We have a new exciting line of work showing that the typical guarantees for univeral approximation for neural networks do not hold for recurrent systems in general. Abel’s new work shows that a subclass of dynamical systems can have guaranteed approximation by Neural ODEs with connection to temporal generalization.

Ságodi, Á. and Park, I. M. (2026). Universal Approximation Theorems for Dynamical Systems with Infinite-Time Horizon Guarantees. http://arxiv.org/abs/2602.08640