Discovering Interpretable Concepts in Large Generative Music Models

The fidelity of generative models presents a scientific opportunity: these systems appear to have learned implicit theories of the content structure through statistical learning alone. This could offer a novel lens on theories of human-generated media. Where these representations align with traditional constructs (e.g. chord progressions in music), they demonstrate how these can be inferred from statistical regularities. Where they diverge, they highlight potential limits in our theoretical frameworks. In this paper, we focus on the specific case of music generators. We introduce an unsupervised method to discover musical concepts using sparse autoencoders (SAEs), extracting interpretable features from the residual stream activations of a transformer model. We evaluate this approach by extracting a large set of features and producing an automatic labeling and evaluation pipeline. Our results reveal both familiar musical concepts and counterintuitive patterns that lack clear counterparts in existing theories or natural language altogether. Our work provides a new empirical tool that might help discover organizing principles in ways that have eluded traditional methods of analysis and synthesis.