Can AI Music Models Think Across Time? Probing MusicGen's Long-Horizon Planning

Most music generation AI models are suspected of stitching together locally plausible audio segments rather than genuinely planning musical structure over extended timeframes. This mechanistic interpretability study directly tests whether MusicGen contains internal neural features that actually track and enforce long-term musical patterns—a crucial question for understanding whether AI truly composes versus merely assembles. If successful, this research could unlock more sophisticated music generation and reveal fundamental principles about how generative models maintain coherence across long-horizon tasks.