The ability to flexibly compose previously acquired skills to execute intelligent behaviors is a hallmark of natural intelligence. This capability is commonly attributed to gating mechanisms that regulate how multiple policies, or primitives, are weighted and composed. In this talk, a principled and theoretically-grounded model in which gating rules emerge from the minimization of free energy is proposed. From this, a continuous-time dynamical system that provably converges to the optimal solution is derived, together with a neural implementation as a soft-competitive recurrent neural circuit. The model is evaluated on collective behavior and human decision-making benchmarks, reproducing key behavioral signatures and providing insights into the data.
January 26, 2026
