Closing the Loop: PID Feedback Control for Interpretable Activation Steering in Symbolic Music Generation

2026-06-17

AI Summary

Existing Transformer-based approaches to symbolic music generation struggle to achieve fine-grained and interpretable control over discrete musical attributes such as pitch and duration. This work proposes a Dual Steering framework that enables deterministic, training-free attribute manipulation during inference by applying activation interventions along attribute-specific directions—identified via DiffMean—in the residual stream of a Multitrack Music Transformer. To mitigate feature entanglement among multiple attributes, the method incorporates Gram-Schmidt orthogonalization for disentanglement and employs a PID feedback mechanism to dynamically optimize intervention strength. Experimental results demonstrate that the proposed approach substantially reduces conceptual interference and signal degradation, achieving highly correlated and precise control over target attributes even under strong autoregressive constraints.

Abstract

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.

Problem

• interpretable control
• symbolic music generation
• activation steering
• feature entanglement
• deterministic modulation

Innovation

• activation steering
• mechanistic interpretability
• Gram-Schmidt orthogonalization
• symbolic music generation
• inference-time control