Researchers have developed a new method for training symbolic music language models to understand the context of entire compositions rather than just short excerpts. The approach, called Full-Horizon Compressed Recurrence (FHCR), reduces the memory requirements of recurrent neural networks while preserving their ability to capture long-range dependencies in musical structure. This is achieved through compression of key-value representations, allowing for efficient whole-piece training on limited GPU memory. Experiments on the MAESTRO dataset show that FHCR models perform better than truncated versions at capturing context beyond local segments. --- Why it matters: This matters because it enables more accurate and efficient modeling of musical structure, which is essential for applications such as music generation and analysis. By preserving long-range dependencies, FHCR can improve the quality of generated music and provide insights into the underlying structure of compositions. Source: https://arxiv.org/abs/2602.19816