English

A Training-Time Diagnostic for Generalization via the Log-Alignment Ratio

Machine Learning 2026-05-29 v1

Abstract

We study the log-alignment ratio (LAR), a measure of parameter-activation alignment, introduced in parameterization theory. We reformulate it as the overlap between a weight spectrum pp of the normalized squared singular values of a matrix and an activation spectrum qq of the normalized squared projections of inputs onto its singular directions. We show that unembedding LAR tracks the transition between memorization and generalization in two different settings by capturing the spread of pp and qq during training. In grokking, LAR predicts the effective dimension of the learned function: kn2(1LAR)k \approx n^{2(1-\text{LAR})}, where nn is the input dimension of the matrix. In 3B-parameter language model pre-training, its deviation from a non-overfitting baseline tracks the generalization gap, and its rate of decline increases as overfitting approaches. LAR is computable from quantities available during the forward pass with negligible computational overhead, and requires no held-out validation data.

Cite

@article{arxiv.2605.28975,
  title  = {A Training-Time Diagnostic for Generalization via the Log-Alignment Ratio},
  author = {Ali Shehper and Ashish Vaswani},
  journal= {arXiv preprint arXiv:2605.28975},
  year   = {2026}
}

Comments

32 pages, 25 figures