We present a rigorous theoretical framework for modeling decision-making as anon-equilibrium dynamical process in open dissipative systems. The architectureexhibits a sequential two-phase structure: (1) stochastic exploration of hypothesesvia Langevin dynamics, and (2) dissipative convergence to an attractor state representing the decision. The transition between phases is governed by an informationtheoretic threshold related to posterior concentration. We establish the connectionto variational Bayesian inference via free energy minimization and analyze the thermodynamic costs of computation. Our analysis demonstrates the existence of anoptimal noise intensity that minimizes decision time, analogous to stochastic resonance. The framework is formulated as an effective theory, independent of specificmicroscopic substrates, making it applicable to diverse systems from neural populations to artificial inference engines.
Kotenko Valeriy (Tue,) studied this question.