A study examines whether pretrained video foundation models hold intuitive-physics information in their frozen representations. Using frozen-feature probing on the IntPhys2 benchmark, the authors compare how this information differs across model families, layers, and probe types. The work is a cross-listed arXiv replacement submission in cs.AI and cs.LG.
This arXiv paper introduces LPA-CWM, a learned adjudicator that weighs the outputs of counterfactual world models when extracting motion from pretrained video predictors. The authors note that predictions produced under different target-frame masks differ in reliability, making uniform weighting suboptimal. Their approach learns to score and combine those predictions rather than treating them equally.
A new arXiv paper asks whether video generators that produce physically implausible motion never learned the correct dynamics or merely fail to use what they learned. The authors introduce a notion of causal writability to probe this, and report that the correct motion is still represented inside the model and can be made to steer generation. The finding points to a gap between internal knowledge and how it is used, rather than a simple absence of physical understanding.