@inproceedings{gao-etal-2025-mola,
title = "{M}o{LA}: {M}o{E} {L}o{RA} with Layer-wise Expert Allocation",
author = "Gao, Chongyang and
Chen, Kezhen and
Rao, Jinmeng and
Liu, Ruibo and
Sun, Baochen and
Zhang, Yawen and
Peng, Daiyi and
Guo, Xiaoyuan and
Subrahmanian, Vs",
editor = "Chiruzzo, Luis and
Ritter, Alan and
Wang, Lu",
booktitle = "Findings of the Association for Computational Linguistics: NAACL 2025",
month = apr,
year = "2025",
address = "Albuquerque, New Mexico",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.findings-naacl.284/",
doi = "10.18653/v1/2025.findings-naacl.284",
pages = "5097--5112",
ISBN = "979-8-89176-195-7",
abstract = "Recent efforts to integrate low-rank adaptation (LoRA) with the Mixture-of-Experts (MoE) have managed to achieve performance comparable to full-parameter fine-tuning by tuning much fewer parameters. Despite promising results, research on improving the efficiency and expert analysis of LoRA with MoE is still in its early stages. Recent studies have shown that experts in the MoE architecture have different strengths and also exhibit some redundancy. Does this statement also apply to parameter-efficient MoE? In this paper, we introduce a novel parameter-efficient MoE method, \textit{ \textbf{M}oE-L\textbf{o}RA with \textbf{L}ayer-wise Expert \textbf{A}llocation (MoLA)} for Transformer-based models, where each model layer uses a varying number of LoRA experts. We investigate several architectures with varying layer-wise expert configurations. Experiments on six well-known NLP and commonsense QA benchmarks demonstrate that MoLA achieves equal or superior performance compared to all baselines on top of both LLAMA-2, Mistral, and Gemma. We find that allocating more LoRA experts to middle layers further enhances the effectiveness of models with a certain number of experts in total. The redundancy of the experts is more obvious in the lower layers. With much fewer parameters, this allocation strategy outperforms the setting with the same number of experts in every layer. This work can be widely used as a plug-and-play parameter-efficient tuning approach for various applications. The code has been made available at \url{https://github.com/GCYZSL/MoLA}."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="gao-etal-2025-mola">
<titleInfo>
<title>MoLA: MoE LoRA with Layer-wise Expert Allocation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chongyang</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kezhen</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jinmeng</namePart>
<namePart type="family">Rao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruibo</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Baochen</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yawen</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Daiyi</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaoyuan</namePart>
<namePart type="family">Guo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vs</namePart>
<namePart type="family">Subrahmanian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-04</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: NAACL 2025</title>
</titleInfo>
<name type="personal">
<namePart type="given">Luis</namePart>
<namePart type="family">Chiruzzo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alan</namePart>
<namePart type="family">Ritter</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Albuquerque, New Mexico</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-195-7</identifier>
</relatedItem>
<abstract>Recent efforts to integrate low-rank adaptation (LoRA) with the Mixture-of-Experts (MoE) have managed to achieve performance comparable to full-parameter fine-tuning by tuning much fewer parameters. Despite promising results, research on improving the efficiency and expert analysis of LoRA with MoE is still in its early stages. Recent studies have shown that experts in the MoE architecture have different strengths and also exhibit some redundancy. Does this statement also apply to parameter-efficient MoE? In this paper, we introduce a novel parameter-efficient MoE method, MoE-LoRA with Layer-wise Expert Allocation (MoLA) for Transformer-based models, where each model layer uses a varying number of LoRA experts. We investigate several architectures with varying layer-wise expert configurations. Experiments on six well-known NLP and commonsense QA benchmarks demonstrate that MoLA achieves equal or superior performance compared to all baselines on top of both LLAMA-2, Mistral, and Gemma. We find that allocating more LoRA experts to middle layers further enhances the effectiveness of models with a certain number of experts in total. The redundancy of the experts is more obvious in the lower layers. With much fewer parameters, this allocation strategy outperforms the setting with the same number of experts in every layer. This work can be widely used as a plug-and-play parameter-efficient tuning approach for various applications. The code has been made available at https://github.com/GCYZSL/MoLA.</abstract>
<identifier type="citekey">gao-etal-2025-mola</identifier>
<identifier type="doi">10.18653/v1/2025.findings-naacl.284</identifier>
<location>
<url>https://aclanthology.org/2025.findings-naacl.284/</url>
</location>
<part>
<date>2025-04</date>
<extent unit="page">
<start>5097</start>
<end>5112</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T MoLA: MoE LoRA with Layer-wise Expert Allocation
%A Gao, Chongyang
%A Chen, Kezhen
%A Rao, Jinmeng
%A Liu, Ruibo
%A Sun, Baochen
%A Zhang, Yawen
%A Peng, Daiyi
%A Guo, Xiaoyuan
%A Subrahmanian, Vs
%Y Chiruzzo, Luis
%Y Ritter, Alan
%Y Wang, Lu
%S Findings of the Association for Computational Linguistics: NAACL 2025
%D 2025
%8 April
%I Association for Computational Linguistics
%C Albuquerque, New Mexico
%@ 979-8-89176-195-7
%F gao-etal-2025-mola
%X Recent efforts to integrate low-rank adaptation (LoRA) with the Mixture-of-Experts (MoE) have managed to achieve performance comparable to full-parameter fine-tuning by tuning much fewer parameters. Despite promising results, research on improving the efficiency and expert analysis of LoRA with MoE is still in its early stages. Recent studies have shown that experts in the MoE architecture have different strengths and also exhibit some redundancy. Does this statement also apply to parameter-efficient MoE? In this paper, we introduce a novel parameter-efficient MoE method, MoE-LoRA with Layer-wise Expert Allocation (MoLA) for Transformer-based models, where each model layer uses a varying number of LoRA experts. We investigate several architectures with varying layer-wise expert configurations. Experiments on six well-known NLP and commonsense QA benchmarks demonstrate that MoLA achieves equal or superior performance compared to all baselines on top of both LLAMA-2, Mistral, and Gemma. We find that allocating more LoRA experts to middle layers further enhances the effectiveness of models with a certain number of experts in total. The redundancy of the experts is more obvious in the lower layers. With much fewer parameters, this allocation strategy outperforms the setting with the same number of experts in every layer. This work can be widely used as a plug-and-play parameter-efficient tuning approach for various applications. The code has been made available at https://github.com/GCYZSL/MoLA.
%R 10.18653/v1/2025.findings-naacl.284
%U https://aclanthology.org/2025.findings-naacl.284/
%U https://doi.org/10.18653/v1/2025.findings-naacl.284
%P 5097-5112
Markdown (Informal)
[MoLA: MoE LoRA with Layer-wise Expert Allocation](https://aclanthology.org/2025.findings-naacl.284/) (Gao et al., Findings 2025)
ACL
- Chongyang Gao, Kezhen Chen, Jinmeng Rao, Ruibo Liu, Baochen Sun, Yawen Zhang, Daiyi Peng, Xiaoyuan Guo, and Vs Subrahmanian. 2025. MoLA: MoE LoRA with Layer-wise Expert Allocation. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 5097–5112, Albuquerque, New Mexico. Association for Computational Linguistics.