Agogic: Performance-Timed Music Tokens for LLM-Native Text-to-Symbolic-Music Generation

This paper isolates music tokenization by holding pretrained Qwen3.5 backbones, data, training budget, and decoding fixed across seven representations. The authors report that representation dominates model size for distributional fidelity: a 0.8B model using the released Performance Music Tokens (PMT) reaches FMD 159, versus 272–286 for beat-grid tokenizations, and outperforms a 27B beat-grid model. PMT uses 10 ms timing, per-note velocity, multi-track texture, and a 609-symbol vocabulary. The work also reports replication on a 26M from-scratch backbone and another performance-resolution tokenizer, while explicitly leaving audible superiority to a preregistered human study.