Extend activation-level steering methods in music to time-varying music-theoretic control

Determine whether activation-level steering approaches for music generation, such as Activation Patching for Interpretable Steering in Music Generation and Smitin (Self-monitored inference-time intervention for generative music transformers), can be extended to provide time-varying, music-theoretic control over generated audio content.

Background

The paper situates MusicRFM within a broader landscape of activation-level steering research. In LLMs, several techniques steer internal activations to change style or sentiment without retraining, while in music generation, existing work has largely focused on binary attributes or broad controls such as instrument presence.

The authors explicitly note uncertainty regarding whether these music-focused activation-steering approaches can achieve fine-grained, time-varying control aligned with music theory (e.g., notes, chords, tempo). MusicRFM is proposed to address this gap by discovering and injecting concept-aligned directions in a frozen MusicGen model, but the general question of extending prior activation-steering methods remains explicitly open.