“Generalization and knowledge transfer may be more effective in larger models because they can learn shared representations, whereas smaller models might rely more on memorization.”