“Research by Lovelace et al. (2026) found that using strong weight decay reduces the overfitting penalty caused by data repetition in data-constrained training.”