AI Paper Atlas

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

ICML 2026 · Architectures and efficiency › Mixture of experts · approach: Expert routing and specialization · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace Top-k routing with PI-controlled dynamic Top-p plus routing normalizationMoE-1.3B-6.9B-64E8A · 13-benchmark averageaverage accuracy: 49.0 -> 50.9Improved
replace fixed-threshold Top-p with DTop-pMoE-1.3B-6.9B-64E8A · 13-benchmark averageaverage accuracy: 49.3 -> 50.9Improved
replace fixed Top-k with Top-pMoE-1.3B-6.9B-64E8A · 13-benchmark averageaverage accuracy: 49.0 -> 49.3No clear effect
use DTop-p instead of Top-kMoE-1.3B-6.9B-64E8A · SVAMP 5-shotexact match: 10.3 -> 16.0Improved
use DTop-p instead of Top-kMoE-1.3B-6.9B · NLP validation lossvalidation loss: 2.6384 -> 2.6193Improved
increase expert granularity to 128E16A with DTop-pMoE-1.3B-6.9B · 13-benchmark averageaverage accuracy: 49.7 -> 51.4Improved
use DTop-p instead of Top-k at 128E16AMoE-1.3B-6.9B-128E16A · 13-benchmark averageaverage accuracy: 49.3 -> 51.4Improved
increase total experts 16 to 64 with DTop-pMoE-1.3B · 13-benchmark averageaverage accuracy: 47.3 -> 50.9Improved
scale model from 0.4B to 2.4B activated with DTop-pMoE 64E8A · 13-benchmark averageaverage accuracy: 45.2 -> 54.1Improved
scale training data 100B to 300B tokens with DTop-pMoE-1.3B-6.9B-64E8A · 13-benchmark averageaverage accuracy: 50.9 -> 53.8Improved
add PI controller (to Top-p with DRN) (ablation)DTop-p MoE · 30B-token ablationvalidation loss / activated expertsImproved
add Dynamic Routing Normalization (ablation)DTop-p MoE · 30B-token ablationvalidation lossImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy