Research map / Architectures and efficiency / RNNs and normalization
How does the optimizer implicitly bias the model merging loss landscape?
Chenxiang Zhang, Alexander Theus, Damien Teney, Antonio Orvieto, Jun Pang, Sjouke Mauw
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 |
|---|---|---|---|
| increase learning rate (2e-1 vs 1e-2) at similar single-model accuracy (ablation) | ResNet18 SGD · CIFAR100 | median merge gain (%): +0.2% -> +1.2% | Improved |
| use too large a learning rate (ablation) | ResNet18 / DenseNet121 SGD · CIFAR10, CIFAR100, TinyImageNet | merging success | Hurt |
| increase weight decay (scale-invariant nets) (ablation) | DenseNet121 · TinyImageNet | median merge gain (%): +0.5% -> +1.2% | Improved |
| vary weight decay in non-scale-invariant MLP (ablation) | MLP · SVHN | merge gain | No clear effect |
| decrease batch size from 256 to 16 (ablation) | ResNet18 · CIFAR100 | median merge gain (%): ~0% -> +1% | Improved |
| increase momentum to 0.9 (ablation) | ResNet18 · CIFAR100 | median merge gain (%): +0.2% -> +1.0% | Improved |
| add data augmentation (ablation) | ResNet18 · CIFAR100 | merge gain | Improved |
| increase learning rate (ablation) | CLIP ViT-B/16 transfer · WILDS-FMoW | accuracy gain of merged model: 4x larger gain at 1e-4 than 1e-6 | Improved |
| use largest learning rate for finetuning (ablation) | CLIP ViT-B/16 transfer · WILDS-FMoW | merged model accuracy | Hurt |
| use larger learning rate with non-pretrained initialization (ablation) | ConvNext-T task arithmetic · CIFAR10 non-transfer | loss landscape flatness under alpha-interpolation | Hurt |
| use TIES merging instead of task arithmetic (ablation) | CLIP ViT-B/16 · FMoW+RESISC45 at lr 3e-5 | normalized accuracy (%): 85.9 -> 88.0 | Improved |
| merge models trained with different learning rates (ablation) | CLIP ViT-B/16 · FMoW+RESISC45 | merged accuracy | Hurt |
Closest papers by meaning
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.88)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.88)
- EMR-Merging: Tuning-Free High-Performance Model Merging (NeurIPS 2024 · similarity 0.88)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.88)
- Parameter Competition Balancing for Model Merging (NeurIPS 2024 · similarity 0.88)
- Rethinking Layer-wise Model Merging through Chain of Merges (NeurIPS 2026 · similarity 0.88)
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.
