Test confinement of reward-relevant content

Determine whether reward-relevant content is confined to the reward-relevant groups of the CausalDreamer factored representation.

Background

The reward predictor in CausalDreamer reads only the groups designated as reward-relevant, and the analysis tests whether changes that do not affect reward are mostly routed into the reward-irrelevant groups. However, the paper does not test the converse property: whether information that does affect reward is absent from the reward-irrelevant groups. Establishing this would assess whether the learned factorization reliably isolates reward-relevant information.

References

Our analysis covers changes that do not affect the reward, in two tasks, and does not test whether reward-relevant content is confined to the reward-relevant groups.

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