Kuaishou's UniRec jointly trains pre-ranking and ranking to fix cascade inconsistency

_reachsumit · x · 2026-09-11

Kuaishou proposes UniRec, a unified cross-stage recommendation fusion model addressing cross-stage inconsistency in cascaded recommenders: upstream models may filter out items the downstream ranker prefers, and independently tuned downstream fusion can offset upstream gains.

Key ideas:

Original post →

More from Research

Research channel →