Preprint
Aug 2026
Rethinking Item Tokenization in Generative Recommenders: From Fixed Atoms to Semantic Subwords
Semantic Subword Tokenization (SST) is proposed, which represents historical items as variable-length semantic subwords while preserving fixed-length target decoding and introduces Behavior-induced Co-occurrence Augmentation to inject coarse-grained semantic prefix transition signals.
Xin-Rui Miao, Mingjia Yin, Jiaqing Zhang et al.
· 0 citations