Research map / Architectures and efficiency / Pruning and efficient tuning
M-Loss: Quantifying Model Merging Compatibility with Limited Unlabeled Data
Tiantong Wang, Yiyang Duan, Haoyu Chen, Tiantong Wu, Wei Yang Bryan Lim
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 |
|---|---|---|---|
| replace fixed TIES trimming with M-Loss-guided dynamic pruning (M-TIES) | TIES · 8-task merging, ViT-B/32 | average accuracy: 72.42 -> 73.23 | Improved |
| replace fixed TIES trimming with M-TIES | TIES · 8-task merging, ViT-L/14 | average accuracy: 84.94 -> 85.28 | Improved |
| M-TIES vs ensembling | M-TIES · 8-task merging, ViT-B/32 | average accuracy: 78.56 -> 73.23 | Hurt |
| M-TIES vs ensembling | M-TIES · 8-task merging, ViT-L/14 | average accuracy: 85.56 -> 85.28 | Hurt |
| replace TIES with M-TIES on SVHN | TIES · SVHN, ViT-B/32 | accuracy (%): 86.20 -> 83.06 | Hurt |
| M-TIES on out-of-domain datasets vs TIES | TIES · OOD average, ViT-B/32 | accuracy (%): 57.55 -> 63.46 | Improved |
| increase unlabeled samples for M-Loss from 128 to 512 (ablation) | M-TIES · 8-task average, ViT-B/32 | accuracy (%): 73.232 -> 73.227 | No clear effect |
| vary random seed for sampling M-Loss data (ablation) | M-TIES · 8-task average, ViT-B/32 | accuracy (%): 73.225 / 73.165 / 73.232 | No clear effect |
| apply M-Loss pruning only to first and last few layers (Few-Layer M-TIES) (ablation) | M-TIES · 8-task average, ViT-B/32 | accuracy (%): 73.232 -> 73.216 | No clear effect |
| apply M-Loss pruning only to a few layers (ablation) | M-TIES · 8-task average, ViT-L/14 | accuracy (%): 85.278 -> 85.259 | No clear effect |
Closest papers by meaning
- Parameter Competition Balancing for Model Merging (NeurIPS 2024 · similarity 0.88)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.87)
- EMR-Merging: Tuning-Free High-Performance Model Merging (NeurIPS 2024 · similarity 0.87)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.87)
- Scalable Model Merging with Progressive Layer-wise Distillation (ICML 2025 · similarity 0.86)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.86)
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.
