LeVo 2: Stable and Melodious Song Generation via Hierarchical Representation Modeling and Progressive Post-Training

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Shun Lei, Huaicheng Zhang, Dapeng Wu, Yaoxun Xu, Lishi Zuo, Wei Tan, Hangting Chen, Guangzheng Li, Jianwei Yu, Zhiyong Wu, Dong Yu
• Venue: arXiv: cs.SD
• Published: 2026-06-29
• Collected: 2026-06-30
• URL: http://arxiv.org/abs/2606.30642v1
• PDF: https://arxiv.org/pdf/2606.30642v1
• DOI: UNKNOWN

Plain-Language Brief

The paper presents LeVo 2, a hybrid framework for generating full-length songs that balance coherence, musicality, and detailed acoustics. It uses a language model for semantic planning and a diffusion-based Music Codec for waveform reconstruction. The training schedule is designed to improve quality, controllability, and musicality through various methods, including an aesthetics-guided training schedule, progressive post-training, and modular extension. LeVo 2 outperforms open-source baselines and approaches commercial systems in listening tests.

What the Document Says

• LeVo 2 is a hybrid framework for generating full-length songs with coherence, musicality, and detailed acoustics.
• It uses a language model (LeLM) for semantic planning and a diffusion-based Music Codec for waveform reconstruction.
• The training schedule includes an aesthetics-guided training schedule, progressive post-training, and modular extension to improve quality, controllability, and musicality.
• Expert listening tests and objective evaluations show that LeVo 2 outperforms open-source baselines and approaches leading commercial systems.

Key Concepts Mentioned

• LeVo 2
• full-length song generation
• hierarchical modeling
• language model
• diffusion-based Music Codec
• aesthetics-guided training schedule
• progressive post-training
• modular extension

Claims Made by the Document

• LeVo 2 outperforms open-source baselines across six subjective dimensions.
• LeVo 2 approaches leading commercial systems on several listening metrics.

Evidence Used

• Expert listening tests and objective evaluations show that LeVo 2 outperforms open-source baselines and approaches leading commercial systems.

Methods / Approach

• Hierarchical modeling
• Language model for semantic planning
• Diffusion-based Music Codec for waveform reconstruction
• Aesthetics-guided training schedule
• Progressive post-training
• Modular extension

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

The paper does not discuss potential limitations of the LeVo 2 framework.

Open Questions

• What are the six subjective dimensions used for the evaluations?
• Which leading commercial systems were compared with LeVo 2?

Notes for Future Q&A

May include interpretation. Labeled where it does.

• The paper mentions six subjective dimensions but does not specify what these are.
• The paper compares LeVo 2 to leading commercial systems, but it's unclear which specific systems were used.

Raw Abstract / Excerpt

Full-length song generation must preserve coherence and musicality, render detailed vocal and accompaniment acoustics, and follow lyrics and prompts. Existing language model-based systems face a structural trade-off: mixed-token modeling preserves vocal-instrument coordination but obscures track-specific details, whereas dual-track prediction improves acoustics but requires longer sequences and weakens global planning. We present LeVo 2, a hybrid LLM-Diffusion framework for controllable full-length song generation. LeVo 2 formulates this trade-off as hierarchical modeling: LeLM first predicts mixed tokens for semantic planning, then predicts vocal and accompaniment tokens in parallel for track-specific refinement, while a diffusion-based Music Codec reconstructs full-length waveforms. A central contribution of this extended version is an aesthetics-guided training schedule for alignment. During pre-training, an automated music aesthetic evaluation framework assigns musicality-tier conditions to large-scale data, providing musicality priors before preference alignment. Progressive post-training applies SFT, large-scale offline DPO, and closed-loop semi-online DPO to separately improve generation quality, controllability, and musicality. Modular extension then trains the Track-Specific LM for acoustic refinement while preserving the aligned semantic planner. This schedule separates musicality learning, controllability alignment, and acoustic refinement, mitigating optimization conflict and the limitations of static offline preference pairs. Expert listening tests and objective evaluations show that LeVo 2 outperforms open-source baselines across six subjective dimensions, and approaches leading commercial systems on several listening metrics. Ablations validate the effects of the training strategy, aesthetics guidance, scaling, and hierarchical architecture.