WorldRuler paper proposes generative world models that follow user-defined rules
A research team built a system that keeps custom gameplay rules consistent when virtual worlds are saved, reloaded, and rendered.
In plain words
Generative world models create interactive video worlds, but they usually follow real-world physics rather than rules a player invents. WorldRuler stores rule definitions and object states in a 3D graph so those rules survive saving and loading, and it applies rule effects to the right objects on screen. The authors also released a training dataset built from GTA-V and Habitat.
What was announced
The arXiv paper identifies two obstacles for rule-following world models. Rule-State Persistence requires that rule definitions, dependencies, and world states survive saving and loading, so a restored world still obeys established rules. Rule-to-Visual Realization requires that rule-induced changes appear on the correct objects and attributes, including objects that have not yet been seen. WorldRuler addresses both by arranging rules and object states in an executable 3D rule graph and linking rule execution, video generation, and perceptual feedback in a closed loop, with dynamic object binding tying rule-specified entities to visual instances from their first appearance.
For training data, the authors built an automated collection and annotation pipeline using GTA-V and Habitat, producing RuleVerse, a dataset of rule-driven interactions with state and event annotations, including cases where rules take effect before affected objects become visible. They also introduce object attribute state replacement and counterfactual reconstruction training, which they say enable accurate attribute control while limiting interference with unrelated content. The paper reports experiments showing WorldRuler preserves rule-state consistency across save and load cycles and realizes rule outcomes on designated objects and attributes.
What it does not show
- The source is only the paper's abstract, so the experimental details, benchmarks, and comparisons behind the effectiveness claims cannot be verified from the supplied text.
- The effectiveness results are reported by the paper's authors; no independent evaluation is available in the supplied material.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.