Reproduce “Do Music Generative Models Understand Musical Qualities? Automatic Music Evaluation with Model-Intrinsic Signals” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaosha Li, Chun Liu, Ziyu Wang
The paper predicts human ratings of music from a generator’s expected tokens, surprise, confidence, and sparse-autoencoder features. It combines these traces with a rating-trained scorer and evaluates five benchmarks totaling 17,600 clips and 241 hours. On held-out tests, MusicGen-large SAE-only reaches Pearson r = 0.81 and Spearman rho = 0.80, aligning with benchmark ratings rather than universal musical quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.