Fully latent structural inference without hidden-state supervision

Develop structural inference with fully latent hidden agents when hidden-agent trajectories are unavailable during both training and deployment, without relying on direct hidden-state supervision from complete training trajectories.

Background

The current SIHA formulation assumes that complete multi-agent trajectories are available during training so that hidden-state predictors can be supervised using the trajectories of masked agents. Ground-truth interaction labels are not required, but hidden-state supervision remains essential to the reported method.

Experiments with training objectives based only on visible-agent future prediction perform substantially worse than the supervised formulation. Consequently, the paper identifies structural inference with fully latent hidden agents—where hidden trajectories are unavailable during both training and deployment—as an unresolved problem requiring self-supervised hidden-state reconstruction and structural inference.