Exploring the Internal Mechanisms of Music LLMs: A Study of Root and Quality via Probing and Intervention Techniques
Keywords: representation learning, music AI, interpretability
Abstract: How do music large language models (LLMs) interpret musical concepts? This study investigates the representational abilities of Music- Gen, a transformer-based music LLM, using single chords to assess how these models process structured musical entities. We have developed a novel probe-via-intervene approach to enhance our understanding of the model’s internal interpretability. Our findings indicate that although the model faces challenges in forming linearly separable representations for certain musical concepts such as chord quality, the integration of directional vectors from other musical concepts into the transformer’s residual stream substantially improves the probing results. Notably, significant enhancements are achieved by intervening in just one head across all layers. These insights underscore the differences between human and machine perception of music and suggest important considerations for future design of music LLMs.
Submission Number: 143
Loading