Research map / Large language models / Instruction and fine-tuning
Expert Merging: Model Merging with Unsupervised Expert Alignment and Importance-Guided Layer Chunking
Dengming Zhang, Xiaowen Ma, Zhen-Liang Ni, Zhenkai Wu, Han Shu, Xin Jiang, Xinghao Chen
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 logit alignment (ablation) | Expert Merging (InternVL2.5-1B) · Multi-task merging average | average score: 54.72 -> 58.11 | Improved |
| add logit alignment (ablation) | Expert Merging (Mistral-7B) · Multi-task merging average | average score: 43.62 -> 47.43 | Improved |
| add logit alignment (ablation) | Expert Merging (Qwen2-VL-7B) · Multi-task merging average | average score: 62.71 -> 63.34 | Improved |
| add hidden-state alignment (ablation) | Expert Merging (InternVL2.5-1B) · Multi-task merging average | average score: 57.75 -> 58.11 | Improved |
| add hidden-state alignment (ablation) | Expert Merging (Qwen2-VL-7B) · Multi-task merging average | average score: 62.43 -> 63.34 | Improved |
| add coefficient regularization (ablation) | Expert Merging (InternVL2.5-1B) · Multi-task merging average | average score: 47.57 -> 58.11 | Improved |
| add coefficient regularization (ablation) | Expert Merging (Mistral-7B) · Multi-task merging average | average score: 40.81 -> 47.43 | Improved |
| add coefficient regularization (ablation) | Expert Merging (Qwen2-VL-7B) · Multi-task merging average | average score: 62.17 -> 63.34 | Improved |
| add explicit task weights (ablation) | Expert Merging · Multi-task merging average | average score: 57.71 -> 58.11 (InternVL) | Improved |
| add importance-guided chunk-wise coefficients (ablation) | Expert Merging (Mistral-7B) · Multi-task merging average | average score: 47.43 -> 48.71 | Improved |
| add importance-guided chunk-wise coefficients (ablation) | Expert Merging (InternVL2.5-1B) · Multi-task merging average | average score: 58.11 -> 58.45 | Improved |
| chunk all layers (2x coefficients) instead of importance-guided (ablation) | Expert Merging++ (InternVL2.5-1B) · Multi-task merging average | average score: 58.45 -> 56.76 | Hurt |
Closest papers by meaning
- Model Merging in Pre-training of Large Language Models (NeurIPS 2025 · similarity 0.92)
- Channel Merging: Preserving Specialization for Merged Experts (AAAI 2025 · similarity 0.92)
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.91)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.91)
- OptMerge: Unifying Multimodal LLM Capabilities and Modalities via Model Merging (ICLR 2026 · similarity 0.91)
- Activation-Informed Merging of Large Language Models (NeurIPS 2025 · similarity 0.91)
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.
