AI world models can predict two different futures from the same starting scene, and quietly assume two different sets of physics to do it.
Researchers built a system called OneWorld that generates multiple action-conditioned predictions from a shared starting scene while forcing them to share the same underlying physical rules, like friction and mass, instead of inventing new ones for each branch. A physical mechanism interpreter estimates plausible physics for each predicted outcome, then combines those estimates into a shared consistency signal that shapes how the model trains and samples. The team also built a new evaluation protocol, based on an existing benchmark called ACWM-Phys, that checks whether a model's different predicted futures could actually be explained by one shared set of physical parameters. In controlled test environments, OneWorld improved cross-intervention physical consistency while keeping single-prediction accuracy competitive with existing approaches.
This targets a specific failure mode in world models built for planning and robotics: a model that predicts an object slides easily under one action and sticks under another isn't just wrong, it's incoherent, and that incoherence compounds once the model is used to choose actions rather than just describe them. Consistency, not just raw prediction accuracy, is what makes a world model usable for reasoning about consequences before acting.
The gains so far come from controlled simulation environments, not messy real-world video, so this reads as an early fix for a well-scoped bug rather than proof that world models have solved physics.