pith. machine review for the scientific record. sign in

arxiv: 2604.12841 · v1 · submitted 2026-04-14 · 🪐 quant-ph

Recognition: 2 theorem links

· Lean Theorem

Fast and accurate AI-based pre-decoders for surface codes

Christopher Chamberland, Igor Baratta, Jan Olle, Muyuan Li, Scott Thornton

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:57 UTC · model grok-4.3

classification 🪐 quant-ph
keywords surface codequantum error correctionAI decoderpre-decodermachine learningPyMatchingfault-tolerant quantum computingnoise learning
0
0 comments X

The pith

An AI pre-decoder for surface codes performs fast local parallel corrections that remove most errors before a global decoder finishes the job, cutting runtimes to microseconds per round while lowering logical error rates.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a modular AI-based pre-decoder that handles local error correction across space and time in parallel for the surface code. This leaves a much smaller set of residual syndromes for any downstream global decoder to resolve. The approach achieves end-to-end runtimes of order one microsecond per round at large code distances on GPUs when paired with uncorrelated PyMatching, and it reduces logical error rates compared with global decoding alone. A larger trained model further improves error rates enough to beat correlated PyMatching up to distance 13. The same framework includes a noise-learning component that extracts decoding weights directly from measured syndrome statistics, allowing effective decoding when the underlying noise model is unknown or changing.

Core claim

A scalable AI pre-decoder executes local, block-wise parallel error correction on surface-code syndromes, removing the majority of physical errors before residual data reaches an arbitrary global decoder. When composed with uncorrelated PyMatching this yields O(1 μs) per-round runtimes at large distances on NVIDIA GB300 GPUs together with lower logical error rates than global decoding alone; larger models outperform correlated PyMatching up to distance 13. A separate noise-learning architecture infers graph weights from experimental syndrome statistics alone, producing performance that nearly matches or exceeds standard PyMatching in several regimes without requiring an explicit circuitnoise

What carries the argument

The AI-based pre-decoder: a neural network that predicts local corrections in parallel across space-time blocks and passes only the remaining syndrome to a global decoder.

If this is right

  • End-to-end decoding reaches O(1 μs) per round on single GPUs and can drop well below that with multiple GPUs in block-wise parallel mode.
  • Logical error rates fall below those of global decoding alone, and a larger model beats correlated PyMatching up to distance 13.
  • Purely data-driven weight estimation from syndrome statistics nearly matches uncorrelated PyMatching and exceeds correlated PyMatching in some noise regimes.
  • The modular design works with any existing or future surface-code global decoder without modification.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same local-pre-decoder pattern could be tested on other topological codes or with different global decoders to check whether the runtime and error-rate gains generalize.
  • If the noise-learning component tracks time-varying hardware noise in real time, it could support adaptive decoding on physical devices whose error rates drift.
  • Because the pre-decoder runs in parallel across many blocks, it may allow decoding latency to stay constant even as code distance grows, provided enough GPUs are available.

Load-bearing premise

The pre-decoder trained on finite-distance data continues to remove errors correctly at larger distances without creating logical errors that the global decoder cannot later fix.

What would settle it

Running the full pipeline at distance 17 or higher and measuring a logical error rate higher than that of the global decoder alone would show the pre-decoder is adding uncorrectable errors.

Figures

Figures reproduced from arXiv: 2604.12841 by Christopher Chamberland, Igor Baratta, Jan Olle, Muyuan Li, Scott Thornton.

