Research map / Learning theory and algorithms / Combinatorial and search optimization
From Parameter to Representation: A Closed-Form Approach for Controllable Model Merging
Jialin Wu, Jian Yang, Handing Wang, Jiajun Wen, Zhiyong Yu
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 ReACT correction on AMPP instead of plain AMPP | AdaMerging++ · 8 tasks ViT-B/32 | average accuracy: 85.4 vs. 81.1 | Improved |
| use ReACT instead of Pareto Merging under one-hot preference | AMPP · 8 tasks ViT-B/32 one-hot | average accuracy: 88.9% vs. 83.6% | Improved |
| use ReACT instead of Pareto Merging under equal preference | AM · 8 tasks ViT-B/32 equal pref | average accuracy: 84.9 vs 84.9 | No clear effect |
| use ReACT instead of Pareto Merging on controllability | AM · Cars/RESISC45/DTD priority | HV@8: 60.90 -> 68.49 | Improved |
| use ReACT instead of Pareto Merging on controllability | AMPP · Cars/RESISC45/DTD priority | Uniformity U@8: 26.59 -> 41.58 | Improved |
| restrict calibration data to 10% of unlabeled test data | ReACT (AMPP) · 8 tasks one-hot | average accuracy: 88.9 -> 86.7 | Hurt |
| add orthogonal regularization (beta>0 vs beta=0) (ablation) | ReACT · 8 tasks ViT-B/32 | accuracy | Improved |
| use regularized solution instead of pure orthogonal transformation (beta to infinity) (ablation) | ReACT · 8 tasks ViT-B/32 | accuracy | Improved |
| vary beta across a wide range (ablation) | ReACT · 8 tasks ViT-B/32 | accuracy | No clear effect |
| use data-aware Pareto-optimal aggregation instead of naive averaging (ablation) | ReACT · equal preference | NAcc: 92.2 -> 93.5 | Improved |
| use data-aware Pareto-optimal aggregation instead of polar decomposition (ablation) | ReACT · equal preference | NAcc: 91.5 -> 93.5 | Improved |
| use data-aware aggregation instead of naive averaging (ablation) | ReACT · priority preference | HV: 67.6 -> 70.2 | Improved |
Closest papers by meaning
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.90)
- Pareto Merging: Multi-Objective Optimization for Preference-Aware Model Merging (ICML 2025 · similarity 0.90)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.88)
- Parameter Competition Balancing for Model Merging (NeurIPS 2024 · similarity 0.88)
- Less is More: Efficient Model Merging with Binary Task Switch (CVPR 2025 · similarity 0.88)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (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.
