Prompt-to-Slate: Diffusion Models for Prompt-Conditioned Slate Generation

Spotify Research introduces DMSG (Diffusion Model for Slate Generation), a novel approach that uses diffusion models to generate coherent item recommendations from natural language prompts. Unlike traditional methods that evaluate items individually, DMSG considers entire slates holistically, incorporating an encoding module for continuous embeddings, a conditioning module for prompt processing, and a fast diffusion process. Offline experiments on music playlists and e-commerce bundles showed significant improvements in relevance metrics (+17% NDCGSim) and diversity. A live A/B test with 1 million users demonstrated +6.8% increase in playlist curation and +10.5% in liked songs, while achieving 150ms P99 latency compared to 500ms for existing systems.