A new preprint on arXiv introduces Beckmann World Models, a method for action-conditioned prediction in which the model learns a terminal map directly. The authors argue that while world models are useful for planning and policy evaluation, learning an accurate one-step predictor can be computationally expensive. Their proposed approach is designed to address that bottleneck.
The paper is currently available as an abstract-only announcement, so details of the architecture and experiments are not yet public. The central idea, however, is to replace the usual training procedure with direct terminal-map learning, which may offer a more efficient route to one-step prediction. As with any preprint, the claims have not yet been peer-reviewed, and further evidence will be needed to confirm the practical gains.