Research map / Large language models / Instruction and fine-tuning
From Memorization to Parameter Interference: How Overtraining Experts Harms Model Merging
Stefan Horoi, Guy Wolf, Eugene Belilovsky, Gintare Karolina Dziugaite
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 |
|---|---|---|---|
| fine-tune experts 256 steps instead of 2048 (ablation) | Task Arithmetic merge of FFT ViT-B-32 · 8 vision tasks | merged accuracy (%): 69.8 -> 72.7 | Improved |
| fine-tune LoRA experts 128 steps instead of 2048 (ablation) | TIES merge of LoRA ViT-B-32 · 8 vision tasks | merged accuracy (%): 50.7 -> 66.9 | Improved |
| train experts longer (256 to 2048 steps) (ablation) | FFT ViT-B-32 experts · 8 vision tasks | expert accuracy (%): 88.4% -> 90.0% | Improved |
| train experts longer (256 to 2048 steps) (ablation) | FFT T5-Base experts · 7 NLP tasks | expert accuracy (%): 78.2% -> 81.9% | Improved |
| train experts longer (256 to 2048 steps) (ablation) | Task Arithmetic, TIES, DARE merges of FFT ViT-B-32 · 8 vision tasks | merged accuracy (%): about 3% drop | Hurt |
| train experts longer (ablation) | Average merge of T5-Base · 7 NLP tasks | merged accuracy (%): peak 65.2 maintained | No clear effect |
| use early stopping on expert training | TIES merge of T5-Base · 7 NLP tasks | merged accuracy (%): 72.2 -> 76.9 | Improved |
| use early stopping on expert training | Average merge of T5-Base · 7 NLP tasks | merged accuracy (%): 65.2 -> 63.4 | Hurt |
| use early stopping on LoRA ViT experts | Task Arithmetic merge · 8 vision tasks | merged accuracy (%): 61.1 -> 68.0 | Improved |
| use early stopping on FFT ViT experts | Average merge · 8 vision tasks | merged accuracy (%): 65.9 -> 64.5 | Hurt |
| remove hardest examples (top 1-10% EL2N) from expert training (ablation) | merges of ViT-B-32 experts · 8 vision tasks | merged accuracy | Hurt |
| increase LoRA rank (ablation) | LoRA experts merging · vision/NLP | merged accuracy | Improved |
Closest papers by meaning
- Mitigating Parameter Interference in Model Merging via Sharpness-Aware Fine-Tuning (ICLR 2025 · similarity 0.91)
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.90)
- Parameter Competition Balancing for Model Merging (NeurIPS 2024 · similarity 0.90)
- Model Merging in Pre-training of Large Language Models (NeurIPS 2025 · similarity 0.90)
- LiNeS: Post-training Layer Scaling Prevents Forgetting and Enhances Model Merging (ICLR 2025 · similarity 0.89)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · 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.
