“In data-constrained settings, it is more effective to train an ensemble of smaller models than to train a single large model with the same total parameter count.”