Accurate Leakage Speculation for Quantum Error Correction

Quantum Error Correction (QEC) can correct bit- and phase-flip faults in the |0⟩ /|1⟩ subspace. Unfortunately, physical qubits also leak into higher levels such as |2⟩. These leaked states contaminate every downstream syndrome measurement and can spread from one qubit to another; the counter-measure is to insert leakage-reduction circuits (LRCs) throughout the code cycle. Because LRCs are themselves noisy, they must be scheduled only when leakage is truly present. Prior work, including ERASER, is surface-code-specific and misclassifies many benign syndromes, triggering a large number of LRCs that increases the average QEC cycle time and inflate the logical-error rate. Our proposal, GLADIATOR, is a generalizable and adaptable leakage-speculation framework that works unchanged for surface, color, and LDPC-style codes. Offline, it constructs a code-aware error-propagation graph weighted by device calibration data; online, it classifies each syndrome in a few nanoseconds and schedules an LRC only when the pattern is provably leakage-dominated. This accurate speculation eliminates up to 3× (and on average 2×) un-necessary LRC insertions, shortens the QEC cycle, and suppresses false positives at their source. Across standard fault-tolerant benchmarks, these leaner cycles yield 1.7×–3.9× speed-ups and a 16% relative reduction in logical-error rate.