Event

Research on Causal World Models under Physical Parameter Shifts

1 reports · 1 sources

Chu Lin and Yi Sun propose letting physical parameters enter the dynamics computation as explicit affine coefficients in a world model, and restricting the inputs available to the predictor; results show that this method significantly outperforms control models when physical parameters fall outside the training distribution. In the CausalWorld connecting-rod mass extrapolation task (training and test mass ranges do not overlap), compared with the control that fixes the mass input to the training mean, the model using a known PD controller raised the median joint R² of 150-step rollout prediction from 0.609 to 0.727, with all ten task–random-seed combinations improving (sign test p=0.002). On a synthetic benchmark with shortcut features, a frozen dependency mask derived from paired interventions reduced the out-of-distribution RMSE of a fully connected predictor from 1.769 to 0.697, close to the generator oracle mask's 0.694. In shared 100-step prediction, the median joint R² of the learned torque model (0.755–0.768) was comparable to the strongest recurrent world model configuration (0.752), whereas directly inputting out-of-range mass caused clear degradation in both the MLP and the recurrent predictor.