MAGIC: Transition-Aware Generation of Navigable Multi-Scene Game Worlds with Large Language Models

MAGIC presents a four-stage prompt-to-project pipeline for generating runnable multi-scene game worlds. It creates a shared transition-aware intermediate representation, specifies scenes while validating portal reachability with flood fill, generates scenes and transition scripts jointly, and assembles them into one project. The paper also introduces an evaluation agent that executes transitions in play rather than relying only on static scene metrics. On a benchmark of 100 multi-scene cases, MAGIC reportedly produced an executable project for every case and achieved 0.99 precision, 0.95 recall, and 0.96 F1 for end-to-end transition identification, outperforming an LLM baseline and Holodeck on portal recovery and navigability.