1Tsinghua · 2HK PolyU · 3BTBU · 4NTU · 5SYSU · 6HKU · 7UCAS · 8PKU · 9UESTC · 10SparcAI
† Corresponding authors
Listen to what the representation buys. Every clip is a single unedited generation conditioned only on the caption shown. The captions come from a public benchmark split disjoint from our training data and unlike our own captions, so all of it is zero-shot with respect to caption domain.
A short listening set: 12 of the 31 pieces a human listener kept while marking 52 of a pool of 254 generations. Selection is by ear, not by score: we had ranked candidates by measurable proxies for coherence (tonal clarity, consonance, onsets on a grid, self-similarity, pitch variety), and against the listener's own keep/skip marks that composite is close to uninformative, so we stopped using it to choose. It is therefore the strong end of our output and not a sample from it. 8 were prompted with captions written for this page; 3 come from the controlled-grid split and 1 from the public benchmark split, both zero-shot. Each block states which model produced it. The other sections are not curated this way, each states the fixed rule it used.
A romantic solo piano nocturne in a minor key with a flowing broken-chord accompaniment and an expressive singing melody, rubato, intimate and dark.
Engraved, from the first sounding bar
Acoustic Grand Piano
A slow chorale for string ensemble and organ with continuous four-part sustained harmony, no rests, deep and reverent.
Engraved, from the first sounding bar
Acoustic Grand Piano, Recorder
Ye Banks And Braes, a waltz. The audio clip opens with a solo piano performance of a gentle, lyrical melody presented over simple arpeggiated chords in the left hand. The theme is structured into two
Engraved, from the first sounding bar
Acoustic Grand Piano
A bright and yearning film-score cue for piano, strings and soft synth pad, with a rising chord progression and a lyrical melody that climbs into a warm climax, uplifting and cinematic.
Engraved, from the first sounding bar
Acoustic Grand Piano, Viola
A piano and strings piece with sustained block chords in the strings and rolling arpeggiated chords in the piano, harmony sounding at all times, romantic and expansive.
Engraved, from the first sounding bar
Acoustic Grand Piano, Electric Guitar (clean)
Road To Tralee, The, a slide. The audio clip begins with a solo piano performance featuring a slow to moderate tempo, set within an intimate, dry acoustic space free from reverb or background noise. T
Engraved, from the first sounding bar
Acoustic Grand Piano
A chamber arrangement for piano, cello and violin with an interlocking arpeggio accompaniment and a long-breathed melody, intimate and cinematic.
Engraved, from the first sounding bar
Acoustic Bass, Pizzicato Strings, Trombone, Flute, Cello, Program 114, French Horn, Drum Kit
An orchestral cue where piano, strings, brass and choir pad all sustain chords simultaneously, very thick seven-voice writing, majestic and cinematic.
Engraved, from the first sounding bar
Acoustic Grand Piano
Beoga!, a reel. The audio clip opens with a solo digital piano performance featuring bright, bell-like timbres characteristic of a synthesized instrument, likely sampled from a real grand piano but la
Engraved, from the first sounding bar
Acoustic Grand Piano
A delicate music-box melody over sustained strings, simple diatonic harmony, childlike and a little sad.
Engraved, from the first sounding bar
Acoustic Grand Piano, Drum Kit, Violin
A hymn arrangement for piano, organ and string ensemble in four-part harmony, all voices sounding together throughout, dignified and warm.
Engraved, from the first sounding bar
Acoustic Grand Piano
A slow-paced pop song with a touch of rock, this Christmas-themed composition exudes happiness and love through its melodic lines. Performed by a combination of piano, violin, string ensemble, drums, and clean electric guitar, it's set in the key of A minor with a 4/4 time signature. The chord progression of E7, Am, and Dm repeats frequently throughout the song, creating a sense of familiarity and comfort.
Engraved, from the first sounding bar
Acoustic Grand Piano, Drum Kit, Acoustic Bass, String Ensemble 1, Violin
Press play: the cursor tracks the audio and each note lights as it sounds, so the chord density is something you can watch as well as count. The engraved staves under each roll are the same MIDI put through a notation engraver.