Figure 1
Figure 1. Figure 1: FIG. 1. Example showing the syndrome density being reduced by the pre-decoder for both [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Example of a surface code patch for [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Example of a four-layer fully connected three [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. (a) Example mapping of [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Example illustrations of the computation of [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Circuit for a [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Spacelike homological equivalence convention as shown in a [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Timelike homological equivalence convention for a [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10. Timelike homological equivalence convention for a [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: FIG. 12. Architecture for learning the circuit-level noise parameters of the gates used to implement the surface code. [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: FIG. 13. Plots of per-round LER for uncorrelated PyMatching (dashed lines) vs per-round LER of a pre-decoder model followed [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: FIG. 14. Plots of the syndrome density reduction factor for models 1 and 5 as a function of the physical error rate [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: FIG. 15. Pre-decoder neural network architecture used when the global decoder employs correlated matching. The model [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: FIG. 16. Per-round LERs obtained from using pre-decoder model 6 described in Fig. [PITH_FULL_IMAGE:figures/full_fig_p023_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: FIG. 17. GPU runtime performance on an NVIDIA GB300 GPU using TensorRT with FP8 precision. (a) runtimes measurements [PITH_FULL_IMAGE:figures/full_fig_p024_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: FIG. 18. Pre-decoder runtime as a function of the batch [PITH_FULL_IMAGE:figures/full_fig_p024_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: FIG. 19. End-to-end per-round logical error rates (LER) and [PITH_FULL_IMAGE:figures/full_fig_p025_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: FIG. 20. (a) LER for correlated and uncorrelated PyMatching when using probability vectors in a detector error model (DEM) [PITH_FULL_IMAGE:figures/full_fig_p027_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: FIG. 21. LER of the surface code using the uncorrelated [PITH_FULL_IMAGE:figures/full_fig_p029_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: FIG. 22. (a) Two-dimensional graph for [PITH_FULL_IMAGE:figures/full_fig_p032_22.png] view at source ↗
read the original abstract

Fast, scalable decoding architectures that operate in a block-wise parallel fashion across space and time are essential for real-time fault-tolerant quantum computing. We introduce a scalable AI-based pre-decoder for the surface code that performs local, parallel error correction with low decoding runtimes, removing the majority of physical errors before passing residual syndromes to a downstream global decoder. This modular architecture is backend-agnostic and composes with arbitrary global decoding algorithms designed for surface codes, and our implementation is completely open source. Integrated with uncorrelated PyMatching, the pipeline achieves end-to-end decoding runtimes of order $\mathcal{O}(1 \mu\text{s})$ per round at large code distances on NVIDIA GB300 GPUs while reducing logical error rates (LERs) relative to global decoding alone. In a block-wise parallel decoding scheme with access to multiple GPUs, the decoding runtime can be reduced to well below $\mathcal{O}(1 \mu\text{s})$ per round. We observe further LER improvements by training a larger model, outperforming correlated PyMatching up to distance-13. We additionally introduce a noise-learning architecture that infers decoding weights directly from experimentally accessible syndrome statistics without requiring an explicit circuit-level noise model. We show that purely data-driven graph weight estimation can nearly match uncorrelated PyMatching and exceed correlated PyMatching in certain regimes, enabling highly-optimized decoding when hardware noise models are unknown or time-varying, as well as training pre-decoders with realistic noise models. Together, these results establish a practical, modular, and high-throughput decoding framework suitable for large-distance surface-code implementations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces a scalable AI-based pre-decoder for surface codes that performs local, parallel error correction on patches to remove the majority of physical errors before forwarding residual syndromes to a downstream global decoder such as uncorrelated or correlated PyMatching. It reports end-to-end runtimes of O(1 μs) per round at large distances on NVIDIA GB300 GPUs, LER reductions relative to global decoding alone, and further LER gains with larger models that outperform correlated PyMatching up to distance 13. A separate noise-learning module is presented that infers decoding graph weights directly from syndrome statistics without an explicit circuit-level noise model.

Significance. If the empirical results hold under rigorous validation, the modular pre-decoder architecture would represent a practical advance for real-time, high-throughput decoding in large-distance surface-code fault tolerance, by enabling block-wise parallel execution across space and time while remaining backend-agnostic. The open-source implementation and data-driven weight estimation for unknown or time-varying noise are concrete strengths that support reproducibility and hardware applicability.

