Grevo: A Unified Generative Recommendation Framework with Evolutionary Item Indexing

Generative Recommendation (GR) typically relies on Semantic Identifiers (SIDs) to represent items. Traditional methods either freeze these identifiers using a static tokenizer (which ignores behavioral patterns) or use complex end-to-end coupling (which requires multiple models and unstable optimization). Grevo addresses this by treating the SID assignment as an evolvable variable. It uses a single multitask Transformer that performs two tasks: Behavioral SID Generation (BSG) and Semantic SID Grounding (SSG). By using the trained recommender itself as a posterior evaluator, Grevo periodically updates a subset of identifiers to better align with user behavior without needing a separate tokenizer or complex alignment losses.

Grevo operates in a round-based lifecycle: pre-train, collect, enumerate, evolve, and retrain. After each training round, the model identifies 'high-risk' identifiers—those that the model struggles to predict or that lack consistency between behavioral and semantic tasks. It then generates candidate identifiers by borrowing tokens from behaviorally similar items or under-utilized codebook entries. These candidates are scored based on their potential to improve recommendation accuracy and consistency, and the best candidates are committed under a strict budget. This process ensures that the index remains stable while gradually adapting to the actual consumption patterns of users.

This approach effectively bridges the semantic-behavioral gap that has historically limited generative recommenders. By folding the tokenizer's role into the recommender itself, Grevo eliminates the need for auxiliary models or delicate alternating-optimization schedules. This results in a more robust and efficient framework that consistently outperforms state-of-the-art generative baselines on real-world datasets, demonstrating that a well-organized, evolving index is more valuable than a static one.

Semantic Identifier (SID)

A discrete sequence of tokens used to represent an item, allowing the recommender to treat retrieval as an autoregressive generation task.

Behavioral SID Generation (BSG)

A task where the model learns to predict the target item's SID based on the user's historical interaction sequence.

Semantic SID Grounding (SSG)

A task where the model learns to recover an item's SID from its own semantic content representation, ensuring the index retains descriptive meaning.

Evolutionary Item Indexing

A post-hoc refinement process that updates SID assignments based on posterior signals from a trained model rather than back-propagating gradients into a tokenizer.

Mutable Layers

Specific layers within the hierarchical SID structure that are permitted to change during the evolution process to allow for fine-grained index refinement.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.