Evolving Decoding Algorithms for Quantum Error Correction

Fault-tolerant quantum computation entails robustness engineering: a noisy quantum state is measured continuously, and a classical algorithm —the decoder— must infer corrections that reliably remove the noise. Using a Domain-Specific Language (DSL) consisting of 38 typed operations, we synthesize new decoders by syntactic evolution, yielding inspectable algorithmic programs rather than opaque weights. Sixteen evolved candidates, frozen before their exam data was generated, were judged on 200 000 fresh shots from the [[72,12,6]] bivariate bicycle code against 13 published reference decoders, each at its best budget-admissible setting: 12 cleared a pre-registered dominance bar, from seeded and unseeded arms alike. Three evolved decoders are more accurate than BP+OSD-0 at lower cost, while simultaneously outperforming BP+OSD-CS at under 5% of its cost. The most accurate clearing decoder is evolved from a reference decoder as seed; it is more accurate than every reference, outperforming the leading BP+AC decoder by +0.51 pp (z = 16.7) at equivalent cost. The most original new decoder —evolved sans seeding, with no language model— has a composition matching no reference decoder, yet still dominates the leader (z = 6.11) at equal cost. We argue that the DSL is where the expert knowledge lives, serving as a robust intermediate layer between people and machines.