In music-generation models, an explicit instruction to realize a musical attribute can paradoxically make it less common. How, then, should we assess whether a model follows a musical instruction? We use matched inputs that omit an attribute, specify one target, or specify another target, and compare the resulting audio. We evaluate key and beat grouping in ACE-Step 1.5, Stable Audio 3 and LeVo2 on 2,304 primary outputs. In Stable Audio 3, a four-beat instruction reduces detected four-beat occurrence from 96.9% to 56.3%, even though four beats remain more common than three under their respective instructions. Further experiments show that rephrasing can improve attainment while reducing the difference between target instructions, and that ACE-Step shows larger changes in target occurrence when text instructions retain its native attribute field. Our findings suggest that evaluating musical instruction following requires both target attainment and comparisons with generation with an omitted target instruction and with a different specified target.

翻译：暂无翻译