Expected Free Energy Optimizes Information Gathering in POMDPs.

A new theoretical equivalence has been established, showing that minimizing Expected Free Energy (EFE) in active inference is precisely the same as solving a rho-POMDP (Partially Observable Markov Decision Process) where the utility function directly represents expected information gain. This significant finding eliminates a long-standing practical challenge in rho-POMDPs: the need to manually tune the belief-dependent utility (rho) and its associated exploration weight for every new task. The research proves this equivalence for observe-then-commit POMDPs and extends it to a broader class of factored observation POMDPs, which encompass interleaved observe-act problems like non-destructive testing or mobile sensing where information gathering doesn't alter the hidden state. Crucially, the exploration weight is fixed at w=1 because the variational bound naturally expresses both pragmatic (reward-seeking) and epistemic (information-seeking) value in the same units. Experimental results across diverse environments, from the classic Tiger problem to more complex benchmarks like RockSample and a new Structural Inspection task, consistently support the theory. The untuned EFE approach matched or surpassed reward-only planning at the same horizon, avoided the over-exploration issues of manually tuned bonuses, and positioned itself optimally on the success-reward Pareto frontier. This provides a robust, out-of-the-box exploration objective for agents operating under partial observability.