The paper proposes a neural network–based learned heuristic that accelerates MPE inference in high-treewidth PGMs by predicting impactful variable assignments from solver traces, significantly reducing search space while preserving solution quality.