KGD lifts Shopee GMV per user 1.75%

Shopee’s Knowledge-Geometry Decoupling recommender increased gross merchandise value per user by 1.75% and advertising revenue by 1.53% in a live homepage-search A/B test. The newly featured paper reports that KGD separates a refreshable behavior encoder from downstream task parameters, beat pretrain-transfer baselines by 4–12% across eight public benchmarks, and held its advantage during a 90-day production stream. For operators confronting the stale-context risk in Cognition’s persistent-memory acquisition, the broader design lesson is to give changing knowledge and stable task adaptation separate owners—and to demand online lift before accepting the added architecture.