Teaching LLMs to Speak Spotify — Yves Raimond & Jacqueline Wood, Spotify

Spotify is evolving its recommendation infrastructure from traditional ranking algorithms to "generative personalization," a system that enables dynamic, interactive, and steerable user experiences. This transition relies on the "Large Taste Model," which integrates semantic IDs—quantized content embeddings—into open-source LLMs to bridge natural language understanding with catalog knowledge. The "Neo" training paradigm, consisting of four distinct stages, ensures the model maintains core language capabilities while learning to reason about user intent and content context. By employing grounded LLM judges, the system achieves high alignment with human preferences, allowing for sophisticated features like natural language playlist generation, personalized daily briefings, and user-editable taste profiles. This approach moves beyond simple content matching, enabling the platform to explain recommendations and adapt dynamically to user feedback in real time.