Sparse Distributed Archetypes Reveal Compressible Network Motifs Underlying Naturalistic Cognition

Naturalistic cognition emerges from coordinated interactions among distributed brain systems operating across multiple representational scales. Characterizing this organization remains challenging because cognitively relevant information is embedded within high-dimensional neural activity. Here, we apply Multisubject Archetypal Analysis (MS-AA) [1] to naturalistic fMRI data collected during intact narrative listening, word-scrambled audio, and rest to investigate how condition-relevant information is distributed across archetypal representations. We examine both spatial and temporal formulations of MS-AA as complementary views of naturalistic brain activity. Across analyses, decoding performance consistently followed the hierarchy intact > word-scrambled > rest, indicating that archetypal representations preserve meaningful condition-related structure. Top-m decoding analyses further revealed that this information is highly compressible: relatively small subsets of archetypes frequently recovered substantial fractions of full-model decoding performance. Spatial AA exhibited a stable sparse-decoding regime that persisted across representational scales. Across a broad range of matched component ratios, approximately 5-15 archetypes consistently captured disproportionate amounts of condition-relevant information. These same sparse subsets also organized subjects into condition-aligned clusters more strongly than expected from random archetype subsets, with the strongest joint decoding-clustering effects occurring repeatedly within an intermediate representational regime (K {approx}50 - 88). Network over-representation analyses revealed that informative archetypes were not isolated canonical networks but distributed mixtures of interacting systems. Across the highest-performing decoding- clustering configurations, default mode and frontoparietal systems were consistently overrepresented relative to network size, whereas visual and limbic systems were underrepresented. Together, these findings suggest that the archetypal motifs most informative for distinguishing cognitive states are sparse, distributed subnetworks enriched for higher-order association systems. More broadly, the results demonstrate that MS-AA provides a useful framework for studying the compressibility, geometry, and multiscale organization of cognitive brain states.