Relational Learning as World Modeling

Relational Learning as World Modeling. By Zheyu Zhang and one other, from the Technical University of Munich and the Munich Center for Machine Learning. Published on the twelfth of October. Its code repository is public and real: about thirty source files.

Relational learning usually predicts from the state of a database. Decision making asks something else: how will that state change if an action is taken? In operational systems the answer depends on dynamics specific to each system, and those dynamics may not be visible in the database at all. The paper treats each such system as a relational world, and asks whether a model can infer how that world behaves just by watching executions of it. Its claim is that it can: relational dynamics can be learned in context.

Two pieces make the test possible. WorldRel is a generator of executable relational worlds, used for pretraining. TabWM, short for Tabular World Model, is the model. Given executions from a world it has never seen, TabWM takes them as context and predicts the sparse, structured edits a new action will cause, with no parameter updates.

The bar for a correct prediction is strict. An edit counts only if its operation, its relational location, and its value all match. Across five hundred and twelve held-out worlds, complete-edit F1 is zero point zero one six when no executions are observed. With sixty-four observed executions, it rises to zero point six eight five. Since the predictions are explicit relational edits, they can also be applied recursively, imagining candidate futures and comparing actions before taking them. A broader search over these predicted futures improves the realized return of the action sequences selected.

The evidence is bounded in ways the abstract makes visible. The test worlds are held out, but they are worlds of the kind WorldRel generates, not operational systems drawn from deployment. The headline measure is exact agreement on structured edits, at a stated number of observed executions. For planning, the abstract reports that wider search helps the realized return, and gives no figure for how much. The authors themselves frame the conclusion cautiously, saying the results suggest in-context learning of these dynamics rather than establish it.

If the suggestion holds, the scope of relational learning widens. It moves from describing what a system looks like now to anticipating how it will change, which is the question a decision actually turns on.