Research map / Architectures and efficiency / Mixture of experts
Optimal Sparsity of Mixture-of-Experts Language Models for Reasoning Tasks
Taishi Nakamura, Satoki Ishikawa, Masaki Kawamura, Takumi Okamoto, Daisuke Nohara, Jun Suzuki, Rio Yokota
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 total parameters by adding experts (higher sparsity) at fixed width and top-k (ablation) | Mixtral-style MoE (16 layers) · GSM8K and GSM-Plus | task loss and accuracy | Hurt |
| increase total parameters by adding experts at fixed width and top-k (ablation) | Mixtral-style MoE (16 layers) · TriviaQA and HellaSwag | task loss and accuracy | Improved |
| increase top-k (active parameters) at matched pre-training loss or fixed TPP (ablation) | Mixtral-style MoE · GSM8K and GSM-Plus | reasoning accuracy | Improved |
| change top-k while keeping active parameters constant (ablation) | Mixtral-style MoE · memorization and reasoning benchmarks | performance | No clear effect |
| make MoE denser at large active-parameter budgets (ablation) | Mixtral-style MoE · GSM8K and GSM-Plus (fixed active params) | task loss and accuracy | Improved |
| increase sparsity at fixed active parameters (ablation) | Mixtral-style MoE · TriviaQA and HellaSwag | task loss and accuracy | Improved |
| tune tokens per parameter toward about 20 (ablation) | Mixtral-style MoE · reasoning tasks | accuracy: peak near TPP of 20 | Improved |
| lower tokens per parameter (more parameters) (ablation) | Mixtral-style MoE · TriviaQA and HellaSwag | accuracy | Improved |
| add GRPO post-training on GSM8K (ablation) | MoE models · GSM8K | accuracy | No clear effect |
| add test-time compute (self-consistency with 128 samples) (ablation) | MoE models · GSM8K (zero-shot) | accuracy | No clear effect |
| increase top-k only at inference time (ablation) | MoE d=2048, 128 experts, top-2 · GSM8K with TTC | Pass@1 | No clear effect |
| lower learning rate and smaller initialization scale (ablation) | MoE E=16, k=2 · reasoning benchmarks at matched training loss | downstream accuracy | Improved |
Closest papers by meaning
- Parameters vs FLOPs: Scaling Laws for Optimal Sparsity for Mixture-of-Experts Language Models (ICML 2025 · similarity 0.94)
- Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models (ICLR 2024 · similarity 0.92)
- Mixture of Parrots: Experts improve memorization more than reasoning (ICLR 2025 · similarity 0.91)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.90)
- Joint MoE Scaling Laws: Mixture of Experts Can Be Memory Efficient (ICML 2025 · similarity 0.90)
- Scaling Laws for Fine-Grained Mixture of Experts (ICML 2024 · similarity 0.90)
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.
