The GraphDOP paper mentions that an optional rollout in observational space can be implemented. However, the paper doesn't clearly specify whether GraphDOP was trained solely on latent rollout or using a mixed approach (both latent and observational rollouts). It would be valuable to run a sensitivity test comparing both methods to evaluate how each rollout type affects model performance.
