Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Learn to Merge: Meta-Learning for Adaptive Multi-Task Model Merging
Jun Chen, Qin Zhang, Weizhi Zhang, Xiao Luo, Philip S. Yu, Ziyue Qiao
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 |
|---|---|---|---|
| meta-learn merging coefficients with MetaMerging versus AdaMerging | ViT-B/32 merging · 8 vision tasks | avg accuracy: 80.1 -> 87.2 | Improved |
| MetaMerging versus Surgery on ViT-L/14 | ViT-L/14 merging · 8 vision tasks | avg accuracy: 92.3 -> 93.4 | Improved |
| MetaMerging versus Weight Averaging on GPT-2 | GPT-2 merging · GLUE 7 tasks | avg accuracy: 56.1 -> 76.9 | Improved |
| add task-specific adapters to the meta-learned unified model (ablation) | MetaMerging · ViT-B/32 8 tasks | avg accuracy: 67.8 -> 87.2 | Improved |
| add the same adapters to Weight Averaging merged model (ablation) | Weight Averaging · ViT-B/32 8 tasks | avg accuracy: 65.8 -> 80.0 | Improved |
| meta-learn coefficients instead of AdaMerging coefficients, both with adapters (ablation) | unified model + adapters · ViT-B/32 8 tasks | avg accuracy: 86.1 -> 87.2 | Improved |
| meta-learned unified model before adapters is weaker than AdaMerging before adapters (ablation) | MetaMerging · ViT-B/32 8 tasks, no adapters | avg accuracy: 80.1 -> 67.8 | Hurt |
| use one inner adapter step instead of zero in meta-learning (GPT-2) (ablation) | MetaMerging · GPT-2 GLUE | avg accuracy: 74.9 -> 76.4 | Improved |
| increase inner steps from 1 to 5 or 10 (ablation) | MetaMerging · GPT-2 GLUE | avg accuracy: 76.4 -> 75.7 (5 steps), 75.8 (10 steps) | Hurt |
| set inner step size to 1 instead of 0.1 (ablation) | MetaMerging · ViT-B/32 8 tasks | avg accuracy: 87.15 -> 86.74 | Hurt |
| set meta step size to 1 instead of 0.01 (ablation) | MetaMerging · ViT-B/32 8 tasks | avg accuracy: 87.15 -> 85.74 | Hurt |
| use MetaMerging instead of multi-task learning for training cost | ViT-B/32 merging · 8 vision tasks | running time: 15h 53m -> 1h 23m | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- AdaMerging++ — AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024) · better on 4 of 4 shared results
- DOGE — DOGE: Domain Reweighting with Generalization Estimation (ICML 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024 · similarity 0.90)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.89)
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.88)
- Task Groupings Regularization: Data-Free Meta-Learning with Heterogeneous Pre-trained Models (ICML 2024 · similarity 0.87)
- Provable Meta-Learning with Low-Rank Adaptations (NeurIPS 2025 · similarity 0.87)
- Model Merging in Pre-training of Large Language Models (NeurIPS 2025 · similarity 0.87)
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.
