Closing the Loop: PID Feedback Control for Interpretable Activation Steering in Symbolic Music Generation
It addresses the need for fine-grained, interpretable control over discrete signal attributes in transformer-based music generation, offering a method for deterministic modulation without retraining.
The paper introduces a framework for interpretable activation steering in symbolic music generation, achieving high correlation between steering magnitude and attribute shift for Pitch and Duration, and reducing interference in multi-attribute control via Gram-Schmidt orthogonalization.
Transformer-based architectures have significantly advanced the generation of complex symbolic sequences, yet a significant gap remains in achieving fine-grained, interpretable control over discrete signal attributes. This paper investigates the mechanistic interpretability of the Multitrack Music Transformer (MMT) and proposes a framework for deterministic attribute modulation without retraining to bridge this gap via inference-time activation steering. Utilizing the Difference-in-Means (DiffMean) methodology, we isolate latent directions for signal attributes, specifically Pitch and Duration, within the residual stream. We validate the Linear Representation Hypothesis in this domain, achieving high correlation between steering magnitude and attribute shift. To address the inherent feature entanglement in multi-attribute steering, we introduce a Dual Steering framework utilizing Gram-Schmidt Orthogonalization. Experimental results demonstrate that this geometric decoupling reduces conceptual interference and signal degradation compared to naive vector addition, enabling independent deterministic control even against strong autoregressive conditioning.