New diagnostic reveals how world models encode physics internally

World models predict future states, but how they organize physical information inside their latent dynamics has remained a black box. Researchers Yang Liu and Yuming Chen from an undisclosed institution (paper on arXiv) propose a controlled diagnostic protocol for passive object-state world models. They test whether hidden representations encode event-regime information (free motion, collision, occlusion), whether event contexts reweight non-exclusive physical field readouts (kinematic, contact, object-permanence), and whether field-aligned representational components have functional consequences for prediction. Their balanced dataset includes events like free-motion, collision, and hard occlusion. The models evaluated include recurrent networks, attention-based transformers, and latent state-space models under a fixed-horizon forecasting setup.

The results are striking. Hidden state readouts reliably classify which event regime the model is in—showing that the latent space is organized by event type. Event contexts systematically reweight physical fields: free-motion windows favor kinematic-dominant representations, collision windows combine kinematic and contact structure, and occlusion windows blend motion-related and object-permanence structure. Time-aligned and directional-consistency analyses reveal phase-related shifts in field emphasis. Most importantly, the authors use a causal field effect (CFE) approach: suppressing field-aligned directions degrades prediction accuracy in event-relevant time windows. For example, contact-aligned structure is crucial for collision windows, and object-permanence-aligned structure matters for hard occlusion hidden windows. However, the results do not imply explicit physical modules or context-invariant sliding-window generalization—rather, they show event-conditioned organization and functional sensitivity in a fixed-horizon setting. This work provides a robust framework for interpretability in physical world models, crucial for safer AI robotics.