Brain predictions use phrases, not just next words ^{CEFR B1}

Research published in Nature Neuroscience asked whether brain predictions work like the next-word predictions used in phones and large language models (LLMs). Coauthor David Poeppel explains that while LLMs are trained to predict the next word, the human brain makes predictions by grammatically grouping words into phrases.

The study tested Mandarin Chinese speakers and recorded brain activity with magnetoencephalography (MEG). It also used behavioral Cloze tests, where people fill missing words, and the team examined additional brain data from patients exposed to English to check cross-language consistency.

To compare brain responses and model behavior, the researchers used LLMs to compute entropy (many possible continuations) and surprisal (how unexpected a word is). If the brain worked like an LLM, correlations between brain signals and model predictions would be uniformly high. Instead, brain responses varied with a word’s position inside grammatical structure, showing sensitivity to constituents. The authors conclude human prediction is balanced and modulated by grammatically organized chunks, not just next-word probability.