Content-Sensitive Linguistic Representations in the Human Multiple-Demand Network

Neuroimaging dissociates specialized language regions from the domain-general multiple-demand (MD) network, yet the functional contribution of MD regions to language processing remains unresolved. Because MD recruitment during linguistic tasks is conventionally attributed to domain-general cognitive load, prior research has largely prioritized activation magnitude over representational content. Consequently, it remains unclear whether MD cortices function merely as nonspecific amplifiers of effort or actively encode fine-grained linguistic structures. Here we show that the MD network reliably encodes content-sensitive linguistic representations during naturalistic comprehension independently of peak cognitive demand. Using a voxelwise encoding framework with architecturally identical language models trained on distinct semantic domains (BERT and SciBERT) across three naturalistic fMRI datasets, we find that multiple-demand regions track linguistic structure during both high-demand scientific lectures and low-demand narratives. Furthermore, comparative analyses reveal that these representations are sensitive to domain-specific semantic regularities rather than surface-level form alone. Finally, by evaluating relative model alignment alongside the language network, we demonstrate that the representational balance between these two systems shifts dynamically as a function of task context and domain relevance, revealing a complementary division of labor. Together, these findings reframe the multiple-demand network from a purely extrinsic control mechanism into an active, content-sensitive component of a distributed semantic architecture.