Research map / Architectures and efficiency / Pruning and efficient tuning
Escaping Optimization Stagnation: Taking Steps Beyond Task Arithmetic via Difference Vectors
Jinping Wang, Zhiqiang Gao, Zhang Dinggen, Zhiwu Xie
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 |
|---|---|---|---|
| use difference vector direction instead of random perturbation of same magnitude (ablation) | DV-BASI · Unsupervised model merging (ViT-B/32) | accuracy: 5.6 -> 83.9 | Improved |
| use anisotropic block-wise scaling instead of scalar scaling (ablation) | DV-BASI · Unsupervised model merging (ViT-B/16) | accuracy: 83.6 -> 85.3 | Improved |
| add DV-BASI to aTLAS | aTLAS · Task addition, 8 datasets (unsupervised) | absolute accuracy: 82.8 -> 85.3 (ViT-B/16); 90.2 -> 90.8 (ViT-L/14) | Improved |
| add DV-BASI to tauJP reg | tauJP reg · Task addition, 8 datasets (unsupervised) | absolute accuracy: 84.5 -> 86.7 (ViT-B/32) | Improved |
| add DV-BASI to supervised aTLAS | aTLAS · Task addition, 8 datasets (supervised) | absolute accuracy: 82.9 -> 87.2 (ViT-B/16) | Improved |
| add multi-objective optimization (MOO) to DV-BASI (ablation) | aTLAS + DV-BASI · Task addition, 8 datasets (supervised) | absolute accuracy: 85.9 -> 86.2 (ViT-B/32) | Improved |
| add DV-BASI to aTLAS for task negation | aTLAS · Task negation target accuracy (supervised) | target accuracy: 19.4 -> 10.7 (ViT-B/32) | Improved |
| add DV-BASI to tauJP reg for task negation (unsupervised) | tauJP reg · Task negation target accuracy (unsupervised) | target accuracy: 6.7 -> 5.7 (ViT-B/32) | Improved |
| add DV-BASI to aTLAS for test-time adaptation | aTLAS · Test-time adaptation (22 datasets) | accuracy: 65.16 -> 66.76 (ViT-B/32) | Improved |
| apply DV-BASI after single-task fine-tuning | fine-tuned CLIP ViT-B/32 · 20 datasets average | accuracy: 82.04 -> 82.74 | Improved |
| vary initial scaling coefficient from 0.1 to 1.0 (ablation) | DV-BASI · Unsupervised model merging | accuracy | No clear effect |
Closest papers by meaning
- When is Task Vector Provably Effective for Model Editing? A Generalization Analysis of Nonlinear Transformers (ICLR 2025 · similarity 0.86)
- Training-Free Task Vectors for LLM Behavioral Control (NeurIPS 2026 · similarity 0.86)
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024 · similarity 0.86)
- Efficient Model Editing with Task-Localized Sparse Fine-tuning (ICLR 2025 · similarity 0.86)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.85)
- Task Singular Vectors: Reducing Task Interference in Model Merging (CVPR 2025 · similarity 0.85)
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.
