Research

Paper

AI LLM March 05, 2026

NeuronMoE: Neuron-Guided Mixture-of-Experts for Efficient Multilingual LLM Extension

Authors

Rongzhi Li, Hitomi Yanaka

Abstract

Extending large language models to low-resource languages is essential for global accessibility, but training separate models per language is prohibitively expensive. Mixture-of-Experts (MoE) architectures address this by adding sparse language-specific parameters, but determining how many experts each layer needs remains an open question. Current approaches allocate experts based on layer-level similarity, yet language processing exhibits fine-grained specialization at individual neurons. We propose $\textbf{NeuronMoE}$, a method that analyzes language-specific neurons across all transformer components to guide expert allocation per layer based on empirically measured cross-lingual neuron diversity. Applied to Llama-3.2-3B for low-resource languages (Greek, Turkish, and Hungarian), this approach achieves approximately 40% average parameter reduction while matching the performance of the LayerMoE baseline. We find that low-resource language experts independently develop neuron specialization patterns mirroring the high-resource language, which are concentrated in early and late layers. This reveals potential universal architectural principles in how multilingual models organize linguistic knowledge.

Metadata

arXiv ID: 2603.05046
Provider: ARXIV
Primary Category: cs.CL
Published: 2026-03-05
Fetched: 2026-03-06 14:20

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.05046v1</id>\n    <title>NeuronMoE: Neuron-Guided Mixture-of-Experts for Efficient Multilingual LLM Extension</title>\n    <updated>2026-03-05T10:55:26Z</updated>\n    <link href='https://arxiv.org/abs/2603.05046v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.05046v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Extending large language models to low-resource languages is essential for global accessibility, but training separate models per language is prohibitively expensive. Mixture-of-Experts (MoE) architectures address this by adding sparse language-specific parameters, but determining how many experts each layer needs remains an open question. Current approaches allocate experts based on layer-level similarity, yet language processing exhibits fine-grained specialization at individual neurons. We propose $\\textbf{NeuronMoE}$, a method that analyzes language-specific neurons across all transformer components to guide expert allocation per layer based on empirically measured cross-lingual neuron diversity. Applied to Llama-3.2-3B for low-resource languages (Greek, Turkish, and Hungarian), this approach achieves approximately 40% average parameter reduction while matching the performance of the LayerMoE baseline. We find that low-resource language experts independently develop neuron specialization patterns mirroring the high-resource language, which are concentrated in early and late layers. This reveals potential universal architectural principles in how multilingual models organize linguistic knowledge.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <published>2026-03-05T10:55:26Z</published>\n    <arxiv:primary_category term='cs.CL'/>\n    <author>\n      <name>Rongzhi Li</name>\n    </author>\n    <author>\n      <name>Hitomi Yanaka</name>\n    </author>\n  </entry>"
}