Research map / Label-efficient and robust learning / Domain adaptation and generalization
Training-Free Model Merging for Multi-target Domain Adaptation
Wenyi Li, Huan-ang Gao, Mingju Gao, Beiwen Tian, Rong Zhi, Hao Zhao
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 |
|---|---|---|---|
| merge buffers (BN statistics) in addition to parameters (ablation) | Model merging (ResNet101) · GTA -> Cityscapes+IDD | harmonic mean mIoU: 54.8 -> 56.3 | Improved |
| merge buffers in addition to parameters (ablation) | Model merging (MiT-B5) · GTA -> Cityscapes+IDD | harmonic mean mIoU: 60.2 -> 60.8 | Improved |
| mid-point parameter merging of independently adapted models (ablation) | HRDA STDA models (ResNet101) · GTA -> Cityscapes+IDD | harmonic mean mIoU: 52.1 -> 54.8 | Improved |
| use Git Re-Basin permutation alignment instead of plain mid-point merging (ablation) | Model merging · GTA -> Cityscapes+IDD | harmonic mean mIoU | No clear effect |
| initialize the two models from distinct (not same) weights (ablation) | Linear interpolation between adapted models · GTA -> Cityscapes/IDD | interpolated harmonic mean mIoU | Hurt |
| same initial weights but without pretraining (ablation) | Linear interpolation between adapted models · GTA -> Cityscapes/IDD | interpolated harmonic mean mIoU | Hurt |
| disable EMA teacher (copy student weights) in self-training (ablation) | Linear mode connectivity · GTA -> Cityscapes/IDD | interpolation performance | No clear effect |
| train models on disjoint synthetic subsets (ablation) | Linear mode connectivity · synthetic subsets -> Cityscapes | interpolation performance | No clear effect |
| merge only backbones, keep separate decoder heads | Model merging (ResNet101) · GTA -> Cityscapes+IDD | harmonic mean mIoU: 56.3 -> 57.7 | Improved |
| merge models from more target domains (four vs two) | Model merging · GTA -> C, I, A, D | harmonic mean mIoU: 33.5 -> 36.4 (all four); best 39.3 | Improved |
| merge models rather than training on combined data (SYNTHIA, MiT-B5) | Model merging · SYNTHIA -> Cityscapes+IDD | harmonic mean mIoU*: 62.5 -> 64.8 | Improved |
Closest papers by meaning
- Merge-Friendly Post-Training Quantization for Multi-Target Domain Adaptation (ICML 2025 · similarity 0.88)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.86)
- Invertible Projection and Conditional Alignment for Multi-Source Blended-Target Domain Adaptation (AAAI 2025 · similarity 0.85)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.85)
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.85)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.85)
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.
