Generative recommendation has recently garnered significant industry attention due to its potential for enhanced scalability and stronger model capacity. However, deploying real-time generative recommendation in large-scale advertising environments demands specialized designs that go beyond conventional large-language-model (LLM)-style training and serving methodologies.
Addressing these challenges, Kuaishou has developed GR4AD (Generative Recommendation for ADdvertising), a production-oriented generative recommender system. GR4AD is an integrated solution, co-designed across its architecture, learning mechanisms, and serving infrastructure to meet the rigorous demands of large-scale deployment.
GR4AD introduces several key innovations. For tokenization, it proposes UA-SID (Unified Advertisement Semantic ID), an advanced mechanism engineered to capture and represent complex business information effectively, providing richer and more accurate input for the model.
Furthermore, GR4AD incorporates LazyAR, a novel lazy autoregressive decoder. This decoder relaxes layer-wise dependencies specifically for generating short, multi-candidate sequences. This design preserves recommendation effectiveness while substantially reducing inference costs, thereby facilitating scalable operations under fixed serving budgets.
To ensure that model optimization is closely aligned with business value, GR4AD employs VSL (Value-Aware Supervised Learning). Building upon this, the system introduces RSPO (Ranking-Guided Softmax Preference Optimization), a ranking-aware, list-wise reinforcement learning algorithm. RSPO is designed to optimize value-based rewards under list-level metrics, enabling continuous online updates that maximize business outcomes.
For online inference, GR4AD also implements dynamic beam serving. This technique dynamically adapts the beam width across different generation levels and in response to online load, effectively managing computational resources and ensuring efficiency and stability in high-concurrency scenarios.
Large-scale online A/B tests have demonstrated significant improvements, with GR4AD achieving up to a 4.2% increase in ad revenue compared to an existing DLRM-based recommendation stack. These gains were consistent, stemming from both model scaling and inference-time optimizations. GR4AD has been fully deployed within Kuaishou's advertising system, serving over 400 million users with high-throughput real-time recommendations.