major comments (2)
  1. [Results (d=13 LER comparisons and generalization statements)] The central performance claims (runtime O(1 μs) and LER reduction up to d=13) rest on the assumption that local AI corrections never produce residual syndromes whose combined weight with the pre-decoder output exceeds the global decoder's correction capability at the original code distance. No explicit failure-mode analysis, out-of-distribution test sets at d>13, or distance-extrapolation study is provided to verify that the pre-decoder does not complete logical operators or inflate error chains for unseen configurations.
  2. [Methods and experimental setup] Concrete runtime and LER numbers are reported in the abstract and results, yet the manuscript supplies no training details, validation splits, Monte Carlo sample counts, error bars, hyperparameter search procedure, or ablation studies on model size versus performance. This absence makes it impossible to determine whether the reported gains are robust or influenced by post-hoc model selection.
minor comments (2)
  1. [Figures and abstract] Figure captions and text should explicitly state the number of Monte Carlo shots and the precise definition of 'per round' when quoting O(1 μs) runtimes.
  2. [Noise-learning section] Notation for the noise-learning architecture (e.g., how syndrome statistics map to edge weights) could be formalized with a short equation or pseudocode to improve clarity for readers unfamiliar with the PyMatching interface.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and positive assessment of the work's potential significance. We address the major comments point by point below, with plans to revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Results (d=13 LER comparisons and generalization statements)] The central performance claims (runtime O(1 μs) and LER reduction up to d=13) rest on the assumption that local AI corrections never produce residual syndromes whose combined weight with the pre-decoder output exceeds the global decoder's correction capability at the original code distance. No explicit failure-mode analysis, out-of-distribution test sets at d>13, or distance-extrapolation study is provided to verify that the pre-decoder does not complete logical operators or inflate error chains for unseen configurations.

    Authors: We agree that an explicit failure-mode analysis would provide valuable additional validation. The observed LER reductions relative to global decoding alone, including outperforming correlated PyMatching up to d=13, provide empirical evidence that the pre-decoder does not systematically inflate error chains in the tested regimes. The architecture applies only local corrections within each patch, which by design targets low-weight errors and forwards residuals to the global decoder operating at the full code distance. Nevertheless, we will add a dedicated discussion of potential failure modes, including analysis of residual syndrome weights and any available out-of-distribution tests, in the revised manuscript. revision: partial

  2. Referee: [Methods and experimental setup] Concrete runtime and LER numbers are reported in the abstract and results, yet the manuscript supplies no training details, validation splits, Monte Carlo sample counts, error bars, hyperparameter search procedure, or ablation studies on model size versus performance. This absence makes it impossible to determine whether the reported gains are robust or influenced by post-hoc model selection.

    Authors: We acknowledge that these experimental details are missing from the current version and agree they are necessary for assessing robustness. In the revised manuscript we will include full training details, validation and test splits, the number of Monte Carlo samples used for each LER estimate, error bars computed across independent runs, the hyperparameter search procedure, and ablation studies examining model size versus both LER improvement and runtime. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical training and composition with external decoders

full rationale

The paper's central results rest on training an AI pre-decoder on simulated error data and measuring empirical LER/runtime improvements when composed with PyMatching (uncorrelated or correlated). No equations, uniqueness theorems, or self-citations are invoked to derive the performance claims; the reported gains are direct experimental outcomes on held-out configurations up to distance 13. The noise-learning component similarly infers weights from syndrome statistics without reducing to a fitted quantity defined by the target result. All load-bearing steps remain falsifiable via independent simulation or hardware runs and do not collapse to the inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central performance claims rest on a neural-network model whose weights are fitted to simulated or experimental syndrome data; the abstract does not enumerate the exact number or form of these parameters.

free parameters (1)
  • neural network weights
    Parameters of the AI pre-decoder and noise-learning model are trained on data and not derived from first principles.
axioms (1)
  • domain assumption Standard surface-code stabilizer formalism and syndrome extraction circuit assumptions hold.
    The pre-decoder operates on syndromes produced by the usual surface-code measurement circuits.

pith-pipeline@v0.9.0 · 5587 in / 1307 out tokens · 20165 ms · 2026-05-10T15:57:30.060463+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Real-time Surface-Code Error Correction Using an FPGA-based Neural-Network Decoder

    quant-ph 2026-05 unverdicted novelty 6.0

    An FPGA-based neural-network decoder achieves 550 ns deterministic closed-loop latency for real-time distance-3 surface code error correction on a superconducting processor, matching offline decoding performance.

Reference graph

Works this paper leans on

51 extracted references · 12 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    Sample a base error ratep base from a log-uniform dis- tribution over [ pmin, pmax], then derive the 25 noise parameters with location-specific random multipliers and random Pauli-type distributions (see Section A 1)

  2. [2]

    Generate B independent syndrome samples at the train- ing distancedusing the sampled noise model

  3. [3]

    For each sample k, compute zk = MLP(GAP(CNN(xk)))

  4. [4]

    Average logits: ¯z = 1 B P k zk, then ˆp = BoundedLogSpace(¯z) via Eq. (61)

  5. [5]

    Compute ˆPej =E j( ˆp) and ˆHk =H k( ˆp)

  6. [6]

    PyMatching after model X

    Minimize L = Ledge + Lhyper and backpropagate through the differentiable formulas. The hierarchical noise sampling ensures diverse train- ing data spanning multiple orders of magnitude while maintaining physically reasonable correlations between parameters. E. Inference strategy At inference time, the trained network is applied to syndrome data produced b...

