Music generation stops chasing generic appeal: distilling a listening history into an aesthetic profile makes the same track fit the listener

Personalized music generation can now infer a listener's preference first and control synthesis second, instead of optimizing only population-level quality or explicit prompt control. Existing systems mainly optimize population-level quality or explicit controllability, while personalized preference alignment remains underexplored; a track can satisfy generic quality criteria while remaining poorly matched to the person who receives it. On Personal-100, across four backbone configurations that change the generation MLLM, the evaluation MLLM, and the music generator, PAG-Agent improves ACS by roughly 9 to 19 percent and ASI by 27 to 29 percent over Direct; against Random the gains reach roughly 33 to 37 percent in ACS and 46 to 54 percent in ASI.

The method separates preference inference from music realization: it first builds an explicit Personal Aesthetic Profile, then translates that profile into generator-facing controls. A profiling multimodal LLM distills favorite-song audio together with contextual attributes such as age, occupation, MBTI type, cultural background, and social activity level into a JSON representation holding an aesthetic summary, core stylistic attributes, and negative constraints; a planning LLM then converts it into track-level controls for emotion, vocal form, genre, instrumentation, and style. Evaluation runs on Personal-100, built from 100 participants across multiple countries and occupations, with ablations on the Personal-25 subset; music is generated with Suno-V5 or ACE-Step, and a follow-up human study involves 50 participants.

Successful personalization is split into two questions: agreement with the intended listener and specificity relative to other listeners. Generic audio quality measures do not directly characterize individual preference alignment, so PA-Eval scores alignment with the target user's aesthetic reference through the Aesthetic Consistency Score and contrasts it with cross-user matches through the Aesthetic Specificity Index, reducing the score when the same content also aligns well with other users. Across the six-dimensional aesthetic space of Emotional Valence, Energy and Arousal, Sonic Texture Complexity, Musical Sophistication, Novelty and Unconventionality, and Cultural and Stylistic Rebellion, reference and generated scores stay positively correlated in every dimension, with coefficients from 0.69 to 0.76 and an average of 0.74.

The work targets researchers and engineering teams who want generated music aligned to a specific listener, and its demonstrated scope is Personal-100 with 100 participants, the Suno-V5 and ACE-Step generators, and a 50-participant human evaluation. Because the profile is decoupled from the generator, the same Personal Aesthetic Profile can be used with different music generators without retraining.

In the human evaluation, mean musicality is similar across the three methods, while mean subjective preference rises from 3.68 for Random and 4.12 for Direct to 4.76 for PAG-Agent; ACS correlates positively with subjective preference but only weakly with musicality. Ablations show either modality alone performs below the combined setting: adding a sparse profile to the song history raises ACS from 0.60 to 0.68 and ASI from 0.42 to 0.54, and the full profile reaches 0.74 and 0.62. Performance varies across aesthetic dimensions, with Energy and Arousal and Cultural and Stylistic Rebellion highest at 0.79 ACS and Musical Sophistication lowest at 0.60 ACS and 0.39 ASI.