Music sequence models, which predict the next note from previous notes, often use the entire piece as history. This is computationally costly, and it is unclear how much history is actually needed.
A small model was trained on folk tunes and other music, with maximum history ranging from 6 to 336 notes. The study also tested if longer history helps because the model copies repeated passages exactly.
Most of the improvement from longer history came from exact copying. On folk tunes, prediction error dropped by 71% when using 336 notes versus 6, but on piano performances the drop was only 9%. This suggests that for repetitive music, a simple copy mechanism might be enough, saving computation.