Applications of machine learning (ML) in mixed-integer program (MIP) optimization is an active area of research. The human-designed heuristics used by MIP solvers rely on domain expertise and years of experience with the expectation of applicability to only a specific problem class. The underlying question is whether training can simulate expertise and experience. We consider this question in the problem of finding feasible solutions for MIP instances. We develop an ML agent based on a graph neural network approach to local search for Bool SAT. While we encounter some evidence of learning, we ultimately face issues with scaling to practical instance sizes and learning difficult problems.