Robust Belief Sharing in Federated Active Inference: A Recovery-Tested Generalized-Variational Framework for Categorical Contamination-Aware Consensus

Multi-agent active inference gives a natural account of belief sharing: agents hold local posteriors over a shared latent state, communicate those beliefs, and pool them into a colony-level consensus. The same mechanism is fragile when a member is miscalibrated, corrupted, or strategically wrong. Because the standard pool multiplies the reports together, a single confident-but-wrong broadcast that puts near-zero mass on the true state can pull the whole consensus off it, outweighing many honest members. The colony therefore needs a way to preserve the useful structure of belief sharing while limiting the influence of contaminated beliefs. This paper presents Active Fedference, a discrete-categorical framework that connects robust federated generalized variational inference with active inference belief sharing. The main bridge is structural: standard belief sharing appears as the non-robu