Activation Patching for Interpretable Steering in Music Generation

2025-04-06

AI Summary

Large audio foundation models lack interpretable and controllable representations for musical attributes (e.g., tempo, timbre) in text-to-music generation. Method: We adapt direction vector techniques from mechanistic interpretability to the audio domain, proposing a differential mean steering vector construction method based on residual stream activation patching—enabling continuous, fine-grained control over binary musical attributes. Through ablation studies varying scaling strength and injection layer position, we identify mid-to-late Transformer layers as critical for intervention; the learned steering vectors generalize across diverse text prompts. Contribution/Results: Our approach achieves significant improvements in precision for tempo and timbre control without degrading audio fidelity, establishing the first interpretable and intervenable representation manipulation framework for music generation. The method is model-agnostic, requires no retraining, and enables attribute-specific steering via simple vector arithmetic in latent space.

Technology Category

• Natural Language Processing: Generation
• Computer Vision: Diffusion Models for Vision
• Machine Learning: Mixture of Experts (MoE)

Application Category

• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Understanding how large audio models represent music, and using that understanding to steer generation, is both challenging and underexplored. Inspired by mechanistic interpretability in language models, where direction vectors in transformer residual streams are key to model analysis and control, we investigate similar techniques in the audio domain. This paper presents the first study of latent direction vectors in large audio models and their use for continuous control of musical attributes in text-to-music generation. Focusing on binary concepts like tempo (fast vs. slow) and timbre (bright vs. dark), we compute steering vectors using the difference-in-means method on curated prompt sets. These vectors, scaled by a coefficient and injected into intermediate activations, allow fine-grained modulation of specific musical traits while preserving overall audio quality. We analyze the effect of steering strength, compare injection strategies, and identify layers with the greatest influence. Our findings highlight the promise of direction-based steering as a more mechanistic and interpretable approach to controllable music generation.

Problem

• Understanding music representation in large audio models
• Steering music generation using latent direction vectors
• Controlling musical attributes in text-to-music generation

Innovation

• Latent direction vectors control music attributes
• Difference-in-means method computes steering vectors
• Activation injection preserves audio quality