07, Merging Strategy=Dynamic Merging2026.05 | 99.6 | — | — | — | |
| WUDI-MergingBackbone=ViT-B/32, #Params=x1, Merging Strategy=Static Merging2026.05 | 99.5 | — | — | — | |
| FREE-MergingBackbone=ViT-B/32, #Params=x2.16, Merging Strategy=Dynamic Merging2026.05 | 99.5 | — | — | — | |
| DiDi-Merging-MBackbone=ViT-B/32, #Params=x1.4, Merging Strategy=Dynamic Merging2026.05 | 99.5 | — | — | — | |
| WUDI-Merging w/ FEATCALBackbone=CLIP-ViT-B/32, Upstream Merger=WUDI-Merging, Refinement=FEATCAL2026.05 | 99.4 | — | — | — | |
| Traditional MTLBackbone=CLIP-ViT-B/32, Evaluation Protocol=Traditional MTL2026.05 | 99.3 | — | — | — | |
| WUDI-MergingBackbone=CLIP-ViT-B/32, Upstream Merger=WUDI-Merging2026.05 | 99.3 | — | — | — | |
| WUDI-Merging w/ SurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=WUDI-Merging, Refinement=Surgery2026.05 | 99.3 | — | — | — | |
| WUDI-Merging w/ ProbSurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=WUDI-Merging, Refinement=ProbSurgery2026.05 | 99.3 | — | — | — | |
| TSV-MergingBackbone=ViT-B/32, #Params=x1, Merging Strategy=Static Merging2026.05 | 99.3 | — | — | — | |
| TSV-CBackbone=ViT-B/32, #Params=x2.08, Merging Strategy=Dynamic Merging2026.05 | 99.3 | — | — | — | |
| TeacherArchitecture relationship=Teacher2026.04 | 99.28 | — | — | — | |
| DiDi-Merging-SBackbone=ViT-B/32, #Params=x1.24, Merging Strategy=Dynamic Merging2026.05 | 99.2 | — | — | — | |
| TALL-Mask+TIESBackbone=ViT-B/32, #Params=x2.5, Merging Strategy=Dynamic Merging2026.05 | 99.1 | — | — | — | |
| Twin-MergingBackbone=ViT-B/32, #Params=x2.25, Merging Strategy=Dynamic Merging2026.05 | 99.1 | — | — | — | |
| EMR-MergingBackbone=ViT-B/32, #Params=x4, Merging Strategy=Dynamic Merging2026.05 | 99.1 | — | — | — | |
| AdaMerging w/ FEATCALBackbone=CLIP-ViT-B/32, Upstream Merger=AdaMerging, Refinement=FEATCAL2026.05 | 99 | — | — | — | |
| DivBFKDArchitecture relationship=Same, Knowledge Distillation access type=Black-box, Real image budget (N)=2 K, Synthetic image budget (M)=24 K2026.04 | 98.97 | — | — | — | |
| Student-FullArchitecture relationship=Full2026.04 | 98.91 | — | — | — | |
| FS-BBTArchitecture relationship=Same, Knowledge Distillation access type=Black-box, Real image budget (N)=2 K, Synthetic image budget (M)=24 K2026.04 | 98.91 | — | — | — | |
| Task Arithmetic w/ FEATCALBackbone=CLIP-ViT-B/32, Upstream Merger=Task Arithmetic, Refinement=FEATCAL2026.05 | 98.8 | — | — | — | |
| AdaMerging w/ ProbSurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=AdaMerging, Refinement=ProbSurgery2026.05 | 98.8 | — | — | — | |
| BBKDArchitecture relationship=Same, Knowledge Distillation access type=Black-box, Real image budget (N)=2 K, Synthetic image budget (M)=24 K2026.04 | 98.74 | — | — | — | |
| AdaMerging w/ SurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=AdaMerging, Refinement=Surgery2026.05 | 98.7 | — | — | — | |
| IAdaPID-ADGBackbone=Multi-Layer Neural Network (MLNN), Batch size=128, Epochs=202026.05 | 98.68 | — | — | — | |
| CAT-MergingBackbone=ViT-B/32, #Params=x1, Merging Strategy=Static Merging2026.05 | 98.6 | — | — | — | |
