MuseCritic: Learning Multi-Aspect Song Rewards through Natural-Language Aesthetic Critiques

2026-08-12

AI Summary

This work addresses the lack of interpretable, multi-dimensional aesthetic reward mechanisms in existing long-form music generation models, which hinders alignment with human aesthetic preferences. To this end, we propose MuseCritic—a semi-scalar reward model that, for the first time, incorporates natural language aesthetic critiques as an intermediate representation spanning five aesthetic dimensions. A two-stage training strategy is employed: supervised fine-tuning using critiques from a teacher model followed by reward learning with self-generated critiques, effectively mitigating distributional shift. Evaluated on SongEval, our approach achieves a macro-averaged MSE of 0.2316 (LCC=0.9068, SRCC=0.8838, Kendall’s τ=0.7178) and attains 71.35% accuracy in preference judgments on Music Arena. When integrated with GRPO, MuseCritic significantly enhances the performance of Muse-0.6B across nine aesthetic metrics.

Abstract

Long-form song generation models continue to improve in duration, structural integrity, and acoustic complexity, making reliable aesthetic rewards increasingly important for aligning these models with human preferences. However, reward models for complete songs remain limited, and existing evaluators typically predict scores in a single forward pass without providing readable explanations. We introduce MUSECRITIC, a semi-scalar reward model that generates a natural-language critique covering five aesthetic dimensions and uses it as an intermediate representation to predict continuous reward scores. MUSECRITIC follows a two-stage training pipeline: a teacher model first provides high-quality critiques for supervised fine-tuning, after which the fine-tuned model generates its own critiques for reward learning, mitigating distribution shift between training and inference. On an in-domain test set of 200 SongEval songs, MUSECRITIC reduces macro-averaged mean squared error from 0.2875 to 0.2316 and improves macro-averaged LCC, SRCC, and Kendall's tau to 0.9068, 0.8838, and 0.7178, respectively. On the out-of-domain Music Arena benchmark with 733 preference pairs, it achieves the highest accuracy of 71.35%. Moreover, using MUSECRITIC with GRPO improves Muse-0.6B on all nine aesthetic metrics from SongEval and Audiobox Aesthetics. These results demonstrate that critique-conditioned reward modeling reduces scoring error and provides an effective optimization signal for song generation. The project repository is available at https://github.com/WuqnEl/MuseCritic.

Problem

• song generation
• aesthetic evaluation
• reward modeling
• natural-language critique
• human preference alignment

Innovation

• critique-conditioned reward modeling
• natural-language aesthetic critique
• multi-aspect song evaluation
• semi-scalar reward model
• two-stage training pipeline