Individual Differences in Sustained Attention Tune Semantic Abstraction in Neural Encoding of Narratives

A recent study explored how individual differences in sustained attention influence the neural encoding of semantic information within narratives. Traditional neural encoding modeling often aggregates semantic representation across participants, disregarding the rich individual variations in brain-LLM alignment. To address this, researchers analyzed fMRI data from 49 participants while they listened to Le Petit Prince, using GPT-2 embeddings from different layers to train encoding models.

High-level embeddings (from the 8th layer) demonstrated significant alignment between brain and LLM representations in 73 out of 122 brain regions, primarily in the temporal cortex and default mode network. Functional connectivity in a connectome-based model linked sustained attention performance to neural encoding in 18 ROIs, especially in the insula and networks associated with control and attention.

Notably, neural encoding of sustained attention was linked to a higher proportion of high-level gist information in frontoparietal and attention networks, while the inverse relationship was observed for low-level information. Importantly, sustained attention scores predicted enhanced neural anticipation of upcoming narrative content across various brain networks.

These results indicate that sustained attention does not uniformly enhance semantic encoding but rather redirects neural representations toward more compressed, gist-level formats that facilitate future content anticipation. The study suggests that individual differences in brain-LLM alignment may reflect distinct representational strategies rather than mere noise, with implications for cognitive phenotyping.