NeSySy: a synthetic generator for controlled neurosymbolic learning tasks
Researchers introduce NeSySy, a tool that builds logical constraints and matching datasets with adjustable difficulty, using solvers to certify every generated sample is consistent.
In plain words
Neurosymbolic learning combines neural networks with logical rules. NeSySy is a generator that creates practice tasks made of such rules plus data that follow them, letting researchers dial properties like rule complexity up or down. Solvers check that each task and sample is valid, so experiments can measure what actually makes these problems hard.
What was announced
The arXiv paper describes NeSySy as a synthetic generator that jointly produces logical constraints and consistent datasets for neurosymbolic learning. Unlike existing benchmarks, it offers fine-grained control over properties such as constraint structure, tightness, input dependence, and data generation and alignment strategies. It also supports rich constraints over infinite domains, and uses solvers to keep constraints meaningful and ensure generated samples satisfy them. The authors designed it for both learning with declarative constraints and learning from distant supervision.
Across a range of constraint families, the authors report evidence that the generator's parameters induce systematic and largely independent effects on task difficulty. Using NeSySy, they studied constraint-aware learning methods and identified regimes where explicit constraint enforcement helps, and others where expressive unconstrained models can learn constraint structure directly from data. The pipeline relies on satisfiability checkers CNFgen and Z3 to generate and certify constraints, while constrained optimizers RC2 for MaxSAT and Z3 for OMT compute repairs that guarantee every sample is consistent. The authors position NeSySy as a controlled testbed for analysing how learning interacts with logical structure.
What it does not show
- The evidence is the paper's abstract only; the reported difficulty effects and learning findings are author claims not yet independently verified, and full experimental details are outside the supplied scope.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.