The Signal
CoDecoder Cuts Quantum Decoding Time by Up to 124×
Researchers report that CoDecoder, a neural front end for existing quantum error-correction decoders, reached a 124-fold speed-up in a benchmark on a bivariate-bicycle code. In registered tests, the hybrid approach also reduced decoding fai
Quantum.world Team · Written with AI · fact-checked ·
A 124-fold result on one code
A neural front end cut decoding time by as much as 124-fold in tests of CoDecoder, a system designed to work alongside existing quantum error-correction decoders. The result came from a specific comparison: on the [[72,12,6]] bivariate-bicycle code at a physical error rate of 0.2%, CoDecoder routed 0.245% of shots to full-DEM Relay-BP, reduced failures from 288 to 204 per million shots, and recorded the largest speed-up in the study.
The researchers submitted the paper to arXiv on October 20, 2026. Their broader registered tests found that CoDecoder paired with correlated matching reduced failures by 6.4% to 31.4% across surface codes of distance 7 to 17. With full-DEM Relay-BP on two bivariate-bicycle codes, it cut failures by 22% to 50%; speed-ups across those comparisons ranged from 2.3-fold to 124-fold.
“In registered tests under circuit-level noise, composition reduces failures by 6.4-31.4% relative to correlated matching on surface codes of distance 7-17 and by 22-50% relative to full-DEM Relay-BP on two BB codes, with 2.3-124-fold speed-ups across these comparisons.”
Yicheng Guang, Yuliang Xiong, Xun Gao and Yueqi Chen, authors of “CoDecoder: Make Quantum Decoders Up to 100x Faster and Even More Accurate”
Confident answers stay; difficult shots go on
CoDecoder predicts a shot’s logical outcome and assigns it a confidence score. If that score clears a threshold, the system keeps the neural network’s answer; otherwise, it sends the original detector data to a fallback decoder, whose algorithm remains unchanged. That selective routing aims to avoid running the slower decoder on every shot.
The design also lets two decoders combine their answers on routed shots when the fallback is another neural decoder that provides logits. The authors say this joint decision improved accuracy over their Cascade re-implementations in all four tested settings, with 14% to 18% fewer failures on the [[72,12,6]] code. Those Cascade comparisons used the researchers’ own re-implementations, not the original Cascade code or trained networks, which the paper says are not public.
Benchmark gains, not a live-machine demonstration
The experiments used memory tests under circuit-level noise, with registered settings fixed before the test shots were drawn. Timing involved NVIDIA H100 and RTX PRO 6000 GPUs; CPU components in the classical-decoder comparisons used a timing basis derived from one workstation core and NVIDIA-published CPU timings. The 124-fold figure therefore describes a measured benchmark comparison, not a demonstrated reduction in the runtime of an operating quantum computer.
The headline result also depends on the fallback and code being tested. The paper reports that plain matching was slightly faster than the composition at surface-code distance 7: 11.89 microseconds against 11.94 microseconds, though it failed 1.9 times as often. On the gross bivariate-bicycle code, a different composition took 1.49 times as long as CUDA-Q BP-OSD under the authors’ Grace-based timing conversion.
Tail latency remains a constraint
Mean decoding time is not the only concern when a decoder must return an answer at a particular point in a computation. The authors found higher 99.9th-percentile single-shot latency for every tested CoDecoder-Cascade composition than for Cascade alone on both GPUs: routed shots pass through both networks. On the [[72,12,6]] tests, the increase ranged from 7.2% to 10.3% on the RTX PRO 6000 and from 9.7% to 19.1% on the H100.
The 124-fold improvement is a result for one registered setup, with a small fraction of shots sent to a specific fallback—not a general speed guarantee. In the paper’s largest reported speed-up case, 0.245% of shots were routed, and failures fell from 288 to 204 per million.