The abstract of "EvoSim: Learning to Model, Modeling to Learn" opens by noting that physics-based models are valuable precisely because they connect scientific explanation with quantitative prediction. It then lists what constructing such models demands: selecting physical processes, defining states and governing equations, specifying couplings, and identifying parameters. The available source ends there.
The title suggests the paper's core idea: that modelling is not just a means to an end but a learnable activity in its own right. Read literally, "learning to model" and "modeling to learn" point to a loop in which a system improves its models and, through that process, learns about the system it is studying. This interpretation is inferred from the title and opening lines; the source does not yet describe how EvoSim implements this loop.
Because only the abstract's first sentences are available, this digest can report the paper's motivation but not its contributions. Readers interested in the actual mechanism or experimental results will need the full preprint.