Preprint
Aug 2026
Knowledge-Geometry Decoupling: Refreshable Pretrained Transfer for Streaming Recommendation
Knowledge-Geometry Decoupling (KGD) is proposed and Behavioral Multi-Token Prediction (BMTP) is introduced to retain only collaboratively or semantically related future items as supervision, yielding cleaner and more transferable behavioral knowledge.
Zixuan Wang, Yuhong Chen, Yuxuan Zhu et al.
· 0 citations