Latent Space Disentanglement via Activation Steering for Interpretable Attribute Control in Symbolic Music Generation

2026-05-29

AI Summary

This work addresses the challenge of fine-grained and interpretable control over discrete musical attributes—such as pitch and duration—in symbolic music generation. The authors propose a Dual Steering framework that leverages an activation steering mechanism during inference to locate latent directions associated with specific attributes within the residual stream of a multi-track music Transformer, enabling deterministic control without retraining. By combining Difference-in-Means to identify linear representation directions and Gram-Schmidt orthogonalization to disentangle entangled multi-attribute features, the method substantially reduces conceptual interference and signal degradation. Experimental results demonstrate that, even under strong autoregressive constraints, the approach achieves highly correlated yet independently controllable attribute shifts, significantly enhancing the interpretability and stability of generated music.

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

• Latent Space Disentanglement
• Interpretable Attribute Control
• Symbolic Music Generation
• Feature Entanglement
• Activation Steering

Innovation

• activation steering
• latent space disentanglement
• symbolic music generation
• Gram-Schmidt orthogonalization
• mechanistic interpretability