Code World Model

Code World Model is a research framework that separates world evolution from visual rendering. A coding agent acts as the "world brain": it reasons about events and their consequences and writes executable code that maintains persistent world state and applies rule-consistent updates.

To connect code state to video generation, the authors introduce a proxy representation encoding frame-wise spatiotemporal constraints. It is compiled into a proxy video that conditions a video model to render the visual output. Data pipelines build aligned proxy-observation pairs from gameplay and real-world videos. After fine-tuning on paired gameplay data, MiniMax-H3 follows proxy-based specifications from simple interactive worlds built by the coding agent. The project page also shows long-horizon generation with periodic style changes.

The work is by Yiwen Chen, Guosheng Lin and Chi Zhang and is presented via a project page and an arXiv paper. A YouTube creator reported that code was released, but no repository link appeared in the provided material. Demonstrations are limited to simple interactive worlds, and no public product or hosted service was identified.