Large language models poorly capture how word order shapes sentence meaning in the brain

Original source

Large language models poorly capture how word order shapes sentence meaning in the brain

With 7-tesla fMRI from 30 readers of 108 crafted sentences, transformer models matched brain patterns poorly when meaning depended on roles, while models built on sentence structure did better.

Evidence

Evidence: Peer-reviewed (evidence step 5 of 6)

Published after review by independent experts.

The story in three lines

Key finding

When word choice is held constant, transformer representations track brain responses to role changes poorly; structure-based models fit better.

Significance

Technical: fluent language output does not mean models represent meaning as the brain does.

Takeaway

Hybrid models that encode semantic roles may be closer to human language processing.

Later coverage in the journal

No later coverage in the journal yet.

Corrections and updates

No corrections or updates recorded for this item.