Music transformers' harmonic representations are localized and steerable
altryne · x · 2026-10-15
An ISMIR 2026 paper by Silberschein, Wei & Rott Shaham studies how a LLaMA-based symbolic music transformer (Aria) internally represents harmony, evaluated on Bach chorales. Harmonic decisions are causally localized progressively from middle layers onward; linear probes show mode becomes decodable in the same layer range where activation patching reveals growing causal effect, while key identity emerges later. Crucially, the representations are actionable: steering along latent directions for harmonic changes reliably produces deceptive or minor cadence resolutions while preserving structural coherence — demonstrating direct inference-time control of a model's harmonic output.
More from Research
- OSVerify system finds two substantial issues in OpenAI's quantum geometric Langlands paper — james_y_zou · 2026-10-16
- Intelligence is proportional to data: reorder tasks by data volume and the jagged frontier becomes a spiral — sebkrier · 2026-10-16
- Doc2Harness: New benchmark finds best coding agent passes only 53% across 329 cases — _akhaliq · 2026-10-16
- Stanford's Fei-Fei Li team open-sources OpenWAM, a unified world-action model framework for robots — jiqizhixin · 2026-10-16
- NVIDIA's TacitVLA: robots reason without reading their own thoughts, 3.4x lower latency — arankomatsuzaki · 2026-10-16
- Cohere Labs world model course session 4: DreamerV1-V4 architectures in one go — Cohere_Labs · 2026-10-16