Music Playlist Captioning at Scale with Large Language Models

As music streaming services like Deezer rely on personalized playlists (e.g., Daily Mix) to drive engagement, a major challenge is helping users understand the content of these recommendations. While traditional approaches use generic titles or limited visual cues, this paper investigates whether automatically generating descriptive, natural-language captions for playlists can improve user interaction and perception.

The researchers developed a production-scale pipeline that leverages Large Language Models (LLMs) to generate playlist captions. Instead of captioning millions of individual user playlists, the system operates on approximately 5,000 pre-defined 'artist clusters'—the building blocks of the Daily Mix feature. The pipeline aggregates weighted descriptors (genres, moods, countries, and user-generated playlist titles) for these clusters and feeds them into an LLM (Gemini 2.0 Flash) with a carefully crafted prompt. The system includes strict safety checks and an 'LLM-as-a-Judge' validation step to ensure caption quality and faithfulness before deployment.

The system was evaluated via a large-scale A/B test involving millions of users. The results demonstrated that adding descriptive captions led to a 24.9% increase in playlist adoption, a 16.9% increase in reconnection, and an 11.5% increase in user satisfaction. Crucially, because the underlying recommendation algorithm remained unchanged, these gains are attributed entirely to 'semantic framing'—the way descriptive language shapes how users perceive and interact with the same content.

This study provides empirical evidence that the presentation layer of a recommender system is as critical as the recommendation algorithm itself. By using LLMs to provide context, streaming services can significantly improve user engagement without needing to overhaul their core recommendation logic. The findings also highlight the practical utility of LLMs in industrial settings, provided they are paired with robust validation and quality control mechanisms.