Reasoning Shortcuts in Neurosymbolic Learning Formally Analyzed

A recent preprint on arXiv (2604.23377) establishes reasoning shortcuts in neurosymbolic learning as a problem of constraint satisfaction. The researchers demonstrate that a specific discrimination property—ensuring that no valid concept mapping can be altered into another by merely swapping two concept values—is essential for learning without shortcuts under bijective mappings. However, they provide a counterexample indicating that this property alone is not adequate, even when a connected constraint graph is present. They introduce an ASP-based algorithm that checks if a particular constraint set uniquely defines the desired concept mapping, ensuring both soundness and completeness. Additionally, when shortcuts are identified, a greedy repair algorithm works to eliminate them by expanding the constraint set.