Taste Transfer: Learning from Artistic and Musical Preference Modeling

TL;DR: If music and art apps can learn your unique taste, why can’t AI scientists learn a “personalized” scientific taste? We’ll borrow techniques from music/art recommender systems to model taste diversity in science.

Research Question: Can AI models adapted from art and music recommendation systems capture heterogeneous, field-specific, or even individual variations in scientific taste, thus enabling personalized or sub-community aligned research judgments?

Hypothesis: Methods such as diversified attentive user profiles or content-aware collaborative filtering (from art/music domains) can uncover latent “taste clusters” in science, improving the relevance and diversity of AI-generated research proposals.

Experiment Plan: Adapt models like Diversified Attentive User Profiles or music preference embeddings to the scientific literature domain, using clusters of researchers/editors as “users” and papers/ideas as “items.” Train and test models on datasets annotated with both citation-based and human-judged taste labels. Analyze the emergence of sub-community tastes and their alignment with actual editorial or funding decisions.

References:

• Tong, J., et al. (2026). AI Can Learn Scientific Taste.
• Barkan, O., Shaked, T., Fuchs, Y., & Koenigstein, N. (2023). Modeling users’ heterogeneous taste with diversified attentive user profiles. User modeling and user-adapted interaction.
• Magron, P., & F'evotte, C. (2020). Leveraging the Structure of Musical Preference in Content-Aware Music Recommendation. IEEE International Conference on Acoustics, Speech, and Signal Processing.