Research map / Architectures and efficiency / Pruning and efficient tuning
When Shared Knowledge Hurts: Spectral Over-Accumulation in Model Merging
Yayuan Li, Ze Peng, Jian Zhang, Jintao Guo, Yue Duan, Yinghuan Shi
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 |
|---|---|---|---|
| add SVC calibration to Task Arithmetic | Task Arithmetic · CV 8 tasks | average accuracy (%): 68.9 -> 81.9 | Improved |
| add SVC calibration to Task Arithmetic | Task Arithmetic · CV 14 tasks | average accuracy (%): 46.4 -> 63.1 | Improved |
| add SVC calibration to TSV-M | TSV-M · CV 8 tasks | average accuracy (%): 84.0 -> 84.8 | Improved |
| add SVC calibration to Iso-CTS | Iso-CTS · CV 14 tasks (ViT-B/32) | average accuracy (%): 76.7 -> 76.7 | No clear effect |
| add SVC calibration to Task Arithmetic | Llama2-7B · GSM8K | accuracy: 46.1 -> 52.2 | Improved |
| add SVC calibration to TSV-M | Llama2-7B · GSM8K | accuracy: 51.9 -> 51.9 | No clear effect |
| add SVC calibration to Task Arithmetic | BERT (FT) · NLP 11 classification tasks | average accuracy (%): 56.9 -> 69.0 | Improved |
| add SVC calibration to TIES | T0 (PEFT) · NLP 11 classification tasks | average accuracy (%): 54.1 -> 54.1 | No clear effect |
| use left (output-space) rather than right singular vectors for overlap (ablation) | SVC with Task Arithmetic · CV 8 tasks | average accuracy (%): 64.9 -> 81.9 | Improved |
| use left (output-space) rather than right singular vectors for overlap (ablation) | SVC with TIES · CV 8 tasks | average accuracy (%): 65.7 -> 80.0 | Improved |
| use left (output-space) rather than right singular vectors for overlap (ablation) | SVC with Iso-CTS · CV 8 tasks | average accuracy (%): 85.5 -> 85.6 | No clear effect |
| lower alpha below 1 to also boost under-accumulated subspaces (ablation) | SVC · CV 8 tasks | average accuracy (%) | Mixed |
Closest papers by meaning
- No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces (ICML 2025 · similarity 0.90)
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.89)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.89)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.87)
- EMR-Merging: Tuning-Free High-Performance Model Merging (NeurIPS 2024 · similarity 0.87)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · 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.
