Learning rate collapse prevents training recurrent neural networks at scale - NeurIPS'25 - https://openreview.net/forum?id=pFs3jNjT5f