Researchers have proposed an alternative formulation for active inference, a method used to select actions by minimizing expected free energy. The new approach, based on a Bethe free energy functional, supports message passing treatments of inference procedures. An information constraint is imposed to maintain the epistemic drive, which is essential for the agent's decision-making process. This constraint ensures that the mutual information between future observations and parameters given actions meets or exceeds the entropy of the goal prior. The researchers demonstrate that their approach can recover expected free energy solutions under certain conditions and show its performance on three tasks compared to other methods. --- Why it matters: This work matters for AI engineers because it provides a new framework for active inference, which could improve decision-making in complex environments. By supporting message passing treatments, the proposed method may enable more efficient and scalable implementations of active inference. Source: https://arxiv.org/abs/2608.17167