StarWM: Self-Supervised Trained Attention Routing for Robust World Models
StarWM is proposed, which uses a cross-attention module trained on self-supervised dynamics to decide where reconstruction applies and preserves state attributes with near-perfect fidelity through long-horizon imagination while systematically discarding distractors.
Zeqiang Zhang, Fabian Wurzberger, Maximilian Otte et al.
· 0 citations