TuneJury: An Open Metric for Improving Music Generation Preference Alignment

A new open-source metric called TuneJury aims to standardize how text-to-music generation models are evaluated, providing a pairwise reward model that predicts human preference scores from a text prompt and an audio clip [1]. The model, detailed in a paper submitted on 15 Jun 2026, is trained entirely on publicly available human-preference labels, including arena-style A/B votes, crowdsourced pairwise comparisons, and expert aesthetic ratings [2]. The predicted score margin between two audio clips is well calibrated on a held-out test split, which the authors state supports data filtering via a simple score threshold [2]. TuneJury also generalizes to out-of-distribution benchmarks, remaining competitive with prior baselines [2]. For music generators released after the model's training, the researchers introduce anchor calibration, a post-hoc, per-system Bradley-Terry calibration method that recovers agreement with substantially better data efficiency than retraining from scratch [2]. The same frozen reward model drives consistent gains across three downstream applications: inference-time best-of-N selection, DITTO-style latent optimization, and expert-iteration post-training [2]. The code is available on GitHub [2]. The release of an open preference metric for generative audio arrives amid broader scrutiny of how algorithmic systems encode and reproduce bias. Algorithmic bias describes systematic and repeatable harmful tendencies in computerized systems that create unfair outcomes, often privileging one category over another [3]. Bias can emerge from many factors, including how data is coded, collected, selected, or used to train an algorithm, and it has been observed in search engine results and social media platforms [3]. The study of such bias is most concerned with algorithms that reflect "systematic and unfair" discrimination, an issue only recently addressed in legal frameworks such as the European Union's General Data Protection Regulation and the Artificial Intelligence Act adopted in 2024 [3]. Because algorithms are often considered neutral, they can inaccurately project greater authority than human expertise, a phenomenon known as automation bias [3]. In the context of creative AI, preference models like TuneJury that are trained on human judgments must contend with the cultural, social, and institutional expectations embedded in those labels [3]. A 2021 survey identified multiple forms of algorithmic bias, including historical, representation, and measurement biases, each of which can contribute to unfair outcomes [3]. The TuneJury paper does not explicitly address bias mitigation in its training data, but its reliance on publicly available labels and open release provides a foundation for external auditing of its preference signals [2].