OmniScore-Bench: Evaluating Vision-Language Models on Musical Score Reading
Different models lead on score transcription depending on whether they read excerpts, single pages or multiple pages.
OmniScore-Bench tests 12 vision-language models with 3,000 questions across six aspects and 24 task subtypes. It evaluates symbol localisation, musical consistency and evidence-linked answers, separating answer-format violations from musical errors. On a tested subset, removing score images exposes visual dependence in pitch and repeat interpretation. The results identify specific score-reading strengths and bottlenecks that a transcription score alone leaves unresolved.
ArXiv · AI/CL/LG's note
OmniScore-Bench tests 12 vision-language models with 3,000 questions across six aspects and 24 task subtypes. It evaluates symbol localisation, musical consistency and evidence-linked answers, separating answer-format violations from musical errors. On a tested subset, removing score images exposes visual dependence in pitch and repeat interpretation. The results identify specific score-reading strengths and bottlenecks that a transcription score alone leaves unresolved.
ArXiv · AI/CL/LG's note
score 4