Shopee's KGD paper: 1.75% GMV lift in 90-day live A/B test

Production A/B numbers from a live e-commerce recommender don't turn up in academic papers very often, which is why a new arXiv preprint from a Shopee-affiliated team is worth a closer look than the usual pretrain-transfer paper. The method, called Knowledge-Geometry Decoupling or KGD, pairs a behavioral multi-token prediction objective with an architecture that separates refreshable pretrained knowledge from task-specific geometry. On eight public benchmarks the authors report 4 to 12 percent gains over strong pretrain-transfer baselines.

The more interesting number is the one from Shopee's live traffic. In the reported A/B test, KGD increases GMV per user by 1.75 percent and advertising revenue by 1.53 percent. In production ranking, single-digit lifts of that size are the kind of thing platform teams fight over for months; the paper's larger claim is that KGD sustains its advantage over a 90-day production stream while baselines show no gains, which if it holds is more striking than the point-in-time lift itself. The mechanism the authors credit is continual knowledge refresh without task-gradient interference or invalidating downstream adaptation, which is a real known failure mode in streaming recsys.

The honest caveat is that this is one team's write-up of one platform's experiment. The paper does not spell out, at least not in what's visible from the abstract page, exactly which pretrain-transfer baselines flatlined over those 90 days, whether they were refreshed on the same cadence, or what the extra serving cost of the decoupled architecture looks like at Shopee's scale. Take the specifics as reported by the authors, not as independently settled.

Still, the direction is the useful part. Recommender teams at marketplaces, ads platforms and feed products have been living with the mismatch between big pretrained sequence models and the fact that user behavior keeps moving under them. A concrete design that treats knowledge and geometry as separate things you can refresh at different rates, plus a released codebase and a real GMV number attached to it, is the sort of thing worth reproducing in your own harness before the next planning cycle.