Fast and accurate neural triage decoding of ultra-high-rate quantum LDPC codes

AI summary

Overview

Research area: Quantum error correction — specifically classical decoding algorithms for high-rate quantum low-density parity-check (qLDPC) codes, combining graph neural network message passing with adaptive, syndrome-guided search.

Technical level: Advanced. The paper assumes familiarity with stabilizer codes, detector error models, Tanner graphs, belief propagation, and ordered-statistics decoding, and includes a theorem on the failure of exact bitwise marginal decoding.

Scope (one sentence): The paper introduces and benchmarks "neural triage decoding," a three-stage decoder (neural message passing, iterative residual decoding, rare diversification) for lifted-product and Kasai qLDPC codes encoding 148 and 580 logical qubits, reporting logical error rates comparable to or better than BP-OSD and BP-relay-MLE (Tiered) with roughly one to two orders of magnitude lower amortized decoding latency.

What This Paper Is About

High-rate qLDPC codes can pack hundreds of logical qubits into blocks of roughly a thousand physical data qubits, but exploiting that overhead saving requires a classical decoder that is both accurate and fast enough to keep up with syndrome generation. Accurate algorithmic decoders such as BP-OSD and BP-relay-MLE exist but are computationally expensive, and neural decoders had not previously been demonstrated in the regime of hundreds of logical qubits per block. The paper's goal is a single decoder framework that is accurate, fast, and portable across code families in this ultra-high-rate regime.

Key Contributions

1. A three-stage "Triage" decoding framework. A local neural message-passing network (NeuralMP) produces fast corrections; iterative decoding re-applies the same network to the residual syndrome; a rarely invoked diversification stage (syndrome flips, prior perturbations, symmetric permutations) searches for syndrome-consistent corrections. Computation is allocated adaptively according to syndrome consistency rather than uniformly across shots.

2. A theoretical account of why single-pass marginal decoding must fail. The authors prove (Theorem 1) that the exact bitwise marginal (MAP) decoder has a syndrome-inconsistency probability bounded below by $\Omega(p^{\kappa(\nu)})$ for codes with minimum qubit degree $\nu \geq 2$, no 4-cycles, and no low-weight kernel vectors, under phenomenological noise with $r \geq 5$ rounds. For the LP and Kasai codes used here ($\nu = 3$), this gives $P_{\mathrm{inc}}(p) = \Omega(p^4)$ — a distance-independent error floor — which motivates the iterative and diversification stages.

3. Demonstration on two ultra-high-rate code families at scale. Benchmarks on the lifted-product $[[1122,148,d \leq 20]]$ code (13.2% encoding rate) and the Kasai $[[1152,580,d \leq 12]]$ code (50.3% encoding rate), i.e. hundreds of logical qubits per block, a regime where a 500-logical-qubit block has $2^{500} \approx 10^{150}$ logical classes.

4. A failure-aware training scheme and a robustness result. Training couples online error-weight-conditioned simulation with a dynamically updated mistake bank of logical failures plus symmetry-augmented replay, supervised jointly on fault predictions, logical observables, and syndrome consistency. The trained decoder is also shown to be resilient to mismatched error priors.

Main Findings

• State-of-the-art accuracy at $p = 0.015$: estimated per-qubit-per-round logical error rates of $9.33 \times 10^{-11}$ on LP20 and $5.08 \times 10^{-10}$ on Kasai, compared with $4.43 \times 10^{-10}$ for BP-OSD and $1.76 \times 10^{-9}$ for the BP-relay-MLE Tiered decoder respectively. The abstract summarizes this as roughly $5 \times 10^{-10}$ at $p = 1.5%$ on the Kasai code.

• Large latency and throughput gains: full-decoder amortized cost at batch size 64 is 2.37 ms and 12.3 ms per shot on a single NVIDIA H-series GPU, versus CPU baseline means of 815 ms (BP-OSD) and 349 ms (Tiered) — approximately $343\times$ and $28\times$ speedups in these settings, described overall as one to two orders of magnitude.

