UniRec: Cross-stage Multi-Task Fusion with Preference Alignment for Cascaded Recommender Systems
Lingyuan Kong, Jiaqi Cui, Fanjiao Zeng, Congqi Wang, Yu Li, Yuan Cheng, Jingxin Liu, Xiaoshuang Chen, Kaiqiao Zhan
cs.IR
2026-09-10
UniRec jointly trains Kuaishou pre-rank and rank fusion with cross-stage alignment. Full-traffic A/B lifts app usage 0.616% and watch time 0.675%.
Cascaded recommenders cut thousands of candidates in pre-ranking, then score a much smaller set in ranking. The two stages differ in objectives, features, and latency, so their fusion modules are usually trained apart. Pre-ranking can drop items the ranker would have kept. Independently tuned ranking fusion can then cancel upstream gains.
Prior cross-stage work mostly freezes the ranker as a teacher and adds a distillation term to pre-ranking scores. Jointly training both fusion heads has been largely untouched.
UniRec places one fusion agent at each stage, shares part of the input embedding, and back-propagates through a single graph. Ranking-only pXTRs stay in a private embedding. Pre-ranking uses MLP-Mixer to cover a larger candidate set; ranking uses a slim AutoInt block because the list is smaller. An adapter feeds the pre-ranking representation into ranking with a stop-gradient, so ranking cannot rewrite the upstream tower. Cross-stage coupling lives in the shared embedding.
Supervision has two axes. Vertically, downstream pairwise order is copied onto the upstream score; the loss matches order, not score scale. Horizontally, dozens of pairwise task losses are summed into positive and negative evidence before Softplus, shrinking the backward graph from O(M N²) to O(N²) without changing the algebra. Exposed items add a posterior feedback term.
Unconstrained fusion will exploit attributes such as video duration: long videos collect more watch-time evidence, so lifting a whole bucket reduces loss. AGRR buckets by duration, computes group-relative advantages, and asks the fusion scores to match the within-bucket target. Promoting an entire bucket yields no gain; cross-bucket order is still handled by the pairwise losses.
On RecFlow, UniRec reaches ranking NDCG@10/30/50 of 0.0990 / 0.1870 / 0.2409, above the single-stage fusion model UMRE at 0.0947 / 0.1818 / 0.2376. rankAUC is 0.5912 versus UMRE's 0.6009; AUC pools across requests, NDCG scores within-request order, and truncation cares about the latter. Cross-stage ASH (how much of ranking Top-10 survives pre-ranking Top-50) is 0.9978, against 0.9855 for UMRE and 0.9668 for COPR, which only updates pre-ranking. Kendall τ / Spearman ρ are 0.6318 / 0.8072.
Independently trained EMER lands at 0.7219 ASH, which is the failure mode UniRec is built against: a strong ranker whose preferred items never make the shortlist.
Replacing both production fusion modules on Kuaishou, a one-week 20% A/B test reports +0.616% app usage time, +0.675% total watch time, +0.755% video watch time, +0.189% active users, +2.367% likes, and +1.573% follows. No listed metric fell. The system is fully deployed.
In production, the fusion formula decides who survives, not any single pXTR head. Treating the two fusion modules as one differentiable problem is a better fit than another distillation coefficient on pre-ranking. CPPA makes dozens of ranking objectives trainable; AGRR removes duration gaming as a free lunch. A 0.6% usage-time lift is a full-fleet change at Kuaishou's scale.
The alignment loss does not update ranking-private parameters, and ablations show almost no ranking-NDCG change from joint training. AGRR currently buckets only on video duration. Absolute RecFlow NDCG is low, tied to label processing and cutoffs, so it should not be compared raw across papers. Re-ranking, e-commerce, and ads are left for later.