From Structural Priors to Local Causality in World Models

Object-centric world models represent scenes as distinct entities, but this decomposition alone does not specify how action information should enter the scene and propagate between objects. We introduce a robot-mediated action prior and state-dependent sparse gates to organize information flow within these models. Our design separates interaction admissibility from interaction selection: a physical attention mask restricts direct action–slot exchange to robot slots, while gates select among admissible connections. The gates and predictor are jointly optimized with prediction loss and sparsity regularization, without object–object edge supervision. Multi-layer message passing enables action information to reach other objects through the robot. We define local transition dependencies through input interventions on a fixed predictor, restricting our causal analysis to the learned transition mechanism. We evaluate the approach on four visual planning benchmarks. Our model achieves 100% success on Two-Room and 92.00% on PushT. On OGBench-Cube and OGBench-Scene, it achieves 84.67% and 78.00%, respectively, compared with 75.33% and 72.67% for the reproduced LeWorldModel baseline. Structural ablations on OGBench-Cube show greater prediction-error stability under candidate nuisance-slot removal and lower action-effect error on selected contact samples. Route interventions further identify learned connections carrying local action responses within the tested eligible neighborhoods. Together, these results support compact, robot-mediated transition modeling with competitive planning performance.