Image Classification on GTSRB
99.8AccuracyHAWX
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| HAWXBackbone=EfficientNet, Configuration=1-47:KV8, Power=0.993, Area=0.9742026.02 | 99.8 | — | — | — | — | |
| HAWXBackbone=EfficientNet, Configuration=1-45,47:KV8 | 46:(KV8,KV9), Power=0.993, Area=0.9722026.02 | 99.8 | — | — | — | — | |
| CNN3ST2021.02 | 99.71 | — | — | — | — | |
| Sill-Net2021.02 | 99.68 | — | — | — | — | |
| HAWXBackbone=EfficientNet, Configuration=1-47:KV6, Power=1, Area=12026.02 | 99.65 | — | — | — | — | |
| Multiple modelsBackbone=ViT-L/142022.08 | 99.3 | — | — | — | — | |
| PaRaMSModel Status=Protected Standalone (θ̂_def)2025.11 | 99.24 | — | — | — | — | |
| fine-tunedBackbone=ViT-L/142026.02 | 99.2 | — | — | — | — | |
| theta_B fine-tuneBackbone=ViT-L/142025.10 | 99.16 | — | — | — | — | |
| MMFTModel=ViT, Method=Ours2026.01 | 99.08 | — | — | — | — | |
| Multiple modelsBackbone=ViT-B/162022.08 | 99 | — | — | — | — | |
| fine-tunedBackbone=ViT-B/162026.02 | 99 | — | — | — | — | |
| FTModel=ViT, Method=FT2026.01 | 98.93 | — | — | — | — | |
| Joint patchingBackbone=ViT-L/142022.08 | 98.9 | — | — | — | — | |
| fine-tunedBackbone=ViT-B/322026.02 | 98.9 | — | — | — | — | |
| Fine-TunedBackbone=DINOv3 ViT-B/162025.12 | 98.9 | — | — | — | — | |
| FLYPModel=ViT, Method=FLYP2026.01 | 98.83 | — | — | — | — | |
| θBmode=fine-tune2026.02 | 98.77 | — | — | — | — | |
| θBMode=fine-tune2026.02 | 98.77 | — | — | — | — | |
| theta_A fine-tuneModel=ViT-B/16, Mode=fine-tune2026.02 | 98.77 | — | — | — | — | |
| Fine-Tuned*Note=Upper-bound2025.12 | 98.7 | — | — | — | — | |
| Fine-TunedBackbone=CLIP ViT-B/32, Classes=432025.12 | 98.7 | — | — | — | — | |
| theta_B fine-tuneBackbone=ViT-B/162025.10 | 98.65 | — | — | — | — | |
| theta_A fine-tuneModel=laion2b, Protocol=fine-tune2026.02 | 98.64 | — | — | — | — | |
| theta_B fine-tuneModel=laion400m, Protocol=fine-tune2026.02 | 98.63 | — | — | — | — | |
| Multiple modelsBackbone=ViT-B/322022.08 | 98.5 | — | — | — | — | |
| theta_B fine-tuneModel=ViT-L/14, Mode=fine-tune2026.02 | 98.42 | — | — | — | — | |
| MergeGuardModel Status=Protected Standalone (θ̂_def)2025.11 | 98.25 | — | — | — | — | |
| MMFTModel=RN50, Method=Ours2026.01 | 97.92 | — | — | — | — | |
| ISO-C/LARVBackbone=ViT-L/142026.02 | 97.8 | 0 | — | — | — | |
| FLYPModel=RN50, Method=FLYP2026.01 | 97.77 | — | — | — | — | |
| Joint patchingBackbone=ViT-B/162022.08 | 97.7 | — | — | — | — | |
| ADAProtection Status=None (theta_merge)2025.11 | 97.5 | — | — | — | — | |
| ISO-CTS/LARVBackbone=ViT-L/142026.02 | 97.3 | -0.2 | — | — | — | |
| MDA AMBackbone=ViT-B/32, Rotation Alignment=true2025.11 | 97.3 | — | — | — | — | |