    2048

  7. [7]

    • Measurement errors (2): PmX for X-basis mea- surement,P mZ forZ-basis measurement

    Notation and methodology The circuit-level noise model is parameterized by 25 probabilities: • State preparation errors (2): PSX for |+⟩ prepa- ration,P SZ for|0⟩preparation. • Measurement errors (2): PmX for X-basis mea- surement,P mZ forZ-basis measurement. • Idle errors during CNOT layers (3): P (X) idle,CNOT, P (Y) idle,CNOT, P (Z) idle,CNOT for singl...

  8. [8]

    Arise from data qubit errors

    Edge classification The matching graph contains four categories of edges: • Spacelike edges: Connect different stabilizers within the same measurement round. Arise from data qubit errors. • Timelike edges: Connect the same stabilizer across adjacent measurement rounds. Arise from ancilla/measurement errors. • Diagonal edges: Connect different stabilizers ...

  9. [9]

    These formulas detect ZandYerrors on data qubits

    X-stabilizer graph edge formulas We provide the verified edge probability formulas for the X-stabilizer matching graph. These formulas detect ZandYerrors on data qubits. a. Spacelike edges TypeP (X) S1 : P (X) S1 = Mh P (Y Y) CX +P (ZZ) CX , P (IZ) CX +P (XZ) CX , P (Z) I , P (Z) I , P (Y Z) CX +P (ZY) CX , P (IY) CX +P (XY) CX , P (Y) I , P (Y) I i .(A4)...

  10. [10]

    Similar to the X-graph, it has 18 edge types: 3 spacelike (S1–S3), 4 timelike (T1–T4), 5 diagonal (D1–D5), and 6 boundary (B1–B6)

    Z-stabilizer graph edge formulas The Z-stabilizer matching graph detectsX and Y errors on data qubits. Similar to the X-graph, it has 18 edge types: 3 spacelike (S1–S3), 4 timelike (T1–T4), 5 diagonal (D1–D5), and 6 boundary (B1–B6). The explicit formulas are obtained from the X-stabilizer formulas above by replacing all Z-type Paulis with X-type Paulis, ...

  11. [11]

    The methodology is:

    Summary and verification The formulas were derived by systematically tracing error propagation through the syndrome extraction circuit for each possible Pauli error at each fault location. The methodology is:

  12. [12]

    For each fault location (CNOT, idle, state prepara- tion), activate a single Pauli error

  13. [13]

    Generate the detector error model (DEM) using Stim

  14. [14]

    Identify which DEM patterns contain the target edge’s detector pair

  15. [15]

    Group contributions by pattern and sum Paulis from the same location. 35

  16. [16]

    The formulas aredistance-independent: the same formulas apply identically for d = 5, 7, 9, 11, 13 and be- yond

    XOR-combine all pattern contributions to get the final formula. The formulas aredistance-independent: the same formulas apply identically for d = 5, 7, 9, 11, 13 and be- yond. This is because edge probabilities depend only on local stabilizer geometry, not global code size. Only thecountof each edge type changes with distance. For example, at d = 5 the X-...

  17. [17]

    P. W. Shor, Scheme for reducing decoherence in quantum computer memory, Phys. Rev. A52, R2493 (1995)

    1995

  18. [18]

    Knill, R

    E. Knill, R. Laflamme, and L. Viola, Theory of quantum error correction for general noise, Phys. Rev. Lett.84, 2525 (2000)

    2000

  19. [19]

    Chao and B

    R. Chao and B. W. Reichardt, Quantum error correction with only two extra qubits, Phys. Rev. Lett.121, 050502 (2018)

    2018

  20. [20]

    Chamberland and M

    C. Chamberland and M. E. Beverland, Flag fault-tolerant error correction with arbitrary distance codes, Quantum 2, 53 (2018)

    2018

  21. [21]

    Chao and B

    R. Chao and B. W. Reichardt, Flag fault-tolerant er- ror correction for any stabilizer code, PRX Quantum1, 010302 (2020)

    2020

  22. [22]