| Example | Chord-time | Max poly | Instruments | Notes | Duration |
|---|---|---|---|---|---|
| EX 01 | 100% | 6 | 2 | 220 | 42 s |
| EX 02 | 90% | 9 | 5 | 186 | 28 s |
| EX 03 | 86% | 13 | 4 | 193 | 27 s |
| EX 04 | 92% | 6 | 4 | 171 | 27 s |
| EX 05 | 93% | 8 | 3 | 209 | 40 s |
| EX 06 | 89% | 4 | 4 | 174 | 27 s |
| EX 07 | 88% | 7 | 2 | 210 | 35 s |
| EX 08 | 99% | 9 | 1 | 253 | 41 s |
| EX 09 | 96% | 13 | 1 | 252 | 46 s |
| EX 10 | 95% | 10 | 1 | 232 | 49 s |
| EX 11 | 88% | 7 | 1 | 270 | 55 s |
| EX 12 | 84% | 11 | 1 | 229 | 90 s |
A meditative and relaxing electronic classical composition in F major, featuring acoustic guitar, string ensemble, piano, clean electric guitar, and synth strings. The piece maintains a 3/4 time signature and a Moderato tempo, with a recurring chord progression of Am, Dm, Cm, Am, and C7 that adds to its soothing ambiance.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Acoustic Guitar (nylon), Violin
A meditative and relaxing electronic ambient song that evokes a sense of space and dream-like atmosphere. Synth pads and piano create a soothing soundscape, with the piano playing chords in the key of Eb major. The piece maintains a steady 4/4 time signature and a moderato tempo of 93 beats per minute, guiding the listener through a tranquil journey.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Acoustic Grand Piano, Pad 1 (new age), Synth Strings 1, Electric Bass (finger), Pad 2 (warm)
A melodic pop song with a Christmas vibe, featuring a piano, harmonica, drums, string ensemble, and electric bass, set in the key of G major with a moderate tempo and 4/4 time signature, revolving around the chords D and G throughout its duration.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Drum Kit, Bright Acoustic Piano, Pad 1 (new age)
This lengthy electronic composition in C minor exudes a melodic and relaxing ambiance, with hints of a dreamy, meditative quality. The overdriven guitar, piano, and drums are accompanied by the ethereal tones of a choir and a whistle, creating a unique sonic tapestry. The piece maintains a steady 4/4 time signature and an Andante tempo, allowing the recurring chord progression of Ab, Gm7, C, Ab, and Gm7 to gently guide the listener through the musical journey.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Choir Aahs
A meditative and relaxing ambient piece with elements of easy listening, featuring a harmonica playing the lead melody, accompanied by a synth pad and an acoustic guitar. The song is in the key of C major with a moderate tempo of 120 beats per minute, and the main chord progression is C, D, and G. The track evokes a sense of emotional depth and has an epic, cinematic quality to it.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Acoustic Guitar (steel), Acoustic Bass, Harmonica
A captivating electronic composition infused with classical elements, this song evokes a cinematic atmosphere, inviting relaxation and introspection. With a moderate tempo and a 4/4 time signature, it unfolds gracefully in the key of C major, guided by a recurring chord progression of C, F, C, and G. The ensemble of piano, synth strings, string ensemble, French horn, and trumpet weaves a rich tapestry of sound, transporting the listener to epic and meditative realms.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Music Box, Acoustic Grand Piano, Celesta, Tuba
A melodic and relaxing classical piece, this short song features an acoustic guitar as the sole instrument. Set in the key of F major with a 4/4 time signature, it moves at an Allegro tempo. The chord progression of F, Abm6, and Em7b5 adds a touch of sophistication to the composition, while the overall mood evokes a sense of tranquility and love, making it well-suited for a film soundtrack.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Acoustic Guitar (nylon)
A melodic pop song in the key of G major, evoking a sense of relaxation and love. With a 4/4 time signature, it features a chord progression of G, D, Em, and C. The song's duration is 229 seconds, making it a full-length piece that could be well-suited for a film soundtrack.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Acoustic Grand Piano
A melodic electronic song with a moderate tempo, set in the key of G minor and featuring a chord progression of D/A, Gm, Cm, Gm, and Eb, creating an atmosphere suitable for film and evoking feelings of love, energy, and relaxation throughout its duration.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Acoustic Grand Piano
A slow and melodic electronic ambient song featuring a piano, set in the key of B minor with a 4/4 time signature and an Andante tempo, evoking a relaxing, dark, and spacey atmosphere.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Acoustic Grand Piano
A melodic pop ballad that evokes feelings of love and emotion. Set in the key of C# major with a 4/4 time signature, the song maintains an Allegro tempo. The chord progression features Ab, Bbm7, Ab7/C, and C#.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Acoustic Grand Piano
A short, melodic electronic piece with a touch of pop, featuring a piano that sets a film-like, slightly dark, and emotional atmosphere. The song is in E minor, has a 4/4 time signature, and maintains an Allegro tempo throughout its brief duration.
Engraved, from the first sounding bar · notation quantized to 16ths for legibility
Acoustic Grand Piano
One player, seven sources. Switching system keeps the playback position, so you can flip between systems inside the same phrase, which is the only way a difference becomes audible.
How the captions were chosen. By the chord-richness of the real reference piece, a rule blind to what any system produced, taking a block from the top of that ranking. Caption 16 is included because we lose there: our output reaches max polyphony 7 against the reference's 18 while one baseline matches it exactly. Removing it would turn this into a reel of wins.
