@inproceedings{shabgahi-etal-2026-moes,
title = "{M}o{E}s Are Stronger than You Think: Hyper-Parallel Inference Scaling with {R}o{E}s",
author = "Shabgahi, Soheil Zibakhsh and
Samragh, Mohammad and
Nishu, Kumari and
Hannah, Lauren and
Kundu, Arnav and
Cho, Minsik",
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.2143/",
doi = "10.18653/2026.findings-emnlp.2143",
pages = "45797--45814",
ISBN = "979-8-89176-462-0",
abstract = "The generation quality of large language models (LLMs) is often improved by utilizing inference-time scaling methods such as Chain-of-Thought (sequential scaling) or Self-Consistency (parallel scaling). We introduce Roster of Experts (RoE), a training-free inference algorithm for Mixture-of-Experts (MoE) models that defines a third, orthogonal scaling direction. Rather than generating longer outputs or aggregating multiple sequences, RoE improves the next token prediction distribution by treating a single MoE as a dynamic ensemble of internal subnetworks. It injects controlled stochasticity into the expert routing mechanism via Gumbel-Top-k sampling, enabling the model to aggregate multiple diverse subnetworks into a single, higher-quality token prediction. Crucially, because RoE improves the model{'}s output distribution rather than sampling from it, it is the only approach in this family that is applicable to open-ended generation tasks where sequence-level aggregation is infeasible. Furthermore, RoE is orthogonal to existing scaling methods; combining RoE with Self-Consistency yields greater accuracy than either method alone. To overcome the computational cost of multiple forward passes at inference, we employ an efficient batching strategy that exploits hardware parallelism to minimize latency, requiring no fine-tuning of model parameters."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shabgahi-etal-2026-moes">
<titleInfo>
<title>MoEs Are Stronger than You Think: Hyper-Parallel Inference Scaling with RoEs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Soheil</namePart>
<namePart type="given">Zibakhsh</namePart>
<namePart type="family">Shabgahi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohammad</namePart>
<namePart type="family">Samragh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kumari</namePart>
<namePart type="family">Nishu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lauren</namePart>
<namePart type="family">Hannah</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Arnav</namePart>
<namePart type="family">Kundu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minsik</namePart>
<namePart type="family">Cho</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>The generation quality of large language models (LLMs) is often improved by utilizing inference-time scaling methods such as Chain-of-Thought (sequential scaling) or Self-Consistency (parallel scaling). We introduce Roster of Experts (RoE), a training-free inference algorithm for Mixture-of-Experts (MoE) models that defines a third, orthogonal scaling direction. Rather than generating longer outputs or aggregating multiple sequences, RoE improves the next token prediction distribution by treating a single MoE as a dynamic ensemble of internal subnetworks. It injects controlled stochasticity into the expert routing mechanism via Gumbel-Top-k sampling, enabling the model to aggregate multiple diverse subnetworks into a single, higher-quality token prediction. Crucially, because RoE improves the model’s output distribution rather than sampling from it, it is the only approach in this family that is applicable to open-ended generation tasks where sequence-level aggregation is infeasible. Furthermore, RoE is orthogonal to existing scaling methods; combining RoE with Self-Consistency yields greater accuracy than either method alone. To overcome the computational cost of multiple forward passes at inference, we employ an efficient batching strategy that exploits hardware parallelism to minimize latency, requiring no fine-tuning of model parameters.</abstract>
<identifier type="citekey">shabgahi-etal-2026-moes</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.2143</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2143/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>45797</start>
<end>45814</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T MoEs Are Stronger than You Think: Hyper-Parallel Inference Scaling with RoEs
%A Shabgahi, Soheil Zibakhsh
%A Samragh, Mohammad
%A Nishu, Kumari
%A Hannah, Lauren
%A Kundu, Arnav
%A Cho, Minsik
%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 shabgahi-etal-2026-moes
%X The generation quality of large language models (LLMs) is often improved by utilizing inference-time scaling methods such as Chain-of-Thought (sequential scaling) or Self-Consistency (parallel scaling). We introduce Roster of Experts (RoE), a training-free inference algorithm for Mixture-of-Experts (MoE) models that defines a third, orthogonal scaling direction. Rather than generating longer outputs or aggregating multiple sequences, RoE improves the next token prediction distribution by treating a single MoE as a dynamic ensemble of internal subnetworks. It injects controlled stochasticity into the expert routing mechanism via Gumbel-Top-k sampling, enabling the model to aggregate multiple diverse subnetworks into a single, higher-quality token prediction. Crucially, because RoE improves the model’s output distribution rather than sampling from it, it is the only approach in this family that is applicable to open-ended generation tasks where sequence-level aggregation is infeasible. Furthermore, RoE is orthogonal to existing scaling methods; combining RoE with Self-Consistency yields greater accuracy than either method alone. To overcome the computational cost of multiple forward passes at inference, we employ an efficient batching strategy that exploits hardware parallelism to minimize latency, requiring no fine-tuning of model parameters.
%R 10.18653/2026.findings-emnlp.2143
%U https://aclanthology.org/2026.findings-emnlp.2143/
%U https://doi.org/10.18653/2026.findings-emnlp.2143
%P 45797-45814
Markdown (Informal)
[MoEs Are Stronger than You Think: Hyper-Parallel Inference Scaling with RoEs](https://aclanthology.org/2026.findings-emnlp.2143/) (Shabgahi et al., Findings 2026)
ACL