Brain-LLM Alignment Tracks Training Data, Not Typology

In three linesBrain-LLM alignment depends on training language dominance, not inherent English properties. Test on 112 participants (English, Chinese, French) with 7 LLMs: a Chinese-dominant model (Baichuan2-7B) reverses alignment gradient. Typological distance and tokenization fertility explain remaining variation.

Summary generated by Claude — human-verified