metadata
license: apache-2.0
tags:
- audio
- music
- reward-model
- audio-language-model
MuseCritic
MuseCritic is introduced in MuseCritic: Learning Multi-Aspect Song Rewards through Natural-Language Aesthetic Critiques.
MuseCritic is a semi-scalar reward model for long-form song evaluation. Given
a song and an aesthetic evaluation rubric, the model first generates a
natural-language critique covering five dimensions and then predicts five
continuous scores in the interval (1, 5):
- Overall coherence
- Musicality
- Memorability
- Structural clarity
- Vocal naturalness
Resources
- Code: WuqnEl/MuseCritic
- Paper: arXiv:2608.11755
Please refer to the GitHub repository for installation, inference, training, and evaluation instructions.
Citation
If you find MuseCritic useful, please cite:
@misc{zhuang2026musecriticlearningmultiaspectsong,
title={MuseCritic: Learning Multi-Aspect Song Rewards through Natural-Language Aesthetic Critiques},
author={Jiabao Zhuang and Changhao Jiang and Hanchen Wang and Jiahao Chen and Zhixiong Yang and Zhenghao Xiang and Yifei Cao and Jiajun Sun and Hui Li and Ming Zhang and Tao Ji and Tao Gui and Qi Zhang and Xuanjing Huang},
year={2026},
eprint={2608.11755},
archivePrefix={arXiv},
primaryClass={cs.SD},
url={https://arxiv.org/abs/2608.11755},
}