In the field of music generation, large language models can produce superficially legal twelve-tone scores but often fall into degenerate textures. To address this issue, we introduce a neuro-symbolic harness that wraps a language model proposer in a generate-verify-repair-trace loop, incorporating symbolic verification. This complete pipeline significantly improves event-local consistency without claiming whole-piece legality.
Across 40 controlled tasks and four paired models, the audited delivery yield increases from 13.3% under raw generation to 48.1% with the harness. Meanwhile, the pass rate of a narrower collision and serialization-consistency check rises from 33.5% to 58.3%, while degeneracy remains near 0.05, including under exploratory adversarial prompting. A blinded evaluation by five experts also shows a descriptive aggregate preference for harness candidates over raw generation in adherence, perceived legality, coherence, and overall quality.
Blogger's Review: This paper showcases the potential of neuro-symbolic methods in music generation. By implementing a generate-verify-repair loop, it effectively enhances the quality and consistency of generated works, marking a significant advancement in the field of music generation. This approach not only improves practical application outcomes but also provides new insights and directions for future musical creation.