A 0.8B-parameter model beat a 27B model from the same family on music generation quality — not with more data, longer training, or a cleverer architecture, but by changing how music is encoded into tokens. This is the core finding of the Agogic paper, and it challenges one of AI's biggest beliefs: bigger is better.
A Skipped Variable
The text-to-music pipeline has many stages: backbone, data, training recipe, decoding — and finally, turning music into tokens. That last step is usually a default choice, never seriously ablated.
Agogic pins down every variable and changes only tokenization:
- Backbone: Qwen3.5, from 0.8B to 27B
- Data: identical
- Training budget: identical
- Decoding: identical
- Only variable: how music is tokenized
- PMT (the paper's new representation: 10ms timing precision, per-note dynamics, multi-track texture, 609 symbols): 0.8B model FMD = 159
- Beat grids (REMI and descendants, the current mainstream): 0.8B model FMD = 272–286
- Beat-grid family (REMI and descendants): snaps note onsets to beat grid points. Structurally stable, but quantization destroys micro-timing and per-note dynamics — precisely what separates a "performance" from a mechanical transcription.
- Text-ABC family: fuller information via textual scores, but poorly aligned with LLMs' natural-language pretraining.
Seven tokenizations were tested, with strikingly clean results.
Representation, Not Model Size, Is the Bottleneck
Lower Fréchet Music Distance (FMD) is better. PMT is 1.7–1.8x lower, up to 2.8x under some protocols.
In plain terms: a 0.8B model with PMT beats a 27B model with beat grids. Scaling the backbone from 0.8B to 27B (34x) barely moves FMD; switching tokenization cuts it roughly in half.
The paper's own words: "representation, not model size, is the bottleneck variable for distributional fidelity."
Why Beat Grids Lose
Not Luck, but a Category Effect
Three controls rule out "PMT just got a lucky vocabulary":
1. Smaller backbone: a 26M from-scratch model — PMT still wins. 2. A different performance-resolution tokenizer: the effect persists — it's the class of representation, not PMT specifically. 3. Not just finer resolution: snapping PMT's onsets to beat-grid resolution, PMT still leads by 67–129 FMD.
An Unsettling Side Finding: Imprinting
The "imprinting diagnostic" found:
> Published text-to-MIDI systems produce nearly identical chord-timing distributions across two unrelated domains — 72% vs 71%.
In other words: the model mostly ignores your prompt and regurgitates its training distribution. The appearance of prompt-following masks repetition, echoing findings elsewhere that current LLM "understanding" may be shallower than assumed.
A Lightweight Patch
Decode-time constraints: without them, instrument F1 = 0.28 and correct key = 0.16; with them, instrument F1 = 0.60 and correct key = 0.35 — both more than doubled, with no loss in distribution quality. Prompt-following and distributional fidelity are separable problems; you can fix one without touching the other.
What This Means
The transferable engineering principle: before spending money scaling, check which variable is actually the bottleneck.
Agogic offers a counterexample to "scale is all you need": on music generation, 34x more parameters did almost nothing, while changing representation halved the error. The lesson applies to any multi-variable system — run controlled ablations, find the bottleneck variable, then decide where to invest.
The paper also releases the full evaluation framework, 25+ checkpoints, and two corpora (86.6k aligned caption/MIDI/ABC/audio, plus 6.25M captioned — the largest annotated corpus in the music domain). Any future music representation can now be *measured, not asserted*. That may be the paper's biggest contribution: it turns representation choice from a matter of faith into a measurable question.
---
Paper: https://arxiv.org/abs/2608.03999
Code & data: evaluation framework, 25+ checkpoints, and two corpora as stated in the paper.