Training an LLM-RecSys Hybrid for Steerable Recs with Semantic IDs
Eugene Yan
Researchers trained a hybrid language model and recommendation system using semantic IDs—meaningful tokens that represent items in a product catalog—enabling an LLM to recommend items while supporting natural language steering and reasoning about its choices. The system was trained on 66,000 video game products from Amazon Reviews 2023 with 78,600 user interaction sequences averaging 6.5 items each. The resulting model unifies search, recommendations, and chat in a single interface while sacrificing some precision compared to specialized recommender systems.
Why it matters
An LLM that can converse in English & item IDs, and make recommendations w/o retrieval or tools.
