Study targets topological OOD generalization in DSR

A NeurIPS 2026 paper studies topological out-of-domain generalization in dynamical systems reconstruction and time series forecasting, where regimes shift from cyclic to chaotic, as in climate, epilepsy or sepsis. The authors identify failure modes in prior hierarchical DSR models and fix them with feature splitting and physical sparsity priors, correctly predicting bifurcations and beyond without known control parameters.