Unifying Generative Recall and Multi-Objective Ranking in a Single Decoder-Only Sequence

Modern industrial recommendation systems typically operate as a two-stage cascade: a recall stage for retrieving candidates and a ranking stage for scoring them. This separation leads to objective inconsistency, redundant computation of user context, and information loss at the boundary between stages. The authors investigate whether these two distinct tasks can be unified into a single Transformer-based architecture to improve efficiency and performance.

To address this, the authors propose UniR2, a unified decoder-only Transformer that processes a single heterogeneous sequence containing user context, SID (Semantic ID) trajectories, and item features. The model employs a novel Dual-Query Prefix-Causal Attention (DQ-PCA) mechanism. This mechanism allows the generative recall task to follow an autoregressive, prefix-causal pattern, while the ranking task utilizes bidirectional attention to fuse item features with the recall trajectory. To prevent the ranking task from destabilizing the generative backbone, the authors implement task-specific optimization boundaries using ranking-side LoRA (Low-Rank Adaptation), ensuring that the two tasks share base attention weights while maintaining separate optimization paths.

UniR2 successfully bridges the gap between recall and ranking by allowing the ranking stage to consume the recall decision process directly, rather than relying on a compressed candidate list. By sharing the user-side context computation, the model eliminates redundant encoding of long user histories. Offline experiments on large-scale industrial data demonstrate that UniR2 improves both recall and ranking performance compared to traditional cascaded baselines. Furthermore, long-term online A/B testing on the Kuaishou platform confirms consistent positive gains, validating the model's practicality for large-scale production environments.