Perspective Latents Drive Causal Emergence in AI

Recent studies have linked causal emergence in reinforcement learning agents to the growth of Integrated Information Decomposition (Phi_r) during training, correlating it with reward improvement. For active inference, which focuses on reward-free predictive organization, the relationship to such information-theoretic signatures has been less clear. This paper explores an active inference agent architecture that distinguishes between a fast perception latent (z) and a slow global latent (g), where 'g' is driven by prediction error and decoupled from policy gradients. In reward-free environmental regime-switching scenarios, Phi_r concentrates within this slow global latent 'g'. The study finds that the aggregate magnitude of Phi_r is largely architectural and decreases with training. However, a more detailed atom-compositional analysis reveals that learning significantly impacts the sign of decoupling (flipping from negative to positive and becoming regime-invariant) and how downward causation adjusts to environmental changes. This identifies 'g' as the architectural hub for Phi_r-relevant temporal organization, suggesting that scalar Phi_r alone doesn't fully capture learned integration.