Research map / Architectures and efficiency / Pruning and efficient tuning
Stay Unique, Stay Efficient: Preserving Model Personality in Multi-Task Merging
Kuangpu Guo, Aijing Yu, Jian Liang, Yuhe Ding, Zilei Wang, Ran He, Tieniu Tan
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 |
|---|---|---|---|
| personalize merging with DTS-D instead of basic merging (Task-Arithmetic) | CLIP ViT-B/32 · 8 vision tasks | accuracy: 69.22 -> 90.40 | Improved |
| use DTS-D instead of T-Switch at lower memory | CLIP ViT-B/32 · 8 vision tasks | accuracy: 90.15 -> 90.40 | Improved |
| add SVD decomposition before thresholding (ablation) | DTS-T · 8 vision tasks | memory rate (accuracy nearly equal): AMR 9.37 -> 3.68; accuracy 90.41 -> 90.32 | Improved |
| add per-group scaling factors (ablation) | DTS-T · 8 vision tasks | accuracy: 4.98 -> 90.32 | Improved |
| use 4-group thresholding instead of binarization (ablation) | DTS · 8 vision tasks | accuracy: approximately +2% | Improved |
| increase thresholding groups to eight (ablation) | DTS · 8 vision tasks | accuracy: storage doubled, virtually no gain | No clear effect |
| replace median with K-Means dynamic thresholding (ablation) | DTS-T · ViT-B/32 seen and unseen tasks | accuracy: seen 87.94 -> 88.25; unseen 55.23 -> 55.76 (DTS-D unseen 55.41 -> 55.37) | Mixed |
| use semantic-similarity weighted fusion of task information for unseen tasks | DTS-D · RESISC45+SVHN unseen, ViT-B/32 | accuracy: T-Switch 42.05 -> 55.41 | Improved |
| use DTS reconstruction instead of rank-1 LoRA | CLIP ViT-B/32 · 8 vision tasks | accuracy: 68.61 -> 90.32 (switching 1.01 -> 1.06 ms) | Improved |
| cap DTS storage below 1% (DTS-D*) instead of r=0.3 (ablation) | DTS-D · 8 vision tasks, ViT-B/32 | accuracy: 90.40 -> 90.17 | No clear effect |
| apply personalized merging on unseen tasks without adaptive weighting (baselines) | T-Switch · unseen tasks, ViT-B/32 | accuracy: 42.05 vs Ties-Merging 53.85 | Hurt |
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 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
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.89)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.89)
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024 · similarity 0.89)
- Representation Surgery for Multi-Task Model Merging (ICML 2024 · similarity 0.89)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.89)
- Localizing Task Information for Improved Model Merging and Compression (ICML 2024 · 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.
