bibkey: “liumichaudtegmark2022omnigrok” authors: “Ziming Liu; Eric J. Michaud; Max Tegmark” year: 2022 title: “Omnigrok: Grokking Beyond Algorithmic Data” doi: null claim: “Initialization norm and weight decay control delayed generalization across image, language and molecular tasks (the LU mechanism).” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://arxiv.org/abs/2210.01117”
Omnigrok: Grokking Beyond Algorithmic Data
Verified locator
ICLR 2023; arXiv:2210.01117. Empirical dependence of grokking on initialization scale and weight decay. Cited as directionally consistent with the initialization-scaling prediction of the fiber-law volume; not used as verification of that prediction.
Declared identifiers: https://arxiv.org/abs/2210.01117.