Research map / Architectures and efficiency / Pruning and efficient tuning
GLOBA: Rethinking Parameter Conflicts in Model Merging
Zehao Liu, Kun Li, Wei Zhou
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 GLOBA instead of Task Arithmetic when merging Math and Code | 7B merge · MBPP (Math+Code, 7B) | pass@1 (%): 15.00 -> 18.80 | Improved |
| use GLOBA when merging Instruct and Code | 7B merge · MBPP (Instruct+Code, 7B) | pass@1 (%): 20.80 -> 31.60 | Improved |
| use GLOBA instead of Task Arithmetic when merging Math and Code | 13B merge · MBPP (Math+Code, 13B) | pass@1 (%): 8.80 -> 21.80 | Improved |
| use GLOBA instead of TIES-Merging when merging Math and Code | 13B merge · MBPP (Math+Code, 13B) | pass@1 (%): 15.60 -> 21.80 | Improved |
| use GLOBA instead of TA when merging Instruct and Code | 13B merge · AlpacaEval (Instruct+Code, 13B) | win rate (%): 77.97 -> 82.12 | Improved |
| use GLOBA instead of TIES when merging Math and Code (GSM8K) | 7B merge · GSM8K (Math+Code, 7B) | accuracy (%): 4.32 -> 51.48 | Improved |
| use GLOBA instead of TIES for merging Instruct and Math | 7B merge · AlpacaEval (Instruct+Math, 7B) | win rate (%): 39.93 -> 39.14 | No clear effect |
| add completely orthogonal (Type A) patterns after pruning (ablation) | pruned base model · Math+Instruct merge, 7B | GSM8K accuracy (%): 49.20 -> 50.27 | No clear effect |
| add same-sign overlap (Type D+) patterns to WizardMath-7B (ablation) | WizardMath-7B · GSM8K (Math+Code, 7B) | accuracy (%): 51.63 -> 5.38 | Hurt |
| add opposite-sign (Type D-) patterns to WizardMath-7B (ablation) | WizardMath-7B · GSM8K (Math+Code, 7B) | accuracy (%): 51.63 -> 0 | Hurt |
| add all overlap patterns for Math+Code merge (ablation) | WizardMath-7B · GSM8K (Math+Code, 7B) | accuracy (%): 51.63 -> 43.14 | Hurt |
| add all patterns for Instruct into Math model (ablation) | WizardMath-7B · AlpacaEval (Math+Instruct, 7B) | win rate (%): 13.00 -> 38.57 | Improved |
Closest papers by meaning
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.89)
- Parameter Competition Balancing for Model Merging (NeurIPS 2024 · similarity 0.88)
- No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces (ICML 2025 · similarity 0.88)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.87)
- CAT Merging: A Training-Free Approach for Resolving Conflicts in Model Merging (ICML 2025 · similarity 0.87)
- Mixture-Trained Merging for Unified Multi-Objective Models (NeurIPS 2026 · 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.
