bibkey: “mohamadiliwusutherland2024grok” authors: “Mohamad Amin Mohamadi; Zhiyuan Li; Lei Wu; Danica J. Sutherland” year: 2024 title: “Why Do You Grok? A Theoretical Analysis of Grokking Modular Addition” doi: null claim: “For modular addition, kernel-regime predictors need far more samples than appropriately bounded interpolating networks, so grokking reflects escape from kernel-like behaviour to the limiting behaviour of gradient descent on deep networks.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://arxiv.org/abs/2407.12332”

Why Do You Grok? A Theoretical Analysis of Grokking Modular Addition

Verified locator

ICML 2024; arXiv:2407.12332. Theorems 3.4, 3.6, 4.3 and 4.5 compare kernel sample lower bounds with sufficient sample sizes for bounded networks. Cited only as background on grokking outside the fixed-feature setting.

Declared identifiers: https://arxiv.org/abs/2407.12332.