Emergence of a temporal processing gradient from naturalistic inputs and network connectivity

Natural language unfolds over multiple nested timescales: Words form sentences, sentences form paragraphs, and paragraphs build into full narratives. Correspondingly, the brain exhibits a hierarchy of processing timescales, spanning from lower- to higher-order regions. During narrative comprehension, neural activation patterns have been shown to propagate along this cortical hierarchy with increasing temporal delays (lags). To investigate the mechanisms underlying this lag gradient, we systematically manipulate the structure of a recurrent reservoir network. In the biologically inspired “Limited-Canal” configuration, word embeddings are received by a limited set of sensory neurons and transmitted through a series of local connections to the distal end of the network. This configuration endows the network with an intrinsic lag gradient, inducing a cascade of activity as information propagates along the network. We found that, similar to the human brain, this intrinsic lag gradient is enhanced by naturalistic narratives. The interaction between naturalistic input and network structure becomes evident when manipulating local connectivity through the “canal width” parameter, which determines how closely the Limited-Canal model mirrors the human brain’s sensitivity to narrative structure. In addition, we found that processing cost, as a computational proxy for the BOLD signal, increases more slowly in later neurons, which can account for the emergence of the lag gradient. Our results demonstrate that narrative-driven neural dynamics can emerge from macroscale anatomical topology alone without task-specific training. These fundamental topological properties of the human cortex may have evolved to effectively process the hierarchical structures ubiquitous in the natural environment.