Research map / Label-efficient and robust learning / Recognition and long-tail
AdaMerging: Adaptive Model Merging for Multi-Task Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AdaMerging, which learns task-wise or layer-wise coefficients for merging task vectors by minimizing prediction entropy on unlabeled test data, improving multi-task model merging over Task Arithmetic and Ties-Merging.
Problem
Task-vector-based model merging for multi-task learning is highly sensitive to the merging coefficient, and grid-searching coefficients is impractical with many tasks and layers, all without access to original training data.
Approach
Treat merging coefficients as learnable parameters and optimize them without labels by minimizing Shannon entropy of predictions on unlabeled multi-task test samples, justified by a strong positive correlation between entropy and prediction loss; applied per task (Task-wise) or per layer per task (Layer-wise), and also on Ties-merged vectors (AdaMerging++).
Key results
- Layer-wise AdaMerging on ViT-B/32 reaches 80.1% average accuracy vs. 69.1% for Task Arithmetic (about 11% gain); Layer-wise AdaMerging++ reaches 81.1% vs. 72.4% for Ties-Merging.
- On unseen tasks (ViT-B/32), AdaMerging improves average accuracy over Task Arithmetic by roughly 8.3% on MNIST/EuroSAT and 4.4% on RESISC45/SVHN.
- On ViT-L/14, AdaMerging reaches 90.8% and AdaMerging++ 91.0% average accuracy, vs. 84.5% for Task Arithmetic and 86.0% for Ties-Merging.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace fixed shared merging coefficients with adaptive layer-wise coefficients | Task Arithmetic merging on ViT-B/32 · Eight image tasks | Avg Acc: 69.1 -> 80.1 | Helped |
| Replace fixed shared merging coefficients with adaptive layer-wise coefficients | Ties-Merging on ViT-B/32 · Eight image tasks | Avg Acc: 72.4 -> 81.1 | Helped |
| Replace fixed shared merging coefficients with adaptive layer-wise coefficients | Task Arithmetic merging on ViT-L/14 · Eight image tasks | Avg Acc: 84.5 -> 90.8 | Helped |
| Replace fixed shared merging coefficients with adaptive layer-wise coefficients | Ties-Merging on ViT-L/14 · Eight image tasks | Avg Acc: 86.0 -> 91.0 | Helped |
| Use layer-wise rather than task-wise adaptive coefficients (ablation) | AdaMerging on ViT-B/32 · Eight image tasks | Avg Acc: 71.1 -> 80.1 | Helped |
| Use Ties-Merging task vectors rather than original task vectors (ablation) | Task-wise AdaMerging on ViT-B/32 · Eight image tasks | Avg Acc: 71.1 -> 73.7 | Helped |
| Use Ties-Merging task vectors rather than original task vectors (ablation) | Layer-wise AdaMerging on ViT-B/32 · Eight image tasks | Avg Acc: 80.1 -> 81.1 | Helped |
| Use adaptive merging for unseen tasks instead of fixed Task Arithmetic coefficients | AdaMerging · Unseen MNIST and EuroSAT, ViT-B/32 | Avg Acc: 61.7 -> 70.0 | Helped |
| Use adaptive merging for unseen tasks instead of fixed Task Arithmetic coefficients | AdaMerging · Unseen RESISC45 and SVHN, ViT-B/32 | Avg Acc: 51.1 -> 55.5 | Helped |
| Use adaptive merging instead of fixed Task Arithmetic coefficients | AdaMerging on ViT-B/32 · Clean test, four tasks | Avg Acc: 79.8 -> 88.0 | Helped |
| Use only 0.10% rather than 100.0% of the unlabeled test set for coefficient learning (ablation) | Layer-wise AdaMerging on ViT-B/32 · Eight image tasks | Avg Acc: 80.1 -> 74.0 | Hurt |
| Use only 5.00% rather than 100.0% of the unlabeled test set for coefficient learning (ablation) | Layer-wise AdaMerging on ViT-B/32 · Eight image tasks | Avg Acc: 80.1 -> 80.1 | No clear effect |
Surprising findings
- Unsupervised AdaMerging comes close to the supervised-coefficient version (e.g., 71.1% vs 71.3% for Task-wise), supporting entropy as a proxy for cross-entropy.
- Only 0.1% of unlabeled test data already yields about 4.9% gain over Task Arithmetic on ViT-B/32.
Limitations
- Requires access to unlabeled test data from all merged tasks at merging time.
- Gains are reported mainly for image classification with CLIP ViT models; other architectures are left to future work.
- Task-wise AdaMerging underperforms on some tasks (e.g., Cars, DTD) relative to the layer-wise variant.
Experimental setup
Datasets: SUN397 Cars RESISC45 EuroSAT SVHN GTSRB MNIST DTD
Benchmarks: SUN397 Cars RESISC45 EuroSAT SVHN GTSRB MNIST DTD
Metrics: average accuracy
Baselines: Weight Averaging Fisher Merging RegMean Task Arithmetic Ties-Merging Traditional MTL Individual
Builds on: Task Arithmetic Ties-Merging test-time adaptation via entropy minimization Fisher Merging RegMean
Closest papers by meaning
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.91)
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.90)
- CALM: Consensus-Aware Localized Merging for Multi-Task Learning (ICML 2025 · similarity 0.89)
- Representation Surgery for Multi-Task Model Merging (ICML 2024 · similarity 0.89)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.89)
- CAT Merging: A Training-Free Approach for Resolving Conflicts in Model Merging (ICML 2025 · 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.