MAGIC: Generating Multi-Room Game Worlds
CUHK-CSE · hf · 2026-07-16
This paper introduces MAGIC, a prompt-to-project system that converts natural language prompts into runnable, multi-scene game worlds. It focuses on solving three major challenges in multi-scene generation:
- Cross-scene consistency: Entrances and exits across different rooms/scenes must align perfectly.
- Intra-scene navigability: Generated scenes must remain fully traversable.
- Transition evaluation: Effectiveness cannot be judged by single scenes alone; actual switching execution must be verified.
Methodologically, MAGIC employs a four-stage pipeline:
- Planning an intermediate representation with "transitions" in mind.
- Generating each scene and constraining door/entrance reachability using a flood-fill validator.
- Simultaneously generating scenes and transition scripts to assemble a complete project.
- Running actual transition tests within the game using a dedicated transition evaluation agent.
The authors also created a benchmark comprising 100 multi-scene cases. On this dataset, MAGIC successfully generated executable projects for every sample, achieving 0.99 precision / 0.95 recall / 0.96 F1 in end-to-end transition recognition, ultimately producing more navigable layouts than LLM baselines and Holodeck. The code is open-source.
More from Research
- Structural ensembles beat single predictions in TCR:pMHC generalization study — quaidmorris · 2026-07-22
- RSS launches under OMSF to push structural biology data modeling at scale — MoAlQuraishi · 2026-07-22
- enFoldX tops 8 neoantigen scans and an unseen-peptide benchmark — quaidmorris · 2026-07-22
- enFoldX reaches AUC 0.82 on human VDJdb and transfers to mouse at 0.76 — quaidmorris · 2026-07-22
- enFoldX gains accuracy as AF3 ensemble disagreement rises for non-binders — quaidmorris · 2026-07-22
- A 3D ray plot shows how hard this Jacobian counterexample is to read — moultano · 2026-07-22