Composer Vector: Style-steering Symbolic Music Generation in a Latent Space

Symbolic music generation has recently achieved impressive progress, yet fine-grained and flexible control over composer style remains underexplored. Existing training-based approaches for style conditioning require large labeled datasets and are limited to single-composer generation, making them impractical for scenarios demanding blended or continuous stylistic control. In this work, we introduce \textbf{Composer Vector}, a training-free, inference-time steering method that enables controllable symbolic music generation. Experiments demonstrate that Composer Vector effectively shifts generations toward desired composer styles, provides interpretable and continuous control across steering coefficients, and highlights challenges in evaluating fused stylistic outputs. Our work shows that simple latent-space steering offers a powerful and efficient tool for controllable symbolic music generation, paving the way for more flexible and interactive creative applications.