causal-emergence

This paper investigates the architectural locus of causal emergence in active inference agents, finding that a slow global latent variable (g) serves as the primary site for integrated information (Φr) and that scalar Φr should not be interpreted as a direct measure of learned integration.