Eugene Yan has developed a novel approach to recommender systems by training a hybrid language model that understands both natural language and item IDs. This model, which extends the vocabulary of a language model with semantic ID tokens, can generate recommendations based on user history and also respond to conversational prompts to steer suggestions. The system aims to combine the world knowledge of LLMs with the catalog awareness of traditional recommender systems, offering steerability and reasoning capabilities. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
RANK_REASON This is a research paper detailing a novel method for training an LLM-recommender hybrid.