Research map / Architectures and efficiency / Pruning and efficient tuning
Revisiting the Role of Pretrained Weights in Model Merging: On Near-Optimality within the Core Subspace
Wenju Sun, Qingyong Li, Tiancheng Li, Yangliao Geng, Boyang Li
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 |
|---|---|---|---|
| add CSP projection | Task Arithmetic · 8 vision tasks, ViT-B/32 | average accuracy: 69.1 -> 72.7 | Improved |
| add CSP projection | TSV · 8 vision tasks, ViT-B/32 | average accuracy: 84.1 -> 85.1 | Improved |
| add CSP projection | ISO-CTS · 8 vision tasks, ViT-B/32 | average accuracy: 84.3 -> 84.7 | Improved |
| add CSP projection | Task Arithmetic · 8 vision tasks, ViT-L/14 | average accuracy: 84.5 -> 86.2 | Improved |
| add CSP projection | TSV · 20 vision tasks, ViT-B/32 | average accuracy: 73.20 -> 75.06 | Improved |
| add CSP projection | Task Arithmetic · OKVQA (merged BLIP) | accuracy: 17.71 -> 24.67 | Improved |
| add CSP projection | Task Arithmetic · TextVQA (merged BLIP) | accuracy: 0.49 -> 10.28 | Improved |
| add CSP projection | Task Arithmetic · COCO Caption (merged BLIP) | CIDEr: 0.86 -> 0.83 | Hurt |
| add CSP projection | TSV · COCO Caption (merged BLIP) | CIDEr: 0.84 -> 0.81 | Hurt |
| add CSP projection | Task Arithmetic · Banking77 (Qwen3-0.6B) | accuracy: 74.67% -> 74.16% | Hurt |
| add CSP projection | Task Arithmetic and TSV · HotpotQA (Qwen3-8B) | NDCG@10: 0.64 -> 0.65 (TA); 0.75 -> 0.76 (TSV) | Improved |
| add CSP projection | KnOTS · 7 vision tasks (LoRA merging) | average accuracy: 54.51 -> 56.38 | 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.
- RobustMerge — RobustMerge: Parameter-Efficient Model Merging for MLLMs with Direction Robustness (NeurIPS 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
- No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces (ICML 2025 · similarity 0.93)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.90)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.89)
- Training-Free Pretrained Model Merging (CVPR 2024 · similarity 0.89)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.89)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.89)
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.
