Code World Model: Coding Agent as World Brain
A coding agent keeps the simulated world state, while a video model turns that state into frames.
The paper proposes separating world evolution from visual rendering. Its “Code World Model” has a language-model agent reason about events, write executable code, and maintain persistent consequences in the world. A proxy representation converts that state into frame-by-frame spatiotemporal constraints for a video model. The authors report that, after fine-tuning on paired gameplay data, MiniMax-H3 followed those proxy specifications while preserving visual detail and motion. ArXiv · AI/CL/LG's note
The paper proposes separating world evolution from visual rendering. Its “Code World Model” has a language-model agent reason about events, write executable code, and maintain persistent consequences in the world. A proxy representation converts that state into frame-by-frame spatiotemporal constraints for a video model. The authors report that, after fine-tuning on paired gameplay data, MiniMax-H3 followed those proxy specifications while preserving visual detail and motion. ArXiv · AI/CL/LG's note
score 5