Research map / Learning theory and algorithms / Kernels and regression theory
Generalization and Scaling Laws for Mixture-of-ExpertsTransformers
Mansour MAYAKI
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 expert pool M from 16 to 64 at fixed k=4 (ablation) | MoE LM · routing ablation | validation loss: 2.73 -> 2.37 | Improved |
| increase routing term k log(eM/k) in moderate regime (M/k <= 8) (ablation) | MoE LM · routing ablation | validation loss | Hurt |
| increase number of experts E from 1 to 32 (ablation) | MoE scaling laws (JMSL data) · per-E identity check | residual r_D: 0.0427 -> 0.1304 | Hurt |
| increase number of experts E from 1 to 32 (ablation) | MoE scaling laws (JMSL data) · per-E identity check | ratio alpha_D/alpha_N: 1.08 -> 1.70 | Improved |
| increase E from 1 to 32 in amplification kappa (ablation) | MoE scaling laws (JMSL data) · per-E identity check | kappa(E): 1.28 -> 1.97 | Improved |
| raise beta from 1.0 to 1.5 in theory | exponent fit · OpenWebText | alpha_N: 0.044 -> 0.067 | Improved |
| measure model exponent against active parameters | MoE LM · WikiText-103 | alpha_N empirical vs theory: 0.063 -> 0.060 | No clear effect |
| measure data exponent against theory | MoE LM · TinyStories | alpha_D empirical vs theory: 0.080 -> 0.081 | No clear effect |
| raise compute-optimal tokens per active parameter via more experts, E from 1 to 16 at budget 1e20 vs 1e21 (ablation) | JMSL allocations · compute-optimal D*/N* | D*/N* gain: x4.5 vs x3.0 | Mixed |
Closest papers by meaning
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.90)
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks (NeurIPS 2025 · similarity 0.89)
- Joint MoE Scaling Laws: Mixture of Experts Can Be Memory Efficient (ICML 2025 · similarity 0.89)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.89)
- Tight Clusters Make Specialized Experts (ICLR 2025 · similarity 0.88)
- Scaling Beyond the GPU Memory Limit for Large Mixture-of-Experts Model Training (ICML 2024 · 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.
