A new arXiv preprint studies how to improve decision policies from observational data when the actions themselves are complex, structured objects such as text messages or images. This setting differs from standard policy improvement, where actions are typically simple choices from a small set.

The authors note that the task faces two major challenges, though the abstract does not specify what they are in the available text. The work appears to target domains where generating better actions requires understanding high-dimensional, structured outputs.

If successful, such an approach could enable policy improvement without costly online trials, using existing logs of decisions and outcomes. The paper is available on arXiv under the identifier 2610.23436.