Tutorial on Generative Recommendation: Foundations and Frontiers
TL;DR
This work comprehensively surveys recent generative recommendation advances through a tri-decoupled perspective, summarize the evolution of tokenization strategies, analyze the trade-offs of major generative architectures, and summarize the transition from supervised next-token prediction to reinforcement-learning-based strategies.
Abstract
In the current digital ecosystem, recommender systems serve as the core infrastructure for navigating large-scale content catalogs and delivering personalized services, typically following multi-stage discriminative pipelines (e.g., retrieval, ranking, and re-ranking). However, their fragmented architectures cause cascading cross-stage error propagation and suboptimal hardware utilization. This motivates a paradigm shift toward Generative Recommendation (GR). GR mitigates these issues through end-to-end unified generative modeling, optimizing for multi-dimensional preference objectives beyond local user behaviors. This work comprehensively surveys recent generative recommendation advances through a tri-decoupled perspective centered on tokenization, architecture, and optimization, the three foundational components shaping these systems. Specifically, we summarize the evolution of tokenization strategies, analyze the trade-offs of major generative architectures, and summarize the transition from supervised next-token prediction to reinforcement-learning-based strategies. Connecting these technical developments to practical deployment patterns and open challenges, we provide researchers and practitioners a foundational reference and actionable blueprint for building next-generation generative recommender systems. An updated collection of relevant papers and resources is accessible in https://github.com/Kuaishou-RecModel/Tri-Decoupled-GenRec.