Research map / Architectures and efficiency / Pruning and efficient tuning
Saliency-Aware Model Merging
Jungin Park, Jiyoung Lee, Kwanghoon Sohn
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 |
|---|---|---|---|
| use SA-Merging instead of WUDI-Merging for data-free merging | CLIP ViT-B/32 8-task merging · 8-task vision suite | average accuracy: 85.2 -> 85.9 | Improved |
| use SA-Merging instead of WUDI-Merging | CLIP ViT-L/14 8-task merging · 8-task vision suite | average accuracy: 92.6 -> 93.4 | Improved |
| use SA-Merging instead of WUDI-Merging | RoBERTa-Base GLUE merging · GLUE 8 tasks | normalized score: 85.3 -> 87.1 | Improved |
| use SA-Merging instead of WUDI-Merging | RoBERTa-Large GLUE merging · GLUE 8 tasks | normalized score: 88.8 -> 90.2 | Improved |
| use rank-wise saliency for LoRA merging instead of WUDI | Flan-T5-base LoRA GLUE merging · GLUE LoRA 8 tasks | average score: 81.2 -> 82.5 | Improved |
| use rank-wise saliency for LoRA merging instead of WUDI | Qwen-14B LoRA merging · MMLU/TruthfulQA/BBQ/CNN | average score: 55.69 -> 56.33 | Improved |
| set pruning ratio p to 0.2 (vs other ratios) (ablation) | SA-Merging · vision merging | average accuracy | Improved |
| increase number of pruning iterations up to about 10 (ablation) | SA-Merging · vision merging | per-task accuracy | Improved |
| use SA-Merging pruning to reduce cross-model sign conflicts | SA-Merging · sign conflicts across models | parameter conflicts | Improved |
| SA-Merging vs test-time AdaMerging++ on ViT-B/32 | CLIP ViT-B/32 · 8-task vision suite | average accuracy: 81.1 -> 85.9 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TIES-Merging — TIES-Merging: Resolving Interference When Merging Models (NeurIPS 2023) · better on 4 of 4 shared results
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
- Contribution-Aware Structured Sparsity for Model Merging (NeurIPS 2026 · similarity 0.90)
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.89)
- No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces (ICML 2025 · similarity 0.89)
- Towards Minimizing Feature Drift in Model Merging: Layer-wise Task Vector Fusion for Adaptive Knowledge Integration (NeurIPS 2025 · similarity 0.89)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.88)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · 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.