Read Δ, not chord-time. Higher chord-time is not better; distance from the reference is. Several baselines reach high chord-time by over-texturing, and one emits a fixed note count for every caption regardless of what was asked.
A meditative and relaxing electronic ambient song that evokes a sense of space and dream-like atmosphere. Synth pads and piano create a soothing soundscape, with the piano playing chords in the key of Eb major. The piece maintains a steady 4/4 time signature and a moderato tempo of 93 beats per minute, guiding the listener through a tranquil journey.
Switching system keeps the playback position, so you can flip mid-phrase. ← → switch · space play
Texture against the reference · Δ is max polyphony minus the reference's
| System | Chord-time | Max poly | Δ | Notes | Instr. |
|---|---|---|---|---|---|
| Real reference | 100% | 12 | — | 256 | 1 |
| Ours | 90% | 9 | -3 | 186 | 5 |
| MIDI-LLM | 94% | 15 | +3 | 241 | 2 |
| MIDILM | 98% | 15 | +3 | 469 | 5 |
| Amadeus | 100% | 6 | -6 | 2046 | 2 |
| text2midi | 89% | 28 | +16 | 234 | 19 |
| ChatMusician | 0% | 1 | -11 | 68 | 1 |
Source MIDI, every system
↓ Real reference↓ Ours↓ MIDI-LLM↓ MIDILM↓ Amadeus↓ text2midi↓ ChatMusician
A slow and meditative pop Christmas song in G major, featuring a melodic and relaxing atmosphere. With a 4/4 time signature, the piece is characterized by a chord sequence of G, G7, Cmaj7, and C6. Its soothing ambiance evokes a sense of love and tranquility throughout the song's duration.
Switching system keeps the playback position, so you can flip mid-phrase. ← → switch · space play
Texture against the reference · Δ is max polyphony minus the reference's
| System | Chord-time | Max poly | Δ | Notes | Instr. |
|---|---|---|---|---|---|
| Real reference | 100% | 14 | — | 256 | 1 |
| Ours | 86% | 15 | +1 | 188 | 5 |
| MIDI-LLM | 100% | 21 | +7 | 333 | 6 |
| MIDILM | 62% | 6 | -8 | 435 | 2 |
| Amadeus | 98% | 5 | -9 | 2046 | 1 |
| text2midi | 52% | 31 | +17 | 233 | 13 |
| ChatMusician | 0% | 1 | -13 | 83 | 1 |
Source MIDI, every system
↓ Real reference↓ Ours↓ MIDI-LLM↓ MIDILM↓ Amadeus↓ text2midi↓ ChatMusician
A slow-paced classical soundtrack piece in 6/8 time signature, set in the key of C minor, featuring a blend of synth effects, oboe, pizzicato strings, bassoon, and clarinet, creating a cinematic atmosphere suitable for films, documentaries, or adventure scenes, with a subtle Christmas undertone.
Switching system keeps the playback position, so you can flip mid-phrase. ← → switch · space play
Texture against the reference · Δ is max polyphony minus the reference's
| System | Chord-time | Max poly | Δ | Notes | Instr. |
|---|---|---|---|---|---|
| Real reference | 100% | 6 | — | 256 | 1 |
| Ours | 89% | 11 | +5 | 111 | 4 |
| MIDI-LLM | 85% | 15 | +9 | 328 | 6 |
| MIDILM | 62% | 201 | +195 | 490 | 9 |
| Amadeus | 100% | 16 | +10 | 2046 | 8 |
| text2midi | 53% | 27 | +21 | 232 | 16 |
| ChatMusician | 0% | 1 | -5 | 118 | 1 |
Source MIDI, every system
↓ Real reference↓ Ours↓ MIDI-LLM↓ MIDILM↓ Amadeus↓ text2midi↓ ChatMusician
A melodic pop ballad that evokes feelings of love and emotion. Set in the key of C# major with a 4/4 time signature, the song maintains an Allegro tempo. The chord progression features Ab, Bbm7, Ab7/C, and C#.
Switching system keeps the playback position, so you can flip mid-phrase. ← → switch · space play
Texture against the reference · Δ is max polyphony minus the reference's
| System | Chord-time | Max poly | Δ | Notes | Instr. |
|---|---|---|---|---|---|
| Real reference | 100% | 18 | — | 256 | 1 |
| Ours | 88% | 7 | -11 | 270 | 1 |
| MIDI-LLM | 27% | 6 | -12 | 332 | 1 |
| MIDILM | 51% | 7 | -11 | 427 | 1 |
| Amadeus | 100% | 5 | -13 | 2046 | 1 |
| text2midi | 93% | 18 | +0 | 223 | 7 |
| ChatMusician | 27% | 6 | -12 | 578 | 3 |
Source MIDI, every system
↓ Real reference↓ Ours↓ MIDI-LLM↓ MIDILM↓ Amadeus↓ text2midi↓ ChatMusician
Our weakest axis is doing what the caption explicitly asks. Across the 126 captions in this test split that name a key, the generated piece is in that exact key and mode 17% of the time. The three cases below are real misses, and they are picked to be unambiguous: keys that are not the relative or parallel of what was asked, so they cannot be excused as near misses. Roughly twice as many captions land on a related key, which is the gap between the two key metrics in the paper.
