AI news ·
WorldRuler preserves user-defined rules and world states in interactive 3D environments
WorldRuler is a generative world model that follows user-defined rules and preserves them across game saves. The system uses a 3D rule graph to maintain rule-state consistency, even applying changes to objects not yet visible on screen.

Researchers have developed a new generative world model that can follow user-defined rules, not just real-world physics, and preserve those rules and world states when a game is saved and loaded again. The system, called WorldRuler, tackles a core problem for customizable gameplay: how to make AI-generated interactive worlds that respect personalized rule sets and maintain consistency across sessions.
The work comes from a team that built WorldRuler around an executable 3D rule graph. This graph organizes rule definitions, dependencies, and object states so that rule execution, video generation, and perceptual feedback operate in a closed loop. The approach addresses two specific technical challenges the researchers call Rule-State Persistence and Rule-to-Visual Realization.
How the 3D rule graph works
Rule-State Persistence means the system must keep rule definitions, their dependencies, and the current world state intact when saving and reloading. Without this, a restored world might not follow the rules the user originally set. WorldRuler's graph supports joint saving and restoration of both rules and states. Dynamic object binding grounds rule-specified entities in visual instances and uses stored states to constrain generation from the first moment an object appears.
Rule-to-Visual Realization requires the model to accurately show what happens when a rule takes effect on specific objects and their attributes. This gets harder when the affected objects are not yet visible on screen. The researchers introduced object attribute state replacement and counterfactual reconstruction training to handle this. These techniques enable precise attribute control and compatible appearance and motion changes while limiting interference with unrelated content.
Building a dataset for rule-driven interactions
Training such a system demands data that captures rule-driven interactions with detailed state and event annotations. The team built an automated data collection and annotation pipeline using GTA-V and Habitat to create RuleVerse. This dataset includes cases where rules take effect before the affected objects become visible - a scenario that tests whether the model can maintain consistency even when the player cannot yet see the consequences.
The researchers report that extensive experiments demonstrate WorldRuler's effectiveness in preserving rule-state consistency across saving and loading and accurately realizing rule outcomes on designated objects and attributes. The system shows particular strength in handling cases where rule-induced changes must be remembered and applied to objects that appear later in a scene.
Why this matters for creatives, educators, and product developers
For professionals building interactive experiences, WorldRuler points toward a future where AI-generated worlds can follow arbitrary, user-defined logic rather than being limited to mimicking real-world behaviors. A game designer could define a rule that "all red objects become blue when the player touches the altar," and the system would maintain that rule across saves while correctly rendering the color change - even for red objects the player has not yet encountered.
This has practical implications for rapid prototyping in game development, interactive storytelling, and simulation-based training. The ability to define rules and have the generative model respect them consistently reduces the need to hand-author every possible interaction. For those exploring generative video tools in their workflows, the underlying concepts connect to broader advances in controllable generation - a topic covered in resources like AI Video Creation Courses and Text to Video AI Courses that examine how creative professionals can work with these emerging capabilities.