Compositional recombination is facilitated by a distributed cortico-cerebellar network

Human cognition depends on the ability to flexibly recombine existing knowledge in new ways. Although this capacity for compositionality has traditionally been attributed to cortical networks, its broader neural basis remains unclear. We combine dimensionality reduction of task-based fMRI with recurrent neural network modeling to dissociate two processes underlying compositional cognition: specialized components; and the more general process of recombination. Across 87 participants performing a well-established compositional task, domain-specific cortical and anterior cerebellar regions support component processes, whereas a low-dimensional, highly integrated cortico-cerebellar network engages recombination processes consistently across diverse contexts. Recurrent neural networks trained to perform multiple cognitive tasks present similar functional signatures to the fMRI data, suggesting that low-dimensional recombination is a general solution for flexible compositional cognition. Our findings revise existing models of compositional cognition, highlighting interactions between component and recombination processes with cortico-cerebellar interactions serving as a mechanism for flexible, integrative task generalization.