Unifying Generative Recall and Multi-Objective Ranking in a Single Decoder-Only Sequence
Ruochen Yang, Shuang Wen, Pengbo Xu, Yusheng Huang, Jiangxia Cao, Shuang Yang, Zhaojie Liu, Jiawei Sheng, Tingwen Liu
cs.IR
2026-07-27
UniR2 unifies generative recall and multi-objective ranking in one decoder-only Transformer over a heterogeneous sequence, using Dual-Query Prefix-Causal Attention for task-specific visibility and ranking-side LoRA to preserve adaptability. Two-week online A/B on Kuaishou: play volume +1.177%, like rate +2.560%, total gifting +2.569%.
Industrial recommendation usually splits recall and ranking into two independent pipelines. The cascade supports corpus-level retrieval and fine-grained multi-objective scoring, but the costs are real: objective inconsistency between stages, information loss at the candidate hand-off, and user-side context computed twice. Generative recall and multi-objective ranking share a Transformer backbone, so there is a natural opportunity to merge. The hard part is that direct sharing breaks, because the two stages need different information visibility and different optimization. UniR2 unifies generative recall and multi-objective ranking in one decoder-only Transformer.
The model works over one heterogeneous sequence S = [P || Tgen || Trank]: P is the user segment (profile plus per-behavior history for click, long-view, gift and so on), Tgen is the SID trajectory (3-layer Res-Kmeans, codebook 8129), and Trank holds item-feature tokens.
The core is Dual-Query Prefix-Causal Attention, which gives each task different visibility. The generative query attends only to the user segment and its own preceding positions; the ranking query uses scoped bidirectional self-attention over the user, generative, and ranking segments. Ranking can therefore consume the entire recall decision process, not only its discrete candidate output.
The generated SID trajectory acts as a representation bridge between the two. For optimization, the base projections are shared, but the recall and ranking parameter sets are disjoint; ranking-side LoRA residuals preserve adaptability without corrupting the generative backbone, giving forward coupling with optimization isolation.
Offline on large-scale industrial data, recall (vs OneLive, NEZHA, PROMISE) reaches HR@128 of 0.8114 on Show (+3.05%) and 0.8741 on Click (+3.86%), with MRR ahead across the board. Ranking CTR AUC is 0.8513 (+0.75%), UAUC +1.45%.
Online A/B ran two weeks on Kuaishou (June 2026, 5% traffic). Kuaishou main app: play volume +1.177%, follow rate +0.655%, like rate +2.560%. Kuaishou Lite: gifting users +0.717%, gifting intention +1.567%, total gifting amount +2.569%. For efficiency, the unified architecture reuses user context and KV cache computed during recall, runs ranking and strategy filtering in parallel, and removes cross-service data transfer between two independent models.
For industrial recommendation teams, this is a workable blueprint for unifying recall and ranking. Dual-Query Attention splits visibility and LoRA isolates the optimization boundary, so you get representation coupling without ranking gradients wrecking the generative backbone. A 2.569% lift in total gifting on live traffic shows the architecture holds at billion-user scale. It also reflects a trend: once recall and ranking are both Transformer-based, the engineering dividend of merging them is worth counting seriously.
The paper has no dedicated limitations section, common for industrial-system papers. A few points are readable between the lines. Scaling returns show clear diminishing gains. When SID semantics are unstable, two-stage training is needed to prevent negative transfer. Dropping the long history from the ranking query saves FLOPs but demands that the generated SID trajectory condense enough of the user-item matching process. An unverified concern: tying recall and ranking into one model enlarges the single-point failure surface, and training and serving complexity plus blast radius are not discussed; production stability has to be worked out separately.