arXiv Study Proposes Bethe Free Energy for Active Inference with Information Constraints
Researchers have introduced a Bethe free energy functional as an alternative to the expected free energy approach in active inference. Standard formulations lose their Kullback-Leibler structure when averaging over unobserved outcomes, complicating message passing. The new method imposes an information constraint alongside normalization, marginalization, and form constraints, requiring that mutual information among future observations, states, and parameters given actions at least matches the entropy of the goal prior. At a particular value of the Karush-Kuhn-Tucker multiplier, the constrained Bethe Lagrangian's stationary point equates to the expected free energy solution. The paper, listed as arXiv:2608.17167, demonstrates that as the information demand increases, the approach behaves accordingly.
Key facts
- The paper is identified as arXiv:2608.17167.
- Active inference typically minimizes an expected free energy functional.
- Adding an expectation over unobserved outcomes removes the Kullback-Leibler structure.
- The authors propose a Bethe free energy functional instead.
- The Bethe formulation fully supports inference by message passing.
- An information constraint maintains the epistemic drive.
- The constraint requires mutual information to be at least as large as the entropy of the goal prior.
- For a specific Karush-Kuhn-Tucker multiplier, the stationary point recovers the expected free energy solution.
Entities
—