Tag
The paper proposes that grokking in deep neural networks arises from noise-driven escape from metastable phases in first-order L2 phase transitions, demonstrating that delayed generalization follows Arrhenius scaling and reproduces canonical grokking curves.