Research map / Architectures and efficiency / Mixture of experts
Mixture-of-Experts Can Surpass Dense LLMs Under Strictly Equal Resource
Houyi Li, Ka Man Lo, Shijie Xuyang, Ziqi Wang, Wenzhen Zheng, Haocheng Zhang, Zhao Li, Shuigeng Zhou, Xiangyu Zhang, Daxin Jiang
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 MoE at 20% activation rate instead of dense at equal compute | 2B model · Validation BPC | BPC: 0.4921 -> 0.4857 | Improved |
| use MoE at 20% activation rate instead of dense at equal compute | 7B model · Validation BPC | BPC: 0.4736 -> 0.4543 | Improved |
| use MoE at 20% AR instead of dense with 2x compute | 7B model · Validation BPC | BPC: 0.4594 -> 0.4543 | Improved |
| use very sparse MoE (AR 8.7%) instead of AR ~20% (ablation) | 2B MoE · Validation BPC | BPC: 0.4857 -> 0.5048 | Hurt |
| use denser MoE (AR >50%) instead of AR ~20% (ablation) | 2B MoE · Validation BPC | BPC: 0.4857 -> 0.4933 (50.80%), 0.4934 (57.73%) | Hurt |
| use 1dense+SE layer arrangement instead of interleave (ablation) | 2B MoE · pretraining loss | loss: 1.8620 -> 1.8557 | Improved |
| add gate score normalization (ablation) | 2B MoE · pretraining loss | loss: 1.6712 -> 1.6726 | No clear effect |
| increase shared expert fraction of activated FFN (ablation) | 2B MoE · pretraining loss | loss: 1.6752 -> 1.6712 | No clear effect |
| use fine-grained top-K 11 instead of K=1 (AR ~27%) (ablation) | 2B MoE · pretraining loss | loss: 2.0470 -> 2.0338 | Improved |
| use fine-grained experts (K=22 of 88) instead of K=2 of 8 at AR 44% (ablation) | 2B MoE · pretraining loss | loss: 1.9996 -> 2.0266 | Hurt |
| add strict data reuse (68B unique tokens, multi-epoch) (ablation) | 7B MoE AR 20% · Validation BPC | BPC: 0.4543 -> 0.4590 | Hurt |
| add strict data reuse in pretraining (ablation) | 7B MoE AR 20% · MMLU (SFT) | accuracy: 32.92 -> 24.57 | Hurt |
Closest papers by meaning
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.93)
- Mixture Compressor for Mixture-of-Experts LLMs Gains More (ICLR 2025 · similarity 0.93)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.92)
- MoE-CAP: Benchmarking Cost, Accuracy and Performance of Sparse Mixture-of-Experts Systems (NeurIPS 2025 · similarity 0.92)
- Scaling Laws for Upcycling Mixture-of-Experts Language Models (ICML 2025 · similarity 0.92)
- $\textit{Read-ME}$: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design (NeurIPS 2024 · similarity 0.91)
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.
