Research map / Label-efficient and robust learning / Domain adaptation and generalization
Bridging Domains through Subspace-Aware Model Merging
Levy Chaves, Chao Zhou, Rebekka Burkholz, Eduardo Valle, Sandra Avila
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 |
|---|---|---|---|
| trim off-diagonal outliers while keeping diagonal, versus diagonal only (ablation) | SCORE · DG 8 datasets avg, ViT-B-32 | accuracy (%): 63.62 -> 65.69 (+2.07) | Improved |
| trim off-diagonal outliers while keeping diagonal, versus diagonal only (ablation) | SCORE · DG 8 datasets avg, ViT-B-16 | accuracy (%): 67.38 -> 69.97 (+2.59) | Improved |
| trim off-diagonal outliers while keeping diagonal, versus diagonal only (ablation) | SCORE · DG 8 datasets avg, ViT-L-14 | accuracy (%): 71.50 -> 73.04 (+1.53) | Improved |
| add trimming to the full matrix (instead of keeping all off-diagonal entries) (ablation) | SCORE · DG 8 datasets avg, ViT-B-32 | accuracy (%): 7.59 -> 65.69 | Improved |
| add off-diagonal components to the diagonal (trimmed) rather than off-diagonal only (ablation) | SCORE · DG 8 datasets avg, ViT-B-32 | accuracy (%): 58.41 -> 65.69 | Improved |
| keep only off-diagonal instead of only diagonal (ablation) | SCORE · DG 8 datasets avg, ViT-B-32 | accuracy (%): 63.62 -> 58.41 (-5.21) | Hurt |
| replace logit ensemble with SCORE merging | fine-tuned CLIP models · DG 8 datasets avg, ViT-B-32 | accuracy (%): 64.57 -> 65.69 (+1.12) | Improved |
| replace logit ensemble with SCORE merging | fine-tuned CLIP models · DG 8 datasets avg, ViT-B-16 | accuracy (%): 68.07 -> 69.97 (+1.90) | Improved |
| replace logit ensemble with SCORE merging | fine-tuned CLIP models · DG 8 datasets avg, ViT-L-14 | accuracy (%): 71.81 -> 73.04 (+1.24) | Improved |
| merge with Task Arithmetic instead of using zero-shot model | CLIP ViT-B-32 · DG 8 datasets avg | accuracy (%): 55.63 -> 42.47 | Hurt |
| merge with SCORE instead of using zero-shot model | CLIP ViT-B-32 · DG 8 datasets avg | accuracy (%): 55.63 -> 65.69 | Improved |
| use SCORE instead of TSV in multi-task merging | CLIP ViT-L-14 · Multi-task learning, 8 tasks avg | accuracy (%): 90.53 -> 91.17 | Improved |
Closest papers by meaning
- No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces (ICML 2025 · similarity 0.91)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.91)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.90)
- 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)
- Training-Free Pretrained Model Merging (CVPR 2024 · 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.
