Densing law for user representation learning: tokenizing behavior breaks billion-scale scaling bottleneck

_reachsumit · x · 2026-08-25

The paper proposes a User Behavioral Densing Law characterizing the quantitative relationship between data scale and minimum sufficient tokenization capacity. On billion-scale Alipay data, raw data scaling hits a bottleneck, while tokenization yields sustained gains. They find an approximately linear log-log relationship between minimum sufficient tokenization capacity and input data size, with slope varying by method and data source.

Original post →

More from Research

Research channel →