Computational Lesions in Multilingual Language Models Separate Shared and Language-specific Brain Alignment

Researchers used computational lesions on multilingual large language models to identify how the brain processes language across different languages. By selectively disabling parameters, they found that a shared computational core handles 60% of multilingual processing, while language-specific components fine-tune predictions for individual languages, providing new insights into how multilingual AI aligns with human neurobiology.

This research bridges neuroscience and AI by treating large language models as tools for understanding the human brain's multilingual architecture. The computational lesioning approach—systematically removing parameter groups and measuring the impact on fMRI prediction accuracy—offers a novel causal framework that neuroimaging alone cannot provide. Traditional brain imaging identifies which regions activate during language processing, but cannot determine whether those regions perform shared or specialized functions across languages. By lesioning models and observing how brain predictivity changes, researchers can reverse-engineer the functional organization of both the model and the brain.

The findings have significant implications for multilingual AI development. The discovery of a compact shared backbone supporting 60% of cross-language processing suggests that multilingual models could be more efficient than previously assumed. Rather than requiring separate processing streams for each language, models maintain a unified semantic and syntactic foundation with embedded language-specific refinements. This mirrors how human brains appear to organize multilingual competence: a core language faculty enhanced by language-specific specializations.

For the AI industry, this research validates the architectural assumptions underlying current multilingual LLMs and suggests that model compression and transfer learning between languages might be more effective than scaling separate language-specific components. The work also demonstrates how neuroscience can guide AI development by identifying optimal computational structures. As multilingual AI becomes increasingly critical for global applications, understanding the brain's solution to multilingual processing provides a blueprint for building more efficient, interpretable systems that align with human cognition rather than brute-force scaling.

• →Computational lesions revealed that 60% of multilingual processing relies on a shared core, with language-specific components handling the remaining 40%.
• →The research demonstrates a causal framework for studying brain-model alignment that goes beyond correlational neuroimaging approaches.
• →Findings suggest multilingual LLMs could be optimized for efficiency by prioritizing shared computational backbone over language-specific parameters.
• →Human neuroscience and AI development can inform each other through systematic model ablation studies comparing predictions against fMRI data.
• →The shared-core-plus-specialization architecture observed in both brains and models may represent an optimal solution for multilingual processing.