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:
- The two fusion modules partially share input embeddings and train in a single computation graph, so gradients from either stage propagate through shared representations.
- A dual-axis preference alignment objective: a vertical cross-stage consistency term transfers downstream pairwise preferences upstream, and a horizontal compact aggregation term reorganizes dozens of pairwise objectives into bidirectional preference evidence.
- The authors also identify degeneration risks in unconstrained end-to-end fusion optimization and mitigate them.
More from Research
- Survey proposes PAPAV: a unified framework for multimodal embodied agents — MikeShou1 · 2026-09-11
- Terence Tao warns AI math capability strongly correlates with cyberattack ability — sytelus · 2026-09-11
- European Math Society hails OpenAI's Navier–Stokes solution, flags closed-model access concern — i_dg23 · 2026-09-11
- Debate over "slop math": academia must adapt verification norms for AI output — _sathvikr · 2026-09-11
- Alibaba's RecGPT Brings Three LLMs into Taobao's Recommender Pipeline — _reachsumit · 2026-09-11
- Meta's Auto-RecSys Deploys Autonomous Research Agents for Multi-Day Recommender Experiments — _reachsumit · 2026-09-11