Cosine annealing is a learning rate schedule that decreases the rate following a cosine curve, spending most of training at moderate rates and decaying smoothly to near zero. This smooth schedule avoids abrupt drops that can destabilize training and has become the default schedule for most large language model training runs.