@inproceedings{tian-etal-2026-moe,
title = "{M}o{E} Sparsity Begets Monosemanticity: Insights into Understanding and Fine-Tuning",
author = "Tian, Mingxue and
Zhang, Qi and
Wang, Xiaohan and
Chai, Jiajun and
Yin, Guojun and
Lin, Wei and
Chen, Cailian and
Xu, Qimin and
Wang, Siyu and
Wang, Yisen",
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.975/",
pages = "20613--20630",
ISBN = "979-8-89176-462-0",
abstract = "Mixture-of-Experts (MoE) architectures enhance the scalability and efficiency of large language models (LLMs) by activating only a subset of parameters. However, the interpretability of individual experts and corresponding strategies for post-training adaptation to domain-specific tasks remain underexplored. In this work, we develop an interpretability framework for expert-level monosemanticity in MoE models, drawing inspiration from sparse autoencoders and offering new insights into the specialization patterns of domain experts. We demonstrate that the inherent sparsity of MoE models encourages stronger monosemantic behavior at the expert level, which allows for the identification of experts responsible for particular downstream tasks and aids in preserving cross-domain performance. Building on this, we propose an expert-frozen fine-tuning method that selectively updates domain-specific experts while freezing the rest, which also alleviates catastrophic forgetting. Experiments on specialized domains, including medicine and finance, show that our approach performs on par with or better than fully fine-tuned models on in-domain tasks, while achieving relative improvements of 19.58{\%} and 37.09{\%} in retaining performance on out-of-domain benchmarks. We also demonstrate consistent advantages over PEFT baselines such as LoRA, achieving relative improvements of 11.49{\%} and 8.60{\%} in other domains."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tian-etal-2026-moe">
<titleInfo>
<title>MoE Sparsity Begets Monosemanticity: Insights into Understanding and Fine-Tuning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mingxue</namePart>
<namePart type="family">Tian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qi</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaohan</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiajun</namePart>
<namePart type="family">Chai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Guojun</namePart>
<namePart type="family">Yin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Lin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cailian</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qimin</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siyu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yisen</namePart>
<namePart type="family">Wang</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>Mixture-of-Experts (MoE) architectures enhance the scalability and efficiency of large language models (LLMs) by activating only a subset of parameters. However, the interpretability of individual experts and corresponding strategies for post-training adaptation to domain-specific tasks remain underexplored. In this work, we develop an interpretability framework for expert-level monosemanticity in MoE models, drawing inspiration from sparse autoencoders and offering new insights into the specialization patterns of domain experts. We demonstrate that the inherent sparsity of MoE models encourages stronger monosemantic behavior at the expert level, which allows for the identification of experts responsible for particular downstream tasks and aids in preserving cross-domain performance. Building on this, we propose an expert-frozen fine-tuning method that selectively updates domain-specific experts while freezing the rest, which also alleviates catastrophic forgetting. Experiments on specialized domains, including medicine and finance, show that our approach performs on par with or better than fully fine-tuned models on in-domain tasks, while achieving relative improvements of 19.58% and 37.09% in retaining performance on out-of-domain benchmarks. We also demonstrate consistent advantages over PEFT baselines such as LoRA, achieving relative improvements of 11.49% and 8.60% in other domains.</abstract>
<identifier type="citekey">tian-etal-2026-moe</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.975/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>20613</start>
<end>20630</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T MoE Sparsity Begets Monosemanticity: Insights into Understanding and Fine-Tuning
%A Tian, Mingxue
%A Zhang, Qi
%A Wang, Xiaohan
%A Chai, Jiajun
%A Yin, Guojun
%A Lin, Wei
%A Chen, Cailian
%A Xu, Qimin
%A Wang, Siyu
%A Wang, Yisen
%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-moe
%X Mixture-of-Experts (MoE) architectures enhance the scalability and efficiency of large language models (LLMs) by activating only a subset of parameters. However, the interpretability of individual experts and corresponding strategies for post-training adaptation to domain-specific tasks remain underexplored. In this work, we develop an interpretability framework for expert-level monosemanticity in MoE models, drawing inspiration from sparse autoencoders and offering new insights into the specialization patterns of domain experts. We demonstrate that the inherent sparsity of MoE models encourages stronger monosemantic behavior at the expert level, which allows for the identification of experts responsible for particular downstream tasks and aids in preserving cross-domain performance. Building on this, we propose an expert-frozen fine-tuning method that selectively updates domain-specific experts while freezing the rest, which also alleviates catastrophic forgetting. Experiments on specialized domains, including medicine and finance, show that our approach performs on par with or better than fully fine-tuned models on in-domain tasks, while achieving relative improvements of 19.58% and 37.09% in retaining performance on out-of-domain benchmarks. We also demonstrate consistent advantages over PEFT baselines such as LoRA, achieving relative improvements of 11.49% and 8.60% in other domains.
%U https://aclanthology.org/2026.findings-emnlp.975/
%P 20613-20630
Markdown (Informal)
[MoE Sparsity Begets Monosemanticity: Insights into Understanding and Fine-Tuning](https://aclanthology.org/2026.findings-emnlp.975/) (Tian et al., Findings 2026)
ACL
- Mingxue Tian, Qi Zhang, Xiaohan Wang, Jiajun Chai, Guojun Yin, Wei Lin, Cailian Chen, Qimin Xu, Siyu Wang, and Yisen Wang. 2026. MoE Sparsity Begets Monosemanticity: Insights into Understanding and Fine-Tuning. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 20613–20630, Budapest, Hungary. Association for Computational Linguistics.