bibkey: “lyujinlidulee2023dichotomy” authors: “Kaifeng Lyu; Jikai Jin; Zhiyuan Li; Simon S. Du; Jason D. Lee; Wei Hu” year: 2023 title: “Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokking” doi: null claim: “With large initialization and small weight decay, homogeneous networks first converge to a kernel predictor and later transition to a min-norm or max-margin predictor, producing grokking.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://arxiv.org/abs/2311.18817”

Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokking

Verified locator

ICLR 2024; arXiv:2311.18817. Two-phase (kernel then rich) picture under large initialization and small weight decay. Cited as background for the boundary of the fiber-law volume’s linearized bookkeeping; no theorem from it is re-proved and no breakdown time is inferred from it.

Declared identifiers: https://arxiv.org/abs/2311.18817.