In a manuscript submitted to arXiv on 18 November 2026, researchers describe an experimental AI orchestra whose practice can become part of a composition. The agents can save the musical variations produced during training for a composer to select and arrange. For experimental musicians and music educators, that offers a different kind of AI output to consider: editable ideas collected along the way.
The work is by Eduardo Reck Miranda and Moein Fahmideh Vatandoost at the University of Plymouth. Its arXiv record describes it as accepted for Computer Music Journal but still subject to revision before final publication. This is a research account, not a consumer music-service launch.
How saved practice becomes musical material
The concrete idea appears in the paper’s section on reusing learning history: preserve the material generated after training episodes, then use those intermediate outputs for offline composition. Think of it as a musical sketchbook. The saved pages contain variations, not a transcript of an AI’s thoughts.
The system works with MIDI: instructions about musical events, rather than a finished sound recording. Its representation is monophonic—one note at a time. The agents transform short passages by changing pitch, note duration or both. An “orchestra” here is a collection of these software agents, not a generated recording of a full ensemble.
The authors also describe a batch workflow that saves variations separately or joins them into sequences for further work in a score editor or audio workstation. Agents can be chained, so one transforms the output of another. Selection and arrangement remain meaningful composing tasks; producing alternatives does not decide which ones belong in a piece.
The composer helps write the musical brief
These agents learn through reinforcement learning: they try permitted changes, receive numerical rewards and update how they choose actions. In this system, adjustable quantum circuits participate in a learning loop that also uses conventional computing. They are not chatbots assigned instrumental roles.
Two choices shape what they explore: the moves they are allowed to make and the results they are rewarded for. The paper’s rhythmic examples distinguish rearranging existing note lengths from changing those lengths. That is a concrete creative choice: reshuffle the rhythm’s existing ingredients, or let the system introduce different ones.
Rewards also encode preferences. One of the defined measures, Diversity, rewards a larger proportion of distinct symbols in a sequence. The analytical catch is simple: if a composer wants a recurring motif, avoiding repetition may work against the brief. A higher score is not automatically a better musical decision.
Not training on a large collection of songs does not remove human influence. Someone still chooses the starting phrase, available transformations, reward weights and eventual arrangement. Human guidance and machine exploration can coexist—a broader learning question explored in how IBMARL learns beyond imitation, a separate study of simulated cooperating agents.
Why the last output is not the whole result
One simulator experiment shows why retaining the trail could matter. The authors compared a trained rhythmic agent, Shin, with random action selection and an untrained circuit whose settings stayed fixed. All three used the same input, permitted actions and reward, with 50 training episodes or equivalent exploration per run, across 30 random seeds.
In Table 4, the trained agent’s mean best-found reward was 0.6276, compared with 0.4132 for random actions; it also exceeded the fixed circuit. But the final-episode means were much closer: 0.3218 for training and 0.3061 for random actions. These are the researchers’ scores for their chosen objectives, not listener ratings.
That difference matters to a composer. A search can discover an interesting candidate and then move on. Keeping only the last output throws away earlier alternatives, including ones that scored better. The experiment therefore supports assessing what the search found along the way, not assuming its last answer must be its best.
There is a second judgment still to make: whether any of those candidates serves the music. The study does not report a listener evaluation of this orchestra. It cannot turn a reward advantage into a claim that audiences prefer the result.
What the quantum demonstration establishes
The numerical comparison above ran in simulation. In a separate demonstration, the authors trained three agents on IBM quantum hardware for six episodes each, then used chained transformations to produce musical material. That demonstrates their experimental workflow; it is not evidence that quantum computing makes better music than a comparable conventional learning system.
The authors also say they did not systematically measure end-to-end response times. The saved-variation workflow is therefore easier to assess from this evidence than dependable live responsiveness. The limited examples applying trained agents to unseen snippets likewise do not establish reliable results across genres or arbitrary compositions.
Reward-shaped music has a longer history. In 2016, Magenta’s RL Tuner research combined patterns learned from songs with music-theory rewards. It is historical context, not a head-to-head competitor in this experiment. The new paper does not establish a general advantage over classical reinforcement learning in musical quality, speed or cost.
What composers can take from it
For a composer, the useful distinction is between wanting finished audio and wanting material to develop. Readers considering production controls can turn to the separate coverage of Lyria’s music-production tools. This experiment asks a different question: what happens if the learning attempts themselves are worth keeping?
The cited evidence does not establish a ready-to-use consumer app, a complete downloadable implementation or a product price. It is a reason to consider a creative workflow, not a reason to buy quantum-computing access.
A practical implication for future creative tools is to preserve alternatives alongside the settings that produced them. Composers could then judge whether a collection offers useful variety, whether its parts can be edited and whether selecting them is worth the effort. Those are proposed criteria, not benefits measured by this study.
The appeal is in the sketches: a musician can choose, reject and rearrange what the AI leaves along the way. The computer’s scoring rule can guide the search. It need not get the final say on the music.
