AXIOM (Active eXpanding Inference with Object-centric Models) is designed as a general-purpose framework, with its capabilities tested across ten diverse 2D arcade-style environments from the Gameworld-10k challenge. In these settings, agents learn to control characters or objects across a range of tasks. These environments are specifically designed to evaluate an agent’s general control capabilities, testing its ability to learn efficiently from limited data, sustain reasoning and planning over time, and handle delayed or sparse rewards. Our validation assessed the theoretical foundations of AXIOM, its empirical performance metrics, and comparison with DreamerV3. This evaluation considered the underlying mixture models and the Active Inference framework within the target environments, along with a review of key hyperparameters and design components. AXIOM’s efficiency stems from its integration of Variational Inference, which approximates Bayesian posteriors using human-like priors, and Active Inference, a framework where agents minimize expected free energy—or surprise—to reduce epistemic uncertainty and support pragmatic behavior. Additionally, the use of core priors within a fast structure-learning mixture-model architecture, combined with gradient-free computations and Bayesian simplification mechanisms, further enhances learning efficiency.
Read the Full Report here
