Find the Neurons, Control the Feature? Selection Strategies for Network Bending Audio Models

Neural audio synthesisers like RAVE enable high-quality generation, but designing effective control interfaces requires understanding where musical features are encoded and whether those encodings can be manipulated. We present evidence that these questions yield different answers: neurons that "know about" a feature are not necessarily the neurons that control it. Through analysis of three RAVE models, we identify functionally-coherent neuron groups that encode BPM more strongly than individual layers. However, manipulation experiments across 288,000 samples find little compelling evidence for a relationship between control and encoding (all |ρ| ≤ 0.16, n.s.). Encoding strength instead predicts effect magnitude, with the direction feature-dependent. We find strongly BPM-encoding groups produce larger effects (cross-layer clusters ρ = +0.24; layer-wise ρ = +0.50), while strongly pitch-encoding layers produce smaller effects (layer-wise ρ = −0.28). Scaling whole layers achieves superior control, emerging as the best method across over 6× as many feature/section combinations as clustering, whereas clusters provide nearly 4× better efficiency and stronger magnitude.

• https://aimc2026.org/proceedings