AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

Grouter: Decoupling Routing from Representation for Accelerated MoE Training

Yuqi Xu, Rizhen Hu, Zihan Liu, Mou Sun, Kun Yuan

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
add positional information to the MHA Transformer encoder Grouter (ablation)Grouter · Mini-DS-V2-Lite, 10B tokensvalidation loss: 3.367 -> 3.318Improved
use MHA Transformer encoder instead of decoder (ablation)Grouter · Mini-DS-V2-Lite, 10B tokensvalidation loss: 3.342 -> 3.318Improved
use MHA instead of MLA in the encoder Grouter (ablation)Grouter · Mini-DS-V2-Lite, 10B tokensvalidation loss: 3.356 -> 3.318Improved
use Transformer encoder instead of residual linear layer (ablation)Grouter · Mini-DS-V2-Lite, 10B tokensvalidation loss: 3.375 -> 3.318Improved
add positional information to the residual linear Grouter (ablation)Grouter · Mini-DS-V2-Lite, 10B tokensvalidation loss: 3.359 -> 3.375Hurt
add positional information to the MLA decoder Grouter (ablation)Grouter · Mini-DS-V2-Lite, 10B tokensvalidation loss: 3.387 -> 3.335Improved
add positional information to the MLA encoder Grouter (ablation)Grouter · Mini-DS-V2-Lite, 10B tokensvalidation loss: 3.360 -> 3.356No clear effect
use Grouter routing prior instead of Z-Loss + Aux LossMoE pretraining · Mini-Qwen3 (50B tokens)validation loss: 2.740 -> 1.763Improved
use Grouter routing prior instead of Z-Loss + Aux LossMoE pretraining · Mini-GPT-OSS (30B tokens)validation loss: 3.346 -> 2.940Improved
use Grouter routing prior instead of Z-Loss + Aux LossMoE pretraining · Downstream tasks (6 benchmarks)average score: average improvement of 2.80 across six benchmarksImproved
use Grouter with expert-placement-based communication reduction on top of DeepEPMoE training throughput · 1, 2 and 4 nodesthroughput speedup: 33.5%, 19.3%, and 15.5% on 1, 2, and 4 nodesImproved
add lightweight load-balancing expert tuning (50M tokens) to Grouter Raw (ablation)Grouter · MoE pretraining (30B tokens)validation loss / load balanceImproved

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