MAS-SAT: Synergizing ML-Assisted and Standalone Solvers for SAT Solving

Chengdi Cao, Cho-Jui Hsieh, Jason Cong

Keywords: Boolean Satisfiability, Graph Neural Networks

Abstract: Machine learning has emerged as a promising approach to accelerate the solving of Boolean satisfiability (SAT), with prior research exploiting graph neural networks (GNNs) either as standalone solvers or to assist existing SAT solvers. Despite their contributions, the two paradigms mainly suffer from poor scalability and limited compute budgets, respectively. To address these challenges, we propose **M**achine learning **A**ssisting and **S**olving **SAT** (MAS-SAT), a novel framework that synergizes the strengths of both paradigms while mitigating their weaknesses. In MAS-SAT, standalone solvers demonstrate improved scalability when solving sub-problems generated by ML-assisted solvers, and ML-assisted solvers achieve better performance under limited compute budgets by leveraging the parallel search ability of standalone solvers. In addition, we develop an efficient asynchronous deployment strategy with influences heuristics to further optimize MAS-SAT. Extensive experiments across diverse domains and architectures demonstrate that MAS-SAT consistently outperforms both paradigms. When deployed, MAS-SAT achieves up to $2.4\times\$ median speedup on hard instances and solves $3$ more instances on SAT Competition 2023 compared to the base state-of-the-art solver, kissat.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 24377