PAPER · 2026 · Sole author
Steering Music Generation via Activation Addition
The first rigorous statistical proof that a single musical concept can be steered into a 3.5-billion-parameter audio model, and that the effect is the concept itself, not noise.
Contrastive Activation Addition extracts a concept direction by contrasting concept-present and concept-absent activations at one transformer block, then adds it back at inference, no training, no fine-tuning. Across 2,250 generated clips, the piano direction's dose-response is significantly steeper than a matched-magnitude random vector on both evaluators (CLAP: β=0.035, p<1e-34; MuQ: β=0.047, p<1e-60 vs random β≈0.004, n.s.). That contrast is the evidence that steering is concept-specific rather than generic perturbation.
LIMITATION ·Single-layer steering (tf6 only; the tf7 hook did not transfer cleanly), a single random-vector draw per concept, and no human listening evaluation yet. All deferred to future work.
Accepted at ISMIR 2026. Presented in Abu Dhabi, November 2026. Link added on publication.