Knowledge-Geometry Decoupling: Refreshable Pretrained Transfer for Streaming Recommendation

The paper introduces Knowledge-Geometry Decoupling (KGD) for pretrain-then-transfer recommendation under behavioral drift. Behavioral Multi-Token Prediction (BMTP) selects collaboratively or semantically related future items instead of treating every adjacent transition as supervision. A refreshable encoder stores behavioral knowledge, while read-only cross-attention and an Anchored Calibration Residual (ACR) let a task learner write task-specific geometry separately. The abstract reports 4–12% gains over strong transfer baselines on eight public benchmarks, sustained gains over a 90-day production stream, and Shopee Homepage Search A/B improvements of 1.75% GMV per user and 1.53% advertising revenue. Core code is released as KGD4REC.