Brain predictions use phrases, not just next words ^{CEFR B2}

Scientists examined whether human anticipatory language processing works the same way as next-word prediction in large language models (LLMs). The new study, published in Nature Neuroscience, reports that the brain predicts upcoming language using larger grammatical units, or constituents, rather than relying solely on the single most likely next word. "While LLMs are trained and optimized to predict the next word, the human brain makes predictions by grammatically grouping words into phrases," says coauthor David Poeppel.

The researchers ran several experiments with Mandarin Chinese speakers and recorded neural responses with magnetoencephalography (MEG). They complemented those recordings with behavioral Cloze tests, where readers supply missing words, and they reanalyzed additional brain data from patients exposed to English to test language generality.

• They used LLMs to quantify predictability via entropy and surprisal.
• High entropy means many possible next words; high surprisal means a word is unexpected.
• They compared model-based predictions to time-locked brain responses for the same sentences.
• Instead of a uniform match, brain correlations depended on a word’s position in grammatical phrases.

These results indicate that human prediction is modulated by grammatically organized chunks, a sensitivity that LLM next-word probabilities do not capture. The findings raise questions about how computational models should represent linguistic structure to better reflect human language processing.