A Survey of Item Identifiers in Generative Recommendation: Construction, Alignment, and Generation

The paradigm shift from discriminative to generative recommendation has revolutionized the field, empowering systems to autoregressively generate the next item in a user's interaction sequence. Central to this new paradigm is the representation of items, specifically, the item identifier (item ID), which acts as the fundamental bridge connecting discrete items in a catalog to the continuous generation space of Large Language Models (LLMs) and Transformer-based architectures. This survey provides a comprehensive and systematic review of item IDs in Generative Recommender Systems (Gen-RecSys), covering both LLM-as-Gen-RecSys and Semantic ID (SID)-based Gen-RecSys. We propose a novel taxonomy that categorizes existing works based on the lifecycle of Item IDs: Construction, Alignment, and Generation. We first posit six critical quality criteria for ideal item IDs, including uniqueness, semantics, etc., and use them to evaluate various ID types, ranging from numeric and textual IDs to learned semantic codes. Subsequently, we delve into the technical methodologies for constructing SIDs, analyzing diverse tokenization schemes such as hierarchical clustering and codebookbased quantization (e.g., VQ, RQ, and PQ). Furthermore, we systematically review strategies for aligning these IDs with collaborative and multi-modal signals, as well as techniques for ensuring validity and accelerating decoding during the generation phase. Finally, we highlight open challenges regarding ID collision, semantic flatness, and inference efficiency for SIDs, and outline promising future directions to advance the design of robust and scalable item IDs.