Spotify Research presents two complementary systems for LLM-based recommendation at scale, both accepted at KDD 2026. NEO is a unified, language-steerable model that handles recommendation, search, explanation, and user understanding within a single set of weights by treating catalog items as Semantic IDs — discrete tokens derived from item content. GLIDE is a production 1B-parameter LLM deployed for podcast discovery, introducing the concept of non-habitual listening and combining short-term history with dense long-term user embeddings as soft tokens. In a 21-day A/B test covering ~20 million impressions, GLIDE increased non-habitual podcast listening by 5.4% and new-show discovery by 14.3%. Both systems rely on staged training, beam-search decoding, and hybrid evaluation combining offline metrics, human judges, and LLM-based judges to bridge the gap between retrieval accuracy and real recommendation quality.
