Juliana Li

Language models can forget learned rules during training, with their fate determined by how frequently the training data supports those rules, not by the model's size or complexity.