Probe and Canonicalize: Active Identification of Hidden Physical Scale for Out-of-Distribution World Models

Physical world models often fail to extrapolate when a physical factor leaves the training distribution, even when the governing interaction is unchanged. We study whether such a factor should be learned jointly with dynamics or identified separately and applied through known structure. Our active hidden-physics canonicalization predicts a factor-independent raw effect, estimates an unknown scale from a controlled visual probe, and applies it through a fixed analytical operator. In an object-centric environment where mass is absent from appearance, inverse mass is estimated from pixel displacement after a five-step probe calibrated only on training masses in [1, 4]. Across five backbone families and five seeds each, canonicalization reduces normalized final-position error by 58.8–60.9% on unseen masses in [6, 10] and improves every evaluated seed. For the mirror-attention backbone, error falls from 1.670 ± 0.028 to 0.680 ± 0.004, matching a true-mass oracle at 0.679 ± 0.003. Crucially, supplying the same estimate as a learned input yields only 1.585 ± 0.086, and supplying oracle mass as input gives the same result. Fixed analytical application therefore reduces error by 57.1% relative to learned conditioning: access to the factor alone does not ensure extrapolation. Interventions further show robustness to estimation noise while revealing that this benchmark depends primarily on recovering the unseen physical scale rather than precise per-object attribution. These results isolate a specific OOD entanglement that is prevented by placing identified physical structure outside the learned predictor.