A new technique trains AI world models to tell the difference between what their actions cause and what just happens to occur alongside them.
The method, called Do-JEPA, comes from a paper titled 'Do-JEPA: From Masking to Intervention in Latent World Models,' posted to arXiv 2609.37378 on September 30, 2026. Instead of masking parts of a scene and asking a model to guess what's hidden - the approach used by systems like C-JEPA - Do-JEPA reruns a saved simulator state under two different actions and trains the model on the difference between the two outcomes. In a synthetic test, this let the model correctly identify which object an action affected 99.95% of the time, versus essentially never for the masking-based approach. On real pixel data, the added effect loss cut latent prediction error by 28.4% and physical effect error by 13.5%, with gains as high as 20% and 66% on separate physics-shift benchmarks.
World models are the backbone of robotics and planning systems that need to predict what happens if they take an action, not just what tends to happen next. Today's dominant approach, masking objects for a model to reconstruct, can't separate causation from correlation, so a robot's model might credit an object's motion to something it happened to see rather than to its own grip. Do-JEPA's fix is conceptually simple: change what happens in the simulation, not what the model is shown.
The catch: training a model from scratch with this method actually costs some factual accuracy. The gains only appear when it's used to fine-tune a model that already works, which makes this look like a useful patch rather than a wholesale replacement for how world models learn.