Towards Effective Updating of Pretrained Symbolic Music Models for Fine-Grained Bar-Level Control

Yangyang Shu, Haiming Xu, Ziqin Zhou, Anton van den Hengel, Lingqiao Liu

Keywords: Symbolic Music Model, Fine-Grained Bar-Level Control

Abstract: Automatically generating symbolic music scores tailored to specific user needs offers significant benefits for musicians and enthusiasts alike. Pretrained symbolic music autoregressive models have demonstrated promising results, thanks to large datasets and advanced transformer architectures. However, in practice, the control provided by such models is often limited, particularly when fine-grained controls are needed at the level of individual bars. While fine-tuning the model with newly introduced control tokens may seem like a straightforward solution, our research reveals challenges in this approach, as the model frequently struggles to respond effectively to these precise bar-level control signals. To overcome this issue, we propose two novel strategies. First, we introduce a pre-training task that explicitly links control signals with their corresponding musical tokens, enabling a more effective initialization for fine-tuning. Second, we develop a unique counterfactual loss function that enhances alignment between the generated music and the specified control prompts. These combined methods substantially improve bar-level control, yielding a 13.06\% improvement over the fine-tuning baseline. Importantly, subjective evaluations confirm that this increased control does not compromise the musical quality produced by the original pretrained model.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5376