Research map / Vision-language and multimodal / CLIP and zero-shot
Task Alignment: A simple and effective proxy for model merging in computer vision
Pau de Jorge Aranda, César DE SOUZA, Björn Michele, Mert Bulent SARIYILDIZ, Philippe Weinzaepfel, Florent Perronnin, Diane Larlus, Yannis Kalantidis
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 |
|---|---|---|---|
| select hyperparameters with TAP instead of labeled evaluation | Task Arithmetic · CLIP classification (avg over ViT-B-32/ViT-L-14, T=8/14/20) | average accuracy: 72.7 -> 72.5 | No clear effect |
| select hyperparameters with TAP instead of labeled evaluation | Iso-CTS · CLIP classification (avg over ViT-B-32/ViT-L-14, T=8/14/20) | average accuracy: 86.9 -> 86.8 | No clear effect |
| select hyperparameters with TAP instead of labeled evaluation | TSV · CLIP classification (avg over ViT-B-32/ViT-L-14, T=8/14/20) | average accuracy: 85.2 -> 85.1 | No clear effect |
| replace entropy loss with TAP in AdaMerging | AdaMerging · CLIP classification (avg over ViT-B-32/ViT-L-14, T=8/14/20) | average accuracy: 81.5 -> 81.1 | Mixed |
| merge fine-tuned LiDAR encoders with TAP-selected hyperparameters | AdaMerging · LiDAR segmentation (4 datasets) | normalized performance: 0.841 -> 0.985 | Improved |
| merge task-specific encoders on top of DUNE with TAP | AdaMerging · DUNE heterogeneous tasks (base DUNE) | normalized performance: 0.929 -> 0.971 | Improved |
| merge with TIES instead of Task Arithmetic on DUNE tasks | TIES · DUNE heterogeneous tasks (base DUNE) | normalized performance: 0.956 -> 0.934 | Hurt |
| merge DINOv2 fine-tuned models with TAP instead of distilled DUNE | Iso-CTS · DUNE heterogeneous tasks (base DINOv2) | normalized performance: 0.847 -> 0.915 | Improved |
| use TAP selection to cut hyperparameter search cost | TAP · LiDAR segmentation | gpu time per candidate: 30 gpu hours -> 7 minutes | Improved |
| vary number of images used to compute TAP (very few images) (ablation) | TAP · DUNE tasks | selected model | No clear effect |
| vary feature distance (l1, l2, cosine) in TAP (ablation) | TAP · DUNE tasks | TAP curve | No clear effect |
Closest papers by meaning
- No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces (ICML 2025 · similarity 0.89)
- ZipIt! Merging Models from Different Tasks without Training (ICLR 2024 · similarity 0.89)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.88)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.88)
- CABS: Conflict-Aware and Balanced Sparsification for Enhancing Model Merging (ICML 2025 · similarity 0.88)
- Localizing Task Information for Improved Model Merging and Compression (ICML 2024 · 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.
