We proposes a novel method that enables Graph Neural Networks (GNNs) to solve SAT problems by leveraging a technique developed for applying GNNs to Mixed Integer Linear Programming (MILP). Specifically, k-CNF formulae are mapped into MILP problems, which are then encoded as weighted bipartite graphs and subsequently fed into a GNN for training and testing. From a theoretical perspective: (i) we establish permutation and equivalence invariance results, demonstrating that the method produces outputs that are stable under reordering of clauses and variables; (ii) we identify a theoretical limitation, showing that for a class of formulae called foldable formulae, standard GNNs cannot always distinguish satisfiable from unsatisfiable instances; (iii) we prove a universal approximation theorem, establishing that with Random Node Initialization (RNI), the method can approximate SAT solving to arbitrary precision on finite datasets, that is, the GNN becomes approximately sound and complete on such datasets. Furthermore, we show that for unfoldable formulae, the same approximation guarantee can be achieved without the need for RNI. Finally, we conduct an experimental evaluation of our approach, which show that, despite the simplicity of the neural architecture, the method achieves promising results.
MILP-SAT-GNN: yet another neural SAT solver
Cardillo F. A.;Straccia U.
2025
Abstract
We proposes a novel method that enables Graph Neural Networks (GNNs) to solve SAT problems by leveraging a technique developed for applying GNNs to Mixed Integer Linear Programming (MILP). Specifically, k-CNF formulae are mapped into MILP problems, which are then encoded as weighted bipartite graphs and subsequently fed into a GNN for training and testing. From a theoretical perspective: (i) we establish permutation and equivalence invariance results, demonstrating that the method produces outputs that are stable under reordering of clauses and variables; (ii) we identify a theoretical limitation, showing that for a class of formulae called foldable formulae, standard GNNs cannot always distinguish satisfiable from unsatisfiable instances; (iii) we prove a universal approximation theorem, establishing that with Random Node Initialization (RNI), the method can approximate SAT solving to arbitrary precision on finite datasets, that is, the GNN becomes approximately sound and complete on such datasets. Furthermore, we show that for unfoldable formulae, the same approximation guarantee can be achieved without the need for RNI. Finally, we conduct an experimental evaluation of our approach, which show that, despite the simplicity of the neural architecture, the method achieves promising results.| Campo DC | Valore | Lingua |
|---|---|---|
| dc.authority.orgunit | Istituto di linguistica computazionale "Antonio Zampolli" - ILC | en |
| dc.authority.orgunit | Istituto di Scienza e Tecnologie dell'Informazione "Alessandro Faedo" - ISTI | en |
| dc.authority.people | Cardillo F. A. | en |
| dc.authority.people | Khyari H. | en |
| dc.authority.people | Straccia U. | en |
| dc.authority.project | corda_____he::86c21b1aa82d5bdc53411947d7ebd9f8 | en |
| dc.authority.project | Future Artificial Intelligence Research | en |
| dc.collection.id.s | 95773a9f-8d06-4466-a951-5d4e15d70690 | * |
| dc.collection.name | 08.04 Rapporto tecnico | * |
| dc.contributor.appartenenza | Istituto di Scienza e Tecnologie dell'Informazione "Alessandro Faedo" - ISTI | * |
| dc.contributor.appartenenza | Istituto di linguistica computazionale "Antonio Zampolli" - ILC | * |
| dc.contributor.appartenenza.mi | 918 | * |
| dc.contributor.appartenenza.mi | 973 | * |
| dc.contributor.area | Non assegn | * |
| dc.contributor.area | Non assegn | * |
| dc.date.accessioned | 2026/03/04 14:51:48 | - |
| dc.date.available | 2026/03/04 14:51:48 | - |
| dc.date.firstsubmission | 2026/03/04 12:47:15 | * |
| dc.date.issued | 2025 | - |
| dc.date.submission | 2026/03/09 11:58:23 | * |