| DivBFKDArchitecture relationship=Smaller, Knowledge Distillation access type=Black-box, Real image budget (N)=2 K, Synthetic image budget (M)=24 K2026.04 | 98.58 | — | — | — | |
| FS-BBTArchitecture relationship=Smaller, Knowledge Distillation access type=Black-box, Real image budget (N)=2 K, Synthetic image budget (M)=24 K2026.04 | 98.42 | — | — | — | |
| Kernel CurvSSLBackbone=ResNet-18, Evaluation Protocol=Linear evaluation2025.11 | 98.4 | — | — | — | |
| AdamWweight decay=0.01, learning rate=10^−3, batch size=128, seeds=32026.05 | 98.22 | — | — | — | |
| DiffGradBackbone=Multi-Layer Neural Network (MLNN), Batch size=128, Epochs=202026.05 | 98.22 | — | — | — | |
| AdaMergingBackbone=CLIP-ViT-B/32, Upstream Merger=AdaMerging2026.05 | 98.2 | — | — | — | |
| GONOlearning rate=10^−3, batch size=128, seeds=32026.05 | 98.15 | — | — | — | |
| AMSGradBackbone=Multi-Layer Neural Network (MLNN), Batch size=128, Epochs=202026.05 | 98.15 | — | — | — | |
| AdaPIDBackbone=Multi-Layer Neural Network (MLNN), Batch size=128, Epochs=202026.05 | 98.11 | — | — | — | |
| Task Arithmetic w/ SurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=Task Arithmetic, Refinement=Surgery2026.05 | 98 | — | — | — | |
| BPL=42026.06 | 98 | — | — | — | |
| Simple Averaging w/ FEATCALBackbone=CLIP-ViT-B/32, Upstream Merger=Simple Averaging, Refinement=FEATCAL2026.05 | 97.9 | — | — | — | |
| Task Arithmetic w/ ProbSurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=Task Arithmetic, Refinement=ProbSurgery2026.05 | 97.9 | — | — | — | |
| CurvSSLBackbone=ResNet-18, Evaluation Protocol=Linear evaluation2025.11 | 97.9 | — | — | — | |
| DBPTTModel variant=DBPTT2026.05 | 97.87 | — | — | — | |
| Vanilla PCL=42026.06 | 97.8 | — | — | — | |
| HEPL=42026.06 | 97.8 | — | — | — | |
| BPL=82026.06 | 97.8 | — | — | — | |
| BPL=162026.06 | 97.7 | — | — | — | |
| Topology-guided HDCSetting=Baseline (clean), Training status=After train2026.05 | 97.68 | — | — | — | |
| Topology-guided HDCSetting=Rotation (20°), Training status=After train2026.05 | 97.61 | — | — | — | |
| EPModel variant=EP2026.05 | 97.45 | — | — | — | |
| DEPModel variant=DEP2026.05 | 97.43 | — | — | — | |
| ISO-CTSBackbone=ViT-B/32, #Params=x1, Merging Strategy=Static Merging2026.05 | 97.4 | — | — | — | |
| SGD-Momentummomentum=0.9, learning rate=10^−3, batch size=128, seeds=32026.05 | 97.23 | — | — | — | |
| MLPHyperpars=h = 64, Params=50,8902026.06 | 97.14 | — | — | — | |
| BPL=322026.06 | 97.1 | — | — | — | |
| Adamlearning rate=10^−3, batch size=128, seeds=32026.05 | 97.08 | — | — | — | |
| Topology-guided HDCSetting=Cut out (size = 4), Training status=After train2026.05 | 97.03 | — | — | — | |
| MLPParams=643,538, Time (s)=194.662026.04 | 96.95 | 96.92 | 96.92 | 96.92 | |
| Simple Averaging w/ SurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=Simple Averaging, Refinement=Surgery2026.05 | 96.8 | — | — | — | |
| Simple Averaging w/ ProbSurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=Simple Averaging, Refinement=ProbSurgery2026.05 | 96.5 | — | — | — | |
| Public DP-KFCepsilon=8, Backbone=CNN, Seeds=52026.05 | 96.4 | — | — | — | |
