A Modular Survey for Semantic ID-Based Generative Recommendation

Traditional discriminative recommenders score and rank items indexed by single item IDs, whereas Semantic ID-based generative recommendation formulates recommendation as conditional generation of Semantic ID token sequences. This shift offers a unified view of retrieval and ranking and shows promising scaling properties, but the literature is fragmented across tokenization and quantization choices, model backbones, and training and decoding protocols, making systematic comparison difficult. To address this, we present the first survey that organizes the field, with four pivotal contributions. First, we introduce a unified five-stage reference pipeline: Representation Layer, Tokenization, Generative Backbone, Training, and Inference. This pipeline standardizes terminology and exposes shared structure. Second, grounded in this pipeline, we map existing methods into a fine-grained typology along semantic granularity, architectural coupling, and learning objectives. Third, based on this structured view, we provide a scaling-oriented perspective that connects component-level decisions to expressiveness, efficiency, and empirical performance, clarifying trade-offs. Finally, we synthesize open challenges and concrete directions that follow from the identified bottlenecks. To support reproducibility and controlled ablations across stages, we release UniGenRec (https://github.com/hupeiyu21/UniGenRec-A-universal-generative-recommendation-toolbox), an open-source modular toolbox implementing the proposed pipeline.