| FTModel=RN50, Method=FT2026.01 | 97.2 | — | — | — | — | |
| MDA TABackbone=ViT-B/322025.11 | 97 | — | — | — | — | |
| Joint patchingBackbone=ViT-B/322022.08 | 96.9 | — | — | — | — | |
| TSV-M/LARVBackbone=ViT-L/142026.02 | 96.9 | 0.4 | — | — | — | |
| MDA AMBackbone=ViT-B/32, Rotation Alignment=false2025.11 | 96.9 | — | — | — | — | |
| TSV TABackbone=ViT-B/322025.11 | 96.7 | — | — | — | — | |
| ISO-CTS/LARVBackbone=ViT-B/162026.02 | 96.6 | 1.9 | — | — | — | |
| Seq. patchingBackbone=ViT-L/14, Seed=22022.08 | 96.2 | — | — | — | — | |
| EMR-MERGINGBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Number of Merged Models=302024.05 | 95.98 | — | — | — | — | |
| AdaMerging (Layer) + DeVecApproach=Our Method, Granularity=Layer-based, Optimization=Unsupervised2025.12 | 95.9 | — | — | — | — | |
| IndividualBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Number of Merged Models=12024.05 | 95.74 | — | — | — | — | |
| TSV-M/LARVBackbone=ViT-B/322026.02 | 95.1 | — | — | — | — | |
| LoRAParameter Count=0.67M2025.12 | 95.04 | — | — | — | — | |
| ISO-C/LARVBackbone=ViT-B/162026.02 | 94.9 | -0.1 | — | — | — | |
| Partial-LoRAParameter Count=0.26M2025.12 | 94.72 | — | — | — | — | |
| DoRAParameter Count=0.77M2025.12 | 94.62 | — | — | — | — | |
| ISO-C/LARVBackbone=ViT-B/322026.02 | 94.6 | 2.2 | — | — | — | |
| AdaLoRAParameter Count=0.67M2025.12 | 94.13 | — | — | — | — | |
| Layer-wise AdaMergingBackbone=ViT-B/322026.02 | 93.8 | — | — | — | — | |
| AdaMergingBackbone=ViT-B/322025.11 | 93.8 | — | — | — | — | |
| Partial-AdaLoRAParameter Count=0.25M2025.12 | 93.79 | — | — | — | — | |
| TSV-M/LARVBackbone=ViT-B/162026.02 | 93.7 | -0.4 | — | — | — | |
| Seq. patchingBackbone=ViT-L/14, Seed=12022.08 | 93.5 | — | — | — | — | |
| ISO-CTS/LARVBackbone=ViT-B/322026.02 | 93.5 | — | — | — | — | |
| AdaMergingBackbone=ViT-L/142026.02 | 93.2 | — | — | — | — | |
| RegMean++Backbone=ViT-B/322026.02 | 93.2 | — | — | — | — | |
| ISO-C/LARVBackbone=ViT-B/322026.02 | 93.2 | — | — | — | — | |
| LoRA+Parameter Count=0.67M2025.12 | 93.03 | — | — | — | — | |
| HAWXBackbone=EfficientNet, Configuration=1-47:KV9, Power=0.965, Area=0.9392026.02 | 92.8 | — | — | — | — | |
| HAWXBackbone=EfficientNet, Configuration=1-45,47:KV8 | 46:(KV9,L2J), Power=0.963, Area=0.9352026.02 | 92.8 | — | — | — | — | |
| TSV-M/LARVBackbone=ViT-B/322026.02 | 92.7 | 0.2 | — | — | — | |
| AdaMerging (Layer)Approach=Task Vectors, Granularity=Layer-based2025.12 | 91.9 | — | — | — | — | |
| Seq. patchingBackbone=ViT-L/14, Seed=02022.08 | 91.8 | — | — | — | — | |
| AdaMergingBackbone=ViT-B/162026.02 | 91.6 | — | — | — | — | |
