The Signal
Neural Triage Posts 28×–343× qLDPC Speedups in Benchmark
A neural decoder called Triage cuts reported decoding time for two high-rate quantum LDPC codes in simulations. The October 29 preprint reports speedups of 28× and 343×, but compares GPU batch processing with CPU baselines.
Quantum.world Team · Written with AI · fact-checked ·
“Real-time operation also requires measuring single-shot latency, its distribution, and sustained throughput against the timing requirements of the hardware.”
Hong-Ye Hu, Zhenlai Wang, Binzhao Luo, Tao Xiang and Chen Zhao, authors of the paper
The warning comes from a new decoder study posted to arXiv on October 29. The authors’ neural “Triage” method reports substantially lower decoding times than selected conventional algorithms on two high-rate quantum low-density parity-check (qLDPC) codes. But the headline speedups come from an amortized GPU-versus-CPU benchmark, not a demonstration of real-time feedback on a quantum computer.
Triage gives difficult syndromes extra work
Quantum error-correction decoders infer likely physical errors from syndrome measurements: checks that reveal whether the measured state has violated the code’s parity constraints. Triage starts with NeuralMP, a neural message-passing network that proposes corrections across the code’s sparse graph. It then checks whether each proposed correction resolves the syndrome.
Shots that the network resolves return from the initial or iterative stages; unresolved cases receive additional attempts. The later “diversification” stage changes the inference conditions through techniques including syndrome flips and perturbed error priors, then checks candidate corrections against the original syndrome. In the Kasai test at a physical error rate of 1.5%, the search resolved 99.6% of shots left unresolved after iteration, reducing the unresolved fraction from 2.6 × 10-4 to 1.0 × 10-6 per shot.
Two codes, hundreds of encoded qubits
The team tested a lifted-product code with 1,122 data qubits and 148 logical qubits, and a Kasai code with 1,152 data qubits and 580 logical qubits. The latter has a reported encoding rate of 50.3%. Both tests used phenomenological noise models rather than a full model of errors in a hardware circuit; the simulations included repeated noisy measurements and ended with a perfect syndrome extraction.
At a physical error rate of 1.5%, the estimated logical error rate per qubit per round was 9.33 × 10-11 for the lifted-product code and 5.08 × 10-10 for Kasai. The authors report that Triage’s accuracy was comparable to or better than the algorithmic decoders they tested. The neural models were trained at a single physical error rate of 2% and evaluated at other rates as well.
The speed comparison changes hardware
For the lifted-product code, the paper reports 2.37 milliseconds per shot for Triage, compared with 815 milliseconds for BP-OSD, a belief-propagation decoder paired with ordered-statistics processing. For Kasai, Triage’s reported 12.3 milliseconds per shot compares with 349 milliseconds for the Tiered algorithmic decoder. The authors calculate speedups of approximately 343× and 28×, respectively.
Those figures average processing across batches of 64 shots. Triage ran on a single NVIDIA H-series GPU; the BP-based baselines ran on a CPU, with belief propagation single-threaded and the Tiered decoder’s exact-solver stage using eight threads. The measurements therefore compare complete implementations under different hardware and parallelism conditions, rather than isolating the neural method’s algorithmic speed advantage.
Real-time performance remains untested
The authors identify the next test themselves: single-shot latency and its distribution, sustained throughput, and performance against the timing requirements of a quantum processor. They also call for evaluation on circuit-level workloads and experimental data, as well as work on GPU optimization and possible FPGA deployment. Until those measurements exist, the reported figures describe batch-processing costs in simulations—not proof that Triage can keep pace with a live error-correction cycle.
The paper’s code is available from the corresponding author upon request, and its arXiv page says the full supplementary information is expected in a later update. For now, the concrete result is a neural decoder benchmarked on two simulated qLDPC codes, with its fastest reported comparison—343×—measured against BP-OSD on a CPU.