Grokking Metrics Overstate Compression; Lag Between Accuracy and Representation

This research challenges the conventional understanding and measurement of "grokking," a phenomenon where neural networks suddenly generalize to unseen data long after achieving high training accuracy. The authors demonstrate that metrics commonly used to assess representation compression during grokking significantly overstate the actual converged values. For instance, effective rank measurements taken at the grokking transition point can be 3 to 5 times higher than the true converged value for MLPs and 1.3 to 1.5 times higher for transformers. A key finding is that the compression of network representations lags the accuracy transition by tens of thousands of steps, rather than coinciding with it. The study identifies factors like LayerNorm as influencing this lag. To address these measurement inaccuracies, the paper proposes a new audit toolkit designed to separate the onset of generalization from the actual compression, flag censoring, and ensure reference floors have plateaued, providing a more valid assessment of grokking.