Dynamical alignment: a principle for adaptive neural computation

Abstract Computational capabilities are typically viewed as fixed by static network architecture. We establish ‘dynamical alignment’: a principle demonstrating that a fixed computational substrate can be steered into fundamentally different operational modes solely by controlling its inputs’ temporal dynamics. By encoding static data into dynamical system trajectories, we use continuous-time neural networks as a transparent testbed to reveal a bimodal computational landscape. This landscape’s structure, including a controllable phase transition in efficiency, is governed by the input’s global phase space volume evolution, not local chaotic sensitivity. A ‘dissipative’ mode (contracting dynamics) enables sparse, energy-efficient processing, while an ‘expansive’ mode (expanding dynamics) unlocks high-performance states. We show this bifurcation arises mechanistically from timescale alignment between input and substrate dynamics, which dictates distinct coding strategies. This ‘dynamic software on fixed hardware’ paradigm establishes time itself as a controllable computational resource, shifting focus in AI from architectural design towards mastering adaptive computational principles.