Active inference paper reveals causal emergence hides in slow latent, not scalar metrics

A recent line of research has used Integrated Information Decomposition (Φ_r) to measure causal emergence in reinforcement learning agents, finding that Φ_r increases with training and correlates with reward improvement. Hongju Pae from arXiv (paper 2607.20708) extends this inquiry to active inference — a reward-free framework where agents learn predictive models of the world. The agent architecture separates a fast perception latent (z) from a slow global latent (g), where g is driven by prediction error and structurally decoupled from policy gradients. In a regime-switching environment, the author measures Φ_r at multiple scales.

The core finding is that Φ_r's aggregate magnitude is largely an artifact of architecture — it concentrates in g and actually decreases with training. The true effect of learning only becomes visible at the atom-compositional level: decoupling flips from negative to positive sign and becomes regime-invariant under environmental change, while downward causation carries the regime-dependent adjustment. This identifies g as the architectural locus of temporal organization and argues strongly against reading scalar Φ_r as a direct index of learned integration. For AI researchers, this means that popular information-theoretic measures of emergence may conflate architectural structure with genuine learning, especially in active inference systems.