Resgru: Syndrome decoder for heavy-hexagon quantum error correction code

Abstract Decoding is a central bottleneck in near-term quantum error correction (QEC), where repeated syndrome extraction must be translated into reliable logical outcomes under realistic noise and hardware constraints. We study heavy-hexagon (HH) quantum memory experiments and propose Resgru , a lightweight recurrent neural decoder that maps time-ordered detection events to a binary logical-class prediction for each shot. We benchmark Resgru against a standard minimum-weight perfect matching (MWPM) decoder instantiated on the same detector error model for HH code at code distances d ∈ { 3 , 5 , 7 } under circuit-level depolarizing noise with physical error rate p = 0.1 % . Resgru consistently achieved higher logical fidelity than MWPM decoder and yields lower per-round logical error rates, including a 28% reduction for d = 5 and a 14% reduction for d = 7 . Despite being trained only on short sequences, Resgru maintains an advantage when evaluated on longer simulated memory experiments up to 300 QEC rounds.