Event-Centric Active Inference: Agents That Form Their Own Intermediate Goals

Agents that pursue distant outcomes need intermediate goals, yet in most systems these are either hand-designed or learned as opaque latent targets. We introduce event-centric active inference (EC-AIF), in which an agent forms its own vocabulary of intermediate goals from experience and chooses among them by expected free energy (EFE). An event is defined by its effect: the change a stretch of behaviour caused, outcome included. The vocabulary of events is formed online by minimising variational free energy, with Bayesian model reduction merging redundant terms. Each event carries an earned precision, the rate at which it comes true when pursued times the evidence that achieving it is followed by better outcomes, so it is zero until earned. An event layer chooses the next event in the current context and hands it down to the agent that acts. On bsuite catch, EC-AIF raises a standard discrete active-inference agent from 0.09 to 0.78. As a goal-former for a deep reinforcement learner (PPO on Craftax-Classic), it improves the Crafter score on 20 of 20 fresh pre-registered seeds (+1.33, 95% CI [1.00, 1.70]), and registered controls indicate that the gain comes from the events themselves, not from the added reward or from novelty. This report introduces the idea and these first results; the work is ongoing, and a full account will follow.