    Chamberland and A

    C. Chamberland and A. W. Cross, Fault-tolerant magic state preparation with flag qubits, Quantum3, 143 (2019)

    2019

  23. [23]

    Chamberland and K

    C. Chamberland and K. Noh, Very low overhead fault- tolerant magic state preparation using redundant ancilla encoding and flag qubits, npj Quantum Information6, 91 (2020)

    2020

  24. [24]

    B. M. Terhal, Quantum error correction for quantum memories, Rev. Mod. Phys.87, 307 (2015)

    2015

  25. [25]

    Chamberland, L

    C. Chamberland, L. Goncalves, P. Sivarajah, E. Peterson, and S. Grimberg, Techniques for combining fast local decoders with global decoders under circuit-level noise, Quantum Science and Technology8, 045011 (2023)

    2023

  26. [26]

    Skoric, D

    L. Skoric, D. E. Browne, K. M. Barnes, N. I. Gillespie, and E. T. Campbell, Parallel window decoding enables scalable fault tolerant quantum computation, Nature Communica- tions14, 7040 (2023)

    2023

  27. [27]

    X. Tan, F. Zhang, R. Chao, Y. Shi, and J. Chen, Scalable Surface-Code Decoders with Parallelization in Time, PRX Quantum4, 040344 (2023), arXiv:2209.09219 [quant-ph]

    work page arXiv 2023

  28. [28]

    Chamberland and E

    C. Chamberland and E. T. Campbell, Universal quantum computing with twist-free and temporally encoded lattice surgery, PRX Quantum3, 010331 (2022)

    2022

  29. [29]

    New magic state distillation factories optimized by temporally encoded lattice surgery , publisher =

    P. Prabhu and C. Chamberland, New magic state dis- tillation factories optimized by temporally encoded lat- tice surgery, arXiv e-prints , arXiv:2210.15814 (2022), arXiv:2210.15814 [quant-ph]

    work page arXiv 2022

  30. [30]

    Chamberland and P

    C. Chamberland and P. Ronagh, Deep neural decoders for near term fault-tolerant experiments, Quantum Science and Technology3, 044002 (2018)

    2018

  31. [31]

    Baireuther, M

    P. Baireuther, M. D. Caio, B. Criger, C. W. J. Beenakker, and T. E. O’Brien, Neural network decoder for topological color codes with circuit level noise, New Journal of Physics 21, 013003 (2019)

    2019

  32. [32]

    Bausch, A

    J. Bausch, A. W. Senior, F. J. H. Heras, T. Edlich, A. Davies, M. Newman, C. Jones, K. Satzinger, M. Y. Niu, S. Blackwell, G. Holland, D. Kafri, J. Atalaya, C. Gidney, D. Hassabis, S. Boixo, H. Neven, and P. Kohli, Learn- ing high-accuracy error decoding for quantum processors, Nature635, 834 (2024)

    2024

  33. [33]

    A. W. Senior, T. Edlich, F. J. H. Heras, L. M. Zhang, O. Higgott, J. S. Spencer, T. Applebaum, S. Black- well, J. Ledford, A. ˇZemgulyt˙ e, A. ˇZ´ ıdek, N. Shutty, A. Cowie, Y. Li, G. Holland, P. Brooks, C. Beattie, M. Newman, A. Davies, C. Jones, S. Boixo, H. Neven, P. Kohli, and J. Bausch, A scalable and real-time neural decoder for topological quantum ...

    work page arXiv 2025

  34. [34]

    Zhang, Z

    K. Zhang, Z. Yi, S. Guo, L. Kong, S. Wang, X. Zhan, T. He, W. Lin, T. Jiang, D. Gao, Y. Zhang, F. Liu, F. Zhang, Z. Ji, F. Chen, and J. Chen, Learning to De- code in Parallel: Self-Coordinating Neural Network for Real-Time Quantum Error Correction, arXiv e-prints , arXiv:2601.09921 (2026), arXiv:2601.09921 [quant-ph]

    work page arXiv 2026

  35. [35]

    A. G. Fowler and C. Gidney, Low overhead quan- tum computation using lattice surgery, arXiv e-prints , arXiv:1808.06709 (2018), arXiv:1808.06709 [quant-ph]

    work page arXiv 2018

  36. [36]

    arXiv preprint arXiv:1808.02892 , year=

    D. Litinski, A Game of Surface Codes: Large-Scale Quan- tum Computing with Lattice Surgery, Quantum3, 128 (2019), 1808.02892

    work page arXiv 2019

  37. [37]

