World models are a powerful means of reasoning and planning, but many action-conditioned predictive models are high-dimensional and computationally expensive. This limits the number of action candidates that can be considered, which is a key bottleneck for planning and decision-making.
The new paper, posted on arXiv, introduces "Action Jets" as a method to learn the local shape of the action-conditioned data manifold. Rather than modeling the full high-dimensional space directly, the approach appears to target the manifold's local geometry, potentially reducing computational cost.
The abstract is truncated, so specific methodological details and experimental results are not available from the source. The contribution, as described, is a way to make action-conditioned world models more efficient by exploiting the structure of the data manifold.