Research map / Architectures and efficiency / Pruning and efficient tuning
MergOPT: A Merge-Aware Optimizer for Robust Model Merging
Enneng Yang, Qun Yang, Peng Wang, Anke Tang, Guibing Guo, Xiaochun Cao, Li Shen
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 with MergOPT before merging with DARE | DARE merging · TraceBench 7 tasks (Llama-3.2-1B) | average score: 0.3786 -> 0.4147 | Improved |
| fine-tune experts with MergOPT before merging with Task Arithmetic | Task Arithmetic · TraceBench 7 tasks (Llama-3.2-3B) | average score: 0.4871 -> 0.5045 | Improved |
| fine-tune experts with MergOPT before merging with TIES-Merging | TIES-Merging · TraceBench 7 tasks (Llama-3.2-3B) | average score: 0.4898 -> 0.5098 | Improved |
| use MergOPT fine-tuning on Qwen2.5-1.5B with Task Arithmetic | Task Arithmetic · TraceBench 7 tasks (Qwen2.5-1.5B) | average score: 0.4484 -> 0.4731 | Improved |
| use MergOPT at 8B scale (4 tasks) | Weight Averaging, Task Arithmetic, TIES, DARE · TraceBench 4 tasks (Llama-3.1-8B) | average score: +1.08%, +1.40%, +0.46%, +0.51% | Improved |
| merge fewer tasks (4-task groups) with MergOPT | Task Arithmetic · TraceBench 4-task groups (Llama-3.2-1B) | average score: 0.3708 -> 0.3851; 0.4959 -> 0.5130 | Improved |
| MergOPT fine-tuning changes single-expert performance (before merging) | individual experts · TraceBench 7 tasks (Llama-3.2-3B) | average score: 0.6057 -> 0.5836 | Hurt |
| replace SAFT-Merge (SAM) with MergOPT at equal epochs | merging with Weight Averaging, DARE vs Task Arithmetic, TIES · TraceBench 7 tasks (Llama-3.2-1B) | average score: +2.59%, +5.20% vs -2.23%, -1.29% | Mixed |
| replace SAFT-Merge with MergOPT at comparable training time | merging methods · TraceBench 7 tasks (Llama-3.2-1B) | average score: +4.08%, +0.70%, +0.95%, +3.99% | Improved |
| tangent-space fine-tuning instead of standard fine-tuning | Weight Averaging on ViT-B/32 · CLIP-ViT-B/32 10 tasks | average accuracy: 54.9 -> 50.5 | Hurt |
| MergOPT instead of standard fine-tuning on ViT-B/32 with Weight Averaging | Weight Averaging · CLIP-ViT-B/32 10 tasks | average accuracy: 54.9 -> 56.2 | Improved |
| vary Laplace scale b in MergOPT (ablation) | Task Arithmetic w/ MergOPT · TraceBench 7 tasks (Llama-3.2-1B) | average score: 0.4055 -> 0.4164 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TIES-Merging — TIES-Merging: Resolving Interference When Merging Models (NeurIPS 2023) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.90)
- EMR-Merging: Tuning-Free High-Performance Model Merging (NeurIPS 2024 · similarity 0.90)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.90)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.90)
- CAT Merging: A Training-Free Approach for Resolving Conflicts in Model Merging (ICML 2025 · similarity 0.90)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.90)
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.
