Generative Recommendation in Dynamic Catalogs: A Survey of Item Admission, Semantic ID Evolution, and Model Adaptation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Generative recommendation must connect predicted identifiers to a catalog that changes after training. New items precede feedback, identifier assignments evolve, and request rules constrain which outputs can be served. This survey compares item admission, semantic-ID evolution, and model adaptation through their update dependencies and dynamic operating conditions. A coupled-state framework connects catalog events to item assignments, model parameters, histories, and access structures. We use it to explain when metadata retrieval, targeted editing, continual learning, recoding, and request filtering address the relevant bottleneck. Three findings synthesize interface dependencies and their operating conditions: access and feedback learning require separate tests; recoding requires coordinated model and resolver handoffs; candidate access, item resolution, and eligibility expose different failures. From these findings and the heterogeneous cost evidence, we derive an evaluation principle: assess update quality against complete cost within the item's available lifetime. Six source-located comparison groups show when cohort choice reverses a baseline preference, mapping repair leaves learning unresolved, and evaluation or filtering changes the meaning of a gain. A conditional method-selection table connects these contrasts to the information, migration work, and service window required by each route. The resulting evaluation guidance preserves early access-only intervals, measures quality by item age, and accounts for preparation, updates, and serving. Open questions concern when to switch or combine update routes, how to migrate identifier versions, and how admission translates into useful exposure.