QAOA Query Cost Cut via Graph Neural Trust Regions
arXiv paper proposes a graph-conditioned trust-region method to reduce objective evaluations in shallow QAOA circuits, where query count dominates cost.
A GNN predicts a Gaussian distribution over angles; the mean seeds a local optimizer, covariance bounds the search region, and predicted uncertainty sets an adaptive evaluation budget.
Theoretical bounds derived under smoothness, curvature, calibration, and noise assumptions.