Research map / Architectures and efficiency / Mixture of experts
ExpertWeaver: Unlocking the Inherent MoE in Dense LLMs with GLU Activation Patterns
Ziyu Zhao, Tong Zhu, Xin Yu, Zhi Zhang, Tiantian Fan, Jinluan Yang, Kun Kuang, zhongyu wei, Fei Wu, Yu Cheng
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 multi-task Flan-v2 calibration instead of C4 (ablation) | ExpertWeaver · 5-benchmark avg, Qwen2.5-7B 25% sparsity | average accuracy: 65.8 -> 67.0 | Improved |
| increase calibration tasks from 25% to 50% of Flan-v2 (ablation) | ExpertWeaver · Qwen2.5-7B 25% sparsity avg | average accuracy: 63.9 -> 66.1 | Improved |
| increase calibration tasks from 50% to 100% of Flan-v2 (ablation) | ExpertWeaver · Qwen2.5-7B 25% sparsity avg | average accuracy: 66.1 -> 67.0 | Improved |
| increase calibration tasks from 50% to 75% (ablation) | ExpertWeaver · Qwen2.5-7B 25% sparsity avg | average accuracy: 66.1 -> 65.6 | Hurt |
| layer-aware dynamic shared-expert allocation instead of uniform ratio (ablation) | ExpertWeaver · MMLU, Qwen2.5-7B | MMLU accuracy | Improved |
| use AbsTopk-GLU gating at 50% sparsity instead of Wanda | LLaMA3-8B · 5-benchmark avg | average accuracy: 66.3 -> 68.2 | Improved |
| use AbsTopk-GLU instead of magnitude pruning | LLaMA3-8B 50% sparsity · 5-benchmark avg | average accuracy: 49.6 -> 68.2 | Improved |
| use ExpertWeaver instead of CMoE | LLaMA3-8B 25% sparsity · 5-benchmark avg | average accuracy: 57.2 -> 60.4 | Improved |
| use ExpertWeaver instead of LLM-Pruner | Qwen2.5-7B 25% sparsity · 5-benchmark avg | average accuracy: 65.0 -> 67.0 | Improved |
| use ExpertWeaver instead of LLM-Pruner on GSM8K | Qwen2.5-7B 25% sparsity · GSM8K | accuracy: 2.0 -> 34.9 | Improved |
| downcycle Qwen2.5-7B with ExpertWeaver vs OLMoE-1B-7B* trained from scratch | MoE 200B tokens · 8-benchmark avg | average accuracy: 58.9 -> 63.5 | Improved |
| ExpertWeaver downcycling vs sparse upcycling and from-scratch training | OLMo-1.3B · training loss and downstream performance | loss/accuracy | Improved |
Closest papers by meaning
- Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning (ICLR 2024 · similarity 0.89)
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.89)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.88)
- Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models (ICLR 2024 · similarity 0.88)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.88)
- DeRS: Towards Extremely Efficient Upcycled Mixture-of-Experts Models (CVPR 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.
