Analysis of 818,904 measurements across ten open-weight models reveals that middle layers form a high-relative-hyperbolicity plateau, while final layers become more tree-like. The study finds that layer depth is a more consistent predictor of hierarchical structure than parameter scale.
HOW THIS AFFECTS YOU
●
researcherYou can use Gromov Hyperbolicity as a diagnostic tool to map the hierarchical distance structure within different model layers.