Research map / Architectures and efficiency / Pruning and efficient tuning
Towards Performance Consistency in Multi-Level Model Collaboration
Qi Li, Runpeng Yu, Xinchao Wang
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 NeuLig CoopVecs instead of simple averaging for merging | CLIP-ViT-B/32 5 models · 5-dataset avg | merging accuracy: 78.19 -> 95.44 | Improved |
| use NeuLig instead of Simple-Averaging | CLIP-RN50 5 models · 5-dataset avg | merging accuracy: 50.39 -> 92.07 | Improved |
| use NeuLig instead of WeMoE | CLIP-ViT-B/32 5 models · 5-dataset avg | merge-ensemble gap: 1.32 -> 0.02 | Improved |
| use NeuLig instead of Simple-Averaging | CLIP-RN50 5 models · 5-dataset avg | merge-ensemble gap: 39.34 -> 0.61 | Improved |
| use supervised instead of semi-supervised boosting term (ablation) | NeuLig CLIP-ViT-B/32 · 5-dataset avg | merging accuracy: 95.44 -> 95.37 | No clear effect |
| use supervised instead of semi-supervised boosting term (ablation) | NeuLig CLIP-RN50 · 5-dataset avg | merge-ensemble gap: 0.61 -> 1.22 | Hurt |
| increase collaborating CLIP-RN50 models from 2 to 5 (ablation) | NeuLig · avg accuracy | merge-ensemble gap: 0.17 -> 1.22 | Hurt |
| reduce unlabeled data scale from 1.0 to 0.01 (ablation) | NeuLig CLIP-ViT-B/32 · semi-supervised | merging accuracy: 95.44 -> 91.43 | Hurt |
| reduce unlabeled data scale from 1.0 to 0.01 (ablation) | NeuLig CLIP-RN50 · semi-supervised | merging accuracy: 92.07 -> 80.93 | Hurt |
| reduce unlabeled data scale from 1.0 to 0.1 (ablation) | NeuLig CLIP-ViT-B/32 · semi-supervised | merging accuracy: 95.44 -> 93.46 | Hurt |
| train models from random initialization instead of shared checkpoint (ablation) | NeuLig · 2 models GTSRB-MNIST | merge vs ensemble: 82.17 vs 82.19 | No clear effect |
| compare ViT with ResNet backbone | Simple-Averaging · 5-dataset avg | merging accuracy: 50.39 -> 78.19 | Mixed |
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 2 of 2 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
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.89)
- Parameter Competition Balancing for Model Merging (NeurIPS 2024 · similarity 0.89)
- CALM: Consensus-Aware Localized Merging for Multi-Task Learning (ICML 2025 · similarity 0.89)
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024 · similarity 0.88)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.88)
- No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces (ICML 2025 · similarity 0.88)
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.
