Research map / Architectures and efficiency / Pruning and efficient tuning
Model Stock: All we need is just a few fine-tuned models
Dong-Hwan Jang, Sangdoo Yun, Dongyoon Han
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| average more fine-tuned weights (approach the weight-space center) (ablation) | CLIP ViT-B/32 fine-tuned on ImageNet · ImageNet / avg. shifts | accuracy: ImageNet 79.72 -> 80.59; shifts 46.37 -> 48.85 | Improved |
| use two fine-tuned models plus the pre-trained anchor with geometric layer-wise interpolation instead of 48-model Model Soup | CLIP ViT-B/32 · ImageNet (zero-shot init) | top-1 accuracy: 79.76 (uniform soup) -> 79.89 at cost 48 -> 2 | No clear effect |
| Model Stock instead of Uniform Model Soup on distribution shifts | CLIP ViT-B/32 (zero-shot init) · avg. shifts | accuracy: 52.08 -> 50.99 | Hurt |
| Model Stock* with LP init and strong augmentation versus vanilla fine-tuning | CLIP ViT-B/32 · ImageNet (LP init) | top-1 accuracy: 79.72 -> 81.19 | Improved |
| Model Stock instead of vanilla fine-tuning | CLIP ViT-L/14 · avg. distribution shifts | accuracy: 68.0 -> 73.5 (Model Stock* vs Vanilla FT*) | Improved |
| increase number of merged models N from 2 to 4 at fixed training cost (ablation) | Model Stock · ImageNet / avg. shifts (ViT-B/32) | accuracy: IN 80.1 -> 80.4; shifts 48.8 -> 48.9 | Improved |
| vary merging period from 1000 or 5000 iterations to 1 epoch (ablation) | Model Stock · ImageNet / avg. shifts (ViT-B/32) | accuracy: avg shifts 48.7 / 48.5 / 48.8 | No clear effect |
| use geometric interpolation with pre-trained anchor (post-training) instead of uniform averaging (ablation) | Model Stock · avg. shifts (ViT-B/32, N=2) | accuracy: 47.8 -> 50.4 | Improved |
| use geometric interpolation instead of uniform averaging on ImageNet accuracy (ablation) | Model Stock (post-training) · ImageNet (ViT-B/32, N=3,4) | top-1 accuracy: +0.0 for N=3 and N=4 | No clear effect |
| random-perturbed weights nearer the weight center (ablation) | CLIP ViT-B/32 · ImageNet | accuracy vs distance | Improved |
Later papers that compared with this method
2 later accepted papers report ModelStock as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- MODEL SOUPS NEED ONLY ONE INGREDIENT (ICML 2026 · better on 1 of 1 shared result)
- DaWin: Training-free Dynamic Weight Interpolation for Robust Adaptation (ICLR 2025 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Robust Fine-tuning of Zero-shot Models via Variance Reduction (NeurIPS 2024 · similarity 0.86)
- LT-Soups: Bridging Head and Tail Classes via Subsampled Model Soups (NeurIPS 2025 · similarity 0.85)
- Upweighting Easy Samples in Fine-Tuning Mitigates Forgetting (ICML 2025 · similarity 0.85)
- Understanding Overadaptation in Supervised Fine-Tuning: The Role of Ensemble Methods (ICML 2025 · similarity 0.85)
- Parameter Efficient Fine-tuning via Explained Variance Adaptation (NeurIPS 2025 · similarity 0.85)
- WAVE: Weight Templates for Adaptive Initialization of Variable-sized Models (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
