AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

Towards Greater Leverage: Scaling Laws for Efficient Mixture-of-Experts Language Models

Changxin Tian, Kunlong Chen, Jia Liu, Ziqi Liu, Zhiqiang Zhang, JUN ZHOU

ICLR 2026 · Architectures and efficiency › Mixture of experts · approach: Efficient MoE language model systems · 36 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
lower the expert activation ratio (more sparsity) at fixed FLOPs per token (ablation)MoE LM · IsoFLOPs 1e18-3e20training loss / Efficiency Leverage: loss decreases monotonically; lowest ratio 0.8% bestImproved
increase compute budget for a given MoE architecture (ablation)MoE LM · Efficiency LeverageELImproved
vary expert granularity from 2 to 16 (ablation)MoE LM · IsoFLOPs 1e18-1e20training loss: U-shaped; optimum around 12 (range 8 to 12)Mixed
add a small non-zero shared expert ratio (vs none) (ablation)MoE LM · IsoFLOPstraining loss: best S=16.7% below 1e20 FLOPs, 8.3% aboveImproved
raise shared expert ratio up to 83.3% (ablation)MoE LM · IsoFLOPstraining lossHurt
replace first 1-3 MoE layers with dense layers (ablation)MoE LM (60 layers) · up to 3e20 FLOPsEfficiency Leverage: EL close to 1No clear effect
vary attention FLOPs ratio between 20% and 50% (ablation)MoE LM · pretraining lossperformanceNo clear effect
poor routing balance (ablation)MoE LM · granularity studyoptimal granularity / lossHurt
use 17.5B MoE (3.4% activation, G=12) instead of Dense-6.1B at 1T tokensLing-mini-beta · downstream suite overall averageaverage score: 44.0 -> 45.5Improved
use MoE instead of dense on BBHLing-mini-beta · BBHscore: 48.0 -> 35.7Hurt
use MoE with larger batch size and slightly lower LR than dense at large computehyperparameter scaling laws · optimal hyperparametersB_opt, eta_opt: B exponent 0.3644, LR exponent -0.1529Mixed

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