bibkey: “prietobarsbeymedianobirdal2025grokking” authors: “Lucas Prieto; Melih Barsbey; Pedro A. M. Mediano; Tolga Birdal” year: 2025 title: “Grokking at the Edge of Numerical Stability” doi: null claim: “Without regularization, grokking tasks approach numerical limits of the softmax; grokking can be induced without weight decay by removing the naive loss-minimization gradient component.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://arxiv.org/abs/2501.04697”

Grokking at the Edge of Numerical Stability

Verified locator

ICLR 2025; arXiv:2501.04697. Cited only as a report of delayed generalization without weight decay, i.e. outside the fiber-law volume’s fixed-feature theorems.

Declared identifiers: https://arxiv.org/abs/2501.04697.