Brain predictions use phrases, not just next words ^{CEFR A2}

Scientists asked whether the human brain predicts words in the same simple way as phones and LLMs. They ran experiments with Mandarin speakers and recorded brain activity with magnetoencephalography (MEG). Participants also did Cloze tests, where words are removed and people fill the blanks. The team analyzed additional brain data from patients listening to English sentences to check if results hold across languages.

The researchers used LLMs to measure how predictable a word is, using two ideas called entropy and surprisal. High entropy means many possible next words and high surprisal means a word is unexpected. The brain reacted differently depending on a word’s place in grammatical groups. The authors conclude the brain predicts by grouping words into phrases, not only by next-word probability, and LLMs do not show the same sensitivity.