AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

EasyBalance: Cross-Layer Load Balancing in Distributed MoE Inference

Yize Wu, KE GAO, Ling Li, Yanjun Wu

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 EasyBalance cross-layer schedulingQwen3-30B · 2wikimqa EP=8GPU under-utilization: 0.37 -> 0.21Improved
add EasyBalance cross-layer schedulingMoonlight-16B · repobench-p EP=8GPU under-utilization: 0.46 -> 0.29Improved
add EasyBalance cross-layer schedulingQwen3-30B · trec EP=2GPU under-utilization: 0.08 -> 0.04Improved
add micro-batching (4 splits) (ablation)Qwen3-30B · 2wikimqa EP=8end-to-end latency (s): 10.15 -> 8.60Improved
add micro-batching (4 splits) (ablation)Moonlight-16B · 2wikimqa EP=8end-to-end latency (s): 5.06 -> 4.40Improved
use MaxUtil instead of linear heuristics CumUtil/DiffPeak (ablation)EasyBalance · 2wikimqa Qwen3-30Bend-to-end latency (s): 8.27 -> 8.21Improved
use cheaper DiffPeak scheduler instead of MaxUtil (ablation)EasyBalance · 2wikimqa Qwen3-30Bscheduling latency per step (ms): 0.277 -> 0.130Improved
add EasyBalance on top of EPLB(128) (ablation)EPLB · 2wikimqa Qwen3-30BGPU under-utilization: 0.23 -> 0.14Improved
add EasyBalance on top of EPLB(136) (ablation)EPLB with replicas · trec Qwen3-30BGPU under-utilization: 0.28 -> 0.17Improved
add 8 redundant experts to EPLB (128 to 136 physical) (ablation)EPLB · 2wikimqa Qwen3-30BGPU under-utilization: 0.23 -> 0.22No clear effect
increase expert-parallel size from 2 to 8 without EasyBalance (ablation)Qwen3-30B · 2wikimqaGPU under-utilization: 0.10 -> 0.37Hurt
use very small micro-batch count (2) (ablation)EasyBalance · LongBenchimbalance mitigationHurt

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