Overview
Spotify is rebuilding personalization around a deceptively hard matching problem: connecting roughly 760 million monthly active users across 184 markets with more than 100 million music tracks, plus podcasts, videos, and audiobooks. Its historical progression runs from human curation, through scalable recommendation systems such as Discover Weekly, to generative personalization—experiences that reason about recommendations, explain them, and let users directly steer the outcome. Products including Spotify DJ, prompted playlists, editable taste profiles, and personal podcasts demonstrate different expressions of this shift. Underneath them sits the Large Taste Model, a shared system that combines interaction history, catalog understanding, prediction, reasoning, and real-time user control. A core technical method, NEO, converts catalog embeddings into discrete semantic IDs, adds them to an open-weight LLM, grounds the new tokens while freezing the backbone, and then teaches Spotify-specific capabilities through multitask instruction tuning. Ablations indicate that pretraining, staged grounding, and beam-search inference are important, while multitask learning can improve newer or sparse content categories. Finally, Spotify argues that generative recommendations require grounded LLM judges because conventional engagement metrics cannot assess intent alignment or explanation quality. Together, these techniques create one production system for search, recommendation, explanation, and controllable experience generation.
Sections
Strategic Insights
Higher-level implications of Spotify's transition to generative personalization.
- Spotify is turning its recommender from a collection of ranking pipelines into a user-facing interaction model. The key product shift is not merely adopting an LLM, but exposing personalization as something users can inspect, correct, and direct.
- Semantic IDs create a shared interface between natural-language reasoning and catalog-scale retrieval. This allows one model to treat users' words, behavioral histories, and recommendable entities as compatible inputs and outputs.
- The strong audiobook results imply that multitask recommendation models may redistribute knowledge from mature content domains to newer ones, reducing dependence on category-specific interaction volume.
- The evaluation results suggest that an LLM judge is not automatically a reliable evaluator merely because it is capable of reasoning. Its usefulness depends heavily on grounding it with behavioral context and task-specific evidence, especially when a query is ambiguous.
Technical Architecture and Findings
Specific implementation choices, model-training decisions, and reported evaluation results.
- Spotify quantizes existing content embeddings into discrete semantic-ID tokens, expands an open-weight LLM's vocabulary with those tokens, and fine-tunes the model to process natural language and catalog entities together.
- Domain grounding trains bidirectional mappings between text and semantic IDs while freezing the original LLM weights and embeddings; only the new semantic-ID embeddings are updated.
- Capability induction unfreezes the complete model and applies either full-parameter fine-tuning or LoRA fine-tuning across Spotify tasks such as next-item recommendation and retrieval.
- Unconstrained generation produced valid semantic IDs 98% of the time. Constrained decoding added latency but supported catalog restrictions, while top-p sampling significantly reduced accuracy relative to beam search.
- Grounded evaluation produced several reported results: 75% LLM-to-human preference alignment with textual user profiles, a 5% overall alignment increase from behavioral grounding, a 91% increase for ambiguous queries, and 0.87 agreement with human system rankings in scaled Cranfield-style evaluation.
Evolution of Spotify Personalization
The chronological progression from manual curation to production-scale generative personalization.
- Spotify personalization initially centered on people manually assembling playlists for particular tastes.
- Discover Weekly launched as an early large-scale recommendation use case.
- Spotify launched DJ as an early generative-personalization experience and subsequently added the ability for listeners to steer a session.
- Spotify launched Taste Profile in New Zealand, allowing users to inspect and edit the algorithm's understanding of their preferences in natural language.
- Spotify announced Personal Podcast, an experience that generates recurring personalized content such as a daily community brief.
- The NEO-based system is operating in production and powers use cases including podcast discovery.
Key Definitions
Core concepts used to describe Spotify's model and evaluation approach.
- Generative personalization: personalization that goes beyond matching users to content by reasoning about matches, accepting natural-language direction, explaining choices, shaping experiences dynamically, and potentially generating new content.
- Large Taste Model: Spotify's internal shared system for understanding historical interactions and catalog content, combining prediction with reasoning, and generating steerable personalized experiences in real time.
- Semantic ID: a discrete token produced by quantizing an existing content embedding, used to represent a Spotify catalog entity directly inside an LLM's vocabulary.
- Domain grounding: the NEO stage that aligns new semantic-ID embeddings with the pretrained model's language-embedding space through text-to-ID and ID-to-text mappings while keeping the backbone frozen.
- Capability induction: multitask instruction tuning that unfreezes the model and teaches Spotify-specific tasks such as retrieval and next-item recommendation.
- Catastrophic forgetting: degradation of a pretrained model's original language or world-knowledge capabilities while it learns a new domain; Spotify's frozen-backbone grounding stage is designed to mitigate it.
- Cranfield-style collection: an evaluation set built by pooling candidates from multiple sources and ranking the resulting pool, traditionally using human assessors.