Personalization in the Era of LLMs - Shivam Verma, Spotify

Spotify is evolving its recommendation architecture by shifting from traditional, siloed machine learning pipelines toward unified, transformer-based foundational models. This transition centers on three pillars: foundational user modeling, catalog understanding, and durable personalization. By utilizing semantic IDs, the platform compresses complex content vectors into hierarchical tokens, allowing LLMs to autoregressively predict user preferences. Furthermore, the integration of soft tokenization projects individual user representations directly into the LLM’s latent space, facilitating highly personalized, steerable outputs. These advancements enable features like the AI DJ and the Taste Profile, which provide users with greater transparency and control over their listening history. This shift represents a broader industry movement toward leveraging generative AI to create more responsive and context-aware recommendation systems for millions of users.