@inproceedings{qiu-etal-2025-bag,
title = "Bag of Tricks for Sparse Mixture-of-Experts: A Benchmark Across Reasoning, Efficiency, and Safety",
author = "Qiu, Mufan and
Shen, Zheyu and
Li, Pingzhi and
Li, Ang and
Chen, Tianlong",
editor = "Christodoulopoulos, Christos and
Chakraborty, Tanmoy and
Rose, Carolyn and
Peng, Violet",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2025",
month = nov,
year = "2025",
address = "Suzhou, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.findings-emnlp.43/",
doi = "10.18653/v1/2025.findings-emnlp.43",
pages = "817--835",
ISBN = "979-8-89176-335-7",
abstract = "Mixture-of-Experts (MoE) has emerged as a promising approach for scaling large language models efficiently. However, how to design a desired MoE architecture given performance, efficiency, or safety goals remains absent. Existing benchmarks often focus on isolated aspects (e.g., reasoning, efficiency, safety), and there is a lack of consensus on optimal design choices, such as the number and size of experts, the type of routers, and the regularization during pre-training, or strategies like freezing, learning rate adjustments, and limiting expert collaboration during fine-tuning, with prior works often yielding conflicting conclusions. Motivated by this research gap, we introduce MoEBench, the first comprehensive assessment of MoE designs across the three dimensions of reasoning ability, efficiency, and safety. Our benchmark systematically evaluates optimal architectural choices during both pre-training and fine-tuning phases. We evaluate two popular MoE backbones across four dimensions of design choices on over eight metrics. Our empirical findings uncover hidden underlying correlations among MoE design choices. Specifically, we observe that (1) token-level routing and z-loss regularization improve reasoning performance; (2) shared experts enhance training stability but reduce specialization; and (3) collaboration-constrained routing and freezing strategies significantly influence load balance, specialization, and safety alignment. Furthermore, we propose three ``sweet point'' combinations of optimal strategies tailored to different scenarios. We hope this study provides actionable insights for building more robust, efficient, and secure MoE models. Code, checkpoints, and raw data will be released upon acceptance of the paper."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="qiu-etal-2025-bag">
<titleInfo>
<title>Bag of Tricks for Sparse Mixture-of-Experts: A Benchmark Across Reasoning, Efficiency, and Safety</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mufan</namePart>
<namePart type="family">Qiu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zheyu</namePart>
<namePart type="family">Shen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pingzhi</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ang</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tianlong</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2025</title>
</titleInfo>
<name type="personal">
<namePart type="given">Christos</namePart>
<namePart type="family">Christodoulopoulos</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tanmoy</namePart>
<namePart type="family">Chakraborty</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Carolyn</namePart>
<namePart type="family">Rose</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Violet</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Suzhou, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-335-7</identifier>
</relatedItem>
<abstract>Mixture-of-Experts (MoE) has emerged as a promising approach for scaling large language models efficiently. However, how to design a desired MoE architecture given performance, efficiency, or safety goals remains absent. Existing benchmarks often focus on isolated aspects (e.g., reasoning, efficiency, safety), and there is a lack of consensus on optimal design choices, such as the number and size of experts, the type of routers, and the regularization during pre-training, or strategies like freezing, learning rate adjustments, and limiting expert collaboration during fine-tuning, with prior works often yielding conflicting conclusions. Motivated by this research gap, we introduce MoEBench, the first comprehensive assessment of MoE designs across the three dimensions of reasoning ability, efficiency, and safety. Our benchmark systematically evaluates optimal architectural choices during both pre-training and fine-tuning phases. We evaluate two popular MoE backbones across four dimensions of design choices on over eight metrics. Our empirical findings uncover hidden underlying correlations among MoE design choices. Specifically, we observe that (1) token-level routing and z-loss regularization improve reasoning performance; (2) shared experts enhance training stability but reduce specialization; and (3) collaboration-constrained routing and freezing strategies significantly influence load balance, specialization, and safety alignment. Furthermore, we propose three “sweet point” combinations of optimal strategies tailored to different scenarios. We hope this study provides actionable insights for building more robust, efficient, and secure MoE models. Code, checkpoints, and raw data will be released upon acceptance of the paper.</abstract>
<identifier type="citekey">qiu-etal-2025-bag</identifier>
<identifier type="doi">10.18653/v1/2025.findings-emnlp.43</identifier>
<location>
<url>https://aclanthology.org/2025.findings-emnlp.43/</url>
</location>
<part>
<date>2025-11</date>
<extent unit="page">
<start>817</start>
<end>835</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Bag of Tricks for Sparse Mixture-of-Experts: A Benchmark Across Reasoning, Efficiency, and Safety
%A Qiu, Mufan
%A Shen, Zheyu
%A Li, Pingzhi
%A Li, Ang
%A Chen, Tianlong
%Y Christodoulopoulos, Christos
%Y Chakraborty, Tanmoy
%Y Rose, Carolyn
%Y Peng, Violet
%S Findings of the Association for Computational Linguistics: EMNLP 2025
%D 2025
%8 November
%I Association for Computational Linguistics
%C Suzhou, China
%@ 979-8-89176-335-7
%F qiu-etal-2025-bag
%X Mixture-of-Experts (MoE) has emerged as a promising approach for scaling large language models efficiently. However, how to design a desired MoE architecture given performance, efficiency, or safety goals remains absent. Existing benchmarks often focus on isolated aspects (e.g., reasoning, efficiency, safety), and there is a lack of consensus on optimal design choices, such as the number and size of experts, the type of routers, and the regularization during pre-training, or strategies like freezing, learning rate adjustments, and limiting expert collaboration during fine-tuning, with prior works often yielding conflicting conclusions. Motivated by this research gap, we introduce MoEBench, the first comprehensive assessment of MoE designs across the three dimensions of reasoning ability, efficiency, and safety. Our benchmark systematically evaluates optimal architectural choices during both pre-training and fine-tuning phases. We evaluate two popular MoE backbones across four dimensions of design choices on over eight metrics. Our empirical findings uncover hidden underlying correlations among MoE design choices. Specifically, we observe that (1) token-level routing and z-loss regularization improve reasoning performance; (2) shared experts enhance training stability but reduce specialization; and (3) collaboration-constrained routing and freezing strategies significantly influence load balance, specialization, and safety alignment. Furthermore, we propose three “sweet point” combinations of optimal strategies tailored to different scenarios. We hope this study provides actionable insights for building more robust, efficient, and secure MoE models. Code, checkpoints, and raw data will be released upon acceptance of the paper.
%R 10.18653/v1/2025.findings-emnlp.43
%U https://aclanthology.org/2025.findings-emnlp.43/
%U https://doi.org/10.18653/v1/2025.findings-emnlp.43
%P 817-835
Markdown (Informal)
[Bag of Tricks for Sparse Mixture-of-Experts: A Benchmark Across Reasoning, Efficiency, and Safety](https://aclanthology.org/2025.findings-emnlp.43/) (Qiu et al., Findings 2025)
ACL