Research map / Architectures and efficiency / Pruning and efficient tuning
FW-Merging: Scaling Model Merging with Frank-Wolfe Optimization
Hao Mark Chen, Shell Xu Hu, Wayne Luk, Timothy Hospedales, Hongxiang Fan
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 layer-wise LMO instead of task-wise (ablation) | FW-TA · NLP discriminative tasks | avg score: 78.2 -> 85.4 | Improved |
| use layer-wise LMO instead of task-wise (ablation) | FW-TA · vision 4 CV tasks | avg accuracy: 73.7 -> 74.0 | Improved |
| use task-wise LMO instead of layer-wise (ablation) | FW-AM · vision 4 CV tasks | avg accuracy: 79.7 -> 80.1 | Improved |
| optimize merging coefficients lambda (ablation) | FW-AM · vision 4 CV tasks | avg accuracy: 70.3 -> 80.1 | Improved |
| use soft FW (multiple top-k vertices) instead of hard FW (ablation) | FW-AM vs FW-TA · vision merging many models | avg accuracy: up to 6.7 points | Improved |
| initialize FW-Merging from Adamerging instead of pre-trained model (ablation) | FW-Merging · vision benchmark | accuracy | Improved |
| initialize FW-Merging from Task Arithmetic instead of pre-trained model (ablation) | FW-Merging · vision benchmark | accuracy | Hurt |
| apply FW-Merging on top of Ties-Merging as merging function (ablation) | FW-Merging · NLP and vision | accuracy | Improved |
| add 16 irrelevant models to merging pool | FW-AM · 4 CV tasks | avg accuracy: 73.9 to 74.1 | No clear effect |
| add irrelevant models to merging pool | Task Arithmetic · 4 CV tasks | avg accuracy: 49.1% drop | Hurt |
| add irrelevant models to merging pool | DARE · 4 CV tasks | avg accuracy: 64.4% drop | Hurt |
| add relevant models to merging pool | FW-AM · 20 CV tasks | avg accuracy: 59.2 -> 68.3 | 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 5 of 5 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
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.91)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.90)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.90)
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024 · similarity 0.89)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.89)
- OptMerge: Unifying Multimodal LLM Capabilities and Modalities via Model Merging (ICLR 2026 · 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.
