New findings reveal how smaller learning rates are key to efficient training for large language models, offering a rule-of-thumb for transferring hyperparameters and improving overall performance. In ...