• Marginal decoding has an irreducible inconsistency floor: the proof shows that when a syndrome has two equally likely minimum-weight explanations on complementary halves of a low-weight space-time degeneracy, exact marginals sit near $1/2$ and bitwise decisions can violate parity. Single-pass NeuralMP residual-syndrome probabilities are consistent with this mechanism, following a $p^4 = p^{w_{\min}/2}$ reference scaling ($w_{\min}$ = minimal weight of space-time degeneracy).

• Iteration and diversification close the gap: residual-guided iteration with at most $K_{\max} = 4$ outer passes, followed where needed by syndrome flips guided by marginal uncertainty and repeated corrections, resolves the remaining shots. The paper gives a worked argument that a single-qubit syndrome flip changes the relative weights of complementary candidates from $w/2$ and $w/2$ to $w/2 - 1$ and $w/2 + 1$, a factor $[(1-p)/p]^2$ in likelihood, without altering their relative likelihood in the original problem.

• Robustness to mismatched priors: with priors perturbed as $p_f \to p_f(1 + \sigma_r \mathcal{N}(0,1))$, the four-iteration NeuralMP frontend keeps the block-LER ratio near unity — approximately 1.01 for LP20 and 0.98 for Kasai at $\sigma_r = 0.5$ — whereas algorithmic baselines rise by roughly 1.6–2.1 times relative to matched-prior decoding.

• Single-rate training generalizes: both codes are trained at a single physical error rate of $p = 2%$ and evaluated across previously unseen error rates.

• Favorable scaling of the architecture: one NeuralMP evaluation costs $\mathcal{O}(R[E\chi + N\chi^2])$ for $N = m + F$ nodes and $E$ edges, linear in $N$ when $E = \mathcal{O}(N)$ at fixed $R, \chi$, versus $\mathcal{O}(N^2)$ for attention-based global decoders.

Methodology in Plain English

The decoder is built on a detector error model: a sparse bipartite "Tanner graph" linking fault mechanisms to detectors, with a parity-check matrix $H$ and a logical-observable matrix $L$. NeuralMP encodes each detector's binary outcome and each mechanism's prior probability into learned state vectors, then runs $R$ rounds of alternating updates (fault-to-detector, detector-to-fault) using small multilayer perceptrons shared across nodes of the same type within a round. Nodes of the same type update in parallel, which suits GPU execution. A linear readout plus sigmoid produces one error probability per mechanism.

Because independently thresholded marginals can produce a correction that does not reproduce the measured syndrome, the decoder repeatedly applies the same network to the leftover ("residual") syndrome, sampling proposal bits and XOR-accumulating them, for up to four outer iterations. Shots still unresolved enter diversification: the syndrome is deliberately flipped on a small guided set of mechanisms, or the priors are perturbed, or symmetry-related relabelings of the graph are used so the network sees a different view; candidates are mapped back, checked for syndrome consistency, and ranked by negative log-likelihood under the original priors. If the budget runs out, the decoder falls back to the iterative-stage correction and the shot counts as unresolved.

Training uses fresh online simulation mixed with a replay bank of hard examples that are added on logical failure and removed after repeated successes, augmented by code symmetries that permute faults and syndromes together. The loss is the unweighted sum of three terms with unit coefficients: a binary cross-entropy fault loss with positive-class weight eight, a logical loss on a differentiable Bernoulli-factorized parity estimate, and a syndrome-consistency penalty. The deployed Kasai network uses hidden dimension $\chi = 64$ and $R = 80$ MP-rounds with round-specific parameters; sampling temperature is $T = 1$.

Benchmarks use phenomenological noise with a final perfect syndrome extraction: five noisy rounds with independent $X$ and $Z$ errors at probability $p$ and measurement flips at $p$ for LP20; eight noisy rounds with depolarizing noise ($X$, $Y$, $Z$ each at $p/3$) and measurement flips at $4p/3$ for Kasai. Statistics use roughly 0.04–40.96 million shots per point for LP20 and 0.24–20.48 million for Kasai, with 95% Clopper–Pearson intervals; latency error bars are 95% bootstrap intervals from 2,000 resamples, with CPU baselines on one thread except the eight-thread COPT stage.

