TLDR: A new hypothesis, “Transfer Neurons,” explains how multilingual LLMs process different languages. Specific neurons in the model’s MLP module act as bridges, converting language-specific inputs into a shared semantic space for reasoning (Type-1 neurons) and then back into language-specific outputs (Type-2 neurons). Empirical evidence shows these neurons are crucial for language transformation and reasoning, with their specificity varying by linguistic distance. Deactivating them severely impairs multilingual understanding and generation.
Multilingual Large Language Models (LLMs) have transformed how we interact with AI, enabling them to understand and generate text in various languages. A common understanding of how these models process multilingual inputs suggests a three-stage framework: initial layers convert diverse inputs into a more universal, often English-centric, representation; middle layers then perform the core reasoning; and finally, the output layers translate these processed representations back into the specific language requested. However, the precise internal mechanisms driving these crucial language transformations have remained somewhat mysterious.
Unveiling the “Transfer Neurons Hypothesis”
A recent study by Hinata Tezuka and Naoya Inoue introduces and validates a compelling explanation for these internal dynamics: the “Transfer Neurons Hypothesis.” This hypothesis proposes that specific neurons within the LLM’s MultiLayer Perceptron (MLP) module are directly responsible for facilitating the movement of information between language-specific latent spaces and a shared, universal semantic latent space. Essentially, these “transfer neurons” act as bridges, enabling the model to switch between understanding a concept in a specific language and grasping its universal meaning.
The researchers categorize these critical neurons into two types:
- Type-1 Transfer Neurons: Located in the initial layers of the LLM, these neurons are tasked with converting input representations from their original language-specific latent space into the shared semantic latent space.
- Type-2 Transfer Neurons: Found predominantly in the final layers, these neurons perform the reverse function, transforming the reasoned representations from the shared semantic latent space back into the appropriate language-specific latent space for generating the final output.
Mapping the Model’s Multilingual Mind
To support their hypothesis, Tezuka and Inoue conducted extensive empirical investigations. They observed that in the early layers of LLMs, each language forms its own distinct “latent space” – a region where its unique linguistic features are processed. As the input moves through the model, these language-specific spaces gradually converge into a single, shared semantic latent space in the middle layers, where the model performs its core reasoning, largely independent of the original input language. Finally, in the output layers, these representations diverge again, re-establishing language-specific latent spaces to generate responses in the target language.
Interestingly, the study found that languages linguistically closer to English (like Dutch and Italian) maintain a closer proximity to the English-centric shared latent space even in initial and final layers, compared to more distant languages like Japanese and Korean. This suggests that the degree of “transfer” required varies based on linguistic distance.
The Critical Role of Transfer Neurons
The researchers developed a scoring methodology to identify these transfer neurons and then performed intervention experiments by deactivating them. The results were striking:
- Deactivating a small fraction (0.2%) of Type-1 neurons significantly impaired the model’s ability to map input representations to the correct positions within the shared semantic latent space. This meant the model struggled to understand the meaning of parallel sentences across languages, indicating that Type-1 neurons are crucial for the initial language-to-meaning translation.
- Deactivating Type-2 neurons in the final layers severely inhibited the model’s ability to transform reasoned representations back into language-specific outputs. In some cases, the model would generate coherent text, but in a different language than the input, highlighting the role of Type-2 neurons in directing the output language.
Further analysis revealed insights into the nature of these neurons. While most Type-1 neurons showed less language specificity, Japanese Type-1 neurons exhibited stronger language-specific correlations, likely due to the greater linguistic distance Japanese has from English. Conversely, Type-2 neurons were generally more language-specific, as they are responsible for directing output to a particular language. The study also found evidence of “language-family specificity,” where linguistically similar languages shared more overlapping transfer neurons.
Also Read:
- Unveiling the Image-to-Text Information Pathways in Large Vision-Language Models
- Tracking the Internal Development of Language Models
Implications for Reasoning and Future LLMs
The study demonstrated the critical importance of transfer neurons for reasoning. When Type-1 neurons were deactivated, the model’s performance on multilingual knowledge QA tasks (like MKQA and MMLU-ProX) significantly degraded, even with only a tiny fraction of neurons affected. This suggests that these neurons are essential for the model to correctly align inputs with the shared semantic space necessary for effective reasoning.
This groundbreaking research not only deepens our understanding of how multilingual LLMs function internally but also opens new avenues for improving them. The authors suggest that selectively fine-tuning these transfer neurons could enhance multilingual capabilities, especially for low-resource languages, by better aligning their hidden states with the English-centric reasoning space. For more details, you can read the full research paper here.
