New neural network design handles longer sequences without retraining

Original: Length Generalization with Log-Depth Recurrent Units

Source: arxiv.org ↗