    Chamberland and E

    C. Chamberland and E. T. Campbell, Circuit-level pro- tocol and analysis for twist-based lattice surgery, Phys. Rev. Research4, 023090 (2022)

    2022

  38. [38]

    Gicev, L

    S. Gicev, L. C. L. Hollenberg, and M. Usman, A scalable and fast artificial neural network syndrome decoder for surface codes, Quantum7, 1058 (2023)

    2023

  39. [39]

    Fully convolutional 3D neural network decoders for surface codes with syndrome circuit noise

    S. Gicev, L. C. L. Hollenberg, and M. Usman, Fully convo- lutional 3D neural network decoders for surface codes with syndrome circuit noise, arXiv e-prints , arXiv:2506.16113 (2025), arXiv:2506.16113 [quant-ph]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  40. [40]

    arXiv preprint arXiv:2509.03954 , year=

    K. Zhang, J. Xu, F. Zhang, L. Kong, Z. Ji, and J. Chen, LATTE: A Decoding Architecture for Quantum Comput- ing with Temporal and Spatial Scalability, arXiv e-prints , arXiv:2509.03954 (2025), arXiv:2509.03954 [quant-ph]

    work page arXiv 2025

  41. [41]

    & Campbell, E

    L. Caune, B. Reid, J. Camps, and E. Campbell, Belief propagation as a partial decoder (2023), arXiv:2306.17142 [quant-ph]

    work page arXiv 2023

  42. [42]

    Dennis, A

    E. Dennis, A. Kitaev, A. Landahl, and J. Preskill, Topolog- ical quantum memory, Journal of Mathematical Physics 43, 4452 (2002)

    2002

  43. [43]

    A. G. Fowler, M. Mariantoni, J. M. Martinis, and A. N. Cleland, Surface codes: Towards practical large-scale quantum computation, Phys. Rev. A86, 032324 (2012)

    2012

  44. [44]

    Tomita and K

    Y. Tomita and K. M. Svore, Low-distance surface codes under realistic quantum noise, Phys. Rev. A90, 062320 (2014)

    2014

  45. [45]

    Higgott, Pymatching: A python package for de- coding quantum codes with minimum-weight perfect matching, ACM Transactions on Quantum Computing3, 36 10.1145/3505637 (2022)

    O. Higgott, Pymatching: A python package for de- coding quantum codes with minimum-weight perfect matching, ACM Transactions on Quantum Computing3, 36 10.1145/3505637 (2022)

    work page doi:10.1145/3505637 2022

  46. [46]

    Litinski and F

    D. Litinski and F. v. Oppen, Lattice surgery with a twist: Simplifying Clifford gates of surface codes, Quantum2, 62 (2018)

    2018

  47. [47]

    Delfosse and N

    N. Delfosse and N. H. Nickerson, Almost-linear time de- coding algorithm for topological codes, Quantum5, 595 (2021)

    2021

  48. [48]

    Edmonds, Paths, trees, and flowers, Canadian Journal of Mathematics17, 449–467 (1965)

    J. Edmonds, Paths, trees, and flowers, Canadian Journal of Mathematics17, 449–467 (1965)

    1965

  49. [49]

    Higgott and C

    O. Higgott and C. Gidney, Sparse Blossom: correcting a million errors per core second with minimum-weight matching, Quantum9, 1600 (2025)

    2025

  50. [50]

    Distilling the Knowledge in a Neural Network

    G. Hinton, O. Vinyals, and J. Dean, Distilling the knowl- edge in a neural network, arXiv preprint arXiv:1503.02531 (2015)

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  51. [51]

    S. A. Caldwell, M. Khazraee, E. Agostini, T. Las- siter, C. Simpson, O. Kahalon, M. Kanuri, J.-S. Kim, S. Stanwyck, M. Li, J. Olle, C. Chamberland, B. Howe, B. Schmitt, J. G. Lietz, A. McCaskey, J. Ye, A. Li, A. B. Magann, C. I. Ostrove, K. Rudinger, R. Blume- Kohout, K. Young, N. E. Miller, Y. Xu, G. Huang, I. Sid- diqi, J. Lange, C. Zimmer, and T. Humbl...

    work page arXiv 2025