AI Paper Atlas

Research map / Federated learning and privacy / Personalized and heterogeneous FL

FedEMoE: Improving Personalization on Heterogeneous Federated Learning via Elastic Mixture of Experts Architecture

Haizhou Du, Lixin Huang, Zonghan Wu, Huan Huo

ICML 2026 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Bayesian personalized FL

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 knowledge exchange between personalized and shared experts (ablation)FedEMoE · CIFAR-100 (K=50)accuracy: 43.85 -> 47.69Improved
add personalized expert group multi-scale extraction (ablation)FedEMoE · CIFAR-100 (K=50)accuracy: 35.46 -> 47.69Improved
add shared experts with elastic EMoE architecture (ablation)FedEMoE · CIFAR-100 (K=50)accuracy: 42.25 -> 47.69Improved
use FedEMoE instead of best baseline FedMRLHtFL · CIFAR-100 (K=50)accuracy: 36.01 -> 47.69Improved
use FedEMoE instead of best baseline FedSAHtFL · Tiny-ImageNet (K=50)accuracy: 21.27 -> 29.91Improved
use FedEMoE instead of best baseline FedMRLHtFL · CIFAR-10 (K=50)accuracy: 78.37 -> 79.83Improved
use FedEMoE instead of FedGC in homogeneous settingFL · CIFAR-100 homogeneous (K=50)accuracy: 38.71 -> 47.53Improved
use FedEMoE instead of FedGC in homogeneous settingFL · CIFAR-10 homogeneous (K=50)accuracy: 78.68 -> 79.22Improved
increase clients from 20 to 100 (ablation)FedEMoE · CIFAR-100accuracy: 49.03 -> 41.76Hurt
restrict gate to topK=1 instead of topK=2 at N=3 (ablation)FedEMoE · CIFAR-100 (K=20)accuracy: 49.03 -> 49.60No clear effect
increase shared experts from N=3 to N=5 (ablation)FedEMoE · CIFAR-100 (K=20)accuracy: 49.03 -> 46.88Hurt
expansion threshold v0=0.85 rather than 0.95 at gamma=0.07 (ablation)FedEMoE · CIFAR-100 (K=20)accuracy: 47.26 -> 49.03Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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