YuE2 puts editable symbolic planning at the centre of AI song generation
The system turns lyrics and musical style into a modifiable score before generating songs, while its developers report leading results across three evaluation areas.
YuE2 is a music-generation system that uses symbolic planning before producing a complete song with vocals and accompaniment. Its plan allows users to inspect or change elements including melody, rhythm, chords, lyrics, tempo and arrangement.
The developers report that YuE2 achieved a 6.9632 mean score in a 5 September 2026 WildSongBench comparison, ahead of Suno v5’s 6.8721. The evaluation covered 192 prompts and 15 system settings, with a “best-of-8” process selecting one of eight generations based on musicality, prompt control and lyric accuracy.
The project also reports results for related music models. MERT2 led on 14 of 15 MARBLE metrics for music representation, while SheetSage2 led on 10 of 13 transcription metrics. The comparisons included published baselines and task-specific systems, according to the YuE2 project page cited in the Hacker News source material.
The developers say their models were trained primarily on CC0 music and synthetic data, with most synthetic training data supplied under licence by Tokenwave.AI. They also state a commitment to the ethical and responsible use of data.
The benchmark results and “state-of-the-art” claims come from the developers and were not independently verified in the supplied material. The source does not establish statistical significance, reproducibility, licensing compliance or whether all comparisons used identical generation conditions.
