Text-to-music systems are producing increasingly convincing audio, but a new paper argues that existing evaluation methods reveal little about whether the generated result actually matches what a user intended. A single global text-audio relevance score, the authors note, can overlook implicit intent in underspecified prompts and mask failures in specific requirements such as instrumentation, structure, rhythm, or mood progression.
To address this, the researchers formulate text-to-music intent alignment as satisfying a per-request rubric of independently verifiable items that cover both a request's explicit requirements and its implied musical intent. Scoring items individually, they argue, makes evaluation diagnostic by intent source and musical dimension rather than relying on one opaque score.
They instantiate this idea as MuRA-Bench, a benchmark of 100 real-world platform requests curated by music experts with expert-revised rubrics covering explicit and inferred musical intent.
The paper also proposes MIRA (Musical Intent Refinement Agent), a test-time agent that first grounds a request's intent into rubrics, then searches over prompt revisions for a black-box generator under a bounded budget. It iteratively generates music, verifies it against the rubrics, and uses that feedback to guide a trajectory-aware tree search with branch-specific memory.
Experiments across open-source and commercial backends show that MIRA improves intent alignment without updating generator parameters, enabling an open-source generator to reach performance comparable to representative commercial systems such as Suno and Mureka. The project page offers further details and audio examples.