Engineering · 4 min read · April 29, 2026
Graph Neural Networks Cut QAOA Query Cost by 87%
A trust-region method using GNNs to predict QAOA parameter distributions reduces circuit evaluations while preserving solution quality on small graphs.
Graph neural networks predict parameter distributions to guide QAOA optimization with fewer circuit evaluations.
- — GNN predicts mean and covariance of QAOA angles, defining search policy not just initial guess.
- — Trust region constrains parameter space; predicted uncertainty sets instance-dependent evaluation budget.
- — Method reduces mean evaluations from 343 (random) to 45 on MaxCut depth-2 problems.
- — Maintains approximation ratios within 3 percentage points of baseline heuristics.
- — Learned trust regions transfer to untrained graph sizes without retraining.
- — Theoretical bounds derived for local smoothness, gradient variance, and noise robustness.
- — Advantage is query efficiency, not improved absolute solution quality.
Frequently asked
- It is not faster in wall-clock time; it reduces the number of circuit evaluations by 87% (from 343 to 45 on small MaxCut problems). Since each evaluation is expensive on quantum hardware, fewer evaluations means lower total cost. Absolute solution quality remains comparable to baselines.