Mind-Studio: Executable World Models with Lookahead Evaluation for Partially Observable Games

Mind-Studio is a new framework that synthesizes executable world models for partially observable games using large language models, transforming state-action-next-state trajectories into independent programs. It employs a K-step lookahead fidelity protocol to evaluate synthesis quality, achieving a significant improvement in next-state prediction accuracy from 0.3% to 48.7% on Montezuma's Revenge and demonstrating superior branch-level fidelity on other games like Alien and Assault. This advancement is crucial for practitioners as it enhances the development of more accurate and autonomous AI agents in complex environments.