Why This Matters

Impact on research: The paper supplies both a theoretical reason why marginal neural decoders plateau — a distance-independent $\Omega(p^4)$ inconsistency floor for $\nu=3$ codes — and a practical architecture that escapes it. It also opens the ultra-high-rate regime (hundreds of logical qubits per block) to neural decoding, which had previously only been shown for blocks encoding tens of logical qubits.

Real-world applications:

• Real-time decoding pipelines for fault-tolerant quantum processors, where decoder latency must match syndrome-extraction timing.
• Hardware-aware control stacks that need decoders tolerant of drifting or miscalibrated noise, given the demonstrated robustness to prior mismatch.
• Deployment on GPU or FPGA accelerators co-designed with quantum hardware, where the linear-in-$N$ message-passing cost matters.
• Reducing qubit overhead budgets for early fault-tolerant machines by making high-rate codes practically decodable.

Industry relevance: Decoder throughput is a first-order constraint on any superconducting, neutral-atom, or trapped-ion fault-tolerance roadmap. A decoder that is code-family-agnostic, trainable at one error rate, and several hundred times faster than CPU algorithmic baselines addresses a recognized bottleneck for systems builders.

Future Directions

• Bake symmetries into the network. The authors suggest incorporating symmetric permutations directly into NeuralMP through symmetry-respecting parameter sharing (as in Cascade) to shrink network size and inference cost, rather than using symmetries only in diversification.
• Systematically diagnose decoding failures. Relating the effectiveness of syndrome flips, prior perturbations, and permutations to graph structure and degeneracy could explain why the proposals work differently on LP20 versus Kasai and guide per-shot budget allocation.
• Automated self-improvement. The discussion proposes using AI agents and meta-learning to analyze failures and propose architectural, training, or diversification changes, evaluated against logical error rate and total decoding cost.
• Circuit-level and experimental validation. Extending from phenomenological quantum memory to circuit-level logical gadgets and algorithms or to experimental data, plus measuring single-shot latency distributions, sustained throughput, and FPGA deployment — none of which are reported here — are stated as central next steps. Coupling adaptive decoding to hardware self-calibration in a closed loop is also proposed but not demonstrated.

Target Audience

Researchers and engineers working on quantum error correction and fault-tolerant architectures, particularly those building classical decoders for qLDPC codes; machine-learning researchers applying graph neural networks to decoding problems; and quantum hardware teams assessing whether high-rate codes with hundreds of logical qubits per block are compatible with their latency budgets. Readers without a background in stabilizer codes, detector error models, and BP-based decoding will find the theoretical sections and benchmark comparisons difficult.

Authors’ abstract

Quantum error correction is essential for fault-tolerant quantum computing. High-rate quantum low-density parity-check (qLDPC) codes have emerged as a promising approach to reducing the substantial physical qubit overhead of quantum error correction, enabling encoding rates exceeding $1/2$ in blocks of approximately one thousand physical data qubits. However, their practical implementation requires decoders that combine high accuracy with low latency to keep pace with quantum hardware. Existing decoders of high-rate qLDPC codes can achieve high accuracy but often incur substantial computational latency, whereas neural network decoders have yet to be demonstrated for high-rate qLDPC codes. Here, we introduce \emph{neural triage decoding}, a framework that uses syndrome consistency to adaptively allocate neural computation. A local neural message-passing network rapidly predicts error corrections, additional iterations reduce residual syndrome violations, and a rarely invoked diversification stage searches for consistent corrections to the remaining shots. We demonstrate that the \texttt{Triage} decoder can decode ultra-high-rate qLDPC codes encoding a few hundred logical qubits, including lifted-product and Kasai codes, achieving state-of-the-art accuracy comparable to the near optimal algorithmic decoders while reducing decoding latency by approximately one to two orders of magnitude. Our framework is inspired by a rigorous analysis of the exact bitwise marginal decoder, revealing a distance-independent syndrome-inconsistent error floor. In addition, we show the robustness of neural network decoders to mismatched error priors. Our results establish adaptive neural decoding as a promising route towards fast, accurate, and scalable quantum error correction in the near term.