Brain-LLM Alignment Depends on Training Language, Not Typology
A recent investigation published on arXiv (2605.23032) explores the extent to which the correlation between brain activity and large language models (LLMs) is consistent across different languages. The study utilized fMRI data from 112 individuals fluent in English, Chinese, and French (derived from the Le Petit Prince corpus) to examine seven LLMs, which included English-dominant, Chinese-dominant, and multilingual variants. A significant discovery indicates that the dominance of the training language influences alignment rather than being an intrinsic characteristic of English. The Chinese-focused model Baichuan2-7B, which is architecturally similar to LLaMA-2-7B, demonstrated the strongest alignment with Chinese brain activity and the weakest with English. Furthermore, alignment degradation correlates with formal typological distance, with syntax-related brain areas (IFG) exhibiting 2.3× steeper typological gradients compared to lexico-semantic regions (PTL). This research implies that the alignment between brain activity and LLMs is not universal but influenced by the training data, challenging existing beliefs about neural language processing across languages.
Key facts
- Study uses fMRI data from 112 participants across English, Chinese, and French.
- Data comes from the Le Petit Prince corpus.
- Seven LLMs tested: English-dominant, Chinese-dominant, and multilingual.
- Baichuan2-7B is a Chinese-dominant model architecture-matched to LLaMA-2-7B.
- Training-language dominance, not English, drives alignment pattern.
- Baichuan2-7B aligns best with Chinese brains and worst with English.
- Formal typological distance independently covaries with alignment degradation.
- Syntax-associated brain regions (IFG) show 2.3× steeper typological gradients than lexico-semantic regions (PTL).
Entities
Institutions
- arXiv