| DOGE TABackbone=ViT-B/322025.11 | 91.6 | — | — | — | — | |
| ISO-CTS/LARVBackbone=ViT-B/322026.02 | 91.4 | 1 | — | — | — | |
| Seq. patchingBackbone=ViT-B/16, Seed=22022.08 | 91.3 | — | — | — | — | |
| VeRAParameter Count=0.10M2025.12 | 91.23 | — | — | — | — | |
| X-FM_baseLinear evaluation=true, Model size=Base, Patch size=16*16, Resolution=224*2242023.01 | 90.9 | — | — | — | — | |
| ISO TABackbone=ViT-B/322025.11 | 90.8 | — | — | — | — | |
| DeepBackbone=ResNet-18, Visual Reprogramming Method=Watermarking-based, Learning Method Category=Deep learning-based2024.10 | 89.5 | — | — | — | — | |
| RegMeanBackbone=ViT-B/322026.02 | 89.4 | — | — | — | — | |
| AdaMerging++ (Layer)Approach=Task Vectors, Granularity=Layer-based2025.12 | 89 | — | — | — | — | |
| TA/LARVBackbone=ViT-L/142026.02 | 88.8 | 5.3 | — | — | — | |
| TIES/LARVBackbone=ViT-L/142026.02 | 88.8 | 3.6 | — | — | — | |
| theta_B + delta_starBackbone=ViT-L/142025.10 | 88.65 | — | — | — | — | |
| ThermoLion2025.12 | 87.98 | — | — | — | — | |
| EnsemblingBackbone=ViT-L/142026.02 | 87.92 | — | — | — | — | |
| Seq. patchingBackbone=ViT-B/16, Seed=12022.08 | 87.6 | — | — | — | — | |
| TIES/LARVBackbone=ViT-B/162026.02 | 87.6 | 11.3 | — | — | — | |
| ISO-CLS TABackbone=ViT-B/322025.11 | 87.5 | — | — | — | — | |
| HAWXBackbone=EfficientNet, Configuration=1-29,34,42,44-47:KV9 | 30-33,35-41:KVP | 43:L2J, Power=0.928, Area=0.9162026.02 | 87.4 | — | — | — | — | |
| Task MatrixBackbone=CLIP ViT-B/32, Best Layer=11, Classes=432025.12 | 87.2 | — | — | — | — | |
| Seq. patchingBackbone=ViT-B/32, Seed=22022.08 | 87.1 | — | — | — | — | |
| Linear ProbeBackbone=CLIP ViT-B/32, Classes=432025.12 | 86.8 | — | — | — | — | |
| TAProtection Status=None (theta_merge)2025.11 | 86.7 | — | — | — | — | |
| CLIPLinear evaluation=true, Model size=Base, Patch size=16*16, Resolution=224*2242023.01 | 86.6 | — | — | — | — | |
| Seq. patchingBackbone=ViT-B/16, Seed=02022.08 | 86.4 | — | — | — | — | |
| Task MatrixBackbone=DINOv3 ViT-B/16, best layer=112025.12 | 86.3 | — | — | — | — | |
| DaVinciLinear evaluation=true, Model size=Base, Patch size=16*16, Resolution=224*2242023.01 | 85.8 | — | — | — | — | |
| Linear ProbeBackbone=DINOv3 ViT-B/162025.12 | 85.7 | — | — | — | — | |
| AdaMerging (Task) + DeVecApproach=Our Method, Granularity=Task-based, Optimization=Unsupervised2025.12 | 84.9 | — | — | — | — | |
| AdaMerging (Task)Approach=Task Vectors, Granularity=Task-based2025.12 | 84.6 | — | — | — | — | |
| EsViTBackbone=Swin-T, Evaluation Protocol=Linear probe2021.06 | 84.3 | — | — | — | — | |
| TA/LARVBackbone=ViT-B/162026.02 | 84.3 | 2.3 | — | — | — |