Introduces a novel Active Inference architecture that accelerates learning in low-data environments by embedding object-centric priors and modeling sparse object-object interactions. Unlike traditional Active Inference models designed for single domains, AXIOM (Active eXpanding Inference with Object-centric Models) generalizes across tasks by learning from individual events and refining its structure through Bayesian model growth and reduction. It achieves human-level data efficiency, mastering tasks within 10,000 interactions, while avoiding the high data demands and poor generalization typical of deep reinforcement learning.