Distribution-aware Language Neuron Identification in Multilingual Large Language Models 多语言大语言模型中的分布感知语言神经元识别

TL;DR

• Existing methods for identifying language-specific neurons in mLLMs rely on entropy of binary activation probabilities, which fails to capture the distributional and mutually related nature of multilingual representations
• The authors propose Distribution-aware Language Neuron selection, which analyzes pairwise relationships between per-language activation distributions across the full activation range (including negative values)
• Language specificity is quantified by clustering languages using pairwise overlap coefficients between their activation distributions
• Evaluated across two mLLMs and two held-out corpora, the new method yields up to 4.9× higher on-target language damage per neuron while preserving off-target language performance
• The approach more effectively isolates language-specific causal effects compared to prior entropy-based methods

Why It Matters

This work addresses a critical gap in mechanistic interpretability for multilingual models, where understanding how languages are represented and separated within a shared model is essential for both safety and efficiency. For practitioners building or fine-tuning mLLMs, being able to precisely identify and manipulate language-specific neurons enables more targeted interventions, better multilingual control, and reduced risk of unintended cross-lingual interference during model editing or pruning.

Technical Details

• Problem with prior methods: Previous approaches define a neuron as "active" when its activation value is positive (binary thresholding) and measure language specificity using entropy of language-wise activation probabilities, ignoring the continuous and overlapping nature of multilingual representations
• Proposed method: Distribution-aware Language Neuron selection analyzes the full activation distribution (including negative values) for each language, computing pairwise overlap coefficients between per-language activation distributions to quantify how distinct or shared each neuron's representation is across languages
• Clustering-based specificity quantification: Languages are clustered based on pairwise overlap coefficients, allowing the method to capture mutual relationships between languages rather than treating them as independent categories
• Evaluation: Tested across two multilingual LLMs and two held-out corpora, with causal intervention experiments measuring on-target language damage and off-target language preservation
• Key result: The proposed identifier achieves up to 4.9× higher on-target language damage per neuron while maintaining off-target language performance, demonstrating superior isolation of language-specific causal effects

Industry Insight

• The finding that binary activation thresholding significantly underestimates language specificity suggests that interpretability tooling for multilingual models should adopt distribution-aware methods rather than relying on simple activation binarization
• As multilingual models become more prevalent in production, the ability to precisely target language-specific neurons enables more efficient multilingual fine-tuning, safer model editing, and better cross-lingual transfer control without degrading unrelated language capabilities
• The 4.9× improvement in causal isolation demonstrates that subtle methodological shifts in interpretability can yield substantial practical gains, encouraging investment in more nuanced analysis techniques for multilingual representation understanding

摘要

现有识别多语言大模型（mLLMs）中语言特异性神经元的方法依赖于二元激活概率的熵，无法捕捉多语言表征的分布特性和相互关联性。
作者提出了分布感知语言神经元选择方法，该方法分析全激活范围（包括负值）内各语言激活分布之间的成对关系。
语言特异性通过聚类语言来量化，使用其激活分布之间的成对重叠系数。
在两个 mLLM 和两个保留语料库上的评估表明，新方法在保持非目标语言性能的同时，每个神经元的目标语言损伤率最高提升 4.9 倍。
与之前的基于熵的方法相比，该方法能更有效地隔离语言特异性的因果效应。

深度分析

一句话总结

• 现有识别多语言大模型中语言特异性神经元的方法依赖于二元激活概率的熵，无法捕捉多语言表征的分布特性和相互关联性
• 作者提出了分布感知语言神经元选择方法，该方法分析全激活范围（包括负值）内各语言激活分布之间的成对关系
• 语言特异性通过聚类语言来量化，使用其激活分布之间的成对重叠系数
• 在两个 mLLM 和两个保留语料库上的评估表明，新方法在保持非目标语言性能的同时，每个神经元的目标语言损伤率最高提升 4.9 倍
• 与之前的基于熵的方法相比，该方法能更有效地隔离语言特异性的因果效应

为何重要

这项工作填补了多语言模型机械可解释性领域的一个关键空白，理解语言如何在共享模型中被表征和分离，对于安全性和效率都至关重要。对于构建或微调多语言大模型的研究者而言，能够精确识别和操作语言特异性神经元，可以实现更有针对性的干预、更好的多语言控制，并在模型编辑或剪枝过程中降低意外跨语言干扰的风险。

技术细节

• 现有方法的缺陷：之前的方法在激活值为正时将神经元定义为"激活"（二元阈值化），并使用熵来衡量语言特异性