Training Data Drives Brain-LLM Alignment Across Languages
An arXiv preprint by Dongxin Guo et al., posted 21 May 2026, reports that brain-LLM alignment was measured using fMRI data from 112 participants across English, Chinese, and French reading the Le Petit Prince corpus (arXiv:2605.23032). The study evaluated seven LLMs spanning English-dominant, Chinese-dominant, and multilingual models and finds that training-language dominance, rather than typology alone, primarily predicts cross-linguistic alignment patterns, according to the preprint. The paper reports that a Chinese-dominant model, Baichuan2-7B, architecture-matched to LLaMA-2-7B, reverses the alignment gradient by aligning best with Chinese brains and worst with English; it also reports that typological distance covaries with alignment degradation, that syntax-associated regions (IFG) show a 2.3x steeper typological gradient than lexico-semantic regions (PTL), and that tokenization fertility accounts for roughly 60% of a cross-linguistic shift in optimal encoding layer. Editorial analysis: The results suggest that previously reported "English advantage" in brain-LLM studies may reflect training-data composition rather than an inherent language-specific alignment property, with implications for multilingual evaluation and cognitive interpretation.
What happened
The paper titled Brain-LLM Alignment Tracks Training Data, Not Typology was posted to arXiv on 21 May 2026 and archived on OpenReview for CoNLL 2026 (arXiv:2605.23032; OpenReview entry published 18 May 2026). The authors report analyses using fMRI recordings from 112 participants across English, Chinese, and French who read aligned texts from the Le Petit Prince corpus. The preprint evaluates seven LLMs described as English-dominant, Chinese-dominant, or multilingual and assesses how model representations map to the human language network.
Technical details
Per the arXiv preprint, the authors compare architecture-matched models to isolate the effect of training-language dominance; the paper highlights Baichuan2-7B (Chinese-dominant) versus LLaMA-2-7B (English-dominant) as a key contrast that reverses the alignment gradient, aligning best with Chinese brains and worst with English. The study reports that formal typological distance independently covaries with alignment degradation, that syntax-associated brain regions (IFG) exhibit a 2.3x steeper typological gradient than lexico-semantic regions (PTL), and that tokenization fertility statistically accounts for about 60% of a cross-linguistic shift in the optimal encoding layer, according to the reported analyses.
Context and significance
Editorial analysis: The paper reframes the common observation of an "English advantage" in brain-LLM alignment as largely driven by training-data composition rather than language-intrinsic alignment properties. For researchers comparing brain and model representations across languages, the finding underscores the need to control for training-language dominance, tokenizer design, and corpus composition when interpreting alignment metrics. The reported concentration of residual typological effects in syntactic regions points to a plausible dissociation between syntax and lexico-semantic representation in cross-linguistic alignment studies.
What to watch
For practitioners: replication on more typologically diverse language sets and larger models, systematic variation of tokenizer schemes, and analyses that decouple architecture from training-corpus composition will clarify how general these patterns are. Observers should also track whether future encoding studies confirm the reported 2.3x IFG/PTL gradient difference and the approximately 60% contribution of tokenization fertility to layer-shift effects.
Scoring Rationale
The paper reports a clear, reproducible effect linking model training-language dominance to brain alignment, which matters for cognitive modeling and multilingual evaluation. It is notable but specialized to brain-model comparison and requires replication across more languages and model scales.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
