Research map / Architectures and efficiency / Mixture of experts
Towards Greater Leverage: Scaling Laws for Efficient Mixture-of-Experts Language Models
Changxin Tian, Kunlong Chen, Jia Liu, Ziqi Liu, Zhiqiang Zhang, JUN ZHOU
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 |
|---|---|---|---|
| lower the expert activation ratio (more sparsity) at fixed FLOPs per token (ablation) | MoE LM · IsoFLOPs 1e18-3e20 | training loss / Efficiency Leverage: loss decreases monotonically; lowest ratio 0.8% best | Improved |
| increase compute budget for a given MoE architecture (ablation) | MoE LM · Efficiency Leverage | EL | Improved |
| vary expert granularity from 2 to 16 (ablation) | MoE LM · IsoFLOPs 1e18-1e20 | training loss: U-shaped; optimum around 12 (range 8 to 12) | Mixed |
| add a small non-zero shared expert ratio (vs none) (ablation) | MoE LM · IsoFLOPs | training loss: best S=16.7% below 1e20 FLOPs, 8.3% above | Improved |
| raise shared expert ratio up to 83.3% (ablation) | MoE LM · IsoFLOPs | training loss | Hurt |
| replace first 1-3 MoE layers with dense layers (ablation) | MoE LM (60 layers) · up to 3e20 FLOPs | Efficiency Leverage: EL close to 1 | No clear effect |
| vary attention FLOPs ratio between 20% and 50% (ablation) | MoE LM · pretraining loss | performance | No clear effect |
| poor routing balance (ablation) | MoE LM · granularity study | optimal granularity / loss | Hurt |
| use 17.5B MoE (3.4% activation, G=12) instead of Dense-6.1B at 1T tokens | Ling-mini-beta · downstream suite overall average | average score: 44.0 -> 45.5 | Improved |
| use MoE instead of dense on BBH | Ling-mini-beta · BBH | score: 48.0 -> 35.7 | Hurt |
| use MoE with larger batch size and slightly lower LR than dense at large compute | hyperparameter scaling laws · optimal hyperparameters | B_opt, eta_opt: B exponent 0.3644, LR exponent -0.1529 | Mixed |
Closest papers by meaning
- Scaling Laws for Fine-Grained Mixture of Experts (ICML 2024 · similarity 0.94)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.93)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.93)
- Parameters vs FLOPs: Scaling Laws for Optimal Sparsity for Mixture-of-Experts Language Models (ICML 2025 · similarity 0.92)
- How to Scale Mixture-of-Experts: From muP to the Maximally Scale-Stable Parameterization (NeurIPS 2026 · similarity 0.92)
- Scaling Laws for Upcycling Mixture-of-Experts Language Models (ICML 2025 · similarity 0.92)
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.
