Multiple-Demand Network encoding geometry balances generalization and dimensionality during novel task assembly.

On the basis of verbal instructions, humans can accomplish novel and diverse demands at the very first try. This complex phenomenon recruits structured brain activity across the frontoparietal Multiple Demand Network (MDN), which is thought to encode upcoming task parameters and guide behavior. Nonetheless, it is still uncertain how novel instructions are translated into efficient neural task representations. To address this, we collected functional magnetic resonance imaging (fMRI) data while participants followed a rich set of novel verbal instructions. These varied along three core dimensions: the overarching task demand (to select or to integrate stimuli information), the relevant target category (animate or inanimate items), and the visual feature that participants responded to (color or shape). Multivariate pattern analysis (MVPA) was used to examine the informational content and format of MDN distributed activity. We contrasted two alternative representational geometries that may underpin novel task coding: low-dimensional spaces based on abstract and generalizable representations and high-dimensional architectures hosting context-unique, conjunctive neural codes. Our results show that anticipatory activity in the MDN was sensitive to the content of instructions. While the selection vs. integration task demands were broadly encoded within this network, coding of the relevant categories and features was restricted to lateral MDN regions, namely, the intraparietal sulcus and the inferior frontal junction. Critically, the representational spaces across the MDN displayed a mixture of geometrical motifs, partially supporting our two alternative hypotheses. On the one hand, Cross-Condition Generalization Performance revealed the presence of abstract and transferable neural codes, although only for task demand information. On the other hand, Shattering Dimensionality showed complex, high-dimensional coding spaces across the MDN, structured around both task-informative and non-informative axes. Still, no evidence of conjunctive neural codes was observed. Overall, these findings highlight that novel instructed behavior may recruit both abstraction and high dimensionality to promote generalization while still maximizing the expressivity of MDN coding spaces. More broadly, they stress the role of the encoding geometry for a computational understanding of cognitive control processes.