bibkey: “guchenzhanghucao2025beyond” authors: “Zihan Gu; Ruoyu Chen; Hua Zhang; Yue Hu; Xiaochun Cao” year: 2025 title: “Beyond Progress Measures: Theoretical Insights into the Mechanism of Grokking” doi: null claim: “Grokking on algorithmic tasks is analysed through a quotient of the input space by equality of targets and through embedding uniformity, rather than weight decay alone.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://arxiv.org/abs/2504.03162”

Beyond Progress Measures: Theoretical Insights into the Mechanism of Grokking

Verified locator

arXiv:2504.03162. Definition 3.1 identifies input pairs with equal targets and works on the quotient of inputs; cited only to record that this quotient language concerns task inputs, not parameter directions invisible to the training readout.

Declared identifiers: https://arxiv.org/abs/2504.03162.