PAG-Agent: Zero-Shot Personalized Music Album Generation via Aesthetic Profile Distillation

Yifeng Mao; Ziye Wang; Yutong Shen; Feiran Liu; Wenjuan Han

cs.MMOct 10, 2026

PAG-Agent: Zero-Shot Personalized Music Album Generation via Aesthetic Profile Distillation

Authors: Yifeng Mao, Ziye Wang, Yutong Shen, Feiran Liu, Wenjuan Han

Organizations: Beijing University of Technology · Beijing Jiaotong University

Abstract

Large Music Models (LMMs) can synthesize high-fidelity audio and support increasingly flexible prompt control, yet modeling the subjective preferences of an individual listener remains underexplored. We present PAG-Agent, a zero-shot multimodal framework that uses an explicit Personal Aesthetic Profile to bridge user preference inference and music generation control. A multimodal LLM distills favorite-song audio and contextual user information into this structured representation, which a planning LLM translates into generation instructions for existing music synthesizers. We further introduce PA-Eval, which evaluates personalization through the Aesthetic Consistency Score (ACS) for target-user alignment and the Aesthetic Specificity Index (ASI) for cross-user specificity. Experiments with Personal-100 on multiple MLLM and music-generation backbones show consistent improvements over generic and direct conditioning. A 50-participant follow-up study further shows a correlation $r=0.64$ between ACS and subjective preference.

Figures & tables

Figure 1: Overview of PAG-Agent. The profiling module distills favorite songs and contextual user information into a Personal Aesthetic Profile. The planning module converts that profile into track-level specifications for a music generator. The same profile can also condition cover art and a short album narrative, while the audio tracks remain the primary evaluation target in this work.

Figure 2: The PA-Eval pipeline. Independent reference and reverse streams map user evidence and generated output into a shared six-dimensional aesthetic space. ACS measures target-user consistency, while ASI compares that consistency against cross-user alignment.

PAG MLLM Gemini Eval MLLM Gemini ACE-Step
Method	ACS $\uparrow$	ASI $\uparrow$
Random	$0.51{\pm}0.15$	$0.37{\pm}0.19$
Direct	$0.58{\pm}0.11$	$0.44{\pm}0.15$
PAG-Agent	$\mathbf{0.69{\pm}0.10}$	$\mathbf{0.56{\pm}0.14}$
PAG MLLM Gemini Eval MLLM Gemini Suno-V5
Random	$0.56{\pm}0.12$	$0.41{\pm}0.12$

Table 1: PA-Eval results on Personal-100. Values report mean

\pm

standard deviation across users. PAG-Agent improves both metrics across every displayed backbone configuration.

Method	Musicality	Preference
Random	4.30	3.68
Direct	4.42	4.12
PAG-Agent	4.40	4.76
$r_{\mathrm{ACS/Human}}$	$-0.16$	0.64

Table 2: Mean human ratings from 50 participants. PAG-Agent receives the highest mean preference while musicality remains similar across methods. ACS is positively associated with subjective preference and shows little association with generic musicality.

Configuration	ACS	ASI
Songs only $S_{u}$	0.60	0.42
Profile only $P_{u}$	0.61	0.47
Songs + sparse $P_{u}$	0.68	0.54
Songs + full $P_{u}$	0.74	0.62

Table 3: Modality ablation with Gemini. Combining song history with the full profile gives the highest ACS and ASI.

Dimension	ACS	ASI	$r_{\mathrm{ref,gen}}$
Emotional Valence (EV)	0.68 $\pm$ 0.12	0.49 $\pm$ 0.12	0.72
Energy & Arousal (EA)	0.79 $\pm$ 0.04	0.64 $\pm$ 0.07	0.76
Sonic Texture (STC)	0.77 $\pm$ 0.05	0.62 $\pm$ 0.09	0.75
Musical Sophistication (MS)	0.60 $\pm$ 0.14	0.39 $\pm$ 0.18	0.69
Novelty (NU)	0.70 $\pm$ 0.08	0.57 $\pm$ 0.10	0.73
Cultural Rebellion (CSR)	0.79 $\pm$ 0.05	0.65 $\pm$ 0.10	0.76

Table 4: Per-dimension results with Gemini and Suno-V5.

r_{\mathrm{ref,gen}}

is computed across users within each dimension.

Music Generation

Music generation studies computational methods for composing, synthesizing, and editing music from symbolic, audio, textual, and multimodal inputs.

See all 107 papers

APEX: Large-scale Multi-task Aesthetic-Informed Popularity Prediction for AI-Generated Music
May 5, 2026Jaavid Aktar Husain, Dorien HerremansMulti-Task Learning Music Information Retrieval
AMAAI Lab, Singapore University of Technology and Design
MusicTwins: Training Audio-Language Models to Listen like Their Persona
Oct 23, 2026Milan Liessens Dujardin, Kevin MiaoAudio-Language Models Personalized Language Model Alignment
Bryel Labs, UC Berkeley · Bryel Labs
Enhancing Zero-shot Personalized Image Aesthetics Assessment with Profile-aware Multimodal LLM
Apr 19, 2026Chun Wang, Chenfeng Wei, Chenyang Liu +1Zero-Shot Learning Personalized Image Aesthetics Assessment
Mashang Consumer Finance Co., Ltd., Chongqing, China · Xi’an Jiaotong-Liverpool University, Suzhou, China · Beijing University of Posts and Telecommunications, Beijing, China

May 5, 2026J/K move · Enter open · S save

APEX: Large-scale Multi-task Aesthetic-Informed Popularity Prediction for AI-Generated Music

Jaavid Aktar Husain, Dorien Herremans

AMAAI Lab, Singapore University of Technology and Design

Multi-Task Learning Music Information Retrieval

Open paper AskPDF arXiv

Music popularity prediction has attracted growing research interest, with relevance to artists, platforms, and recommendation systems. However, the explosive rise of AI-generated music platforms has created an entirely new and largely unexplored landscape, where a surge of songs is produced and consumed daily without the traditional markers of artist reputation or label backing. Key, yet unexplored in this pursuit is aesthetic quality. We propose APEX, the first large-scale multi-task learning framework for AI-generated music, trained on over 211k songs (10k hours of audio) from Suno and Udio, that jointly predicts engagement-based popularity signals - streams and likes scores - alongside five perceptual aesthetic quality dimensions from frozen audio embeddings extracted from MERT, a self-supervised music understanding model. Aesthetic quality and popularity capture complementary aspects of music that together prove valuable: in an out-of-distribution evaluation on the Music Arena dataset, comprising pairwise human preference battles across eleven generative music systems unseen during training, including aesthetic features consistently improves preference prediction, demonstrating strong generalisation of the learned representations across generative architectures.

Abstract

Music Generation

Explore similar work

APEX: Large-scale Multi-task Aesthetic-Informed Popularity Prediction for AI-Generated Music

MusicTwins: Training Audio-Language Models to Listen like Their Persona

Enhancing Zero-shot Personalized Image Aesthetics Assessment with Profile-aware Multimodal LLM