19:40Teaching LLMs to Speak Spotify
Yves Raimond and Jacqueline Wood explain how Spotify is moving from ranked recommendations to generative personalization: experiences that users can steer and that can explain their recommendations. They introduce the Large Taste Model and show how semantic IDs let open-weight LLMs represent Spotify catalog items alongside natural language. The NEO training approach grounds those IDs while preserving the model’s language abilities, then tunes the model across recommendation tasks; experiments compare training and inference choices. The talk also describes production uses, including podcast discovery, and how grounded LLM judges help evaluate recommendation quality and explanations.
