Certifying Physical Language in Multimodal AI Models.

This paper addresses a critical challenge in world models: ensuring that compact multimodal representations truly serve as effective interfaces between perception and physical interaction. Current evaluation methods often fall short in verifying if different sensory inputs convey the same executable meaning or if this meaning remains consistent when new actions are composed. The researchers propose an operational capability hierarchy and the Disjoint-Bridge Operator-Substitution Certificate (DBOSC). This certificate evaluates whether independently trained modality compilers can interchangeably utilize a frozen response chart based on unseen data. Experiments on a haptic dataset showed that audio and acceleration representations of the same unseen surface were significantly closer in response space than incorrect pairings, indicating a degree of alignment. However, further tests on ordered execution in an elastoplastic system revealed limitations. While a converged model could execute held-out programs, the ability to "clear the gate" for execution was a property of the executor, not merely the response chart itself. The study concludes that simple compression and fusion of modalities are insufficient to determine novel composition laws, highlighting the need for more robust certification of physical language understanding in AI.