| dc.description.abstracteng | We proposes a novel method that enables Graph Neural Networks (GNNs) to solve SAT problems by leveraging a technique developed for applying GNNs to Mixed Integer Linear Programming (MILP). Specifically, k-CNF formulae are mapped into MILP problems, which are then encoded as weighted bipartite graphs and subsequently fed into a GNN for training and testing. From a theoretical perspective: (i) we establish permutation and equivalence invariance results, demonstrating that the method produces outputs that are stable under reordering of clauses and variables; (ii) we identify a theoretical limitation, showing that for a class of formulae called foldable formulae, standard GNNs cannot always distinguish satisfiable from unsatisfiable instances; (iii) we prove a universal approximation theorem, establishing that with Random Node Initialization (RNI), the method can approximate SAT solving to arbitrary precision on finite datasets, that is, the GNN becomes approximately sound and complete on such datasets. Furthermore, we show that for unfoldable formulae, the same approximation guarantee can be achieved without the need for RNI. Finally, we conduct an experimental evaluation of our approach, which show that, despite the simplicity of the neural architecture, the method achieves promising results. | - |
| dc.description.allpeople | Cardillo, F. A.; Khyari, H.; Straccia, U. | - |
| dc.description.allpeopleoriginal | Cardillo F.A.; Khyari H.; Straccia U. | en |
| dc.description.fulltext | open | en |
| dc.description.note | DOI ISTI - 10.32079/ISTI-TR-2025/020 | en |
| dc.description.numberofauthors | 3 | - |
| dc.identifier.doi | 10.48550/arXiv.2507.01825 | en |
| dc.identifier.source | arxiv | * |
| dc.identifier.uri | https://hdl.handle.net/20.500.14243/571029 | - |
| dc.identifier.url | http://arxiv.org/abs/2507.01825v1 | en |
| dc.language.iso | eng | en |
| dc.relation.medium | ELETTRONICO | en |
| dc.relation.numberofpages | 19 | en |
| dc.relation.projectAcronym | STARWARS | en |
| dc.relation.projectAcronym | FAIR | en |
| dc.relation.projectAwardNumber | 101086252 | en |
| dc.relation.projectAwardNumber | - | en |
| dc.relation.projectAwardTitle | STormwAteR and WastewAteR networkS heterogeneous data AI-driven management | en |
| dc.relation.projectAwardTitle | Future Artificial Intelligence Research | en |
| dc.relation.projectFunderName | European Commission | en |
| dc.relation.projectFunderName | - | en |
| dc.relation.projectFundingStream | Horizon Europe Framework Programme | en |
| dc.relation.projectFundingStream | - | en |
| dc.subject.keywordseng | Computer Science - Learning | - |
| dc.subject.keywordseng | Computer Science - Artificial Intelligence | - |
| dc.subject.singlekeyword | Computer Science - Learning | * |
| dc.subject.singlekeyword | Computer Science - Artificial Intelligence | * |
| dc.title | MILP-SAT-GNN: yet another neural SAT solver | en |
| dc.type.driver | info:eu-repo/semantics/other | - |
| dc.type.full | 08 Report e Working Paper::08.04 Rapporto tecnico | it |
| dc.type.miur | 298 | - |
| iris.mediafilter.data | 2026/03/28 03:41:15 | * |
| iris.orcid.lastModifiedDate | 2026/03/27 18:23:36 | * |
| iris.orcid.lastModifiedMillisecond | 1774632216265 | * |
| iris.sitodocente.maxattempts | 1 | - |
| iris.unpaywall.metadataCallLastModified | 29/03/2026 07:39:27 | - |
| iris.unpaywall.metadataCallLastModifiedMillisecond | 1774762767645 | - |
| iris.unpaywall.metadataErrorDescription | 0 | - |
| iris.unpaywall.metadataErrorType | ERROR_NO_MATCH | - |
| iris.unpaywall.metadataStatus | ERROR | - |
| Appare nelle tipologie: | 08.04 Rapporto tecnico | |
| File | Dimensione | Formato | |
|---|---|---|---|
|
Copertina_Frontespizio_2025-TR-020.pdf
accesso aperto
Descrizione: MILP-SAT-GNN: yet another neural SAT solver
Tipologia:
Altro materiale allegato
Licenza:
Altro tipo di licenza
Dimensione
292.61 kB
Formato
Adobe PDF
|
292.61 kB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


