AI world models appear to carve up physics on their own — and a new paper shows how.
Researchers developed a controlled diagnostic protocol to probe what is actually happening inside the hidden states of passive object-state world models — systems that predict how physical scenes will unfold. Rather than just measuring prediction accuracy, the team tested whether latent representations encode different event types (free motion, collisions, occlusions) and whether those event contexts shift how the models weight physical information. Using recurrent, attention-based, and latent state-space models, they found that hidden states reliably distinguish event regimes and reweight physical readouts accordingly: free motion leans on kinematic structure, collisions combine kinematic and contact information, and occlusions blend motion-related cues with object-permanence structure.
The finding matters because it moves the conversation past benchmark scores. If world models are spontaneously organizing physical knowledge by event context — not just fitting a loss curve — that has implications for how researchers design, interpret, and trust these systems in robotics, simulation, and planning applications. Understanding the internal organization is a prerequisite for knowing when a model will fail.
The authors are careful not to overclaim: the results do not imply discrete physics modules or clean causal circuits inside the network — just that suppressing field-aligned directions can degrade event-relevant predictions, with the strongest evidence in collision-contact windows. It is a diagnostic finding, not a blueprint. But in a field that often treats model internals as a black box to be benchmarked around, that counts as progress.