| Synthetic DP-KFC + DP-AdamBCepsilon=8, Backbone=CNN, Seeds=52026.05 | 96.4 | — | — | — | |
| HEPL=82026.06 | 96.4 | — | — | — | |
| HEPL=162026.06 | 96.4 | — | — | — | |
| LTBs-KANParams=203,378, Time (s)=244.272026.04 | 96.32 | 96.31 | 96.27 | 96.28 | |
| FastKANParams=230,346, Time (s)=259.132026.04 | 96.19 | 96.18 | 96.13 | 96.14 | |
| Task ArithmeticBackbone=CLIP-ViT-B/32, Upstream Merger=Task Arithmetic2026.05 | 96.1 | — | — | — | |
| Synthetic DP-KFC + DP-AdamBCepsilon=2, Backbone=CNN, Seeds=52026.05 | 96.1 | — | — | — | |
| FasterKANParams=204,896, Time (s)=237.572026.04 | 96.02 | 95.99 | 95.95 | 95.97 | |
| Synthetic DP-KFCepsilon=8, Backbone=CNN, Seeds=52026.05 | 95.9 | — | — | — | |
| VicRegBackbone=ResNet-18, Evaluation Protocol=Linear evaluation2025.11 | 95.9 | — | — | — | |
| HEPL=642026.06 | 95.9 | — | — | — | |
| GottliebKANParams=105,859, Time (s)=263.172026.04 | 95.81 | 95.75 | 95.78 | 95.76 | |
| HEPL=1282026.06 | 95.8 | — | — | — | |
| Public DP-KFCepsilon=2, Backbone=CNN, Seeds=52026.05 | 95.7 | — | — | — | |
| HEPL=322026.06 | 95.6 | — | — | — | |
| Synthetic DP-KFC + DP-AdamBCepsilon=1, Backbone=CNN, Seeds=52026.05 | 95.5 | — | — | — | |
| Naive HDCSetting=Baseline (clean), Training status=After train2026.05 | 95.5 | — | — | — | |
| Standard-KDArchitecture relationship=Smaller, Knowledge Distillation access type=Black-box, Real image budget (N)=2 K, Synthetic image budget (M)=24 K2026.04 | 95.36 | — | — | — | |
| DP-AdamBCepsilon=8, Backbone=CNN, Seeds=52026.05 | 95.3 | — | — | — | |
| Public DP-KFCepsilon=1, Backbone=CNN, Seeds=52026.05 | 95.3 | — | — | — | |
| HalfNetHyperpars=k = 16, Params=14,1062026.06 | 95.3 | — | — | — | |
| Student-AloneArchitecture relationship=Smaller, Knowledge Distillation access type=N/A, Real image budget (N)=2 K, Synthetic image budget (M)=24 K2026.04 | 95.14 | — | — | — | |
| MLPHyperpars=h = 20, Params=15,9102026.06 | 95.04 | — | — | — | |
| Synthetic DP-KFCepsilon=2, Backbone=CNN, Seeds=52026.05 | 95 | — | — | — | |
| BPL=1282026.06 | 95 | — | — | — | |
| Barlow TwinsBackbone=ResNet-18, Evaluation Protocol=Linear evaluation2025.11 | 94.9 | — | — | — | |
| BPL=642026.06 | 94.9 | — | — | — | |
| DP-AdamBCepsilon=2, Backbone=CNN, Seeds=52026.05 | 94.8 | — | — | — | |
| EfficientKANParams=254,080, Time (s)=366.652026.04 | 94.38 | 94.89 | 94.44 | 94.48 | |
| HalfNetHyperpars=k = 8, Params=7,8342026.06 | 94.34 | — | — | — | |
| DiSKepsilon=8, Backbone=CNN, Seeds=52026.05 | 94.3 | — | — | — | |
| BSRBF-KANParams=306,528, Time (s)=517.562026.04 | 94.23 | 94.37 | 94.14 | 94.19 | |
| Synthetic DP-KFCepsilon=1, Backbone=CNN, Seeds=52026.05 | 94.2 | — | — | — | |
| DiSKepsilon=2, Backbone=CNN, Seeds=52026.05 | 94.1 | — | — | — | |
| DP-AdamBCepsilon=1, Backbone=CNN, Seeds=52026.05 | 94 | — | — | — | |
| Topology-guided HDCSetting=Baseline (clean), Training status=Before train2026.05 | 93.96 | — | — | — | |
| Topology-guided HDCSetting=Rotation (20°), Training status=Before train2026.05 | 93.89 | — | — | — | |
| DP-SGDepsilon=8, Backbone=CNN, Seeds=52026.05 | 93.7 | — | — | — |