Research map / Optimization / Nonconvex and smooth optimization
Rethinking Convergence in MoE Training: The Role of Routing Sparsity
Weihao Zhu, Long Shi, Kang Wei, Zhe Wang, Yipeng Zhou, Haixia Zhang
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 |
|---|---|---|---|
| increase activated experts K from 1 to 2 (ablation) | MoE Transformer · WikiText-103 | iterations to reach loss 1.71: 10000 -> 3200 | Improved |
| increase K from 1 to 2 at fixed 10000 iterations (ablation) | MoE Transformer · WikiText-103 | loss: 1.71 -> 0.93 | Improved |
| increase K from 2 to 4 at fixed 10000 iterations (ablation) | MoE Transformer · WikiText-103 | loss: 0.93 -> 0.83 | Improved |
| increase K beyond 2 under a fixed FLOPs budget (ablation) | MoE Transformer · WikiText-103 and Alpaca (M=16, fixed R) | loss | Hurt |
| increase K under a fixed FLOPs budget (ablation) | MoE CNN · CIFAR-10 (fixed R) | accuracy | Mixed |
| increase K from 1 (K>1 range) for routing stability (ablation) | MoE Transformer · Alpaca, Dolly, GSM8K | loss gap sparse vs dense | Improved |
Closest papers by meaning
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.92)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.89)
- Scaling Beyond the GPU Memory Limit for Large Mixture-of-Experts Model Training (ICML 2024 · similarity 0.88)
- Mixture of Experts Provably Detect and Learn the Latent Cluster Structure in Gradient-Based Learning (ICML 2025 · similarity 0.88)
- ReMoE: Fully Differentiable Mixture-of-Experts with ReLU Routing (ICLR 2025 · similarity 0.88)
- Advancing Expert Specialization for Better MoE (NeurIPS 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.
