Beyond Musical Descriptors: Extracting Preference-Bearing Intent in Music Queries

Researchers introduced MusicRecoIntent, a new benchmark dataset of 2,291 annotated Reddit music requests. It labels musical descriptors across seven categories with positive, negative, or referential user intent. The study found that while large language models (LLMs) can capture explicit musical descriptors, they significantly struggle with context-dependent ones that reveal a user's true preference. This highlights a key limitation in current AI-based music understanding and recommendation systems.