Image Classification on SVHN (test) (Top-1 and Top-5 Accuracy)
97.3Top-1 AccuracyFine-tuned (STL)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Fine-tuned (STL)Backbone=CLIP-ViT-B/32, Evaluation Protocol=Fine-tuned (STL)2026.05 | 97.3 | — | |
| WUDI-Merging w/ FEATCALBackbone=CLIP-ViT-B/32, Upstream Merger=WUDI-Merging, Refinement=FEATCAL2026.05 | 96.7 | — | |
| WUDI-Merging w/ SurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=WUDI-Merging, Refinement=Surgery2026.05 | 95.6 | — | |
| WUDI-Merging w/ ProbSurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=WUDI-Merging, Refinement=ProbSurgery2026.05 | 95.6 | — | |
| Traditional MTLBackbone=CLIP-ViT-B/32, Evaluation Protocol=Traditional MTL2026.05 | 95.5 | — | |
| Task Arithmetic w/ FEATCALBackbone=CLIP-ViT-B/32, Upstream Merger=Task Arithmetic, Refinement=FEATCAL2026.05 | 95 | — | |
| Ours (DS-7B) SVHNSearch Cost=22 cycles, Generation=Delta, Training Protocol=LLM-Based NAS on LEMUR (fine-tuned, 1-epoch proxy)2026.05 | 94.9 | — | |
| Ours (Qwen-7B) SVHNSearch Cost=22 cycles, Generation=Delta, Training Protocol=LLM-Based NAS on LEMUR (fine-tuned, 1-epoch proxy)2026.05 | 94.9 | — | |
| Ours (Mistral-7B) SVHNSearch Cost=22 cycles, Generation=Delta, Training Protocol=LLM-Based NAS on LEMUR (fine-tuned, 1-epoch proxy)2026.05 | 94.9 | — | |
| WUDI-MergingBackbone=CLIP-ViT-B/32, Upstream Merger=WUDI-Merging2026.05 | 94.8 | — | |
| AdaMerging w/ FEATCALBackbone=CLIP-ViT-B/32, Upstream Merger=AdaMerging, Refinement=FEATCAL2026.05 | 94.2 | — | |
| Simple Averaging w/ FEATCALBackbone=CLIP-ViT-B/32, Upstream Merger=Simple Averaging, Refinement=FEATCAL2026.05 | 92.6 | — | |
| AdaMerging w/ SurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=AdaMerging, Refinement=Surgery2026.05 | 90.5 | — | |
| AdaMerging w/ ProbSurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=AdaMerging, Refinement=ProbSurgery2026.05 | 90.2 | — | |
| AdaMergingBackbone=CLIP-ViT-B/32, Upstream Merger=AdaMerging2026.05 | 87.6 | — | |
| Task Arithmetic w/ ProbSurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=Task Arithmetic, Refinement=ProbSurgery2026.05 | 85.3 | — | |
| Khalid v7 SVHNSearch Cost=22 cycles, Generation=Full code, Training Protocol=LLM-Based NAS on LEMUR (fine-tuned, 1-epoch proxy)2026.05 | 84.5 | — | |
| Task Arithmetic w/ SurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=Task Arithmetic, Refinement=Surgery2026.05 | 83.6 | — | |
| Task ArithmeticBackbone=CLIP-ViT-B/32, Upstream Merger=Task Arithmetic2026.05 | 77.9 | — | |
| Simple Averaging w/ SurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=Simple Averaging, Refinement=Surgery2026.05 | 70.8 | — | |
| Simple Averaging w/ ProbSurgeryBackbone=CLIP-ViT-B/32, Upstream Merger=Simple Averaging, Refinement=ProbSurgery2026.05 | 70.4 | — | |
| Simple AveragingBackbone=CLIP-ViT-B/32, Upstream Merger=Simple Averaging2026.05 | 64.5 | — | |
| Pre-trainedBackbone=CLIP-ViT-B/32, Evaluation Protocol=Zero-shot2026.05 | 31.6 | — | |
| MetaErrLabeled Training Data=10%2026.04 | 22.83 | 70.13 | |
| VATLabeled Training Data=10%2026.04 | 21.13 | 54.04 | |
| ICTLabeled Training Data=10%2026.04 | 20.41 | 67.21 | |
| LPLabeled Training Data=10%2026.04 | 19.71 | 64.52 | |
| CLLabeled Training Data=10%2026.04 | 19.47 | 64.24 | |
| PL-CBLabeled Training Data=10%2026.04 | 19.04 | 62.88 |