“Learning Length-Extrapolatable Recurrent Models” Recurrent models can naturally run beyond their training length, but backpropagation through time often gives weak long-range learning signals. This paper proposes Credit Stabilization through Time (CST). Instead of modifying parameter gradients after signals are mapped and aggregated, CST stabilizes state credit—the upstream signal through which future losses reach earlier recurrent states. On GDN-based models, CST improves length up to 128×…