Modern machine learning architectures are predominantly associative, excelling at pattern recognition but lacking mechanisms for causal structure identification. We introduce a framework for autonomous structural discovery using the Free Energy Principle. By representing competing causal hypotheses as generative models, an Active Inference agent performs model selection via the minimization of Variational Free Energy (VFE). Across controlled simulation environments, the agent reliably distinguishes between additive and multiplicative causal structures using fewer than 20 adaptive queries. Compared to standard baselines (GP-BO and Random Search) under identical query budgets, the approach demonstrates significantly improved sample efficiency and robustness under observational noise. These results provide a potential foundation for autonomous scientific reasoning in sparse-data regimes.
Chouhan et al. (Thu,) studied this question.