Neural Policy Composition via Free Energy Minimization

A new normative framework from arXiv introduces policy composition emerging from variational free energy minimization. The approach derives a continuous-time gradient flow with guaranteed convergence, offering a principled objective for context-dependent gating. This addresses the lack of clear normative objectives in existing methods, which rely on prespecified design choices tied to specific architectures or datasets. The work provides a broadly applicable solution for composing previously acquired skills into intelligent behaviors.