Research map / Architectures and efficiency / Mixture of experts
Training-Free Task Classification for Multi-Task Model Merging
Jungyong Son, Jinwook Jung, 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 |
|---|---|---|---|
| replace Euclidean distance with SVD projection residual for task classification (ablation) | SiM · 8 vision tasks (CLIP ViT-B/32) | task classification accuracy: 95.8 -> 99.0 | Improved |
| replace projection residual with Mahalanobis distance (ablation) | SiM · 8 vision tasks | task classification accuracy: 99.0 -> 96.5 | Hurt |
| replace projection residual with 1-NN (ablation) | SiM · 8 vision tasks | task classification accuracy: 99.0 -> 92.0 | Hurt |
| add SiM routing to compressed task vectors (TM-TA) | TALL-Mask + Task Arithmetic · 20 vision tasks, ViT-B/32 | avg accuracy: 60.5 (Task Arithmetic) -> 89.6 | Improved |
| use SiM with larger task sets (30 and 50 tasks) | TSV-C + SiM · 50 vision tasks, ViT-B/32 | avg accuracy: 80.2 vs best baseline 66.6 (TSV-M) | Improved |
| SiM with ViT-L/14 on 8 tasks, TM-TA variant | TM-TA + SiM · 8 vision tasks, ViT-L/14 | avg accuracy: 92.0 vs 94.9 (MoW-Merging) and 95.4 (TSV-C + SiM) | Mixed |
| use SiM on unseen tasks (manifolds from 8 seen tasks only) | SiM-merged models · unseen 6 vision tasks | avg accuracy: seen 92.7 -> unseen 63.3 (TM-TIES); EMR 90.7 -> 57.9 | Mixed |
| add SiM routing on NLP tasks | EMR/TM-TA/TM-TIES/TSV-C + SiM · 7 NLP tasks, T5-large | avg score: 88.8 fine-tuned vs 85.6 (TM-TIES + SiM); 79.6 TSV-M | Improved |
| use SiM instead of per-sample dynamic interpolation (inference speed) | SiM with group batch inference · 8 vision tasks, ViT-B/32 | latency per input: 0.63s (DaWin) -> 0.003s | Improved |
| use 32 support samples per task instead of full training set (ablation) | SiM · task classification, vision | task classification accuracy | No clear effect |
| replace SiM with oracle routing (ablation) | SiM · vision downstream | avg accuracy: 0.3-0.6% gap | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DaWin — DaWin: Training-free Dynamic Weight Interpolation for Robust Adaptation (ICLR 2025) · 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
- 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)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.90)
- CAT Merging: A Training-Free Approach for Resolving Conflicts in Model Merging (ICML 2025 · similarity 0.90)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.90)
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024 · similarity 0.89)
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.
