A new paper pits three leading theories of artificial curiosity against each other in a simple foraging test, and they do not agree on what curiosity is for.
Researchers extended the Maximum Occupancy Principle, or MOP, to environments where an agent cannot fully observe its surroundings, and paired it with a reformulated version of the Expected Free Energy calculation used in Active Inference. Both approaches now factor in an agent's belief about hidden states, and the Active Inference version can be solved offline through value iteration over the full space of possible beliefs. The team then ran minimal simulations of agents foraging for food from uncertain sources, tracking how each framework decided when to explore versus when to commit. They also added Empowerment, a third intrinsic-motivation approach, as a point of comparison.
Most production AI agents today are trained to maximize one explicit reward, which works until the reward signal is wrong, sparse, or absent. This research asks what an agent does with no goal at all beyond staying curious, and the differences were stark: MOP agents switched between goal-directed feeding and scouting other food sources based on their energy and beliefs, while Active Inference agents parked themselves near a single food source and rarely left, with Empowerment behaving similarly to Active Inference.
It is a toy world with cartoon food sources, not a blueprint for a household robot, but the comparison is a useful reminder that curiosity-driven AI is not one idea, it is at least three, and they do not behave the same way under pressure.