Resolve the contribution of inverse dynamics recovery

Determine whether the inverse-dynamics action-recovery objective provides a reliable incremental benefit beyond residual prediction and normalized mutual-information action recovery in AD-WM.

Background

AD-WM combines residual prediction, inverse dynamics recovery, and normalized mutual-information recovery. The ablations indicate that inverse dynamics has a smaller and weight-dependent effect than the other components. In the planning diagnostics, the successive improvement attributable to the inverse-dynamics-only increment has a confidence interval spanning zero, leaving its contribution unresolved.

References

The Inv-only increment remains unresolved; Res+MI and AD-WM have similar mean regret.

— AD-WM: Action-Discriminative World Models for Counterfactual Model Predictive Control  (2609.30264 - Qiu et al., 24 Sep 2026) in Section 5.3, Planning Diagnostics