Language-Specific Neurons: The Key to Multilingual Capabilities in Large Language Models

2024/02/26 by Tianyi Tang, Tang, Tianyi, Wenyang Luo +13 · 39 citations
Computer Science · #Computation and Language (cs.CL) #Computer science #Computer security #FOS: Computer and information sciences #Key (lock) #Linguistics #Natural Language Processing Techniques #Natural language processing #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2402.16438

published in arXiv (Cornell University) (Cornell University)

openalex publication_date 2024/02/26 · openalex created_date 2024/02/28 · openalex updated_date 2026/07/28

Abstract

Large language models (LLMs) demonstrate remarkable multilingual capabilities without being pre-trained on specially curated multilingual parallel corpora. It remains a challenging problem to explain the underlying mechanisms by which LLMs process multilingual texts. In this paper, we delve into the composition of Transformer architectures in LLMs to pinpoint language-specific regions. Specially, we propose a novel detection method, language activation probability entropy (LAPE), to identify language-specific neurons within LLMs. Based on LAPE, we conduct comprehensive experiments on several representative LLMs, such as LLaMA-2, BLOOM, and Mistral. Our findings indicate that LLMs' proficiency in processing a particular language is predominantly due to a small subset of neurons, primarily situated in the models' top and bottom layers. Furthermore, we showcase the feasibility to "steer" the output language of LLMs by selectively activating or deactivating language-specific neurons. Our research provides important evidence to the understanding and exploration of the multilingual capabilities of LLMs.

Cited by

• Breaking Minds, Breaking Systems: Jailbreaking Large Language Models via Human-like Psychological Manipulation
• CLINIC: Evaluating Multilingual Trustworthiness in Language Models for Healthcare
• Minimal neuron ablation triggers catastrophic collapse in the language core of Large Vision-Language Models
• Where Culture Fades: Revealing the Cultural Gap in Text-to-Image Generation
• Quantification and object perception in Multimodal Large Language Models and human linguistic cognition
• POWSM: A Phonetic Open Whisper-Style Speech Foundation Model
• Transcribe, Translate, or Transliterate: An Investigation of Intermediate Representations in Spoken Language Models
• Hierarchical Frequency Tagging Probe (HFTP): A Unified Approach to Investigate Syntactic Structure Representations in Large Language Models and the Human Brain
• Sparse Subnetwork Enhancement for Underrepresented Languages in Large Language Models
• SASER: Stego attacks on open-source LLMs
• The Achilles' Heel of LLMs: How Altering a Handful of Neurons Can Cripple Language Abilities
• Neuron-Level Analysis of Cultural Understanding in Large Language Models
• Multilingual Routing in Mixture-of-Experts
• Understanding Post-Training Structural Changes in Large Language Models
• Linguistic Neuron Overlap Patterns to Facilitate Cross-lingual Transfer on Low-resource Languages
• What if I ask in alia lingua? Measuring Functional Similarity Across Languages
• RepreGuard: Detecting LLM-Generated Text by Revealing Hidden Representation Patterns
• Isolating Culture Neurons in Multilingual Large Language Models
• Language Arithmetics: Towards Systematic Language Neuron Identification and Manipulation
• Unveiling the Influence of Amplifying Language-Specific Neurons
• What Language(s) Does Aya-23 Think In? How Multilinguality Affects Internal Language Representations
• Detecting and Pruning Prominent but Detrimental Neurons in Large Language Models
• AudioLens: A Closer Look at Auditory Attribute Perception of Large Audio-Language Models
• Pruning General Large Language Models into Customized Expert Models
• How Does Alignment Enhance LLMs' Multilingual Capabilities? A Language Neurons Perspective
• Unveiling Instruction-Specific Neurons & Experts: An Analytical Framework for LLM's Instruction-Following Capabilities
• Paths Not Taken: Understanding and Mending the Multilingual Factual Recall Pipeline
• ExPLAIND: Unifying Model, Data, and Training Attribution to Study Model Behavior
• How Syntax Specialization Emerges in Language Models
• Understanding How Value Neurons Shape the Generation of Specified Values in LLMs
• Semantic Pivots Enable Cross-Lingual Transfer in Large Language Models
• When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual Reasoners
• Pierce the Mists, Greet the Sky: Decipher Knowledge Overshadowing via Knowledge Circuit Analysis
• Language-Specific Latent Process Hinders Cross-Lingual Performance
• Unveiling Knowledge Utilization Mechanisms in LLM-based Retrieval-Augmented Generation
• LangFIR: Discovering Sparse Language-Specific Features from Monolingual Data for Language Steering
• CLaS-Bench: A Cross-Lingual Alignment and Steering Benchmark
• GEB-Bench: Abstract Structures Told in Many Voices
• Localized Cultural Knowledge is Conserved and Controllable in Large Language Models

Related

• Language Arithmetics: Towards Systematic Language Neuron Identification and Manipulation
• How do Large Language Models Handle Multilingualism?
• Language-specific Neurons Do Not Facilitate Cross-Lingual Transfer
• Converging to a Lingua Franca: Evolution of Linguistic Regions and Semantics Alignment in Multilingual Large Language Models
• The Emergence of Abstract Thought in Large Language Models Beyond Any Language