Harmonic representations in symbolic music transformers are localized and controllable
kastnerkyle · x · 2026-10-15
Research Question
How symbolic music transformers internally represent harmonic structure — using Aria, a LLaMA-based symbolic music transformer trained on Bach chorales — and whether that internal knowledge can steer generation.
Key Findings
- Causal localization: Harmonic decisions become causally localized progressively from middle layers onward. Linear probes show mode becomes decodable in the same layer range where patching finds growing causal effect, while key identity emerges later.
- Controllable generation: Steering via latent directions corresponding to harmonic changes reliably produces a deceptive or minor cadence while preserving structural coherence and quality.
Conclusion
The model's internal harmonic knowledge can be directly manipulated at inference time to control harmonic output. 9 pages, accepted at ISMIR 2026.
More from Multimodal
- Priming Models: Teacher-Free One-Step Generation Hits FID 1.19 With 12x Less Compute — du_yilun · 2026-10-15
- Same prompt, three takes: Claude vs ChatGPT vs Figma Make UI showdown — Tegadesigns · 2026-10-15
- Designer finds Opus 5.5 picks up nuanced aesthetic taste with startling speed — floguo · 2026-10-15
- The math: 5 rerolls at 480p plus a 4K upscale costs $5 vs $9 at 720p — HeyAmit_ · 2026-10-15
- Dreamina workflow: generate at 480p, upscale to 4K, cut video costs 44% — HeyAmit_ · 2026-10-15
- One-prompt GPT-6 Astra writes two Palestrina-style motets in LilyPond — deanwball · 2026-10-15