What Type of Inference is Active Inference?

2026-06-03

AI Summary

This work clarifies the relationship between expected free energy (EFE) and variational free energy (VFE) in active inference by reframing EFE minimization as VFE minimization over a generative model endowed with cognitive priors. By explicitly decomposing EFE into an entropy-correction term and a planning-correction term, the authors uniquely deconstruct EFE-based planning into three components: the VFE of the predictive model, an observation-side cognitive correction, and a planning correction essential for policy optimization—thereby fully characterizing its variational structure. Experiments across three grid-world environments demonstrate that when observations are unambiguous, the planning correction alone suffices for effective behavior; however, under perceptual ambiguity, incorporating the cognitive correction significantly enhances performance. This framework unifies the inferential mechanisms underlying goal-directed and information-seeking behaviors and precisely delineates the corrective terms required by distinct planning paradigms.

Abstract

Active inference casts decision-making as inference, with the Expected Free Energy (EFE) unifying goal-directed and information-seeking behavior. Recent work showed that EFE minimization can be written as Variational Free Energy (VFE) minimization on a generative model augmented with epistemic priors. We prove that the VFE of the augmented model can be rewritten as the VFE of the predictive model plus explicit entropy-correction terms, making the EFE contribution transparent. We then show that proper EFE-based planning requires combining these epistemic corrections with a planning correction that turns marginal inference into policy optimization, yielding a full variational characterization of EFE-based planning. This clarifies which corrections are needed for cross-entropy planning and for full EFE-based planning. The same entropy-corrected formulation leads to a detailed message-passing scheme for EFE-based planning together with simpler ablations. Experiments on three grid-world environments show that the planning correction already helps when observations are decisive, whereas the additional observation-side epistemic corrections matter most when observations are merely suggestive.

Problem

• Active Inference
• Expected Free Energy
• Variational Free Energy
• Epistemic Priors
• Policy Optimization

Innovation

• Active Inference
• Expected Free Energy
• Variational Free Energy
• Epistemic Priors
• Policy Optimization