Neural triage decoding speeds correction for large qLDPC codes
Fast and accurate neural triage decoding of ultra-high-rate quantum LDPC codes · arXiv:2610.40490
In plain words
High-rate quantum codes can store many logical qubits in one block, but decoding must find corrections accurately and quickly. Neural Triage Decoding uses a neural network to suggest corrections, then spends more computation on cases that still fail a check.
In simulations of two code families, it had lower estimated logical error rates and lower amortized decoding times than the respective algorithmic decoders. The tests used simulated quantum-memory noise, so they do not establish real-time performance on quantum hardware.
The research
High-rate quantum codes can encode many logical qubits in a block of roughly a thousand physical data qubits, but realizing that advantage depends on decoding accurately and quickly. Can a neural decoder handle these codes without the computational cost associated with stronger algorithmic decoders?
Triage pairs a local neural decoder with repeated updates and a conditional search to resolve syndrome failures on high-rate quantum codes.
Triage uses a local neural network to propose corrections, repeats its updates when measured checks still show a mismatch (a residual syndrome), and searches alternative candidates only for shots that remain unresolved. This design directs more computation toward cases the first stages do not resolve.
In phenomenological quantum-memory simulations at p=0.015, Triage had lower estimated logical error rates than the respective algorithmic decoders on LP20 and Kasai.
Its measured full-decoder time was also lower when amortized over batches of 64 shots, supporting accuracy and lower cost in these tested settings.
These results come from phenomenological quantum-memory simulations. The timing comparison measures amortized costs at batch size 64 on different GPU and CPU hardware; it does not establish single-shot latency or sustained real-time operation.
What the evidence shows
- The decoding challenge
Neural Triage Decoding (Triage) is a neural decoder tested on high-rate quantum low-density parity-check codes, which can encode many logical qubits in one block. The practical question is whether a neural decoder can combine accurate corrections with manageable computation at this scale.
Neural decoders may leave corrections inconsistent with the measured checks, the pattern of check outcomes called the syndrome, while algorithmic decoders can be computationally costly.
- Computation follows unresolved cases
Triage starts with a local message-passing network that proposes corrections from the measured syndrome. If checks remain unsatisfied, it applies the same trained network again to the residual syndrome, for up to four outer iterations.
Shots still unresolved enter a conditional search for alternative corrections. Candidates are checked against the original syndrome and ranked using the original noise probabilities, so additional computation is reserved for unresolved shots.
- A bounded explanation for residuals
The authors establish a theoretical reason syndrome inconsistency can persist: under stated assumptions, exact bitwise most-likely decisions have a probability of at least order p⁴ of violating the syndrome constraint for the LP and Kasai code matrices.
This theorem concerns exact marginal estimates under an independent X/Z noise model. Separately, one-pass NeuralMP measurements of nonzero residual syndromes are described as consistent with that mechanism; they do not extend the theorem to the approximate neural decoder.
- Results on two code families
In simulated quantum-memory tests at p=0.015, Triage had lower estimated logical error rates than the respective algorithmic baseline on each code: BP-OSD for LP20 and BP-relay-MLE Tiered for Kasai. Its full-decoder time was also lower when amortized across batches of 64 shots.
Those timing results compare a GPU implementation with CPU baselines, and the reported error rates and comparisons apply separately to the two code families.
- Prior mismatch has a narrow scope
A separate test perturbed the probabilities given to the four-iteration neural frontend while keeping physical noise fixed. At the largest tested mismatch, its block-error-rate ratios relative to matched probabilities were about 1.01 on LP20 and 0.98 on Kasai.
This supports resilience to the tested prior perturbations, not evidence that Triage follows changing physical noise during operation.
Deep
- Original paper authors
- Hong-Ye Hu, Zhenlai Wang, Binzhao Luo, Tao Xiang, Chen Zhao
- Submitted to arXiv
- License
- arXiv non-exclusive