Q-BRIDGE uses graph learning to denoise quantum program tests with 94% accuracy

Testing quantum programs on today’s noisy NISQ (Noisy Intermediate-Scale Quantum) backends is notoriously unreliable because hardware noise distorts the outcome distributions that determine pass/fail decisions. To solve this, the authors introduce Q-BRIDGE, a graph learning approach that converts noisy observations into denoised distributions suitable for oracle-based verification. Q-BRIDGE employs a graph transformer architecture that encodes the transpiled quantum circuit—capturing gates, connectivity, and physical backend characteristics—while a FiLM (Feature-Wise Linear Modulation) conditioning layer integrates the noisy observations to produce clean outputs.

Evaluated on 23 IBM noise backends across 6 circuit families representative of real workloads, Q-BRIDGE was tested in two settings: per-backend models and a single general model shared across all backends. In both cases, Q-BRIDGE significantly outperformed the previous state-of-the-art baseline in noise mitigation. Specifically, it achieved 93.97%–94.90% precision and 82.50%–83.51% recall in detecting bug-induced test failures. These results demonstrate that leveraging the graph structure of transpiled circuits and backend-specific noise characteristics is a practical path to more reliable quantum software testing.