BeatEdit: Symbolic Music Generation as Explicit Editing

Music creation is fundamentally a process of revision, yet mainstream symbolic music generation paradigms produce complete sequences from scratch with no mechanism for selective modification. In other modalities, analogous tasks are routinely solved by edit-based methods that predict explicit per-token operations, yet such methods remain largely unexplored for symbolic music. We trace this absence to a representational barrier: conventional music encodings lack the structural properties that editing demands. The Beat encoding, originally designed for autoregressive generation, naturally satisfies editing requirements. Building on this, we propose BeatEdit, the first framework for symbolic music generation based on explicit edit operations, with three complementary mechanisms along an edit-complexity spectrum: per-token sequence tagging for error correction, tag-then-fill for segment completion, and iterative refinement for accompaniment editing. All these tasks share an encoding and pre-trained backbone, with music-specific adaptations for cascading pitch dependencies and encoding invariant maintenance. Experiments on 192K piano pieces show that Beat encoding substantially improves correction recovery over alternative encodings under the same architecture, and that edit-based methods outperform autoregressive and diffusion baselines while achieving over 300× faster inference, enabling real-time interactive editing. Cross-evaluation across methods and tasks reveals three-way complementarity and shows that encoding choice significantly influences method rankings, while per-difficulty analysis reveals that editing methods degrade when tasks demand substantial content regeneration. Preliminary multi-track experiments further confirm generalization beyond the piano domain, establishing it as a general-purpose approach to symbolic music editing.