OneRanker: Unified Generation and Ranking with One Model in Industrial Advertising Recommendation

A research team from Tencent has published a paper on OneRanker, a novel AI architecture designed to revolutionize industrial-scale advertising recommendation systems. The model addresses a fundamental tension in modern ad tech: the shift from traditional multi-stage 'cascaded' systems to end-to-end generative AI often creates misalignment between predicting user interest and maximizing business value. OneRanker proposes a deep architectural integration to solve three core challenges: objective misalignment, target-agnostic generation, and the disconnection between generation and ranking stages. This approach moves beyond the trade-offs of previous methods, which either suffered from single-stage optimization tension or lost critical information by decoupling the stages.

The technical innovation lies in three key components. First, a value-aware multi-task decoupling architecture uses task token sequences and causal masks to separate interest coverage and value optimization within shared model representations. Second, a coarse-to-fine collaborative target awareness mechanism employs 'Fake Item Tokens' for implicit guidance during ad generation and a dedicated ranking decoder for explicit value alignment. Finally, input-output consistency is guaranteed through Key/Value pass-through mechanisms and a Distribution Consistency Constraint Loss for end-to-end optimization. The full deployment on Tencent's WeChat (WeiXin) channels advertising platform has yielded a significant 1.34% lift in Gross Merchandise Value (GMV), proving its industrial feasibility and setting a new benchmark for unifying generative and ranking tasks in real-world systems.