Brain-LLM alignment tracks training data, not language typology

A new study from researchers Dongxin Guo, Jikun Wu, and Siu Ming Yiu, accepted to CoNLL 2026, investigates why large language models (LLMs) align with human brain activity differently across languages. Using fMRI data from 112 participants reading the same narrative in English, Chinese, and French (the Le Petit Prince corpus), they tested seven LLMs ranging from English-dominant (like LLaMA-2-7B) to Chinese-dominant (Baichuan2-7B) and multilingual models. The central finding: the apparent 'English advantage' in brain-LLM alignment is an artifact of training data composition. Baichuan2-7B, architecture-matched to LLaMA-2-7B but trained predominantly on Chinese, completely reversed the alignment gradient—best matching Chinese brains and worst with English. This proves that training-language dominance, not any inherent property of English, dictates the pattern.

Beyond training dominance, the study reveals that formal typological distance between languages independently correlates with degradation in neural alignment. Critically, this effect is not uniform: syntax-associated regions in the inferior frontal gyrus (IFG) show a 2.3× steeper drop-off with typological distance compared to lexico-semantic regions (PTL). Additionally, tokenization fertility—how many tokens a model uses to represent the same linguistic content—accounts for roughly 60% of the cross-linguistic shift in which layer of the LLM best predicts brain activity. These results challenge the assumption that English-centric LLMs are universally optimal for modeling human language processing, and suggest that multilingual models must account for both training data composition and typological structure to achieve broad neural alignment.