Math & Optimization · advanced · concept 75 of 311 · Updated Nov 4, 2026

💡Grokking

A network memorizes a small dataset, sits at near-perfect training accuracy with poor test accuracy for a long time, then suddenly generalizes. Power et al. at OpenAI reported it in 2022 on modular addition. Nanda et al. reverse-engineered a grokked transformer and found a Fourier-based algorithm, with weight decay slowly favoring that circuit over the memorized solution. Test loss can stay flat long after most people would stop training.