Research
Brain scans and language models keep converging on the same representations
A sweeping neuroimaging study finds the internal geometry of large language models predicts human brain activity during reading better than any prior model of language.
By Priya Sharma, Research Editor — OXFORD
OXFORD — The internal representations of large language models predict human brain activity during natural reading with record accuracy, according to the largest neuroimaging study of its kind, published Thursday — deepening a convergence between artificial and biological language processing that neither field fully expected.
Across hundreds of participants and dozens of models, one regularity held: the better a model predicts the next word, the better its internal states predict cortical responses — up to a ceiling that the newest reasoning models are the first to approach.
The authors are careful about the word 'understanding'. What the data shows, they write, is shared geometry — a common shape to how meaning is organised — which is either a coincidence of scale or a clue about the nature of language itself.
Enable JavaScript to read the full story on Neural Daily News.