Establish generalization beyond the evaluated regimes

Determine whether the observed separation between argmax accuracy and calibration for Moose, Moose+BEARS, and Moose+NeSyDM generalizes beyond the relational, role-chain, and Pizzäiolo regimes evaluated in the paper.

Background

The experiments show that the relative performance of the reasoning-shortcut mitigation methods depends on the supervision regime. BEARS generally improves family or latent-concept accuracy under symbolic ambiguity, whereas NeSyDM often provides better calibration; in relational settings where forward subsumption identifies the latent vector, plain Moose performs best.

The paper does not establish whether this regime-dependent pattern persists for other ontologies, supervision structures, ambiguity patterns, or domain sizes. The authors explicitly defer that generalization question to future work.