Synthetic Neurosymbolic Task Generation with Controlled Constraint Complexity
Explicit constraint enforcement helps in some generated tasks; in others, unconstrained models learn the logical structure from data.
The authors introduce NeSySy, a generator that produces logical constraints alongside datasets guaranteed to satisfy them. It lets researchers vary constraint structure, tightness, input dependence and data-generation strategies, including constraints over infinite domains. Solvers certify constraints and compute repairs to keep every generated sample consistent. Across the constraint families tested, the authors report systematic and largely independent effects of these controls on task difficulty.
ArXiv · AI/CL/LG's note
The authors introduce NeSySy, a generator that produces logical constraints alongside datasets guaranteed to satisfy them. It lets researchers vary constraint structure, tightness, input dependence and data-generation strategies, including constraints over infinite domains. Solvers certify constraints and compute repairs to keep every generated sample consistent. Across the constraint families tested, the authors report systematic and largely independent effects of these controls on task difficulty.
ArXiv · AI/CL/LG's note
score 4