Kuaishou unifies generative recall and ranking in one decoder-only Transformer

_reachsumit · x · 2026-07-28

Kuaishou’s paper proposes UniR², a decoder-only Transformer that unifies generative recall and multi-objective ranking in one heterogeneous sequence.

• It targets the common two-stage recommendation pipeline, where recall and ranking are separated.
• The authors argue that this split creates objective mismatch, information loss at the hand-off, and duplicated user-context computation.
• UniR² uses a single sequence containing user context, SID trajectory, and item features.
• The generated trajectory acts as a bridge between recall and ranking.
• Dual-Query Prefix-Causal Attention gives the two tasks different visibility.
• The backbone attention weights are shared, while ranking keeps separate optimization boundaries via LoRA so ranking adaptation does not disturb the generative model.
• The paper reports strong offline results on recommendation benchmarks.

Original post →

More from Research

• Kevin Murphy's Model Discovery Agent: LLM-assisted Bayesian experiment design hits SOTA — sirbayes · 2026-08-26
• GVHMR: World-Grounded Human Motion Recovery via Gravity-View Coordinates — rsasaki0109 · 2026-08-26
• François Chollet Releases Tutorial on LLMs and Text Generation — culurciello · 2026-08-26
• Mean scores mislead RL progress; Google Research proposes IQM and confidence intervals — burkov · 2026-08-26
• Paper reveals Agent benchmark unreliability: Harness variance 7.8x model variance — omarsar0 · 2026-08-26
• OpenAI staff claims Transformer is now the bottleneck to AI progress — KyeGomezB · 2026-08-26

Research channel →