AINeutralarXiv – CS AI · 7h ago6/10
🧠
Latent Space Disentanglement via Activation Steering for Interpretable Attribute Control in Symbolic Music Generation
Researchers propose a novel framework for controlling symbolic music generation in Transformer models through activation steering, enabling fine-grained control over musical attributes like pitch and duration without retraining. The approach uses latent space analysis and orthogonalization techniques to independently manipulate multiple attributes while reducing interference and maintaining generation quality.