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

Neural encoding modeling for narrative comprehension -- using LLM embeddings to predict the temporal dynamics of neural activation -- is widely used to evaluate how well the brain represents certain semantic information. This approach, however, typically collapses semantic representation into a scalar fidelity score aggregated across participants, overlooking potentially rich individual differences and qualitative differences in naturalistic neural representations. In the current study, we addressed this gap by investigating how neural semantic encoding varies across individuals with sustained attention, and whether this difference is qualitative in addition to quantitative. Using fMRI data from 49 participants listening to an audiobook, Le Petit Prince, in English, we trained encoding models with GPT-2 embeddings at varying layers and forecasting lengths. Encoding models trained on high-level embeddings (from the 8th layer of GPT-2) showed significant brain-LLM alignment in 73 of 122 cortical and subcortical ROIs, concentrated in the temporal cortex and the default mode network. Functional connectivity strength in a validated connectome-based predictive model of sustained attention correlated with encoding performance in 18 ROIs, most strongly in the insula and in control and attention networks, with no ROI showing significant opposite relationship. Decomposing semantic information by levels of abstraction showed that the neural signature of sustained attention was selectively associated with a greater proportion of high-level gist information encoded in the frontoparietal and attention networks, while the opposite relationship emerged for proportion of low-level information. Sustained attention scores further predicted stronger predictive coding of upcoming narrative content across multiple networks across the whole brain (control, default mode, salience/ventral attention, limbic, and somatomotor networks). These findings suggest that sustained attention does not uniformly amplify neural semantic encoding, but instead shifts neural representations toward more compressed, gist-level formats that support anticipation of upcoming content. Individual differences in brain-LLM alignment can be a signature of representational strategy rather than noise, with potential utility for cognitive phenotyping.