Neural Constraint Reasoning Requires Symbolic Integration

This paper argues that neural solvers for constraint satisfaction problems struggle with distribution shifts despite high confidence. It proposes that neural-only methods must be replaced or augmented by symbolic integration to guarantee correctness under hard constraints.