Research map / Architectures and efficiency / Mixture of experts
Learning to Recover Task Experts from a Multi-Task Merged Model
Jinwook Jung, Taegyu Kim, Kumju Jo, Sungyong Baik
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 ReTeX task-expert recovery on top of Weight Averaging | Weight Averaging · Unseen tasks (MNIST, EuroSAT) | accuracy: 57.8 -> 63.4 | Improved |
| add ReTeX to Task Arithmetic | Task Arithmetic · Unseen tasks | accuracy: 58.5 -> 66.0 | Improved |
| add ReTeX to TIES-Merging | TIES-Merging · Unseen tasks | accuracy: 57.9 -> 68.7 | Improved |
| add ReTeX to AdaMerging | AdaMerging · Unseen tasks | accuracy: 66.8 -> 69.1 | Improved |
| add ReTeX to Weight Averaging | Weight Averaging · Seen tasks | accuracy: 63.7 -> 90.4 | Improved |
| scale task suite from 8 to 30 tasks | ReTeX · CLIP ViT-B/32 | accuracy: 92.0 -> 91.2 | No clear effect |
| scale task suite from 8 to 30 tasks | DaWin · CLIP ViT-B/32 | accuracy: 89.0 -> 40.3 | Hurt |
| scale task suite from 8 to 30 tasks | WEMoE · CLIP ViT-B/32 | accuracy: 90.4 -> 67.1 | Hurt |
| use stronger base merge (Task Arithmetic, TIES) under ReTeX (ablation) | ReTeX · ViT-B/32 | normalized accuracy | Improved |
| increase task embedding dimension (ablation) | ReTeX · 8, 14, 20 tasks | normalized accuracy | Improved |
| factorize generator output (ReTeX-E) (ablation) | ReTeX · ViT-B/32 8 tasks | VRAM and accuracy: 92.2 -> 92.0 avg, 2.5 -> 1.8 GB | Mixed |
| use grouped inference B=64 instead of sample-wise (ablation) | ReTeX · ViT-B/32 8 tasks | latency per input: 0.05s -> 0.0015s | Improved |
Closest papers by meaning
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.88)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.88)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.88)
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.87)
- Towards Minimizing Feature Drift in Model Merging: Layer-wise Task Vector Fusion for Adaptive Knowledge Integration (NeurIPS 2025 · similarity 0.86)
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024 · similarity 0.86)
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.
