put two and two together between SimBA paper and a new grokking paper from iclr 2025, both calling for controlling model complexity for better generalization