Creative
Using large multimodal models to assess creativity and generate game levels in Physics Playground—toward scalable, valid measures of creative production.
Project Description
Overview
We train and prompt LMMs to rate novelty, usefulness, and surprise of human-authored levels, compare to expert rubrics, and study bias and reliability.
Methods
- Multimodal prompting pipelines (image + text)
- Rater-agreement calibration vs. human experts
- Auto-generation with constraint checks for playability
Project Details
- Lead: Dr. Seyedahmad Rahimi
- Sponsor: UF / Internal seed
- Funding: —
- Duration: 2024