Research map / Architectures and efficiency / Mixture of experts
DTop-p MoE: Sparsity-Controlled Dynamic Top-p MoE for Foundation Model Pre-training
Can Jin, Hongwu Peng, Mingcan Xiang, Qixin Zhang, Xiangchi Yuan, Amit Hasan, Ohi Dibua, Yifan Gong, Yan Kang, Dimitris N. Metaxas
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 |
|---|---|---|---|
| replace Top-k routing with PI-controlled dynamic Top-p plus routing normalization | MoE-1.3B-6.9B-64E8A · 13-benchmark average | average accuracy: 49.0 -> 50.9 | Improved |
| replace fixed-threshold Top-p with DTop-p | MoE-1.3B-6.9B-64E8A · 13-benchmark average | average accuracy: 49.3 -> 50.9 | Improved |
| replace fixed Top-k with Top-p | MoE-1.3B-6.9B-64E8A · 13-benchmark average | average accuracy: 49.0 -> 49.3 | No clear effect |
| use DTop-p instead of Top-k | MoE-1.3B-6.9B-64E8A · SVAMP 5-shot | exact match: 10.3 -> 16.0 | Improved |
| use DTop-p instead of Top-k | MoE-1.3B-6.9B · NLP validation loss | validation loss: 2.6384 -> 2.6193 | Improved |
| increase expert granularity to 128E16A with DTop-p | MoE-1.3B-6.9B · 13-benchmark average | average accuracy: 49.7 -> 51.4 | Improved |
| use DTop-p instead of Top-k at 128E16A | MoE-1.3B-6.9B-128E16A · 13-benchmark average | average accuracy: 49.3 -> 51.4 | Improved |
| increase total experts 16 to 64 with DTop-p | MoE-1.3B · 13-benchmark average | average accuracy: 47.3 -> 50.9 | Improved |
| scale model from 0.4B to 2.4B activated with DTop-p | MoE 64E8A · 13-benchmark average | average accuracy: 45.2 -> 54.1 | Improved |
| scale training data 100B to 300B tokens with DTop-p | MoE-1.3B-6.9B-64E8A · 13-benchmark average | average accuracy: 50.9 -> 53.8 | Improved |
| add PI controller (to Top-p with DRN) (ablation) | DTop-p MoE · 30B-token ablation | validation loss / activated experts | Improved |
| add Dynamic Routing Normalization (ablation) | DTop-p MoE · 30B-token ablation | validation loss | Improved |
Closest papers by meaning
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.90)
- From Sparse to Soft Mixtures of Experts (ICLR 2024 · similarity 0.88)
- Mixture-of-Experts Operator Transformer for Large-Scale PDE Pre-Training (NeurIPS 2025 · similarity 0.87)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.87)
- ReMoE: Fully Differentiable Mixture-of-Experts with ReLU Routing (ICLR 2025 · similarity 0.87)
- Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations (NeurIPS 2025 · similarity 0.87)
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.
