Validation of influence neighborhoods against explicit temporal causal graphs

Determine whether the influence neighborhoods inferred by Causal-JEPA—defined as the minimal subsets of contextual variables sufficient to predict a masked object's state under object-level masking—correspond to ground-truth temporal causal relations by directly validating these neighborhoods on datasets that provide explicit temporal causal graphs.

Background

Causal-JEPA introduces object-level masking that induces latent interventions, and the authors formalize the notion of an influence neighborhood as the minimal set of contextual variables needed to predict a masked object state. This constructs a causal inductive bias without assuming an explicit causal graph.

While the paper provides theoretical characterization of influence neighborhoods and empirical benefits, it does not validate whether these neighborhoods align with explicit temporal causal graphs. The authors highlight this gap as future work, motivating a concrete validation study on datasets with known causal structures.