A research team has released LeVo 2, an AI system that generates full-length songs by splitting the hard problem of music into a hierarchy of smaller ones.
Most language model-based music systems face a structural choice: blend vocal and instrumental tokens together for better coordination, or separate them for cleaner acoustics — but not both. LeVo 2 sidesteps this by using a two-stage language model. A planning layer predicts mixed tokens first to handle global structure, then generates vocal and accompaniment tokens in parallel for track-level detail. A diffusion-based codec handles the final waveform reconstruction. The result is a system that can follow lyrics and style prompts while maintaining coherence across a full song.
The more interesting contribution may be the training pipeline. The team calls it progressive post-training: supervised fine-tuning improves raw generation quality, large-scale offline DPO sharpens controllability, and closed-loop semi-online DPO — which feeds the model's own outputs back as a training signal — targets musicality specifically. Each stage tackles a different failure mode, rather than asking one optimization pass to do everything at once. That separation matters because aesthetic quality and technical controllability tend to pull in opposite directions during training.
In listening tests, LeVo 2 outperformed open-source baselines across six subjective dimensions and came close to leading commercial systems on several metrics — though "approaches" is doing real work in that sentence, and the gap to proprietary tools like Suno or Udio isn't quantified here.