MIRA: A Musical Intent Refinement Agent for Aligning Text-to-Music Generation with User Intent
Zekai Liu, Zhilin Wang, Xuzheng He, Yu Cheng, Yang Yang
Shandong University · University of Science and Technology of China · Central Conservatory of Music · Kunlun Tech Co. Ltd. · Shanghai Jiao Tong University
Published October 7, 2026 · on the feed October 9, 2026 · ▲ 7 · arXiv 2610.10355
The abstract
Text-to-music systems produce increasingly convincing audio, yet evaluation reveals little about whether the result matches user intent. A global text-audio relevance score can overlook the implicit intent in underspecified prompts and mask failures in specific requirements, such as instrumentation, structure, rhythm, or mood progression. To bridge this gap, we formulate text-to-music intent alignment as satisfying a per-request rubric of independently verifiable items covering both a request's explicit requirements and its implied musical intent. Scoring items individually makes evaluation diagnostic by intent source and musical dimension, rather than a single opaque score. We instantiate this as MuRA-Bench, a benchmark of real-world platform requests curated by music experts. We further propose 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, iteratively generating music, verifying it against the rubrics, and using this feedback to guide a trajectory-aware tree search. Experiments across open-source and commercial backends show that MIRA improves intent alignment, enabling an open-source generator to achieve performance comparable to representative commercial systems (e.g. Suno and Mureka). Project page: https://mirareview.github.io/.
The Brief
2:06A short podcast summary of the paper. Written and voiced by AI.
Transcript
MIRA. A Musical Intent Refinement Agent for Aligning Text-to-Music Generation with User Intent. By Zekai Liu and four others, from Shandong University, the University of Science and Technology of China, and three other institutions. Published on the seventh of October.
Text-to-music systems now produce audio that is increasingly convincing, yet the usual evaluation says little about whether a track matches what the user wanted. A single global score for how relevant the audio is to the text can miss the intent left implicit in an underspecified prompt. It can also hide failures on specific requirements: the instrumentation, the structure, the rhythm, or how the mood should progress.
The authors recast intent alignment as satisfying a rubric written for each request. The rubric is a set of items that can each be verified independently, and it covers both what the request explicitly asks for and the musical intent it implies. Because items are scored one by one, evaluation becomes diagnostic, broken down by where the intent came from and by musical dimension, instead of collapsing into one opaque number. This is instantiated as MuRA-Bench, a benchmark of real requests from a music platform, curated by music experts.
MIRA, the Musical Intent Refinement Agent, works at test time. It first grounds a request's intent into rubrics. It then searches over revisions of the prompt sent to a black-box generator, within a bounded budget. Each round generates music, verifies it against the rubrics, and feeds that verdict back to steer a tree search that takes the trajectory so far into account. The generator itself is never changed; only what it is asked is.
Experiments run across both open-source and commercial backends, and MIRA improves intent alignment on them. With it, an open-source generator reaches performance comparable to representative commercial systems, such as Suno and Mureka.
Two moves work together here. Breaking intent into checkable items turns a vague notion of matching the request into something that can be scored and diagnosed, and the same items then serve as the feedback that guides a search over prompts, which is how a generator treated as a closed box can still be pushed closer to what a listener meant.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free