BeatEdit

Symbolic Music Generation as Explicit Editing

Haoyu Gu1, Lekai Qian1, Haowu Zhou1, Qi Liu1,*, Shuai Wang2,*

1 School of Future Technology, South China University of Technology
2 School of Intelligence Science and Technology, Nanjing University
* Corresponding authors
Official ACM Multimedia 2026 conference logo
South China University of Technology
Nanjing University

01Why explicit editing?

Music creation is a process of revision.

Correct a wrong note

Change the error; retain the phrase.

Reshape accompaniment

Revise the left hand; preserve the melody.

Complete missing bars

Write new content into existing context.

Source and target overlap. The editor must locate what needs to change and preserve the rest.

02The bottleneck is the representation: R1–R3

R1 · Atomic edit units

Each note should map to a compact, self-contained token group, allowing one edit label to address a complete musical target.

R2 · Source–target alignment

Source and target need stable positional correspondence, so edit labels can be extracted by directly comparing aligned units.

R3 · Edit locality

An edit should affect a bounded token span and preserve a valid encoding, without updating the rest of the sequence.

EncodingR1R2R3Bottleneck
MIDI-Like×××Onset/offset decoupled
REMI×××Insert/delete shifts positions
Structured∼×∼Relative time propagates
CPWord∼×∼No structural grid
BEAT (Rel.)✓✓∼Within-beat pitch cascade
BEAT (Abs.)✓✓✓Meets all three requirements

✓ satisfied · × violated · ∼ partial.

03BEAT supplies a beat-aligned editing space

BEAT compresses within-beat rhythm into ternary patterns and aligns source and target on a fixed beat grid.
Paper figure: piano roll, ternary patterns, beat-level assembly and voice interleaving

04Choose the mechanism by edit density

SeqTag

Sparse · Error correction

Detect erroneous tokens and assign edit tags, keeping low-confidence positions unchanged.

IterEdit

Dense · Accompaniment editing

Delete–insert–predict cycles revise accompaniment while preserving melody and structure.

TagFill

Missing spans · Segment completion

Allocate mask positions, then fill them in two confidence-ranked passes while preserving surrounding musical context.

05Results: accuracy, latency, and encoding

Mechanism and encoding both matter: explicit editing improves accuracy and speed; edit-compatible encoding improves recovery. Correction/editing: beat exact match; completion: POP909 Note F1; encoding: recovery.

Correction1.84×
SeqTag .726 vs AR-Detect .394
65 vs 9,677 ms
Editing1.73×
IterEdit .480 vs AR-Detect .278
655 vs 9,458 ms
Completion4.65×
TagFill-D .279 · Polyffusion .060
Note F1 · POP909 · 35 vs 1,747 ms
Encoding4.81×
BEAT 48.1% vs REMI 10.0%
Recovery · Same SeqTag