Layer- and scale-dependent hierarchical reassignment rules

Determine, for RQ-VAE–based collaborative tokenizers used in generative recommendation (e.g., the DACT framework), whether the hierarchical code reassignment rules should be varied across different codebook layers and dataset scales when applied to large-scale item corpora, rather than using a single uniform policy (e.g., always reassign at layer 1 and conditionally reassign deeper layers only when layer 1 changes).

Background

DACT introduces a hierarchical code reassignment strategy for RQ-VAE tokenizers to balance stability and plasticity: layer 1 is actively reassigned to reflect major semantic shifts, while deeper layers are updated only if the first layer changes, thereby limiting unnecessary code changes and preserving the GRM’s token–embedding alignment.

In the context of large-scale item corpora, the authors explicitly note that it is not yet established whether a uniform reassignment policy across all layers and scales is optimal. Understanding if and how reassignment rules should vary with layer depth and corpus scale is crucial for reliable continual tokenization at production scale.