@inproceedings{tian-etal-2026-dual,
title = "Dual-Bias Load Balancing: Mitigating Expert-Bias Staleness in Auxiliary-Loss-Free Mixture-of-Experts",
author = "Tian, Yuxin and
Zhao, Hailin and
Liu, Jia and
Zhang, Zhiqiang",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1786/",
pages = "38193--38206",
ISBN = "979-8-89176-462-0",
abstract = "Auxiliary-loss-free load balancing (ALF-LB) achieves promising performance in Mixture-of-Experts large language model (MoE-LLM) pre-training as it decouples load balancing from the main objective. However, we observe an unexplored phenomenon: routing imbalance is concentrated in shallow MoE layers. We attribute this phenomenon to the interaction between representation homogeneity, which makes shallow-layer routing decisions highly sensitive, and expert-bias staleness, which applies stale load corrections to the current router state. To break the cycle, we propose \textbf{Dual-Bias Load Balancing (DLB)}, which augments the standard global expert bias with an intra-step micro-bias residual. DLB introduces the effective bias $\mathbf{b}_{\mathrm{eff}}=\mathbf{b}_g+\Delta\mathbf{b}$, where $\mathbf{b}_g$ is updated once per gradient step and $\Delta\mathbf{b}$ is updated across micro-batches within the same step. DLB does not add forward or backward passes and integrates naturally into existing large-scale MoE training pipelines. Experiments on large-scale MoE pre-training show that DLB mitigates expert-bias staleness, alleviates shallow-layer imbalance, and yields lower test perplexity and stronger average downstream performance."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tian-etal-2026-dual">
<titleInfo>
<title>Dual-Bias Load Balancing: Mitigating Expert-Bias Staleness in Auxiliary-Loss-Free Mixture-of-Experts</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yuxin</namePart>
<namePart type="family">Tian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hailin</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jia</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhiqiang</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Auxiliary-loss-free load balancing (ALF-LB) achieves promising performance in Mixture-of-Experts large language model (MoE-LLM) pre-training as it decouples load balancing from the main objective. However, we observe an unexplored phenomenon: routing imbalance is concentrated in shallow MoE layers. We attribute this phenomenon to the interaction between representation homogeneity, which makes shallow-layer routing decisions highly sensitive, and expert-bias staleness, which applies stale load corrections to the current router state. To break the cycle, we propose Dual-Bias Load Balancing (DLB), which augments the standard global expert bias with an intra-step micro-bias residual. DLB introduces the effective bias \mathbfb_eff=\mathbfb_g+Δ\mathbfb, where \mathbfb_g is updated once per gradient step and Δ\mathbfb is updated across micro-batches within the same step. DLB does not add forward or backward passes and integrates naturally into existing large-scale MoE training pipelines. Experiments on large-scale MoE pre-training show that DLB mitigates expert-bias staleness, alleviates shallow-layer imbalance, and yields lower test perplexity and stronger average downstream performance.</abstract>
<identifier type="citekey">tian-etal-2026-dual</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1786/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>38193</start>
<end>38206</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Dual-Bias Load Balancing: Mitigating Expert-Bias Staleness in Auxiliary-Loss-Free Mixture-of-Experts
%A Tian, Yuxin
%A Zhao, Hailin
%A Liu, Jia
%A Zhang, Zhiqiang
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F tian-etal-2026-dual
%X Auxiliary-loss-free load balancing (ALF-LB) achieves promising performance in Mixture-of-Experts large language model (MoE-LLM) pre-training as it decouples load balancing from the main objective. However, we observe an unexplored phenomenon: routing imbalance is concentrated in shallow MoE layers. We attribute this phenomenon to the interaction between representation homogeneity, which makes shallow-layer routing decisions highly sensitive, and expert-bias staleness, which applies stale load corrections to the current router state. To break the cycle, we propose Dual-Bias Load Balancing (DLB), which augments the standard global expert bias with an intra-step micro-bias residual. DLB introduces the effective bias \mathbfb_eff=\mathbfb_g+Δ\mathbfb, where \mathbfb_g is updated once per gradient step and Δ\mathbfb is updated across micro-batches within the same step. DLB does not add forward or backward passes and integrates naturally into existing large-scale MoE training pipelines. Experiments on large-scale MoE pre-training show that DLB mitigates expert-bias staleness, alleviates shallow-layer imbalance, and yields lower test perplexity and stronger average downstream performance.
%U https://aclanthology.org/2026.findings-emnlp.1786/
%P 38193-38206
Markdown (Informal)
[Dual-Bias Load Balancing: Mitigating Expert-Bias Staleness in Auxiliary-Loss-Free Mixture-of-Experts](https://aclanthology.org/2026.findings-emnlp.1786/) (Tian et al., Findings 2026)
ACL