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

TL;DR

NeuronMoE introduces a neuron-guided approach to efficiently extend multilingual large language models by optimizing expert allocation based on neuron diversity, reducing parameters while maintaining performance.

Contribution

The paper presents NeuronMoE, a novel method that analyzes neuron-level specialization to guide expert allocation in multilingual models, improving efficiency over layer-based methods.

Findings

Achieves ~40% parameter reduction with maintained performance.

Low-resource languages develop neuron specialization patterns similar to high-resource languages.

Specialization is concentrated in early and late layers, indicating universal principles.

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.