Research map / Architectures and efficiency / Pruning and efficient tuning
MODEL SOUPS NEED ONLY ONE INGREDIENT
Alireza Abdollahpoorrostam, Nikolaos Dimitriadis, Adam Hazimeh, Pascal Frossard
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 |
|---|---|---|---|
| apply MonoSoup (R=0.8) to the best-OOD checkpoint | FT CLIP ViT-B/32 · ImageNet avg OOD | accuracy: 50.67% -> 51.60% | Improved |
| apply MonoSoup to the worst-OOD checkpoint | FT CLIP ViT-B/32 · ImageNet avg OOD | accuracy: 36.71% -> 44.21% | Improved |
| apply MonoSoup to the worst-ID checkpoint | FT CLIP ViT-B/32 · ImageNet avg OOD | accuracy: 38.64% -> 46.54% | Improved |
| use MonoSoup on a model pair instead of ModelStock | CLIP ViT-B/32 pair (ID+, OOD+) · ImageNet avg OOD | accuracy: 50.53% -> 51.37% | Improved |
| apply MonoSoup instead of LiNeS | Qwen3-0.6B M-1 · GSM_Plus | accuracy: 30.1 -> 30.2 | Improved |
| apply MonoSoup to the fine-tuned Qwen3-0.6B | Qwen3-0.6B M-1 · GSM_Plus | accuracy: 29.5 -> 30.2 | Improved |
| apply MonoSoup post-hoc to English-math fine-tuned model | Qwen3-0.6B · MGSM average | accuracy: 30.6 -> 33.8 | Improved |
| apply MonoSoup to fine-tuned ConvNeXt-Small with fixed R=0.8 | ConvNeXt-Small · ImageNet avg OOD | accuracy: 55.85% -> 56.70% | Improved |
| apply MonoSoup to ZH language | Qwen3-0.6B · MGSM ZH | accuracy: 39.2 -> 38.8 | Hurt |
| use both spectral decay and alignment in the mixing rule instead of one signal (ablation) | MonoSoup · CLIP ViT-L/14 | ID/OOD accuracy | Improved |
| set variance threshold R too small (ablation) | MonoSoup (fixed R) · CLIP ViT-L/14 | ID/OOD accuracy | Hurt |
| use MonoSoup as start point for Wise-FT interpolation | Wise-FT · CLIP ViT-B/32, 70 checkpoints | ID-OOD Pareto front | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- + LiNeS — LiNeS: Post-training Layer Scaling Prevents Forgetting and Enhances Model Merging (ICLR 2025) · better on 4 of 4 shared results
- ModelStock — Model Stock: All we need is just a few fine-tuned models (ECCV 2024) · 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 0.87)
- LT-Soups: Bridging Head and Tail Classes via Subsampled Model Soups (NeurIPS 2025 · similarity 0.87)
- Connect Later: Improving Fine-tuning for Robustness with Targeted Augmentations (ICML 2024 · similarity 0.85)
- DaWin: Training-free Dynamic Weight Interpolation for Robust Adaptation (ICLR 2025 · similarity 0.85)
- Sparse Model Soups: A Recipe for Improved Pruning via Model Averaging (ICLR 2024 · similarity 0.85)
- Detecting Out-of-Distribution Through the Lens of Neural Collapse (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.
