Sophisticated Policies from Epistemic Priors

This paper argues that the key computational advantage of Sophisticated Inference is not necessarily recursive tree search, but closed-loop control within the planning horizon: future actions must remain dependent on future states and observations. Using the stochastic Reactivity Maze, the authors compare several variational objectives with matched state-action posterior families, including Sophisticated Inference, factorized active inference, and standard Expected Free Energy planning. The abstract reports that epistemic motivation alone does not yield reliable goal-reaching, while closed-loop inference without epistemic incentives does not seek information. Both Sophisticated Inference and full-joint epistemic-prior active inference solve the benchmark.