Extend MEND to stochastic dynamics and on-manifold hallucinations

Extend the study of MEND, a label-free conditional score-field method for detecting, localising, and correcting latent hallucinations in frozen world models, to stochastic dynamics and settings in which prediction errors remain close to the valid-state manifold.

Background

MEND assumes that hallucinated predictions lie off the manifold of valid next latent states. The paper notes that this assumption holds in the deterministic Wall and PointMaze navigation environments studied, but may fail when prediction errors remain close to the valid-state manifold. In stochastic environments, multiple future states may be valid, so reachability-based detection and correction must account for alternative legitimate outcomes rather than treating deviations from a single realised successor as hallucinations.

The authors explicitly identify extending the approach to stochastic dynamics as an open direction because the current detection formulation and experimental proxy are not established for that setting.

References

However, two directions remain open. First, the approach assumes that hallucinations lie off-manifold, which holds for the navigation environments studied here but may not hold when a predictor's errors stay close to the manifold. Hence, extending the study to stochastic dynamics is a natural next step.

— MEND: Label-Free Detection, Localisation, and Correction of Latent Hallucination in World Models  (2609.39182 - Alrasheed et al., 30 Sep 2026) in Section Conclusion

Second, correction is established at the representation level, and whether these gains transfer to a downstream controller is left for the future work.

— MEND: Label-Free Detection, Localisation, and Correction of Latent Hallucination in World Models  (2609.39182 - Alrasheed et al., 30 Sep 2026) in Section Conclusion