What YuE2 tries to solve is not only generation quality, but the one-shot gacha problem for which AI music has long been criticized. Its core mechanism is symbolic planning: before generating audio, the model first plans a two-track score draft containing vocal melody, instrumental melody, chords, and section structure, saves it in ABC text notation form, and then generates the final sound based on it.
This mechanism makes melody and harmony no longer black-box results, but something that can be located and modified, thereby supporting Agentic Music Editing. The example given by the source is: a user can use natural language to request keeping the chorus vocal melody, inserting a "Twinkle Twinkle Little Star" saxophone solo, and making the harmony more modern; an Agent connected to YuE2 will read the original song's melody and chords, propose modification plans combined with music theory, locate and rewrite the corresponding chords or notes, and then call YuE2 to generate a new version. For professional musicians, the same mechanism also allows changing only one chord (for example, to a diminished seventh chord) while preserving the unmodified melody and harmony as much as possible.
The experimental data for editing capability is: after targeted modification of melody and chords in a complete work, the target pitch match rate rises from 0.83% to 93.75%, and the target chord consistency rate rises from 0 to 83.13%; the melody and chord similarity of unmodified parts remain above 0.96 and 0.93 respectively. These are metrics from the team's experiments, showing that changes can enter the sound while the rest is relatively well preserved, but they are still source-side self-evaluation.