View a PDF of the paper titled Occasion-Conditioned Diagnostics of Kinematic, Contact, and Object-Permanence Fields in Passive Object-State World Fashions, by Yang Liu and Yuming Chen
View PDF
Summary:World fashions can predict future bodily states, however prediction accuracy alone doesn’t clarify how bodily data is organized and used inside their latent dynamics. We introduce a managed diagnostic protocol for learning event-conditioned latent bodily construction in passive object-state world fashions. The protocol checks whether or not hidden representations encode event-regime data, whether or not occasion contexts reweight non-exclusive bodily discipline readouts, and whether or not field-aligned representational elements have practical penalties for prediction. Utilizing a balanced controlled-generator dataset with free-motion, collision, and occlusion occasions, we consider recurrent, attention-based, and latent state-space transition fashions underneath a fixed-horizon forecasting setup. The fashions be taught helpful predictive dynamics and their hidden states assist dependable event-regime readout. Occasion contexts systematically reweight kinematic, contact, and object-permanence discipline readouts: free movement is kinematic-dominant, collision combines kinematic and get in touch with construction, and occlusion combines motion-related and object-permanence construction. Time-aligned and directional-consistency analyses additional present phase-related shifts in discipline emphasis. Lastly, fixed-horizon projection causal discipline impact (CFE) reveals that suppressing field-aligned instructions can degrade event-relevant prediction, with strongest proof for contact-aligned construction in collision-contact home windows and extra certified proof for object-permanence-aligned construction in hard-occlusion hidden home windows. These outcomes assist event-conditioned group and fixed-horizon practical sensitivity of latent bodily fields, whereas not implying specific bodily modules, remoted causal circuits, or context-invariant sliding-window generalization.
Submission historical past
From: Yang Liu [view email] [v1]
Fri, 26 Jun 2026 10:11:59 UTC (982 KB)
[v2]
Fri, 14 Aug 2026 01:38:03 UTC (971 KB)
![[2606.28455] Occasion-Conditioned Diagnostics of Kinematic, Contact, and Object-Permanence Fields in Passive Object-State World Fashions [2606.28455] Occasion-Conditioned Diagnostics of Kinematic, Contact, and Object-Permanence Fields in Passive Object-State World Fashions](http://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png)
