Mitigating Reasoning Shortcuts in Abductive Learning with Semantic-Layer Initialization

Abductive Learning (ABL) is a promising framework for integrating machine learning with logical reasoning. However, its generalization can be undermined by reasoning shortcuts, where seemingly strong reasoning performance masks unintended meanings of intermediate concepts. We trace this vulnerability to the learning–reasoning interface in ABL: associating input patterns with reasoning concepts is entangled with representation learning, leaving semantic alignment to random initialization and training dynamics. Meanwhile, the prevailing weak-binding practice treats inputs independently during reasoning, weakening cross-example constraints on what a predicted label means. To address these problems, we reformulate ABL with an explicit semantic layer that decouples semantic alignment from representation learning while enabling strong binding across related inputs. Under idealized perfect strong binding, we formalize shortcuts as concept transformations preserving reasoning outcomes and distinguish benign permutations from malignant non-injective collapses, yielding an identifiability limit up to benign equivalence. Guided by this analysis, we propose BASIL (Binding-Alignment for Semantic Initialization via Logic-consistency), a two-stage initialization. Stage I clusters inputs for strong binding; Stage II performs training-free semantic alignment by maximizing logic consistency. Experiments on visual and symbolic tasks show that BASIL substantially improves ABL's robustness to reasoning shortcuts.