AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

$\phi$-Balancing for Mixture-of-Experts Training

Lizhang Chen, Jonathan Li, Qi Wang, Runlong Liao, Shuozhe Li, Chen Liang, Ni Lao, qiang liu

ICML 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
use phi-balancing (negative entropy) instead of Shannon-free l_p p=1 potential (ablation)phi-balancing · Mirror map ablationvalidation loss: 3.142 -> 3.084Improved
use negative Shannon entropy instead of l_p p=2 (ablation)phi-balancing · Mirror map ablationMaxVio_global: 0.610 -> 0.104Improved
use negative Shannon entropy instead of softplus potential (ablation)phi-balancing · Mirror map ablationMaxVio_global: 0.810 -> 0.104Improved
use negative Renyi (alpha=0.95) instead of Shannon (ablation)phi-balancing · Mirror map ablationvalidation loss: 3.091 vs 3.084No clear effect
use phi-balancing instead of ST-MoEDeepSeek-MoE-16B-Chat · 7-benchmark fine-tuning averageaverage accuracy: 51.95 -> 53.92Improved
use phi-balancing instead of ST-MoEDeepSeek-V2-Lite · 7-benchmark fine-tuning averageaverage accuracy: 49.29 -> 51.86Improved
use phi-balancing instead of ST-MoEMoonlight-16B-A3B-Instruct · 7-benchmark fine-tuning averageaverage accuracy: 73.37 -> 76.07Improved
use phi-balancing instead of ST-MoEMoonlight-16B-A3B-Instruct · Math500accuracy: 64.80 -> 66.20Hurt
use phi-balancing instead of ST-MoE at global batch 32Gemma-MoE 986M · MMLUaccuracy: 41.96 -> 42.88Improved
increase global batch size from 32 to 512 (ablation)phi-balancing · MMLUaccuracy: 42.88 -> 43.18Improved
track EMA of routing probabilities instead of selection frequencies (ablation)phi-balancing · Validation loss 111Mvalidation loss: 3.089 -> 3.0847No clear effect
track EMA of routing probabilities instead of selection frequencies (ablation)phi-balancing · Validation loss 986Mvalidation loss: 2.598 -> 2.6019No clear effect

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