Two things this is not. It is not tonal incompetence: the pieces are internally consistent in some key, just not the requested one, and a decode-time constraint that restricts pitches and programs to what the caption names doubles both key and instrument agreement without retraining and without changing distributional fidelity. And it is not the axis the paper's main claim rests on, which is distance to the real note distribution. We show these because a page of successes would misrepresent where the method stands.
This uplifting rock and electronic fusion song, set in the key of Eb major, features a lively interplay between the Hammond organ, clean electric guitar, and alto saxophone, punctuated by orchestra hits and propelled by a steady drum beat. With a tempo of 114 BPM and a 4/4 time signature, the piece maintains an energetic Allegro pace throughout its duration. The recurring chord progression of Eb/F, F, Eb/F, and Bb forms the harmonic backbone of this melodic and motivational composition.
An energetic electronic song with a retro vibe, featuring overdriven and distorted guitars, synth lead, drums, and rock organ, all contributing to its melodic and upbeat atmosphere. Set in the key of F# major with a 4/4 time signature, the track maintains an Allegro tempo throughout, with the chords B and F# repeating frequently.
A melodic and relaxing electronic ambient song with a moderate tempo of 114 bpm, featuring drums, piano, electric bass, synth strings, and flute. Set in the key of C minor and a 4/4 time signature, the piece revolves around the frequently recurring chord progression of Cm and G7/B, creating a dreamy and slightly uplifting atmosphere.
No model at all. One real human performance is encoded and then decoded by each representation, so there is no sampling and nothing to select: the difference is what the format can physically hold.
The ladder under each arm draws every onset in the original down to where that tokenizer put it. A vertical line means the timing survived; a slant means it was moved. The beat grids lean.
The error accumulates, which is worth reading carefully. Our representation stores the raw interval between onsets on a fixed 10 ms lattice, so its error is the same at the start and at the end. A beat grid stores positions relative to a tempo, so small rounding compounds as the piece goes on: in the first three seconds it is close, and by twelve seconds it is off by more than a hundred milliseconds. That is why the first-6-s median printed above each ladder is several times smaller than the whole-clip median, and it is the reason we report the whole-clip figure in the paper.
Where each onset moved · first 6 s · median 2.5 ms here, 2.3 ms over the whole clip
Where each onset moved · first 6 s · median 16.0 ms here, 85.9 ms over the whole clip
Where each onset moved · first 6 s · median 16.0 ms here, 85.9 ms over the whole clip
Median onset error against elapsed time · all three representations
No model is involved: the performance is encoded and decoded, nothing more. ← → switch · space play
Where each onset moved · first 6 s · median 5.0 ms here, 4.5 ms over the whole clip
Where each onset moved · first 6 s · median 27.0 ms here, 86.4 ms over the whole clip
Where each onset moved · first 6 s · median 27.0 ms here, 86.4 ms over the whole clip
Median onset error against elapsed time · all three representations
No model is involved: the performance is encoded and decoded, nothing more. ← → switch · space play
Where each onset moved · first 6 s · median 0.0 ms here, 0.0 ms over the whole clip
Where each onset moved · first 6 s · median 30.5 ms here, 82.6 ms over the whole clip
Where each onset moved · first 6 s · median 30.5 ms here, 82.6 ms over the whole clip
Median onset error against elapsed time · all three representations
No model is involved: the performance is encoded and decoded, nothing more. ← → switch · space play
@misc{chen2026agogicperformancetimedmusictokens,
title={Agogic: Performance-Timed Music Tokens for LLM-Native Text-to-Symbolic-Music Generation},
author={Junhao Chen and Mingjin Chen and Jingjia Mao and Lin Chen and Saining Zhang and Minglin Chen and Ruocheng Wu and Liaoyuan Fan and Wenyi Li and Mingju Gao and Henghaofan Zhang and Zhihao Li and Hao Zhao and Yufei Wang and Ruqi Huang},
year={2026},
eprint={2608.03999},
archivePrefix={arXiv},
primaryClass={cs.SD},
url={https://arxiv.org/abs/2608.03999},
}
Paper: arXiv:2608.03999 · Code: github.com/SparcAI-Inc/Agogic · Project page: yisuanwang.github.io/Agogic