AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

Capacity-Aware Inference: Mitigating the Straggler Effect in Mixture of Experts

Shwai He, Weilin Cai, Jiayi Huang, Ang Li

ICLR 2026 · Architectures and efficiency › Mixture of experts · approach: Expert routing and specialization

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
select dropped tokens by gating score instead of order or random (ablation)Token Drop · OLMoE avg 8 tasks, gamma=1.0accuracy: 53.1 -> 61.1 (vs Random)Improved
keep low-load experts and drop tokens instead of skipping the lowest-load experts (ablation)Token Drop vs Expert Drop · OLMoE-Instruct avg 9 tasksaccuracy: 60.5 -> 62.3Improved
reroute overflow tokens to low-load experts (Expanded Drop) instead of discarding (ablation)Token Drop · DeepSeek-V2-Lite-Chat avg 9 tasksaccuracy: 68.2 -> 68.9Improved
reroute overflow tokens to low-load experts (Expanded Drop) (ablation)Token Drop · Mixtral-8x7B-Instruct avg 9 tasksaccuracy: 73.8 -> 74.5Improved
lower the capacity factor gamma from 2.0 to 1.0 (ablation)Token Drop (Score) · OLMoE avg 8 tasksaccuracy: 64.0 -> 61.1Hurt
lower capacity factor gamma (ablation)Token Drop / Expanded Drop · End-to-end inference speedspeedupImproved
apply capacity constraint at device level instead of expert level (ablation)Expanded Drop · Qwen3-MoE avg 9 tasksaccuracy: 73.9 -> 74.8Improved
device-level constraint instead of expert-level (ablation)Expanded Drop · Qwen3-MoE end-to-end speedupspeedup: 1.23x -> 1.31xImproved
drop image tokens first instead of text first (ablation)Capacity-aware inference · MolmoE MME Perceptionscore: 1114.2 -> 1346.5 (Token Drop)Improved
remove the max-k cap on selected experts in Expanded Drop (ablation)Expanded Drop · OLMoE avg 8 tasks, m=2kaccuracy: 61.0 -> 61.2No clear effect
deploy more experts per GPU (ablation)Capacity-aware inference · End-to-end speedupspeedupHurt

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