Research map / Architectures and efficiency / Pruning and efficient tuning
ACE-Merging: Data-Free Model Merging with Adaptive Covariance Estimation
Bo Xu, Haotian Wu, Hehai Lin, Weiquan Huang, Beier Zhu, Yao Shu, Chengwei Qin
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 adaptive covariance normalization (ablation) | basic closed-form merge · GLUE (RoBERTa-Large) | avg score (%): 80.05 -> 88.04 | Improved |
| add adaptive covariance normalization (ablation) | basic closed-form merge · GLUE (GPT-2) | avg score (%): 68.72 -> 71.50 | Improved |
| add collective structural prior (ablation) | ACE-Merging E2 · GLUE (RoBERTa-Large) | avg score (%): 88.04 -> 86.79 | Hurt |
| add collective structural prior (ablation) | ACE-Merging E2 · GLUE (GPT-2) | avg score (%): 71.50 -> 71.51 | No clear effect |
| add collective structural prior (ablation) | ACE-Merging E2 · Vision (ViT-B/16, 8 tasks) | avg score (%): 89.91 -> 90.60 | Improved |
| add spectral refinement (ablation) | ACE-Merging E3 · GLUE (RoBERTa-Large) | avg score (%): 86.79 -> 91.68 | Improved |
| add spectral refinement (ablation) | ACE-Merging E3 · GLUE (GPT-2) | avg score (%): 71.51 -> 74.09 | Improved |
| raise heterogeneity threshold from 0.3 to 0.5 (ablation) | ACE-Merging · GLUE (GPT-2) | avg accuracy (%): 74.09 -> 69.35 | Hurt |
| raise heterogeneity threshold from 0.3 to 0.5 (ablation) | ACE-Merging · GLUE (RoBERTa-Large) | avg accuracy (%): 91.68 -> 88.09 | Hurt |
| vary spectral rank fraction 0.1 to 0.5 (ablation) | ACE-Merging · GLUE (GPT-2) | avg accuracy (%): 74.17 / 74.09 / 73.78 | No clear effect |
| use ACE-Merging instead of WUDI-Merging | data-free merging · GLUE normalized (RoBERTa-Base) | avg score: 85.3 -> 90.4 | Improved |
| use ACE-Merging instead of Iso-CTS | data-free merging · Vision (ViT-L/14, 20 tasks) | avg accuracy (%): 90.1 -> 89.5 | Hurt |
Closest papers by meaning
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.91)
- No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces (ICML 2025 · similarity 0.91)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.90)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.90)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.90)
- CAT Merging: A Training-Free Approach for Resolving Conflicts in Model Merging (ICML 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.
