MIRA: A Musical Intent Refinement Agent for Aligning Text-to-Music Generation with User Intent
MIRA turns vague music prompts into checkable intent rubrics, then uses them to revise prompts against black-box generators.
The paper argues that broad text-audio relevance scores miss failures in instrumentation, structure, rhythm, and mood progression. It introduces MuRA-Bench, built from real-world platform requests and curated by music experts, to score explicit and implied musical intent item by item. MIRA uses those rubrics at test time, generating and verifying music under a bounded search budget before choosing prompt revisions. The authors report gains across open-source and commercial backends, with an open-source generator reaching performance comparable to systems such as Suno and Mureka. HF Daily Papers' note
The paper argues that broad text-audio relevance scores miss failures in instrumentation, structure, rhythm, and mood progression. It introduces MuRA-Bench, built from real-world platform requests and curated by music experts, to score explicit and implied musical intent item by item. MIRA uses those rubrics at test time, generating and verifying music under a bounded search budget before choosing prompt revisions. The authors report gains across open-source and commercial backends, with an open-source generator reaching performance comparable to systems such as Suno and Mureka. HF Daily Papers' note
score 4