bibkey: “xuwangfreivardihu2023benign” authors: “Zhiwei Xu; Yutong Wang; Spencer Frei; Gal Vardi; Wei Hu” year: 2023 title: “Benign Overfitting and Grokking in ReLU Networks for XOR Cluster Data” doi: null claim: “Two-layer ReLU networks trained by gradient descent on noisy XOR cluster data with small initialization and no weight decay first fit the noisy labels with near-chance test accuracy and later reach near-optimal test accuracy while still interpolating.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://arxiv.org/abs/2310.02541”

Benign Overfitting and Grokking in ReLU Networks for XOR Cluster Data

Verified locator

arXiv:2310.02541. Theorem 3.1 under Assumptions A1-A6. Cited only as a proved example of delayed generalization by feature learning without weight decay, outside the fixed-feature theorems of the fiber-law volume.

Declared identifiers: https://arxiv.org/abs/2310.02541.