When word order matters: human brains represent sentence meaning differently from large language models

TL;DR

Human brains represent sentence meaning sensitively to word order, whereas LLMs rely more on context than order for meaning processing, as demonstrated by fMRI and model analysis.

Problem

To determine whether humans and LLMs fundamentally differ in how they process sentence meaning, particularly regarding the impact of word order on meaning representation.

Approach

Participants were presented with normal and scrambled sentences during fMRI scanning, and brain activity patterns were analyzed. Simultaneously, internal representations of several LLMs (e.g., GPT-2, BERT) were extracted and compared with human brain responses. A metric was developed to measure the degree of change in meaning representation due to word order alterations.

Results & Contribution

Human brains showed significantly different meaning representations when word order changed, whereas LLMs were relatively less sensitive to order changes and maintained similar representations using contextual information. This suggests that human sentence processing is based on syntactic structure, while LLMs rely on distributional semantics. The study contributes to understanding differences between human and AI language processing.