Validate controllability of the factored representation

Determine whether the controllable groups of the CausalDreamer factored representation capture the effects of the agent’s actions.

Background

CausalDreamer routes the action input to the two groups designated as controllable, but this routing is imposed architecturally rather than established by a dedicated learning objective. The paper reports that it does not evaluate whether these groups actually encode the consequences of actions, leaving the semantic validity of the controllability factor unresolved.

References

Within the factored representation itself, the routing of the action is an architectural choice without an objective of its own, and we do not evaluate whether the controllable groups capture the effect of the action.

— CausalDreamer: Learning Predictive World Models with Latent Disentanglement  (2610.12016 - Jha et al., 8 Oct 2026) in Appendix, Section “Limitations and Future Work”