The paper addresses a core design choice in visual world models: what information the latent state should be encouraged to preserve. Models trained with only a latent-prediction objective adapt to environment dynamics, but the abstract suggests this approach has shortcomings, though the details are cut off in the available text.
The proposed alternative, implied by the title, is to treat pretrained features as teachers rather than as states. This appears to guide the latent space toward more useful representations, resulting in compact latent world models suited for visual planning.
Because the source is a single truncated abstract, the full method and results are not yet available. The significance lies in the potential to simplify world models by leveraging pretrained features, which could improve planning efficiency without sacrificing predictive accuracy.