Data Scaling as Progressive Coverage of a Predictive Contribution Spectrum

In three linesarXiv paper on data scaling laws: progressive coverage of a latent predictive contribution spectrum (via suffix-automaton representation) strongly correlates with empirical scaling exponent. Across 12 real corpora, log K(N) shows near-linear relationship with log N (R²≈0.96), suggesting training advances an effective frontier through a predictive state spectrum.

Summary generated by Claude — human-verified