Challenge: Despite the impressive multilingual capabilities demonstrated by LLMs, the understanding of how these abilities develop and function remains nascent.
Approach: They propose a novel detection method to pinpoint language-specific neurons within LLMs by selectively activating or deactivating these neurons.
Outcome: The proposed method can “steer” the output language of LLMs by selectively activating or deactivating language-specific neurons.

Similar Papers

Converging to a Lingua Franca: Evolution of Linguistic Regions and Semantics Alignment in Multilingual Large Language Models (2025.coling-main)

Copied to clipboard

Challenge: Recent studies suggest that large language models can transfer skills learned in one language to others, but internal mechanisms behind this ability remain unclear.
Approach: They find that LLMs map semantically identical inputs from different languages into a common semantic latent space that allows for consistent processing across languages.
Outcome: The findings highlight the structural evolution of multilingual models during training and scaling up.
On the Multilingual Ability of Decoder-based Pre-trained Language Models: Finding and Controlling Language-Specific Neurons (2024.naacl-long)

Copied to clipboard

Challenge: Existing decoder-based pre-trained language models demonstrate excellent multilingual capabilities, but it is unclear how they handle multilingualism.
Approach: They propose to examine the neuron-level internal behavior of decoder-based PLMs by finding neurons that fire “uniquely for each language” within decoded PLM models.
Outcome: The proposed models fire “uniquely for each language” and show that language-specific neurons are unique, with a slight overlap (5%) between languages.
Revealing the Parallel Multilingual Learning within Large Language Models (2024.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) can handle multilingual and cross-lingual text within a single input; however, previous studies focusing on using English as the pivot language to enhance language understanding and reasoning focus on using multiple languages.
Approach: They propose to use parallel multilingual input to enhance the model's comprehension of the input and to examine how multilingual processing affects prediction.
Outcome: The proposed model can handle multilingual and cross-lingual text within a single input, but previous studies focused on using English as the pivot language to enhance language understanding and reasoning.
Does Large Language Model Contain Task-Specific Neurons? (2024.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated remarkable capabilities in comprehensively handling various types of natural language processing (NLP) tasks.
Approach: They propose a method for task-specific neuron localization based on Causal Gradient Variation with Special Tokens (CGVST) this method identifies task- specific neurons by concentrating on the most significant tokens during task processing, eliminating redundant tokens and minimizing interference from non-essential neurons.
Outcome: The proposed method can locate task-specific neurons across eight public tasks.
From Neurons to Semantics: Evaluating Cross-Linguistic Alignment Capabilities of Large Language Models via Neurons Alignment (2025.acl-long)

Copied to clipboard

Challenge: Existing alignment benchmarks focus on sentence embeddings, but prior research has shown that neural models tend to induce a non-smooth representation space, which impact of semantic alignment evaluation on low-resource languages.
Approach: They propose a novel cross-lingual alignment evaluation method based on the consistency of parallel sentences to assess model alignment.
Outcome: The proposed method achieves a correlation of 0.9556 with downstream tasks performance and 0.8524 with transferability even with a small dataset.
The Transfer Neurons Hypothesis: An Underlying Mechanism for Language Latent Space Transitions in Multilingual LLMs (2025.emnlp-main)

Copied to clipboard

Challenge: Recent studies suggest a processing framework for multilingual inputs in decoder-based LLMs.
Approach: They propose a framework for multilingual inputs in decoder-based LLMs that enables transfer of representations between latent spaces and shared semantic latent space.
Outcome: The proposed framework is validated by a new study on transfer neurons in multilingual LLMs.
How Programming Concepts and Neurons Are Shared in Code Language Models (2025.findings-acl)

Copied to clipboard

Challenge: Several studies have focused on programming languages in a monolingual setting, but most focus on programming language models.
Approach: They perform a few-shot translation task on 21 PL pairs using two Llama-based models and decode the embeddings of intermediate layers.
Outcome: The proposed model assigns high probability to English tokens in the second half of the intermediate layers and language-specific neurons are concentrated in the bottom layers . the model's concept space is closer to English (including PL keywords) and the model is more efficient at identifying language-related neurons.
The LLM Language Network: A Neuroscientific Approach for Identifying Causally Task-Relevant Units (2025.naacl-long)

Copied to clipboard

Challenge: Recent advances in large language models (LLMs) have revealed their potential to perform far more than language processing tasks, showcasing abilities in reasoning and problem-solving.
Approach: They identify language-selective units within 18 popular LLMs using the same localization approach that is used in neuroscience.
Outcome: The proposed method shows that language-selective units are more aligned to brain recordings from the human language system than random units.
How Vocabulary Sharing Facilitates Multilingualism in LLaMA? (2024.findings-acl)

Copied to clipboard

Challenge: Large Language Models (LLMs) show strong performance on English tasks, but their performance in other languages is limited.
Approach: They conducted an exhaustive analysis of the multilingual capability of LLMs by examining the performance gap before and after embedding fine-tuning across 101 languages.
Outcome: The proposed model improves on the attributes of four quadrants in the model and provides actionable and efficient guidelines for tuning these languages.
Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders (2025.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) exhibit impressive abilities in various domains such as text generation, instruction following, and reasoning.
Approach: They propose a method to decompose the activations of Large Language Models into a sparse linear combination of SAE features.
Outcome: The proposed method shows that some features are strongly related to specific languages, while others are unaffected by ablating them.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations