Researchers have found that GPT-style models, which excel at processing language, don't automatically work well for symbolic music. This is because tokenization, a key component of these models, relies on compression in a specific coordinate system. The team proposes a framework to identify effective coordinates and shows that simply compacting sequences isn't enough for predictive compression. Instead, the model should preserve contextual freedom to allow higher-order musical organization to emerge. This study highlights why GPT-style models don't directly transfer across modalities: it's not just about the architecture, but also about how data is represented. --- Why it matters: This research matters because it reveals a fundamental limitation of current AI architectures when applied to different domains. It shows that simply transferring successful language models to music or other areas won't work without rethinking tokenization and representation. This has implications for future AI development and the design of more generalizable models. Source: https://arxiv.org/abs/2608.18025