Code World Model Rewires World Models
This new research uses a coding agent to maintain executable world state and reason about persistent consequences, while a video model renders the resulting scene. A proxy representation bridges code-defined dynamics with MiniMax-H3’s visual generation.
Code World Model makes a compelling architectural bet: let code handle causality and persistence, while generative video handles appearance.
- –Executable state could preserve object attributes, positions, and consequences across long interactions better than pixel prediction alone
- –The programmable proxy translates abstract code into spatiotemporal constraints without requiring a complete 3D asset pipeline
- –The approach is modular: stronger coding agents improve world logic, while better video models improve visual fidelity
- –The current system remains an early prototype, relying on simple interactive worlds and a closed-source game-code foundation
- –Latency, proxy quality, error accumulation, and the difficulty of reliably generating simulation logic remain major obstacles
DISCOVERED
1h ago
2026-08-28
PUBLISHED
1h ago
2026-08-28
RELEVANCE
AUTHOR
mark_k
