Determine whether successful OOD computation is induced or human-supplied

Determine whether the structure enabling a system to generalize to a novel regime was induced autonomously by the system or supplied in advance by a human designer.

Background

The paper contrasts transductive systems, which consult fitted models during inference, with inductive systems that produce and execute reusable programs. It argues that exact execution explains success on precise computations and compositional tasks, while approximation can suffice for fuzzier perceptual tasks.

The authors leave unresolved the provenance of the exact structure in successful systems: it may be induced through search or learning, or it may reflect representational structure manually supplied by a person. Resolving this question bears directly on whether future architectures can overcome the paper’s claimed dependence on hand-designed structure.

References

The claim is that where a system generalizes to a novel regime something in it computed the mechanism exactly, and that the open question is whether that something was induced by the system or supplied by a person.

— Exactness at Inference: A Representational Criterion for Out-of-Distribution Generalization  (2609.24942 - Rocha et al., 21 Sep 2026) in Section 9.5, “The evidence from abstraction benchmarks” (Section \ref{sec:arc})