News · ai music — By WATMM Desk —
Does notation matter? A new paper says AI music models learn more from scores than from MIDI
A 25 Oct 2026 arXiv paper trains 26 language models on MusicXML, MIDI and ABC and reports that the MusicXML models do better at both understanding and generating music. For producers who work in MIDI, here is what that means and what it doesn't.
Key facts
- What happened: on 25 Oct 2026 Poltronieri, Spanio, Chieppa, Serra and Rocamora posted arXiv 2610.34247, which they describe as the first general-purpose tokenisation framework for MusicXML.
- How: they trained 26 autoregressive and masked language models on MusicXML, MIDI and ABC-based representations, tested with linear probes (composer, difficulty, accompaniment) and a human pairwise preference study.
- Why it matters: the authors report MusicXML models beat MIDI and text-based ones on both understanding and generation, a hint that the information MIDI throws away is useful to models.
Most AI tools that write notes for producers think in MIDI. A paper posted to arXiv on 25 October asks whether that is the right starting point, and its answer is no (arXiv 2610.34247).
What the paper did
Andrea Poltronieri, Matteo Spanio, Simone Chieppa, Xavier Serra and Martín Rocamora built what they call the first general-purpose tokenisation framework for MusicXML, the standard file format for exchanging notated scores (arXiv). Tokenisation is the step that turns music into the sequence of symbols a language model reads, the way text models read word pieces.
They then trained 26 autoregressive and masked language models and compared MusicXML against MIDI and ABC-based (plain-text notation) representations. They also tested several ways of tokenising MusicXML, including whether a tokeniser that knows musical grammar beats treating the score as plain text (arXiv).
Understanding was measured with linear probes on three tasks: composer classification, difficulty estimation and accompaniment suggestion. Generation was measured with statistical metrics and a human pairwise preference study. The headline: models trained on MusicXML tokens did better than MIDI and text-based ones on both understanding and generation (arXiv). The abstract gives no effect sizes, so how big the gap is needs the full paper.
Why a MIDI user should care
MIDI records what was played: note number, velocity, timing. A score also records how it is written: key and time signatures, whether a note is a C sharp or a D flat, beams, voices, ties, dynamics. Those are the differences in "notational richness" the authors set out to test (arXiv).
What we think
Our read is that this is a data story, not a format war. If the extra structure in a score helps a model, then the MIDI-first generators producers use today are learning from a thinner signal than they could. Expect tools that suggest chords, basslines or accompaniment to lean more on score data, then export MIDI to your DAW anyway. You will probably never touch MusicXML, but it may shape what lands in your piano roll.
Two cautions. First, this is one preprint, not yet peer reviewed, and the tasks lean towards notated, largely classical-style music; a beat-maker's quantised loop has little spelling or voicing to recover, so the advantage may shrink there. Second, MIDI 2.0 adds richer per-note data, which narrows part of the gap from the performance side; we covered how that works in our MIDI 2.0 explainer. For another recent look at how text-to-MIDI models work inside, see our piece on MI-MIDI and sparse autoencoders.
The practical takeaway for now: if you export MIDI from notation software, keep the score file too. It holds information your MIDI file does not.