<ul><li>This contribution presents a comprehensive evaluation of graph neural networks used for Boolean satisfiability problems.</li><li>The model incorporates training improvements, such as a novel closest assignment supervision method.</li><li>The study demonstrates the effectiveness of variable-clause graph representations and recurrent neural network updates.</li><li>The network's reasoning process resembles continuous relaxations of MaxSAT, enabling interpretability and scalability beyond the training distribution.</li></ul>

Neural Approaches to SAT Solving: Design Choices and Interpretability

Discover more