Scaling behavior of residual-generation acceleration

Determine how the acceleration ratio of the Residual Cascading Generation module changes as the model size, codebook depth, and candidate budget are scaled.

Background

The Residual Cascading Generation module is designed to improve inference efficiency by avoiding repeated heavy Transformer Decoder passes during hierarchical Semantic ID generation. Its reported efficiency benefits are established for the experimental configuration evaluated in the paper, but the authors identify uncertainty about whether those benefits scale predictably when key system dimensions change. Specifically, the unresolved issue concerns the evolution of the acceleration ratio under increases in model size, codebook depth, and the number of generated candidates, which are important factors for large-scale deployment.