causal-emergence-active-inference-perspective-latents

name	causal-emergence-active-inference-perspective-latents
title	Perspective Latents as an Architectural Condition for Causal Emergence in Active Inference Agents
description	Analysis of causal emergence in active inference agents using Integrated Information Decomposition, showing how architectural separation of fast perception and slow global latents affects Φᵣ dynamics.
arxiv_id	2607.20708
date	2026-07-22T00:00:00.000Z
authors	["Hongju Pae"]
categories	["q-bio.NC","cs.AI","cs.LG"]
trigger_words	["causal emergence","active inference","integrated information decomposition","perspective latents","Φᵣ","regime-switching","temporal organization"]

Perspective Latents as an Architectural Condition for Causal Emergence in Active Inference Agents

Overview

This paper investigates causal emergence in active inference agents through Integrated Information Decomposition (Φᵣ), contrasting with reinforcement learning findings. It examines how architectural separation of fast perception latents (z) from slow global latents (g) affects information-theoretic signatures of integration.

Key Contributions

1. Architectural Framework

• Dual Latent Architecture: Separates fast perception latent (z) from slow global latent (g)
• Structural Decoupling: Global latent g is driven by prediction error and decoupled from policy gradients
• Reward-Free Organization: Tests predictive organization without explicit reward signals

2. Causal Emergence Findings

• Φᵣ Concentration: Φᵣ concentrates in the global latent g rather than distributed across the network
• Architectural Dominance: Aggregate Φᵣ magnitude is largely determined by architecture, not learning
• Training Effect: Φᵣ actually decreases with training in reward-free environmental regimes

3. Atom-Compositional Analysis

• Decoupling Sign Flip: Learning flips decoupling sign from negative to positive
• Regime Invariance: Decoupling becomes invariant under environmental change after learning
• Downward Causation: Carries regime-dependent adjustment information

4. Theoretical Implications

• Architectural Locus: Identifies g as the architectural locus of Φᵣ-relevant temporal organization
• Scalar Φᵣ Limitations: Argues against interpreting scalar Φᵣ as direct index of learned integration
• Active Inference vs RL: Contrasts reward-free predictive organization with reward-driven RL

Applications

AI Safety and Alignment

• Emergent Behavior Monitoring: Use Φᵣ analysis to monitor emergent causal structures in AI agents
• Architectural Safety: Design agent architectures with controlled causal emergence properties
• Predictive vs Reward-Based: Understand differences between predictive and reward-based learning paradigms

Computational Neuroscience

• Hierarchical Processing: Model brain's hierarchical processing with separated fast/slow latents